Deep Learning in Chemistry

Adam C. Mater, Michelle L. Coote*

ARC Centre of Excellence for Electromaterials Science, Research School of Chemistry,

Australian National University, Canberra, Australian Capital Territory 2601, Australia

ABSTRACT

Machine learning enables computers to address problems by learning from data. Deep learning is

a type of machine learning that uses hierarchical recombination of features to extract pertinent

information, and then learn the patterns represented in the data. Over the last eight years, its

abilities have increasingly been applied to a wide variety of chemical challenges, from improving

computational chemistry, to drug and materials design, and even synthesis planning. This review

aims to explain the concepts of deep learning to chemists from any background and will follow

this with an overview of the diverse applications demonstrated in the literature. We hope that this

will empower the broader chemical community to engage with this burgeoning field and foster the

growing movement of deep learning accelerated chemistry.

1
INTRODUCTION

Deep learning has emerged as a dominant force within machine learning over the last ten years

through a series of demonstrations of its frequently superhuman predictive power1-7. These initial

demonstrations have fostered a desire among researchers to harness its abilities to address

challenges in a diverse range of areas. Chemistry stands as one of these areas, with a variety of

immensely complex problems such as retrosynthesis, reaction optimization, and drug design.

Historically, these have presented fierce opposition to computational approaches based on hand

coded heuristics and rules, with these approaches being met with skepticism by chemists8-11. There

are strong analogies between these problems and those which deep learning has come to dominate,

such as computer vision and natural language processing12. As a result of this, chemistry has seen

a steady increase in the deployment of these technologies, with many demonstrating significant

improvements in predictive accuracy and ability to replicate human decision making13-24.

With the prevalence deep learning is likely to achieve within chemistry, it is important that

chemical researchers not familiar with the minutiae of deep learning become comfortable with

how these techniques function. There have been a number of reviews covering subfields of deep

learning in chemistry. Goh et al.’s14 review serves as an excellent overview for theoretical chemists

and has accessible explanations of the core deep learning concepts. While not strictly a review,

Wu, Ramsundar et al.’s13 paper on MoleculeNet provides an extensive summary of the available

descriptors and datasets as well as model comparisons. In addition to this there are a number of

broader reviews covering machine learning for drug design25-26, synthesis planning11, materials

science27, quantum mechanical calculations28, and cheminformatics29. This paper seeks to adopt a

central stance on deep learning in chemistry, explaining the core ideas in the broadest possible

sense, without emphasis on mathematical detail, and with reference to chemical applications. This

2
understanding will then be used to provide a broad overview of the influences deep learning has

so far had across applied and theoretical chemistry.

THE BIG PICTURE

Machine learning is an extremely broad sub-field of artificial intelligence that aims to solve the

problem of computers learning from data. Representation learning is a subset of machine learning

in which computational models learn internal representations of objects that inform the decisions

or predictions that they make. Finally, deep learning is a subset of representation learning in which

multiple layers of internal representations, initially of simple shapes such as edges, are combined

to form increasingly complex objects, like faces30. Chemistry stands as an exemplar of this

phenomenon, with the behavior of molecules determined not simply by atoms, but their immediate

grouping into functional groups, followed by interactions between these groups at increasing

ranges. Ostensibly, this makes chemistry an ideal candidate for these methods. Unfortunately,

molecules also supply a set of challenging problems including sampling sufficiently diverse

molecules and their accompanying conformational space, effectively representing molecules, and

obtaining suitably large datasets.

Understanding how these problems are being addressed requires an introduction to the methods

of deep learning. Machine learning, and thus deep learning, at its core contains three components:

the data (and its associated representation), the model that will learn to interpret the data, and a

prediction space from which we draw utility. The model in deep learning (as well as other

methodologies) represents an optimization cycle of three sub-components: the learner, evaluation,

and optimization. These ideas are summarized in Figure 1. Understanding chemical deep learning

requires familiarity with each of these ideas and the unique challenges chemistry presents in each.

The first section of this review seeks to disambiguate these topics, beginning with an exploration

3
of data and how molecules are represented. This leads into a discussion of three of the dominant

model architectures in chemical deep learning. The prediction space will then be examined, to

explain how chemical problems must be phrased in order to make them amenable to deep learning.

This section will conclude with a brief overview of terms that are frequently referenced in the

literature.

Data Model Prediction Space

Molecule Prop. Learner Predicted
Pred. Data Sq.E.
kJ/mol
-10.2 -19.5 86.5
CCOCN -19.5
C(C=O) -22.5 -20.6 3.6
-20.6
4.3 1.4 8.4
CCCCC 1.4 Training
-5.4 -2.3 9.6 Cycles
CC=CC -2.3
-29.2 -20.9 68.9
CCOCC -20.9 1
-7.9 -5.4 12.3
COC=C -5.4
10
O=CCN 12.2 Evaluation
CC#CC -9.5 100
CCSCC -2.7 Training Cycle
P1 P2 P3 P4
C(Cl)(F) -31.0
Predict
CCSCN -12.9
Evaluate
BrCCO -17.2
Optimization Optimize
ICCCF -2.5 Calculated

Figure 1 - The Big Picture of Deep Learning. The learner shown in this image is a deep

feedforward network, however this same procedure applies to a plethora of learners. The ∆P term

indicates the change to the parameters in each network layer after the input layer. The data in this

image is fictitious and thus labelled simply as property.

The Data. Learning cannot happen without data, and in the case of supervised learning, this data

must be labelled. These labels indicate the ground truth associated with the data point, such as

associating a label of ‘truck’ with an image of a truck. In a chemical sense, the data can be a

representation of a molecule with its free energy of solvation labelled or any other property. This

creates one of the first big challenges of deep learning, how can enough data be obtained? The

most dominant demonstrations of deep learning’s potential are in fields where data is abundant,

typically where millions, if not billions, of data points can be obtained through distributed

4
collection via social networks or even more broadly, the internet1, 31. In the case of science, the

requisite volume of data only exists in certain applications. In chemistry, all levels of data are

present, with extensive data available for successful reactions or ground state energies, a moderate

amount of data for specific properties such as ionization energies, through to relatively small

databases for properties such as free energies of solvation32-34. As a result of this need for data,

chemical deep learning has formed a strong link with computational chemistry due to the latter’s

capacity to generate huge volumes of data significantly faster than it could be obtained in a

laboratory33, 35. This presents challenges however, due to the poorer accuracy of these calculations

relative to experimentally obtained results. Lab-derived datasets are available, and are the gold

standard, but aside from reaction databases, the number of data points they contain is not usually

on the same order of magnitude36.

Additionally, effective assessment of deep learning models requires that the data undergoes

subsequent splitting. Assessing a model on the data it was trained on leads to significant overfitting

in which the model learns to reproduce that specific set of data but not the trends underlying it. To

stop this ‘memorization’ of data, it is common to test the models on data that they have not yet

seen. This is typically done by dividing the data into three separate sets: the training, validation,

and test sets. The training set (typically 60-80% of the data) is given to the network in its entirety

and its labels are used to adjust the network’s parameters in supervised learning. The validation

set (typically 10-20% of the data) is used to ensure that the model is not overfitting by providing

a constant estimate of its performance on unseen examples. In addition to this, when training

multiple models validation data is used to identify the best performing model. Finally, the third

dataset, the test set, is used as the final performance evaluation of the chosen model on the

remainder of the withheld data. In order to remove any bias in the partitioning of the data into these

5
sets, k-fold cross validation is used, in which the data partitioning process is randomized k times37.

Any model is highly dependent on the way in which the data is represented. Due to this, deep

learning has a strong interest in the long-standing cheminformatics problem of how best to

represent chemical structures for a computer.

There are three key invariances that must be captured, two of which are intuitively captured by the

human visual processing. Formally, these are:

• Permutation invariance – the representation must be unaltered by a change in the

specified order of the atoms.

• Translational invariance – the representation must not be changed by a translation in

space.

• Rotational invariance – the representation must be unchanged by a rotation operation.

Familiar examples of these variances are shown in Figure 2 below. An additional requirement

for some models is a fixed size input. This is typically achieved by padding the representation with

zeros for smaller molecules.

6
 Rotation Variance Translation Variance

6 1 0 0 6 0 1 0
1 6 2 1 0 6 1 0
0 2 8 0 1 1 6 2
0 1 0 6 0 0 2 8

Permutation Variance

Figure 2: Three key variances in common molecular descriptors that much be overcome for

deep learning. The top two invariance grids show acetone undergoing rotation and translation in

a fixed reference grid. Permutation invariance shows two equivalent acetone representations as

atom connectivity matrices introduced by Spialter38. The atom connectivity matrix has nuclear

charges listed along the diagonal, with off diagonal elements representing bonds of associated

bond order between the diagonally located atoms that they link. To facilitate the following

discussion of model architectures, a brief exploration of the most widely used molecular

representations is required.

A molecular graph is a set of vertices (atoms) that are connected by edges (bonds). This can be

expressed in matrix form, with an example shown in Figure 2. Originally, deep learning models

utilized extended connectivity fingerprints (ECFP). These involve assigning an integer identifier

to each atom and updating it to include information from neighboring atoms by expanding a

circular radius that analyzed the atoms contained within. Within this circle, the atoms were sorted

to achieve permutation invariance and, by compressing spatial information into integer identifiers,

the two spatial invariances were also satisfied. Each of these integer identifiers were passed

7
through a hashing function to produce a number, which, combined with modulo arithmetic,

allowed a particular index within a fixed vector to be switched to a one39. This vector has a fixed

size, achieves the three invariances, but contains only zeroes and ones and is thus referred to as a

bit vector. This is the basic methodology that inspired the molecular graph-based models that will

be described below. The idea of gathering information about an atom’s local environment while

preserving their invariances was retained, but critically, they encode the molecular information in

a real valued vector allowing for significantly richer information to be embedded.

The Simplified Molecular Input Line Entry System (SMILES) is a classic cheminformatics

representation that uses a set of ordered rules and specialized syntax to encode three dimensional

chemical structures as strings of text40-41. An additional procedure can be applied on top of this to

create permutation invariance, a process known as canonicalization. The other frequently used

text-based identifier, the international chemical identifier (InChI), is not regularly used in deep

learning due to multiple studies finding that its more complex and numeric formulations lead to

deterioration in predictive performance42-43. A reaction variant of SMILES, which contains

specialized grammar to describe chemical transformations is also frequently used in machine

learning for models that operate on reaction datasets44-45.

Graph inputs currently dominate due to their ability to extract higher-level features, and the

increase in predictive performance that comes with this. It must also be noted that there are

additional representations such as point clouds46 and Coulomb matrices47 that are also used.

Finally, regardless of representation, molecules must be entered into datasets order to be

transformed into a model input. To digitize the enormous number of structures in the literary

corpus, deep learning has been used to automate the digitization of these structures48.

8
 The Model. In any given deep learning framework, the model is the component that transforms

the data into a prediction, classification, or action. The model relies on an interplay between its

learner, evaluation, and optimization. The learner contains a set of parameters which define how

each input point is converted into an output. This prediction is then quantitatively compared to the

desired output via an evaluation or cost function. Finally, optimization alters the parameters of the

model to decrease the difference between the predicted and the desired output for each data point.

This cycle of the model making predictions, which are then evaluated, and finally used to optimize

the model’s parameters is bundled into a single training cycle. These ideas are summarized visually

in Figure 1.

