This is an open access article published under an ACS AuthorChoice License, which permits

copying and redistribution of the article or any adaptations for non-commercial purposes.

Article

pubs.acs.org/molecularpharmaceutics

Deep Learning Applications for Predicting Pharmacological

Properties of Drugs and Drug Repurposing Using Transcriptomic
Data
Alexander Aliper,*,† Sergey Plis,‡,§ Artem Artemov,† Alvaro Ulloa,§ Polina Mamoshina,†
and Alex Zhavoronkov*,†,∥
†
Insilico Medicine, ETC, B301, Johns Hopkins University, Baltimore, Maryland 21218, United States
‡
Datalytic Solutions, 1101 Yale Boulevard NE, Albuquerque, New Mexico 87106, United States
§
The Mind Research Network, Albuquerque, New Mexico 87106, United States
See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.

∥
The Biogerontology Research Foundation, Oxford, U.K.
*
S Supporting Information
Downloaded via 117.249.230.177 on February 20, 2023 at 17:05:35 (UTC).

ABSTRACT: Deep learning is rapidly advancing many areas

of science and technology with multiple success stories in
image, text, voice and video recognition, robotics, and
autonomous driving. In this paper we demonstrate how deep
neural networks (DNN) trained on large transcriptional
response data sets can classify various drugs to therapeutic
categories solely based on their transcriptional profiles. We
used the perturbation samples of 678 drugs across A549,
MCF-7, and PC-3 cell lines from the LINCS Project and
linked those to 12 therapeutic use categories derived from
MeSH. To train the DNN, we utilized both gene level transcriptomic data and transcriptomic data processed using a pathway
activation scoring algorithm, for a pooled data set of samples perturbed with different concentrations of the drug for 6 and 24
hours. In both pathway and gene level classification, DNN achieved high classification accuracy and convincingly outperformed
the support vector machine (SVM) model on every multiclass classification problem, however, models based on pathway level
data performed significantly better. For the first time we demonstrate a deep learning neural net trained on transcriptomic data to
recognize pharmacological properties of multiple drugs across different biological systems and conditions. We also propose using
deep neural net confusion matrices for drug repositioning. This work is a proof of principle for applying deep learning to drug
discovery and development.
KEYWORDS: deep learning, DNN, predictor, drug repurposing, drug discovery, confusion matrix, deep neural networks

■ INTRODUCTION
Drug discovery and development is a complicated and time and
mining and recognition,6 improving state of the art perform-
ances by more than 30%, where the prior decade struggled to
resource consuming process, and various computational obtain 1−2% improvements. Traditional machine learning
approaches are regularly being developed to improve it. In approaches have achieved significant levels of classification
silico drug discovery1,2 has evolved over the past decade and accuracy, but at the price of manually selected and tuned
offers a targeted, efficient approach compared to those of the features. Arguably, feature engineering is the dominating
past, which often relied on either identifying active ingredients research component in practical applications of ML. In
in traditional remedies or, in many cases, serendipitous contrast, the power of NNs is in automatic feature learning
discovery. Modern methods include data mining, structure from massive data sets. Not only does it simplify manual and
modeling (homology modeling), traditional machine learning3 laborious feature engineering but also it allows learning task-
(ML), and its biologically inspired branch technique, deep optimal features.
learning (DL).4 Modern biology has entered the era of Big Data, wherein
DL4 methods model high-level representations of data using data sets are too large, high-dimensional, and complex for
deep neural networks (DNNs). DNNs are flexible multilayer classical computational biology methods. The ability to learn at
systems of connected and interacting artificial neurons that
perform various data transformations. They have several hidden Received: March 18, 2016
layers of neurons, which number variation allows adjusting the Revised: May 13, 2016
level of data abstraction. DL now plays a dominant role in the Accepted: May 20, 2016
areas of physics5 and speech, signal, image, video, and text Published: May 20, 2016

