G Model

PRONEU 1500 No. of Pages 25

Progress in Neurobiology xxx (2017) xxx–xxx

Contents lists available at ScienceDirect

Progress in Neurobiology
journal homepage: www.elsevier.com/locate/pneurobio

Review article

Uncertainty and stress: Why it causes diseases and how it is mastered

by the brain
Achim Petersa,* , Bruce S. McEwenb , Karl Fristonc
a
Medical Clinic 1, Endocrinology & Diabetes, University of Luebeck, Luebeck, Germany
b
Harold and Margaret Milliken Hatch Laboratory of Neuroendocrinology, The Rockefeller University, New York, NY, USA
c
The Wellcome Trust Centre for Neuroimaging, University College London, London, UK

A R T I C L E I N F O A B S T R A C T

Article history:
Received 13 March 2017 The term ‘stress’ – coined in 1936 – has many definitions, but until now has lacked a theoretical
Received in revised form 22 May 2017 foundation. Here we present an information-theoretic approach – based on the ‘free energy principle’ –
Accepted 24 May 2017 defining the essence of stress; namely, uncertainty. We address three questions: What is uncertainty?
Available online xxx What does it do to us? What are our resources to master it? Mathematically speaking, uncertainty is
entropy or ‘expected surprise’. The ‘free energy principle’ rests upon the fact that self-organizing
Keywords: biological agents resist a tendency to disorder and must therefore minimize the entropy of their sensory
Allostatic load states. Applied to our everyday life, this means that we feel uncertain, when we anticipate that outcomes
Atherosclerosis
will turn out to be something other than expected – and that we are unable to avoid surprise. As all
Attention
cognitive systems strive to reduce their uncertainty about future outcomes, they face a critical constraint:
Bayesian Brain
Brain energy metabolism Reducing uncertainty requires cerebral energy. The characteristic of the vertebrate brain to prioritize its
Learning own high energy is captured by the notion of the ‘selfish brain’. Accordingly, in times of uncertainty, the
mortality selfish brain demands extra energy from the body. If, despite all this, the brain cannot reduce uncertainty,
Selfish Brain a persistent cerebral energy crisis may develop, burdening the individual by ‘allostatic load’ that
Stress definition contributes to systemic and brain malfunction (impaired memory, atherogenesis, diabetes and
Stress habituation subsequent cardio- and cerebrovascular events). Based on the basic tenet that stress originates from
Uncertainty uncertainty, we discuss the strategies our brain uses to avoid surprise and thereby resolve uncertainty.
Variational free energy
© 2017 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND
license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ...... .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
1.1. What is uncertainty? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ...... .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
1.2. What are our resources to master uncertainty? . . . . . . . . . . ...... .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
1.3. What happens when uncertainty cannot be resolved? . . . . ...... .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
2. The Bayesian brain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ...... .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
2.1. Perceptual inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ...... .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
2.1.1. Two modalities of Bayesian updating . . . . . . . . . . . ...... .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
2.1.2. Precision and attention . . . . . . . . . . . . . . . . . . . . . . ...... .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
2.1.3. The anatomical correlates of hierarchical predictive coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
2.2. Active inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ...... .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
2.2.1. Goal-directed-decision making . . . . . . . . . . . . . . . . ...... .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
2.2.2. The degree of uncertainty about what to do next . ...... .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00

Abbreviations: ACC, anterior cingulate cortex; GABA, g-aminobutyric acid; GLUT1, glucose transporter 1; GLUT4, glucose transporter 4; GR, glucocorticoid receptors; HPA,
hypothalamus pituitary adrenal axis; KL, Kullback Leibler divergence; L1, layer 1; LC, locus coeruleus; LTP, long-term potentiation; LTD, long-term depression; MR,
mineralocorticoid receptors; NE, norepinephrine; pre-SMA, pre-supplementary motor area; OFC, orbitofrontal cortex; PFC, prefrontal cortex; SNS, sympathetic nervous
system; TrkB, tropomyosin receptor kinase B; vmPFC, ventromedial prefrontal cortex.
* Corresponding author.
E-mail addresses: achim.peters@uksh.de (A. Peters), mcewen@mail.rockefeller.edu (B.S. McEwen), k.friston@ucl.ac.uk (K. Friston).

http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
0301-0082/© 2017 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

2 A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx

3. Mastering uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.1. Attention – the procurement of more precise sensory information for Bayesian updating . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.1.1. Noradrenergic regulation of presynaptic glutamate release . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.1.2. Energetic constrains on information transmission . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.1.3. Prediction errors are encoded with higher precision during stress . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.1.4. NE increases the precision with which prediction errors induce Bayesian updating . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.1.5. The ‘selfish brain’ provides the energy for increasing precision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.2. Learning – updating during and after stress . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.2.1. Glucocorticoids gate the time window for cortical plasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.2.2. Stress and surprisal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.2.3. Functional plasticity at the apical tuft . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.2.4. Glucocorticoids gate learning of prediction-error precision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.2.5. Glucocorticoids gate learning of the generative model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.2.6. Structural plasticity at the apical tuft . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.3. Habituation – updating of goal states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.3.1. Habituation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.3.2. Habituators can tolerate stress . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4. Allostatic load . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4.1. The plasticity and vulnerability of the brain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4.2. Non-habituators are fully exposed to toxic stress . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4.3. Comparison of mortality among habituators and non-habituators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
5. Updating the ‘stress definition’ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
6. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00

1. Introduction information theory is cast in terms of the probability theory from

which Bayes' rule is derived. In 1948, the mathematician, electrical
Individuals, who feel threatened by changes in the external engineer, and cryptographer Claude Shannon founded information
environment or their internal body milieu, may find themselves theory in his seminal work (Shannon, 1948). It was he who stated
confronted with the question that ‘information’ is ‘reduction of uncertainty’.
‘What strategy should I select to safeguard my future physical, The Bayesian Brain is essentially a theoretical construct that can
mental and social wellbeing? (Question 1) be traced back to the students of Plato through to the philosophy of
Kant. It inherits much from the writings of people like Helmholtz in
“Stress” arises in those people who are uncertain about the
the 19th century and people like Richard Gregory in the 20th
answer. Thus, stressed individuals lack control. By placing
century. Geoffrey Hinton and colleagues formalized these ideas in
emphasis on the notion of ‘uncertainty’, we have recently proposed
the 1980s. One of us has imported these ideas into the biological
this novel definition of the term ‘stress’ (Peters and McEwen,
sciences in the form of a ‘free energy principle’ (Friston et al.,
2015).
2006). According to that concept, all cognitive or biological
How does our brain make predictions about the world when it
systems are driven to minimize an information-theoretic quantity
only has access to small fragments of it? Apparently, the brain
known as ‘free energy’. More precisely, this quantity is ‘variational
makes use of sophisticated statistical methods. By operating with
free energy’ and differs from ‘thermodynamic free energy’, though
probabilities, the brain is able to make predictions of future
the two are mathematically equivalent. Free energy, defined in the
outcomes even under conditions of uncertainty. How exactly these
current article as the information-theoretic quantity, bounds
procedures are executed is one of the most fascinating questions in
surprise, and is conceived as the difference between an organism’s
neuroscience (Friston, 2010).
predictions about its sensory inputs (embodied in its internal
model of the world) and the sensations it actually encounters.
1.1. What is uncertainty?
To put it concisely, reducing ‘free energy’ inevitably reduces
‘surprise’ – as measured by a violation of predictions. In exactly the
We begin with a brief summary of how mathematicians deal
same vein, reducing ‘expected free energy’ inevitably reduces
with the problem of uncertainty. In the 1740’s, the Reverend
‘expected surprise’ – known as entropy or uncertainty. Thus, in the
Thomas Bayes had an ingenious idea, which led him to the simple
long-term, we are all compelled to avoid surprises and resolve
mathematical rule that carries his name – but then puzzlingly he
uncertainty. Intriguingly, we found ‘variational free energy’ (in
gave it up. Seventy years later, Pierre-Simon Laplace rediscovered it
bits) and ‘thermodynamic energy’ (in Joules) closely coupled, a
independently and gave it its modern mathematical form. Bayes’
finding that emerged first when analyzing their interrelation
rule says: ‘By updating our initial beliefs with new evidence we
mathematically (Sengupta et al., 2013, 2016), and second when
obtain a novel and improved belief’. Bayes’ theorem is particularly
studying experimentally how the brain responds to stress (Harris
interesting, because it allows inferring from an effect to its probable
et al., 2012; Hitze et al., 2010).
cause. Such causal inference works better as more observational
evidence comes available. Current brain research makes use of the
1.2. What are our resources to master uncertainty?
very same principle in the so-called Bayesian Brain concept,
according to which the brain uses effects – that is, the sensory data
Let us take a quick look at how neuroscientists regard our
(or input), which is all the brain has access to – to figure out their
resources to master uncertainty. In the middle of the 20th century,
probable causes. Obviously, this is a fairly brief summary of the
Hans Selye identified a set of reactions, which he termed the
application of the Bayesian vernacular to the brain. Underlying it is
‘general adaptation syndrome’ (Selye, 1956). This uniform pattern
a more substantial view based on the rather uncontroversial idea
of responses that he described included the enlargement of the
that the brain engages with information processing, and that
adrenal cortex, atrophy of the thymus, and gastro-intestinal ulcers.

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx 3

Selye focused on physical, chemical, and microbiological noxae

that caused changes in the external environment and in the
internal body milieu: the lack of oxygen, nutrition or water; heat or
cold; toxins; microorganisms, etcetera. John W. Mason, however,
has criticized that Selye had underestimated the role of
psychosocial influences (Mason, 1959). He emphasized that
novelty, unpredictability, and uncontrollability of a condition
and expectation of adverse sequels are key triggers of a stress
reaction (Mason, 1968). Over the past forty years, many researchers
have defined stressful situations – from a bio-psychological
perspective – as characterized by ‘no information, no control,
uncertainty with a sense of threat’ (Koolhaas et al., 2011; Lyons and
Levine, 1994; Monat et al., 1972). More recently, experiments have
confirmed that humans are capable of computing environmental
uncertainty, and that their beliefs about uncertainty mediate the
strength of their stress responses (de Berker et al., 2016). Moreover,
the better people tune their beliefs about uncertainty, the better
they were able to predict future outcomes – suggesting that the
stress response has an adaptive function (de Berker et al., 2016).
According to these ‘Masonian’ notions, family conflicts at home,
mobbing at work, or highly disordered neighborhoods also
indicate an inhospitable environment. In any case, as suggested
above, the uncertainty about how to deal with such changes in the
internal body milieu or the external environment causes ‘stress’.
The approach to stress proposed here entails a well-defined
question (i.e., Question 1), where we have in mind a set of answers
without necessarily knowing which is correct. This means that we
can treat Mason's aspects of stress – novelty, unpredictability and
uncontrollability – more precisely and formally by using Shannon's
uncertainty (entropy). A second aspect of this novel stress
definition is that a ‘sense of threat’ is required. A third feature is
the inclusion of optional behavioral responses.
To resolve uncertainty, three processes play a crucial role:
attention, learning and habituation. When persons feel uncertain
and threatened, because of a changing internal or external
environment, their brains enter a hypervigilant status to decrease Fig. 1. Repeated stress responses in times of uncertainty. Stress responses are
uncertainty (about strategy selection) as fast as possible. To garner elicited, when the environment changes and the individual becomes uncertain
about which strategy to select in order to safeguard his/her future wellbeing. In this
the required information, extra cerebral energy is needed. From the
situation, the brain computes the probabilities for each available strategy (S1, S2, S3)
physical point of view, the following applies: Obtaining any that it will safeguard future wellbeing (upper insert). If all available strategies
information costs energy, erasing information produces energy display equal probabilities, then uncertainty (entropy) is maximal. Novel
(Berut et al., 2012; Brillouin, 1953; Toyabe et al., 2014; Tribus and information is needed to resolve uncertainty. To get novel information energy is
required. Thus, the stress responses contain a key component, namely, the
McIrvine, 1972). With respect to the brain, this means: Reducing
procurement of additional energy for the brain. With the help of the extra cerebral
uncertainty and providing the brain with the necessary energy energy, information processing is enhanced. If uncertainty is resolved, the
entails neuroendocrine and neuroenergetic responses that consti- individual may certainly select a suitable strategy from a repertoire (T1, T2, T3)
tute the stress response (Fig. 1). The Selfish Brain theory – founded (lower insert). In this case, the certainty has been regained, and stress reactions
by one of us in 1998 – describes this characteristic of the vertebrate have subsided (*). Non-habituators are people who show full neuroenergetic,
neuroendocrine, emotional and cardiovascular responses when repeatedly exposed
brain to cover its own, relatively high, energy requirements with
to an inhospitable environment. Chronically activated stress responses also exert
the highest priority when controlling energy fluxes within the adverse effects that lead to damage of the body and the brain-referred to as
organism (Peters et al., 2004). In this respect, the brain behaves allostatic load (**).
‘selfishly’ – as has been confirmed experimentally (Hitze et al.,
2010; Kubera et al., 2012b; Oltmanns et al., 2008; Peters et al.,
2011). Such a prioritization of brain energy metabolism is an descending pathways from the ACC are numerous and include the
inherent feature in vertebrates – a feature that is evident on short amygdala, as well as midbrain (e.g., periaqueductal grey) and
and long time-scales for preserving cerebral function, energy and brainstem nuclei, which contribute to multiple elements of the
mass during stress or food deprivation (Gong et al., 1998; Kind ‘stress’ response (Barrett and Simmons, 2015). The ACC-amygdala
et al., 2005; Miller et al., 2002; Muhlau et al., 2007). complex can stimulate two important descending projections:
The key mechanisms of the acute stress response – that have first, the pathway to the locus coeruleus (LC); this activation causes
been disclosed so far – are as follows: The anterior cingulate cortex a hypervigilant state (Hermans et al., 2011; Reyes et al., 2011;
(ACC) assesses the degree of uncertainty about whether future Valentino and Van Bockstaele, 2008), which has been linked to an
outcomes are uncertain (Fig. 2) (Behrens et al., 2007; Feinstein augmented brain energy consumption (Hitze et al., 2010; Kubera
et al., 2006; Karlsson et al., 2012; Liljeholm et al., 2013; Paulus et al., et al., 2012b); second, the pathway to the ventromedial hypothal-
2002; Sarinopoulos et al., 2010). The amygdala (by exchanging amus and the paraventricular nucleus, thereby stimulating the
information with the orbitofrontal cortex) may play a key role in sympathetic nervous system (SNS) and the hypothalamus pituitary
responding to threats to wellbeing (Schulkin et al., 1994). People adrenal (HPA) axis, which in turn provide the additional energy
who are not certain about their future wellbeing exhibit correlated required for the brain (Fig. 3) (Hitze et al., 2010; Kubera et al.,
activity in the ACC and amygdala (Sarinopoulos et al., 2010). The 2012b; Peters et al., 2004).

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

4 A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx

Fig. 2. Generating a stress response. The brain receives sensory and viscerosensory data about the hidden states of the world and the body. In turn, the brain acts on the
world and the body through its behavioral responses and stress responses. Both perception and action aim at minimizing prediction errors (free energy). Beliefs about the
current states of the world/body, about the attainable states, and about the goal states are proposed to be represented in the lateral PFC, the pre-SMA and the vmPFC/OFC,
respectively. The anterior cingulate cortex (ACC) compares attainable states with goal states. From that comparison, the risks for alternative strategies can be assessed. Based
on the risk assessment the ACC may select the best strategy for safeguarding future wellbeing: If strategy selection is certain, the pre-SMA and the primary motor cortex
initiate the respective behavioral response. If action selection is uncertain and threatening, the amygdala initiates a stress response.

Activated by the first path, the LC in the brain stem feeds back to Some people – but not all – adapt to chronic stress by
the cerebral hemispheres, where norepinephrine (NE) acts at habituating. Stress habituation can be regarded as a form of
cortical synapses (Aston-Jones and Cohen, 2005; Berridge and adaptation when living under uncertainty. We define ‘habituators’
Waterhouse, 2003). In this way, norepinephrine increases attention as those who display attenuated autonomic, endocrine and
and enhances information transmission at neuronal synapses. metabolic reactions, when being repeatedly exposed to the same
Crucially, enhanced information transmission at synapses is itself hostile environment. As we shall see later on, stress habituation
particularly expensive in terms of energy (Harris et al., 2012). reduces the uncertainty about what strategy to select.
Activated by the second path, the HPA-axis releases cortisol into
the general circulation. Cortisol passes the blood-brain-barrier 1.3. What happens when uncertainty cannot be resolved?
easily and binds to mineralocorticoid receptors (MR) and
glucocorticoid receptors (GR) located in and on neurons of the Next, we look at how neuroscientists and physicians are
hippocampus, the amygdala, and the cerebral cortex (Arriza et al., concerned with the health effects that arise from the failure to
1988; McEwen et al., 1968; Patel et al., 2000; Sanchez et al., 2000). resolve uncertainty. When the acute stress responses turn out to be
Here, cortisol regulates synaptic plasticity (long-term potentiation insufficient for resolving uncertainty, the critical situation may
and long-term depression) and in this way gates learning (Maggio become chronic, and the organism is burdened by ‘allostatic load’.
and Segal, 2007, 2009; Pavlides et al., 1995). Learning after stress According to the concept of ‘allostatic load’ – developed by one of us
can be interpreted as updating flawed beliefs about the world, (McEwen and Stellar, 1993) – the neuroendocrine, cardiovascular,
thereby furnishing better predictions of future outcomes (Fig. 3). neuroenergetic, and emotional responses become persistently

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx 5

Fig. 3. Three stress responses for reducing uncertainty. During uncertainty about what to do next, the ACC-amygdala complex activates three subsystems that feed back to
the brain: the LC, SNS and HPA-axis. First, LC-drive leads to cortical NE release, which enhances cortical information transmission. Second, the SNS allocates the required extra
energy to the brain that is necessary to support enhanced and more precise neuronal message passing. Third, the HPA-axis releases glucocorticoids, which suspend cortical
plasticity.

activated so that blood flow turbulences in the coronary and 2. The Bayesian brain
cerebral arteries, high blood pressure, atherogenesis, cognitive
dysfunction and depressed mood accelerate disease progression. The Bayesian brain casts neuronal processing as the process of
All of these long-lasting effects of persistently activated stress inferring the causes of sensations (e.g., this pattern of visual input
reactions are called ‘allostatic load’ (McEwen, 1998). When serious is caused by someone smiling at me). This inference conforms to
damage manifests, the effects are referred to as ‘allostatic overload’ Bayes rule, which says that (sensory) evidence is used to update
(McEwen and Wingfield, 2003). Uncertainty can lead to a vicious ‘prior beliefs’ into more informed, evidence-based ‘posterior
cycle of altered brain architecture and systemic pathophysiology, beliefs’. Accordingly, the Bayesian Brain iteratively updates its
which further damages the capability of the subject to cope with prior beliefs based on emerging evidence. All past experiences over
uncertainty. People living in a volatile and insecure environment the life course have ultimately formed the current prior beliefs,
(e.g., an insecure job, unhappy relationship, poverty, etc) have a which form the basis for how the Bayesian Brain makes predictions
high risk of depression, cognitive impairment, myocardial infarc- or decisions. Thus, early life adversity and failed attachment to
tion, and stroke (McEwen, 1998; Peters and McEwen, 2015). parental, but also any strong positive or negative experience in
In the current paper, the mathematical, neurobiological and school, work and personal life influence the way the Bayesian Brain
medical aspects of uncertainty are combined. On this background, selects its strategy for securing the future wellbeing.
stress is regarded as a form of uncertainty: We often make false To understand the Bayesian Brain, two key processes have to be
predictions about the world, and our prediction errors are considered in relation to each other: perceptual and active
therefore large. When stressed, we find ourselves unable to avoid inference. We begin with the first process, which focuses on
prediction errors or resolve uncertainty. In short, brains in distress how we perceive. This perceptual aspect of the Bayesian Brain is
do not work at their minimum of ‘free energy’. Again and again often referred to as ‘predictive coding’.
there are surprises. Although relating the Bayesian Brain concept to
‘stress’ appears plausible, the underlying neuroendocrine mecha- 2.1. Perceptual inference
nisms of such a link have yet to be established. The aim of the
current paper is to disclose those neuroendocrine mechanisms Our senses are bombarded with information about objects in
that could fulfill Bayesian functions in regulating attention, the world. On the basis of that sensory input, we perceive what is
learning, and habituation. In brief, we will consider the resources out there. The problem that will concern us is how the brain
the Bayesian brain has at its disposal to master uncertainty – but accomplishes this feat of perception – turning sensations into
also what happens when uncertainty cannot be resolved. percepts.

