PLOS COMPUTATIONAL BIOLOGY

RESEARCH ARTICLE

Systems biology informed deep learning for

inferring parameters and hidden dynamics
Alireza Yazdani ID1☯, Lu Lu ID2☯, Maziar Raissi ID3, George Em Karniadakis ID1*
1 Division of Applied Mathematics, Brown University, Providence, Rhode Island, USA, 2 Department of
Mathematics, Massachusetts Institute of Technology, Cambridge, Massachusetts, USA, 3 Department of
Applied Mathematics, University of Colorado, Boulder, Colorado, USA

☯ These authors contributed equally to this work.

* george_karniadakis@brown.edu
a1111111111
a1111111111
a1111111111 Abstract
a1111111111
a1111111111 Mathematical models of biological reactions at the system-level lead to a set of ordinary dif-
ferential equations with many unknown parameters that need to be inferred using relatively
few experimental measurements. Having a reliable and robust algorithm for parameter infer-
ence and prediction of the hidden dynamics has been one of the core subjects in systems
OPEN ACCESS biology, and is the focus of this study. We have developed a new systems-biology-informed
Citation: Yazdani A, Lu L, Raissi M, Karniadakis GE deep learning algorithm that incorporates the system of ordinary differential equations into
(2020) Systems biology informed deep learning for the neural networks. Enforcing these equations effectively adds constraints to the optimiza-
inferring parameters and hidden dynamics. PLoS tion procedure that manifests itself as an imposed structure on the observational data.
Comput Biol 16(11): e1007575. https://doi.org/
Using few scattered and noisy measurements, we are able to infer the dynamics of unob-
10.1371/journal.pcbi.1007575
served species, external forcing, and the unknown model parameters. We have success-
Editor: Vassily Hatzimanikatis, Ecole Polytechnique
fully tested the algorithm for three different benchmark problems.
Fédérale de Lausanne, SWITZERLAND

Received: November 28, 2019

Accepted: October 11, 2020

Published: November 18, 2020 Author summary

Copyright: © 2020 Yazdani et al. This is an open The dynamics of systems biological processes are usually modeled using ordinary differ-
access article distributed under the terms of the ential equations (ODEs), which introduce various unknown parameters that need to be
Creative Commons Attribution License, which
estimated efficiently from noisy measurements of concentration for a few species only. In
permits unrestricted use, distribution, and
reproduction in any medium, provided the original this work, we present a new “systems-informed neural network” to infer the dynamics of
author and source are credited. experimentally unobserved species as well as the unknown parameters in the system of
equations. By incorporating the system of ODEs into the neural networks, we effectively
Data Availability Statement: All relevant data are
within the manuscript, its Supporting information add constraints to the optimization algorithm, which makes the method robust to noisy
files and available from https://github.com/ and sparse measurements.
alirezayazdani1/SBINNs.

Funding: This work was supported by the National

Institutes of Health U01HL142518 (GEK). The
Introduction
funders had no role in study design, data collection
and analysis, decision to publish, or preparation of Systems biology aims at a system-level understanding of biological systems, which is a holistic
the manuscript. approach to deciphering the complexity of biological systems [1]. To understand the biological
Competing interests: The authors have declared systems, we must understand the structures of the systems (both their components and the
that no competing interests exist. structural relationships), and their dynamics [2]. The dynamics of a biological system is usually

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1007575 November 18, 2020 1 / 19

PLOS COMPUTATIONAL BIOLOGY Systems biology informed deep learning

modeled by a system of ordinary differential equations (ODEs), which describes the time evo-
lution of the concentrations of chemical and molecular species in the system. After we know
the pathway structure of the chemical reactions, then the ODEs can be derived using some
kinetic laws, e.g., the law of mass action or the Michaelis-Menten kinetics [3].
System-level biological models usually introduce some unknown parameters that are
required to be estimated accurately and efficiently. Thus, one central challenge in computa-
tional modeling of these systems is the estimation of model parameters (e.g., rate constants or
initial concentrations) and the prediction of model dynamics (e.g., time evolution of experi-
mentally unobserved concentrations). Hence, a lot of attention has been given to the problem
of parameter estimation in the systems biology community. In particular, extensive research
has been conducted on the applications of different optimization techniques, such as linear
and nonlinear least-squares fitting [4], genetic algorithms [5], evolutionary computation [6],
and more [7]. Considerable interest has also been raised by Bayesian methods [8, 9], which
could extract information from noisy data. The main advantage of Bayesian methods is the
ability to infer the whole probability distributions of the unknown parameters, rather than just
a point estimate. More recently, parameter estimation for computational biology models has
been tackled by the algorithms in the framework of control theory. These algorithms were
originally developed for the problem of estimating the time evolution of the unobserved com-
ponents of the state of a dynamical system. In this context, extended Kalman filtering [10],
unscented Kalman filtering [11], and ensemble Kalman methods [12] have been applied as
well. In addition, different methods have also been developed to address the issue of hidden
variables and dynamics [13, 14], but in their examples the number of observable variables is
almost one half of the number of total variables, while as we will show in the results below our
proposed method requires less observable variables (e.g., one out of eight in the cell apoptosis
model) to correctly infer unknown parameters and predict all unobserved variables.
Due to technical limitations, however, biological reaction networks are often only partially
observable. Usually, experimental data are insufficient considering the size of the model,
which results in parameters that are non-identifiable [15] or only identifiable within confi-
dence intervals (see more details in [16]). Furthermore, a large class of models in systems biol-
ogy are sensitivity to the parameter values that are distributed over many orders of magnitude.
Such sloppiness is also a factor that makes parameter estimation more difficult [17]. In the pro-
cess of parameter inference, two issues accounting for system’s (non-)identifiability have to be
considered: structural identifiability that is related to the model structure independent of the
experimental data [18, 19]; and practical identifiability that is related to the amount and quality
of measured data. The a priori structural identifiability can be used to address the question of
unique estimation of the unknown parameters based on the postulated model. However, a
parameter that is structurally identifiable may still be practically non-identifiable assuming
that the model is exact, but the measurements are noisy or sparse [20].
In this work, we introduce a new deep learning [21] method—systems-informed neural
networks, based on the method of physics-informed neural networks [22, 23], to infer the hid-
den dynamics of experimentally unobserved species as well as the unknown parameters in the
system of equations. By incorporating the system of ODEs into the neural networks (through
adding the residuals of the equations to the loss function), we effectively add constraints to the
optimization algorithm, which makes the method robust to measurement noise and few scat-
tered observations. In addition, since large system-level biological models are typically encoun-
tered, our algorithm is computationally scalable and feasible, and its output is interpretable
even though it depends on a high-dimensional parameter space.