© 2016 American Chemical Society 2524 DOI: 10.1021/acs.molpharmaceut.6b00248

Mol. Pharmaceutics 2016, 13, 2524−2530
Molecular Pharmaceutics Article

Figure 1. Study design. Gene expression data from LINCS Project was linked to 12 MeSH therapeutic use categories. DNN was trained separately
on gene expression level data for “landmark genes” and pathway activation scores for significantly perturbed samples, forming input layers of 977 and
271 neural nodes, respectively.

the higher levels of abstraction made DL a promising and categories solely based on their transcriptional profiles. In total,
effective tool for working with biological and chemical data.7 we analyzed 26,420 drug perturbation samples for three cell
Methods using DL architecture are capable of dealing with lines from the Broad LINCS database. All samples were
sparse and complex information, which is especially demanded assigned to 12 specific therapeutic use categories according to
in the analysis of high-dimensional gene expression data. “Curse MeSH classification of the particular drug (Supplementary
of dimensionality” is one of the major problems of gene Table 1). Since a number of drugs were present in multiple
expression data that can be solved by feature selection categories, we considered only those drugs that belong only to
implementing standard data projection methods as PCA or one category. To increase the number of samples in each of the
more biologically relevant as pathway analysis.8 DNNs categories and to make the classification more robust for each
demonstrate the state-of-the-art performance extracting features given drug, we aggregated all samples corresponding to all
from sparse transcriptomics data (both mRNA and miRNA possible perturbation time, perturbation concentration, and cell
data),9 in classifying cancer using gene expression data10 and line parameters (Supplementary Table 2).
predicting splicing code patterns.11 DL has been effectively When dealing with transcriptional data at the gene level, a
applied in biomodeling and structural genomics to predict common problem is the so-called “curse of dimensionality”.
protein 3-D structure using protein sequence (ordered or Indeed, when we applied DNN on gene level data for whole
disordered protein (with lack of fixed 3-D structure))12,13 and data set of 12,797 genes, it did not perform very well, achieving
may become an essential tool for development of new drugs.14 only 0.24 mean F1 score on 12 classes. So our first step was
DL approaches were successfully implemented to predict proper feature selection. Here we investigated two approaches:
drug−target interactions,15 model reaction properties of pathway activation scoring and using “landmark genes” as new
molecules,16 and calculate toxicity of drugs.17 As deep networks features.
incorporate more features from biology,18 application breadth Pathway Level. For pathway level analysis we used a
and accuracy will likely increase. previously established pathway analysis method called Onco-
Drug repurposing or target extension allows prediction of Finder.25−30 It preserves biological function and allows for
new potential applications of medications or even new dimensionality reduction at the same time. In contrast to other
therapeutic classes of drugs using gene expression data before pathway analysis tools, which mostly implement pathway
and after treatment (e.g., before and after incubation of a cell enrichment analysis, OncoFinder performs quantitative estima-
line with multiple drugs). There are multiple in silico tion of signaling pathway activation strength, and the sign of the
approaches to drug discovery and classification,19−21 and resulting value indicates how significantly the pathway is up- or
many attempts were made to predict transcriptional response downregulated. All perturbation samples were analyzed with
with functional properties of drugs.22−24 In this study we
this tool and for each sample we calculated pathway activation
addressed this problem by classifying various drugs to
profiles for 271 signaling pathways. Samples with zero pathway
therapeutic categories with DNN solely based on their
activation score for all of the pathways were considered as
transcriptional profiles. We used the perturbation samples of
insignificantly perturbed and were excluded from further
X drugs across A549, MCF-7, and PC-3 cell lines from the
analysis. That resulted in a final data set containing 308, 454,
LINCS Project and linked those to 12 therapeutic use
and 433 drugs for A549, MCF7, and PC3 cell lines,
categories derived from MeSH therapeutic use section (Figure
1). After that we independently used both gene expression level respectively, and totalling 9352 samples (Supplementary
data for “landmark genes” and pathway activation scores to Table 2).
train DNN classifier. Using this data set we built a deep learning classifier based