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

6 A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx

A basic and useful formulation of the problem of perceptual approximate Bayesian inference (Box 2), which is typically used in
inference is in terms of cause and effect. States of affairs in the ‘predictive coding’ and has been extensively studied in the field of
world have effects on the brain – objects and processes in the world machine learning. The latter approach describes how a ‘generative
are the causes of the sensory input. The problem of perception is model’ is used to infer from an effect to its probable cause. The
the problem of using the effects – that is the sensory data that the generative model consists of a ‘prior’ belief and the ‘likelihood’ (for
brain has access to – to figure out the causes. This represents a details see Box 2). With the help of the generative model one can
problem of causal inference for the brain, analogous in many predict an effect given the cause that one considers as most likely.
respects to our everyday reasoning about cause and effect, and to In other words, one can generate an effect from a cause.
scientific methods of causal inference (Hohwy, 2013). The problem Accordingly, generative models in the brain are capable of
of perception is a problem because it is not easy to reason from predicting sensory data. The predicted effect is then compared
only the known effects back to their hidden causes. This is because to the observed effect, and the difference between the two
the same effect can arise from many different causes. In other constitutes the so called ‘prediction error’. The prediction error is
words, in many situations we face an ill posed problem that can used in turn to update the prior belief to transform it into an
only be solved using prior expectations. Inferring the causes of approximate posterior belief. Unlike the mathematical procedure
sensations – and making that inference easier by resolving in exact Bayesian updating, the procedure used in predictive
uncertainty – is essential for navigating our world and, on some coding prescribes a particular process that gets the same results.
views, a necessary prequel for action. Furthermore, both forms of Bayesian inference can be cast as a
process that minimizes variational free energy, where free energy
2.1.1. Two modalities of Bayesian updating is, effectively, the overall amount of prediction error. It has been
We will illustrate Bayesian inference using the example of a shown that both perceptual inference and learning can be
medical doctor, who infers the diagnosis (cause) of the uncon- described as a minimization of free energy or the suppression of
sciousness (effect) seen in a young patient (Box 1): the doctor has prediction errors (Friston, 2005; Rao and Ballard, 1999). The
to choose between three possible diagnoses. Although a current concept of free energy originates from statistical mechanics; it is
textbook of medicine could extend the list of possible diagnoses, it often used to convert difficult integration problems – inherent in
is typical for humans to consider a limited number of options the direct application of Bayes rules – into an easier optimization
(Ortega and Braun, 2015). The examples used in Box 1 and Box 2 problem (e.g., the minimization of prediction error). ‘Free energy’
describe two distinct formulations of Bayesian inference in easily can be regarded as the information a person is lacking, and which
understandable terms. The first is exact Bayesian inference (Box 1), he/she could use to make his/her internal model as close as
using the explicit formula of the Bayes theorem, in which a prior possible to reality. Thus, free energy minimization corresponds to
belief is directly updated into a posterior belief. The second is prediction error minimization in predictive coding. By minimizing

Box 1. Exact Bayesian approach to medical diagnosis

Here, we give an example of how a clinician could use the Bayesian approach to estimate the probability of an underlying disease.
In this first example, the doctor is mathematically talented and makes use the explicit formula of Bayes (Eq. (1)). As an emergency
physician, she has been called because an occasional passer-by found a 17-year-old comatose patient at the roadside. When seeing
the young patient lying unconscious (what the doctor sees is an effect), she is uncertain about the probable cause, i.e. the
underlying disease. Three possible diagnoses come to her mind: intoxication (e.g. alcohol or drugs), neuroglycopenic coma (i.e.
brain energy deficiency) or cerebral bleeding. Based on her previous experience she assigns so-called ‘prior probabilities’ to each
of her three potential diagnoses (see table). By the way, the brief version of the term ‘prior probability’ is ‘prior’. Note, that these
three prior probabilities add up to 1.0. First of all, the doctor leaves other possible diagnoses out of consideration. At this time point,
‘intoxication’ is her favorite diagnosis, because it has highest prior probability P(X) = 0.5. However, she still remains uncertain. The
physical examination is followed by a blood glucose check, which indicates 25 mg/dl (normal 70–110 mg/dl). The measurement can
also be regarded as an effect (Y) of the underlying disease (X). After the blood glucose value is obtained, the doctor considers ‘how
likely it is that the disease (cause) fits the blood glucose measurement (effect)’. This is the ‘likelihood’ P(Y|X): the probability that the
cause described in the diagnosis would actually cause that particular effect. If ‘neuroglycopenic coma’ is the underlying cause in
that patient, then the probability is high that a life-threatening blood glucose measurement of 25 mg/dl is obtained. The likelihood
for the diagnosis ‘neuroglycopenic coma’ is 0.9. The other two diagnoses display a particularly small likelihood (0.1 in both cases).
In Bayesian belief updating, the ‘prior probability’ is multiplied by the ‘likelihood’, and the result is proportional to the so-called
‘posterior probability’ P(X|Y).
 
PðY jX Þ
Bayesian belief updating : PðX jY Þ ¼ PðX Þ ð1Þ
PðY Þ
We call the factor in brackets ‘update factor’, since the prior multiplied with this update factor equals the posterior. Bayes’ inference
leads to a new favorite diagnosis, which is ‘neuroglycopenic coma’ with a posterior probability P(X|Y) = 0.857. The other two
diagnoses become less likely. Of note, the posterior probabilities for all diagnoses again add up to 1.0 (see table). After all, the
posterior becomes the new prior, which can be used in the next diagnostic step that involves further diagnostic evidence.
17-year old patient in coma
Cause Effect Prior Likelihood Posterior
X Y P(X) P(Y|X) P(X|Y)
Intoxication Blood glucose measurement shows 25 mg/dl 0.5 0.1 0.119
Neuroglycopenic coma 0.4 0.9 0.857
Cerebral Bleeding 0.1 0.1 0.024

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx 7

Box 2. Approximate Bayesian approach to medical diagnosis

‘Perceptual inference’ – as it is believed to be implemented in the brain – differs from a methodological point of view from the
explicit (‘exact’) Bayesian updating described in Box 1. To illustrate the way ‘perceptual inference’ works, we give a second
example: A different doctor is also confronted with a comatose 17-year old patient. In contrast to the doctor in the first example
(Box 1), he does not use the explicit Bayes formula, but in his case the process of Bayesian inference is rather based on
subconscious intuition. In fact, ‘perceptual inference’ works in such an unconscious way. The second doctor also has his favorite
diagnosis, which is ‘intoxication’, i.e., the diagnosis with the highest prior probability. The prior probabilities are again as shown in
the table (Box 1). Before the blood glucose measurement is obtained, the doctor would use his favorite diagnosis together with the
likelihood P(Y|X) to predict, which blood glucose measurement is most likely to be observed (given his favorite diagnosis is true).
This means, generally speaking, that the ‘prior distribution over the causes’ together with the ‘likelihood of the effect given the
causes’ is used for predicting observable data values. In probability theory and statistics, such a model used for predicting
observable data values is referred to as a ‘generative model’. In case the favorite diagnosis ‘intoxication’ is true, the doctor would
predict a blood glucose value of Y = 90 mg/dl. When the doctor sees the actually obtained blood glucose value of Y* = 25 mg/dl, it is
different from what he had expected. He is surprised. The difference between the predicted 90 mg/dl (Y) and the actually observed
25 mg/dl (Y*) is called ‘prediction error’. Thereon, the doctor uses the prediction error to estimate a new posterior probability. The
technical problem is that the precise calculation of the new posterior cannot be performed easily. Instead, he may use an
approximation of the true posterior Q(X), which can still be very useful. Put simply, he successively adjusts his values of Q(X) to
minimize (blood glucose and prior) prediction errors. In perceptual inference, the ‘approximate true posterior’ lies close to 0.857 for
the diagnosis of a ‘neuroglycopenic coma’, indicating that this is now the best choice. It should be mentioned that the process of
‘predictive coding’ is a little more complex than illustrated here, since it deals with continuous states while the example with the
doctor deals with discrete states. Although different in mathematical terms, ‘exact Bayesian inference’ and ‘approximate Bayesian
inference’ (used in predictive coding) follow the same line of thought: inferring from the effect to its probable cause. The key utility
of predictive coding is that it prescribes a process theory for doing Bayesian inference (by suppressing prediction errors or
variational free energy) that could be implemented in the brain.

prediction error, the internal model is improved step-by-step predictive coding) the priors at intermediate levels now become
updating posterior beliefs, given the prior beliefs and the ‘empirical priors’. This follows because they become accountable
likelihood of sensory observations. to empirical (sensory) data and can therefore be optimized to
In conclusion, Bayesian inference, as used in perception, reduces minimize prediction errors at each hierarchal level.
our uncertainty about the states of affairs in the world that have Neuroanatomically, the notion of a hierarchy is based on the
caused our sensations. distinction between bottom-up and top-down connections (Salin
and Bullier, 1995). The notion of top-down connections provided a
2.1.2. Precision and attention better explanation of experimental data than the idea that only
A key concept – that we will develop later – is the notion of bottom-up connections are necessary (Garrido et al., 2007). Top-
‘precision’ in predictive coding. In brief, the predictive coding down connections arise largely from layer-5-pyramidal cells and
formulation of the Bayesian Brain says that prediction errors target layer-2-pyramidal cells of lower cortical areas (Fig. 4).
represent the newsworthy information that has yet to be Conversely, bottom-up connections arise largely in layer-2-
explained. Clearly, the brain also has to select the ‘news’ channels pyramidal cells and project to the spiny layer-4-neurons of a
it should attend to. This process of selection corresponds to higher cortical area.
increasing the volume or ‘gain’ of precise or reliable prediction Fig. 4 illustrates our simplified model of the hierarchical
errors. ‘Gain modulation’ is a phenomenon commonly observed in organization of the brain. In agreement with neuroanatomical
neuroscience that alters the amplitude of a neuronal response, but studies (Feldmeyer et al., 2005; Harris and Shepherd, 2015; Lubke
not its selectivity. Computationally, the increase in the gain of et al., 2000; Ramaswamy and Markram, 2015; Thomson and Lamy,
neurons encoding prediction errors ensures that precise informa- 2007) and neurocomputational analyses (Bastos et al., 2012;
tion is used to revise internal models. Psychologically, it provides a Haeusler and Maass, 2007; Shipp, 2016; Shipp et al., 2013), we
nice metaphor for attentional selection. Finally, from a physiologi- propose the following assignment of empirical prior beliefs, the
cal perspective it places neuromodulation and cortical gain control likelihood, the sensory data, and the prediction errors to their
in a key position to influence perceptual synthesis. This follows anatomical substrates. Accordingly, empirical prior beliefs are
from the fact that increasing the influence of prediction errors rests located in layer 3 of the cortex, and the likelihoods or predictions
on increasing the responsivity of the neuronal populations, which are represented in layer 5. The empirical prior expectation encodes
report them. the most likely cause of sensory inputs. The combination of
Put simply, the higher the precision of prediction errors, the greater empirical prior and likelihood furnishes a prediction that is
their influence on belief updating. During stress, neuromodulation conveyed to layer-2-pyramidal cells at the level below. Here, the
amplifies the precision of sensory prediction errors, endowing sensory prediction is compared to the sensory data, and in this way, a
information with greater weight, in relation to prior expectations. prediction error is calculated. In return, this prediction error is
conveyed upwards, through layer-4-stellate cells to layer-3-
2.1.3. The anatomical correlates of hierarchical predictive coding pyramidal cells, where it can update the empirical priors. This
The brain’s architecture is hierarchically organized (Felleman brief description has been simplified for clarity; for a more detailed
and Van, 1991). This organization has been intensively investigated account, see legend of Fig. 4. The computational architecture
in the visual system: the lower cortical areas are located closer to depicted in Fig. 4 enables continuous updating of prior beliefs in a
primary sensory input, while the higher areas play an associational hierarchical fashion, in response to fluctuating sensory input at the
role. The hierarchical architecture enables the brain to learn its lowest level.
own priors as well as the intrinsic causal structure of the world that In perceptual inference, the brain makes use of an internal
creates the sensory input. In hierarchical Bayesian inference (or hierarchical model where top-down predictions are continuously

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

8 A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx

Fig. 4. Simplified model of the hierarchical organization of the brain. Here, posterior or empirical prior beliefs P(X), likelihoods P(Y|X), sensory data (Y*), and prediction
errors are associated with anatomical structures. Superficial L3-pyramidal cells encode the sufficient statistics (i.e., the mean) of the empirical prior probability distribution;
deep L5-pyranidal cells encode the sufficient statistics of the likelihood probability distribution. The updating of the empirical prior involves two processes: First, the
empirical prior (L3) and the likelihood (L5) generate a top-down prediction. This prediction is compared to the sensory data (Y*), thereby generating a first prediction error in
L2-pyramidal cell of the lower level (Y*-Y). This lower-level prediction error is conveyed further upwards via L4-stellate cells to L3-pyramidal cells. It contributes to the update
of the empirical prior. Second, a current-level-prediction error is calculated in layer-2-pyramidal cells. This second prediction error results from the comparison between the
higher-level predictions and the current-level-empirical prior. The relative influence of the first and second prediction error on empirical priors is determined by their
precision or reliability. Technically speaking, the precision of a prediction error corresponds to its inverse variability. Physiologically, precision is thought to be encoded by the
responsiveness or gain of cells encoding prediction error while, psychologically, it can be associated with attentional gain that selects precise sources of prediction error;
namely, reliable information that is yet to be explained. The ultimate aim of the belief propagation or neuronal message passing it implied by this circuitry is to minimize
(precision weighted) prediction error or free energy, and thereby optimizing the generative model. For simplicity, we have only shown the lowest sensory level and the first
level of the hierarchy. A similar architecture can be imagined for subsequent levels, where the data become empirical priors at the lower level.

updated by bottom-up prediction errors (Friston et al., 2006). The about the world (Adams et al., 2013; Friston et al., 2006). Put
reciprocal top-down/bottom-up-message passing in this hierarchy simply, one can either change empirical prior beliefs (through
seems able to accommodate the context sensitive and invariant perception) to make predictions more like sensations or one can
aspects of perception, while at the same time explaining many change sensations (through action) to make them more like
neuroanatomical and neurophysiological facts about cortical predictions. The second is ‘active inference’. Experimental
hierarchies (Hohwy, 2013). evidence supports the notion that the principles of active inference
In summary, perceptual inference reduces our uncertainty about underlie the functioning of the visual and auditory systems
what caused our sensory observations. Next, we consider the (Chennu et al., 2013; Kok and de Lange, 2014). In the here and now,
important fact that we can choose which sensations to sample. minimizing certain forms of prediction errors (e.g., proprioceptive
prediction errors) through action can be very simple. For example,
2.2. Active inference motor reflexes can be described as quenching proprioceptive
prediction errors (Adams et al., 2013). Another example concerns
The second process that is important for understanding the the allostatic (predictive) regulation of the internal body milieu
Bayesian Brain is ‘active inference’. ‘Perceptual inference’ and (Sterling, 2012), where viscerosensory prediction errors are
‘active inference’ represent the two ways in which we can reduced (Barrett and Simmons, 2015).
minimize prediction errors. In perceptual inference, individuals There appears to be a dialectic regarding action and uncertainty.
strive to update their internal model of the world, while in active Both exploratory behavior and avoidance behavior show biphasic
inference individuals change their environment (or their sampling effects on uncertainty. Exploratory behavior is costly because it
of the environment) with the aim of better informing their beliefs involves a variety of risks, leading to ‘short-term uncertainty' (e.g.,