Deep learning is named for to the computational depth of its learner, i.e. how many sequential

layers of calculations are required. The learner is thus the defining feature of deep learning

methods, with an intimate link being formed with the field of connectionism. Connectionism is

focused on the development of artificial neural networks (ANNs) and their many variants. These

learners are neurologically inspired systems of interconnected virtual neurons (an example

network is shown as the learner in Figure 1). Due to their prominence in deep learning methods,

the remainder of the model discussion will focus on variants of ANNs. A mathematical discussion

is not the intent of this review, however much of this discussion is inspired by Deep Learning by

Goodfellow et al.30 which contains an extensive and rigorous treatment of deep learning methods.

Despite the enormous diversity in the learner architecture, the evaluation and optimization

procedures are dominated by a few methods. In the case of neural networks, the evaluation step is

typically a simple function that assesses the learner’s performance across batches, or all, of the

data; two common examples are the root mean squared deviation (RMSD) or the cross-entropy

9
cost function. The optimization typically employed for neural networks is the powerful

backpropagation algorithm49. This method propagates the gradients backwards from the outputs

through to the inputs, and using the information contained within these, alters the parameters of

each non-input node in a manner that lowers to deviation between the predicted and true values49.

To highlight what makes the learner networks so different, three of the dominant architectures will

now be discussed.

A deep neural network (DNN) is the prototypical deep learning architecture. DNNs contains

three separate layer types, input, hidden and output. Each layer is comprised of a set of neurons

and in fully connected systems, each hidden layer neuron connects to all neurons in the previous

and following layers. The ‘wiring’ of the network (how many layers there are and how they are

connected), as well as what function each neuron performs is typically referred to as the network’s

topology, and the performance of the network is highly dependent on the chosen topology.

Each neuron in the input layer receives a single, real number from each data point and is thus

represented as a fixed size vector. DNNs were frequently used with ECFP representations, in

which a one indicates the presence of a particular substructural feature which may or may not

correspond to a recognizable function group, and a zero its absence39.

The neurons within the hidden and output layers have two types of trainable parameters. Every

incoming connection has a scalar weight associated with it, that is expressed within a matrix, and

then each neuron has its own scalar term called a bias, collected into a vector for each layer. The

forward data pass is computed by multiplying the input vector with the weight matrix, to produce

an output vector. The bias is then added to this output vector, and it is then passed through an

activation function. This function is critical as it allows the network to model nonlinear

phenomena. One of the simplest and most widely used activation functions is the rectified linear

10
unit (ReLU)50, which simply maps any non-positive number to zero and returns any positive

number unchanged. This vector now becomes the input for the next layer of the network and the

process continues until the output layer is reached.

The output layer is typically either a single real number, indicating that the network is built for

regression (i.e for predicting a property such as the enthalpy of combustion), or a vector that

contains the likelihood of the input being classified as certain objects, and thus a classification

network. In the case of classification tasks, the softmax activation function is commonly used; it

converts a vector of real numbers into a probability distribution where the sum of all terms is one

and all terms are between zero and one. This allows the network to produce a distribution over the

classes, indicating which is most likely. The utilization of matrix operations allows these models

to leverage graphical processing units (GPUs) to massively accelerate the computation51. A

summary of this matrix multiplication process is given in Figure 3.

Input Layer Hidden Layer Output
Layer

NH2

Add Add
Input Molecule
Bias Bias
T
T T
T T

Hidden Layer Output

0 1 0 1 0 ··· 0 0 0 1 1 ⇥
Weights
= ⇥
Weights
=

Input Bit Vector

ReLU Softmax
Activation Function Activation Function

Figure 3: Matrix view of a typical neural network forward pass: The input molecule was

chosen at random, and the bit vector is a simple structural representation that can roughly be

viewed as ones indicating the presence of certain substructural feature, and zeros representing the

11
absence. The bold T’s above the vectors indicate that the transpose is used in the multiplication in

order to make the operation defined.

Learning in these networks involves the backpropagation algorithm, which applies the

multivariate chain rule from calculus to efficiently calculate the gradients of each trainable

parameter in the network, and then uses these to alter the parameters in a way that lowers the cost

function. DNNs have been effective at addressing chemical problems. However, other deep

learning architectures that evolved in two of AI’s largest research areas, computer vision (CV) and

natural language processing (NLP), have largely superseded them.

Graph convolutional neural networks (GCNN). Computer vision is the field of research that

aims to use computers to see in a manner similar to humans. Convolutional neural networks

(CNNs) are networks specialized for interacting with grid like data, such as a 2D image. As

molecules are typically not represented as 2D grids, chemists have focused on a variant of this

approach: graph convolutional neural networks (GCNN) on molecular graphs.

Molecular graphs confer key advantages: they bypass the conformational challenge of using 3D

representations while maintaining invariance to rotation and translation due to their pairwise

definition. A wide variety of molecular graph implementations have developed in recent years18,

22-23, 52-55
and the MoleculeNet paper by Wu, Ramsundar et al.13 offers a concise conceptual

comparison of six major variants. To facilitate the following explanation, the framework of neural

message passing networks put forth by Gilmer et al53. will be used.

Neural message passing networks are a chemically motivated system to understand and compare

these GCNN systems. Fundamentally this approach utilizes a convolutional layer, simply a matrix

of scalar weights, to exchange information between atoms or bonds within a molecule and produce

12
a fixed length, real-valued vector that embeds the molecular information. To begin, they generate

or compute a feature vector for each atom within the molecule; this can contain information such

as how many hydrogens are attached to the atom, its hybridization, whether or not it is aromatic

or in a ring, etc. These feature vectors are then collected into a matrix. Additionally, they generate

a graph topology matrix that specifies the connectivity of the graph, similar to Figure 2 although

often without bond order or atomic number along the diagonal. In a forward convolutional pass,

these three matrices are multiplied together. This allows information to be exchanged between the

feature vectors of each atom with its immediate neighbors, in accordance with the connectivity

specified by the topology matrix. This updates each atom’s feature vector to include information

about its local environment. This updated feature vector matrix is then passed through an activation

function (i.e. ReLU) and can then be iteratively updated by using it as the feature matrix in another

convolutional pass. This propagates information throughout the molecule. Finally, these atom

feature vectors are either summed or concatenated to give a unique, learned representation of the

molecule as a real valued vector (see Figure 4). Alternative approaches to generating this learned

representation have been put forth, such as using traditional computer vision CNNs on 2D grid

images56 of molecules. However, molecular graphs remain the dominant paradigm.

The learned representation in vector form is referred to as a representation in latent space, and

is then used as the input for a traditional fully connected DNN to finally make the classification or

prediction. This process of learning its own molecular representation is the cause of it being in the

broader class of representation learning methods. Backpropagation is once again used to train these

networks by propagating gradients backwards and determining how to change the convolution

matrix weights and the parameters in the DNN.

13
 Recurrent neural networks (RNNs), introduced by Hopfield57 in 1982, are specialized for

dealing with sequences of arbitrary length. This makes them ideally suited to handling textual

representation of chemical information, such as SMILES40. The critical difference is that in the

previous architectures each data input is distinct, while in an RNN each input will influence the

next one. An illustrative example is viewing any particular input, such as a SMILES string, as time

series data. The presence of a carbon atom at one moment in time influences what the next

character is likely to be. This is expressed in the architecture by feeding the output of the hidden

layer for that carbon into the hidden layer of the next atom. As a more complex illustration, this

process can be used to model reactions by utilizing the SMILES reaction strings to encode the

information, and train the network to predict the product (See Figure 4). The feeding of one hidden

state into the next gives the system a recursive relationship within the hidden layer, but it can be

viewed as directional by ‘unfolding’ the network to form of an unfolded, acyclic network graph.

By doing this, it maintains a history of all previous inputs, and they influence its prediction at a

later time. The network can then be trained using a recursive form of backpropagation58. This is

the simplest RNN but more sophisticated and powerful variants such as neural Turing machines59

and long-short term memory networks (LSTM)60 that incorporate memory into the network are the

current leaders. This ability to use previous information has led to their dominance in sequence-

based tasks such as machine translation, as previous words define the context and thus, what the

next word is likely to be.

14
 Updated Hidden Predicted Product
GCNN Graph
Features Weights RNN COc1cncc(Br)c1

Pass 1 ⇥ ⇥
Output ... ...
ReLU
Ot 1 Ot Ot+1
NH2
O Pass 2 ⇥ ⇥

ReLU Unfold
Hidden ... ...
Pass 3 ⇥ ⇥
Ht 1 Ht Ht+1

(1) ReLU
(2)
(3)
(4) Input ... ...
(5) Concatenate
It 1 It It+1
(6)

1. Z 4. Number of Brc1cncc(Br)c1.C[O-]>CN(C)C=O.[Na+]
2. In Ring Hydrogens
Feedforward network Reagent Agent
3. Implicit 5. Aromaticity
Valence 6. Charge

Figure 4: Illustration of the GCNN and RNN architecture for chemical applications. Colored

arrows stemming from the amine group indicate the information transfer from the nitrogen to other

heavy atoms, with the color corresponding to the convolutional pass. Light grey arrows indicate

each atom’s feature vector in the matrix, importantly properties such as atomic number (Z) are

often encoded using one hot vectors, which are binary, but for spatial efficiency the integer is used

in its place. The RNN model shows a simplified ‘many to many’ recurrent network, with the text

above and below the dashed lines indicating a stylized reaction prediction system inspired by the

work of Schwaller et al.61 This system takes in reagent and agent SMILES, and predicts the variable

length product string, however the LSTM architecture they used is significantly more complex

than the one shown above.

The prediction space is the set of all possible outputs for the network. More intuitively, it can be

thought of as the utility of the network or the question that the network can produce an output for.

As discussed above, supervised learning requires labelled data that allows the model to iteratively

improve its predictive performance. This model relies on a quantitative error assessment by the

evaluation component, and thus each deep learning problem needs to be framed in such a way that

15
it can be quantitively evaluated. This creates a significant challenge in chemistry, as questions such

as ‘what is the best synthetic route?’ require systematic analysis to produce a question that can be

numerically evaluated, and thus produce quantitively labelled data. In the broader context of

artificial intelligence, this means that these systems are weak AI, capable of solving only a single,

extremely narrow task, and not being capable of meaningfully answering even slight deviations

from the question it was trained on.

Commonly Used Terms. Before concluding this section, a brief explanation of commonly used

ideas and terms will be provided. Each term is linked to seminal papers and, where appropriate,

accompanied by an example of its application in chemistry.

• Transfer Learning – Transfer learning involves using a network that has been trained on a

related task, and then tweaking its parameters to adapt to a new task, often with less data62.

It has been used to adapt a model trained on DFT to a smaller database of higher fidelity

calculations by Smith et al.63.

• Multitask learning – This involves training a model on multiple prediction tasks at the same

time to decrease the likelihood of overfitting64. It has been demonstrated improvements for

toxicity or bioactivity prediction65.

• One Shot Learning – A technique used to overcome applications with extremely limited

data that uses networks to compress inputs into a continuous latent space and then compares

the representation in this space to a larger, trained latent space66. It has been used in chemistry

for low-data drug discovery67.

• Autoencoders – A network architecture which compresses an input to a real valued vector,

commonly referred to as the latent space. A decoder network then takes this vector as its

16
 input and tries to reproduce the original input data68. It has been used to design molecules by

training the latent space to reflect a particular property, and then navigating it43.

• Generative Adversarial Networks (GANs) – GANs utilize two networks in a competitive

scheme. One network has to generate data, and another has to determine if a particular data

point is a fake generated by the network, or a real one from the dataset. By competing with

one another, the generating network learns to create high quality imitations of the dataset69.

It has been utilized for the inverse molecular design problem70.