■
only on pathway activation scores for drug perturbation profiles
of 3 cell lines: A549, MCF-7, and PC-3. Making a classifier
RESULTS based on a pooled data set with different cell lines, drug
The main aim of this study was to apply and estimate the concentration, and perturbation time, we are able to estimate
accuracy of DL methods to classify various drugs to therapeutic the classification performance in recognizing complex drug
2525 DOI: 10.1021/acs.molpharmaceut.6b00248
Mol. Pharmaceutics 2016, 13, 2524−2530
Molecular Pharmaceutics Article

Figure 2. Classification results. Classification performance of DNN and SVM trained on signaling pathways (a, b, c) and landmark genes (d, e, f) for
3, 5, and 12 drug classes, respectively, after 10-fold cross validation. Training and validation set results are shown in gray and green colors,
respectively.

action patterns across different biological conditions. For the 3- mean F1 scores of 0.372, 0.238, and 0.202 for respective tasks
class classification problem we chose the most abundant (Figure 2d−f).
categories: antineoplastic, cardiovascular, and central nervous DNN as Drug Repurposing Tool. Here we tried to dig a
system agents. DNN achieved 10-fold cross-validation mean F1 bit deeper into classification results on pathway level, since the
score of 0.701. We compared the results of DNN to another DNN model worked best with pathways as features. To
popular classification algorithm called support vector machine determine which of the 12 therapeutic use categories are the
(SVM) trained via nested 3-fold cross validation for several most detectable by DNN, we calculated 10-fold cross-validation
hyperparameters (see Materials and Methods). On 3-class classification accuracy of each category. Antineoplastic agents
classification problem SVM performed with mean F1 score of turned out to be the most “recognizable” category by a large
0.530. margin, with 0.686 accuracy on 12 classes. This was followed by
Addition of gastrointestinal and anti-infective classes anti-infective, central nervous system, and dermatologic
decreased the mean F1 score of DNN to 0.596. Mean F1 categories, with 0.513, 0.506, and 0.505 accuracy, respectively.
score for SVM dropped as well, down to 0.417. The least “recognizable” on the same number of classes was
When all 12 classes were considered, the classification neural hematologic agents, with accuracy of 0.23. On 3- and 5-class
performance decreased in a minor way, with a cross-validation classification problems, the category antineoplastic drugs was
mean F1 score of 0.546. SVM performed with cross-validation on top as well, with accuracy of 0.82 and 0.742. Separability of
mean F1 score of 0.366 on the same 12-class classification therapeutic categories by DNN can be illustrated with
problem. The performance comparison of DNN and SVM on confusion matrices (Figure 3). Here we observed that the
investigated classification problems is depicted in Figure 2a−c. cardiovascular category drugs was relatively often misclassified
These results indicate that our model performance far exceeds as central nervous system and antineoplastic agents. In contrast,
random chance,31 and we can conclude that DNN out- the level of false positives for the antineoplastic category was
performed SVM on every multiclass classification problem. relatively small. If we look even closer into the results,
Landmark Gene Level. In our second feature selection sometimes these misclassified false positive drugs may in fact
approach we used a data set containing normalized gene represent a possibility for drug repurposing. For instance, well-
expression data for 977 “landmark genes”. According to the known muscarinic receptor antagonist otenzepad was mis-
authors of LINCS Project they can capture approximately 80% classified as central nervous system agent, but despite its
of the information and possess great inferential value. For fair obvious role in brain function,32,33 according to the MeSH
comparison we trained DNN exactly the same way we did on therapeutic use section, it is only used against cardiac
the pathway level. We used the same data set of 9352 arrhythmia. Another example includes vasodilator pinacidil, a
significantly perturbed samples and tested the performance of cyanoguanidine drug that opens ATP-sensitive potassium
DNN on the same classification problems. DNN trained on channels, which was misclassified as central nervous system
“landmark gene” data performed with 10-fold cross-validation agent in several cross-validation iterations, although it is used
mean F1 scores of 0.397, 0.285, and 0.234 for 3-, 5-, and 12- only in cardiovascular conditions. It is known that potassium
class classification tasks, respectively. The SVM model showed channels play important roles in different brain regions,34 and
2526 DOI: 10.1021/acs.molpharmaceut.6b00248
Mol. Pharmaceutics 2016, 13, 2524−2530
Molecular Pharmaceutics Article