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx 9

the voyages of Christopher Columbus); but the possible (episte- et al., 2013). Past experiences – such as early life adversity and
mic) benefit of exploratory behavior is the reduction of ‘long-term failed attachment to parental – have a strong impact on what
uncertainty’ (filling in the white areas on the maps). In contrast, future events are foreseen when considering the respective
avoidance behavior can reduce ‘short-term uncertainty’ (retreat strategy. In a long-lasting iterative process, the amygdala- and
may indeed have a relaxing, uncertainty reducing effect); but hippocampus-dependent emotional and declarative memories
people who have suspended appropriate belief-updating during an shape the generative model, which allows the prediction of the
avoidance episode might be very surprised by a world that has attainable states. In this regard, no person or animal is free from
changed profoundly in the meantime. As avoidance behavior such biographical biases towards the prediction of new events.
impairs epistemic foraging, it can promote adherence to outdated Technically speaking, the beliefs about the ‘states that can be
beliefs, thereby causing ‘long-term uncertainty’. In short, there is reached’ are represented by a probability distribution over – or
an optimum balance between approach and avoidance behavior expectations about – (counterfactual) states (Fig. 6A; blue).
that rests upon environmental volatility and, more importantly, The ‘states that agents believe they should occupy’ are
the ability of an agent to estimate volatility and use it to minimize represented in regions like the ventromedial prefrontal cortex
long-term uncertainty (Friston, 2009). By analogy, the exploration prefrontal cortex (vmPFC) and the orbitofrontal cortex (OFC)
for energy faces the same dialectic: Although foraging behavior (Barron et al., 2015; Bechara et al., 2000; Gottfried et al., 2003;
itself is an energy-consuming process, it serves to access energy O'Doherty et al., 2003; Roesch and Olson, 2004). These regions
resources (Peters et al., 2007b). occupy the highest hierarchical position in the Bayesian Brain.
The same notion of minimizing expected prediction errors in They play a key role in defining the expected value (i.e. free energy)
the future (i.e., uncertainty) underlies extension of active inference of future states. These goal or prior preferences provide a point of
into the domain of goal-directed-decision making and planning reference for goal directed behavior but can also be updated. The
(Friston et al., 2015). The implicit action selection is the endpoint of beliefs about the goal states are also represented as a probability
active inference and has particular relevance to our treatment of distribution over states of the world (Fig. 6A; red).
stress. We now take a closer look at how the brain infers the best The ACC is a central region that is in a position to integrate
strategy. beliefs about attainable states and goal states (Lee et al., 2007;
Nguyen et al., 2014). From a theoretical perspective, the difference
2.2.1. Goal-directed-decision making between attainable states and goal states can be formalized by the
Recently, goal-directed-decision making has been considered in so-called ‘Kullback Leibler (KL) divergence’ (Kullback and Leibler,
terms of active inference (Friston et al., 2013, 2014). In other words, 1951). In information theory, the KL divergence measures the
the problem of selecting behavioral strategies can be treated as an difference between two probability distributions. Although it is
inference problem. For such a decision-making process three kinds often intuited as a way of measuring the distance between
of probability distributions are relevant: Probability distributions probability distributions, the KL divergence is not a true metric. It is
over the often used as a measure of the information gained when one
revises one’s beliefs from a prior to a posterior probability
 current states of the world/body, distribution. In the case of the ACC, the KL divergence can be
 states that can be reached, i.e., attainable states, considered as a measure of how much the attainable states differ
 states that the agent believes he/she should occupy, i.e., goal from the goal states (Fig. 6A and B). In engineering and optimal
states. control theory, the process of minimizing the divergence between
predicted and preferred states is called KL control. In economics,
Strategy selection occurs under the prior belief that it will the KL divergence is known as risk; leading to a formal description
minimize the difference between the probability distribution over of risk sensitive behavior.
‘attainable states’ (given beliefs about the current state) and the In summary, an important aspect of resolving uncertainty is the
probability distribution over ‘goal states’. Or, put another way, selection of actions or strategies that reduce expected surprise, in
choices are based upon beliefs about alternative strategies, where relation to prior preferences or goals.
the most likely strategy minimizes the difference between
attainable and desired outcomes (repertoire versus goal) (Friston 2.2.2. The degree of uncertainty about what to do next
et al., 2013, 2014). The principle of how the brain resolves uncertainty about which
The beliefs about the ‘current states of the world’ are action or strategy to select is also found in the most basic forms of
continuously updated during perceptual inference. The lateral life. In the chemotaxis, for example, a bacterium changes its
prefrontal cortex (PFC) is a key brain region where current strategy (i.e., the direction of swimming) until it reaches a
environmental states are thought to be encoded (Panagiotaro- nutrient-rich environment (goal state). This itinerant strategy is
poulos et al., 2012). Thus, this brain region represents updated based on the prior expectation that the agent will only change its
empirical priors or posterior beliefs about the ‘current states of the direction of movement if it is facing unexpected states (Friston,
world’. The lateral PFC occupies a high hierarchical position in the 2011). Likewise, the (ideal) Bayesian Brain entertains alternative
brain. It sends predictions to the sensory cortex, which is located at strategies when it cannot reach goal states. Figuratively speaking,
a lower level, and in turn the lateral PFC receives prediction errors the nutrient-deprived bacterium shows the same restless behavior
from the sensory cortex. The viscerosensory cortex evaluates – which is an indication of ‘uncertainty’ where it should go – as
interoceptive signals (e.g. pain, cutaneous light ‘sensual’ touch and many stressed humans who are uncertain about what to do next.
thermal sensations) that result from changes in the internal body Among the various brain regions, the ACC is in a prime position
milieu (viscera, muscles and skin) (Barrett and Simmons, 2015; to integrate information from the pre-SMA and the vmPFC/OFC and
Chanes and Barrett, 2016). evaluate the relative risk (KL divergence) (Fig. 2). Action selection
The beliefs about the ‘states that can be reached’ may be is based upon these divergence measures, where the most likely
represented in the pre-supplementary motor area (pre-SMA) strategy minimizes the KL divergence or relative risk (Fig. 6C). In
(Fig. 2) (Nguyen et al., 2014; Rushworth et al., 2004). On this view, this way, the ACC plays a central role in selecting a strategy.
the pre-SMA comprises a generative model that predicts outcomes If an individual feels certain about the answer to question 1
that can be reached by the use of alternative strategies (strategy1, (‘What strategy should be selected to safeguard future wellbe-
strategy2, . . . , strategyn) taken from a given repertoire (Friston ing?’), strategies are available for achieving the desired goals (i.e.,

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

10 A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx

strategies with small KL divergences), and the strategy that initiates an emergency program to ensure inferences about the state of
produces best approximates the desired outcome (smallest KL the world are properly informed.
divergence) can be selected and conveyed from the pre-SMA to the To prevent all these short- and long-term risks associated with
primary motor cortex. In turn, the primary motor cortex generates uncertainty, the brain starts its emergency program as soon as a
proprioceptive ‘predictions’, which are fulfilled by transforming person gets into such a precarious situation. Pivotal to this
peripheral proprioceptive ‘prediction errors’ into movement formulation is the activation of the amygdalae. These brain regions
(Adams et al., 2013; Shipp et al., 2013). In this way, the selected – located within the temporal lobes – organize the stress responses
behavior is instantiated (Fig. 2). The certainty about what to do that are ultimately aimed at resolving the risk and uncertainty
next manifests itself through a strong sense of control. portended by the ACC. The reduction of uncertainty in such states
Clearly, it is possible that none of the strategies in the game will of emergency involves the beneficial actions of the stress
achieve the desired goal with a sufficiently high probability (i.e., no hormones. Both cortisol and catecholamines are important in
strategies in play are likely to reach the desired goal). We suggest determining memory of significant things to avoid danger in the
that in such a risky and uncertain state of affairs, the ACC initiates future. Stress hormones are necessary to update our beliefs about
an emergency program comprising a set of coordinated stress the world (and our plans), which are no longer fit for purpose.
responses (Fig. 2). Accordingly, the ACC issues visceromotor
‘predictions’ to the brainstem and spinal cord via connections 3. Mastering uncertainty
that cascade through the amygdala. Such visceromotor ‘predic-
tions’ are fulfilled by converting visceromotor ‘prediction errors’ As mentioned, the brain uses of three processes to master
into neuromodulatory, autonomic and hormonal action (Barrett uncertainty: attention, learning, and habituation. Crucially, this
and Simmons, 2015). In this way, the allostatic network is repertoire of uncertainty resolving processes is closely intertwined
activated. The ‘visceromotor regions’ controlling the allostatic with cerebral and systemic energy metabolism.
processes (the amygdala, ventral striatum, insula, orbitofrontal
cortex, anterior cingulate cortex, medial prefrontal cortex) are 3.1. Attention – the procurement of more precise sensory information
commonly regarded as the ‘circuits for emotions’ (Barrett, 2017). for Bayesian updating
With such an emergency activation of the ACC-amygdala complex,
the individual experiences feelings of threat, uncertainty and lack The first and immediate response to a stressful challenge is
of control. As has been shown experimentally, persons who display arousal. Arousal includes an increase of attention and vigilance. In
the largest stress responses exhibit lowest levels of self-esteem an uncertain situation, the brain switches from the normal vigilant
and locus of control, i.e. self-concept of own competence state during wakefulness into a hypervigilant state. Here, we
(Kirschbaum et al., 1995; Pruessner et al., 2005, 1999). In extreme review evidence that speaks to how the neuroendocrine mecha-
cases, however, when it appears precluded that any of the available nisms that lead to hypervigilance procure more precise sensory
strategies can achieve the goal state (i.e., every strategy exhibits an information. As mentioned above, information is required to
extremely large KL divergence), the individual may despair and reduce uncertainty (Shannon, 1948). To get the more precise
abandon his/her goal. information, extra cerebral energy is needed. The relation between
In the following, we will focus on the risky state of affairs – in information and thermodynamic cost follows from the fact that
which the individual experiences feelings of threat, uncertainty, energy is required to change the information encoded in any
and loss of control. The associated risk is twofold: first, there is the system. From a neuronal perspective, the increase in the gain of
risk of surprising outcomes (e.g., physical injury, loss of social neurons encoding prediction errors is metabolically costly. From
position, financial loss, separation from life partner, etc.). Secondly, the point of view of predictive coding, an increase in precision or
there is a risk that the lack control associated with ‘toxic stress’ and attentional gain necessarily entails an increase in neuronal energy
allostatic overload might progress to disease. Once exposed to supply.
threatening changes in the external or internal environment, the During a psychosocial challenge (e.g., oral examination,
individual is confronted with three possible outcomes: The first dispute), the amygdala and the bed nucleus of the stria terminalis
outcome indicates ‘good stress’; it represents a satisfying result; provide input to the LC, the SNS and the HPA axis (Fig. 3) (Swanson,
certainty could be regained and the individual experiences a sense 2000). As a result, corticotrophin-releasing factor is released from
of mastery and good self-esteem; wellbeing is restored completely nerve terminals targeting LC neurons, thereby regulating their
(Fig. 1; asterisk) (McEwen and Gianaros, 2010). The second firing patterns (Valentino and Van Bockstaele, 2008). This
outcome specifies ‘tolerable stress’; in this case, the individual neuropeptide leads to an increase in tonic and a decrease in
could not undo the changes in the inhospitable environment; phasic discharge of LC neurons. Phasic discharge involves only a
however, uncertainty could be reduced through buffering mecha- limited number of LC neurons and characterizes the mode of
nisms such as habituation. These people show only low stress ‘focused attention’. In this mode, LC projections to cortical areas
responses and intermediate levels of self-esteem and locus of lead to enhancement of only the few parts of the cortex involved in
control (Kirschbaum et al., 1995; Pruessner et al., 2005, 1999). This a specific task (Aston-Jones and Cohen, 2005; Berridge and
second outcome is discussed later on in the chapter ‘Habituation – Waterhouse, 2003). In contrast, tonic discharges of locus coeruleus
updating of goal states’. The third outcome characterizes ‘toxic neurons involve large number of these neurons. This mode leads to
stress’; in this case, the buffering mechanisms failed, and the widespread activation of cortical areas (Aston-Jones and Cohen,
individuals remain trapped in the inhospitable environment; their 2005; Berridge and Waterhouse, 2003). The tonic mode is activated
stress responses are maximal whereas their levels of self-esteem during acute stress and facilitates the interaction of many cortical
and locus of control are minimal (Kirschbaum et al., 1995; areas (i.e., by increasing the influence of ascending prediction
Pruessner et al., 2005, 1999); these persons are at high risk for errors), which – in concert – aim at revising beliefs in order to
physical and mental morbidity and mortality (Fig. 1; two asterisks) reduce uncertainty.
(McEwen, 2012). In the following, we will look at how attention optimizes the
In conclusion, if we feel confident that we can reach our goal states processing of precise sensory information. Specifically, we consider
(i.e., one course of action has a particularly low risk), the ACC informs how norepinephrine increases cortical synaptic information trans-
the motor system about the best action. However, if we feel uncertain mission, how particular information flows are augmented, what
about what to do next (all strategies are equally risky) then the ACC energy costs arise, and how the extra energy required is provided.

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx 11

3.1.1. Noradrenergic regulation of presynaptic glutamate release accounts of perceptual inference. In other words, selecting certain
LC projections target cortical synapses, where they release streams of ascending prediction errors (arising in hotspots) affords
norepinephrine in a paracrine manner (Fig. 3) (Berridge and precision and influence to those prediction errors – so that they
Waterhouse, 2003). Here, norepinephrine modulates the release exert a greater effect higher in the cortical hierarchy. In this way,
probability of glutamate and GABA from presynaptic nerve NE-mediated presynaptic gain control is in a prime position to
terminals. We focus on the modulation of inputs arriving at the selectively procure unexplained or newsworthy sensory informa-
basal dendrites of layer-2-error-units-pyramidal cells (Fig. 5). Mara tion that is necessary to revise empirical prior beliefs about the
Mather and coworkers have described a positive feedback loop and world.
have coined the term ‘hot spot’, describing the coincidence of a In summary, stress ignites multiple NE hotspots, thereby selectively
‘high-frequency-arriving-spike train’ with a ‘high LC-drive’ enhancing the transmission of precise sensory information.
(Mather et al., 2015). Upon such a coincidence, a ‘hot spot’ can
be observed where glutamate release probability is enhanced (for 3.1.2. Energetic constrains on information transmission
details see Box 3). Neural information transfer incurs an exceptional amount of
The concept of NE hot spots (Mather et al., 2015) nicely explains energy (Harris et al., 2012). The brain uses 20% of the total energy
two experimental observations: First, increasing the frequency of that is available in the human organism, although its mass
LC stimulation either leads to facilitation or suppression of sensory contributes only 2% to the total body mass (Peters et al., 2004).
evoked neuronal responses (Devilbiss and Waterhouse, 2004). For, Thus, the brain occupies a privileged metabolic position. Its
according to Mather’s concept, a NE hot spot facilitates synaptic primary fuel is glucose. It takes up more than 60% of the circulating
transmission of a high-frequency-spike train, while it suppresses glucose at rest (Reinmuth et al., 1965). Remarkably, an experimen-
synaptic transmission of a low-frequency-spike train. Second, high tal mental or psychosocial challenge increases whole-brain-
LC activation increases the likelihood that an action potential glucose uptake by more than 10% (Hitze et al., 2010; Madsen
results in the release of glutamate, while low LC activation et al., 1995). So-called ‘brain pull mechanisms’ – a term which
decreases that likelihood (Chiu et al., 2011; Kobayashi et al., 2009). originates from logistics – allow such a rapid systemic (re-)
These neuromodulatory hotspots appear to offer the perfect allocation of energy resources within the human organism (Peters
mechanism for the gain control implicit in predictive coding and Langemann, 2009). Even on the cell-to-cell level, it has been

Fig. 5. LC activity enhances synaptic information transmission at multiple release sites. Upon high LC activity, hotspots are ignited both at glutamatergic and GABAergic
synapses. In the case of multiple release sites from one axon onto a postsynaptic cell, the energetically optimal release probability of neurotransmitters is 25%–50%. By igniting
hotspots at multiple release sites, NE increases the share of action potentials that release neurotransmitter vesicles (glutamate or GABA). Hotspots decrease energy efficiency
of synaptic transmission, and increase total synaptic energy consumption. By expending extra energy, hotspots allow for a selection or boosting of specific prediction errors;
thereby endowing them with greater precision.

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

12 A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx

Box 3. Norepinephrine ignites local ‘hot spots’ of neuronal excitation

If a high LC activity leads to NE-release at a glutamatergic synapse, NE binds to adrenergic receptors located on the presynaptic
glutamatergic neuron: it binds to the high-affine a2-adrenoreceptors and the low-affine a1- and b-adrenoreceptors (Fig. 5)(Ramos
and Arnsten, 2007). At the presynaptic site of cortical neurons, high-affinity and low-affinity adrenoreceptors have been shown to
exert opposing actions. While low-affine a1- and b-adrenoreceptors increase the glutamate release probability (Ferrero et al., 2013;
Kobayashi et al., 2009), high-affine a2-adrenoreceptors receptors decrease it (Chiu et al., 2011). Furthermore, at low NE
concentrations the actions of high-affine a2-adrenoreceptors prevail, while at high NE concentrations the actions of low-affine a1-
and b-adrenoreceptors prevail (Nai et al., 2009, 2010). These findings are in line with the biphasic NE-dose-response curves on
patch clamp recordings (Linster et al., 2011).
Once glutamate has been released from the presynaptic site into the synaptic cleft, it binds on the one hand to postsynaptic AMPA
and NMDA receptors and on the other hand – as glutamate spillover – to NMDA-receptors located on the norepinephrine release
sites. In this way, we find a modulatory loop with NE regulating glutamate release probability, and glutamate spillover regulating
NE release (Mather et al., 2015). This modulatory loop acts in a positive feedback manner, and has been termed ‘hot spot’ (Mather
et al., 2015).
Thus, high LC-drive leads to high concentrations of NE at the presynaptic site, which lead to an increase of glutamate release
probability. If coincidently there is a high-frequency-spike train arriving at the presynaptic site, a particularly high amount of
glutamate is released into the synaptic cleft. A high glutamate spillover in turn amplifies further NE release, resulting in maximal
glutamate release probabilities. In contrast, in case of a low LC-drive and spontaneous glutamate release, the low concentrations of
the glutamate spillover are insufficient to promote further release of NE – the glutamate release probability is even more reduced.
High LC activity also ignites hotspots at GABAergic release sites. NE enhances GABA-release probabilities by low-affine a1-
adrenoreceptors, whereas NE suppresses GABA-release probabilities by high-affine a2-adrenoreceptors (Hirono and Obata, 2006;
Salgado et al., 2012). In summary, from among the action potentials arriving at the synapse, only a few actually elicit the release of
the neurotransmitter vesicles (glutamate or GABA). Only in the case of high LC activity do hotspots occur, where most of the
incoming action potentials of a high-frequency train actually produce a neurotransmitter release (glutamate or GABA).
Noteworthy, NE also affects postsynaptic excitability of cortical neurons (by suppressing excitatory and inhibitory postsynaptic
potentials) (Kobayashi et al., 2009; Salgado et al., 2011), but the facilitatory NE effects on presynaptic neurotransmitter release
(glutamate and GABA) have been shown to override the less effective NE-effects on postsynaptic excitability (Salgado et al., 2011).
In summary, NE ignites hotspots at neurotransmitter release sites, thus increasing the transmitted information (bits per second).

demonstrated that neurons take up energy ‘on demand’ (Magis-

tretti et al., 1999; Pellerin and Magistretti, 1994). This demand
process is caused by other ‘brain pull mechanisms’, which allow a
rapid local (re)allocation of energy resources within the brain itself
(Peters and Langemann, 2009). Most brain energy is spent on
synaptic transmission (Harris et al., 2012).
Not every action potential arriving at the synapse leads to the
release of a glutamate vesicle. Levy and Baxter used an informa-
tion-theoretical approach to show that the brain optimizes the
quotient of ‘information transmitted to energy expended’, rather
than optimizing its ‘coding capacity’ (Levy and Baxter, 1996). In
other words, the brain is, in principle, able to transmit information
at a much higher rate than it actually does, because operating at a
lower transmission capacity is more economical; i.e. energy
efficient (for details see Box 4).
Under non-stress conditions, LC-activation is low and the brain
operates in an economic energy-efficient mode, abstaining from
the [mis]use of its potentially higher information processing
capacity. Under stress conditions, however, LC-activation is high
and the brain operates in an energetically expensive mode,
exploiting its full information processing capacity, while forsaking
optimal energy efficiency.
In short, the use of the energetically expensive mode is restricted to
times of stress and uncertainty.

3.1.3. Prediction errors are encoded with higher precision during stress
Now the question arises, for which task is extra energy actually
needed. On the predictive coding view, uncertainty or loss of
Fig. 6. Selecting the best strategy. Panel A. During stress, the ACC monitors the
confidence about the top-down predictions calls for an increase in
divergence between the probability distribution over ‘attainable states’ and ‘goal
the precision of sensory prediction errors – that induce Bayesian states’ for plausible strategies that constitute a repertoire. Such divergences are
belief updating. Prediction error units assess how much bottom-up called Kullback-Leibler divergence (DKL). The greater the divergences or relative
information about sensory evidence differs from top-down risk, the greater the uncertainty about which strategy to select – and the stronger is
the activation of amygdala. Panel B. Each strategy displays a different relative risk
information on expectations (Fig. 5). But what neurobiological
(DKL). Panel C. Based upon the risk distribution, the ACC assesses for each strategyi
mechanisms help layer-2-error-unit-pyramidal cells to evaluate a (i = 1,2, . . . ,n) the probability that it will ensure wellbeing. In the example depicted
mismatch between descending predictions and the expected state here, none of these probabilities is high. Thus, the individual remains uncertain
of affairs at the current level? about the answer to question 1. Because of the high degree of uncertainty or
entropy, a stress response is initiated.