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1007575 November 18, 2020 2 / 19

PLOS COMPUTATIONAL BIOLOGY Systems biology informed deep learning

Materials and methods

Throughout this paper, we assume that the systems biological process can be modeled by a sys-
tem of ordinary differential equations (ODEs) of the following form
dx
¼ f ðx; t; pÞ; ð1aÞ
dt

xðT0 Þ ¼ x0 ; ð1bÞ

y ¼ hðxÞ þ �ðtÞ; �ðtÞ � N ð0; s2 Þ; ð1cÞ

where the state vector x = (x1, x2, . . ., xS) represents the concentration of S species, and p = (p1,
p2, . . ., pK) are K parameters of the model, which remain to be determined. Hence, the system
of ODEs will be identified once p is known. y = (y1, y2, . . ., yM) are the M measurable signals
(consistent with the ODE system), which we can measure experimentally and could possibly
be contaminated with a white noise � of Gaussian type with zero mean and standard deviation
σ. The output function h is determined by the design of the experiments that are used for
parameter inference. While h could, in general, be any function, it is assumed to be a linear
function with M � S in most models as follows:
0 1 0 1 0 1
y1 xs1 �s1
B C B C B C
B C B C B C
B y2 C B xs2 C B �s2 C
B C B C B C
B C¼B C þ B C; ð2Þ
B...C B ... C B...C
B C B C B C
@ A @ A @ A
yM xsM �sM

i.e., y1, y2, . . ., yM are the noisy measurements of the species xs1 ; xs2 ; . . . ; xsM among all S species
(1 � s1 < s2 < � � � < sM � S).

Systems-informed neural networks and parameter inference

Based on the method of physics-informed neural networks proposed in [22], we introduce a
deep learning framework that is informed by the systems biology equations that describe the
kinetic pathways (Eq (1a)). A neural network with parameters θ takes time t as the input and
outputs a vector of the state variables x^ðt; θÞ ¼ ð^
x 1 ðt; θÞ; x^2 ðt; θÞ; . . . ; x^S ðt; θÞÞ as a surrogate
of the ODE solution x(t) (Fig 1). In our network, in addition to the standard layers, e.g., the
fully-connected layer, we add three extra layers to make the network training easier, described
as follows.
• Input-scaling layer. Because the ODE system could have a large time domain, i.e., the input
time t could vary by orders of magnitude, we first apply a linear scaling function to t, i.e.,
~t ¼ t=T, such that ~t � Oð1Þ. We can simply choose T as the maximum value of the time
domain.
• Feature layer. In many models, the ODE solution may have a certain pattern, e.g., periodicity
in the yeast glycolysis model and fast decay in the cell apoptosis model, and thus it is benefi-
cial to the network training by constructing a feature layer according to these patterns.
Specifically, we employ the L functions e1(�), e2(�), . . ., eL(�) to construct the L features
e1 ð~t Þ; e2 ð~t Þ; . . . ; eL ð~t Þ. The choice of the features is problem dependent, and we can select the
features based on the available observations. For example, if the observed dynamics has

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1007575 November 18, 2020 3 / 19

PLOS COMPUTATIONAL BIOLOGY Systems biology informed deep learning

Fig 1. Neural network architecture. The network consists of an input-scaling layer, a feature layer, several fully-
connected layers, and an output-scaling layer. The input-scaling layer and output-scaling layer are used to linearly
scale the network input and outputs such that they are of order one. The feature layer is used to construct features
explicitly as the input to the first fully-connected layer.
https://doi.org/10.1371/journal.pcbi.1007575.g001

certain periodicity (e.g., the yeast glycolysis model), then we can use sin(kt) as the features,
where k is selected based on the period of the observed dynamics; if the observed dynamics
decays fast (e.g., the cell apoptosis model), then we can use e−kt as a feature. However, if there
is no clear pattern observed, the feature layer can also be removed.
• Output-scaling layer. Because the outputs x^1 ; x^2 ; . . . ; x^S may have different magnitudes, sim-
ilar to the input-scaling layer, we add another scaling layer to transform the output of the last
fully-connected layer x~1 ; x~2 ; . . . ; x~S (of order one) to x^1 ; x^2 ; . . . ; x^S , i.e., x^1 ¼ k1 x~1 , x^2 ¼ k2 x~2 ,
. . . ; x^S ¼ kS x~S . Here, k1, k2, . . ., kS are chosen as the magnitudes of the mean values of the
ODE solution x1, x2, . . ., xS, respectively.

The next key step is to constrain the neural network to satisfy the scattered observations of
y as well as the ODE system (Eq (1a)). This is realized by constructing the loss function by con-
sidering terms corresponding to the observations and the ODE system. Specifically, let us
assume that we have the measurements of y1, y2, . . ., yM at the time t1 ; t2 ; . . . ; tN data , and we
enforce the network to satisfy the ODE system at the time point t1 ; t2 ; . . . ; tN ode . We note that
the times t1 ; t2 ; . . . ; tN data and t1 ; t2 ; . . . ; tN ode are not necessarily on a uniform grid, and they
could be chosen at random. Then, the total loss is defined as a function of both θ and p:

Lðθ; pÞ ¼ Ldata ðθÞ þ Lode ðθ; pÞ þ Laux ðθÞ; ð3Þ

where
" data
#
data
X
M
data
X
M
1 X N
2
L ðθÞ ¼ m Lm
wdata ¼ wdata
m ðym ðtn Þ x^sm ðtn ; θÞÞ ; ð4Þ
m¼1 m¼1
N data
n¼1

" �2 #
X X N ode �
S S
1 X d^
xs
Lode ðθ; pÞ ¼ s Ls
wode ode
¼ wode j fs ðx^s ðtn ; θÞ; tn ; pÞ ; ð5Þ
s¼1
s
s¼1
N ode n¼1 dt tn

X
S X
S
ðxs ðT0 Þ
2
x^s ðT0 ; θÞÞ þ ðxs ðT1 Þ x^s ðT1 ; θÞÞ
2
Laux ðθÞ ¼ s Ls
waux aux
¼ waux
s : ð6Þ
s¼1 s¼1
2

Ldata is associated with the M sets of observations y given by Eq (1c), while Lode enforces the

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1007575 November 18, 2020 4 / 19

PLOS COMPUTATIONAL BIOLOGY Systems biology informed deep learning

structure imposed by the system of ODEs given in Eq (1a). We employ automatic differentia-
tion (AD) to analytically compute the derivative d^dtx s jtn in Lode (see more details of AD in [23]).
The third auxiliary loss term Laux is introduced as an additional source of information for the
system identification, and involves two time instants T0 and T1. It is essentially a component of
the data loss; however, we prefer to separate this loss from the data loss, as in the auxiliary loss
data are given for all state variables at these two time instants. We note that Ldata and Laux are
the discrepancy between the network and measurements, and thus they are supervised losses,
while Lode is based on the ODE system, and thus is unsupervised. In the last step, we infer the
neural network parameters θ as well as the unknown parameters of the ODEs p simultaneously
by minimizing the loss function via gradient-based optimizers, such as the Adam optimizer
[24]:
θ� ; p� ¼ arg min Lðθ; pÞ: ð7Þ
θ;p