• Data Augmentation – This involves expanding a dataset by creating new training examples

through reasonable manipulations of the data. One of the simplest demonstrations of this is

rotating images in a dataset, but maintaining the same label in a way that is obvious to

humans, i.e., a car is still a car at different angles71. This has been used with SMILES to

enumerate the different potential orderings and increase the predictive performance72.

• Reinforcement Learning – When the model learns iteratively through trial and error by

making its cost function measure its progress towards a particular goal73. It has been used to

train a model to optimize reactions74.

• Supervised Learning – Supervised learning involves giving the model a labelled dataset,

effectively telling it what it needs to learn. While this is currently the dominant learning

paradigm in machine learning, it is not representative of how humans tend to learn.

• Unsupervised Learning – Unsupervised learning is learning in which the model is not told

what to reproduce and instead tries to separate the data into its underlying clusters.

Algorithms such as k-mean clustering fall into this category and it is much closer to how

humans learn75.

17
DEEP LEARNING APPLICATIONS

This section reviews the multiple areas of chemistry that deep learning has thus far impacted,

presenting examples in each that highlight particular achievements. To create a logical narrative,

this discussion will follow an idealized chemical workflow. To build a molecule with a particular

property would first require developing methods to accurately correlate any given structure to the

property. These can then be used to intelligently design a molecule that maximizes the desired

property. The final step is to design an efficient synthesis from readily available starting materials

(Figure 5). This creates a closed feedback cycle in which the synthesized molecule can be

experimentally analyzed, and this information can then improve the models that link molecules to

properties. Deep learning has influenced every stage of this workflow, beginning with

understanding molecules.

Designing Molecules

Understanding Molecules Synthesizing Molecules

Cat. X
Prop. Val. 60o C

µ A B
4.653 41%

↵ 76.57 Materials Drug Cat. Y

90o C
zpve 0.149 Design Design A B
68%
✏HOMO -0.274
hv
✏LUMO 0.0267 Molecular 25o C

✏Gap A B
0.301 Design 93%

Accelerated Conformational Reaction

Retrosynthesis
QM Exploration Prediction

Reaction
QSPR
Optimization

Figure 5: Deep learning influence on the idealized chemical workflow. Illustrative examples of

each task are shown in the dialogue boxed with arrows indicating the closed cycle that is contained

18
within the framework. The property values in the blue panel were obtained from the QM9 dataset

for a randomly chosen molecule33.

Accelerated Computational Models. Computational modelling in chemistry seeks to use

physics-based calculations to determine the properties and behavior of a given molecular system.

There are two distinct ways which deep learning can be used within this space. The first is to

integrate the deep learning method with physics style approaches to alleviate computational

bottlenecks. The second is directly predicting properties from molecular structures, thereby by-

passing physical laws altogether.

Integrating deep learning methodologies with physics-based approaches involves training the

network to predict a key component of the overall calculation. These include using the deep

learning model to predict potential energy surfaces52, 76-77, force fields78, add corrections to ab initio

calculations79 and to bypass expensive stages in both density functional and wavefunction

methods80-81. There is an excellent review and tutorial on using neural networks for the prediction

of potential energy surfaces by Behler82-83. Many of these methods adapt a method introduced by

Behler and Parrinello84 in 2007 that determines the energy of the system by summing the energetic

contribution of each atom. This method transforms the cartesian coordinates of a molecule using

radial symmetry functions, which capture the information of each atom’s immediate environment.

This transformed representation is then passed through a neural network that predicts the

contribution of this atom to the total energy. This general method of using functions to capture an

atom’s local environment, then predicting its energy through a network and finally summing these

contributions has been refined in a variety of ways. Notable work in the field includes that of

Schutt et al.24 which produced size extensive predictions with an average error of 1 kCal/mol, and

19
the work of Smith et al.52 which produced errors below 1 kCal/mol and generalization to larger

molecules. Schutt et al.’s85 work has been further refined, and developed into an open source

software package (SchNetPack) that can be used to predict properties.

The advantages of this approach are that it is more flexible than mapping a structure to a

property, and it is more interpretable due to its physical basis. The difficulty is that, as there are

typically still physics-based calculations involved, such methods cannot achieve the same speed

as those that map purely from a structure to a particular property. It is important to note that there

is a large literature base for using kernel ridge regression as the ML method. This approach has

achieved excellent results but is not a deep learning method, and thus is outside the scope of this

review. For an overview of these methods the reader is referred to von Lilienfeld’s excellent

review86.

Quantitative Structure Property/Activity Relationships. The alternative approach to deep

learning in computational chemistry is training a direct map from a simple representation of the

molecule through to the desired property. This is a diverse field of research that can broadly be

captured under the two fields of quantitative structure property relationship (QSPR) and

quantitative structure activity relationships (QSAR). Broadly speaking, QSPR seeks to predict

properties of molecular systems, such as thermochemistry, while QSAR seeks to predict the

activity of that molecule within a broader context, such as toxicity within biological systems. The

goal of these methods is to maximize accuracy of prediction, with chemical accuracy for QSPR

commonly being set to 1 kCal/mol or approximately 4 kJ/mol87-88. The properties that can be

predicted are entirely determined by the available training data, and there are many databases

available. There are summaries of available databases in both the review by Butler et al.27 and the

20
MoleculeNet paper by Wu, Ramsundar et al.13. Typically for properties that can be readily

computed, such as ground state energies, ionization energies, or dipole moments, computational

datasets are the norm. These are typically computed with a DFT method in order to maximize

speed and allow for as much data as possible to be generated. Some of the most commonly utilized

are QM933, ANI-135, and the Materials Project89. Properties that are difficult or currently impossible

to compute accurately, such as toxicity, free energies of solvation, biological activity, or binding

affinities rely instead on experimental datasets that typically contain significantly fewer entries

due to the challenge in obtaining them. Frequently used datasets include ChEMBL90, PubChem91,

and FreeSol34.

For this type of problem. DNNs were the most widely used network architecture for the first half

of this decade. They have been used to effectively predict electronic properties19, 87, 92, bioactivity21,
93-95
, toxicity15, 96-97, reactivity92, as well as other physical properties98. Multitask networks are also

frequently used due to the increase in predictive performance, as well as increased robustness to

overfitting15, 21, 93-94. RNNs have been more widely used as the generative networks that produce

novel molecules which will be discussed later. For predictive purposes, however, they have

utilized both graph type input structures similar to GCNNs to predict aqueous solubility18 and drug

toxicities99 as well as the more traditional text based inputs of SMILES for general property

prediction100-101.

In almost all cases however, GCNNs and their many variants have demonstrably better

predictive performance than either of the other two classes of methods. Due to the focus on

improving network architecture, convolutional models are often tested against a variety of

benchmarks. However, there has been a particular push to improve the predictions of electronic

properties in order to ease the computational stress imparted by physics-based calculations56, 102-103.

21
In addition to this GCNNs have shown dominance in predicting bioactivity104, polymer property

predictions105, and physical properties55, 106. Work to increase their predictive abilities is ongoing,

but errors below 1 kcal/mol are routinely achieved. The accuracy of these methods brings into

question the validity of the training data, particularly the accuracy of the labels, as well as potential

bias in the data. DFT is known to have large errors107-108, while the gold standard methods such as

coupled cluster with singles, doubles, and perturbative triples (CCSD(T)) are currently

prohibitively expensive for datasets of this size109. In order to overcome this deficit, transfer

learning has been utilized to fine tune these networks on smaller datasets of calculations performed

at significantly higher levels of theory, such as CCSD(T)63, 110. Additionally, bias in chemical

datasets is a well-known problem111-112. While there has been recent work to intelligently design

them using deep learning113, genetic algorithms114, or techniques such as query by committee115,

the large datasets required for chemical deep learning are largely restricted to small molecules

containing only carbon, nitrogen, oxygen, fluorine and hydrogen. As the coverage of chemical

space expands, it is critical that the datasets are intelligently designed to maximize coverage of the

rapidly expanding combinatorial space.

The final topic to address is interpretability. Deep learning has a reputation for being a ‘black

box’, as it is almost impossible to understand why the network made the decision that it did116.

Recent work has attempted to overcome this in chemical deep learning by cleverly designing the

architectures to allow for extraction of chemical insights from its decision making. In recent work

from Goh et al.117, by changing the information available to the network in their descriptor, they

were able to infer that the network was learning a different approach to solve different chemical

prediction challenges. Schutt et al. 102 on the other hand demonstrate not how the network is making

22
decisions, but rather that its predictions align with an understanding of chemical ideas such as

aromaticity.

Conformational Exploration. Regardless of how deep learning influences chemical property

mappings, effective exploration of chemical space involves navigating not only the species space,

but also the conformational space of those species. Conformational screening is an immense

challenge in chemistry, as with each new atom, multiple additional local minima appear on the

potential energy surface. The aforementioned neural network potentials offer a rapid way to

explore the conformational space of a molecule. The leading potential at the moment is the ANI-

1 potential that achieved errors below 1 kCal/mol and is trained using off-equilibrium geometries52.

The dataset it was trained on contains approximately 20 million energies of ~57,000 molecules in

different stances35.

The inverse of conformational screening is to develop a system that can generate equilibrium

conformers for a given molecule. This challenge has been undertaken by Gebauer et al.118 which

demonstrated deep learning’s ability to generate equilibrium conformers. This method is an

adaption of SchNet architecture developed by Schutt et al. that was able to regenerate molecular

geometries with a root mean squared deviation of approximately 0.4 Å. Additionally, a novel, but

not as rigorously tested method was introduced by Thomas et al.46 in which 3D point clouds were

used to regenerate molecule geometries. This work didn’t place the same emphasis on minima

structures, but was able to achieve very low errors of approximately 0.15 Å. This field of research

is still very young but holds immense potential to minimize the conformational screening

bottleneck.

MOLECULAR DESIGN

23
 The second stage of the idealized workflow is the problem of molecular design. This problem,

sometimes referred to as inverse QSPR has a history of machine learning applications including

Bayesian optimization119 and genetic algorithms120. Recent years have seen the application of

generative deep learning models to design molecules. One of the seminal demonstrations of this

method is the work of Gómez-Bombarelli et al.43 which used an autoencoder with a latent space

that was optimized by an additional network to reflect a particular property. This ‘landscape’ can

then be explored to identify candidate molecules that maximize the property. There are many other

approaches that also use autoencoders121-124, generative adversarial networks (GANs)70, or

reinforcement learning agents125-128 to navigate chemical space structured around a particular

property. Finally RNNs have also been used for molecular library generation by an adaptation of

their text generation capabilities129. An excellent review of molecular design is provided by

Sanchez-Lengeling and Aspuru-Guzik130.

General molecular design is seeing a surge of activity, however, there are two special classes of

molecules that deserve particular attention: materials and drugs. These are arguably the two most

challenging molecule classes to design and optimize, but also offer the greatest potential benefits.

Therefore, they have motivated significant research efforts with deep learning.

Materials Design. Many modern technologies such as batteries, aerospace, and renewable

energy relying of advanced materials. Deep learning has only recently begun to influence the field,

but there has been a rapid growth in applications in the last few years. The distinction between

discrete small molecules and crystalline structures has led to a separate set of convolutional

descriptors that seek to capture the crystalline structure. Crystal graph convolutional neural

networks (CGCNNs) as introduced by Xie and Grossman131 show much potential in this field.

24
There has been a push, however, to reconcile the representation systems for these two classes of

molecules. SchNet103 has been demonstrated on both, and MEGNet by Chen, et al.132 has been

developed for this specific purpose.

