Professional Documents
Culture Documents
Goh Et Al-2017-Journal of Computational Chemistry
Goh Et Al-2017-Journal of Computational Chemistry
ORG REVIEW
The rise and fall of artificial neural networks is well docu- applicability to a wide range of challenges in the field, includ-
mented in the scientific literature of both computer science ing quantitative structure activity relationship, virtual screen-
and computational chemistry. Yet almost two decades later, ing, protein structure prediction, quantum chemistry, materials
we are now seeing a resurgence of interest in deep learning, a design, and property prediction. In reviewing the performance
machine learning algorithm based on multilayer neural net- of deep neural networks, we observed a consistent outper-
works. Within the last few years, we have seen the transforma- formance against non-neural networks state-of-the-art models
tive impact of deep learning in many domains, particularly in across disparate research topics, and deep neural network-
speech recognition and computer vision, to the extent that based models often exceeded the “glass ceiling” expectations
the majority of expert practitioners in those field are now reg- of their respective tasks. Coupled with the maturity of GPU-
ularly eschewing prior established models in favor of deep accelerated computing for training deep neural networks and
learning models. In this review, we provide an introductory the exponential growth of chemical data on which to train
overview into the theory of deep neural networks and their these networks on, we anticipate that deep learning algo-
unique properties that distinguish them from traditional rithms will be a valuable tool for computational chemistry.
C 2017 Wiley Periodicals, Inc.
V
machine learning algorithms used in cheminformatics. By pro-
viding an overview of the variety of emerging applications of
deep neural networks, we highlight its ubiquity and broad DOI: 10.1002/jcc.24764
Figure 1. a) Schematic representation of a traditional feedforward artificial neural network (ANN) with one hidden layer. Each neuron denoted as circles
accepts a series of n input values and maps it to an output using a nonlinear function, with a bias term (i.e., output of the neural network when it has
zero input) applied to all neurons in the hidden layer. b) Deep neural network (DNN) differ from ANN by having multiple (n > 3) hidden layers as depicted
in the schematic diagram, the bias term is omitted here for simplicity.
many terms if the nonlinear transformations are chosen poorly. exclusively. In the first half of this review, we will provide a
For example, a simple power series expansion will require an brief nontechnical introduction to deep learning, starting with
extremely large amount of terms (and parameters) to fit func- a basic background on artificial neural networks and highlight-
tions with large amounts of oscillations. ing the key technical developments in the last decade that
The mapping of features into an output using the function enabled deep neural networks. In addition, we will focus on
provided is the task of a processing unit, and deep learning how deep learning differs from traditional machine learning
algorithms are constructed from a collection of these process- algorithms that are used in computational chemistry, and how
ing units, arranged in a layered and hierarchical fashion. There- the ongoing resurgence of deep learning differs from artificial
fore, unlike simpler machine learning algorithms, it maps neural network models in the 1980s, which may be regarded
features through a series of nonlinear functions that are as its “parent” algorithm. In the next half of the review, we will
stitched together in a combinatorial fashion to optimally maxi- include a survey of recent developments of deep learning
mize the accuracy of the model. As a result of this complex applications across the field of computational chemistry, where
architecture, a deep learning algorithm learns multiple levels we will examine its performance against existing machine
of representations that correspond to different levels of learning models, and future prospects for contributing to the
abstraction, forming a hierarchy of concepts. By constructing field. This review was written primarily to serve as an introduc-
these so-called “hierarchical representations,” deep learning tory entry point for computational chemists who wish to
has an internal state that may be transferred to new problems, explore or integrate deep learning models in their research
partially overcoming the template matching problem. Finally, from an applications standpoint, and additional references to
information can take many paths through the network, and, as existing literature reviews will be provided to cover the more
a result, expressive power grows exponentially with depth.[10] technical aspects of deep learning neural network architecture
Ultimately, these unique characteristics of deep learning ena- and optimization.
bles it to utilize raw data directly as opposed to engineered
features, and often the resulting models constructed produce
Deep Learning 101
a comparable level of predictive accuracy. Deep learning
achieves this ability because within the multiple layers of the Artificial neural networks (ANNs), on which most deep learning
nonlinear functions, the algorithm transforms the raw data algorithms are based on, are a class of machine learning algo-
and maps it to intermediate “output” that serve as input (fea- rithm inspired by biological neural networks, used to estimate
tures) for the latter layers in the algorithm, in the process or approximate functions by translating a large number of
gradually transforming raw data into learned features. In short, inputs into a target output (Fig. 1a). ANNs are constructed
deep learning algorithms are potentially capable of automati- from a series of layers, and each layer comprises many
cally (i.e., without expert intervention) engineering the neces- “neurons.” Each neuron accepts an input value from the previ-
sary features that are relevant to optimally predict the output ous layer, and maps it onto a nonlinear function. The output
of interest. of this function is used as the input for the next layer in the
The majority of deep learning algorithms currently devel- ANN, until it reaches the last layer, where the output corre-
oped are based off artificial neural networks, and for the pur- sponds to the objective that is to be predicted. In addition, a
pose of this review we will focus on deep neural networks tunable parameter, the “weight” (or coefficient) of each
Figure 2. Plot of a) sigmoidal and b) rectified-linear (ReLU) function (in blue) and their corresponding first derivative (in red). Unlike the sigmoidal function,
where its derivative varies with respect to the value of x, in the ReLU function, the first derivative is either 0 or 1. [Color figure can be viewed at wileyonli-
nelibrary.com]
neuron’s function is adjusted in the construction of this model deeper into the model, an increasing number of neurons and
to minimize the error of the predicted value, a process known weights are associated with a given error. Until recently, this
as “training” the neural network. Figuratively, the collection of made it difficult to train many layers efficiently; anything more
these neurons in ANNs mimics the way neurons work in bio- than a few layers required a long time to converge with a
logical systems, hence its name, artificial neural network. high probability of overfitting, especially for the layers closest
The power of ANNs, as hinted earlier, lies in their ability to to the output. In addition, the nonlinear transformation func-
make multiple nonlinear transformations through many hidden tions, such as sigmoids, had finite dynamic range, so error sig-
layers of neurons, where the “hidden” term refers to layers nals tends to decay as they passed through multiple layers,
that are not directly adjacent to the input or output. In this which is more commonly known as the “vanishing gradient
process, increasingly complex and abstract features can be problem.”[13]
constructed, through the addition of more layers and/or Since 1986, several key algorithms, including unsupervised
increasing the width of layers (i.e., increasing the number of pretraining,[14] rectified linear functions,[15] and dropout,[16]
neurons per layer). Correspondingly, the model can learn have been developed to improve the training process for
increasingly complex and abstract representations (i.e., ANNs, to address the vanishing gradient problem, and to
“concepts” if the term is used loosely). However, for one to reduce overfitting which ANNs are particularly susceptible to.
use more than a single hidden layer, it is necessary to deter- Perhaps the largest impediment to training deep neural net-
mine how to assign error attribution and make corrections to works (DNN), was the vanishing gradient problem as it practi-
its weights by working backwards originating from the pre- cally capped the depth of the neural network. Pretraining,
dicted output, and back through the neural network. This discovered by Hinton et al. in 2006 is a fast, greedy, algorithm
backwards propagation of errors is known formally as that uses an unsupervised layer-wise approach to train a DNN
“backpropagation.” Although the conceptual foundation of one layer at a time.[14] After the pretraining phase is complete,
backpropagation was discovered in 1963,[11] it was not until a more subtle fine-tuning process, such as stochastic gradient
1986 that Hinton and coworkers discovered a way for this descent, is used to train the model. Using the pretraining
algorithm to be applied to neural networks,[12] which was a approach, the model would have already learnt the features
key historical development that enabled practically usable before backpropagation begins, mitigating the vanishing gra-
ANNs. dient problem. An alternative solution emerged in 2011, where
During the process of backpropagation, an algorithm known Bengio and coworkers demonstrated that the rectified linear
as gradient descent is used to find the minimum in the error activation (ReLU) function sidesteps the vanishing gradient
surface caused by each respective neuron when generating a problem entirely. The ReLU’s first derivative is precisely unity
corresponding output. Conceptually, gradient descent is no or 0, generally ensuring that error signals can back-propagate
different from the steepest descent algorithm used in classical without vanishing or exploding. (Fig. 2).[15]
molecular dynamics simulation. The major difference is instead As these methods enabled the training of deeper and more
of iteratively minimizing an energy function and updating complex neural network architecture, overfitting also became
atomic coordinates for each step, an error function of the tar- more of an issue, which led to the development of the drop-
get output of the ANN is iteratively minimized and the out algorithm. In dropout, for each epoch of the training pro-
weights of the neurons are updated each step, which are also cess, a fixed proportion of neurons are randomly selected to
known as “iteration” in the ANN literature. The data in the be temporarily excluded from the model. The net effect of
training set may be iterated over multiple times, with a com- dropout is that it simulates many different architectures during
plete pass over the data being called an “epoch.” training, which prevents codependency among neurons and
A key issue with backpropagation is that the error signals reduces overfitting.[16] While the architecture of modern DNNs
become progressively more diffused as the signal goes back vary widely, a popular configuration is ReLU-based neural net-
through each hidden layer. This is because, as the signal goes works. When coupled with dropout and early stopping, such
Figure 3. a) Schematic diagram of a convolutional neural network (CNN). CNNs are designed with the explicit assumption that the input is in the form of
image data. Each convolutional layer extracts and preserves the spatial information and learns a representation which is then typically passed onto a tradi-
tional fully connected feedforward neural network before the output layer. b) Schematic diagram of a recurrent neural network (RNN). RNNs at its simplest
implementation are a modification of the standard feedforward neural network where each neuron in the hidden layer receives an additional input from
the output from the previous iteration of the model, denoted as “t 2 1” circular arrows. c) Schematic diagram of an autoencoder, which are neural net-
works used in unsupervised learning. In autoencoders, the objective is to learn the identity function of the input layer, and in the process, a compressed
representation of the original data in the hidden layers is learned.