We note that our proposed method is different from meta-modeling [25, 26], as we optimize θ
and p simultaneously.
The M + 2S coefficients ðwdata data data ode ode ode
1 ; w2 ; . . . ; wM Þ in Eq (4), ðw1 ; w2 ; . . . ; wS Þ in Eq (5), and
aux aux aux
ðw1 ; w2 ; . . . ; wS Þ in Eq (6) are used to balance the M + 2S loss terms. In this study, we man-
ually select these weight coefficients such that the weighted losses are of the same order of mag-
nitude during the network training. We note that this guideline makes the weight selection
much easier, although there are many weights to be determined. These weights may also be
automatically chosen, e.g., by the method proposed in [27]. In this study, the time instants
t1 ; t2 ; . . . ; tN data for the observations are chosen randomly in the time domain, while the time
instants t1 ; t2 ; . . . ; tN ode used to enforce the ODEs are chosen in an equispaced grid. Addition-
ally, in the auxiliary loss function, the first set of data is the initial conditions at time T0 for the
state variables. The second set includes the values of the state variables at any arbitrary time
instant T1 within the training time window (not too close to T0); in this work, we consider the
midpoint time for the cell apoptosis model, and the final time instant for the yeast glycolysis
model and ultradian endocrine model. If data is available at another time point, alternatively
this point can be considered.

Analysis of system’s identifiability

A parameter pi is identifiable if the confidence interval of its estimate p^i is finite. In systems
identification problems, two different forms of identifiability namely, structural and practical
are typically encountered. Structural non-identifiability arises from a redundant parameteriza-
tion in the formal solution of y due to insufficient mapping h of internal states x to observables
y in Eq (1) [15]. A priori structural identifiability has been studied extensively, e.g., using
power series expansion [28] and differential algebraic methods [29], yet mostly limited to lin-
ear models as the problem is particularly difficult for nonlinear dynamical systems. Further-
more, practical non-identifiability cannot be detected with these methods, since experimental
data are disregarded.
A parameter that is structurally identifiable may still be practically non-identifiable. Practi-
cal non-identifiability is intimately related to the amount and quality of measured data and
manifests in a confidence interval that is infinite. Different methods have been proposed to
estimate the confidence intervals of the parameters such as local approximation of the Fisher-
Information-Matrix (FIM) [20] and bootstrapping approach [30]. Another approach is to
quantify the sensitivity of the systems dynamics to variations in its parameters using a

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1007575 November 18, 2020 5 / 19

PLOS COMPUTATIONAL BIOLOGY Systems biology informed deep learning

Table 1. Hyperparameters for the problems in this study. The fist and second number in the number of iterations correspond to the first and second training stage.
Model NN depth NN width #Iterations
Yeast glycolysis Noiseless 4 128 1000, 9 × 104
Noisy 4 128 1000, 2 × 106
Cell apoptosis Survival 5 256 0, 1.5 × 106
Death 5 256 0, 1.5 × 106
Ultradian endocrine Parameters only 4 128 2000, 6 × 105
Hidden nutrition 4 128 2000, 1.5 × 106
https://doi.org/10.1371/journal.pcbi.1007575.t001

probabilistic framework [31]. For identifiability analysis, we primarily use the FIM method,
which is detailed in S1 Text.

Implementation
The algorithm is implemented in Python using the open-source library DeepXDE [23]. The
width and depth of the neural networks (listed in Table 1) depend on the size of the system of
equations and the complexity of the dynamics. We use the SwishðxÞ ¼ 1þex x function [32] as
the activation function σ shown in Fig 1, and the feature layer is listed in Table 2.
For the training, we use an Adam optimizer [24] with default hyperparameters and a learn-
ing rate of 10−3, where the training is performed using the full batch of data. As the total loss
consists of two supervised losses and one unsupervised loss, we perform the training using the
following two-stage strategy:
Step 1. Considering that supervised training is usually easier than unsupervised training, we
first train the network using the two supervised losses Ldata and Laux for some itera-
tions, such that the network can quickly match the observed data points.
Step 2. We further train the network using all the three losses.
We found empirically that this two-stage training strategy speeds up the network conver-
gence. The number of iterations for each stage is listed in Table 1.

Results
Yeast glycolysis model
The model of oscillations in yeast glycolysis [33] has become a standard benchmark problem
for systems biology inference [34, 35] as it represents complex nonlinear dynamics typical of
biological systems. We use it here to study the performance of our deep learning algorithm
used for parsimonious parameter inference with only two observables. The system of ODEs
for this model as well as the target parameter values and the initial conditions are included in
S2 Text. To represent experimental noise, we corrupt the observation data by a Gaussian noise

Table 2. The feature layer used in the network for each problem.
Model Features
Yeast glycolysis ~t , sinð~t Þ, sinð2~t Þ, sinð3~t Þ, sinð4~t Þ, sinð5~t Þ, sinð6~t Þ
Cell apoptosis ~t , e ~t
Ultradian endocrine ~t , sinð~t Þ, sinð2~t Þ, sinð3~t Þ, sinð4~t Þ, sinð5~t Þ

https://doi.org/10.1371/journal.pcbi.1007575.t002

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1007575 November 18, 2020 6 / 19

PLOS COMPUTATIONAL BIOLOGY Systems biology informed deep learning

with zero mean and the standard deviation of σ� = cμ, where μ is the standard deviation of each
observable over the observation time window and c = 0 − 0.1.
We start by inferring the dynamics using noiseless observations on two species S5 (the con-
centration of NADH) and S6 (the concentration of ATP) only. These two species are the mini-
mum number of observables we can use to effectively infer all the parameters in the model. S1
Fig shows the noiseless synthetically generated data by solving the system of ODEs in S2 Text
with the parameters listed in S1 Table. We sample data points within the time frame of 0 − 10
minutes at random and use them for training of the neural networks, where the neural net-
work is informed by the governing ODEs of the yeast model as explained above. S2 Fig shows
the inferred dynamics for all the species predicted by the systems-informed neural networks,
and plotted against the exact dynamics that are generated by solving the system of ODEs. We
observe excellent agreement between the inferred and exact dynamics within the training time
window. The neural networks learn the input data given by scattered observations (shown by
symbols in S2 Fig) and is able to infer the dynamics of other species due to the constraints
imposed by the system of ODEs.
Next, we verify the robustness of the algorithm to noise. For that purpose, we introduce
Gaussian additive noise with the noise level c = 10% to the observational data. The input train-
ing data are shown in Fig 2 for the same species (S5 and S6) as the observables, where similar to
the previous test, we sample random scattered data points in time. Results for the inferred
dynamics are shown in Fig 3. The agreement between the inferred and exact dynamics is excel-
lent considering the relatively high level of noise in the training data. Our results show that the
enforced equations in the loss function Lode act as a constraint of the neural networks that can
effectively prevent the overfitting of the network to the noisy data. One advantage of encoding
the equations is their regularization effect without using any additional L1 or L2 regularization.
Our main objective in this work, however, is to infer the unknown model parameters p.
This can be achieved simply by training the neural networks for its parameters θ as well as the
model parameters using backpropagation. The results for the inferred model parameters along
with their target values are given in Table 3 for both test cases (i.e., with and without noise in
the observations). First thing to note is that the parameters can be identified within a confi-
dence interval. Estimation of the confidence intervals a priori is the subject of structural iden-
tifiability analysis, which is not in the scope of this work. Second, practical identifiability