Figure 3. Validation confusion matrix representing deep neural network classification performance over a set of drugs profiled for A549, MCF7, and
PC3 cell lines, belonging to 3 (a), 5 (b), and 12 (c) therapeutic classes. C(i,j) element is a sample count of how many times i was the truth and j was
predicted.

pinacidil might influence some of them. Aforementioned cases drugs into therapeutic categories solely based on their
hint to the fact that imperfect accuracy here might not be a bad transcriptomic data. To our knowledge, this is the first DL
thing and the DNN model could serve as powerful drug model to map transcriptomic data onto therapeutical category.
repositioning tool. DNN trained on gene level data did not perform very well,

■ DISCUSSION
With increasing availability of big data and GPU computing, the
achieving only 0.24 F1 score on 12 classes. Thus, as a way to
reduce dimensionality and keep biological relevance, we
decided to apply pathway activation scoring.27 Translation of
entire field of deep learning is experiencing very rapid perturbation profiles onto the pathway level turned out to be
development, and the breadth of DNN applications goes far very beneficial. Pathways served as excellent features, and we
beyond text, voice, and image recognition problems. In this were able to exclude insignificantly perturbed samples and
paper we explored the possibility of using DL to classify various demonstrate the ability of deep neural network to recognize
2527 DOI: 10.1021/acs.molpharmaceut.6b00248
Mol. Pharmaceutics 2016, 13, 2524−2530
Molecular Pharmaceutics Article

sophisticated drug action mechanisms on the pathway level. comparison using Student’s t test, generates the list of
The power of this approach is further highlighted by the fact significantly differentially expressed genes, and calculates the
that it performs with high accuracy on a pooled data set with pathway activation strength (PAS), a value which serves as a
samples from cells treated by different drug concentrations and qualitative measure of pathway activation. Positive and negative
perturbation times. Furthermore, the DNN achieves significant PAS values indicate pathway up- and downregulation,
classification accuracy even across different cell lines. Its respectively. In this study the genes with FDR-adjusted p-
performance turned out to be better than SVM on every value <0.05 were considered significantly differentially ex-
multiclass classification problem. When we used the same set of pressed. Samples with zero pathway activation score for all of
significantly perturbed samples that we selected with the the pathways were considered as insignificantly perturbed and
pathway activation approach, and trained DNN on a data set were excluded from further analysis. The filtered data set
with gene expression data for “landmark genes”, the results contained 308, 454, and 433 drugs for A549, MCF7, and PC3
turned out to be significantly worse. Hence, we can conclude cell lines, respectively, and comprised 9352 samples total
that pathway level data is more complementary for DNN and (Supplementary Tables 1 and 2).
more suitable for classifying drugs into therapeutic use Classification Methods. Among the multitude of available
categories. Proper comparison to reference group of samples classification methods we have employed two that are highly
plays an essential role, and merely normalized gene expression robust and widely successful in fields outside drug prediction:
data from perturbed samples is not sufficient for complex SVM36,37 and deep neural network.38 SVMs are a celebrated
classification tasks. classification method for their flexibility and ease of use, while
It is possible to interpret the classification results from deep learning approaches are continuing to break records in
different angles. For instance, as it was shown, in confusion many pattern recognition tasks.
matrices (Figure 3), that the “misclassified” samples for a Flexibility of SVM, as a maximum margin classifier, is in part
certain drug might in fact be an indication of its potential for reflected in provided ability to select a kernel that fits the data
novel use, or repurposing, in these exact “incorrectly” assigned best and choose a soft-margin parameter that allows for best
conditions. Misclassification, therefore, may lead to unexpected
generalization. However, these parameters are not evident for
new discoveries. This approach opens a great avenue for
any given data set. It has been previously shown that radial basis
application of DL in the drug repurposing field.