ReLU networks have often been enough to regularize the Williams.[12] Deeper neural networks beyond a few hidden
model (i.e., prevent overfitting).[17] layers (Fig. 1b) was only achievable with more recent algorith-
Having provided a summary of the key developments in mic developments in the last few years.[14–16] Therefore, how
ANNs and its associated algorithms, we note that it is by no is DNNs not just the mere rebranding of ANNs of the last cen-
means comprehensive. In addition to the traditional feedfor- tury, and how is it better than the traditional machine learning
ward DNN (Fig. 1b) that has been discussed thus far, more algorithms that are already successfully used in various chem-
recent developments include alternative architectures, notably informatics applications?
convolutional neural networks (Fig. 3a),[18,19] recurrent neural Many decades of chemistry research has led to the develop-
networks (Fig. 3b),[19,20] and autoencoders (Fig. 3c) that have ment of several thousand molecular descriptors that describe
been highly successful in computer vision and natural lan- a range of properties of conceivably any compound. Molecular
guage processing applications. A technical discussion of vari- descriptors thus serve as features constructed using chemical
ous DNN architectures while informative to understanding the knowledge and intuition (i.e., domain expertise) that can be
deep learning literature, is beyond the scope of this review, used in traditional machine learning models, which have
therefore, we refer our readers to the following prior publica- achieved reasonable success in computational chemistry appli-
tions summarizing this research topic.[21–24] By now, it should cations.[26–31] Traditional machine learning algorithms such as
be evident that ANNs itself are not a new invention. Indeed, linear regression and decision trees are intuitive and create
the mathematical algorithm for ANNs was developed by simple models that humans can understand. Nevertheless, as
McCulloch and Pitts in 1943,[25] and practically trainable ANNs we progress to the prediction of more complex properties
dates as far back to 1986, coinciding with invention of back- with nonlinear relationship, typically those associated with bio-
propagation for neural networks by Rumelhart, Hinton, and logical processes and materials engineering, it is often
Figure 5. Growth of a) deep learning publications as indexed by ISI, and b) annual google trends score of major deep learning software packages since
2010. [Color figure can be viewed at wileyonlinelibrary.com]
from its spike in google search rankings (Fig. 5b), and the fact neural networks, which has significantly reduced the barrier of
that its Github has been starred and forked over 33,000 and entry for new deep learning practitioners.
14,000 times, respectively, despite it being only released for a Unquestionably, the computer science domain has been the
little over a year. In addition, APIs, such as Keras released in main benefactor of the surge of mineable data obtained from
2015, has greatly simplified the construction and training of the internet (Fig. 6a), and not surprisingly has also been the
Figure 6. The growth of a) global data generated, b) number of structures deposited in the Protein Data Bank, c) number of compounds deposited in Pub-
Chem, and d) GPU computing power for scientific computing (GPU computing power data points were obtained from the reported double-precision com-
puting power of NVIDIA Tesla Series GPUs for the flagship model released each year: C2070 (2010), M2090 (2011), K20 (2012), K40 (2013), K80 (2014), P100
(2015)), all share similar parallels in their upwards trajectory.
field where deep learning had the largest impact. In chemistry, in their case was the results from 19 assays. The dataset used
we have also seen a corresponding growth of data in publical- was curated from PubChem and included over 100,000 data
ly accessible databases, such as the Protein Data Bank (Fig. 6b) points. Molecular descriptors totaling 3764 descriptors per
and PubChem (Fig. 6c), with more data being generated from molecule were generated using Dragon,[63] and they were
recent developments in high-throughput omics technolo- used as the input features for the DNN. In an accuracy perfor-
gies.[53] It is for these reasons that we are optimistic that the field mance benchmark against other traditional machine learning
of computational chemistry is starting to experience the same algorithms, such as gradient-boosted decision trees and logis-
confluence of events, and this will greatly facilitate deep learning tic regression, the DNN-based model outperformed all others
applications in our field. We can leverage on the algorithmic in 14 of 19 assay predictions by a statistically significant mar-
breakthroughs in the computer science domain, the increasing gin and was comparable in terms of performance in the
availability of chemical data, and the now matured GPU- remaining 5 assay prediction.[62] In addition, the advantages of
accelerated computing technologies [Fig. 6d; GPU computing a multitask neural network was noted by the authors, particu-
power data points were obtained from the reported double- larly in the fact that it develops a shared, learned feature
precision computing power of NVIDIA Tesla Series GPUs for the extraction pipeline for multiple tasks. This means that not only
flagship model released each year: C2070 (2010), M2090 (2011), can learning more general features produce better models,
K20 (2012), K40 (2013), K80 (2014), P100 (2015)]. but weights in multitask DNNs are also constrained by more
data cases, sharing statistical strength.[62] Lastly, an interesting
Computer-Aided Drug Design observation from that study was how DNNs were able to han-
dle thousands of correlated input features, which goes against
In computer-aided drug design, traditional machine learning traditional QSAR wisdom as highlighted by Winkler in 2002,[64]
algorithms have a long history in the field of cheminformatics, although we note that the observations published by Winkler
notably in their contribution to quantitative structure activity at that time was prior to the development of DNNs. In Dahl’s
relationship (QSAR) applications. In QSAR, the output to be work, the authors observed that halving the input features did
predicted is usually the biological activity of a compound. Usu- not led to any performance degradation.
ally regression models are used, and the input data are molec- A subsequent study in 2015 published by Merck, compre-
ular descriptors, which are precomputed physicochemical hensively analyzed the training of DNNs and compared their
properties of the molecule, designed from chemistry domain performance to the current state of the art used in the field,
knowledge. Early work in QSAR applications used linear regres- RF-based models, on an expanded Merck challenge dataset.[65]
sion models, but these were quickly supplanted by Bayesian The authors concluded that DNNs could be adopted as a prac-
neural networks,[54–56] followed by RFs[26] and SVMs.[31] Practi- tical QSAR method, and easily outperformed RF models in
tioners in the field have historically favored models that allow most cases. In terms of practical adoption, the authors empha-
for variable selection so that an informed chemist can deter- sized the dramatic advance in GPU hardware that DNNs lever-
mine if selected features made sense. In addition, models that age and also the economic cost advantages of deploying GPU
allowed assessment of uncertainty of output predictions were resources as opposed to conventional CPU clusters that are
also preferred. The field of QSAR is vast, and we refer readers used by traditional machine learning models.[65] The key issue
to the following list of reviews for key historical technical associated with training deep neural networks, particularly in
developments.[57–60] For the purpose of this review, we will the number of tunable parameters was also investigated. The
limit the scope of discussion to the performance of DNN- authors discovered that most single task problems could be
based QSAR models and appropriate comparisons to tradition- run on architectures with two hidden layers, using only 500–
al machine learning models. 1000 neurons per layer and 75 training epochs. More complex
The first foray of deep learning into QSAR was the Merck architecture and/or longer training time yielded incremental
challenge in 2012.[61] In this publically available challenge, but diminishing returns in model accuracy.[65] Despite the
teams were provided precomputed molecular descriptors for overall promising performance of DNNs in the Merck challenge
compounds and their corresponding experimentally measured and associated studies as summarized above, the results were
activity for a total of 15 drug targets. Submitted models were received with skepticism by some practitioners in the research
evaluated on their ability to predict activity against a test set community.[66] Common concerns include the small sample
not released to participants. The winning group used DNN size, and that the incremental improvements in predictive
models, led by Dahl who was part of Hinton’s research accuracy was difficult to justify in the face of increase in model
team.[61] Notably, it should be emphasized that the team had complexity.
no formally trained computational chemist in the group; they In 2014, Hochreiter and coworkers published a peer-
were from the computer science department. reviewed paper at the Neural Information Processing Systems
In 2014, Dahl et al., submitted an arxiv paper exploring the (NIPS) conference on the application of multitask DNNs for
effectiveness of multitask neural networks for QSAR applica- QSAR application on a significantly larger dataset.[67] In this
tions, based on the algorithms used in the Merck challenge.[62] study, the authors curated the entire ChEMBL database, which
In this work, the authors used a multitask DNN model. Here, was almost two orders of magnitude larger than the original
“multitask” refers to a model that predicts not just a single Merck challenge dataset. This dataset included 743,336 com-
output of interest, but multiple outputs simultaneously, which pounds, approximately 13 million chemical features, and 5069
by conventional means, making toxicological studies of DILI molecules toward soft nucleophiles such as gluthaione
difficult. A recent study from Xu et al. used DNNs to predict (GSH).[83] By training only on qualitative reactivity data, they
DILI toxicity.[75] The authors used both explicit molecular were able to construct a DNN-based model that identified
descriptors, computed from Mold[76] and PaDEL,[77] as well as sites of reactivity within reactive molecules with 90.8% accura-
the URGNN method for molecular structuring encoding devel- cy, and separate reactive and unreactive molecules with 80.6%
oped by Lusci et al.[78] as input data for the DNNs. The model accuracy.[83] In addition, the model’s predictions correlated
was trained on 475 drugs, with an external test set of 198 well with quantitative GSH reactivity measurements in external
drugs, and the best model that utilized a DNN had accuracy of data sets that were more chemically diverse, indicating the
86.9%, sensitivity of 82.5%, specificity of 92.9%, and AUC of model’s generalizability across a larger area of chemical
0.955.[75] In comparison, traditional models have lower abso- space.[83] A subsequent publication expanded the scope of the
lute performance metrics by 10–20%.[75] Interestingly, using model to encompass reactivity toward GSH, cyanide, protein,
input from the URGNN molecular structural encoding method, and DNA. The resulting model yielded a cross-validated AUC
the authors created a model with the highest performance performance of 89.8% for DNA and 94.4% for protein, and sep-
(AUC 0.955), outperforming similarly trained DNNs that used arated electrophilically reactive molecules with DNA and pro-
calculated molecular descriptors from Mold (AUC 0.931) and tein from nonreactive molecules with a cross-validated AUC
PaDEL (AUC 0.895).[75] This suggests that a good molecular performance of 78.7% and 79.8%, respectively.[84] Furthermore,
encoding method such as UGRNN may be more effective in the model’s performance also significantly outperformed reac-
providing the necessary features to DNNs, as deep learning tivity indices calculated from QM methods.[84] As drug toxicity
algorithms have the ability to automatically extract the neces- is often caused by electrophilic reactive metabolites, models
sary features, and this ability may be on par or perhaps even that assist in the study of identifying site reactivity, which has
better than domain-expert feature engineering through the been up to now conspicuously absent in the literature, can
development of explicit molecular descriptors. potentially be utilized to construct a mechanism-based predic-
Another application for DNN modeling toxicity was pub- tion of molecule toxicity.