Fig 2. Glycolysis oscillator noisy observation data given to the algorithm for parameter inference. 500
measurements are corrupted by a zero-mean Gaussian noise and standard deviation of σ = 0.1μ. Only two observables
S5 and S6 are considered and the data are randomly sampled in the time window of 0 − 10 minutes.
https://doi.org/10.1371/journal.pcbi.1007575.g002

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1007575 November 18, 2020 7 / 19

PLOS COMPUTATIONAL BIOLOGY Systems biology informed deep learning

Fig 3. Glycolysis oscillator inferred dynamics from noisy measurements compared with the exact solution. 500
scattered observations are plotted using symbols for the two observables S5 and S6.
https://doi.org/10.1371/journal.pcbi.1007575.g003

analysis can be performed to identify the practically non-identifiable parameters based on the
quality of the measurements and the level of the noise. We have performed local sensitivity
analysis by constructing the Fisher Information Matrix (FIM) (S1 Text) and the correlation
matrix R derived from the FIM.
The inferred parameters from both noiseless and noisy observations are in good agreement
with their target values. The most significant difference (close to 30% difference) can be seen

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1007575 November 18, 2020 8 / 19

PLOS COMPUTATIONAL BIOLOGY Systems biology informed deep learning

Table 3. Parameter values for yeast glycolysis model and each corresponding inferred values. The standard deviations are estimated using Eq. (S3) in S1 Text as practi-
cal non-identifiability analysis based on the FIM.
Parameter Target value Inferred value (Noiseless observations) Inferred value (Noisy observations) Standard deviation
J0 2.5 2.50 2.49 0.18
k1 100 99.9 86.1 62.0
k2 6 6.01 4.55 21.3
k3 16 15.9 14.0 21.9
k4 100 100.1 97.1 103.6
k5 1.28 1.28 1.24 0.25
k6 12 12.0 12.7 5.1
k 1.8 1.79 1.55 4.34
κ 13 13.0 13.4 25.9
q 4 4.00 4.07 0.27
K1 0.52 0.520 0.550 0.091
ψ 0.1 0.0994 0.0823 0.317
N 1 0.999 1.29 2.94
A 4 4.01 4.25 2.28
https://doi.org/10.1371/journal.pcbi.1007575.t003

for the parameter N (the total concentration of NAD+ and NADH). However, given that the
glycolysis system (S2 Text) is identifiable (c.f. [33, 35] and S3 Fig), and the inferred dynamics
shown in S2 Fig and Fig 3 show that the learned dynamics match very well with the exact
dynamics, the inferred parameters are valid. We used Eq. (S3) in S1 Text to estimate the stan-
dard deviations of the model parameters. The σi estimates for the parameters are the lower
bounds, and thus, may not be informative here. Further, these estimates are derived based on a
local sensitivity analysis. A structural/practical identifiability analysis [15] or a bootstrapping
approach to obtain the parameter confidence intervals is probably more relevant here. Using
the FIM, we are able to construct the correlation matrix R for the parameters. Nearly perfect
correlations (|Rij| � 1) suggest that the FIM is singular and the correlated parameters may not
be practically identifiable. For the glycolysis model, as shown in S3 Fig, no perfect correlations
can be found in R (except for the anti-diagonal elements), which suggests that the model
described by S2 Text is practically identifiable. In the example above, we considered 500 data
measurements, but in systems biology we often lack the ability to observe system dynamics at a
fine-time scale. To investigate the performance of our method to a set of sparse data points, we
used only 200 data points, and still have a good inferred dynamics of the species S1, S2, S5 and
S6 (S4 Fig).

Cell apoptosis model

Although the glycolysis model is highly nonlinear and difficult to learn, we have shown that its
parameters can be identified. To investigate the performance of our algorithm for non-identi-
fiable systems, we study a cell apoptosis model, which is a core sub-network of the signal trans-
duction cascade regulating the programmed cell death-against-survival phenotypic decision
[36]. The equations defining the cell apoptosis model and the values of the rate constants for
the model are taken from [36] and listed in S2 Table.
Although the model is derived using simple mass-action kinetics and its dynamics is easy to
learn with our algorithm, most of the parameters are not identifiable due to both structural
and practical non-identifiability. To infer the dynamics of this model, we only use 120 random
samples of measurements collected for one observable (x4), where we assume that the mea-
surements are corrupted by a zero-mean Gaussian noise and 5% standard deviation as shown

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1007575 November 18, 2020 9 / 19

PLOS COMPUTATIONAL BIOLOGY Systems biology informed deep learning

Fig 4. Cell apoptosis noisy observation data given to the algorithm for parameter inference. 120 measurements are
corrupted by a zero-mean Gaussian noise and standard deviation of σ = 0.05μ. Data for the observable x4 only are
randomly sampled during the time window of 0 − 60 hours for two scenarios: (top) cell survival with the initial
condition x7(0) = 2.9 × 104 (molecules/cell) and (bottom) cell death with x7(0) = 2.9 × 103 (molecules/cell).
https://doi.org/10.1371/journal.pcbi.1007575.g004

in Fig 4. Furthermore, it is possible to use different initial conditions in order to produce dif-
ferent cell survival outcomes. The initial conditions for all the species are given in S3 Text,
while we use x7(0) = 2.9 × 104 (molecules/cell) to model cell survival (Fig 4(top)) and x7(0) =
2.9 × 103 (molecules/cell) to model cell death (Fig 4(bottom)).
Using the systems-informed neural networks and the noisy input data, we are able to infer
most of the dynamics (including x3, x4, x6, x7 and x8) of the system as shown in S5 Fig and Fig
5. These results show a good agreement between the inferred and exact dynamics of the cell
survival/apoptosis models using one observable only.
We report the inferred parameters for the cell apoptosis model in Table 4, where we have
used noisy observations on x4 under two scenarios of cell death and survival for comparison.
The results show that four parameters (k1, kd2, kd4 and kd6) can be identified by our proposed
method with relatively high accuracy, as indicated by the check mark (✓) in the last column of
Table 4. We observe that the standard deviations for most of the parameter estimates are
orders of magnitude larger than their target values, and thus the standard deviations estimated
using the FIM are not informative in the practical identifiability analysis. The only informative
standard deviation is for kd6 (indicated by the symbol †), and kd6 is inferred with relatively
high accuracy by our method.
To have a better picture of the practical identifiability analysis, we have plotted the correla-
tion matrix R in S6 Fig. We observe perfect correlations |Rij| � 1 between some parameters.
Specifically, parameters k1 − kd1, and k3 − kd3 have correlations above 0.99 for cell survival
model, which suggests that these parameters may not be identified. This is generally in agree-
ment with the parameter inference results in Table 4 with some exceptions. Our parameter
inference algorithm suggests that k1 is identifiable, whereas kd1 is not for the cell survival
model. Thus, in order to increase the power of the practical identifiability analysis and to com-
plement the correlation matrix, we have computed the FIM null eigenvectors and for each
eigenvector we identified the most dominant coefficients, which are plotted in Fig 6. We