■
function (RBF) kernel SVMs perform well on largely different
MATERIALS AND METHODS data. In order to be more objective in selection of kernel we
have allowed our nested cross validation choice of three. We
Data Collection. In this study, we performed the analysis of used grid search for hyperparameter optimization. We have
gene expression data produced by the LINCS Project trained the SVM via nested 3-fold cross validation for
participants (http://www.lincsproject.org/). We utilized the hyperparameters that include kernel type (linear, RBF, or
level 3 (Q2NORM) gene expression data for three cell lines: polynomial) and soft-margin parameter C embedded in the
MCF7, A549, and PC3. Q2NORM data contains the outer 10-fold cross-validation loop. For each fold, the algorithm
expression levels of both directly measured landmark tran- could have selected different kernels and different soft-margin
scripts plus inferred genes, which were normalized using parameters. Nevertheless, RBF was indeed the most preferred.
invariant set scaling followed by quantile normalization. The deep learning method used in our work was the standard
Mapping probes onto official HGNC35 gene symbols resulted
fully connected multilayer perceptron with 977 input nodes for
in a gene expression data set comprising 12,797 genes total.
gene expression level data and 271 for pathway activation
Drug Selection. To link the drugs profiled by LINCS
scores. Similarly to SVM, we used grid search for hyper-
Project to medical conditions, we utilized MeSH (medical
parameter optimization. We used cross entropy as cost function
subject headings) classification (https://www.nlm.nih.gov/
mesh/). We picked only those drugs that had a link to a and AdaDelta39 as cost function optimizer. We searched for the
disease in the “therapeutic use” section of the classification tree. optimal number of layers, number of hidden units, and dropout
Drugs belonging to several categories simultaneously were rejection rate in a nested cross-validation framework. For the
excluded from the analysis. hyperparameter search, the number of layers varied from 3 to 8,
Gene Expression Analysis. All of the selected drugs’ where the number of hidden units per layer was reduced with
samples collected for three cell lines were grouped by the the depth of the network to half the previous layer. We set the
combination of the following parameters: cell line, drug, search for the starting hidden layer from 300 to 900 in steps of
perturbation concentration, and perturbation time. This 150. Each layer was initialized with the Glorot uniform
resulted in a total number of 26,420 samples. For gene level approach.40 The experiment shows that the best combination
analysis we trained our models on both whole data set of of parameters was 3 hidden layers with 200 in each with
12,797 genes and a subset of 977 so-called “landmark genes” rectified linear activation function. The dropout rejection ratio
defined in the LINCS Project. was tested for 20% and 80% at each layer. We chose an
In pathway level analysis, for each given case sample group antisymmetric activation function, hyperbolic tangent, as the
we generated a reference group consisting of samples perturbed nonlinear function of hidden neurons because the data is
with DMSO, that came from the same RNA plate as samples normalized to zero mean and unit variance, thus we rely on
from the case group. After that, each case sample group was deviations from the mean to train the network. Also, it has been
independently analyzed using an algorithm called Onco- reported that the hyperbolic tangent speeds up convergence
Finder.27 Taking the preprocessed gene expression data as an compared to sigmoid functions.39 The number of output nodes
input, it allows for cross-platform data set comparison with low was equal to the number of classes to explore in each particular
error rate and has the ability to obtain functional features of experiment with a softmax activation function. The code that
intracellular regulation using mathematical estimations. For implements the feed-forward neural network used in our
each investigated sample group it performs a case-reference experiments is publicly available at https://github.com/
2528 DOI: 10.1021/acs.molpharmaceut.6b00248
Mol. Pharmaceutics 2016, 13, 2524−2530
Molecular Pharmaceutics Article

alvarouc/mlp (commit: 8b07a1a18b17ca530fdcb482fce- (13) Wang, S.; Weng, S.; Ma, J.; Tang, Q. DeepCNF-D: Predicting
c24e26e36b27a). Protein Order/Disorder Regions by Weighted Deep Convolutional