lished by Swamidass and coworkers in 2015.[79] One common A larger scale study on chemical toxicity was also recently
mechanism of drug toxicity stems from electrophilic reactive published by the Hochreiter group in 2016.[85] In this work,
metabolites that covalently bind to proteins. Epoxides are a the authors reported on the application of DNN models on
functional group of this nature, which are often formed by the Tox21 Data Challenge released by NIH in 2014. The data-
cytochrome P450 metabolism of drug molecules, which acts base consisted of 12,000 environmental chemicals and drugs,
on aromatic or double bonds. Swamidass and coworkers and their corresponding measurements on 12 different assays
results were particularly distinctive, because they developed a designed to measure a variety of toxicity effects. Not surpris-
DNN model to predict the specific location on a molecule that ingly, the DeepTox model developed by Hochreiter and cow-
undergoes epoxidation, that is, its site of epoxidation (SOE). orkers had the highest performance of all methods submitted
This work was based off an earlier model, Xenosite, an ANN- to the Tox21 challenge.[9] Further analysis of their model indi-
based model for P450 metabolism on small-molecules, which cated that using a multitask DNN model led to consistent out-
despite being a shallow network, was already outperforming performance against single-task models in 10 out of 12 assay
the accuracy of SVM-based models by as much as 5%.[80] Fur- predictions.[85] Additional benchmarks to traditional machine
ther improvements were subsequently achieved by investigat- learning algorithms, including SVM, RF, and Elastic Net, also
ing the effect of using different types of molecular fingerprints demonstrated that DNN outperformed in 10 out of 15
for modeling P450 metabolism, where they discovered that cases.[85] Lastly, while the original DeepTox model used molec-
further accuracy gains can be achieved using a consensus ular descriptors provided by NIH in the Tox21 challenge, the
model utilizing different fingerprint types,[81] and a related sis- authors also showed that a similarly trained DNN model devel-
ter model that predicted the site of glucoronidation metabo- oped using only ECFP4 fingerprint as input data had similar
lism.[82] In their more recent work on predicting epoxide-based performance to those trained on explicit molecular descriptors,
toxicity, Swamidass and coworkers designed a 4-layer DNN which is similar to the observations made by Xu et al. in their
architecture, and trained the model on a database of 702 DILI toxicity model.[75] Interestingly, on visualization of the first
epoxidation reactions, and identified SOEs with 94.9% AUC hidden layer of these DNNs, the author observed that 99% of
performance, and separated (i.e., classified) epoxidized and neurons in that layer had a significant association with at least
non-epoxidized molecules with 79.3% AUC.[79] Moreover, with- one known toxicophore feature, suggesting that deep learning
in epoxidized molecules, the model was able to provide can possibly support the discovery of new chemical knowl-
atomic-level precise information, by separating aromatic or edge in its hidden layers.[85]
double bond SOEs from all other aromatic or double bonds In line with the progress in QSAR and toxicity prediction,
with AUCs of 92.5% and 95.1%, respectively.[79] This makes the deep learning algorithms have also started to make an impact
DNN model the first mechanistic model in the literature, which in other aspects of computer-aided drug design. In 2013, Baldi
not only predicts the formation of reactive epoxides of drug and coworkers reported using a DNN model to predict mole-
candidates but also accurately identifies the specific epoxi- cule solubility.[78] More recent research developments in this
dized bonds in the molecule. Using a similar DNN model, Swa- direction was also submitted to arxiv by Pande and coworkers,
midass and coworkers modeled the site of reactivity of small- where they developed a multitask DNN model for predicting
not just solubility, but the entire spectrum of ADMET proper- Lastly, a DNN architecture was used to tune the prediction of
ties.[86] Deep learning may also have a future in virtual screen- all the residue–residue contact probabilities by integrating spa-
ing as a viable alternative or complement to existing docking tial and temporal information.[105] CMAPpro was trained on a
methods. In 2016, an arxiv paper was submitted by AtomNet, 2356-member training set derived from the ASTRAL data-
a startup that developed a DNN model to classify the activity base.[106] For cross-validation purposes, the set was segmented
of small molecules docked in protein binding pockets.[87] into 10 disjoint groups belonging to different SCOP fold, which
Remarkably, the AtomNet DNN model was able to achieve meant that neither training nor validation set shared sequence
AUC metrics ranging between 0.7 to 0.9 depending on the or structural similarity. The resulting model performance was
test set used, which significantly outperforms conventional then tested against a 364-member test set of new protein
docking methods, specifically Smina,[88] a fork of AutoDock folds reported between version 1.73 and 1.75 release of the
Vina[89] by 0.1 to 0.2.[87] For additional recent developments of ASTRAL database. CMAPpro performance was compared
deep learning in applications that are more closely aligned to against several permutations of the multistage machine learn-
computational biology, we refer our readers to the following ing model, including a single hidden layer neural network
reviews that focuses on that research topic.[90] (NN), a single hidden layer neural network that utilized the
coarse contact/orientation and alignment predictors, which is
Computational Structural Biology generated by the 2D recursive neural network and the energy-
based neural network (NN 1 CA), and a deep neural network
Predicting the spatial proximity of any two residues of a pro- but without CA features (DNN). Based on the relative perfor-
tein sequence when it is folded in its 3D structure is known as mance, both the deep network architecture and CA features
protein contact prediction. The prediction of contacts between were required to achieve an accuracy of 36%; DNN and
sequentially distinct residues thus imposes strong constraints NN 1 CA each achieved 32%, while NN which represents the
on its 3D structure, making it particularly useful for ab initio previous state-of-the-art only achieved 26% accuracy.[106]
protein structure prediction or engineering. While the use of A different implementation of DNN for protein contact pre-
physics-based simulation methods, such as long-timescale diction was also reported by Eickholt and Cheng in 2012.[107]
molecular dynamics[91,92] can be used for ab initio protein In their algorithm, DNCON, it combined deep learning with
structure prediction, the computational demands are formida- boosting techniques that was used to develop an ensemble
ble. Complementary methods, such as knowledge-based physi- predictor. A 1426-member dataset derived from the Protein
cal approaches developed by the groups of Wolynes, Onuchic Data Bank was used to train DNCON, with a random split
and others are also an option,[93,94] although their computa- between the training (1230-member) and validation (196-
tional expense while lower is still sufficiently demanding that member) set. Explicitly engineered features were used as input
it cannot be used for large-scale studies. Therefore, machine for the DNN. Specifically, three classes of features were used:
learning approaches are viable alternatives, including those (i) those from two windows centered on the residue pair in
based off ANNs,[95–97] SVMs,[27] and hidden Markov model.[98] question (e.g., predicted secondary structure and solvent
Other approaches include template-based approaches that use accessibility, information and likelihoods from the PSSM, and
homology or threading methods to identify structurally similar Acthley factors, etc.), (ii) pairwise features, (e.g., Levitt’s contact
templates to base an inference of protein contact predic- potential, Jernigan’s pairwise potential, etc.), and (iii) global
tion.[99,100] The assessment of these various models for contact features (e.g., protein length, percentage of predicted exposed
predictors is one of the highlights of the critical assessment of alpha helix and beta sheet residues, etc.).[107] Using these engi-
protein structure prediction (CASP) challenge which started in neered features, the DNN model was tasked to predict wheth-
1996. Despite improvements over the years, the long-range er or not a particular residue pair was in contact. In addition,
contact prediction has historically hit a glass ceiling of just boosted ensembles of classifiers was created by training sever-
below 30% accuracy. The key historical developments of com- al different DNNs using a sample of 90,000 long-range resi-
putational protein structure prediction is voluminous, and we due–residue pairs from a larger pool obtained from the
refer interested readers to existing reviews on this top- training set. In evaluating its performance, cross-validated
ic.[101–104] For the purpose of this review, we will limit the accuracy of DNCON was 34.1%. The model’s performance
scope of discussion to the performance of recent DNN-based transferability was demonstrated in its performance bench-
models, and how they have been critical to breaching the his- marks against the two best predictors of CASP9,[108]
torical glass ceiling expectations in the field. ProC_S3,[28] and SVMcon,[27] which are based off RF and SVM
In 2012, Baldi and coworkers developed CMAPpro, a multi- algorithms, respectively. In that assessment, the respective test
stage machine learning approach, which improved contact set was used for each software. While the improvement was
prediction accuracy to 36%.[105] Three specific improvements not as dramatic as that reported by Baldi and coworkers,
were implemented in CMAPpro over earlier models. The first is DNCON performance was 3% better than the state-of-the-art
the use of a 2D recursive neural network to predict coarse algorithms for its time; ProC_S3 (32.6% vs. 29.7%) and SVMcon
contacts and orientations between secondary structure ele- (32.9% vs. 28.5%).[107]
ments. In addition, a novel energy-based neural network Both DNN-based protein contact prediction models were
approach was used to refine the prediction from the first net- noteworthy, as it enabled the community to breakthroughs
work and used to predict residue-residue contact probabilities. the 30% accuracy barrier that was not possible in prior years.