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1007575 November 18, 2020 10 / 19

PLOS COMPUTATIONAL BIOLOGY Systems biology informed deep learning

Fig 5. Cell apoptosis inferred dynamics from noisy observations compared with the exact solution. Predictions are
performed on equally-spaced time instants in the interval of 0 − 60 hours. The scattered observations are plotted using
symbols only for the observable x4. The exact data and the scattered observations are computed by solving the system
of ODEs given in S3 Text.
https://doi.org/10.1371/journal.pcbi.1007575.g005

observe that there are six null eigenvectors associated with the zero eigenvalues of the FIM for
both the cell survival and cell death models. The most dominant coefficient in each null eigen-
vector is associated with a parameter that can be considered as practically non-identifiable.
The identifiable parameters include k1, kd4 and kd6 (indicated by the symbol ? in Table 4),

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1007575 November 18, 2020 11 / 19

PLOS COMPUTATIONAL BIOLOGY Systems biology informed deep learning

Table 4. Parameter values for cell apoptosis model and their corresponding inferred values. The standard deviations are estimated using Eq. (S3) in S1 Text as practical
identifiability analysis using the Fisher Information Matrix. The symbols ✓, † and ? in the last column denote that the corresponding variable is identifiable using our pro-
posed method, FIM standard deviation, and null-eigenvector analysis.
Parameter Target Value Cell Survival Cell Death Identifiable
Inferred Value Standard Deviation Inferred Value Standard Deviation
k1 2.67 × 10−9 0.59 × 10−9 6.2 × 10−6 0.34 × 10−9 1.9 × 10−5 ✓?
−2 −10 −7
kd1 1 × 10 7.06 × 10 35.5 3.37 × 10 96.7
kd2 8 × 10−3 1.72 × 10−3 4.9 2.38 × 10−3 23.8 ✓
k3 6.8 × 10−8 0.15 × 10−8 1.0 × 10−4 0.16 × 10−8 2.1 × 10−4
kd3 5 × 10−2 4.19 × 10−10 62.8 4.23 × 10−10 78.0
−3 −3
kd4 1 × 10 0.92 × 10 0.20 1.28 × 10−3 1.2 ✓?
k5 7 × 10−5 1.49 × 10−7 0.019 7.25 × 10−8 0.37
kd5 1.67 × 10−5 6.92 × 10−11 0.034 8.51 × 10−11 13.7
−4 −4 −4
kd6 1.67 × 10 1.81 × 10 0.35 × 10 1.57 × 10−4 1.57 × 10−4 ✓†?
https://doi.org/10.1371/journal.pcbi.1007575.t004

which agree well with the results of our algorithm. On the contrary, our algorithm successfully
infers one more parameter kd2 than the above analysis. This could be due to the fact that check-
ing practical identifiability using the FIM can be problematic, especially for partially observed
nonlinear systems [37]. We have similar results for the cell death model.

Ultradian endocrine model

The final test case for assessing the performance of the proposed algorithm is to infer parame-
ters of the ultradian model for the glucose-insulin interaction. We use a relatively simple ultra-
dian model [38] with 6 state variables and 30 parameters. This model is developed in a non-
pathophysiologic context and represents relatively simple physiologic mechanics. In this
model, the main variables are the plasma insulin concentration Ip, the interstitial insulin con-
centration Ii, the glucose concentration G, and a three stage filter (h1, h2, h3) that reflects the
response of the plasma insulin to glucose levels [38]. The resulting system of ODEs, the nomi-
nal values for the parameters of the ultradian model along with the initial conditions for the 6
state variables are given in S4 Text.
The nutritional driver IG(t) is the systematic forcing of the model that represents the exter-
nal sources of glucose from nutritional intake. Although the nutritional intake (modeled by

Fig 6. Fisher information matrix null eigenvectors of the cell apoptosis model. The most dominant component in each null eigenvector associated
with a specific parameter suggests that the parameter may not be practically identifiable: (left) cell survival and (right) cell death.
https://doi.org/10.1371/journal.pcbi.1007575.g006

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1007575 November 18, 2020 12 / 19

PLOS COMPUTATIONAL BIOLOGY Systems biology informed deep learning

Fig 7. Ultradian glucose-insulin model observation data given to the algorithm for parameter inference. 360
noiseless measurements on glucose level (G) only are randomly sampled in the time window of 0 − 1800 minutes (*
one day).
https://doi.org/10.1371/journal.pcbi.1007575.g007

the N discrete nutrition events) is required to be defined and properly recorded by the patients,
it is not always accurately recorded or may contain missing values. Therefore, it would be use-
ful to employ systems-informed neural networks to not only infer the model parameters given
the nutrition events, but also to assume that the intake is unknown (hidden forcing) and infer
the nutritional driver in Eq. S7f (S4 Text) as the same time.
Model parameter inference given the nutrition events. We consider an exponential
PN
decay functional form for the nutritional intake IG ðtÞ ¼ j¼1 mj k expðkðtj tÞÞ, where the
decay constant k is the only unknown parameter and three nutrition events are given by (tj,
mj) = [(300, 60) (650, 40) (1100, 50)] (min, g) pairs. The only observable is the glucose level
measurements G shown in Fig 7 (generated here synthetically by solving the system of ODEs),
which are sampled randomly to train the neural networks for the time window of 0 − 1800
minutes. Because we only use the observation of G and have more than 20 parameters to infer,
we limit the search range for these parameters. The range of seven parameters is adopted from
[38], and the range for other parameters is set as (0.2x, 1.8x), where x is the nominal value of
that parameter (S3 Table).
For the first test case, we set the parameters Vp, Vi and Vg to their nominal values and infer
the rest of the parameters. The inferred values are given in Table 5 (column Test 1), where we
observe good agreement between the target and inferred values. For the second test, we also
infer the values of Vp, Vi and Vg (Table 5). Although the inferred parameters are slightly worse
than the Test 1, when using the inferred parameters, we are able to solve the equations for
unseen time instants with high accuracy. We perform forecasting for the second test case after
training the algorithm using the glucose data in the time interval of t = 0 − 1800 min and infer-
ring the model parameters. Next, we consider that there is a nutrition event at time tj = 2000
min with carbohydrate intake of mj = 100 g. As shown in Fig 8, we are able to forecast with
high accuracy the glucose-insulin dynamics, more specifically, the glucose levels following the
nutrition intake.
Model parameter inference with hidden nutrition events. As detailed in the following,
one of the significant advantages of the systems-informed neural network is its ability to infer
the hidden systematic forcing in the model. For example, in the glucose-insulin model, the
nutritional driver IG is the forcing that we aim to infer as well. Here, we use the glucose mea-
surements to train the model for the time interval t = 0 − 1800 min shown in Fig 7, while we
assume that the time instants and quantities of three nutritional events are additionally
unknown.
We found that it is difficult to infer all the parameters as well as the the timing and carbohy-
drate content of each nutrition event. However, given Vp, Vi, Vg and the timing of each