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx 13

Box 4. Energetic costs limit presynaptic release probabilities

In neurons with a maximal firing rate of 400 Hz, transmission capacity is maximal, if they fire at half of their maximum rate; i.e.
200 Hz. Yet in practice, the mean firing rates of neurons in vivo is much lower – around 4 Hz (Harris et al., 2012). Against this
background, it seems plausible that such a low basal glutamate release probability could be enhanced on demand. An action
potential is much more likely to result in the release of glutamate, if locus coeruleus output is high. According to the calculations of
Levy and Baxter, such a high glutamate release probability – that is induced by high LC activation – is energetically much more
expensive than synaptic transmission occurring with low LC activation, which constitutes the economically optimal mode (Levy
and Baxter, 1996).
With respect to energetic constrains, the number of release sites from an axon onto a postsynaptic cell is also important. This
number is larger than 6 for cortical pyramidal to interneuron synapses, larger than 4 for pyramidal cell to pyramidal cell synapses in
cortex, and 6 for excitatory synapses onto pyramidal cells in hippocampal area CA1 (Deuchars and Thomson, 1995; Larkman et al.,
1997; Markram et al., 1997). Julia Harris and David Attwell showed that in case of multiple release sites – from one axon onto a
postsynaptic cell – the optimal release probability is 25%–50%; optimal with respect to the quotient of information transmitted to
energy expended (in bits per joule) (Harris et al., 2012).

It has long been recognized that cortical neurons collectively Gaussian posterior distribution (the so-called ‘Laplace approxima-
exhibit synchronous activity patterns (Salinas and Sejnowski, tion’), their firing rate is taken to represent the posterior mean or
2001). Correlated (i.e., coherent) fluctuations play a crucial role in expectation. However, if neuronal firing encodes a mean what
many cortical processes. If one neuron fires, and the other one is encodes the standard deviation; i.e. uncertainty? The premise of
more (or less) likely to fire, we call their activity ‘temporarily predictive coding is that precision is encoded by the response
correlated’. Many neurons are able to sense temporarily correlated amplitude or gain of the neurons encoding prediction errors. The
input patterns. A neuron can detect coincident firing patterns, if key point here is that the noradrenergic modulation affects
spikes from two inputs arrive within a short time interval. But presynaptic gain control – and therefore contributes encoding of
neurons can also can sum up their inputs to elicit an action precision: NE increases the probability of presynaptic neurotrans-
potential (Konig et al., 1996; Shadlen and Newsome, 1994). A mitter release onto L2-neurons, and in so doing renders the
neuron is called ‘balanced’, if both excitatory and inhibitory inputs postsynaptic L2-neuron more responsive to non-correlated inputs.
are strong and cancel each other out. In this way, the mean input Thus, NE increases the number of action potentials that the L2-
current approximates zero, and the average steady-state voltage is error unit generates when it receives non-correlated input signals.
insufficient to generate an action potential. Yet, such a balanced Such an increase in the number of action potentials of the L2 error
neuron might still generate an action potential, since stochastic unit forwards more weight to the ascending prediction error; i.e.,
voltage fluctuations always occur that are large enough to cross the to the bottom-up sensory information flow.
necessary threshold. Balanced neurons are more sensitive to In summary, NE enhances information transmission at the
coherent (i.e. correlated) presynaptic input than unbalanced synapses, and in so doing enables the L2-error-unit-pyramidal
neurons (Salinas and Sejnowski, 2001). populations to endow prediction errors with higher precision. In this
A few years ago, it was shown that if the input from an excitatory way, NE acting on the sensory cortex – that occurs in situations of
neuron A and the input from an inhibitory neuron B are correlated, uncertainty – adds value and weight to the bottom-up-information
then the fluctuations of the postsynaptic neuron decrease (Salinas flow. Therefore, during stress, the sensory evidence becomes more
and Sejnowski, 2001). Let us assume that L2-error-unit-pyramidal influential than (relatively imprecise) prior expectations. Such
cells receive inhibitory input (conveying predictions from the level selective increases in precision are key for updating beliefs about
above) and excitatory input (encoding the empirical priors from the world that engender uncertainty and stress.
the current level). If predictions from the level above and priors
from the current level are highly correlated, then fluctuations of 3.1.5. The ‘selfish brain’ provides the energy for increasing precision
L2-error units will be attenuated. Conversely, if there is a Where does the extra brain energy needed during uncertainty
mismatch, then synaptic fluctuations will ensue, and the L2-error and stress come from? In many people – but not in habituators (see
unit will start firing. In this way, the L2-error units are capable of below) – the brain demands supplementary cerebral energy from
encoding and conveying prediction errors. the body stores (Peters and McEwen, 2015). The ACC-amygdala
The mechanisms described here for the detection of non- complex not only stimulates LC neurons, but in parallel stimulates
correlated inputs can explain how superficial pyramidal cells the SNS and the HPA-axis (Fig. 3). SNS and HPA-axis activations
extract newsworthy prediction errors for broadcasting deeper into immediately suppress insulin secretion from the pancreatic b-cells
the brains hierarchy. However, the precision of these prediction (Ahren, 2000; Chan et al., 2007; Frühwald-Schultes et al., 2000;
errors depends upon how L2-error units respond to their opposing Tong et al., 2007). Consequently, insulin-dependent GLUT4-
presynaptic inputs (Brown and Friston, 2012). It is here that NE mediated glucose uptake in muscle and fat is prevented,
(and other neuromodulators) comes into play in a very special way. rebalancing energy consumption in favor of the insulin-indepen-
dent GLUT1-mediated glucose uptake at the blood-brain barrier
3.1.4. NE increases the precision with which prediction errors induce (Hasselbalch et al., 1999). In this way, the brain foreshadows its
Bayesian updating own increased cerebral energy need induced by stress.
If the brain is in the business of inferring what its body should In conclusion: During arousal, noradrenergic regulation of
do next, it has to encode – at some level – probability distributions presynaptic neurotransmitter release probabilities leads to precise,
(Richmond and Wiener, 2007). Several probabilistic neuronal high gain transmission of sensory evidence required to update deep,
codes are mathematically possible, but most of the available hierarchical beliefs about changes in the external and internal milieu.
evidence points to predictive coding, in which the activity of single At the same time, the selfish brain supplies itself with the extra energy
units or populations encodes the mean of a Gaussian probability required for precision-engineered belief updating.
distribution (Friston, 2009). When assuming that neurons encode a

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

14 A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx

3.2. Learning – updating during and after stress

The second key process for mastering uncertainty is learning.

Glucocorticoids control functional and structural plasticity in
many brain regions (Fig. 3).

3.2.1. Glucocorticoids gate the time window for cortical plasticity

In the case of uncertainty about what to do next (i.e., large
divergences between the beliefs about ‘attainable states’ and ‘goal
states’ under all available strategies), the ACC-amygdala complex
stimulates the HPA-axis and – in so doing – increases the
concentration of glucocorticoids in both blood and brain
(Fig. 7A). In the hippocampus, amygdala, and cerebral cortex,
glucocorticoids bind to MR and GR receptors that are located both
within neurons (cytosol, nucleus) and on neuronal surfaces
(membranes) (Arriza et al., 1988; Patel et al., 2000; Sanchez
et al., 2000).
Here, we focus on intracellular MRs and GRs. These two
intracellular receptors differ in their affinity for cortisol: MR binds
cortisol with high affinity, GR with low affinity (Arriza et al., 1988).
Once cortisol has activated these receptors in the cytosol, MR and
GR enter the cell nucleus where they exert differential effects on
gene expression (de Kloet et al., 1998). Fig. 7B shows the MR and GR
binding characteristics of glucocorticoids in a pyramidal cell. With
respect to synaptic plasticity; i.e., long-term potentiation (LTP) and
long-term depression (LTD), MR and GR have been shown to act in
an opposing manner on gene expression (Diamond et al., 1992;
Pavlides et al., 1994, 1996). Here, we focus on how declarative
memories are formed under stress, which is distinct from the way
emotional memories are conserved (Maggio and Segal, 2012;
Fig. 7. Glucocorticoids enable or preclude cortical plasticity. Panel A. In case the
Quirarte et al., 1997; Zhou et al., 2010).
ACC fails to select among plausible strategies, uncertainty arises about which
At low cortisol concentrations, the facilitatory effect of MRs on strategy to pursue, leading to amygdala activation, which results in the increase of
plasticity prevails, while at high cortisol concentrations, the glucocorticoid concentrations in blood and brain. Panel B. Glucocorticoids bind with
inhibitory effect of GRs prevails. Due to the opposing actions of MR high affinity to intracellular MR and with a low affinity to intracellular GR.
Intracellular MR and GR exert opposing actions on the expressions of a subset of
and GR, the difference in effects on gene expression becomes
genes involved in long-term plasticity, e.g. the gene encoding the TrkB receptor
paramount (Datson et al., 2001). The MR-GR difference depends on (Schaaf et al., 1997). Thus, the combined effect of MR and GR on gene expression
the glucocorticoid concentration and shows a ‘bell-shaped’ (or corresponds to the difference between the MR-dose-response curve and the GR-
‘inverted U-shaped’) dependency (Fig. 7C) (de Kloet et al., 1998; dose-response curve. Panel C. The combined effect of MR and GR is represented by a
Joels, 2006). A bell-shaped dependence on glucocorticoids is ‘bell-shaped’ curve. For instance, the production rate of TrkB receptors follows such
a bell-shaped function. Since protein production is necessary for maintaining long-
evident, for example, for the probability occurrence of LTP (Fig. 7C)
term plasticity, the dependency of LTP-probability on glucocorticoid concentrations
(Diamond et al., 1992; Joels, 2006; Pavlides et al., 1994, 1996). also shows a bell-shaped form. Of note, the bell-shaped curve depicted here can be
Consolidation of declarative memories is most likely when interpreted as a probability over consolidation of experience dependent plasticity.
glucocorticoid concentrations are low in the normal range (where The mode m indicates the glucocorticoid concentration, where probability of LTP
has its peak.
the bell-shaped curve has its peak). In contrast, consolidation is
unlikely when glucocorticoids are absent (left-hand side of the
bell-shaped curve) (Wagner et al., 2005) or when glucocorticoid itself to be optimized to produce the best generative models.
concentrations are high (right-hand side of the bell-shaped curve) Clearly, if we experience the world as uncertain or ambiguous, we
(Plihal et al., 1999). Likewise, retrieval of memories also shows a want to suspend learning. Conversely, if we experience it as
bell-shaped dependency on glucocorticoids, i.e. retrieval is optimal predictable and lucid, we want to consolidate what we have
at low (normal range) glucocorticoid concentrations, but impaired learned. Given that LTP is selectively enabled over a carefully
at very low and very high glucocorticoid concentrations (Rimmele controlled window of MR-GR difference (Fig. 7), glucocorticoids
et al., 2013). open and close the time windows for ‘learning’ (Joels, 2006;
For an inclusive understanding of the role of glucocorticoids in McEwen, 2015). They are therefore in a key strategic position to
updating the ‘internal model of the world’, we have to consider selectively consolidate when and what we learn.
learning in the context of predictive coding. Previously, we have In conclusion: High glucocorticoid concentrations create a ‘phase
talked about inferring states of the world in terms of neuronal of change’, revising the current model of the world (including its
activity and how this depends on the precision of prediction errors strategies). Low glucocorticoid concentrations create a ‘phase of
encoded by activity dependent changes synaptic efficacy mediated, conservation’, stabilizing the current model of the world.
in this setting, by NE. However, this inference depends upon having As mentioned, LTP shows a bell-shaped dependency on
a good model of the world encoded in the hierarchical connections glucocorticoids. The crucial insight now lies in the fact that TrkB
that convey descending predictions. This model is learned over signaling is required for the production of proteins that maintain
long timescales through experience dependent plasticity; such as LTP (Langemann et al., 2008; Minichiello et al., 1999; Zhang and
LTP and LTD. On this view, the MR-GR difference is therefore in a Poo, 2002), and that TrkB receptor production rate displays a bell-
prime position to control the rate of learning – in the same way that shaped-dose-response dependency on glucocorticoids (Schaaf
NE controls the rate of inference and evidence accumulation by et al., 1997; Shi and Mocchetti, 2000). Consequently, TrkB
boosting the precision of prediction errors. This learning rate has constitutes a link that mediates the bell-shaped-dose-response

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx 15

dependency of LTP probability on glucocorticoids – and deter- uncertainty over the timescales of experience-dependent plastici-
mines when changes to our internal models should or can be ty.
consolidated. Heuristically, this means that we strive to achieve states that go
Many other biological functions also show a bell-shaped along with low (i.e., normal) glucocorticoid concentrations (i.e., the
dependency on glucocorticoids (de Kloet et al., 1998). In the reference level m in Fig. 7C). A special and interesting case presents
current paper, we regard this bell-shaped dependency on itself, where the glucocorticoid-dependent biological function is
glucocorticoids as embodying a reference point for the optimal the ‘glucocorticoid release into the blood circulation’. In this case,
level of experience-dependent plasticity. As uncertainty or stress the glucocorticoid concentration itself depends on the glucocorti-
is, by definition, a sign that our generative models are not fit for coid-dependent bell-shaped relationship. Thus, a setpoint is
purpose (i.e., have a high free energy), we should not endorse generated that depends primarily on the functioning of the MR
anything learned under these models by consolidating associative in coordination with GR (de Kloet et al., 1998; Peters et al., 2007a).
plasticity. This is consistent with a suspension of (maintenance of) As a result, the HPA-axis always strives to retain a low blood
LTP during high glucocorticoid and stress levels. Turning this concentration of cortisol. Therefore, those states with low (normal)
argument on its head, the minimum free energy state must be glucocorticoid concentration are the most probable states that we
associated with low glucocorticoid levels that permit learning. find ourselves in. It works like this: Decreased glucocorticoid
These optimal levels therefore reflect a reference signal or setpoint concentrations stimulate the HPA-axis, while increased glucocor-
to which our bodies (and brains) must aspire. In short, ticoid concentrations (circadian or stress-induced) inhibit it
glucocorticoid levels can be interpreted as a correlate of (Akana et al., 1988; Fehm et al., 1977; Jacobson et al., 1988; Peters

Fig. 8. Learning the prediction error precision. Panel A. The apical tuft of an L2-pyramidal neuron is involved in precision control. The apical tuft receives excitatory and
inhibitory input from layer 1 that modulates the postsynaptic gain of the pyramidal neuron. Excitatory modulation input increases the postsynaptic gain, while inhibitory
modulation input decreases it. Panel B. The output of the L2-neuron, which is the precision weighted prediction error, depends on the difference between excitatory and
inhibitory inputs. Modulatory input onto the apical tuft increases or decreases the slope of the input-output function; therefore, controlling the gain of postsynaptic responses
encoding prediction error. LTP occurring at glutamatergic input synapses of the apical tuft fixes an increased slope of the input-output function. In this case, the prediction
error is encoded with a high precision, i.e. it gets more weight as compared to the corresponding higher-level prediction error. LTP occurring at GABAergic input synapses of
the apical tuft fixes a decreased slope of the input output function. In this case, the prediction error is endowed with low precision, i.e., is weighted lower.

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

16 A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx

et al., 2007a). In this way, an equilibrium cortisol concentration is In conclusion, the probability that a synapse undergoes LTP shows
reached (typically during deep sleep) that represents the mode of a bell-shaped dependency on the glucocorticoid concentration. There
the reference distribution shown in Fig. 7C. In contrast, states with is an optimal glucocorticoid concentration that favors LTP. This
high glucocorticoid concentrations are to be avoided (e.g., when optimum reflects the absence of prediction errors, or in other words,
learning to drive a car). the free energy minimum.
In addition to the self-stabilizing MR/GR feedback, the HPA-axis
receives input from the cerebral cortex. A substantial input derives 3.2.2. Stress and surprisal
from the ACC, which is in a position to convey messages about In information theory, the term ‘surprisal’ – coined by Myron
rising uncertainty (i.e., expected free energy), and in so doing Tribus – is used to describe the deviation from one’s expectations
stimulates the HPA-axis. Under the assumption that our choices (Tribus, 1961). For a given probability distribution over outcomes Y,
are informed by biological and prosocial goals (e.g., satiety, surprisal is defined as the negative logarithm of P(Y): i.e., log(P
affiliative touch, pleasant temperature, etc.), then glucocorticoid (Y)). Thus, surprisal represents the improbability or ‘surprise’ of
concentrations will remain low if these goal states are attainable. observing an outcome; for example, reaching into your pocket for
Conversely, if these goals are deemed unattainable, then the your phone and finding it is not there. Crucially, the ‘expected
glucocorticoid concentrations rise. In short, glucocorticoid con- surprisal’ is ‘entropy’. Entropy is the mathematical measure of
centration is a direct reflection of whether the world is unfolding uncertainty, which is approximated by expected free energy. This
according to expectations or not. means that glucocorticoid levels may not only underwrite the

Fig. 9. Learning the generative model. Panel A. Triple pathways control the updating of the generative model. The prior-L3 neuron sends three pathways to the likelihood-L5
neuron: the driving-input path, the excitatory-modulation-input path, and the inhibitory-modulation-input path (which includes an inhibitory interneuron). As long as
glucocorticoid concentrations are high (during stress), LTP at glutamatergic and GABAergic synapses is prevented. LTP occurs at the L5 neuron’s tuft only if glucocorticoid
concentrations fall (after stress) and can ensue either in the excitatory or in the inhibitory path. Panel B. The generative model consists of the prior and the likelihood. The
slope of the input-output function of the layer-5 neuron represents functional gain between expectation and prediction. Excitatory modulatory input increases the
postsynaptic gain, while inhibitory modulatory input decreases it. Once the generative model is updated it can be conserved through LTP. In contrast, LTD would reset the
slope of the input-output function into a neutral ‘default slope’.