Apart from protein contact prediction, DNNs have also been public PBM, RNAcompete, ChIP-seq, and HT-SELEX experi-
successfully applied to the prediction of various protein angles, ments, the raw data was used as an input into a DNN algo-
dihedrals, and secondary structure from only sequence data. rithm to compute a predictive binding score. DeepBind’s
Using DNNs, Zhou, Yang, and coworkers published a series of ability to characterize DNA-binding protein specificity was
sequence-based predictions for Ca-based angles and tor- demonstrated on the PBM data from the revised DREAM5 TF-
sions.[109–111] Unlike protein contact prediction, backbone tor- DNA Motif Recognition Challenge by Weirauch et al.[120] Nota-
sions are arguably better restraints for use in ab initio protein bly, DeepBind outperformed all existing 26 algorithms based
structure prediction and other modeling purposes.[112] In the on Pearson correlations and AUC metrics, and was ranked first
development of these DNN-based models, Zhou, Yang, and amongst 15 teams in the DREAM5 submission.[118] Interesting-
coworkers used a 4590-member training set and a 1199 inde- ly, their results also indicated that models trained on in vitro
pendent test set obtained from the protein sequence culling data worked well at scoring in vivo data, suggesting that the
server PISCES.[113] Input data included specifically engineered DNNs has captured a subset of the properties of nucleic bind-
features obtained from the position specific scoring matrix ing itself.
generated by PSI-BLAST,[114,115] as well as several other physi- As with the repeated occurrence of deep learning outper-
cochemical properties related to residue identity, including ste- forming traditional machine learning algorithms in other
ric, hydrophobicity, volume, polarizability, isoelectric point, fields,[18,32–35] as well as in computer aided drug design
helix probability, amongst others.[116] itself,[62,67,69] the utilization of DNNs in pushing the “glass
In the development of the SPINE-X algorithm, a DNN was ceiling” boundaries of protein contact prediction and second-
used to predict secondary structure, residual solvent-accessible ary structure prediction should come as no surprise. Conspicu-
surface area (ASA), u and w torsions directly.[110] A six-step ously absent from this review is the application of deep
machine learning architecture was developed where outputs learning for RNA structure prediction and modeling, which to
such as ASA were used as subsequent inputs for other proper- the best of our knowledge has yet to be reported. Compared
ties to be predicted, such as the torsions. Based on the evalua- to the protein database, available structural data on RNA is
tion of the model’s performance on the independent test set, smaller. Furthermore, most RNA structural data are not crystal-
it achieved a mean absolute error of 228 and 338, respectively, lographic but are instead NMR-based, which itself is subjected
for the u and w dihedrals. Secondary structure prediction to a higher uncertainty by virtue of the fact that NMR-
accuracy on independent datasets were ranging from 81.3% structures themselves are approximations resolved using
to 82.3%, and this achievement is noteworthy, considering physics-based force field against experimentally bounded
that the field of secondary structure prediction from sequence restraints.[121] Nevertheless, it will be interesting to see how
data has stagnated just under 80% accuracy in the recent deep learning can benefit the RNA modeling community.
decade, some of which utilize traditional machine learning Lastly, an interesting contrast in the use of deep learning in
algorithms.[117] In a similar fashion, for the SPIDER algorithm computational structural biology applications compared to
that was developed later, a DNN was used to predict Ca computer-aided drug design is the exclusive use of engineered
angles (h) and torsions (s) directly.[109] Based on the evaluation features, and for some cases, the engineering of the architec-
of the model’s performance, it achieved a mean absolute error ture of the multistage machine learning algorithm itself. While
of 98 and 348 for h and s, respectively, and the authors the findings from the computer-aided drug design field is pre-
observed that the model’s error increased from helical residues liminary, there are some indications that explicitly engineered
to sheet residues to coil residues, following the trend in features do not necessarily perform better against chemical
unstructuredness. Using these predicted angles and torsions fingerprints, which arguably require less chemical domain
as restraints, the authors were able to model the 3D structure knowledge to construct. While we concede that proteins are
of the proteins with an average RMSD of 1.9A between the considerably more complex than small molecules, it would be
predicted and native structure.[109] The SPINE-X and SPIDER interesting to determine if the performance of DNN models
algorithm was subsequently retrained as a parallel multistep that uses input data that includes only basic structural and
algorithm that predicted simultaneously the following proper- connectivity information, without any specifically engineered
ties: secondary structure, ASA, u, w, h and s.[111] This resulted features, can accurately predict properties such as protein sec-
in a modest improvement in overall accuracy of secondary ondary structure, and long-range contacts.
structure by 2%, and reduction of MAE by 1–38 for the angles/
torsions, while maintaining the same level of ASA
Quantum Chemistry
performance.
Apart from protein structure modeling, deep learning has Using machine learning to supplement or replace traditional
also been utilized to predict other properties of interest based quantum mechanical (QM) calculations has been emerging in
on sequence data. For example, predicting sequence specific- the last few years. In this section, we will examine some
ities for DNA- and RNA-binding proteins was recently machine learning applications to quantum chemistry, and
reported.[118,119] In the seminal work by Frey and cow- examine the relative performance of similar DNN-based mod-
orkers,[118] the DeepBind algorithm was developed to predict els. In 2012, von Lilienfeld and coworkers developed a
the sequence specificities of DNA and RNA-binding proteins. machine learning algorithm based on nonlinear statistical
Using 12 terabases of sequence data, spanning thousands of regression to predict the atomization energies of organic
molecules.[29] This model used a 7000-member subset of the energies, but several other electronic ground and excited state
molecular generated database (GDB), a library of 109 stable properties.[123] In this work, they attempted to capitalize on
and synthetically tractable organic compounds. The target the advantages of multitask learning, by predicting several
data used for training was atomization energies of the 7000 electronic properties and potentially capturing correlations
compounds calculated using the PBE0 hybrid functional. No between seemingly unrelated properties and levels of theory.
explicit molecular descriptors were used as input data, instead The data was represented using the randomized variant of the
only the Cartesian coordinates and nuclear charge were used “Coulomb matrix.”[122] The target data was atomization ener-
in a “Coulomb” matrix representation. Arguably, without gies, static polarizabilities, frontier orbital eigenvalues HOMO
explicitly engineered features, this type of representation in and LUMO, ionization potential, and electron affinity calculated
the input data would be of the same level as that provided by using several different level of theory such as PBE0, ZINDO,
molecular fingerprints used in classical molecular modeling GW, and SCS. The atomization energy maintained a similar
approaches. Using only 1000 compounds for the training set, accuracy of MAE of 0.16 eV (3.6 kcal/mol) and achieved com-
von Lilienfeld and coworkers achieved a mean absolute error parable accuracy of MAE of 0.11 to 0.17eV (2.5 to 3.9 kcal/
(MAE) accuracy of 14.9 kcal/mol. Further tests on an external mol) for the other energy predictions, including HOMO, LUMO,
6000 compound validation set yielded similar accuracy of 15.3 ionization potential, and electron affinity.[123] Furthermore, this
kcal/mol, demonstrating the transferability of the model within level of accuracy was similar to the error of the corresponding
“in class” compounds. What was particularly groundbreaking level of theory used in QM calculations for constructing the
about this work was the ability to reasonably recapitulate QM- training set.
calculated energies, with a mean-absolute error of 15 kcal/ While using machine learning algorithms to replace QM cal-
mol, without having any implementation of the Schrodinger culations is enticing, an alternative more “first principles
Equation in the machine learning algorithm at all. More impor- grounded” approach is to use machine learning algorithms to
tantly, considering that this work used a traditional machine supplement existing QM algorithms. As first reported by von
learning algorithm that lacks the advantages of DNN, and Lilienfeld and coworkers in 2015, they demonstrated the D-
based on DNN’s historical performance, it suggests that a learning approach, whereby a machine learning “correction
DNN-based model should perform even better. term” was developed.[124] In that study, the authors used DFT
A subsequent publication by Hansen et al. investigated a calculated properties and were able to predict the correspond-
number of established machine learning algorithms, and the ing quantity at the G4MP2 level of theory using the D-learning
influence of molecular representation on the performance of correction term. This composite QM/ML approach combines
atomization energy predictions on the same dataset as used approximate but fast legacy QM approximations with modern
in von Lilienfeld work.[122] Amongst the key findings was that big-data-based QM estimates trained on expensive and accu-
using a randomized variant of the “Coulomb matrix” greatly rate results across chemical space.[124] However, we noted that
improved the accuracy of atomization energies to achieve as this approach has thus far been only demonstrated using tra-
low a MAE as 3.0 kcal/mol.[122] Apart from being an inverse ditional machine learning algorithms. If the performance boost
atom-distance matrix representation of the molecule, the ran- using multitask DNNs that we have observed on numerous
domized variant is unique and retains invariance with respect instances applies to this example, a DNN-based approach
to molecular translation and rotation. An added “side effect” would potentially yield superior results, but that has yet to be
of this improved representation was that it was the richest reported in the literature.