GCNNs, as well as the CGCNN variant have been used to predict the properties of bulk

materials133, predict thermoelectric properties134, optimize polymer properties135, and to explore

chemical materials space136. These applications are still young; however, it has moved beyond

predictive models as demonstrated in work by Li et al.135 in which they successfully used

reinforcement learning to train an agent to experimentally control polymer weight distributions,

and thus the polymer’s properties. Additionally, the exploration of chemical materials space by

Xie and Grossman136 demonstrated the potential of these methods to uncover previously

undetected pockets of materials space. Beyond the properties of materials, work has been done to

optimize their synthesis parameters137 and perform defect detection138. Finally, a deep learning

method that utilizes tensor networks, similarly to Schutt et al.24, demonstrated generative design

of chiral metamaterials139. Most of these applications remain theoretical in nature, and effectively

incorporating them with an experimental workflow, such as in the polymer optimization workflow

of Li et al.135 is a key next step to determine their efficacy.

One key subfield of materials design is catalysis design. Machine learning has seen increased

use in catalytic research140-141, however deep learning has seen limited application in this field due

to the limited data available, the unique nature of each catalytic process, and the difficulty of

representing multimolecular systems. The applications of deep learning within catalyst design

largely center around using neural network potentials to model the catalytic system. Recent

examples of this include Shakouri et al’s.142 work to model nitrogen gas on a ruthenium surface

and the optimization of platinum clusters by Zhai and Alexandrova143. Extending this work beyond

25
using neural network potentials will likely require increased data gathering efforts, as well as the

development of newer descriptors to describe interacting, multimolecular systems.

Drug Design. Drug design is arguably one of chemistry’s most important applications.

Fundamentally it involves identifying molecules that achieve a particular biological function with

maximum efficacy. These can either be obtained from natural sources or built from the ground up.

In either case the goal typically starts with one, or a set of molecules and the challenge is to

optimize its properties to improve potency, specificity, decrease side effects, and decrease

production costs. There are a number of reviews on deep learning’s impact on this field as it is of

great interest to the community25, 144-145.

The generative models in drug design follow the same trends as general molecular design, with

autoencoders146, GANs147, and reinforcement learning148 all being used to try and generate potent

drug molecules. In addition to these, there are some novel approaches to drug development rather

than molecules design that include predicting anti-cancer drug synergy149 and developing a

benchmarking for generative models in drug design150. Drug design approaches struggle from

limited data, possibly more so than any other fields due to the expense of obtaining it. Work by

Altae-tran et al.67 utilized one shot learning to address this deficiency and make informed

predictions about drug candidates with limited data. Finally, while not a molecule optimizing

generative system, work by Segler et al.151 developed methods to develop focused libraries of drug

candidates for screening using RNNs.

SYNTHESIS PLANNING
Synthesis planning is the final stage in this idealized workflow. It can be simplified into three

separate components. Retrosynthesis, in which the product is known, and is broken down into a

26
series of simpler starting materials from which it can be made. Reaction prediction, in which

reagents are known, and the dominant product must be determined. Finally, reaction optimization,

which involves taking a reaction with known reagents and products and trying to maximize the

yield or efficiency of this process. One important distinction to note here is that reaction

optimization and reaction prediction both have well established computational approaches, kinetic

models and quantum calculations respectively. Both of these can however be expensive, and in the

case of quantum calculations, enormously so.

Computational retrosynthesis on the other hand has a long and turbulent history. The original

retrosynthesis program was Pensak and Corey’s8 work on the LHASA software. From this point

there have been a multitude of assistive software packages152-154. The beginning of the 21st century

saw a loss of interest in this field due to a variety of factors, but it is largely attributed to a

widespread belief that computers could not capture the art of synthesis. This field has had a second

wind with the advent of deep learning, with the models beginning to challenge the notion of

computational inferiority in synthesis planning10.

Retrosynthesis. The great challenge of retrosynthesis is the exponential scaling of possible

moves in synthetic space from any point. This is a property it shares with traditional board games

such as Chess or Go. Formally, this can be expressed as a tree search, where the branching factor

is how many possible steps you can take from a particular point. The depth is how many steps it

takes to reach the desired position. Compared to the aforementioned games, retrosynthesis has a

significantly greater branching factor, but lower depth155. Retrosynthesis may present a far greater

challenge due to the immense challenging in knowing a priori whether a reaction will be successful

and produce the desired material, whereas Chess and Go have a perfectly defined set of possible

27
moves. However, these games represent a good starting point to consider the problem, and

fortunately, both have succumbed to artificial intelligence approaches. It is not surprising then that

one of the dominant displays of retrosynthetic AI was heavily inspired by AlphaGo, the seminal

AI system Deepmind developed to achieve superhuman Go playing ability3.

Work by Segler et al.16 adapted the AlphaGo methodology (Monte Carlo Tree Search with deep

neural network policy) to design a state of the art retrosynthetic AI. This system was trained on

over 12 million reactions from the Reaxys156 database and produced human accepted synthesis

routes. Assessing synthesis plans is a thorny challenge, and in order to do this, they performed a

double-blind study in which graduate chemists were shown the machine’s synthetic plan and the

original, literature plan. There was no statistically significant difference in their preferences, thus

giving a preliminary indication that its synthetic routes are ‘human level’. It is also possible,

however, to argue that the graduate chemists’ do not yet have the necessary expertise to distinguish

the human route. Thus, determining when computers achieve human ability in synthesis planning

is a decision that can only be made by the entire field. While this method showed great potential,

there are other avenues of research such as the use of RNNs in an encoder/decoder setup to perform

retrosynthetic analysis of small molecules157.

Computational retrosynthesis is making enormous strides; however, many problems persist.

Firstly, planning a retrosynthesis that looks valid, and experimentally verifying its predictions are

different challenges and until these methods are rigorously tested it is unknown whether or not

they are useful to chemists. This challenge would likely benefit from a user-friendly software

package in order to get chemists’ feedback on the computer-generated syntheses. These are

beginning to appear with an example being the ASKCOS software developed by the machine

learning for pharmaceutical discovery and synthesis consortium158.

28
 Reaction Prediction. Reaction prediction is the process of taking a set of known reagents and

conditions and predicting what products will form; as such, it typically requires greater exploration

into uncharted chemical space. Current methods to perform this, such as quantum calculations are

exceedingly expensive and thus limited to smaller molecules. Deep learning methods represent an

opportunity to alleviate this computational expense, and free up time of trained computational

chemists.

Reaction prediction exemplifies the challenge of predicting outliers, due to the frequent need to

predict outside of the training space. As a result of this, the majority of reaction prediction machine

learning methods either integrate the model with a physics based scheme or apply reaction

templates159. One of the early works that applied deep learning to reaction prediction involved

DNNs with molecular fingerprints to predict what product would form44. Additional work has

utilized RNN variants61, as well as more specialized architectures such as neural machine

translation160-161 and Siamese architectures (which take two identical networks given different

inputs and determine the similarity between them)17. One of the striking challenges for this field is

the immense literature bias towards successful reactions. Recently Coley et al.45 presented by a

clever approach to overcome this by recognizing that a successful reaction implicitly defines a

large number of unsuccessful reactions that can be added to the database. This was performed by

identifying high yielding reactions, and generating viable alternative products that are thus not

formed in high yield. These can then be added to the dataset to augment it with negative examples.

The current state of the art that also stresses interpretability uses a GCNN to predict reaction

outcomes in a manner similar to human intuition162.

29
 Due to deep learning’s relatively new arrival to reaction prediction, there is a history of non-

deep learning methods for reaction prediction that is reviewed by Coley et al.11. Current

developments are reaching a level that is competitive with humans. With further advancements in

predictive ability and transitioning it into user friendly software, this is likely to become a key

addition to the chemical toolset.

Reaction Optimization. Reaction optimization involves tuning the conditions of a reaction to

increase its efficiency. This is often performed via kinetic models, or experimentally through the

use of flow chemistry or high throughput combinatorial chemistry. Despite the maturity of these

methods, there is scope for a system which can rapidly produce idealized synthetic conditions

given a molecule and reaction type. Deep learning has the potential fill this niche, and research

has begun to adapt it to this challenge.

The potential of this approach was demonstrated by Zhou et al.74 in which an RNN variant

learned to optimize the conditions of reactions. Their model used an RNN that learned to evolve

the conditions of a reaction towards an optimized state. It was trained on simulated reactions and

then outperformed other software-based approaches for multiple experimental reaction setups. It

is important to acknowledge here that due to limited availability of data, and the need to flexibly

update the model, deep learning methods may not be the best choice here, instead a method that

uses alternative machine learning methodologies such as random forests has been demonstrated to

be a potent alternative163.

FUTURE DIRECTIONS
To summarize, deep learning is a subfield of machine learning that uses subsequent layers to

extract higher level features and use them to learn the patterns present in a dataset so as to predict

future behavior. Supervised learning requires large volumes of labelled data and a quantitatively

30
assessable goal or question. With this, a model uses an interplay of a predictive learner, evaluation,

and optimization, in the form of a training cycle to iteratively improve its performance until it

begins to overfit the training set, at which point training stops and the model is evaluated.

The last decade has seen explosive growth in the application of these methods across chemistry.

Through its applications, deep learning shows promise of being a game changer within chemistry.

This review has demonstrated that deep learning has and will continue to impact every stage of the

idealized chemistry workflow. Realization of its potential will require a concerted effort to address

the major challenges deep learning still faces, many of which have been discussed throughout this

review. The three main challenges that must be addressed to maximize the potential of this

technique within chemistry are:

1. Obtaining large amounts of high-quality data

2. Developing a standardized framework

3. Effective integration with the broader chemistry community

The first two challenges will be immensely benefitted by increased collaboration, and in

particular, continued open sourcing. The push for open sourcing has increased and there is strong

evidence of it occurring within deep learning particularly through software packages such as

DeepChem13, TensorMol164, ANI52, SchNetPack85, and chemprop165. Addressing the problem of

high-quality data also relies on continued advancements in physics based computational chemistry

and the accompanying software packages166-167.

The final challenge requires concerted action from specialists and the broader community. Open

sourcing software packages is a step in the right direction, but the chemical community has a long

history of resisting assistive software either due to poor usability or unreliable software

performance. The latter is demonstrably addressed by these powerful methods, but the former

31
requires conscious development of usable software packages with feedback from the community.

These methods are built to empower chemists first and foremost, and that must be a priority as this

field matures.

This review hopes to serve as a gateway to this burgeoning field and encourage chemists,

regardless of their specialization, to consider how deep learning could be applied to their work.

The following are a set of guidelines to assist in the initial application of these methods;

• Python has become the coding language of choice for deep learning and finding someone

proficient in it is invaluable.

• Deep learning requires large volumes of data to outperform traditional machine learning

methods. Unless transfer learning is an option, a few thousand data points is a minimum.

• To begin with, employ the open source software packages referenced above with default

settings to get a baseline.

• From this baseline, adapt the network architecture using techniques presented in the

literature referenced in this review to try and improve performance.

• Utilize the wealth of informative online courses and user-friendly software packages168-
169
provided by the deep learning community to aid in learning these techniques.

Deep learning’s contributions to chemistry to date demonstrate that it has a bright future within

chemistry, but through effective collaboration between specialists and the broader community, it

has the potential to offer a revolution.

AUTHOR INFORMATION

Corresponding Author

*Email: [email protected]

32
Author Contributions

The manuscript was written through contributions of all authors. All authors have given approval

to the final version of the manuscript.