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1007575 November 18, 2020 13 / 19

PLOS COMPUTATIONAL BIOLOGY Systems biology informed deep learning

Table 5. Parameter values for the ultradian glucose-insulin model and their corresponding inferred values.
Parameter Nominal value Inferred value (Test 1) Inferred value (Test 2)
Vp 3 – 2.97
Vi 11 – 9.25
Vg 10 – 11.3
E 0.2 0.209 0.216
tp 6 6.58 6.30
ti 100 96.6 136
td 12 11.8 11.6
k 0.0083 0.00837 0.00833
Rm 209 232 198
a1 6.6 6.56 6.48
C1 300 321 277
C2 144 52.6 44.9
C4 80 67.1 73.9
C5 26 25.2 25.6
Ub 72 68.3 73.6
U0/C3 0.04 0.0464 0.0463
Um/C3 0.9 0.790 0.975
Rg 180 182 182
α 7.5 7.89 7.94
β 1.772 1.91 1.85
https://doi.org/10.1371/journal.pcbi.1007575.t005

nutrition event, the algorithm is capable of inferring the other model parameters as well as the
carbohydrate content (S4 Table column Test 1). Having the nutrition events as well as all other
unknown parameters estimated, we are able to forecast the glucose levels for t = 1800 − 3000
min assuming there has been a nutritional intake of (tj, mj) = (2000, 100). The predictions for
the glucose G and the nutritional driver IG are shown in S7 Fig, which show excellent agree-
ment in the forecasting of glucose levels. For the second test, we also infer the values of Vp, Vi
and Vg, and the result is slightly worse (S4 Table column Test 2 and S8 Fig).
If both the timing and carbohydrate content of each nutrition event are unknown, the algo-
rithm is also capable to infer them by assuming that certain model parameters are known. We
found that the selection of the known parameters is important. As shown in S4 Table, we con-
sider different combinations of parameters to be known in Test 3 and Test 4; Test 3 leads to
good prediction accuracy (S9 Fig) while Test 4 does not.

Discussion
We presented a new and simple to implement “systems-biology-informed” deep learning algo-
rithm that can reliably and accurately infer the hidden dynamics described by a mathematical
model in the form of a system of ODEs. The system of ODEs is encoded into a plain “unin-
formed” deep neural networks and is enforced through minimizing the loss function that
includes the residuals of the ODEs. Enforcing the equations in the loss function adds addi-
tional constraints in the learning process, which leads to several advantages of the proposed
algorithm: first, we are able to infer the unknown parameters of the system of ODEs once the
neural network is trained; second, we can use a minimalistic amount of data on a few observ-
ables to infer the dynamics and the unknown parameters; third, the enforcement of the equa-
tions adds a regularization effect that makes the algorithm robust to noise (we have not used

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1007575 November 18, 2020 14 / 19

PLOS COMPUTATIONAL BIOLOGY Systems biology informed deep learning

Fig 8. Ultradian glucose-insulin inferred dynamics and forecasting compared with the exact solution given
nutrition events. 600 scattered observations of glucose level are randomly sampled from 0 − 1800 min and used for
training. Note that the parameter k in the intake function IG is considered to be unknown, while the timing and
carbohydrate content of each nutrition event are given. Given the inferred parameters, we can accurately forecast the
glucose levels following the event at time t = 2000 min.
https://doi.org/10.1371/journal.pcbi.1007575.g008

any other regularization technique); and lastly, the measurements can be scattered, noisy and
just a few.
The problem of structural and practical non-identifiability (such as the one encountered in
the cell apoptosis model) is a long-standing problem in the field of systems identification, and
has been under extensive research, e.g., [39]. Structural non-identifiabilities originate from
incomplete observation of the internal model states. Because the structural non-identifiability
is independent of the accuracy of experimental data measurements, we cannot resolve it by a
refinement of existing measurements, and one possible way to resolve this issue is increasing
the number of observed species. Our focus in this study is mostly on practical identifiability,
which can guide us to redesign the experiment, improve the model, or collect more experi-
mental measurements. In this study, we used FIM and local sensitivity analysis for the iden-
tifiability analysis, but we note that FIM has many limitations and can be problematic,
especially for partially observed nonlinear systems [37], and hence other advanced alternatives
[15, 40] should be used in future works. However, our goal in this work was not to do system-
atic identifiability analysis, but rather to use identifiability analysis to explain some of our
findings.

Conclusion
We have used three benchmark problems to assess the performance of the algorithm including
a highly nonlinear glycolysis model, a non-identifiable cell apoptosis model, and an ultradian
glucose-insulin model for glucose forecasting based on the nutritional intake. Given the system

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1007575 November 18, 2020 15 / 19

PLOS COMPUTATIONAL BIOLOGY Systems biology informed deep learning

of ODEs and initial conditions of the state variables, the algorithm is capable of accurately
inferring the whole dynamics with one or two observables, where the unknown parameters are
also inferred during the training process. An important and very useful outcome of the algo-
rithm is its ability to infer the systematic forcing or driver in the model such as the nutritional
intake in the glucose-insulin model. In this work, we considered the synthetic data of three
small problems to test the performance and limitation of the proposed method. We will apply
our method to larger problems and real data (e.g., the dataset introduced in [41]) in future
work.