one developed, as it is both high-dimensional and accounting To the best of our knowledge, the fewer examples of DNN
for multiple indexing of atoms.[122] The authors discovered in quantum chemistry applications seem to indicate that it is
that sorting various representation by information did yield a in an earlier stage of development compared to computer-
correspondingly lower accuracy across all machine learning aided drug design and computational structural biology. From
algorithms tested,[122] which highlighted the importance of the literature, we know that traditional machine learning mod-
good data representation in QM applications. In fairness, it els have been used in other QM applications, such as model-
should also be noted that the authors did benchmark ANNs, ing electronic quantum transport,[125] learning parameters for
and while they performed satisfactorily with a MAE of 3.5 kcal/ accurate semiempirical quantum chemical calculations,[126] and
mol, it was not considerably better than nonlinear regression so forth. In addition, new representation and fingerprints for
methods of MAE of 3.0 kcal/mol. Nevertheless, we highlight QM applications are also being developed.[127,128] Given the
the neural network used was “shallow” with a few layers, and observed superior accuracy of DNN-based models against tra-
together with the lack of a larger dataset, does not represent ditional machine learning models in other fields of computa-
a true DNN implementation. One particularly illuminating con- tional chemistry, we suggest that the development of DNN-
jecture from this article is by extrapolating the performance based model for these examples of machine learning QM
(MAE error) with respect to the size of the dataset used, the applications would be beneficial for the field.
authors concluded that 3 kcal/mol was probably the “baseline”
error that one could achieve regardless of the machine learn-
Computational Material Design
ing algorithm used.[122]
In 2013, von Lilienfeld reported the application of the first The logical extension of DNN applications in the field of quan-
multitask DNN model that not only predicted atomization tum chemistry is to predict and design material properties
that are correlated to or based on QM-calculated properties. While DNN applications in material design is still at its infan-
Quantitative structure property relationship (QSPR), which is cy, it would be interesting to see how its application will fare
the analogous version of QSAR in the non-biological domain, against traditional QSPR applications and upcoming rational
is the science of predicting physical properties from more materials design endeavors, such as in the prediction of spec-
basic physiochemical characteristics of the compound, and it tral properties of fluorophores,[135,136] properties of ionic
has been extensively reviewed in prior publications.[129,130] liquids,[137] and nanostructure activity.[138]
Similar to the early years of modern drug development, mate-
rial discovery is primarily driven by serendipity and institution- Reservations about Deep Learning and of
al memory.[131] This has relegated the field to exploratory trial- Being a Black Box
and-error experimental approaches, and the key bottleneck in
molecular materials design is the experimental synthesis and Machine learning algorithms, while they may not be the first
tool of choice for many practitioners in our field, undeniably
characterization. In recent years, the paradigm of computation-
possess a rich history in the cheminformatics field and in
al and rational materials design has been encapsulated under
applications like QSAR and protein structure prediction. While
the materials genome initiative.[132,133] Due to the newness of
it may be argued that deep learning in some sense is a resur-
this field, in this section, we will examine a few key accom-
gence of the previous artificial neural network, the algorithmic
plishments of using machine learning for computational mate-
and technological breakthroughs in the last decade has
rial design, and highlighting deep learning applications where
enabled the development of staggeringly complex deep neural
available.
networks, allowing training of networks with hundreds of mil-
A recent high profile example of using machine learning
lions of weights. Coupled with the growth of data and GPU-
models to accelerate materials property research was pub-
accelerated scientific computing, deep learning has overturned
lished by Raccuglia et al. in 2016.[30] The synthesis of inorgan-
many applications in computer science domains, such as in
ic–organic hybrid materials, such as metal organic frameworks
speech recognition and computer vision. Given the similar par-
(MOFs), have been extensively studied for decades, but the
allels in the chemistry world, it suggests that deep learning
theoretical understanding of the formation of these com-
may be a valued tool to be added to the computational chem-
pounds are only partially understood. In the work by Raccuglia
istry toolbox. As summarized in Table 1, which presents key
et al., the authors used a SVM-based model to predict the
preliminary publications of DNN-based models, we have noted
reaction outcomes for the crystallization of templated vanadi-
the broad application of deep learning in many subfields of
um selenites. What was interesting about their work, was the
computational chemistry. In addition, the performance of
inclusion of “dark” reactions in training the model, which are
DNN-based model is almost always equivalent to existing
failed or unsuccessful reactions collected from archived labora- state-of-the-art non neural-network models, and at times pro-
tory notebooks. The resulting model had an 89% success rate, vided superior performance. Nevertheless, we have noticed
as defined by the synthesis of the target compound type. that the performance lift in many cases are not as significant,
Notably, this exceeded the human intuition success rate of if one is to make a comparison to the improvements DNN has
78%.[131] While a DNN-based model was not used in the study brought to its “parent” field of speech recognition and com-
per se, there is no technical reason why it could not be used puter vision. One mitigating factor that explains the lack of a
in place of SVM as a tool used for computational materials revolutionary advance in chemistry could be the relative scarci-
synthesis prediction. ty of data. Unlike the computer science domain where data is
One example of how DNN has been used to accelerate cheap, especially when obtained from the internet or social
materials discovery was reported by Aspuru-Gizik and cow- media, the quantity of usable data in chemistry is understand-
orkers in 2015.[134] Here, the authors used the dataset ably smaller and more expensive since actual experiments or
obtained from the Harvard Clean Energy Project—a high- computations are needed to generate useful data. In addition,
throughput virtual screening effort for the discovery of high- the field of chemistry has been around for centuries and given
performance organic photovoltaic materials. The metric to be the fact that chemical principles are based on the laws of
predicted is power conversion efficiency (PCE) which is a func- physics, it is not unconceivable that the development of fea-
tion of the HOMO an LUMO energies and several other empiri- tures such as molecular descriptors to explain compound solu-
cal parameters.[134] As no high quality 3D data was available bility, for example, would be an easier task than developing
to generate Coulomb matrices, the authors decided to use fin- features to explain the difference between a dog and a cat, a
gerprints based on molecular graphs as input representation. common task in computer vision. Therefore, with more accu-
Four different representations were tested and the results rate and better engineered features in chemistry, it is also
showed generally consistent accuracy (within the same order plausible that we might not see such a large initial perfor-
of magnitude) across HOMO, LUMO, and PCE predictions. The mance improvement, especially for the relatively simpler
dataset consisted of 200,000 compounds randomly selected chemical principles or concepts.
from the CEPDB database and another 50,000 was extracted Furthermore, as computational chemists, there is a greater
as the test set. Testing errors of HOMO and LUMO was 0.15 emphasis placed on conceptual understanding compared to
and 0.12eV, respectively, which is comparable to the DNN engineers or technologists which is arguably the more preva-
model developed previously by von Lilienfeld and coworkers. lent mindset in the computer science field. In this regard,
Table 1. Meta-analysis of DNN-based model performance relative to state-of-the-art non-DNN models in various computational chemistry applications.
Prediction/ Non-DNN
competition DNN models Comments models Comments
Merck Kaggle 0.494 R2 DNN-based model was the top 0.488 R2 Best non-DNN model in the competition.[139]
challenge performing model in the
(activity) competition.[61]
0.465 R2 Median DNN-based model 0.423 R2 Best non-DNN model (RF-based) by Merck
recreated by Merck post-competition.[65]
post-competition.[65]
Activity 0.830 AUC MT-DNN-based model trained 0.816 AUC Best non-DNN model (SVM) trained on the
on the ChEMBL database.[67] ChEMBL database.[66]
0.873 AUC MT-DNN-based model trained 0.800 AUC Best non-DNN model (RF)-based model
on the PCBA database.[69] trained on the PCBA database.[69]
0.841 AUC MT-DNN-based model trained 0.774 AUC Best non-DNN model (RF)-based model
on the MUV database.[69] trained on the MUV database.[69]
NIH Tox21 0.846 AUC DeepTox (MT-DNN-based model) 0.824 AUC Best non-DNN model (multitree ensemble
challenge was the top performing model.[85] model) was placed third in the Tox21
(toxicity) challenge.[140]
0.838 AUC Runner up in Tox21 challenge was
based off associative neural
networks (ASNN).[141]
0.818 AUC Post-competition MT-DNN model.[69] 0.790 AUC Post-competition RF model.[69]
Atom-level 0.949 AUC DNN-based model that predicts site – No comparable model in the literature
reactivity/ of epoxidation, a proxy for toxicity.[79] that can identify site of reactivity
toxicity or toxicity.
0.898 AUC DNN-based model that predicts site
of reactivity to DNA.[83]
0.944 AUC DNN-based model that predicts site
of reactivity to protein.[83]
Protein 36.0% acc. CMAPpro (DNN-based model).[105]
contact
34.1% acc. DNCON (DNN-based model).[107] 29.7% acc. Best non-DNN model reported in CASP9,
28.5% acc. ProC_S3 (RF-based model)[28] and
SVMcon (SVM-based model)[27]
are listed, respectively.