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx 17

expected metabolic free energy associated with stressful (e.g., and the glucocorticoid concentrations. LTP is consolidated if there
flight or fight) responses, it may also reflect changes in is highly correlated activity between the pre- and postsynaptic
informational or variational free energy. The above analysis also neuron and the glucocorticoid concentrations remain in the low
suggests that levels of high free energy preclude a consolidation of (normal) range (Maggio and Segal, 2007, 2009). In all other cases,
experience dependent plasticity – until the world becomes more LTP is attenuated or even LTD occurs (Maggio and Segal, 2007,
predictable and glucocorticoid levels return to their equilibrium 2009).
levels. The consolidation of a generative model through LTP or LTD
In uncertain and threatening situations, descending predictions occurring at the apical tuft of pyramidal neurons is selectively enabled
of stress-related responses are broadcast by the ACC to the anterior when the world is learnable; i.e. when expected surprise, uncertainty
insular and onto autonomic centers to elicit autonomic reflexes and glucocorticoid levels are low. In what follows, we consider this
(Barrett and Simmons, 2015; Behrens et al., 2007; Feinstein et al., learning in more detail.
2006; Karlsson et al., 2012; Liljeholm et al., 2013; Paulus et al.,
2002; Sarinopoulos et al., 2010). The ACC-amygdala complex 3.2.4. Glucocorticoids gate learning of prediction-error precision
activates the HPA-axis, cortisol concentrations increase, cortisol Why would the brain learn the precision of prediction errors? It
binds primarily to cerebral GRs. Only if the glucocorticoid is often difficult to decide whether one should rely on sensory
concentrations return to low concentrations, which bind primarily input or prior expectations. At dusk, for example, it is better to rely
to cerebral MRs, will long-term plasticity in cortical neurons (e.g. on prior expectations than on sensory input (the latter provides
layer-5-pyramidal cells) be induced and maintained. Thus, learning imprecise information due to the low levels of illumination).
of our internal model of the world can only occur at a low However, in bright daylight, it may be better to trust reliable and
glucocorticoid concentration, when our models are fit for purpose precise sensory input than prior expectations. Thus, the precision
– according to our ACC. of prediction errors encoded at lower and higher levels of the
In conclusion, if the internal model of the world makes false cortical hierarchy allow the optimal weighting of the bottom-up
predictions, high glucocorticoid values indicate a high free energy (i.e., and top-down information flow. Depending on what is more
uncertainty); the model is changed and functional plasticity is precise, either the sensory input or the prior expectation is
precluded. Once the updated model succeeds in making correct afforded greater weight. The key point here is that precision can be
predictions, low glucocorticoid values indicate a low free energy (i.e., learned on the basis of past experiences that enable the brain to
predictability); the model is consolidated and functional plasticity re- predict when sensory input will be precise or imprecise.
emerges. As mentioned above, NE increases the probability of presynap-
tic neurotransmitter release at basal dendrites, making the L2
3.2.3. Functional plasticity at the apical tuft neuron more sensitive to non-correlated presynaptic inputs. In this
In the preceding treatment, we distinguished between infer- way, NE effectively increases the precision of the reported
ence and learning; where the former entails optimizing expect- prediction error, which manifests itself as increased attention. In
ations about states of the world through synaptic activity and the contrast, glucocorticoids alter the efficacy of synapses located on
latter entails optimizing expectations about parameters of the the apical tuft of pyramidal neurons, resulting in long-term
model through synaptic plasticity. In the following, we will focus on changes in the way that prediction error is evaluated – and the way
the specific mechanisms that the glucocorticoids use to optimize that it is weighted according its precision. LTP at glutamatergic
long-term synaptic efficacy. synapses of layer-1 inputs increases the precision of the prediction
Glucocorticoids act, among other things, on synapses that are error (Fig. 8B). LTP at GABAergic synapses of layer-1 inputs
located in the apical tuft of pyramidal cells. A layer-1-input axon decreases the prediction-error precision. However, if LTD is
excites either the dendrites of apical tuft or inhibits them using involved in glutamatergic or GABAergic synapses on the apical
axonal collaterals with an interposed GABA interneuron (Figs. 8 A tuft, such a long-term depression returns the input-output
and 9 A ) (Jiang et al., 2013). Glucocorticoids control the function to a default mode with a ‘neutral default slope’
maintenance of LTP or LTD at these glutamatergic and GABAergic (Fig. 8B). Therefore, high glucocorticoid concentrations may lead
synapses. The excitatory and inhibitory inputs arriving at the apical to ‘unlearning’ of the precision of prediction errors, while low
tuft play an important role in regulating the postsynaptic gain of concentrations of glucocorticoids may facilitate the learning of
the pyramidal neurons. Thus, glucocorticoids are in an ideal precision weighting (Liston et al., 2013; Liston and Gan, 2011). This
position to modulate the synaptic gain of pyramidal neurons in the may sound complicated; however, the brain has to (i) optimize the
long term. precision of ascending prediction errors (e.g., through creating NE
By controlling the postsynaptic gain of pyramidal neurons, the hotspots), it has to (ii) learn the right predictions (e.g., through
glucocorticoids may play a key role in learning the ‘precision of synaptic plasticity that is selectively consolidated during low
prediction errors’ (encoded by L2 neurons) and the ‘generative (normal) levels of glucocorticoid) and, finally, it has to (iii) learn
model’ (encoded by L5 neurons) (Figs. 8 A and 9 A). Dendritic gain how to optimize the precision of prediction errors (e.g., through
modulation at the apical tuft has been shown to increase the gain the interaction between NE and glucocorticoid levels described
of layer-5-pyramidal neurons (Larkum et al., 2004, 1999). The above).
combination of input to basal dendrites or soma with distal input In conclusion, glucocorticoids govern how the brain learns the
into the apical tuft increases the slope of the input-output function precision of prediction errors. Such a learning process allows us to
of the neuron. This increase in gain is due to the fact that the back- discriminate between trustworthy and imprecise sources of informa-
propagating Na+ action potentials interact with the weak distal tion.
synaptic input; such an interaction generates forward propagated
Ca++ action potentials that in turn elicit a burst of multiple action 3.2.5. Glucocorticoids gate learning of the generative model
potentials originating from the axonal spike initiation zone We can also optimize and learn the way we translate an
(Larkum, 2013; Larkum et al., 2004). Excitatory input to the apical expectation into a prediction. As noted above an empirical prior is
tuft increases the postsynaptic gain of the pyramidal neuron, while encoded by an expectation (i.e., the firing rates in layer-3-pyramidal
inhibitory input to the apical tuft reduces it (Figs. 8 B and 9 B). cells), and the likelihood by a prediction (represented in layer-5-
Whether LTP or LTD occurs at a synapse depends on two factors: pyramidal cells). The L3-‘empirical prior’ neuron provides the
first, correlated activity between the pre- and postsynaptic neuron input to the L5-‘likelihood’ neuron, which in turn responds with

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

18 A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx

the prediction. It has been shown experimentally that three input shrinkage of the apical dendrites prevent dendritic gain modula-
paths can influence the output of the pyramidal cell: driving input, tion (Liston and Gan, 2011; McEwen and Gianaros, 2011). In this
excitatory modulation input and inhibitory modulation input way, the specification of both the generative model and the
(Mehaffey et al., 2005; Silver, 2010). The driving input targets the precision of prediction errors are suspended. Thus, chronic stress
basal dendrites or the soma of the pyramidal cell. The excitatory results in functional and structural alterations, which can be
modulation input targets the apical tuft of the pyramidal cells that regarded as the deconstruction of the ‘internal representation of
is located in layer 1. The inhibitory modulation input – which the world’ – that was no longer appropriate. Such a deconstruction
consists of an interposed GABAergic elongated neuroglia form cell seems to be a prerequisite for rebuilding a new model of a world
(Jiang et al., 2013) – also targets the apical tuft. These modulation that is more fit for purpose.
input pathways allow modifying the slope of the input-output In summary, both functional and structural plasticity of pyramidal
function or postsynaptic gain of the pyramidal cell (Mehaffey et al., neurons show a bell-shaped dependency on glucocorticoids. Thus, the
2005; Silver, 2010). high concentrations of glucocorticoids during stress and uncertainty
Here, we apply this triple input concept to the connections allow the synaptic efficacy of apical tuft inputs to change over time.
between L3 empirical prior neurons and L5 prediction neurons Moreover, stress and uncertainty lead to the shrinkage of the distal
(Fig. 9) (Jiang et al., 2013; Thomson and Lamy, 2007). The excitatory apical dendrites. In this way, both the generative model and the
and inhibitory modulation paths increase or decrease the slope of synaptic mechanisms of precision or gain control are disassembled.
the L5-input-output function. Low (normal) glucocorticoid con- When stress is resolved and the situation is eased, the glucocorticoid
centrations facilitate LTP at glutamatergic and GABAergic synapses concentrations fall: Then the (synaptic efficacy) parameters that learn
on the apical tuft, thereby fixing an increased or decreased slope of the precision of prediction errors and the generative model are
the input-output function (Fig. 9B). An increase in the slope means consolidated – enabling the learning of an internal model when, and
that a given expectation (encoded by the L3-driving input) results only when, they are capable of resolving uncertainty and stress.
in a larger value of the prediction (encoded by the L5-output rate).
Thus, the functional relationship between expectation and 3.3. Habituation – updating of goal states
prediction can be learned. In contrast, LTD, which occurs at high
glucocorticoid concentrations, resets the slope of the input-output 3.3.1. Habituation
function to a default mode. Such a long-term depression may The third process for coping with uncertainty is stress
correspond to an ‘unlearning’ of the functional relationship habituation. How individuals react during stressful episodes
between expectation and prediction (Bennett et al., 1964). may change during the life course. When exposed to threatening
In summary, glucocorticoids also govern expectations generating changes in the external environment or the body milieu, people
predictions. A beneficial effect of glucocorticoids is that they enable us show two distinct genetically predisposed response patterns
to learn how to make optimal predictions in a particular situation. (Kirschbaum et al., 1995). Non-habituators maintain their high
stress responses when the stressful episodes recur. In contrast,
3.2.6. Structural plasticity at the apical tuft habituators show attenuation of their autonomic and endocrine
In pyramid cells, the glucocorticoids also influence how the responses over time. Stress habituation can be viewed as a special
structure of the dendritic tree is shaped. Brain circuitry can be form of learning, in which not only the glucocorticoids, but also the
remodeled by experience (Bennett et al., 1964), and stressful endocannabinoids play a central role (Hill et al., 2010; Patel and
experiences have functionally relevant effects on synapse number, Hillard, 2008). Crucially, habituation enables us to better
dendritic spine formation, and dendritic arbor shaping in many discriminate between conditions that should be avoided and
brain regions, including the hippocampus, amygdala, and the PFC conditions that could be tolerated. Habituation not only occurs in
(Liston et al., 2013; McEwen and Gianaros, 2011). Plasticity and the stress system but also in many other biological processes at the
remodeling are processes that consume a considerable amount of cellular and systemic level. With respect to habituation, the
energy (Placais and Preat, 2013). question arises whether the typical repetition-induced attenuation
In dendritic spine remodeling, learning leads to the induction of is caused by inhibitory mechanisms or by more sophisticated
spine formation, and successively, a portion of novel spines is (central) processes like ‘predictive coding’. In fact, there is
stabilized and a portion of existing spines is pruned (Hubener and experimental evidence supporting the view that habituation has
Bonhoeffer, 2010; Yang et al., 2009). Spine stabilization shows a all the hallmarks of predictive coding (Ramaswami, 2014;
bell-shaped dependency on glucocorticoids. High glucocorticoid Wacongne et al., 2012).
concentrations have been shown to favor postsynaptic dendritic
spine formation (GRs exert trophic effects via TrkB signaling) 3.3.2. Habituators can tolerate stress
(Ikeda et al., 2015; Jeanneteau et al., 2008), whereas low Problems occur, if we cannot appropriately update of our beliefs
glucocorticoid concentrations are required for the stabilization about current states and attainable states of the world. As a final
of freshly formed spines, the latter process being essential for resort, we can alleviate uncertainty by updating the beliefs about
memory consolidation (Liston et al., 2013). our goal states. This option can be regarded as a last option, because
Dendritic arbor shaping also depends on stress exposure. In the it entails revising our primary preferences or goals (that are usually
PFC of young animals, chronic stress leads to shrinkage of the distal held with high confidence or precision).
apical dendrites; after cessation of chronic stress, dendritic trees If habituators revise their prior preferences by broadening the
regrow (McEwen and Morrison, 2013). When animals recovered probability distribution over goal states (i.e., attenuating their
from stress, distal dendrites were not fully rebuilt, but proximal precision), KL divergences will decrease. As the goal state
dendrites showed hyperextension and spine growth, and deficits in probability distribution is broadened, beliefs about the attainable
synaptic plasticity were complete restored (Goldwater et al., states and the goal states start to overlap more. As can be seen from
2009). Such recovery after stress cessation is blunted by middle age Fig. 10, a less precise prior preference enables a greater overlap
and disappears in aged animals (McEwen and Morrison, 2013). between the beliefs about ‘attainable states’ and ‘goal states’
Long-term glucocorticoid treatment mimics the effects of chronic (compare Fig. 10A and B). Thus, habituation reduces the relative
stress; it leads to retraction of apical dendrites in the PFC risk (KL divergence) of one or more strategies (Fig. 10C). These
(Cerqueira et al., 2005). If uncertainty and stress persist, diverse changes in relative risk also ensure that one or more strategy can
processes like LTD, suspended spine stabilization, and the secure future wellbeing; thereby reducing uncertainty about

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx 19

future outcomes (Fig. 10D). Through these adaptive changes in the 4. Allostatic load
probability distribution over strategies, the habituated individual
becomes more confident about which strategy to select (Fig. 11). 4.1. The plasticity and vulnerability of the brain
Because a broadening of the beliefs about goal states reduces
uncertainty in many situations, habituators exhibit smaller Uncertainty about our responses to situations can be resolved
glucocorticoid and cardiovascular responses than non-habituators. by updating our beliefs about current states, attainable states, and
Accordingly, the glucocorticoid and cardiovascular responses are goal states. A major problem arises if Bayesian belief updating fails
lower in habituators as compared to non-habituators (Fig. 11; three to resolve uncertainty, leading to surprising or aversive outcomes.
asterisk) (Kirschbaum et al., 1995). In parallel, stress habituation A successful update would require a reorganization of the brain
improves the levels of self-esteem and locus of control; i.e. the self- architecture and neuronal circuits that includes synaptic plasticity
concept of own competence (Pruessner et al., 2005, 1999). In the (LTP) and structural plasticity (spine remodeling; rebuilding of the
context of habituators there is another tangible long-term benefit; dendritic trees) (McEwen et al., 2016). However, when updates
namely, escape from the tyranny of allostatic load. cannot avoid surprise, learning processes on different levels are
In short, allostatic load can be averted – at a subpersonal level – by suspended. Evidence accumulated that learning is suspended
reducing the precision of one's prior preferences; heuristically, when glucocorticoids are persistently secreted during chronic
adopting more realistic expectations about what can be achieved. stress (Bangasser and Shors, 2007; De Quervain et al., 1998; Lupien
et al., 1998). With high glucocorticoid concentrations, postsynaptic
dendritic spines are lost and dendritic branches shrink in various
parts of the cortex and the hippocampus (Dias-Ferreira et al., 2009;
Liston and Gan, 2011; Liston et al., 2006; Radley et al., 2006;
Watanabe et al., 1992; Wellman, 2001). Fluctuating concentrations
of glucocorticoids support a fine-tuned interplay between spine
formation, pruning and maintenance, whereas states of prolonged
glucocorticoid exposure interrupt this interplay (Liston et al.,
2013).
Thus, inappropriate updates of our internal model lead to high
glucocorticoid concentrations, and high glucocorticoid concen-
trations in turn prevent the conservation of such inappropriate
updates. Likewise, inappropriate updates can lead to disturbed
sleep or bad dreams (Antonijevic, 2008; Rodenbeck and Hajak,
2001). While sleep normally serves to optimize and conserve our
generative models (Hobson and Friston, 2012), poor sleep is likely
to preclude the revision of inappropriate models (Wagner and
Born, 2008). Similarly, if someone is on therapeutic or recreational
drugs that interfere with the cerebral mechanisms for mastering

Fig. 10. Habituation as ‘updating of goal states’. Panel A. In non-habituators the

goal states are fixed and remain unchanged. Panel B. In habituators the precision of
goal states is relaxed. Even though, for a given strategy, the habituators’ beliefs
about attainable states and the beliefs about goal states do not completely overlap,
they still exhibit more overlap than non-habituators. Panel C. The KL divergences for
each strategy are is smaller in habituators than in non-habituators. Thus, Fig. 11. Stress habituation. We define habituators as those who show repetition-
habituation decreases the risk. Panel D. Because a broadening of the beliefs about induced-response attenuation (neuroenergetic, neuroendocrine, emotional and
‘goal states’ reduces KL divergences, it changes the probability of each strategy that cardiovascular), when being chronically exposed to an inhospitable environment.
it may secure wellbeing. These changes in the probability distribution over When habituators are repeatedly exposed to the same homotypic stressor, they can
strategies mean that habituators become more confident about what strategy to reduce their uncertainty about which strategy they should select by redefining their
select. This explains why habituators exhibit smaller glucocorticoid responses than goal states (***).
non-habituators.