Supporting information
S1 Text. Fisher information matrix.
(PDF)
S2 Text. Yeast glycolysis model.
(PDF)
S3 Text. Cell apoptosis model.
(PDF)
S4 Text. Ultradian endocrine model.
(PDF)
S1 Fig. Glycolysis oscillator noiseless observation data given to the algorithm for parame-
ter inference. 500 noiseless measurements of two observables S5 and S6 are randomly sampled
in the time window of 0 − 10 minutes.
(PDF)
S2 Fig. Glycolysis oscillator inferred dynamics compared with the exact solution. Predic-
tions are performed on equally-spaced time instants in the interval of 0 − 10 minutes. The scat-
tered observations are plotted using symbols for the two observables S5 and S6. The exact data
and the scattered observations are computed by solving the system of ODEs given in S2 Text.
(PDF)
S3 Fig. Correlation matrix for the parameters of glycolysis model. The correlation matrix is
computed using the local sensitivity analysis and the FIM for the parameters involved in the
glycolysis oscillator model assuming 10% noise in the observation data. We observe no perfect
correlations suggesting that FIM is not singular and the parameters in the glycolysis model are
practically identifiable.
(PDF)
S4 Fig. Glycolysis oscillator inferred dynamics from noisy measurements compared with
the exact solution. 200 scattered observations are plotted using symbols for the two observ-
ables S5 and S6.
(PDF)
S5 Fig. Cell survival inferred dynamics from noisy observations compared with the exact
solution. Predictions are performed on equally-spaced time instants in the interval of 0 − 60
hours. The scattered observations are plotted using symbols only for the observable x4. The
exact data and the scattered observations are computed by solving the system of ODEs given
in S3 Text.
(PDF)

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1007575 November 18, 2020 16 / 19

PLOS COMPUTATIONAL BIOLOGY Systems biology informed deep learning

S6 Fig. Correlation matrix for the parameters of the cell apoptosis model. The correlation
matrix is computed using FIM for the practical identifiablity analysis of parameters involved
in the cell apoptosis model assuming 5% noise in the observation data for two scenarios: (left)
cell survival and (right) cell death. We observe perfect correlations of � 1.0 between the
parameters suggesting that the FIM is singular and some parameters in the cell apoptosis
model are practically non-identifiable.
(PDF)
S7 Fig. Ultradian glucose-insulin inferred dynamics with hidden nutritional driver (Test 1
in S4 Table). Scattered observations of glucose level are randomly sampled from 0 − 1800 min
and used for training. The parameter k in the intake function IG as well as carbohydrate con-
tent (mj) of each nutrition event are treated as unknown, while Vp, Vi, Vg and the timing (tj) of
each nutrition event are given.
(PDF)
S8 Fig. Ultradian glucose-insulin inferred dynamics with hidden nutritional driver (Test 2
in S4 Table). Scattered observations of glucose level are randomly sampled from 0 − 1800 min
and used for training. The parameter k in the intake function IG as well as carbohydrate con-
tent (mj) of each nutrition event are treated as unknown, while the timing (tj) of each nutrition
event are given.
(PDF)
S9 Fig. Ultradian glucose-insulin inferred dynamics with hidden nutritional driver (Test 3
in S4 Table). Scattered observations of glucose level are randomly sampled from 0 − 1800 min
and used for training. The timing (tj) and carbohydrate content (mj) of each nutrition event
are treated as unknown, while the parameter k in the intake function IG is given.
(PDF)
S1 Table. Full list of parameters for glycolytic oscillator model [33].
(PDF)
S2 Table. Full list of parameters for cell apoptosis model [36].
(PDF)
S3 Table. Full list of parameters for the ultradian glucose-insulin model [42]. The search
range for the first 7 parameters is adopted from [38], and the range for the other parameters is
(0.2x, 1.8x), where x is the nominal value of that parameter.
(PDF)
S4 Table. Parameter values for the ultradian glucose-insulin model with hidden nutri-
tional driver and their corresponding inferred values.
(PDF)

Author Contributions
Conceptualization: Alireza Yazdani, Maziar Raissi, George Em Karniadakis.
Data curation: Alireza Yazdani.
Formal analysis: Alireza Yazdani, Lu Lu.
Funding acquisition: George Em Karniadakis.
Investigation: Alireza Yazdani, Lu Lu.

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1007575 November 18, 2020 17 / 19

PLOS COMPUTATIONAL BIOLOGY Systems biology informed deep learning

Methodology: Alireza Yazdani, Lu Lu, Maziar Raissi.

Software: Alireza Yazdani, Lu Lu, Maziar Raissi.
Validation: Lu Lu, Maziar Raissi.
Visualization: Alireza Yazdani, Lu Lu.
Writing – original draft: Alireza Yazdani.
Writing – review & editing: Lu Lu, Maziar Raissi, George Em Karniadakis.

References
1. Wilkinson DJ. Stochastic modelling for systems biology. CRC press; 2018.
2. Kitano H. Systems biology: a brief overview. Science. 2002; 295(5560):1662–1664. https://doi.org/10.
1126/science.1069492 PMID: 11872829
3. Cornish-Bowden A, Cornish-Bowden A. Fundamentals of enzyme kinetics. vol. 510. Wiley-Blackwell
Weinheim, Germany; 2012.
4. Mendes P, Kell D. Non-linear optimization of biochemical pathways: applications to metabolic engineer-
ing and parameter estimation. Bioinformatics (Oxford, England). 1998; 14(10):869–883. https://doi.org/
10.1093/bioinformatics/14.10.869 PMID: 9927716
5. Srinivas M, Patnaik LM. Genetic algorithms: A survey. Computer. 1994; 27(6):17–26. https://doi.org/10.
1109/2.294849
6. Moles CG, Mendes P, Banga JR. Parameter estimation in biochemical pathways: a comparison of
global optimization methods. Genome research. 2003; 13(11):2467–2474. https://doi.org/10.1101/gr.
1262503 PMID: 14559783
7. Villaverde AF, Fröhlich F, Weindl D, Hasenauer J, Banga JR. Benchmarking optimization methods for
parameter estimation in large kinetic models. Bioinformatics. 2019; 35(5):830–838. https://doi.org/10.
1093/bioinformatics/bty736 PMID: 30816929
8. Wilkinson DJ. Bayesian methods in bioinformatics and computational systems biology. Briefings in bio-
informatics. 2007; 8(2):109–116. https://doi.org/10.1093/bib/bbm007 PMID: 17430978
9. Calvetti D, Somersalo E. An introduction to Bayesian scientific computing: ten lectures on subjective
computing. vol. 2. Springer Science & Business Media; 2007.
10. Lillacci G, Khammash M. Parameter estimation and model selection in computational biology. PLoS
computational biology. 2010; 6(3):e1000696. https://doi.org/10.1371/journal.pcbi.1000696 PMID:
20221262
11. Quach M, Brunel N, d’Alché Buc F. Estimating parameters and hidden variables in non-linear state-
space models based on ODEs for biological networks inference. Bioinformatics. 2007; 23(23):3209–
3216. https://doi.org/10.1093/bioinformatics/btm510 PMID: 18042557
12. Albers DJ, Blancquart PA, Levine ME, Seylabi EE, Stuart A. Ensemble Kalman methods with con-
straints. Inverse Problems. 2019; 35(9):095007. https://doi.org/10.1088/1361-6420/ab1c09
13. Engelhardt B, Frőhlich H, Kschischo M. Learning (from) the errors of a systems biology model. Scientific
reports. 2016; 6(1):1–9. https://doi.org/10.1038/srep20772 PMID: 26865316
14. Engelhardt B, Kschischo M, Fröhlich H. A Bayesian approach to estimating hidden variables as well as
missing and wrong molecular interactions in ordinary differential equation-based mathematical models.
Journal of The Royal Society Interface. 2017; 14(131):20170332. https://doi.org/10.1098/rsif.2017.
0332 PMID: 28615495
15. Raue A, Kreutz C, Maiwald T, Bachmann J, Schilling M, Klingmüller U, et al. Structural and practical
identifiability analysis of partially observed dynamical models by exploiting the profile likelihood. Bioin-
formatics. 2009; 25(15):1923–1929. https://doi.org/10.1093/bioinformatics/btp358 PMID: 19505944
16. Villaverde AF, Tsiantis N, Banga JR. Full observability and estimation of unknown inputs, states and
parameters of nonlinear biological models. Journal of the Royal Society Interface. 2019; 16
(156):20190043. https://doi.org/10.1098/rsif.2019.0043 PMID: 31266417
17. Gutenkunst RN, Waterfall JJ, Casey FP, Brown KS, Myers CR, Sethna JP. Universally sloppy parame-
ter sensitivities in systems biology models. PLoS computational biology. 2007; 3(10):e189. https://doi.
org/10.1371/journal.pcbi.0030189 PMID: 17922568
18. Chis OT, Banga JR, Balsa-Canto E. Structural identifiability of systems biology models: a critical com-
parison of methods. PloS one. 2011; 6(11):e27755. https://doi.org/10.1371/journal.pone.0027755
PMID: 22132135