Only appropriate comparisons are summarized; models trained on similar/identical datasets, using either information extracted from publications by
the same group that reported multiple ML models or publically available competition.
deep learning algorithms currently fall short on two accounts. modeling makes), more recent reparameterization have started
First, it lacks the conceptual elegance of a first principles mod- modifying dihedral parameters to empirically fit experimental
el that is based on the actual laws of physics, and second, NMR distribution, even though that may lead to deviations
DNNs are essentially a black box; it is difficult to understand from the QM-calculated values.[142,143] Similarly, the choice of
what the neural network has “learned” or exactly how it is pre- columbic interactions to model electrostatic forces is only
dicting the property of interest. approximately correct, and recent parameter development of
To address the first issue of conceptual elegance, from a cer- modeling charged ion interactions have started fitting to vari-
tain perspective, this objection may be more of a philosophical ous experimental observables such as osmotic pressure values,
argument of scientific preferences. In most computational and the introduction of nonphysical correction terms when
chemistry applications, unless one is solving the Schrodinger modeling specific pairs of electrostatic interactions.[144–146] In
Equation exactly, which we know is impossible for anything these examples, approximation from first principles have to be
but a two body system, one must make approximations to the made, and this process is a human decision that is based on
model. In that sense, almost all of computational chemistry is empirical data or at times “chemical intuition”—which as Rac-
an empirically determined, and at times even intuitively deter- cuglia et al. have shown, is not infallible and not always more
mined, approximation of the “true” first principles Schrodinger accurate.[131] At the risk of oversimplification of the work that
Equation. To illustrate this point, let us examine the historical computational chemist do, the development of existing com-
development of classical molecular modeling force fields, such putational chemistry models may be viewed as an elaborate
as CHARMM[42] and AMBER.[43] For example, the parameteriza- curve fitting exercise. Instead of using human expert knowl-
tion of dihedral angle force constants have historically been edge, a conceivable alternative may be to use a deep learning
targeted to QM-calculated values, the “true” values grounded algorithm to “suggest,” or perhaps even help us “decide” what
in validated physical principles. However, because the dynam- approximations should be made to achieve the desired results,
ics of real molecules do not behave in an additive fashion in a move toward a future paradigm of a DNN-based artificial
(which itself is another approximation that classical molecular intelligence (AI) assisted chemistry research. This naturally
leads to the second drawback of deep learning as the inevita- increases, we have also seen deep learning progressing from
ble question surfaces—How do we know that the deep learn- frequently outperforming to always outperforming traditional
ing model is learning the correct physics or chemistry? machine learning models. In addition, some preliminary find-
We will concede that in its current implementation deep ings indicate that explicitly engineered features such as molec-
learning algorithms is still a black box and interrogating what ular descriptors may not be necessary to construct a high
it “learns” is an extremely challenging task. Nevertheless, black performing DNN model, and simpler representations in the
box algorithms such as SVM and RF are also used in several form of molecular fingerprint or coulomb matrices may suffice.
computational chemistry applications, notably in examples This is because of DNN’s ability to extract out its own features
where they are used primarily as a tool, and/or for prediction through its hidden layers. There is even indication that the fea-
of properties that are so complex that even a first principles tures “learned” by DNNs correspond to actual chemical con-
understanding of the problem will not necessarily aid in its cepts such as toxicophores. Coupled with recent research on
prediction. We acknowledge that to advance deep learning to improving interpretability of neural networks, it suggest that
be more than just another tool in the chemist’s toolkit, and for the future role of DNN in computational chemistry may not
it to gain more widespread applicability and adoption for sci- just be only a high-performance prediction tool, but perhaps
entific research, it is evident that improvement in interpretabil- as a hypothesis generation device as well.
ity of DNN is of paramount interest. While interpretability of
neural networks has historically not been a strong research Acknowledgments
focus for practitioners in this field, it is noteworthy that several
recent developments on improving interpretability has been The authors thank Dr. Nathan Baker for critically reviewing the
reported.[147,148] Other viable options include the use of differ- manuscript and providing helpful comments and insights on
ent neural network-based machine learning models, such as machine learning for computational chemistry applications.
influence-relevance voters (IVR) that are designed for interpret-
Keywords: deep learning machine learning molecular mod-
ability. As demonstrated on a few computational chemistry
applications from the work of Baldi and coworkers,[149,150] the eling quantitative structure activity relationship quantum
IRV is a low-parameter neural network which refines a k-near- chemistry materials genome toxicology protein structure
est neighbor classifier by nonlinearly combining the influences prediction cheminformatics artificial intelligence
of a chemical’s neighbors in the training set. IRV influences are
decomposed, also nonlinearly, into a relevance component
and a vote component. Therefore, the predictions of the IRV is How to cite this article: G. B. Goh, N. O. Hodas, A. Vishnu. J.
by nature transparent, as the exact data used to make a pre- Comput. Chem. 2017, 38, 1291–1307. DOI: 10.1002/jcc.24764
diction can be extracted from the network by examining each
prediction’s influences, making it closer to a “white-box” neural
network method.[149,150] [1] D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. van den
Driessche, J. Schrittwieser, I. Antonoglou, V. Panneershelvam, M.
Lanctot, S. Dieleman, D. Grewe, J. Nham, N. Kalchbrenner, I. Sutskever,
Conclusion T. Lillicrap, M. Leach, K. Kavukcuoglu, T. Graepel, D. Hassabis, Nature
2016, 529, 484
Unlike traditional machine learning algorithms currently used [2] M. Campbell, A. J. Hoane, F. H. Hsu, Artif. Intell. 2002, 134, 57
in computational chemistry, deep learning distinguishes itself [3] J. L. Reymond, M. Awale, ACS Chem. Neurosci. 2012, 3 ,649
in its use of a hierarchical cascade of nonlinear functions. This [4] R. S. Bohacek, C. McMartin, W. C. Guida, Med. Res. Rev. 1996, 16, 3
[5] H. D. Li, Y. Z. Liang, Q. S. Xu, Chemometr. Intell. Lab. 2009, 95, 188
allows it to learn representations and extract out the necessary [6] G. Hautier, C. C. Fischer, A. Jain, T. Mueller, G. Ceder,Chem. Mater.
features from raw unprocessed data needed to predict the 2010, 22, 3762.
desired physicochemical property of interest. It is this distin- [7] K. R. Muller, G. Ratsch, S. Sonnenburg, S. Mika, M. Grimm, N. Heinrich,
J. Chem. Inf. Model. 2005, 45, 249
guishing feature that has enabled deep learning to make sig-
[8] A. P. Bartok,M. J. Gillan, F. R. Manby, G. Csanyi, Phys. Rev. B 2013, 88,
nificant and transformative impact in its “parent” field of 054104.
speech recognition and computer vision. In computational [9] NIH, Available at: https://ncats.nih.gov/news/releases/2015/tox21-chal-
chemistry, its impact is more recent and more preliminary. lenge-2014-winners
[10] M. Raghu, B. Poole, J. Kleinberg, S. Ganguli, J. Sohl-Dickstein,
Nevertheless, based on the results from a number of recent
arXiv:1606.05336, 2016.
studies, we have noted the broad application of deep learning [11] A. E. Bryson, W. F. Denham, S. E. Dreyfus, AIAA J. 1963, 1, 2544
in many subfields of computational chemistry, including com- [12] D. E. Rumelhart, G. E. Hinton, R. J. Williams, Nature 1986, 323, 533
puter aided drug design, computational structural biology, [13] S. Hochreiter, Y. Bengio, P. Frasconi, J. Schmidhuber, Gradient Flow in
Recurrent Nets: the Difficulty of Learning Long-term Dependencies. In
quantum chemistry and materials design. In almost all applica-
A Field Guide to Dynamical Recurrent Neural Networks; S. C. Kremer,
tions we have examined, the performance of DNN-based mod- J. F. Kolen, Eds.; 2001.
el is frequently superior to traditional machine learning [14] G. E. Hinton, S. Osindero, Y. W. Teh, Neural Comput. 2006, 18, 1527
algorithms. [15] X. Glorot, A. Bordes, Y. Bengio, In Proc. 14th International Conference
on Artificial Intelligence and Statistics 315–323 (2011).
As the complexity of the problem increases that enables the
[16] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, R. Salakhutdinov, J.
application of multitask learning (i.e., more predictions of dif- Mach. Learn Res. 2014, 15, 1929.
ferent properties are required), and as the size of the dataset [17] Y. LeCun, Y. Bengio, G. Hinton, Nature 2015, 521, 436.
[18] A. Krizhevsky, I. Sutskever, G. E. Hinton, In Advances in Neural Informa- [49] R. Collobert, K. Kavukcuoglu, C. Farabet, In Proceedings of the NIPS
tion Processing Systems (NIPS), 2012, pp. 1106–1114. BigLearn Workshop, 2011.
[19] W. Xiong, J. Droppo, X. Huang, F. Seide, M. Seltzer, A. Stolcke, D. Yu, G. [50] R. Al-Rfou, et al. arXiv:1605.02688, 2016.
Zweig, arXiv:1610.05256, 2016. [51] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S.
[20] Y. Wu, et al. arXiv:1609.08144, 2016. Guadarrama, T. Darrell, arXiv:1408.5093, 2014.
[21] Y. Bengio, A. Courville, P. Vincent, IEEE Trans. Pattern Anal. Mach. Intell. [52] M. Abadi, et al. arXiv:1605.08695, 2016.
2013, 35, 1798. [53] W. W. Soon, M. Hariharan, M. P. Snyder, Mol. Syst. Biol. 2013, 9, 640.
[22] J. Schmidhuber, Neural Netw. 2015, 61, 85. [54] W. P. Ajay Walters, M. A. Murcko, J. Med. Chem. 1998, 41, 3314.
[23] I. Arel, D. C. Rose, T. P. Karnowski, IEEE Comput. Intell. M. 2010, 5, 13. [55] F. R. Burden, D. A. Winkler, J. Med. Chem. 1999, 42, 3183.
[24] E. Gawehn, J. A. Hiss, G. Schneider, Mol. Inf. 2016, 35, 3. [56] F. R. Burden, M. G. Ford, D. C. Whitley, D. A. Winkler, J. Chem. Inf. Com-
[25] W. S. Mcculloch, W. Pitts, B Math. Biol. 1990, 5,2, 99 put. Sci. 2000, 40, 1423.
[26] V. Svetnik, A. Liaw, C. Tong, J. C. Culberson, R. P. Sheridan, B. P. [57] M. Karelson, V. S. Lobanov, A. R. Katritzky, Chem. Rev. 1996, 96, 1027.
Feuston, J. Chem. Inf. Comput. Sci. 2003, 43, 1947. [58] P. Gramatica, QSAR Comb. Sci. 2007, 26, 694.
[27] J. Cheng, P. Baldi, BMC Bioinform. 2007, 8. 113. [59] J. Verma, V. M. Khedkar, E. C. Coutinho, Curr. Top. Med. Chem. 2010,
[28] Y. Q. Li, Y. P. Fang, J. W. Fang, Bioinformatics 2011, 27, 3379. 10, 95.
[29] M. Rupp, A. Tkatchenko, K. R. Muller, O. A. von Lilienfeld, Phys. Rev. [60] A. Tropsha, Mol. Inf. 2010, 29, 476.