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

20 A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx

uncertainty, a temporary well-being may be achieved by relieving behaviors lead to inefficient mitochondrial metabolism, resulting
the allostatic network, but a successful Bayesian updating of his/ in reactive oxygen species (ROS) and inflammation and worsen
her internal model is suspended. systemic and brain pathology (Picard et al., 2014). These
As mentioned above, high glucocorticoid concentrations create pathologies include memory impairment, depression, myocardial
a ‘phase of change’ revising the current model of the world infarction, stroke, visceral fat accumulation, type 2 diabetes,
(including its strategies). If no appropriate update can be found, the muscle loss, osteoporosis, disturbed growth and reproduction
ACC continues to report high uncertainty (or entropy) about which (McEwen, 1998).
strategy to select. We suggest that the ACC-amygdala complex then On the one hand, a long-lasting energy crisis of the brain
sustains a hypervigilant state and hyperactivity of the allostatic damages the vascular system. To safeguard its high-energy
network; in particular, the key components SNS and the HPA axis – demand, the brain (via the amygdalae) sends a sympathetic
the latter maintaining high glucocorticoid concentrations. The message to the heart, thereby increasing heart rate, and in so doing
recurrent or persistent activation of the allostatic network then increases cardiac output in order to procure extra energy for itself.
leads to damaging adverse effects that lead to systemic and brain As stress-induced tachycardia increases flow velocity in the arterial
pathology. This is allostatic load (McEwen and Stellar, 1993). vascular system, high flow speed in turn increases the risk of
The healthy brain is resilient in the face of stressors and arterial turbulences (Falsetti et al., 1983) – particularly at
epigenetic cellular and molecular mechanisms produce continu- branching sites of the blood vessel system (Malek et al., 1999).
ous changes in gene expression. Thus, one cannot ‘roll back the ‘Adaptive vascular remodeling’ describes processes in the vascula-
clock’ after stress is over, so that we must speak of ‘resilience’ and ture that ameliorate such turbulences (Chatzizisis et al., 2007).
‘recovery’ rather than ‘reversal’ even though the alterations in However, if the capacity of ‘adaptive vascular remodeling’ is
neuronal structure and function may appear to have been overwhelmed, turbulences are likely to persist. In this case, there is
‘reversed’; yet they are not the same as before (McEwen et al., an increased risk that turbulences occur at predilection sites,
2015a,b; McEwen and Morrison, 2013). Acute and chronic stress thereby leading to atherosclerosis (Stone et al., 2012). Atheroscle-
interferes with cognition, decision making, anxiety and mood, and rosis in turn often causes myocardial infarction or stroke  and
in so doing affects systemic physiology through neuroendocrine, thus leads to an increased cardiovascular mortality. Just recently, a
autonomic, immune and metabolic mediators and multi-morbidi- team of cardiologists, psychiatrists and psychologists showed that
ty of disorders frequently occurs (McEwen, 2007; McEwen et al., resting amygdala activity independently and robustly predicted
2015b; Rasgon and McEwen, 2016). In the short run, increased cardiovascular disease events (Tawakol et al., 2017).
vigilance or anxiety in a hostile environment may be adaptive; On the other hand, a long-lasting energy crisis of the brain may
however, when the danger passes and the behavioral state and the damage the mitochondria, and in this way toxic products may
changes in neural circuitry become chronic, which get ‘stuck’, such accumulate, which can lead to systemic inflammation and
maladaptation may require an external intervention to get it accelerated cellular ageing (Du et al., 2009; Picard et al., 2014).
‘unstuck’, as is the case for chronic anxiety or depressive disorders Mitochondrial allostatic load can be brought about by elevated
(McEwen, 2007; McEwen et al., 2015b). glucocorticoid levels, even though low (i.e., normal) levels cause
Structural and functional allostatic plasticity is particularly translocation of glucocorticoid receptors into mitochondria to
evident in the hippocampus, a key structure for episodic and promote Ca++ sequestration and maintain low levels of free
spatial memory and mood regulation (McEwen, 2007; McEwen radicals (Du et al., 2009).
et al., 2015a). The hippocampus was the first brain structure
outside of the hypothalamus found to possess stress and sex 4.3. Comparison of mortality among habituators and non-habituators
hormone receptors and it provided a gateway into the hormone
sensitivity of the rest of the brain (McEwen et al., 2015b). Randomized controlled trials have shown that psychosocial
The amygdala involved in fear, anxiety and aggression and the stress is a factor that causes cardiovascular mortality, and that
prefrontal cortex, important for working memory and executive decreasing allostatic load through stress-relief programs reduces
function, both show functional and structural allostatic plasticity. cortisol responses and cardiovascular mortality (Gaab et al., 2003;
In the amygdala, overlapping waves of excessively high concen- Gulliksson et al., 2011; Hammerfald et al., 2006; Orth-Gomer et al.,
trations of glucocorticoids and norepinephrine cause an extended 2009; Storch et al., 2007). The special feature of these stress-relief
window of excitability (Karst and Joels, 2016). Such a prolonged programs is that participants can learn, among other things, to
window of excitability is thought to contribute to the development update their outdated models of the world in a cognitive manner,
of pathological conditions; e.g., posttraumatic stress disorder while the Bayesian Brain often updates its beliefs at a subpersonal
(Karst and Joels, 2016). Basolateral amygdala neurons expand level. There are also observational studies supporting the notion
dendrites from chronic stress (Chattarji et al., 2015) while, as noted that people who respond strongly to stressful challenges display a
earlier, medial PFC neurons, as well as hippocampal neurons, show high risk of atherosclerosis and cardiovascular mortality (Carroll
dendritic shrinkage from the same stress (McEwen and Morrison, et al., 2012; Everson et al., 1997; Hamer et al., 2010; Lynch et al.,
2013). 1998; Seldenrijk et al., 2012). Conversely, people who respond
weakly to stressful challenges exhibit a lower cardiovascular
4.2. Non-habituators are fully exposed to toxic stress mortality, even if they stay in an inhospitable environment. In this
way, habituators can alleviate their allostatic load when repeatedly
Because of continued uncertainty, the brain is constantly exposed to the same stressor. As a consequence of an alleviated
demanding for extra energy. Such an energy crisis with lack of allostatic load, habituators can tolerate an inhospitable environ-
habituation leads to allostatic load contributing to systemic and ment and are sheltered against cerebro- and cardiovascular events
brain pathology. This energy crisis has two consequences: first, (Carroll et al., 2012; Everson et al., 1997; Hamer et al., 2010; Lynch
SNS/HPA-axis hyperactivity and second, metabolic alterations and et al., 1998; Seldenrijk et al., 2012). Such a tolerance leads to side
stress-related health damaging behaviors (tobacco smoking, effects on the systemic energy metabolism of the individual, which
drinking alcohol, sleep deprivation). SNS/HPA-axis hyperactivity we will report elsewhere. Non-habituators, however, are fully
increases the risk of arterial blood flow turbulences, leading to exposed to toxic stress and thus are at increased risk of
atherosclerosis, thereby causing systemic and brain pathology cardiovascular death.
(Peters and McEwen, 2015). Metabolic alterations and poor health

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx 21

Box 5. Open Questions

There are still gaps in our knowledge of the neurobiological underpinnings of the Bayesian Brain concept. Future research may
address the following open questions:

 What is the functional role of fast-acting membrane MR and GR in the Bayesian Brain concept?
 How does the Bayesian Brain control the encoding and retrieval of emotional memories?
 What happens in the Bayesian Brain when posttraumatic stress disorder develops?
 How do alterations/adaptations of the Bayesian Brain affect systemic energy metabolism, i.e. promote the development of
anorexia or obesity?

In summary, habituators can update their ‘goal states’; as a consequently the LC, SNS and HPA-axis; also in this case, stress
consequence, they show attenuated responses when recurrently facilitates a shift in strategy (Figs. 2 and 3). The Selfish Bayesian
challenged, and in this way – although they continue to live in the Brain procures itself with extra thermodynamic energy (enhanced
inhospitable environment – they show a barely limited life expectancy. glucose allocation to the brain) in order to minimize variational
In contrast, non-habituators display a drastically shortened life free energy (prediction errors) by searching for a novel strategy.
expectancy; but they are the individuals who would benefit most from In light of the foregoing, we define ‘stress’ as the individual state of
stress-relief programs that include cognitive restructuring, social skills uncertainty about what needs to be done to safeguard physical,
development, and mindfulness (Gulliksson et al., 2011; Orth-Gomer mental or social well-being.
et al., 2009).
6. Conclusions
5. Updating the ‘stress definition’
We have introduced an information-theoretic account of stress
As suggested in the current paper, stress occurs, if we are and allostatic load based upon recent developments in theoretical
surprised by our sensations and we are uncertain about what to do neurobiology. In particular, we have established the link between
to safeguard our physical, mental or social wellbeing. Surprises can the Bayesian Brain and the Selfish Brain in terms of minimizing
be manifold and can concern our internal body milieu (lack of variational and metabolic free energy respectively. When
energy, lack of oxygen, loss of blood, infection, toxins, trauma, unpacked, the theoretical considerations provide a remarkable
myocardial infarction, etc.) or our external environment (social level of explanatory detail; particularly in relation to the role of
conflicts, overload/underload, disordered neighborhood, mobbing, norepinephrine in nuancing perceptual inference (and action)
discrimination, social defeat, etc.). Activation of the SNS or the HPA through its effects on presynaptic gain control – and implicit
axis is typical in all these situations, but is not considered here to effects on the ability of sensory evidence to revise beliefs about
be a sufficient criterion for ‘stress’. hidden causes in the world or the body. Having established the
Two examples illustrate the principle how we react upon close relationship between the roles of synaptic activity and
changes in the internal and external environment: fasting and efficacy (i.e., precision) in optimizing perceptual inference in
marital dispute. Fasting – as referred to as the metabolic state situations of uncertainty, we then went on to look at the permissive
achieved after complete digestion and absorption of a meal – goes role of glucocorticoids in learning. This analysis suggests that there
along with the activation of the SNS and HPA-axis. Such SNS- and is an optimum glucocorticoid level that enables the consolidation
HPA-axis activations serve to adequately supply the brain with of activity-dependent plasticity (where the ‘bell-shaped’ glucocor-
energy (glucose, ketones, or lactate) (Kubera et al., 2012a, 2014). ticoid-dependency curve has its peak) (Joels, 2006) that mediates
Since the brain-energy concentrations are tightly regulated experience-dependent learning – when and only when, our
(Oltmanns et al., 2008), such increases in SNS and HPA-axis generative models are sufficient to resolve uncertainty, stress
activity are common in everyday life (Peters and Langemann, and elevated glucocorticoid levels.
2009). The Selfish Brain procures itself with energy (Peters et al., Finally, we considered long-term processes that could minimize
2007b). After a few hours of fasting, we feel hunger and perceive variational free energy and stress by looking at the ultimate cause;
SNS-induced interoceptive signals like nervousness, weakness, namely, the discrepancy between states that we can attain by
tremor, tachycardia, dizziness, and sweating. The (Selfish) Bayesian acting on the world and the states we a priori expect to occupy (i.e.,
Brain uses perceptual inference to infer the cause (lack of interoceptive, proprioceptive, emotional and prosocial goals). Our
thermodynamic energy) from the effect (the interoceptive signals). key observation is that exposure to chronic stress – and the
Then it uses active inference (food seeking behavior) to minimize allostatic load that this entails – can be remediated by revising our
variational free energy; i.e., to eliminate the prediction errors highest-level prior beliefs; namely, prior expectations about the
(hunger, autonomic symptoms). If we are certain that food would states we aspire to. This provides a nice metaphor that
be available soon, the appropriate action is selected and ‘no stress’ distinguishes between habituators and non-habituators in re-
occurs. However, if we are uncertain about whether we might get sponse to chronic stress.
food at all, stress occurs – as is the case in ‘food insecurity’ Functional and structural remodeling of the neural architecture
(Bhattacharya et al., 2004). In such a case, uncertainty (entropy) often makes it possible to avoid or master states of uncertainty,
monitored by the ACC stimulates the amygdala, and in so doing anxiety, and hypervigilance. ‘Good Stress’ denotes an episode of
increases LC, SNS and HPA-axis activity; in this way, stress uncertainty in which the beneficial effects of stress responses
facilitates the search for a novel strategy (Figs. 2 and 3). support a successful Bayesian updating and learning of the internal
In a marital dispute, surprising external sensations may model of the world. As a result, uncertainty is resolved. ‘Tolerable
immediately evoke a state of uncertainty, in which one does not stress’ characterizes a situation in which habituation leads to a
know how to resolve the situation. This is also ‘stress’. The partial reduction in uncertainty; this reduction is achieved by the
uncertainty monitored by the ACC stimulates the amygdala, and adjustment of the primary goals. In this case, the damaging effects

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

22 A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx

of persistent stress reactions can be kept at bay. Of note, individuals Cerqueira, J.J., Pego, J.M., Taipa, R., Bessa, J.M., Almeida, O.F., Sousa, N., 2005.
cannot always resolve uncertainty by reconstructing their internal Morphological correlates of corticosteroid-induced changes in prefrontal
cortex-dependent behaviors. J. Neurosci. 25, 7792–7800.
model of the world. The inhospitable environment may also limit Chan, O., Lawson, M., Zhu, W., Beverly, J.L., Sherwin, R.S., 2007. ATP-sensitive K(+)
such a resolution, e.g. when the individuals live in poverty, in war- channels regulate the release of GABA in the ventromedial hypothalamus
afflicted areas, are long-term unemployed or victims of discrimi- during hypoglycemia. Diabetes 56, 1120–1126.
Chanes, L., Barrett, L.F., 2016. Redefining the role of limbic areas in cortical
nation (Kubera et al., 2016; Ludwig et al., 2012; Puhl and Heuer, processing. Trends Cogn. Sci. 20, 96–106.
2009). If the resolution of uncertainty is achieved too late or is not Chattarji, S., Tomar, A., Suvrathan, A., Ghosh, S., Rahman, M.M., 2015. Neighborhood
possible at all, the adverse effects of the futile and brain-energy- matters: divergent patterns of stress-induced plasticity across the brain. Nat.
Neurosci. 18, 1364–1375.
consuming efforts for the resolution come to the fore, and the Chatzizisis, Y.S., Coskun, A.U., Jonas, M., Edelman, E.R., Feldman, C.L., Stone, P.H.,
ongoing brain-energy crisis leads to allostatic load that contributes 2007. Role of endothelial shear stress in the natural history of coronary
to systemic and brain pathology. ‘Toxic stress’ refers to such a atherosclerosis and vascular remodeling: molecular cellular, and vascular
behavior. J. Am. Coll. Cardiol. 49, 2379–2393.
chronic condition in which the damaging effects of the stress
Chennu, S., Noreika, V., Gueorguiev, D., Blenkmann, A., Kochen, S., Ibanez, A., Owen,
responses prevail and the uncertainty can neither be resolved by a A.M., Bekinschtein, T.A., 2013. Expectation and attention in hierarchical
successful Bayesian update nor reduced by habituation. If, in such a auditory prediction. J. Neurosci. 33, 11194–11205.
situation, the brain ‘gets stuck’ and does not recover, external Chiu, K.M., Lin, T.Y., Lu, C.W., Wang, S.J., 2011. Inhibitory effect of glutamate release
from rat cerebrocortical nerve terminals by alpha2 adrenoceptor agonist
intervention is required. dexmedetomidine. Eur. J. Pharmacol. 670, 137–147.
The potential utility of formulating stress and allostatic load in Datson, N.A., van der, P.J., de Kloet, E.R., Vreugdenhil, E., 2001. Identification of
terms of a Selfish Bayesian Brain is that one might use the corticosteroid-responsive genes in rat hippocampus using serial analysis of
gene expression. Eur. J. Neurosci. 14, 675–689.
principles afforded by theoretical neurobiology to organize de Berker, A.O., Rutledge, R.B., Mathys, C., Marshall, L., Cross, G.F., Dolan, R.J.,
existing and future empirical results (Box 5; ‘Open questions’). Bestmann, S., 2016. Computations of uncertainty mediate acute stress
Furthermore, the interrelationships brain function and physiology responses in humans. Nat. Commun. 7, 10996.
de Kloet, E.R., Vreugdenhil, E., Oitzl, M.S., Joels, M., 1998. Brain corticosteroid
are laid bare in the sense that everything appears to be the game of receptor balance in health and disease. Endocr. Rev. 19, 269–301.
reducing uncertainty, variational free energy, surprise, or more De Quervain, D.J., Roozendaal, B., McGaugh, J.L., 1998. Stress and glucocorticoids
simply stress. Conceptual principles of this sort may be useful, not impair retrieval of long-term spatial memory. Nature 394, 787–790.
Deuchars, J., Thomson, A.M., 1995. Innervation of burst firing spiny interneurons by
just from a scientific perspective, but as organizing frameworks for
pyramidal cells in deep layers of rat somatomotor cortex: paired intracellular
cognitive behavioral therapy in stress disorders. recordings with biocytin filling. Neuroscience 69, 739–755.
Devilbiss, D.M., Waterhouse, B.D., 2004. The effects of tonic locus ceruleus output on
sensory-evoked responses of ventral posterior medial thalamic and barrel field
References
cortical neurons in the awake rat. J. Neurosci. 24, 10773–10785.
Diamond, D.M., Bennett, M.C., Fleshner, M., Rose, G.M., 1992. Inverted-U
Adams, R.A., Shipp, S., Friston, K.J., 2013. Predictions not commands: active relationship between the level of peripheral corticosterone and the magnitude
inference in the motor system. Brain Struct. Funct. 218, 611–643. of hippocampal primed burst potentiation. Hippocampus 2, 421–430.
Ahren, B., 2000. Autonomic regulation of islet hormone secretion-implications for Dias-Ferreira, E., Sousa, J.C., Melo, I., Morgado, P., Mesquita, A.R., Cerqueira, J.J., Costa,
health and disease. Diabetologia 43, 393–410. R.M., Sousa, N., 2009. Chronic stress causes frontostriatal reorganization and
Akana, S.F., Jacobson, L., Cascio, C.S., Shinsako, J., Dallman, M.F., 1988. Constant affects decision-making. Science 325, 621–625.
corticosterone replacement normalizes basal adrenocorticotropin (ACTH) but Du, J., Wang, Y., Hunter, R., Wei, Y., Blumenthal, R., Falke, C., Khairova, R., Zhou, R.,
permits sustained ACTH hypersecretion after stress in adrenalectomized rats. Yuan, P., Machado-Vieira, R., McEwen, B.S., Manji, H.K., 2009. Dynamic
Endocrinology 122, 1337–1342. regulation of mitochondrial function by glucocorticoids. Proc. Natl. Acad. Sci. U.
Antonijevic, I., 2008. HPA axis and sleep: identifying subtypes of major depression. S. A. 106, 3543–3548.
Stress 11, 15–27. Everson, S.A., Lynch, J.W., Chesney, M.A., Kaplan, G.A., Goldberg, D.E., Shade, S.B.,
Arriza, J.L., Simerly, R.B., Swanson, L.W., Evans, R.M., 1988. The neuronal Cohen, R.D., Salonen, R., Salonen, J.T., 1997. Interaction of workplace demands
mineralocorticoid receptor as a mediator of glucocorticoid response. Neuron 1, and cardiovascular reactivity in progression of carotid atherosclerosis:
887–900. population based study. BMJ 314, 553–558.
Aston-Jones, G., Cohen, J.D., 2005. An integrative theory of locus coeruleus- Falsetti, H.L., Carroll, R.J., Swope, R.D., Chen, C.J., 1983. Turbulent blood flow in the
norepinephrine function: adaptive gain and optimal performance. Annu. Rev. ascending aorta of dogs. Cardiovasc. Res. 17, 427–436.
Neurosci. 28, 403–450. Fehm, H.L., Voigt, K.H., Lang, R.E., Beinert, K.E., Kummer, G.W., Pfeiffer, E.F., 1977.
Bangasser, D.A., Shors, T.J., 2007. The hippocampus is necessary for enhancements Paradoxical ACTH response to glucocorticoids in Cushing's disease. N. Engl. J.
and impairments of learning following stress. Nat. Neurosci. 10, 1401–1403. Med. 297, 904–907.
Barrett, L.F., Simmons, W.K., 2015. Interoceptive predictions in the brain. Nat. Rev. Feinstein, J.S., Stein, M.B., Paulus, M.P., 2006. Anterior insula reactivity during
Neurosci. 16, 419–429. certain decisions is associated with neuroticism. Soc. Cogn. Affect. Neurosci. 1,
Barrett, L.F., 2017. The theory of constructed emotion: an active inference account of 136–142.
interoception and categorization. Soc. Cogn. Affect. Neurosci. 12, 1–23. Feldmeyer, D., Roth, A., Sakmann, B., 2005. Monosynaptic connections between
Barron, H.C., Garvert, M.M., Behrens, T.E., 2015. Reassessing VMPFC: full of pairs of spiny stellate cells in layer 4 and pyramidal cells in layer 5A indicate that
confidence? Nat. Neurosci. 18, 1064–1066. lemniscal and paralemniscal afferent pathways converge in the infragranular
Bastos, A.M., Usrey, W.M., Adams, R.A., Mangun, G.R., Fries, P., Friston, K.J., 2012. somatosensory cortex. J. Neurosci. 25, 3423–3431.
Canonical microcircuits for predictive coding. Neuron 76, 695–711. Felleman, D.J., Van, E.D.C., 1991. Distributed hierarchical processing in the primate
Bechara, A., Tranel, D., Damasio, H., 2000. Characterization of the decision-making cerebral cortex. Cereb. Cortex 1, 1–47.
deficit of patients with ventromedial prefrontal cortex lesions. Brain 123 (Pt 11), Ferrero, J.J., Alvarez, A.M., Ramirez-Franco, J., Godino, M.C., Bartolome-Martin, D.,
2189–2202. Aguado, C., Torres, M., Lujan, R., Ciruela, F., Sanchez-Prieto, J., 2013. beta-
Behrens, T.E., Woolrich, M.W., Walton, M.E., Rushworth, M.F., 2007. Learning the Adrenergic receptors activate exchange protein directly activated by cAMP
value of information in an uncertain world. Nat. Neurosci. 10, 1214–1221. (Epac) translocate Munc13-1, and enhance the Rab3A-RIM1alpha interaction to
Bennett, E.L., Diamond, M.C., Krech, D., Rosenzweig, M.R., 1964. Chemical and potentiate glutamate release at cerebrocortical nerve terminals. J. Biol. Chem.
anatomical plasticity brain. Science 146, 610–619. 288, 31370–31385.
Berridge, C.W., Waterhouse, B.D., 2003. The locus coeruleus-noradrenergic system: Frühwald-Schultes, B., Kern, W., Born, J., Fehm, H.L., Peters, A., 2000. Comparison of
modulation of behavioral state and state-dependent cognitive processes. Brain the inhibitory effect of insulin and hypoglycemia on insulin secretion in
Res. Brain Res. Rev. 42, 33–84. humans. Metabolism 49, 950–953.
Berut, A., Arakelyan, A., Petrosyan, A., Ciliberto, S., Dillenschneider, R., Lutz, E., 2012. Friston, K., Kilner, J., Harrison, L., 2006. A free energy principle for the brain. J.
Experimental verification of Landauer's principle linking information and Physiol. Paris 100, 70–87.
thermodynamics. Nature 483, 187–189. Friston, K., Schwartenbeck, P., Fitzgerald, T., Moutoussis, M., Behrens, T., Dolan, R.J.,
Bhattacharya, J., Currie, J., Haider, S., 2004. Poverty, food insecurity, and nutritional 2013. The anatomy of choice: active inference and agency. Front. Hum. Neurosci.
outcomes in children and adults. J. Health Econ. 23, 839–862. 7, 598.
Brillouin, L., 1953. The negentropy principle of information. J. Appl. Phys. 24, 1152– Friston, K., Schwartenbeck, P., Fitzgerald, T., Moutoussis, M., Behrens, T., Dolan, R.J.,
1163. 2014. The anatomy of choice: dopamine and decision-making. Philos. Trans. R.
Brown, H.R., Friston, K.J., 2012. Dynamic causal modelling of precision and synaptic Soc. Lond. B Biol. Sci. 369.
gain in visual perception—an EEG study. Neuroimage 63, 223–231. Friston, K., Rigoli, F., Ognibene, D., Mathys, C., Fitzgerald, T., Pezzulo, G., 2015. Active
Carroll, D., Ginty, A.T., Der, G., Hunt, K., Benzeval, M., Phillips, A.C., 2012. Increased inference and epistemic value. Cogn. Neurosci. 6, 187–214.
blood pressure reactions to acute mental stress are associated with 16-year Friston, K., 2005. A theory of cortical responses. Philos. Trans. R. Soc. Lond. B Biol. Sci.
cardiovascular disease mortality. Psychophysiology 49, 1444–1448. 360, 815–836.