Funding Sources

Australian Research Council (FL170100041)

ACKNOWLEDGMENT

MLC gratefully acknowledges an Australian Research Council Georgina Sweet Laureate

Fellowship, while ACM thanks the Australian National University and the Westpac Scholars

Trust for PhD scholarships

ABBREVIATIONS
Artificial Neural Networks – ANNs
Coupled Cluster Singles Doubles with perturbative triples - CCSD(T)
Crystal Graph Convolutional Neural Networks - CGCNN
Convolutional Neural Network – CNN
Graph Convolutional Neural Network – GCNN
Long Short Term Memory - LSTM
Recurrent Neural Network – RNN
Rectified Linear Unit - ReLU
Simplified Molecular Input Line Entry System – SMILES

33
REFERENCES

1. Krizhevsky, A.; Sutskever, I.; Hinton, G.E., ImageNet classification with deep
convolutional neural networks. In Proceedings of the 25th International Conference on Neural
Information Processing Systems - Volume 1, Curran Associates Inc.: Lake Tahoe, Nevada, 2012;
pp 1097-1105.
2. Graves, A., Generating Sequences With Recurrent Neural Networks. eprint
arXiv:1308.0850 2013, arXiv:1308.0850.
3. Silver, D.; Huang, A.; Maddison, C.J.; Guez, A.; Sifre, L.; van den Driessche, G.;
Schrittwieser, J.; Antonoglou, I.; Panneershelvam, V.; Lanctot, M.; Dieleman, S.; Grewe, D.;
Nham, J.; Kalchbrenner, N.; Sutskever, I.; Lillicrap, T.; Leach, M.; Kavukcuoglu, K.; Graepel, T.;
Hassabis, D., Mastering the game of Go with deep neural networks and tree search. Nature 2016,
529, 484-489.
4. Taigman, Y.; Yang, M.; Ranzato, M.; Wolf, L. In DeepFace: Closing the Gap to Human-
Level Performance in Face Verification, 2014 IEEE Conference on Computer Vision and Pattern
Recognition, 23-28 June 2014; 2014; pp 1701-1708.
5. Sutskever, I.; Vinyals, O.; Le, Q.V., Sequence to Sequence Learning with Neural Networks.
eprint arXiv:1409.3215 2014, arXiv:1409.3215.
6. Hinton, G.; Deng, L.; Yu, D.; Dahl, G.E.; Mohamed, A.; Jaitly, N.; Senior, A.; Vanhoucke,
V.; Nguyen, P.; Sainath, T.N.; Kingsbury, B., Deep Neural Networks for Acoustic Modeling in
Speech Recognition: The Shared Views of Four Research Groups. IEEE Signal Processing
Magazine 2012, 29, 82-97.
7. Szegedy, C.; Toshev, A.; Erhan, D., Deep neural networks for object detection. In
Proceedings of the 26th International Conference on Neural Information Processing Systems -
Volume 2, Curran Associates Inc.: Lake Tahoe, Nevada, 2013; pp 2553-2561.
8. Pensak, D.A.; Corey, E.J., LHASA—Logic and Heuristics Applied to Synthetic Analysis.
In Computer-Assisted Organic Synthesis, AMERICAN CHEMICAL SOCIETY: 1977; Vol. 61, pp
1-32.
9. Warr, W.A., A Short Review of Chemical Reaction Database Systems, Computer-Aided
Synthesis Design, Reaction Prediction and Synthetic Feasibility. Molecular Informatics 2014, 33,
469-476.
10. Cook, A.; Johnson, A.P.; Law, J.; Mirzazadeh, M.; Ravitz, O.; Simon, A., Computer-aided
synthesis design: 40 years on. Wiley Interdisciplinary Reviews: Computational Molecular Science
2012, 2, 79-107.
11. Coley, C.W.; Green, W.H.; Jensen, K.F., Machine Learning in Computer-Aided Synthesis
Planning. Accounts of Chemical Research 2018, 51, 1281-1289.
12. LeCun, Y.; Bengio, Y.; Hinton, G., Deep learning. Nature 2015, 521, 436.
13. Wu, Z.; Ramsundar, B.; Feinberg, E.N.; Gomes, J.; Geniesse, C.; Pappu, A.S.; Leswing,
K.; Pande, V. MoleculeNet: A Benchmark for Molecular Machine Learning ArXiv e-prints
[Online], 2017. https://ui.adsabs.harvard.edu/#abs/2017arXiv170300564W (accessed March 01,
2017).
14. Goh, G.B.; Hodas, N.O.; Vishnu, A., Deep learning for computational chemistry. Journal
of Computational Chemistry 2017, 38, 1291-1307.
15. Mayr, A.; Klambauer, G.; Unterthiner, T.; Hochreiter, S., DeepTox: Toxicity Prediction
using Deep Learning. Frontiers in Environmental Science 2016, 3.

34
16. Segler, M.H.S.; Preuss, M.; Waller, M.P., Planning chemical syntheses with deep neural
networks and symbolic AI. Nature 2018, 555, 604.
17. Fooshee, D.; Mood, A.; Gutman, E.; Tavakoli, M.; Urban, G.; Liu, F.; Huynh, N.; Van
Vranken, D.; Baldi, P., Deep learning for chemical reaction prediction. Molecular Systems Design
& Engineering 2018.
18. Lusci, A.; Pollastri, G.; Baldi, P., Deep Architectures and Deep Learning in
Chemoinformatics: The Prediction of Aqueous Solubility for Drug-Like Molecules. Journal of
Chemical Information and Modeling 2013, 53, 1563-1575.
19. Yao, K.; Herr, J.E.; Brown, S.N.; Parkhill, J., Intrinsic Bond Energies from a Bonds-in-
Molecules Neural Network. The Journal of Physical Chemistry Letters 2017, 8, 2689-2694.
20. Faber, F.A.; Hutchison, L.; Huang, B.; Gilmer, J.; Schoenholz, S.S.; Dahl, G.E.; Vinyals,
O.; Kearnes, S.; Riley, P.F.; von Lilienfeld, O.A., Prediction errors of molecular machine learning
models lower than hybrid DFT error. Journal of Chemical Theory and Computation 2017.
21. Ma, J.; Sheridan, R.P.; Liaw, A.; Dahl, G.E.; Svetnik, V., Deep Neural Nets as a Method
for Quantitative Structure–Activity Relationships. Journal of Chemical Information and Modeling
2015, 55, 263-274.
22. Duvenaud, D.; Maclaurin, D.; Aguilera-Iparraguirre, J.; Gómez-Bombarelli, R.; Hirzel, T.;
Aspuru-Guzik, A.; Adams, R.P., Convolutional Networks on Graphs for Learning Molecular
Fingerprints. eprint arXiv:1509.09292 2015, arXiv:1509.09292.
23. Kearnes, S.; McCloskey, K.; Berndl, M.; Pande, V.; Riley, P., Molecular graph
convolutions: moving beyond fingerprints. Journal of Computer-Aided Molecular Design 2016,
30, 595-608.
24. Schütt, K.T.; Arbabzadah, F.; Chmiela, S.; Müller, K.R.; Tkatchenko, A., Quantum-
chemical insights from deep tensor neural networks. Nature Communications 2017, 8, 13890.
25. Chen, H.; Engkvist, O.; Wang, Y.; Olivecrona, M.; Blaschke, T., The rise of deep learning
in drug discovery. Drug Discovery Today 2018, 23, 1241-1250.
26. Ekins, S., The Next Era: Deep Learning in Pharmaceutical Research. Pharmaceutical
research 2016, 33, 2594-603.
27. Butler, K.T.; Davies, D.W.; Cartwright, H.; Isayev, O.; Walsh, A., Machine learning for
molecular and materials science. Nature 2018, 559, 547-555.
28. Rupp, M., Machine learning for quantum mechanics in a nutshell. International Journal of
Quantum Chemistry 2015, 115, 1058-1073.
29. Lo, Y.-C.; Rensi, S.E.; Torng, W.; Altman, R.B., Machine learning in chemoinformatics
and drug discovery. Drug Discovery Today 2018.
30. Goodfellow, I.; Bengio, Y.; Courville, A., Deep Learning. MIT Press: 2016.
31. Le, Q.V.; Ranzato, M.A.; Monga, R.; Devin, M.; Chen, K.; Corrado, G.S.; Dean, J.; Ng,
A.Y., Building high-level features using large scale unsupervised learning. eprint arXiv:1112.6209
2011, arXiv:1112.6209.
32. Lowe, D.M. Extraction of Chemical Structures and Reactions from the Literature.
University of Cambridge, 2012.
33. Ramakrishnan, R.; Dral, P.O.; Rupp, M.; von Lilienfeld, O.A., Quantum chemistry
structures and properties of 134 kilo molecules. Scientific Data 2014, 1, 140022.
34. Mobley, D.L.; Guthrie, J.P., FreeSolv: a database of experimental and calculated hydration
free energies, with input files. Journal of computer-aided molecular design 2014, 28, 711-720.
35. Smith, J.S.; Isayev, O.; Roitberg, A.E., ANI-1, A data set of 20 million calculated off-
equilibrium conformations for organic molecules. Scientific Data 2017, 4, 170193.

35
36. Wang, Y.; Xiao, J.; Suzek, T.O.; Zhang, J.; Wang, J.; Zhou, Z.; Han, L.; Karapetyan, K.;
Dracheva, S.; Shoemaker, B.A.; Bolton, E.; Gindulyte, A.; Bryant, S.H., PubChem's BioAssay
Database. Nucleic acids research 2012, 40, D400-D412.
37. Domingos, P., A few useful things to know about machine learning. Commun. ACM 2012,
55, 78-87.
38. Spialter, L., The Atom Connectivity Matrix (ACM) and its Characteristic Polynomial
(ACMCP): A New Computer-Oriented Chemical Nomenclature. Journal of the American
Chemical Society 1963, 85, 2012-2013.
39. Rogers, D.; Hahn, M., Extended-Connectivity Fingerprints. Journal of Chemical
Information and Modeling 2010, 50, 742-754.
40. Weininger, D., SMILES, a chemical language and information system. 1. Introduction to
methodology and encoding rules. Journal of Chemical Information and Computer Sciences 1988,
28, 31-36.
41. Weininger, D.; Weininger, A.; Weininger, J.L., SMILES. 2. Algorithm for generation of
unique SMILES notation. Journal of Chemical Information and Computer Sciences 1989, 29, 97-
101.
42. Heller, S.; McNaught, A.; Stein, S.; Tchekhovskoi, D.; Pletnev, I., InChI - the worldwide
chemical structure identifier standard. Journal of Cheminformatics 2013, 5, 7.
43. Gómez-Bombarelli, R.; Wei, J.N.; Duvenaud, D.; Hernández-Lobato, J.M.; Sánchez-
Lengeling, B.; Sheberla, D.; Aguilera-Iparraguirre, J.; Hirzel, T.D.; Adams, R.P.; Aspuru-Guzik,
A., Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules.
ACS Central Science 2018, 4, 725-732.
44. Wei, J.N.; Duvenaud, D.; Aspuru-Guzik, A., Neural Networks for the Prediction of Organic
Chemistry Reactions. ACS Central Science 2016, 2, 725-732.
45. Coley, C.W.; Barzilay, R.; Jaakkola, T.S.; Green, W.H.; Jensen, K.F., Prediction of Organic
Reaction Outcomes Using Machine Learning. ACS Central Science 2017, 3, 434-443.
46. Thomas, N.; Smidt, T.; Kearnes, S.; Yang, L.; Li, L.; Kohlhoff, K.; Riley, P. Tensor field
networks: Rotation- and translation-equivariant neural networks for 3D point clouds ArXiv e-prints
[Online], 2018. https://ui.adsabs.harvard.edu/#abs/2018arXiv180208219T (accessed February 01,
2018).
47. Rupp, M.; Tkatchenko, A.; Müller, K.-R.; von Lilienfeld, O.A., Fast and Accurate
Modeling of Molecular Atomization Energies with Machine Learning. Physical Review Letters
2012, 108, 058301.
48. Staker, J.; Marshall, K.; Abel, R.; McQuaw, C., Molecular Structure Extraction From
Documents Using Deep Learning. eprint arXiv:1802.04903 2018, arXiv:1802.04903.
49. Rumelhart, D.E.; Hinton, G.E.; Williams, R.J., Learning representations by back-
propagating errors. Nature 1986, 323, 533.
50. Xavier, G.; Antoine, B.; Yoshua, B., Deep Sparse Rectifier Neural Networks. PMLR: 2011;
pp 315-323.
51. Raina, R.; Madhavan, A.; Ng, A.Y., Large-scale deep unsupervised learning using graphics
processors. In Proceedings of the 26th Annual International Conference on Machine Learning,
ACM: Montreal, Quebec, Canada, 2009; pp 873-880.
52. Smith, J.S.; Isayev, O.; Roitberg, A.E., ANI-1: an extensible neural network potential with
DFT accuracy at force field computational cost. Chemical Science 2017, 8, 3192-3203.
53. Gilmer, J.; Schoenholz, S.S.; Riley, P.F.; Vinyals, O.; Dahl, G.E., Neural Message Passing
for Quantum Chemistry. eprint arXiv:1704.01212 2017, arXiv:1704.01212.