Lett, 2012, 108, 058301. [61] Kaggle, Available at: http://blog.kaggle.com/2012/11/01/deep-learning-
[30] P. Raccuglia, K. C. Elbert, P. D. Adler, C. Falk, M. B. Wenny, A. Mollo, M. how-i-did-it-merck-1st-place-interview/
Zeller, S. A. Friedler, J. Schrier, A. J. Norquist, Nature 2016, 533, 73. [62] G. E. Dahl, N. Jaitly, R. Salakhutdinov, arXiv:1406.1231, 2014.
[31] H. Y. Du, J. Wang, Z. D. Hu, X. J. Yao, X. Y. Zhang, J. Agric. Food Chem. [63] A. Mauri, V. Consonni, M. Pavan, R. Todeschini, Match-Commun. Math.
2008, 56, 10785. Co. 2006, 56, 237.
[32] A. Graves, J. Schmidhuber, Neural Netw. 2005, 18, 602. [64] D. A. Winkler, J. Phys. Chem. Lett. 200, 23, 73.
[33] K. He, X. Zhang, S. Ren, J. Sun, arXiv:1502.01852, 2015. [65] J. Ma, R. P. Sheridan, A. Liaw, G. E. Dahl, V. Svetnik, J. Chem. Inf. Model.
[34] S. Ioffe, C. Szegedy, arXiv:1502.03167, 2015. 2015, 55, 263.
[35] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, [66] D. Lowe, Available at: http://blogs.sciencemag.org/pipeline/archives/
V. Vanhoucke, A. Rabinovich, arXiv:1409.4842, 2014. 2012/12/11/did_kaggle_predict_drug_candidate_activities_or_not
[36] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, [67] T. Unterthiner, A. Mayr, G. Klambauer, M. Steijaert, H. Ceulemans, J.
A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, L. Fei-Fei, Int. J. Com- Wegner, S. Hochreiter, In NIPS Workshop on Deep Learning and Repre-
put. Vis. 2015, 115, 211. sentation Learning (Montreal, QC), 2014.
[37] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. [68] D. Rogers, M. Hahn, J. Chem. Inf. Model. 2010, 50, 742.
Karpathy, A. Khosla, M. Bernstein, A. C. Berg, L. Fei-Fei, arXiv:1409.0575, [69] B. Ramsundar, S. Kearnes, P. Riley, D. Webster, D. Konerding, V. Pande,
2015. arXiv:1502.02072, 2015.
[38] P. Baldi, P. Sadowski, D. Whiteson, Nat. Commun. 2014, 5, 4308 [70] Y. Wang, J. Xiao, T. O. Suzek, J. Zhang, J. Wang, Z. Zhou, L. Han, K.
[39] D. Chicco, P. Sadowski, P. Baldi, In Proceedings of the 5th ACM Confer- Karapetyan, S. Dracheva, B. A. Shoemaker, E. Bolton, A. Gindulyte, S. H.
ence on Bioinformatics Computational Biology, and Health Informatics, Bryant, Nucleic Acids Res. 2012, 40, D400.
2014; p. 533. [71] S. G. Rohrer, K. Baumann, J. Chem. Inf. Model. 2009, 49, 169.
[40] J. E. Stone, J. C. Phillips, P. L. Freddolino, D. J. Hardy, L. G. Trabuco, K. [72] M. M. Mysinger, M. Carchia, J. J. Irwin, B. K. Shoichet, J. Med. Chem.
Schulten, J. Comput. Chem. 2007, 28, 2618. 2012, 55, 6582.
[41] K. Yasuda, J. Chem. Theory Comput. 2008, 4, 1230. [73] R. Huang, S. Sakamuru, M. T. Martin, D. M. Reif, R. S. Judson, K. A.
[42] B. R. Brooks, C. L. Brooks, A. D. Mackerell, L. Nilsson, R. J. Petrella, B. Houck, W. Casey, J. H. Hsieh, K. R. Shockley, P. Ceger, J. Fostel, K. L.
Roux, Y. Won, G. Archontis, C. Bartels, S. Boresch, A. Caflisch, L. Caves, Witt, W. Tong, D. M. Rotroff, T. Zhao, P. Shinn, A. Simeonov, D. J. Dix,
Q. Cui, A. R. Dinner, M. Feig, S. Fischer, J. Gao, M. Hodoscek, W. Im, K. C. P. Austin, R. J. Kavlock, R. R. Tice, M. Xia, Sci. Rep. 2014, 4, 5664.
Kuczera, T. Lazaridis, J. Ma, V. Ovchinnikov, E. Paci, R. W. Pastor, C. B. [74] D. N. Assis, V. J. Navarro, Expert Opin. Drug. Metab. Toxicol. 2009, 5, 463.
Post, J. Z. Pu, M. Schaefer, B. Tidor, R. M. Venable, H. L. Woodcock, X. [75] Y. Xu, Z. Dai, F. Chen, S. Gao, J. Pei, L. Lai, J. Chem. Inf. Model. 2015,
Wu, W. Yang, D. M. York, M. Karplus, J. Comput. Chem. 2009, 30, 1545. 55, 2085.
[43] D. A. Case, T. E. Cheatham, T. Darden, H. Gohlke, R. Luo, K. M. Merz, A. [76] H. Hong, Q. Xie, W. Ge, F. Qian, H. Fang, L. Shi, Z. Su, R. Perkins, W.
Onufriev, C. Simmerling, B. Wang, R. J. Woods, J. Comput. Chem. 2005, Tong, J. Chem. Inf. Model. 2008, 48, 1337.
26, 1668. [77] C. W. Yap, J. Comput. Chem. 2011, 32, 1466.
[44] J. C. Phillips, R. Braun, W. Wang, J. Gumbart, E. Tajkhorshid, E. Villa, C. [78] A. Lusci, G. Pollastri, P. Baldi, J. Chem. Inf. Model. 2013, 53, 1563.
Chipot, R. D. Skeel, L. Kale, K. Schulten, J. Comput. Chem. 2005, 26, 1781. [79] T. B. Hughes, G. P. Miller, S. J. Swamidass, ACS Cent. Sci. 2015, 1, 168.
[45] B. Hess, C. Kutzner, D. van der Spoel, E. Lindahl, J. Chem. Theory Com- [80] J. Zaretzki, M. Matlock, S. J. Swamidass, J. Chem. Inf. Model. 2013, 53,
put. 2008, 4, 435. 3373.
[46] M. Valiev, E. J. Bylaska, N. Govind, K. Kowalski, T. P. Straatsma, H. J. J. [81] J. Zaretzki, K. M. Boehm, S. J. Swamidass, J. Chem. Inf. Model. 2015, 55,
Van Dam, D. Wang, J. Nieplocha, E. Apra, T. L. Windus, W. de Jong, 972.
Comput. Phys. Commun. 2010, 181, 1477. [82] N. L. Dang, T. B. Hughes, V. Krishnamurthy, S. J. Swamidass, Bioinfor-
[47] Y. Shao, L. F. Molnar, Y. Jung, J. Kussmann, C. Ochsenfeld, S. T. Brown, A. matics 2016, 32, 3183.
T. B. Gilbert, L. V. Slipchenko, S. V. Levchenko, D. P. O’Neill, R. A. [83] T. B. Hughes, G. P. Miller, S. J. Swamidass, Chem. Res. Toxicol. 2015, 28, 797.
DiStasio, R. C. Lochan, T. Wang, G. J. O. Beran, N. A. Besley, J. M. [84] T. B. Hughes, N. L. Dang, G. P. Miller, S. J. Swamidass, ACS Cent. Sci.
Herbert, C. Y. Lin, T. Van Voorhis, S. H. Chien, A. Sodt, R. P. Steele, V. A. 2016, 2, 529.
Rassolov, P. E. Maslen, P. P. Korambath, R. D. Adamson, B. Austin, J. [85] A. Mayr, G. Klambauer, T. Unterthiner, S. Hochreiter, Front. Env. Sci.
Baker, E. F. C. Byrd, H. Dachsel, R. J. Doerksen, A, Dreuw, B. D. Dunietz, 2016, 3, 1.
A. D. Dutoi, T. R. Furlani, S. R. Gwaltney, A. Heyden, S. Hirata, C. P. Hsu, [86] S. Kearnes, B. Goldman, V. Pande, arXiv:1606.08793, 2016.
G. Kedziora, R. Z. Khalliulin, P. Klunzinger, A. M. Lee, M. S. Lee, W. Liang, [87] I. Wallach, M. Dzamba, A. Heifets, arXiv:1510.02855, 2016.
I. Lotan, N. Nair, B. Peters, E. I. Proynov, P. A. Pieniazek, Y. M. Rhee, J. [88] D. R. Koes, M. P. Baumgartner, C. J. Camacho, J. Chem. Inf. Model.
Ritchie, E. Rosta, C. D. Sherrill, A. C. Simmonett, J. E. Subotnik, H. L. 2013, 53, 1893.
Woodcock, W. Zhang, A. T. Bell, A. K. Chakraborty, D. M. Chipman, F. J. [89] O. Trott, A. J. Olson, J. Comput. Chem. 2010, 31, 455.
Keil, A. Warshel, W. J. Hehre, H. F. Schaefer, J. Kong, A. I. Krylov, P. M. W. [90] P. Mamoshina, A. Vieira, E. Putin, A. Zhavoronkov, Mol. Pharm. 2016,
Gill, M. Head-Gordon, Phys. Chem. Chem. Phys. 2006, 8, 3172. 13, 1445.
[48] M. W. Schmidt, K. K. Baldridge, J. A. Boatz, S. T. Elbert, M. S. Gordon, J. [91] D. E. Shaw, R. O. Dror, J. K. Salmon, J. P. Grossman, K. M. Mackenzie, J.