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx 23

Friston, K., 2009. The free-energy principle: a rough guide to the brain? Trends Cogn. repeated psychological stress in a subpopulation of healthy men. Psychosom.
Sci. 13, 293–301. Med. 57, 468–474.
Friston, K., 2010. The free-energy principle: a unified brain theory? Nat. Rev. Kobayashi, M., Kojima, M., Koyanagi, Y., Adachi, K., Imamura, K., Koshikawa, N.,
Neurosci. 11, 127–138. 2009. Presynaptic and postsynaptic modulation of glutamatergic synaptic
Friston, K., 2011. Embodied Inference: or I think therefore I am, if I am what I think. transmission by activation of alpha(1)- and beta-adrenoceptors in layer V
In: Tschacher, W., Bergomi, C. (Eds.), The Implications of Embodiment pyramidal neurons of rat cerebral cortex. Synapse 63, 269–281.
(Cognition and Communication). Imprint Academic Exeter. Kok, P., de Lange, F.P., 2014. Shape perception simultaneously up- and
Gaab, J., Blattler, N., Menzi, T., Pabst, B., Stoyer, S., Ehlert, U., 2003. Randomized downregulates neural activity in the primary visual cortex. Curr. Biol. 24, 1531–
controlled evaluation of the effects of cognitive-behavioral stress management 1535.
on cortisol responses to acute stress in healthy subjects. Konig, P., Engel, A.K., Singer, W., 1996. Integrator or coincidence detector?: The role
Psychoneuroendocrinology 28, 767–779. of the cortical neuron revisited. Trends Neurosci. 19, 130–137.
Garrido, M.I., Kilner, J.M., Kiebel, S.J., Friston, K.J., 2007. Evoked brain responses are Koolhaas, J.M., Bartolomucci, A., Buwalda, B., de Boer, S.F., Flugge, G., Korte, S.M.,
generated by feedback loops. Proc. Natl. Acad. Sci. U. S. A. 104, 20961–20966. Meerlo, P., Murison, R., Olivier, B., Palanza, P., Richter-Levin, G., Sgoifo, A.,
Goldwater, D.S., Pavlides, C., Hunter, R.G., Bloss, E.B., Hof, P.R., McEwen, B.S., Steimer, T., Stiedl, O., van Dijk, G., Wohr, M., Fuchs, E., 2011. Stress revisited: a
Morrison, J.H., 2009. Structural and functional alterations to rat medial critical evaluation of the stress concept. Neurosci. Biobehav. Rev. 35, 1291–1301.
prefrontal cortex following chronic restraint stress and recovery. Neuroscience Kubera, B., Hubold, C., Otte, S., Lindenberg, A.S., Zeiss, I., Krause, R., Steinkamp, M.,
164, 798–808. Klement, J., Entringer, S., Pellerin, L., Peters, A., 2012a. Rise in plasma lactate
Gong, Q.Y., Roberts, N., Garden, A.S., Whitehouse, G.H., 1998. Fetal and fetal brain concentrations with psychosocial stress: a possible sign of cerebral energy
volume estimation in the third trimester of human pregnancy using gradient demand. Obes. Facts 5, 384–392.
echo MR imaging. Magn. Reson. Imaging 16, 235–240. Kubera, B., Hubold, C., Zug, S., Wischnath, H., Wilhelm, I., Hallschmid, M., Entringer,
Gottfried, J.A., O'Doherty, J., Dolan, R.J., 2003. Encoding predictive reward value in S., Langemann, D., Peters, A., 2012b. The brain's supply and demand in obesity.
human amygdala and orbitofrontal cortex. Science 301, 1104–1107. Front. Neuroenergetics 4, 4. doi:http://dx.doi.org/10.3389/fnene.2012.00004.
Gulliksson, M., Burell, G., Vessby, B., Lundin, L., Toss, H., Svardsudd, K., 2011. Kubera, B., Hubold, C., Wischnath, H., Zug, S., Peters, A., 2014. Rise of ketone bodies
Randomized controlled trial of cognitive behavioral therapy vs standard with psychosocial stress in normal weight men. Psychoneuroendocrinology 45,
treatment to prevent recurrent cardiovascular events in patients with coronary 43–48.
heart disease: Secondary Prevention in Uppsala Primary Health Care project Kubera, B., Klement, J., Wagner, C., Radel, C., Eggeling, J., Fullbrunn, S., Kaczmarek, M.
(SUPRIM). Arch. Intern. Med. 171, 134–140. C., Levinsky, R., Peters, A., 2016. Differences in fairness and trust between lean
Haeusler, S., Maass, W., 2007. A statistical analysis of information-processing and corpulent men. Int. J. Obes. (Lond.) 40, 1802–1808.
properties of lamina-specific cortical microcircuit models. Cereb. Cortex 17, Kullback, S., Leibler, R.A., 1951. On information and sufficiency. Ann. Math. Stat. 22,
149–162. 79.
Hamer, M., O'Donnell, K., Lahiri, A., Steptoe, A., 2010. Salivary cortisol responses to Langemann, D., Pellerin, L., Peters, A., 2008. Making sense of AMPA receptor
mental stress are associated with coronary artery calcification in healthy men trafficking by modeling molecular mechanisms of synaptic plasticity. Brain Res.
and women. Eur. Heart J. 31, 424–429. 1207, 60–72.
Hammerfald, K., Eberle, C., Grau, M., Kinsperger, A., Zimmermann, A., Ehlert, U., Larkman, A.U., Jack, J.J., Stratford, K.J., 1997. Quantal analysis of excitatory synapses
Gaab, J., 2006. Persistent effects of cognitive-behavioral stress management on in rat hippocampal CA1 in vitro during low-frequency depression. J. Physiol. 505
cortisol responses to acute stress in healthy subjects-a randomized controlled (Pt 2), 457–471.
trial. Psychoneuroendocrinology 31, 333–339. Larkum, M.E., Zhu, J.J., Sakmann, B., 1999. A new cellular mechanism for coupling
Harris, K.D., Shepherd, G.M., 2015. The neocortical circuit: themes and variations. inputs arriving at different cortical layers. Nature 398, 338–341.
Nat. Neurosci. 18, 170–181. Larkum, M.E., Senn, W., Luscher, H.R., 2004. Top-down dendritic input increases the
Harris, J.J., Jolivet, R., Attwell, D., 2012. Synaptic energy use and supply. Neuron 75, gain of layer 5 pyramidal neurons. Cereb. Cortex 14, 1059–1070.
762–777. Larkum, M., 2013. A cellular mechanism for cortical associations: an organizing
Hasselbalch, S.G., Knudsen, G.M., Videbaek, C., Pinborg, L.H., Schmidt, J.F., Holm, S., principle for the cerebral cortex. Trends Neurosci. 36, 141–151.
Paulson, O.B., 1999. No effect of insulin on glucose blood-brain barrier transport Lee, D., Rushworth, M.F., Walton, M.E., Watanabe, M., Sakagami, M., 2007. Functional
and cerebral metabolism in humans. Diabetes 48, 1915–1921. specialization of the primate frontal cortex during decision making. J. Neurosci.
Hermans, E.J., van Marle, H.J., Ossewaarde, L., Henckens, M.J., Qin, S., van Kesteren, 27, 8170–8173.
M.T., Schoots, V.C., Cousijn, H., Rijpkema, M., Oostenveld, R., Fernandez, G., 2011. Levy, W.B., Baxter, R.A., 1996. Energy efficient neural codes. Neural Comput. 8, 531–
Stress-related noradrenergic activity prompts large-scale neural network 543.
reconfiguration. Science 334, 1151–1153. Liljeholm, M., Wang, S., Zhang, J., O'Doherty, J.P., 2013. Neural correlates of the
Hill, M.N., McLaughlin, R.J., Bingham, B., Shrestha, L., Lee, T.T., Gray, J.M., Hillard, C.J., divergence of instrumental probability distributions. J. Neurosci. 33, 12519–
Gorzalka, B.B., Viau, V., 2010. Endogenous cannabinoid signaling is essential for 12527.
stress adaptation. Proc. Natl. Acad. Sci. U. S. A. 107, 9406–9411. Linster, C., Nai, Q., Ennis, M., 2011. Nonlinear effects of noradrenergic modulation of
Hirono, M., Obata, K., 2006. Alpha-adrenoceptive dual modulation of inhibitory olfactory bulb function in adult rodents. J. Neurophysiol. 105, 1432–1443.
GABAergic inputs to Purkinje cells in the mouse cerebellum. J. Neurophysiol. 95, Liston, C., Gan, W.B., 2011. Glucocorticoids are critical regulators of dendritic spine
700–708. development and plasticity in vivo. Proc. Natl. Acad. Sci. U. S. A. 108, 16074–
Hitze, B., Hubold, C., van Dyken, R., Schlichting, K., Lehnert, H., Entringer, S., Peters, 16079.
A., 2010. How the selfish brain organizes its ‘Supply and demand’. Front. Liston, C., Miller, M.M., Goldwater, D.S., Radley, J.J., Rocher, A.B., Hof, P.R., Morrison, J.
Neuroenergetics 2 doi:http://dx.doi.org/10.3389/fnene.2010.00007. H., McEwen, B.S., 2006. Stress-induced alterations in prefrontal cortical
Hobson, J.A., Friston, K.J., 2012. Waking and dreaming consciousness: dendritic morphology predict selective impairments in perceptual attentional
neurobiological and functional considerations. Prog. Neurobiol. 98, 82–98. set-shifting. J. Neurosci. 26, 7870–7874.
Hohwy, J., 2013. The Predictive Brain. Oxford University Press, Oxford. Liston, C., Cichon, J.M., Jeanneteau, F., Jia, Z., Chao, M.V., Gan, W.B., 2013. Circadian
Hubener, M., Bonhoeffer, T., 2010. Searching for engrams. Neuron 67, 363–371. glucocorticoid oscillations promote learning-dependent synapse formation and
Ikeda, M., Hojo, Y., Komatsuzaki, Y., Okamoto, M., Kato, A., Takeda, T., Kawato, S., maintenance. Nat. Neurosci. 16, 698–705.
2015. Hippocampal spine changes across the sleep-wake cycle: corticosterone Lubke, J., Egger, V., Sakmann, B., Feldmeyer, D., 2000. Columnar organization of
and kinases. J. Endocrinol. 226, M13–27. dendrites and axons of single and synaptically coupled excitatory spiny neurons
Jacobson, L., Akana, S.F., Cascio, C.S., Shinsako, J., Dallman, M.F., 1988. Circadian in layer 4 of the rat barrel cortex. J. Neurosci. 20, 5300–5311.
variations in plasma corticosterone permit normal termination of Ludwig, J., Duncan, G.J., Gennetian, L.A., Katz, L.F., Kessler, R.C., Kling, J.R.,
adrenocorticotropin responses to stress. Endocrinology 122, 1343–1348. Sanbonmatsu, L., 2012. Neighborhood effects on the long-term well-being of
Jeanneteau, F., Garabedian, M.J., Chao, M.V., 2008. Activation of Trk neurotrophin low-income adults. Science 337, 1505–1510.
receptors by glucocorticoids provides a neuroprotective effect. Proc. Natl. Acad. Lupien, S.J., de Leon, M., de Santi, S., Convit, A., Tarshish, C., Nair, N.P., Thakur, M.,
Sci. U. S. A. 105, 4862–4867. McEwen, B.S., Hauger, R.L., Meaney, M.J., 1998. Cortisol levels during human
Jiang, X., Wang, G., Lee, A.J., Stornetta, R.L., Zhu, J.J., 2013. The organization of two aging predict hippocampal atrophy and memory deficits. Nat. Neurosci. 1, 69–
new cortical interneuronal circuits. Nat. Neurosci. 16, 210–218. 73.
Joels, M., 2006. Corticosteroid effects in the brain: u-shape it. Trends Pharmacol. Sci. Lynch, J.W., Everson, S.A., Kaplan, G.A., Salonen, R., Salonen, J.T., 1998. Does low
27, 244–250. socioeconomic status potentiate the effects of heightened cardiovascular
Karlsson, M.P., Tervo, D.G., Karpova, A.Y., 2012. Network resets in medial prefrontal responses to stress on the progression of carotid atherosclerosis? Am. J. Public
cortex mark the onset of behavioral uncertainty. Science 338, 135–139. Health 88, 389–394.
Karst, H., Joels, M., 2016. Severe stress hormone conditions cause an extended Lyons, D.M., Levine, S., 1994. Socioregulatory effects on squirrel monkey pituitary-
window of excitability in the mouse basolateral amygdala. Neuropharmacology adrenal activity: a longitudinal analysis of cortisol and ACTH.
110, 175–180. Psychoneuroendocrinology 19, 283–291.
Kind, K.L., Roberts, C.T., Sohlstrom, A.I., Katsman, A., Clifton, P.M., Robinson, J.S., Madsen, P.L., Hasselbalch, S.G., Hagemann, L.P., Olsen, K.S., Bulow, J., Holm, S.,
Owens, J.A., 2005. Chronic maternal feed restriction impairs growth but Wildschiodtz, G., Paulson, O.B., Lassen, N.A., 1995. Persistent resetting of the
increases adiposity of the fetal guinea pig. Am. J. Physiol. Regul. Integr. Comp. cerebral oxygen/glucose uptake ratio by brain activation: evidence obtained
Physiol. 288, R119–R126. with the Kety-Schmidt technique. J. Cereb. Blood Flow Metab. 15, 485–491.
Kirschbaum, C., Pruessner, J.C., Stone, A.A., Federenko, I., Gaab, J., Lintz, D., Maggio, N., Segal, M., 2007. Striking variations in corticosteroid modulation of long-
Schommer, N., Hellhammer, D.H., 1995. Persistent high cortisol responses to term potentiation along the septotemporal axis of the hippocampus. J. Neurosci.
27, 5757–5765.