■
Neural Fields. Int. J. Mol. Sci. 2015, 16 (8), 17315−17330.
(14) Lusci, A.; Pollastri, G.; Baldi, P. Deep Architectures and Deep
ASSOCIATED CONTENT Learning in Chemoinformatics: The Prediction of Aqueous Solubility
*
S Supporting Information for Drug-like Molecules. J. Chem. Inf. Model. 2013, 53 (7), 1563−1575.
The Supporting Information is available free of charge on the (15) Wang, C.; Caihua, W.; Juan, L.; Fei, L.; Yafang, T.; Zixin, D.;
ACS Publications website at DOI: 10.1021/acs.molpharma- Qian-Nan, H. Pairwise Input Neural Network for Target-Ligand
ceut.6b00248. Interaction Prediction. In 2014 IEEE International Conference on
Bioinformatics and Biomedicine (BIBM); 2014.
MeSH category stratification binary matrix (XLSX) (16) Hughes, T. B.; Miller, G. P.; Swamidass, S. J. Modeling
Number of drugs selected for A549, MCF7, and PC3 cell Epoxidation of Drug-like Molecules with a Deep Machine Learning
Network. ACS Cent. Sci. 2015, 1 (4), 168−180.
lines (XLSX) (17) Xu, Y.; Dai, Z.; Chen, F.; Gao, S.; Pei, J.; Lai, L. Deep Learning

■ AUTHOR INFORMATION
Corresponding Author
for Drug-Induced Liver Injury. J. Chem. Inf. Model. 2015, 55 (10),
2085−2093.
(18) Solovyeva, K. P.; Karandashev, I. M.; Zhavoronkov, A.; Dunin-
Barkowski, W. L. Models of Innate Neural Attractors and Their
*E-mail: aliper@insilicomedicine.com. Applications for Neural Information Processing. Front. Syst. Neurosci.
Notes 2015, DOI: 10.3389/fnsys.2015.00178.
The authors declare no competing financial interest. (19) Newby, D.; Freitas, A. A.; Ghafourian, T. Comparing Multilabel

■
Classification Methods for Provisional Biopharmaceutics Class
Prediction. Mol. Pharmaceutics 2015, 12 (1), 87−102.
ACKNOWLEDGMENTS (20) Wenlock, M. C.; Barton, P. In Silico Physicochemical Parameter
We would like to thank Dr. Leslie C. Jellen of Insilico Medicine Predictions. Mol. Pharmaceutics 2013, 10 (4), 1224−1235.
for editing this manuscript for language and style. Authors (21) Broccatelli, F.; Cruciani, G.; Benet, L. Z.; Oprea, T. I. BDDCS
thank Mark Berger and NVIDIA Corporation for providing Class Prediction for New Molecular Entities. Mol. Pharmaceutics 2012,
9 (3), 570−580.
valuable advice and high performance GPU equipment for deep (22) Herrera-Ruiz, D.; Faria, T. N.; Bhardwaj, R. K.; Timoszyk, J.;
learning applications.