H. Jensen, S. Koseki, N. Matsunaga, K. A. Nguyen, S. J. Su, T. L. Windus, A. Bank, C. Young, M. M. Deneroff, B. Batson, K. J. Bowers, E. Chow, M.
M. Dupuis, J. A. Montgomery, J. Comput. Chem. 1993, 14, 1347. P. Eastwood, D. Ierardi, J. L. Klepeis, J. S. Kuskin, R. H. Larson, K.
Lindorff-Larsen, P. Maragakis, M. A. Moraes, S. Piana, Y. B. Shan, B. [123] G. Montavon, M. Rupp, V. Gobre, A. Vazquez-Mayagoitia, K. Hansen,
Towles, In Proceedings of the Conference on High Performance Com- A. Tkatchenko, K. R. M€ uller, O. Anatole von Lilienfeld, New J. Phys.
puting, Networking, Storage and Analysis, 2009. 2013, 15, 095003.
[92] K. Lindorff-Larsen, S. Piana, R. O. Dror, D. E. Shaw, Science 2011, 334, [124] R. Ramakrishnan, P. O. Dral, M. Rupp, O. A. von Lilienfeld, J. Chem.
517. Theory Comput. 2015, 11, 2087.
[93] J. N. Onuchic, Z. Luthey-Schulten, P. G. Wolynes, Annu. Rev. Phys. [125] A. Lopez-Bezanilla, O. A. von Lilienfeld, Phys. Rev. B 2014, 89, 235411.
Chem. 1997, 48, 545. [126] P. O. Dral, O. A. von Lilienfeld, W. Thiel, J. Chem. Theory Comput.
[94] P. G. Wolynes, Philos. T. Roy. Soc. A 2005, 363, 453. 2015, 11, 2120.
[95] M. Punta, B. Rost, Bioinformatics 2005, 21, 2960. [127] F. Faber, A. Lindmaa, O. A. von Lilienfeld, R. Armiento, Int. J. Quantum
[96] G. Shackelford, K. Karplus, Proteins 2007, 69, 159. Chem. 2015, 115, 1094.
[97] P. Fariselli, O. Olmea, A. Valencia, R. Casadio, Proteins 2001, Suppl 5, [128] O. A. von Lilienfeld, R. Ramakrishnan, M. Rupp, A. Knoll, Int. J. Quan-
157. tum Chem. 2015, 115, 1084.
[98] P. Bjorkholm, P. Daniluk, A. Kryshtafovych, K. Fidelis, R. Andersson, T. R. [129] A. R. Katritzky, V. S. Lobanov, M. Karelson, Chem. Soc. Rev. 1995, 24,
Hvidsten, Bioinformatics 2009, 25, 1264. 279.
[99] K. M. S. Misura, D. Chivian, C. A. Rohl, D. E. Kim, D. Baker, Proc. Natl. [130] T. Le, V. C. Epa, F. R. Burden, D. A. Winkler, Chem. Rev. 2012, 112, 2889.
Acad. Sci. USA 2006, 103, 5361. [131] S. V. Kalinin, B. G. Sumpter, R. K. Archibald, Nat. Mater. 2015, 14, 973.
[100] J. Skolnick, D. Kihara, Y. Zhang, Proteins 2004, 56, 502. [132] C. M. Breneman, L. C. Brinson, L. S. Schadler, B. Natarajan, M. Krein, K.
[101] M. Y. Shen, A. Sali, Protein Sci. 2006, 15, 2507. Wu, L. Morkowchuk, Y. Li, H. Deng, H. Xu, Adv. Funct. Mater. 2013, 23,
[102] J. Moult, Curr. Opin. Struct. Biol. 2005, 15, 285. 5746.
[103] Y. Zhang, Curr. Opin. Struct. Biol. 2008, 18, 342. [133] A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards, S. Dacek, S.
[104] D. S. Marks, T. A. Hopf, C. Sander, Nat. Biotechnol. 2012, 30, 1072. Cholia, D. Gunter, D. Skinner, G. Ceder, K. A. Persson, APL Mater.
[105] P. Di Lena, K. Nagata, P. Baldi, Bioinformatics 2012, 28, 2449. 2013, 1, 011002.
[106] J. M. Chandonia, G. Hon, N. S. Walker, L. Lo Conte, P. Koehl, M. Levitt, [134] E. O. Pyzer-Knapp, K. Li, A. Aspuru-Guzik, Adv. Funct. Mater. 2015, 25,
S. E. Brenner, Nucleic Acids Res. 2004, 32, D189. 6495.
[107] J. Eickholt, J. Cheng, Bioinformatics 2012, 28, 3066. [135] V. Venkatraman, S. Abburu, B. K. Alsberg, Phys. Chem. Chem. Phys.
[108] B. Monastyrskyy, K. Fidelis, A. Tramontano, A. Kryshtafovych, Proteins 2015, 17, 27672.
2011, 79, 119. [136] A. Schuller, G. B. Goh, H. Kim, J. S. Lee, Y. T. Chang, Mol. Inf. 2010, 29,
[109] J. Lyons, A. Dehzangi, R. Heffernan, A. Sharma, K. Paliwal, A. Sattar, Y. 717.
Zhou, Y. Yang, J. Comput. Chem. 2014, 35, 2040. [137] M. Deetlefs, K. R. Seddon, M. Shara, Phys. Chem. Chem. Phys. 2006, 8,
[110] E. Faraggi, T. Zhang, Y. Yang, L. Kurgan, Y. Zhou, J. Comput. Chem. 642.
2012, 33, 259. [138] D. Fourches, D. Q. Y. Pu, C. Tassa, R. Weissleder, S. Y. Shaw, R. J.
[111] R. Heffernan, K. Paliwal, J. Lyons, A. Dehzangi, A. Sharma, J. Wang, A. Mumper, A. Tropsha, ACS Nano. 2010, 4, 5703.
Sattar, Y. Yang, Y. Zhou, Sci. Rep. 2015, 5, 11476. [139] Kaggle, Available at: http://www.kaggle.com/c/MerckActivity/
[112] E. Faraggi, Y. D. Yang, S. S. Zhang, Y. Q. Zhou, Structure 2009, 17, leaderboard
1515. [140] G. Barta, Front. Env. Sci. 2016, 4, 1.
[113] G. L. Wang, R. L. Dunbrack, Nucleic Acids Res. 2005, 33, W94. [141] A. Abdelaziz, H. Spahn-Langguth, K. W. Schramm, I. V. Tetko, Front.
[114] S. F. Altschul, T. L. Madden, A. A. Schaffer, J. H. Zhang, Z. Zhang, W. Env. Sci. 2016, 4, 2.
Miller, D. Lipman, J. Nucleic Acids Res. 1997, 25, 3389. [142] J. Huang, A. D. MacKerell, J. Comput. Chem. 2013, 34, 2135.
[115] S. Altschul, T. Madden, A. Schaffer, J. H. Zhang, Z. Zhang, W. Miller, D. [143] K. A. Beauchamp, Y. S. Lin, R. Das, V. S. Pande, J. Chem. Theory Com-
Lipman, FASEB J. 1998, 12, A1326. put. 2012, 8, 1409.
[116] J. Meiler, M. Muller, A. Zeidler, F. Schmaschke, J. Mol. Model. 2001, 7, 360. [144] G. B. Goh, D. M. Eike, B. P. Murch, C. L. Brooks, J. Phys. Chem. B 2015,
[117] B. Rost, J. Struct. Biol. 2001, 134, 204. 119, 6217.
[118] B. Alipanahi, A. Delong, M. T. Weirauch, B. J . Frey, Nat. Biotechnol. [145] M. B. Gee, N. R. Cox, Y. F. Jiao, N. Bentenitis, S. Weerasinghe, P. E.
2015, 33, 831. Smith, J. Chem. Theory Comput. 2011, 7, 1369.
[119] S. Zhang, J. Zhou, H. Hu, H. Gong, L. Chen, C. Cheng, J. Zeng, Nucleic [146] Y. Luo, B. Roux, J. Phys. Chem. Lett. 2010, 1, 183.
Acids Res. 2016, 44, e32. [147] M. T. Ribeiro, S. Singh, C. Guestrin, arXiv:1602.04938, 2016.
[120] M. T. Weirauch, A. Cote, R. Norel, M. Annala, Y. Zhao, T. R. Riley, J. [148] M. T. Ribeiro, S. Singh, C. Guestrin, arxiv:1606.05386, 2016.
Saez-Rodriguez, T. Cokelaer, A. Vedenko, S. Talukder, H. J. [149] A. Lusci, D. Fooshee, M. Browning, J. Swamidass, P. Baldi, J. Chemin-
Bussemaker, Q. D. Morris, M. L. Bulyk, G. Stolovitzky, T. R. Hughes, D. form. 2015, 7.
Consortium, Nat. Biotechnol. 2013, 31, 126. [150] S. J. Swamidass, C. A. Azencott, T. W. Lin, H. Gramajo, S. C. Tsai, P.
[121] Y. Shen, O. Lange, F. Delaglio, P. Rossi, J. M. Aramini, G. H. Liu, A. Baldi, J. Chem. Inf. Model. 2009, 49, 756.
Eletsky, Y. B. Wu, K. K. Singarapu, A. Lemak, A. Ignatchenko, C. H.
Arrowsmith, T. Szyperski, G. T. Montelione, D. Baker, A. Bax, Proc.
Natl. Acad. Sci. USA 2008, 105, 4685.
Received: 14 September 2016
[122] K. Hansen, G. Montavon, F. Biegler, S. Fazli, M. Rupp, M. Scheffler, O. Revised: 9 January 2017
A. von Lilienfeld, A. Tkatchenko, K. R. Muller, J. Chem. Theory Comput. Accepted: 18 January 2017
2013, 9, 3404. Published online on 8 March 2017