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

24 A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx

Maggio, N., Segal, M., 2009. Differential modulation of long-term depression by Stockholm Women's Intervention Trial for Coronary Heart Disease (SWITCHD).
acute stress in the rat dorsal and ventral hippocampus. J. Neurosci. 29, 8633– Circ. Cardiovasc. Qual. Outcomes 2, 25–32.
8638. Panagiotaropoulos, T.I., Deco, G., Kapoor, V., Logothetis, N.K., 2012. Neuronal
Maggio, N., Segal, M., 2012. Steroid modulation of hippocampal plasticity: switching discharges and gamma oscillations explicitly reflect visual consciousness in the
between cognitive and emotional memories. Front. Cell. Neurosci. 6, 12. lateral prefrontal cortex. Neuron 74, 924–935.
Magistretti, P.J., Pellerin, L., Rothman, D.L., Shulman, R.G., 1999. Energy on demand. Patel, S., Hillard, C.J., 2008. Adaptations in endocannabinoid signaling in response to
Science 283, 496–497. repeated homotypic stress: a novel mechanism for stress habituation. Eur. J.
Malek, A.M., Alper, S.L., Izumo, S., 1999. Hemodynamic shear stress and its role in Neurosci. 27, 2821–2829.
atherosclerosis. JAMA 282, 2035–2042. Patel, P.D., Lopez, J.F., Lyons, D.M., Burke, S., Wallace, M., Schatzberg, A.F., 2000.
Markram, H., Lubke, J., Frotscher, M., Roth, A., Sakmann, B., 1997. Physiology and Glucocorticoid and mineralocorticoid receptor mRNA expression in squirrel
anatomy of synaptic connections between thick tufted pyramidal neurones in monkey brain. J. Psychiatr. Res. 34, 383–392.
the developing rat neocortex. J. Physiol. 500 (Pt 2), 409–440. Paulus, M.P., Hozack, N., Frank, L., Brown, G.G., 2002. Error rate and outcome
Mason, J.W., 1959. Hormones and metabolism: psychological influences on the predictability affect neural activation in prefrontal cortex and anterior cingulate
pituitary-adrenal cortical. In: Pincus, C. (Ed.), Recent Progress In Hormone during decision-making. Neuroimage 15, 836–846.
Research: Proceedings of the Laurentian Hormone Conference 1958. Academic Pavlides, C., Kimura, A., Magarinos, A.M., McEwen, B.S., 1994. Type I adrenal steroid
Press, New York and London. receptors prolong hippocampal long-term potentiation. Neuroreport 5, 2673–
Mason, J.W., 1968. A review of psychoendocrine research on the pituitary-adrenal 2677.
cortical system. Psychosom. Med. 30 (Suppl), 576–607. Pavlides, C., Kimura, A., Magarinos, A.M., McEwen, B.S., 1995. Hippocampal
Mather, M., Clewett, D., Sakaki, M., Harley, C.W., 2015. Norepinephrine ignites local homosynaptic long-term depression/depotentiation induced by adrenal
hot spots of neuronal excitation: how arousal amplifies selectivity in perception steroids. Neuroscience 68, 379–385.
and memory. Behav. Brain Sci. 1–100. Pavlides, C., Ogawa, S., Kimura, A., McEwen, B.S., 1996. Role of adrenal steroid
McEwen, B.S., Gianaros, P.J., 2010. Central role of the brain in stress and adaptation: mineralocorticoid and glucocorticoid receptors in long-term potentiation in the
links to socioeconomic status, health, and disease. Ann. N. Y. Acad. Sci. 1186, CA1 field of hippocampal slices. Brain Res. 738, 229–235.
190–222. Pellerin, L., Magistretti, P.J., 1994. Glutamate uptake into astrocytes stimulates
McEwen, B.S., Gianaros, P.J., 2011. Stress- and allostasis-induced brain plasticity. aerobic glycolysis: a mechanism coupling neuronal activity to glucose
Annu. Rev. Med. 62, 431–445. utilization. Proc. Natl. Acad. Sci. U. S. A. 91, 10625–10629.
McEwen, B.S., Morrison, J.H., 2013. The brain on stress: vulnerability and plasticity Peters, A., Langemann, D., 2009. Build-ups in the supply chain of the brain: on the
of the prefrontal cortex over the life course. Neuron 79, 16–29. neuroenergetic cause of obesity and type 2 diabetes mellitus. Front.
McEwen, B.S., Stellar, E., 1993. Stress and the individual: mechanisms leading to Neuroenergetics 1, 2. doi:http://dx.doi.org/10.3389/neuro.3314.3002.2009.
disease. Arch. Intern. Med. 153, 2093–2101. Peters, A., McEwen, B.S., 2015. Stress habituation: body shape and cardiovascular
McEwen, B.S., Wingfield, J.C., 2003. The concept of allostasis in biology and mortality. Neurosci. Biobehav. Rev. 56, 139–150.
biomedicine. Horm. Behav. 43, 2–15. Peters, A., Schweiger, U., Pellerin, L., Hubold, C., Oltmanns, K.M., Conrad, M.,
McEwen, B.S., Weiss, J.M., Schwartz, L.S., 1968. Selective retention of corticosterone Schultes, B., Born, J., Fehm, H.L., 2004. The selfish brain: competition for energy
by limbic structures in rat brain. Nature 220, 911–912. resources. Neurosci. Biobehav. Rev. 28, 143–180.
McEwen, B.S., Bowles, N.P., Gray, J.D., Hill, M.N., Hunter, R.G., Karatsoreos, I.N., Nasca, Peters, A., Conrad, M., Hubold, C., Schweiger, U., Fischer, B., Fehm, H.L., 2007a. The
C., 2015a. Mechanisms of stress in the brain. Nat. Neurosci. 18, 1353–1363. principle of homeostasis in the hypothalamus-pituitary-adrenal system: new
McEwen, B.S., Gray, J.D., Nasca, C., 2015b. 60 YEARS OF NEUROENDOCRINOLOGY: insight from positive feedback. Am. J. Physiol. Regul. Integr. Comp. Physiol. 293,
Redefining neuroendocrinology: stress: sex and cognitive and emotional 83–98.
regulation. J. Endocrinol. 226, T67–83. Peters, A., Pellerin, L., Dallman, M.F., Oltmanns, K.M., Schweiger, U., Born, J., Fehm, H.
McEwen, B.S., Nasca, C., Gray, J.D., 2016. Stress effects on neuronal structure: L., 2007b. Causes of obesity: looking beyond the hypothalamus. Prog. Neurobiol.
hippocampus amygdala, and prefrontal cortex. Neuropsychopharmacology 41, 81, 61–88.
3–23. Peters, A., Bosy-Westphal, A., Kubera, B., Langemann, D., Goele, K., Later, W., Heller,
McEwen, B.S., 1998. Protective and damaging effects of stress mediators. N. Engl. J. M., Hubold, C., Müller, M.J., 2011. Why doesn't the brain lose weight, when obese
Med. 338, 171–179. people diet? Obes. Facts doi:http://dx.doi.org/10.1159/000327676.
McEwen, B.S., 2007. Physiology and neurobiology of stress and adaptation: central Picard, M., Juster, R.P., McEwen, B.S., 2014. Mitochondrial allostatic load puts the
role of the brain. Physiol. Rev. 87, 873–904. ‘gluc' back in glucocorticoids. Nat. Rev. Endocrinol. 10, 303–310.
McEwen, B.S., 2012. Brain on stress: how the social environment gets under the skin. Placais, P.Y., Preat, T., 2013. To favor survival under food shortage: the brain disables
Proc. Natl. Acad. Sci. U. S. A. 109 (Suppl. 2), 17180–17185. costly memory. Science 339, 440–442.
McEwen, B.S., 2015. Preserving neuroplasticity: role of glucocorticoids and Plihal, W., Pietrowsky, R., Born, J., 1999. Dexamethasone blocks sleep induced
neurotrophins via phosphorylation. Proc. Natl. Acad. Sci. U. S. A. 112, 15544– improvement of declarative memory. Psychoneuroendocrinology 24, 313–331.
15545. Pruessner, J.C., Hellhammer, D.H., Kirschbaum, 1999. Low self-esteem, induced
Mehaffey, W.H., Doiron, B., Maler, L., Turner, R.W., 2005. Deterministic failure and the adrenocortical stress response. Personal. Indiv. Diff. 27, 477–489.
multiplicative gain control with active dendrites. J. Neurosci. 25, 9968–9977. Pruessner, J.C., Baldwin, M.W., Dedovic, K., Renwick, R., Mahani, N.K., Lord, C.,
Miller, S.L., Green, L.R., Peebles, D.M., Hanson, M.A., Blanco, C.E., 2002. Effects of Meaney, M., Lupien, S., 2005. Self-esteem locus of control, hippocampal
chronic hypoxia and protein malnutrition on growth in the developing chick. volume, and cortisol regulation in young and old adulthood. Neuroimage 28,
Am. J. Obstet. Gynecol. 186, 261–267. 815–826.
Minichiello, L., Korte, M., Wolfer, D., Kuhn, R., Unsicker, K., Cestari, V., Rossi-Arnaud, Puhl, R.M., Heuer, C.A., 2009. The stigma of obesity: a review and update. Obesity
C., Lipp, H.P., Bonhoeffer, T., Klein, R., 1999. Essential role for TrkB receptors in (Silver Spring) 17, 941–964.
hippocampus-mediated learning. Neuron 24, 401–414. Quirarte, G.L., Roozendaal, B., McGaugh, J.L., 1997. Glucocorticoid enhancement of
Monat, A., Averill, J.R., Lazarus, R.S., 1972. Anticipatory stress and coping reactions memory storage involves noradrenergic activation in the basolateral amygdala.
under various conditions of uncertainty. J. Pers. Soc. Psychol. 24, 237–253. Proc. Natl. Acad. Sci. U. S. A. 94, 14048–14053.
Muhlau, M., Gaser, C., Ilg, R., Conrad, B., Leibl, C., Cebulla, M.H., Backmund, H., Radley, J.J., Rocher, A.B., Miller, M., Janssen, W.G., Liston, C., Hof, P.R., McEwen, B.S.,
Gerlinghoff, M., Lommer, P., Schnebel, A., Wohlschlager, A.M., Zimmer, C., Morrison, J.H., 2006. Repeated stress induces dendritic spine loss in the rat
Nunnemann, S., 2007. Gray matter decrease of the anterior cingulate cortex in medial prefrontal cortex. Cereb. Cortex 16, 313–320.
anorexia nervosa. Am. J. Psychiatry 164, 1850–1857. Ramaswami, M., 2014. Network plasticity in adaptive filtering and behavioral
Nai, Q., Dong, H.W., Hayar, A., Linster, C., Ennis, M., 2009. Noradrenergic regulation of habituation. Neuron 82, 1216–1229.
GABAergic inhibition of main olfactory bulb mitral cells varies as a function of Ramaswamy, S., Markram, H., 2015. Anatomy and physiology of the thick-tufted
concentration and receptor subtype. J. Neurophysiol. 101, 2472–2484. layer 5 pyramidal neuron. Front. Cell. Neurosci. 9, 233.
Nai, Q., Dong, H.W., Linster, C., Ennis, M., 2010. Activation of alpha1 and alpha2 Ramos, B.P., Arnsten, A.F., 2007. Adrenergic pharmacology and cognition: focus on
noradrenergic receptors exert opposing effects on excitability of main olfactory the prefrontal cortex. Pharmacol. Ther. 113, 523–536.
bulb granule cells. Neuroscience 169, 882–892. Rao, R.P., Ballard, D.H., 1999. Predictive coding in the visual cortex: a functional
Nguyen, V.T., Breakspear, M., Cunnington, R., 2014. Reciprocal interactions of the interpretation of some extra-classical receptive-field effects. Nat. Neurosci. 2,
SMA and cingulate cortex sustain premovement activity for voluntary actions. J. 79–87.
Neurosci. 34, 16397–16407. Rasgon, N.L., McEwen, B.S., 2016. Insulin resistance-a missing link no more. Mol.
O'Doherty, J., Critchley, H., Deichmann, R., Dolan, R.J., 2003. Dissociating valence of Psychiatry .
outcome from behavioral control in human orbital and ventral prefrontal Reinmuth, O.M., Scheinberg, P., Bourne, B., 1965. Total cerebral blood flow and
cortices. J. Neurosci. 23, 7931–7939. metabolism. Arch. Neurol. 12, 49–66.
Oltmanns, K.M., Melchert, U.H., Scholand-Engler, H.G., Howitz, M.C., Schultes, B., Reyes, B.A., Carvalho, A.F., Vakharia, K., Van Bockstaele, E.J., 2011. Amygdalar
Schweiger, U., Hohagen, F., Born, J., Peters, A., Pellerin, L., 2008. Differential peptidergic circuits regulating noradrenergic locus coeruleus neurons: linking
energetic response of brain vs. skeletal muscle upon glycemic variations in limbic and arousal centers. Exp. Neurol. 230, 96–105.
healthy humans. Am. J. Physiol. Regul. Integr. Comp. Physiol. 294, R12–R16. Richmond, B.J., Wiener, M.C., 2007. Combining order statistics with Bayes theorem
Ortega, P.A., Braun, D.A., 2015. What is epistemic value in free energy models of for millisecond-by-millisecond decoding of spike trains. In: Doya, K., Ishii, S.,
learning and acting? A bounded rationality perspective. Cogn. Neurosci. 6, 215– Pouget, A., Rao, R.P.N. (Eds.), Bayesian Brain—Probabilistic Approaches to Neural
216. Coding. The MIT Press, Cambridge, Massachusetts.
Orth-Gomer, K., Schneiderman, N., Wang, H.X., Walldin, C., Blom, M., Jernberg, T.,
2009. Stress reduction prolongs life in women with coronary disease: the

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004
G Model
PRONEU 1500 No. of Pages 25

A. Peters et al. / Progress in Neurobiology xxx (2017) xxx–xxx 25

Rimmele, U., Besedovsky, L., Lange, T., Born, J., 2013. Blocking mineralocorticoid Stone, P.H., Saito, S., Takahashi, S., Makita, Y., Nakamura, S., Kawasaki, T., Takahashi,
receptors impairs: blocking glucocorticoid receptors enhances memory A., Katsuki, T., Nakamura, S., Namiki, A., Hirohata, A., Matsumura, T., Yamazaki,
retrieval in humans. Neuropsychopharmacology 38, 884–894. S., Yokoi, H., Tanaka, S., Otsuji, S., Yoshimachi, F., Honye, J., Harwood, D., Reitman,
Rodenbeck, A., Hajak, G., 2001. Neuroendocrine dysregulation in primary insomnia. M., Coskun, A.U., Papafaklis, M.I., Feldman, C.L., 2012. Prediction of progression
Rev. Neurol. (Paris) 157, S57–61. of coronary artery disease and clinical outcomes using vascular profiling of
Roesch, M.R., Olson, C.R., 2004. Neuronal activity related to reward value and endothelial shear stress and arterial plaque characteristics: the PREDICTION
motivation in primate frontal cortex. Science 304, 307–310. Study. Circulation 126, 172–181.
Rushworth, M.F., Walton, M.E., Kennerley, S.W., Bannerman, D.M., 2004. Action sets Storch, M., Gaab, J., Kuttel, Y., Stussi, A.C., Fend, H., 2007. Psychoneuroendocrine
and decisions in the medial frontal cortex. Trends Cogn. Sci. 8, 410–417. effects of resource-activating stress management training. Health Psychol. 26,
Salgado, H., Garcia-Oscos, F., Patel, A., Martinolich, L., Nichols, J.A., Dinh, L., 456–463.
Roychowdhury, S., Tseng, K.Y., Atzori, M., 2011. Layer-specific noradrenergic Swanson, L.W., 2000. Cerebral hemisphere regulation of motivated behavior. Brain
modulation of inhibition in cortical layer II/III. Cereb. Cortex 21, 212–221. Res. 886, 113–164.
Salgado, H., Garcia-Oscos, F., Martinolich, L., Hall, S., Restom, R., Tseng, K.Y., Atzori, Tawakol, A., Ishai, A., Takx, R.A., Figueroa, A.L., Ali, A., Kaiser, Y., Truong, Q.A.,
M., 2012. Pre- and postsynaptic effects of norepinephrine on gamma- Solomon, C.J., Calcagno, C., Mani, V., Tang, C.Y., Mulder, W.J., Murrough, J.W.,
aminobutyric acid-mediated synaptic transmission in layer 2/3 of the rat Hoffmann, U., Nahrendorf, M., Shin, L.M., Fayad, Z.A., Pitman, R.K., 2017. Relation
auditory cortex. Synapse 66, 20–28. between resting amygdalar activity and cardiovascular events: a longitudinal
Salin, P.A., Bullier, J., 1995. Corticocortical connections in the visual system: and cohort study. Lancet .
structure and function. Physiol. Rev. 75, 107–154. Thomson, A.M., Lamy, C., 2007. Functional maps of neocortical local circuitry. Front.
Salinas, E., Sejnowski, T.J., 2001. Correlated neuronal activity and the flow of neural Neurosci. 1, 19–42.
information. Nat. Rev. Neurosci. 2, 539–550. Tong, Q., Ye, C., McCrimmon, R.J., Dhillon, H., Choi, B., Kramer, M.D., Yu, J., Yang, Z.,
Sanchez, M.M., Young, L.J., Plotsky, P.M., Insel, T.R., 2000. Distribution of Christiansen, L.M., Lee, C.E., Choi, C.S., Zigman, J.M., Shulman, G.I., Sherwin, R.S.,
corticosteroid receptors in the rhesus brain: relative absence of glucocorticoid Elmquist, J.K., Lowell, B.B., 2007. Synaptic glutamate release by ventromedial
receptors in the hippocampal formation. J. Neurosci. 20, 4657–4668. hypothalamic neurons is part of the neurocircuitry that prevents hypoglycemia.
Sarinopoulos, I., Grupe, D.W., Mackiewicz, K.L., Herrington, J.D., Lor, M., Steege, E.E., Cell Metab. 5, 383–393.
Nitschke, J.B., 2010. Uncertainty during anticipation modulates neural Toyabe, S., Sagawa, T., Ueda, M., Muneyuki, E., Sano, M., 2014. Experimental
responses to aversion in human insula and amygdala. Cereb. Cortex 20, 929– demonstration of information-to-energy conversation and validation of the
940. generalized Jarzynski equalita. Nat. Phys. 6, 988–992.
Schaaf, M.J., Hoetelmans, R.W., de Kloet, E.R., Vreugdenhil, E., 1997. Corticosterone Tribus, M., McIrvine, E.C., 1972. Energy and information. Sci. Am. 225, 179–186.
regulates expression of BDNF and trkB but not NT-3 and trkC mRNA in the rat Tribus, M., 1961. Thermodynamics and Thermostatics: An Introduction to Energy,
hippocampus. J. Neurosci. Res. 48, 334–341. Information and States of Matter, with Engineering Applications. Van Nostrand
Schulkin, J., McEwen, B.S., Gold, P.W., 1994. Allostasis amygdala, and anticipatory Company Inc, New York.
angst. Neurosci. Biobehav. Rev. 18, 385–396. Valentino, R.J., Van Bockstaele, E.J., 2008. Convergent regulation of locus coeruleus
Seldenrijk, A., Hamer, M., Lahiri, A., Penninx, B.W., Steptoe, A., 2012. Psychological activity as an adaptive response to stress. Eur. J. Pharmacol. 583, 194–203.
distress: cortisol stress response and subclinical coronary calcification. Wacongne, C., Changeux, J.P., Dehaene, S., 2012. A neuronal model of predictive
Psychoneuroendocrinology 37, 48–55. coding accounting for the mismatch negativity. J. Neurosci. 32, 3665–3678.
Selye, H., 1956. What is stress? Metabolism 5, 525–530. Wagner, U., Born, J., 2008. Memory consolidation during sleep: interactive effects of
Sengupta, B., Stemmler, M.B., Friston, K.J., 2013. Information and efficiency in the sleep stages and HPA regulation. Stress 11, 28–41.
nervous system-a synthesis. PLoS Comput. Biol. 9, e1003157. Wagner, U., Degirmenci, M., Drosopoulos, S., Perras, B., Born, J., 2005. Effects of
Sengupta, B., Tozzi, A., Cooray, G.K., Douglas, P.K., Friston, K.J., 2016. Towards a cortisol suppression on sleep-associated consolidation of neutral and emotional
neuronal gauge theory. PLoS Biol. 14, e1002400. memory. Biol. Psychiatry 58, 885–893.
Shadlen, M.N., Newsome, W.T., 1994. Noise: neural codes and cortical organization. Watanabe, Y., Gould, E., McEwen, B.S., 1992. Stress induces atrophy of apical
Curr. Opin. Neurobiol. 4, 569–579. dendrites of hippocampal CA3 pyramidal neurons. Brain Res. 588, 341–345.
Shannon, C.E., 1948. A mathematical theory of communication. Bell Syst. Tech. J. 27 Wellman, C.L., 2001. Dendritic reorganization in pyramidal neurons in medial
379/623-423/656. prefrontal cortex after chronic corticosterone administration. J. Neurobiol. 49,
Shi, B., Mocchetti, I., 2000. Dexamethasone induces TrkA and p75NTR 245–253.
immunoreactivity in the cerebral cortex and hippocampus. Exp. Neurol. 162, Yang, G., Pan, F., Gan, W.B., 2009. Stably maintained dendritic spines are associated
257–267. with lifelong memories. Nature 462, 920–924.
Shipp, S., Adams, R.A., Friston, K.J., 2013. Reflections on agranular architecture: Zhang, X., Poo, M.M., 2002. Localized synaptic potentiation by BDNF requires local
predictive coding in the motor cortex. Trends Neurosci. 36, 706–716. protein synthesis in the developing axon. Neuron 36, 675–688.
Shipp, S., 2016. Neural elements for predictive coding. Front. Psychol. 7, 1792. Zhou, M., Bakker, E.H., Velzing, E.H., Berger, S., Oitzl, M., Joels, M., Krugers, H.J., 2010.
Silver, R.A., 2010. Neuronal arithmetic. Nat. Rev. Neurosci. 11, 474–489. Both mineralocorticoid and glucocorticoid receptors regulate emotional
Sterling, P., 2012. Allostasis: a model of predictive regulation. In: McEwen, B., Peters, memory in mice. Neurobiol. Learn. Mem. 94, 530–537.
A. (Eds.), Allostasis and Allostatic Load/Special Issue. Physiology and Behavior,
pp. 5–15.

Please cite this article in press as: A. Peters, et al., Uncertainty and stress: Why it causes diseases and how it is mastered by the brain, Prog.
Neurobiol. (2017), http://dx.doi.org/10.1016/j.pneurobio.2017.05.004