36
54. Schütt, K.T.; Kindermans, P.-J.; Sauceda, H.E.; Chmiela, S.; Tkatchenko, A.; Müller, K.-
R., SchNet: A continuous-filter convolutional neural network for modeling quantum interactions.
eprint arXiv:1706.08566 2017, arXiv:1706.08566.
55. Cho, H.; Choi, I.S., Three-Dimensionally Embedded Graph Convolutional Network
(3DGCN) for Molecule Interpretation. eprint arXiv:1811.09794 2018, arXiv:1811.09794.
56. Goh, G.B.; Siegel, C.; Vishnu, A.; Hodas, N.O.; Baker, N. Chemception: A Deep Neural
Network with Minimal Chemistry Knowledge Matches the Performance of Expert-developed
QSAR/QSPR Models ArXiv e-prints [Online], 2017.
https://ui.adsabs.harvard.edu/#abs/2017arXiv170606689G (accessed June 01, 2017).
57. Hopfield, J.J., Neural networks and physical systems with emergent collective
computational abilities. Proceedings of the National Academy of Sciences 1982, 79, 2554.
58. Lipton, Z.C.; Berkowitz, J.; Elkan, C., A Critical Review of Recurrent Neural Networks
for Sequence Learning. eprint arXiv:1506.00019 2015, arXiv:1506.00019.
59. Graves, A.; Wayne, G.; Danihelka, I., Neural Turing Machines. eprint arXiv:1410.5401
2014, arXiv:1410.5401.
60. Hochreiter, S.; Schmidhuber, J., Long Short-term Memory. 1997; Vol. 9, p 1735-80.
61. Schwaller, P.; Gaudin, T.; Lanyi, D.; Bekas, C.; Laino, T., "Found in Translation":
Predicting Outcomes of Complex Organic Chemistry Reactions using Neural Sequence-to-
Sequence Models. eprint arXiv:1711.04810 2017, arXiv:1711.04810.
62. Pratt, L.Y., Discriminability-Based Transfer between Neural Networks. In Advances in
Neural Information Processing Systems 5, [NIPS Conference], Morgan Kaufmann Publishers Inc.:
1993; pp 204-211.
63. Smith, J.S.; Nebgen, B.T.; Zubatyuk, R.; Lubbers, N.; Devereux, C.; Barros, K.; Tretiak,
S.; Isayev, O.; Roitberg, A., Outsmarting Quantum Chemistry Through Transfer Learning. 2018.
64. Caruana, R., Multitask Learning. Machine Learning 1997, 28, 41-75.
65. Ramsundar, B.; Kearnes, S.; Riley, P.; Webster, D.; Konerding, D.; Pande, V., Massively
Multitask Networks for Drug Discovery. eprint arXiv:1502.02072 2015, arXiv:1502.02072.
66. Fei-Fei, L.; Fergus, R.; Perona, P., One-shot learning of object categories. IEEE
transactions on pattern analysis and machine intelligence 2006, 28, 594-611.
67. Altae-Tran, H.; Ramsundar, B.; Pappu, A.S.; Pande, V., Low Data Drug Discovery with
One-Shot Learning. ACS Central Science 2017, 3, 283-293.
68. Kingma, D.P.; Welling, M. Auto-Encoding Variational Bayes ArXiv e-prints [Online],
2013. https://ui.adsabs.harvard.edu/#abs/2013arXiv1312.6114K (accessed December 01, 2013).
69. Goodfellow, I.J.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.;
Courville, A.; Bengio, Y. Generative Adversarial Networks ArXiv e-prints [Online], 2014.
http://adsabs.harvard.edu/abs/2014arXiv1406.2661G (accessed June 1, 2014).
70. Sanchez-Lengeling, B.; Outeiral, C.; Guimaraes, G.L.; Aspuru-Guzik, A., Optimizing
distributions over molecular space. An Objective-Reinforced Generative Adversarial Network for
Inverse-design Chemistry (ORGANIC). 2017.
71. Wong, S.C.; Gatt, A.; Stamatescu, V.; McDonnell, M.D. In Understanding Data
Augmentation for Classification: When to Warp?, 2016 International Conference on Digital Image
Computing: Techniques and Applications (DICTA), 30 Nov.-2 Dec. 2016; 2016; pp 1-6.
72. Bjerrum, E.J. SMILES Enumeration as Data Augmentation for Neural Network Modeling
of Molecules ArXiv e-prints [Online], 2017.
https://ui.adsabs.harvard.edu/#abs/2017arXiv170307076J (accessed March 01, 2017).

37
73. Li, Y., Deep Reinforcement Learning: An Overview. eprint arXiv:1701.07274 2017,
arXiv:1701.07274.
74. Zhou, Z.; Li, X.; Zare, R.N., Optimizing Chemical Reactions with Deep Reinforcement
Learning. ACS Central Science 2017, 3, 1337-1344.
75. Längkvist, M.; Karlsson, L.; Loutfi, A., A review of unsupervised feature learning and deep
learning for time-series modeling. Pattern Recognition Letters 2014, 42, 11-24.
76. Behler, J., Atom-centered symmetry functions for constructing high-dimensional neural
network potentials. The Journal of Chemical Physics 2011, 134, 074106.
77. Behler, J., Neural network potential-energy surfaces in chemistry: a tool for large-scale
simulations. Physical Chemistry Chemical Physics 2011, 13, 17930-17955.
78. Zhang, L.; Han, J.; Wang, H.; Car, R.; E, W., Deep Potential Molecular Dynamics: A
Scalable Model with the Accuracy of Quantum Mechanics. Physical Review Letters 2018, 120,
143001.
79. McGibbon, R.T.; Taube, A.G.; Donchev, A.G.; Siva, K.; Hernández, F.; Hargus, C.; Law,
K.-H.; Klepeis, J.L.; Shaw, D.E., Improving the accuracy of Møller-Plesset perturbation theory
with neural networks. The Journal of Chemical Physics 2017, 147, 161725.
80. Mills, K.; Spanner, M.; Tamblyn, I., Deep learning and the Schrodinger equation. Physical
Review A 2017, 96, 042113.
81. Yao, K.; Parkhill, J., Kinetic Energy of Hydrocarbons as a Function of Electron Density
and Convolutional Neural Networks. Journal of Chemical Theory and Computation 2016, 12,
1139-1147.
82. Behler, J., First Principles Neural Network Potentials for Reactive Simulations of Large
Molecular and Condensed Systems. Angewandte Chemie International Edition 2017, 56, 12828-
12840.
83. Behler, J., Constructing high-dimensional neural network potentials: A tutorial review.
International Journal of Quantum Chemistry 2015, 115, 1032-1050.
84. Behler, J.; Parrinello, M., Generalized Neural-Network Representation of High-
Dimensional Potential-Energy Surfaces. Physical Review Letters 2007, 98, 146401.
85. Schütt, K.T.; Kessel, P.; Gastegger, M.; Nicoli, K.A.; Tkatchenko, A.; Müller, K.R.,
SchNetPack: A Deep Learning Toolbox For Atomistic Systems. Journal of Chemical Theory and
Computation 2019, 15, 448-455.
86. von Lilienfeld, O.A., Quantum Machine Learning in Chemical Compound Space.
Angewandte Chemie International Edition 2018, 57, 4164-4169.
87. Montavon, G.; Rupp, M.; Gobre, V.; Vazquez-Mayagoitia, A.; Hansen, K.; Tkatchenko, A.;
Müller, K.-R.; von Lilienfeld, O.A., Machine learning of molecular electronic properties in
chemical compound space. New Journal of Physics 2013, 15, 095003.
88. Hansen, K.; Biegler, F.; Ramakrishnan, R.; Pronobis, W.; von Lilienfeld, O.A.; Müller, K.-
R.; Tkatchenko, A., Machine Learning Predictions of Molecular Properties: Accurate Many-Body
Potentials and Nonlocality in Chemical Space. The Journal of Physical Chemistry Letters 2015, 6,
2326-2331.
89. Jain, A.; Ong, S.P.; Hautier, G.; Chen, W.; Richards, W.D.; Dacek, S.; Cholia, S.; Gunter,
D.; Skinner, D.; Ceder, G.; Persson, K.A., Commentary: The Materials Project: A materials
genome approach to accelerating materials innovation. APL Materials 2013, 1, 011002.
90. Gaulton, A.; Bellis, L.J.; Bento, A.P.; Chambers, J.; Davies, M.; Hersey, A.; Light, Y.;
McGlinchey, S.; Michalovich, D.; Al-Lazikani, B.; Overington, J.P., ChEMBL: a large-scale
bioactivity database for drug discovery. Nucleic Acids Res 2012, 40, D1100-7.