■
Gudmundsson, O. S.; Moench, P.; Wall, D. A.; Smith, R. L.; Knipp, G.
T. A Novel hPepT1 Stably Transfected Cell Line: Establishing a
REFERENCES Correlation between Expression and Function. Mol. Pharmaceutics
(1) Loging, W.; Harland, L.; Williams-Jones, B. High-Throughput 2004, 1 (2), 136−144.
Electronic Biology: Mining Information for Drug Discovery. Nat. Rev. (23) Iskar, M.; Zeller, G.; Blattmann, P.; Campillos, M.; Kuhn, M.;
Drug Discovery 2007, DOI: 10.1038/nrd2345. Kaminska, K. H.; Runz, H.; Gavin, A.-C.; Pepperkok, R.; van Noort,
(2) Kirchmair, J.; Göller, A. H.; Lang, D.; Kunze, J.; Testa, B.; Wilson, V.; Bork, P. Characterization of Drug-Induced Transcriptional
I. D.; Glen, R. C.; Schneider, G. Predicting Drug Metabolism: Modules: Towards Drug Repositioning and Functional Under-
Experiment and/or Computation? Nat. Rev. Drug Discovery 2015, 14, standing. Mol. Syst. Biol. 2013, 9, 662.
387. (24) Kutalik, Z.; Beckmann, J. S.; Bergmann, S. A Modular Approach
(3) Schirle, M.; Jenkins, J. L. Identifying Compound Efficacy Targets for Integrative Analysis of Large-Scale Gene-Expression and Drug-
in Phenotypic Drug Discovery. Drug Discovery Today 2016, 21, 82. Response Data. Nat. Biotechnol. 2008, 26 (5), 531−539.
(4) LeCun, Y.; Bengio, Y.; Hinton, G. Deep Learning. Nature 2015, (25) Spirin, P. V.; Lebedev, T. D.; Orlova, N. N.; Gornostaeva, A. S.;
521 (7553), 436−444. Prokofjeva, M. M.; Nikitenko, N. A.; Dmitriev, S. E.; Buzdin, A. A.;
(5) Baldi, P.; Sadowski, P.; Whiteson, D. Searching for Exotic Borisov, N. M.; Aliper, A. M.; Garazha, A. V.; Rubtsov, P. M.; Stocking,
Particles in High-Energy Physics with Deep Learning. Nat. Commun. C.; Prassolov, V. S. Silencing AML1-ETO Gene Expression Leads to
2014, 5, 4308. Simultaneous Activation of Both pro-Apoptotic and Proliferation
(6) Schmidhuber, J. Deep Learning in Neural Networks: An Signaling. Leukemia 2014, 28 (11), 2222−2228.
Overview. Neural Networks 2015, 61, 85−117. (26) Zhu, Q.; Izumchenko, E.; Aliper, A. M.; Makarev, E.; Paz, K.;
(7) Mamoshina, P.; Vieira, A.; Putin, E.; Zhavoronkov, A. Buzdin, A. A.; Zhavoronkov, A. A.; Sidransky, D. Pathway Activation
Applications of Deep Learning in Biomedicine. Mol. Pharmaceutics Strength Is a Novel Independent Prognostic Biomarker for Cetuximab
2016, 13 (5), 1445−1454. Sensitivity in Colorectal Cancer Patients. Hum. Genome Var. 2015, 2,
(8) Hira, Z. M.; Gillies, D. F. A Review of Feature Selection and 15009.
Feature Extraction Methods Applied on Microarray Data. Adv. Bioinf. (27) Buzdin, A. A.; Zhavoronkov, A. A.; Korzinkin, M. B.; Venkova,
2015, 2015, 198363. L. S.; Zenin, A. A.; Smirnov, P. Y.; Borisov, N. M. Oncofinder, a new
(9) Ibrahim, R.; Rania, I.; Yousri, N. A.; Ismail, M. A.; El-Makky, N. method for the analysis of intracellular signaling pathway activation
M. Multi-Level gene/MiRNA Feature Selection Using Deep Belief using transcriptomic data. Front. Genet. 2014, 5, 55.
Nets and Active Learning. In 2014 36th Annual International (28) Artemov, A.; Aliper, A.; Korzinkin, M.; Lezhnina, K.; Jellen, L.;
Conference of the IEEE Engineering in Medicine and Biology Society; Zhukov, N.; Roumiantsev, S.; Gaifullin, N.; Zhavoronkov, A.; Borisov,
2014. N.; Buzdin, A. A Method for Predicting Target Drug Efficiency in
(10) Fakoor, R.; Ladhak, F.; Nazi, A.; Huber, M. Using Deep Learning Cancer Based on the Analysis of Signaling Pathway Activation.
to Enhance Cancer Diagnosis and Classification. In Proceedings of the Oncotarget 2015, 6 (30), 29347−29356.
International Conference on Machine Learning; 2013. (29) Venkova, L.; Aliper, A.; Suntsova, M.; Kholodenko, R.; Shepelin,
(11) Leung, M. K. K.; Xiong, H. Y.; Lee, L. J.; Frey, B. J. Deep D.; Borisov, N.; Malakhova, G.; Vasilov, R.; Roumiantsev, S.;
Learning of the Tissue-Regulated Splicing Code. Bioinformatics 2014, Zhavoronkov, A.; Buzdin, A. Combinatorial High-Throughput
30 (12), i121−i129. Experimental and Bioinformatic Approach Identifies Molecular
(12) Lyons, J.; Dehzangi, A.; Heffernan, R.; Sharma, A.; Paliwal, K.; Pathways Linked with the Sensitivity to Anticancer Target Drugs.
Sattar, A.; Zhou, Y.; Yang, Y. Predicting Backbone Cα Angles and Oncotarget 2015, 6 (29), 27227−27238.
Dihedrals from Protein Sequences by Stacked Sparse Auto-Encoder (30) Makarev, E.; Cantor, C.; Zhavoronkov, A.; Buzdin, A.; Aliper,
Deep Neural Network. J. Comput. Chem. 2014, 35 (28), 2040−2046. A.; Csoka, A. B. Pathway Activation Profiling Reveals New Insights