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1007575 November 18, 2020 18 / 19

PLOS COMPUTATIONAL BIOLOGY Systems biology informed deep learning

19. Majda AJ. Challenges in climate science and contemporary applied mathematics. Communications on
Pure and Applied Mathematics. 2012; 65(7):920–948. https://doi.org/10.1002/cpa.21401
20. Rodriguez-Fernandez M, Mendes P, Banga JR. A hybrid approach for efficient and robust parameter
estimation in biochemical pathways. Biosystems. 2006; 83(2-3):248–265. https://doi.org/10.1016/j.
biosystems.2005.06.016 PMID: 16236429
21. Goodfellow I BY, A C. Deep Learning; 2017.
22. Raissi M, Perdikaris P, Karniadakis GE. Physics-informed neural networks: A deep learning framework
for solving forward and inverse problems involving nonlinear partial differential equations. Journal of
Computational Physics. 2019; 378:686–707. https://doi.org/10.1016/j.jcp.2018.10.045
23. Lu L, Meng X, Mao Z, Karniadakis GE. DeepXDE: A deep learning library for solving differential equa-
tions. arXiv preprint arXiv:190704502. 2019.
24. Kingma DP, Ba J. Adam: A method for stochastic optimization. arXiv preprint arXiv:14126980. 2014.
25. Caughlin D. Parameter identification methods for metamodeling simulations. In: Proceedings of the
28th conference on Winter simulation; 1996. p. 756–763.
26. Song Xm, Kong Fz, Zhan Cs, Han Jw, Zhang Xh. Parameter identification and global sensitivity analysis
of Xin’anjiang model using meta-modeling approach. Water Science and Engineering. 2013; 6(1):1–17.
27. Wang S, Teng Y, Perdikaris P. Understanding and mitigating gradient pathologies in physics-informed
neural networks. arXiv preprint arXiv:200104536. 2020.
28. Pohjanpalo H. System identifiability based on the power series expansion of the solution. Mathematical
biosciences. 1978; 41(1-2):21–33. https://doi.org/10.1016/0025-5564(78)90063-9
29. Ljung L, Glad T. On global identifiability for arbitrary model parametrizations. Automatica. 1994; 30
(2):265–276. https://doi.org/10.1016/0005-1098(94)90029-9
30. Balsa-Canto E, Alonso AA, Banga JR. Computational procedures for optimal experimental design in
biological systems. IET systems biology. 2008; 2(4):163–172. https://doi.org/10.1049/iet-syb:20070069
PMID: 18681746
31. Foo J, Sindi S, Karniadakis GE. Multi-element probabilistic collocation for sensitivity analysis in cellular
signalling networks. IET systems biology. 2009; 3(4):239–254. https://doi.org/10.1049/iet-syb.2008.
0126 PMID: 19640163
32. Ramachandran P, Zoph B, Le QV. Searching for activation functions. arXiv preprint arXiv:171005941.
2017.
33. Ruoff P, Christensen MK, Wolf J, Heinrich R. Temperature dependency and temperature compensation
in a model of yeast glycolytic oscillations. Biophysical chemistry. 2003; 106(2):179–192. https://doi.org/
10.1016/S0301-4622(03)00191-1 PMID: 14556906
34. Schmidt M, Lipson H. Distilling free-form natural laws from experimental data. science. 2009; 324
(5923):81–85. https://doi.org/10.1126/science.1165893 PMID: 19342586
35. Daniels BC, Nemenman I. Efficient inference of parsimonious phenomenological models of cellular
dynamics using S-systems and alternating regression. PloS one. 2015; 10(3):e0119821. https://doi.org/
10.1371/journal.pone.0119821 PMID: 25806510
36. Aldridge BB, Haller G, Sorger PK, Lauffenburger DA. Direct Lyapunov exponent analysis enables
parametric study of transient signalling governing cell behaviour. IEE Proceedings-Systems Biology.
2006; 153(6):425–432. https://doi.org/10.1049/ip-syb:20050065 PMID: 17186704
37. Joshi M, Seidel-Morgenstern A, Kremling A. Exploiting the bootstrap method for quantifying parameter
confidence intervals in dynamical systems. Metabolic engineering. 2006; 8(5):447–455. https://doi.org/
10.1016/j.ymben.2006.04.003
38. Sturis J, Polonsky KS, Mosekilde E, Van Cauter E. Computer model for mechanisms underlying ultra-
dian oscillations of insulin and glucose. American Journal of Physiology-Endocrinology And Metabo-
lism. 1991; 260(5):E801–E809. https://doi.org/10.1152/ajpendo.1991.260.5.E801 PMID: 2035636
39. Hong H, Ovchinnikov A, Pogudin G, Yap C. SIAN: software for structural identifiability analysis of ODE
models. Bioinformatics. 2019; 35(16):2873–2874. https://doi.org/10.1093/bioinformatics/bty1069
PMID: 30601937
40. Gábor A, Villaverde AF, Banga JR. Parameter identifiability analysis and visualization in large-scale
kinetic models of biosystems. BMC systems biology. 2017; 11(1):1–16. https://doi.org/10.1186/s12918-
017-0428-y PMID: 28476119
41. Albers D, Levine M, Sirlanci M, Stuart A. A Simple Modeling Framework For Prediction In The Human
Glucose-Insulin System. arXiv preprint arXiv:191014193. 2019.
42. Albers DJ, Levine M, Gluckman B, Ginsberg H, Hripcsak G, Mamykina L. Personalized glucose fore-
casting for type 2 diabetes using data assimilation. PLoS computational biology. 2017; 13(4):e1005232.
https://doi.org/10.1371/journal.pcbi.1005232 PMID: 28448498

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1007575 November 18, 2020 19 / 19