38
91. Kim, S.; Thiessen, P.A.; Bolton, E.E.; Chen, J.; Fu, G.; Gindulyte, A.; Han, L.; He, J.; He,
S.; Shoemaker, B.A.; Wang, J.; Yu, B.; Zhang, J.; Bryant, S.H., PubChem Substance and
Compound databases. Nucleic acids research 2016, 44, D1202-D1213.
92. Hughes, T.B.; Miller, G.P.; Swamidass, S.J., Modeling Epoxidation of Drug-like Molecules
with a Deep Machine Learning Network. ACS Central Science 2015, 1, 168-180.
93. Unterthiner, T.; Mayr, A.; Klambauer, G.; Steijaert, M.; Ceulemans, H.; Wegner, J.;
Hochreiter, S., Deep Learning as an Opportunity in Virtual Screening. 2014.
94. Dahl, G.E.; Jaitly, N.; Salakhutdinov, R., Multi-task Neural Networks for QSAR
Predictions. eprint arXiv:1406.1231 2014, arXiv:1406.1231.
95. Korotcov, A.; Tkachenko, V.; Russo, D.P.; Ekins, S., Comparison of Deep Learning With
Multiple Machine Learning Methods and Metrics Using Diverse Drug Discovery Data Sets.
Molecular Pharmaceutics 2017, 14, 4462-4475.
96. Unterthiner, T.; Mayr, A.; Klambauer, G.; Hochreiter, S., Toxicity Prediction using Deep
Learning. eprint arXiv:1503.01445 2015, arXiv:1503.01445.
97. Wenzel, J.; Matter, H.; Schmidt, F., Predictive Multitask Deep Neural Network Models for
ADME-Tox Properties: Learning from Large Data Sets. Journal of Chemical Information and
Modeling 2019.
98. Li, M.; Zhang, H.; Chen, B.; Wu, Y.; Guan, L., Prediction of pKa Values for Neutral and
Basic Drugs based on Hybrid Artificial Intelligence Methods. Scientific Reports 2018, 8, 3991.
99. Xu, Y.; Dai, Z.; Chen, F.; Gao, S.; Pei, J.; Lai, L., Deep Learning for Drug-Induced Liver
Injury. Journal of Chemical Information and Modeling 2015, 55, 2085-2093.
100. Goh, G.B.; Hodas, N.O.; Siegel, C.; Vishnu, A., SMILES2Vec: An Interpretable General-
Purpose Deep Neural Network for Predicting Chemical Properties. ArXiv e-prints 2017, 1712,
arXiv:1712.02034.
101. Jastrzębski, S.; Leśniak, D.; Czarnecki, W.M., Learning to SMILE(S). eprint
arXiv:1602.06289 2016, arXiv:1602.06289.
102. Schütt, K.T.; Gastegger, M.; Tkatchenko, A.; Müller, K.-R., Quantum-chemical insights
from interpretable atomistic neural networks. eprint arXiv:1806.10349 2018, arXiv:1806.10349.
103. Schütt, K.T.; Sauceda, H.E.; Kindermans, P.-J.; Tkatchenko, A.; Müller, K.-R., SchNet - a
deep learning architecture for molecules and materials. ArXiv e-prints 2017, 1712,
arXiv:1712.06113.
104. Wallach, I.; Dzamba, M.; Heifets, A. AtomNet: A Deep Convolutional Neural Network for
Bioactivity Prediction in Structure-based Drug Discovery ArXiv e-prints [Online], 2015.
https://ui.adsabs.harvard.edu/#abs/2015arXiv151002855W (accessed October 01, 2015).
105. Zeng, M.; Nitin Kumar, J.; Zeng, Z.; Savitha, R.; Ramaseshan Chandrasekhar, V.;
Hippalgaonkar, K., Graph Convolutional Neural Networks for Polymers Property Prediction.
eprint arXiv:1811.06231 2018, arXiv:1811.06231.
106. Coley, C.W.; Barzilay, R.; Green, W.H.; Jaakkola, T.S.; Jensen, K.F., Convolutional
Embedding of Attributed Molecular Graphs for Physical Property Prediction. Journal of Chemical
Information and Modeling 2017, 57, 1757-1772.
107. Wodrich, M.D.; Corminboeuf, C.; Schleyer, P.v.R., Systematic Errors in Computed Alkane
Energies Using B3LYP and Other Popular DFT Functionals. Organic Letters 2006, 8, 3631-3634.
108. Cohen, A.J.; Mori-Sánchez, P.; Yang, W., Challenges for Density Functional Theory.
Chemical Reviews 2012, 112, 289-320.
109. Purvis, G.D.; Bartlett, R.J., A full coupled-cluster singles and doubles model: The inclusion
of disconnected triples. The Journal of Chemical Physics 1982, 76, 1910-1918.

39
110. Goh, G.B.; Siegel, C.; Vishnu, A.; Hodas, N.O., Using Rule-Based Labels for Weak
Supervised Learning: A ChemNet for Transferable Chemical Property Prediction. ArXiv e-prints
2017, 1712, arXiv:1712.02734.
111. Ryan-Rhys, G.; Philippe, S.; Alpha, L., Dataset Bias in the Natural Sciences: A Case Study
in Chemical Reaction Prediction and Synthesis Design. 2018.
112. Swann, E.T.; Fernandez, M.; Coote, M.L.; Barnard, A.S., Bias-Free Chemically Diverse
Test Sets from Machine Learning. ACS Comb Sci 2017, 19, 544-554.
113. Segler, M.H.S.; Kogej, T.; Tyrchan, C.; Waller, M.P., Generating Focussed Molecule
Libraries for Drug Discovery with Recurrent Neural Networks. ArXiv e-prints 2017, 1701,
arXiv:1701.01329.
114. Browning, N.J.; Ramakrishnan, R.; von Lilienfeld, O.A.; Roethlisberger, U., Genetic
Optimization of Training Sets for Improved Machine Learning Models of Molecular Properties.
The Journal of Physical Chemistry Letters 2017, 8, 1351-1359.
115. Smith, J.S.; Nebgen, B.; Lubbers, N.; Isayev, O.; Roitberg, A.E., Less is more: Sampling
chemical space with active learning. The Journal of Chemical Physics 2018, 148, 241733.
116. Shwartz-Ziv, R.; Tishby, N., Opening the Black Box of Deep Neural Networks via
Information. eprint arXiv:1703.00810 2017, arXiv:1703.00810.
117. B. Goh, G.; Siegel, C.; Vishnu, A.; O. Hodas, N.; Baker, N., How Much Chemistry Does a
Deep Neural Network Need to Know to Make Accurate Predictions? 2017.
118. Gebauer, N.W.A.; Gastegger, M.; Schütt, K.T., Generating equilibrium molecules with
deep neural networks. eprint arXiv:1810.11347 2018, arXiv:1810.11347.
119. Ikebata, H.; Hongo, K.; Isomura, T.; Maezono, R.; Yoshida, R., Bayesian molecular design
with a chemical language model. Journal of Computer-Aided Molecular Design 2017, 31, 379-
391.
120. Kawai, K.; Nagata, N.; Takahashi, Y., De Novo Design of Drug-Like Molecules by a
Fragment-Based Molecular Evolutionary Approach. Journal of Chemical Information and
Modeling 2014, 54, 49-56.
121. Blaschke, T.; Olivecrona, M.; Engkvist, O.; Bajorath, J.; Chen, H., Application of
Generative Autoencoder in De Novo Molecular Design. Molecular Informatics 2018, 37, 1700123.
122. Jin, W.; Barzilay, R.; Jaakkola, T., Junction Tree Variational Autoencoder for Molecular
Graph Generation. eprint arXiv:1802.04364 2018, arXiv:1802.04364.
123. Dai, H.; Tian, Y.; Dai, B.; Skiena, S.; Song, L., Syntax-Directed Variational Autoencoder
for Structured Data. eprint arXiv:1802.08786 2018, arXiv:1802.08786.
124. Lim, J.; Ryu, S.; Kim, J.W.; Kim, W.Y., Molecular generative model based on conditional
variational autoencoder for de novo molecular design. eprint arXiv:1806.05805 2018,
arXiv:1806.05805.
125. Olivecrona, M.; Blaschke, T.; Engkvist, O.; Chen, H., Molecular de-novo design through
deep reinforcement learning. Journal of Cheminformatics 2017, 9, 48.
126. You, J.; Liu, B.; Ying, R.; Pande, V.; Leskovec, J., Graph Convolutional Policy Network
for Goal-Directed Molecular Graph Generation. eprint arXiv:1806.02473 2018,
arXiv:1806.02473.
127. Putin, E.; Asadulaev, A.; Ivanenkov, Y.; Aladinskiy, V.; Sanchez-Lengeling, B.; Aspuru-
Guzik, A.; Zhavoronkov, A., Reinforced Adversarial Neural Computer for de Novo Molecular
Design. Journal of Chemical Information and Modeling 2018, 58, 1194-1204.
128. Zhou, Z.; Kearnes, S.; Li, L.; Zare, R.N.; Riley, P., Optimization of Molecules via Deep
Reinforcement Learning. eprint arXiv:1810.08678 2018, arXiv:1810.08678.

40
129. Bjerrum, E.J.; Threlfall, R., Molecular Generation with Recurrent Neural Networks
(RNNs). eprint arXiv:1705.04612 2017, arXiv:1705.04612.
130. Sanchez-Lengeling, B.; Aspuru-Guzik, A., Inverse molecular design using machine
learning: Generative models for matter engineering. Science 2018, 361, 360.
131. Xie, T.; Grossman, J.C., Crystal Graph Convolutional Neural Networks for Accurate and
Interpretable Prediction of Material Properties. ArXiv e-prints 2017, 1710, arXiv:1710.10324.
132. Chen, C.; Ye, W.; Zuo, Y.; Zheng, C.; Ong, S.P., Graph Networks as a Universal Machine
Learning Framework for Molecules and Crystals. eprint arXiv:1812.05055 2018,
arXiv:1812.05055.
133. Jain, A.; Bligaard, T., Atomic-position independent descriptor for machine learning of
material properties. Physical Review B 2018, 98, 214112.
134. Laugier, L.; Bash, D.; Recatala, J.; Ng, H.K.; Ramasamy, S.; Foo, C.-S.; Chandrasekhar,
V.R.; Hippalgaonkar, K., Predicting thermoelectric properties from crystal graphs and material
descriptors - first application for functional materials. eprint arXiv:1811.06219 2018,
arXiv:1811.06219.
135. Li, H.; Collins, C.R.; Ribelli, T.G.; Matyjaszewski, K.; Gordon, G.J.; Kowalewski, T.;
Yaron, D.J., Tuning the molecular weight distribution from atom transfer radical polymerization
using deep reinforcement learning. Molecular Systems Design & Engineering 2018, 3, 496-508.
136. Xie, T.; Grossman, J.C., Hierarchical visualization of materials space with graph
convolutional neural networks. The Journal of Chemical Physics 2018, 149, 174111.
137. Kim, E.; Huang, K.; Jegelka, S.; Olivetti, E., Virtual screening of inorganic materials
synthesis parameters with deep learning. npj Computational Materials 2017, 3, 53.
138. Feng, S.; Zhou, H.; Dong, H., Using deep neural network with small dataset to predict
material defects. Materials & Design 2019, 162, 300-310.
139. Ma, W.; Cheng, F.; Liu, Y., Deep-Learning-Enabled On-Demand Design of Chiral
Metamaterials. ACS Nano 2018, 12, 6326-6334.
140. Kitchin, J.R., Machine learning in catalysis. Nature Catalysis 2018, 1, 230-232.
141. Goldsmith, B.R.; Esterhuizen, J.; Liu, J.-X.; Bartel, C.J.; Sutton, C., Machine learning for
heterogeneous catalyst design and discovery. AIChE Journal 2018, 64, 2311-2323.
142. Shakouri, K.; Behler, J.; Meyer, J.; Kroes, G.-J., Accurate Neural Network Description of
Surface Phonons in Reactive Gas–Surface Dynamics: N2 + Ru(0001). The Journal of Physical
Chemistry Letters 2017, 8, 2131-2136.
143. Zhai, H.; Alexandrova, A.N., Ensemble-Average Representation of Pt Clusters in
Conditions of Catalysis Accessed through GPU Accelerated Deep Neural Network Fitting Global
Optimization. Journal of Chemical Theory and Computation 2016, 12, 6213-6226.
144. Smith, J.S.; Roitberg, A.E.; Isayev, O., Transforming Computational Drug Discovery with
Machine Learning and AI. ACS Medicinal Chemistry Letters 2018, 9, 1065-1069.
145. Gawehn, E.; Hiss Jan, A.; Schneider, G., Deep Learning in Drug Discovery. Molecular
Informatics 2015, 35, 3-14.
146. Polykovskiy, D.; Zhebrak, A.; Vetrov, D.; Ivanenkov, Y.; Aladinskiy, V.; Mamoshina, P.;
Bozdaganyan, M.; Aliper, A.; Zhavoronkov, A.; Kadurin, A., Entangled Conditional Adversarial
Autoencoder for de Novo Drug Discovery. Molecular Pharmaceutics 2018, 15, 4398-4405.
147. Kadurin, A.; Nikolenko, S.; Khrabrov, K.; Aliper, A.; Zhavoronkov, A., druGAN: An
Advanced Generative Adversarial Autoencoder Model for de Novo Generation of New Molecules
with Desired Molecular Properties in Silico. Molecular Pharmaceutics 2017, 14, 3098-3104.