2529 DOI: 10.1021/acs.molpharmaceut.6b00248

Mol. Pharmaceutics 2016, 13, 2524−2530
Molecular Pharmaceutics Article

into Age-Related Macular Degeneration and Provides Avenues for

Therapeutic Interventions. Aging 2014, 6 (12), 1064−1075.
(31) Combrisson, E.; Jerbi, K. Exceeding Chance Level by Chance:
The Caveat of Theoretical Chance Levels in Brain Signal Classification
and Statistical Assessment of Decoding Accuracy. J. Neurosci. Methods
2015, 250, 126−136.
(32) Langmead, C. J.; Watson, J.; Reavill, C. Muscarinic Acetylcho-
line Receptors as CNS Drug Targets. Pharmacol. Ther. 2008, 117 (2),
232−243.
(33) Volpicelli, L. A.; Levey, A. I. Muscarinic Acetylcholine Receptor
Subtypes in Cerebral Cortex and Hippocampus. Prog. Brain Res. 2004,
145, 59−66.
(34) Trimmer, J. S. Subcellular Localization of K+ Channels in
Mammalian Brain Neurons: Remarkable Precision in the Midst of
Extraordinary Complexity. Neuron 2015, 85 (2), 238−256.
(35) Gray, K. A.; Yates, B.; Seal, R. L.; Wright, M. W.; Bruford, E. A.
Genenames.org: The HGNC Resources in 2015. Nucleic Acids Res.
2015, 43 (D1), D1079−D1085.
(36) Boser, B. E.; Guyon, I. M.; Vapnik, V. N. A Training Algorithm
for Optimal Margin Classifiers. Proc. 5th Annu. Workshop Comput.
Learn. Theory - COLT ’92 1992, 144.
(37) Cortes, C.; Vapnik, V. Support-Vector Networks. Mach. Learn.
1995, 20 (3), 273−297.
(38) Hinton, G. E.; Salakhutdinov, R. R. Reducing the Dimension-
ality of Data with Neural Networks. Science 2006, 313 (5786), 504−
507.
(39) Zeiler, M. D. ADADELTA: An Adaptive Learning Rate Method
arXiv 2012, 6.
(40) Glorot, X.; Bengio, Y. Understanding the difficulty of training
deep feedforward neural networks. Int. Conf. Artif. Intell. Stat. 2010,
249−256.

2530 DOI: 10.1021/acs.molpharmaceut.6b00248

Mol. Pharmaceutics 2016, 13, 2524−2530