41
148. Popova, M.; Isayev, O.; Tropsha, A., Deep Reinforcement Learning for De-Novo Drug
Design. ArXiv e-prints 2017, 1711, arXiv:1711.10907.
149. Preuer, K.; Lewis, R.P.I.; Hochreiter, S.; Bender, A.; Bulusu, K.C.; Klambauer, G.,
DeepSynergy: predicting anti-cancer drug synergy with Deep Learning. Bioinformatics 2017,
btx806-btx806.
150. Preuer, K.; Renz, P.; Unterthiner, T.; Hochreiter, S.; Klambauer, G., Fréchet ChemNet
Distance: A Metric for Generative Models for Molecules in Drug Discovery. Journal of Chemical
Information and Modeling 2018, 58, 1736-1741.
151. Segler, M.H.S.; Kogej, T.; Tyrchan, C.; Waller, M.P., Generating Focused Molecule
Libraries for Drug Discovery with Recurrent Neural Networks. ACS Central Science 2018, 4, 120-
131.
152. Salatin, T.D.; Jorgensen, W.L., Computer-assisted mechanistic evaluation of organic
reactions. 1. Overview. The Journal of Organic Chemistry 1980, 45, 2043-2051.
153. Satoh, H.; Funatsu, K., SOPHIA, a Knowledge Base-Guided Reaction Prediction System
- Utilization of a Knowledge Base Derived from a Reaction Database. Journal of Chemical
Information and Computer Sciences 1995, 35, 34-44.
154. Socorro, I.M.; Goodman, J.M., The ROBIA Program for Predicting Organic Reactivity.
Journal of Chemical Information and Modeling 2006, 46, 606-614.
155. Segler, M.; Preuß, M.; Waller, M.P., Towards "AlphaChem": Chemical Synthesis Planning
with Tree Search and Deep Neural Network Policies. ArXiv e-prints 2017, 1702,
arXiv:1702.00020.
156. Elsevier Life Sciences, Reaxys. http://www.reaxys.com (accessed March 29, 2019).
157. Liu, B.; Ramsundar, B.; Kawthekar, P.; Shi, J.; Gomes, J.; Luu Nguyen, Q.; Ho, S.; Sloane,
J.; Wender, P.; Pande, V., Retrosynthetic Reaction Prediction Using Neural Sequence-to-Sequence
Models. ACS Central Science 2017, 3, 1103-1113.
158. Machine Learning for Pharmaceutical Discovery and Synthesis Symposium, ASKCOS.
http://askcos.mit.edu/ (accessed May 08, 2019).
159. Kayala, M.A.; Baldi, P., ReactionPredictor: Prediction of Complex Chemical Reactions at
the Mechanistic Level Using Machine Learning. Journal of Chemical Information and Modeling
2012, 52, 2526-2540.
160. Nam, J.; Kim, J., Linking the Neural Machine Translation and the Prediction of Organic
Chemistry Reactions. eprint arXiv:1612.09529 2016, arXiv:1612.09529.
161. Schwaller, P.; Laino, T.; Gaudin, T.; Bolgar, P.; Bekas, C.; Lee, A.A., Molecular
Transformer for Chemical Reaction Prediction and Uncertainty Estimation. eprint
arXiv:1811.02633 2018, arXiv:1811.02633.
162. Coley, C.W.; Jin, W.; Rogers, L.; Jamison, T.F.; Jaakkola, T.S.; Green, W.H.; Barzilay, R.;
Jensen, K.F., A graph-convolutional neural network model for the prediction of chemical
reactivity. Chemical Science 2019, 10, 370-377.
163. Daniel, R.; Gonçalo, B.; Tiago, R., Evolving and Nano Data Enabled Machine Intelligence
for Chemical Reaction Optimization. 2018.
164. Yao, K.; Herr, J.E.; Toth, David W.; McKintyre, R.; Parkhill, J., The TensorMol-0.1 model
chemistry: a neural network augmented with long-range physics. Chemical Science 2018, 9, 2261-
2269.
165. Yang, K.; Swanson, K.; Jin, W.; Coley, C.; Eiden, P.; Gao, H.; Guzman-Perez, A.; Hopper,
T.; Kelley, B.; Mathea, M.; Palmer, A.; Settels, V.; Jaakkola, T.; Jensen, K.; Barzilay, R. Are

42
Learned Molecular Representations Ready For Prime Time? arXiv e-prints [Online], 2019.
https://ui.adsabs.harvard.edu/abs/2019arXiv190401561Y (accessed April 01, 2019).
166. Gan, Z.; Epifanovsky, E.; Gilbert, A.T.B.; Wormit, M.; Kussmann, J.; Lange, A.W.; Behn,
A.; Deng, J.; Feng, X.; Ghosh, D.; Goldey, M.; Horn, P.R.; Jacobson, L.D.; Kaliman, I.; Khaliullin,
R.Z.; Kuś, T.; Landau, A.; Liu, J.; Proynov, E.I.; Rhee, Y.M.; Richard, R.M.; Rohrdanz, M.A.;
Steele, R.P.; Sundstrom, E.J.; Woodcock, H.L.; Zimmerman, P.M.; Zuev, D.; Albrecht, B.; Alguire,
E.; Austin, B.; Beran, G.J.O.; Bernard, Y.A.; Berquist, E.; Brandhorst, K.; Bravaya, K.B.; Brown,
S.T.; Casanova, D.; Chang, C.-M.; Chen, Y.; Chien, S.H.; Closser, K.D.; Crittenden, D.L.;
Diedenhofen, M.; DiStasio, R.A.; Do, H.; Dutoi, A.D.; Edgar, R.G.; Fatehi, S.; Fusti-Molnar, L.;
Ghysels, A.; Golubeva-Zadorozhnaya, A.; Gomes, J.; Hanson-Heine, M.W.D.; Harbach, P.H.P.;
Hauser, A.W.; Hohenstein, E.G.; Holden, Z.C.; Jagau, T.-C.; Ji, H.; Kaduk, B.; Khistyaev, K.; Kim,
J.; Kim, J.; King, R.A.; Klunzinger, P.; Kosenkov, D.; Kowalczyk, T.; Krauter, C.M.; Lao, K.U.;
Laurent, A.D.; Lawler, K.V.; Levchenko, S.V.; Lin, C.Y.; Liu, F.; Livshits, E.; Lochan, R.C.;
Luenser, A.; Manohar, P.; Manzer, S.F.; Mao, S.-P.; Mardirossian, N.; Marenich, A.V.; Maurer,
S.A.; Mayhall, N.J.; Neuscamman, E.; Oana, C.M.; Olivares-Amaya, R.; O’Neill, D.P.; Parkhill,
J.A.; Perrine, T.M.; Peverati, R.; Prociuk, A.; Rehn, D.R.; Rosta, E.; Russ, N.J.; Sharada, S.M.;
Sharma, S.; Small, D.W.; Sodt, A.; Stein, T.; Stück, D.; Su, Y.-C.; Thom, A.J.W.; Tsuchimochi, T.;
Vanovschi, V.; Vogt, L.; Vydrov, O.; Wang, T.; Watson, M.A.; Wenzel, J.; White, A.; Williams,
C.F.; Yang, J.; Yeganeh, S.; Yost, S.R.; You, Z.-Q.; Zhang, I.Y.; Zhang, X.; Zhao, Y.; Brooks, B.R.;
Chan, G.K.L.; Chipman, D.M.; Cramer, C.J.; Goddard, W.A.; Gordon, M.S.; Hehre, W.J.; Klamt,
A.; Schaefer, H.F.; Schmidt, M.W.; Sherrill, C.D.; Truhlar, D.G.; Warshel, A.; Xu, X.; Aspuru-
Guzik, A.; Baer, R.; Bell, A.T.; Besley, N.A.; Chai, J.-D.; Dreuw, A.; Dunietz, B.D.; Furlani, T.R.;
Gwaltney, S.R.; Hsu, C.-P.; Jung, Y.; Kong, J.; Lambrecht, D.S.; Liang, W.; Ochsenfeld, C.;
Rassolov, V.A.; Slipchenko, L.V.; Subotnik, J.E.; Van Voorhis, T.; Herbert, J.M.; Krylov, A.I.; Gill,
P.M.W.; Head-Gordon, M., Advances in molecular quantum chemistry contained in the Q-Chem
4 program package AU - Shao, Yihan. Molecular Physics 2015, 113, 184-215.
167. Frisch, M.J.; Trucks, G.W.; Schlegel, H.B.; Scuseria, G.E.; Robb, M.A.; Cheeseman, J.R.;
Scalmani, G.; Barone, V.; Petersson, G.A.; Nakatsuji, H.; Li, X.; Caricato, M.; Marenich, A.V.;
Bloino, J.; Janesko, B.G.; Gomperts, R.; Mennucci, B.; Hratchian, H.P.; Ortiz, J.V.; Izmaylov, A.F.;
Sonnenberg, J.L.; Williams; Ding, F.; Lipparini, F.; Egidi, F.; Goings, J.; Peng, B.; Petrone, A.;
Henderson, T.; Ranasinghe, D.; Zakrzewski, V.G.; Gao, J.; Rega, N.; Zheng, G.; Liang, W.; Hada,
M.; Ehara, M.; Toyota, K.; Fukuda, R.; Hasegawa, J.; Ishida, M.; Nakajima, T.; Honda, Y.; Kitao,
O.; Nakai, H.; Vreven, T.; Throssell, K.; Montgomery Jr., J.A.; Peralta, J.E.; Ogliaro, F.; Bearpark,
M.J.; Heyd, J.J.; Brothers, E.N.; Kudin, K.N.; Staroverov, V.N.; Keith, T.A.; Kobayashi, R.;
Normand, J.; Raghavachari, K.; Rendell, A.P.; Burant, J.C.; Iyengar, S.S.; Tomasi, J.; Cossi, M.;
Millam, J.M.; Klene, M.; Adamo, C.; Cammi, R.; Ochterski, J.W.; Martin, R.L.; Morokuma, K.;
Farkas, O.; Foresman, J.B.; Fox, D.J. Gaussian 16 Rev. B.01, Wallingford, CT, 2016.
168. Abadi, M.; Barham, P.; Chen, J.; Chen, Z.; Davis, A.; Dean, J.; Devin, M.; Ghemawat, S.;
Irving, G.; Isard, M.; Kudlur, M.; Levenberg, J.; Monga, R.; Moore, S.; Murray, D.G.; Steiner, B.;
Tucker, P.A.; Vasudevan, V.; Warden, P.; Wicke, M.; Yu, Y.; Zheng, X., TensorFlow: A System for
Large-Scale Machine Learning. 2016; pp 265-283.
169. Jia, Y.; Shelhamer, E.; Donahue, J.; Karayev, S.; Long, J.; Girshick, R.; Guadarrama, S.;
Darrell, T., Caffe: Convolutional Architecture for Fast Feature Embedding. eprint
arXiv:1408.5093 2014, arXiv:1408.5093.

43
For table of contents use:

44