Machine Learning Papers: Molecules

Efficient and accurate prediction of molecular properties, such as lipophilicity and solubility, is highly desirable for rational compound design in chemical and pharmaceutical industries. To this end, we build and apply a graph-neural-network framework called self-attention-based message-passing neural network (SAMPN) to study the relationship between chemical properties and structures in an interpretable way. The main advantages of SAMPN are that it directly uses chemical graphs and breaks the black-box mold of many machine/deep learning methods. Specifically, its attention mechanism indicates the degree to which each atom of the molecule contributes to the property of interest, and these results are easily visualized. Further, SAMPN outperforms random forests and the deep learning framework MPN from Deepchem. In addition, another formulation of SAMPN (Multi-SAMPN) can simultaneously predict multiple chemical properties with higher accuracy and efficiency than other models that predict one specific chemical property. Moreover, SAMPN can generate chemically visible and interpretable results, which can help researchers discover new pharmaceuticals and materials. The source code of the SAMPN prediction pipeline is freely available at Github (https://github.com/tbwxmu/SAMPN).

Background
There has been increasing interest in the use of deep neural networks for de novo design of molecules with desired properties. A common approach is to train a generative model on SMILES strings and then use this to generate SMILES strings for molecules with a desired property. Unfortunately, these SMILES strings are often not syntactically valid due to elements of SMILES syntax that must occur in pairs.
 
Results
We describe a SMILES-like syntax called DeepSMILES that addresses two of the main reasons for invalid syntax when using a probabilistic model to generate SMILES strings. The DeepSMILES syntax avoids the problem of unbalanced parentheses by only using close parentheses, where the number of parentheses indicates the branch length. In addition, DeepSMILES avoids the problem of pairing ring closure symbols by using only a single symbol at the ring closing location, where the symbol indicates the ring size. We show that this syntax can be interconverted to/from SMILES with string processing without any loss of information, including stereo configuration.
 
Conclusion
We believe that DeepSMILES will be useful, not just for those using SMILES in deep neural networks, but also for other computational methods that use SMILES as the basis for generating molecular structures such as genetic algorithms.

Deep generative models for graph-structured data offer a new angle on the problem of chemical synthesis: by optimizing differentiable models that directly generate molecular graphs, it is possible to side-step expensive search procedures in the discrete and vast space of chemical structures. We introduce MolGAN, an implicit, likelihood-free generative model for small molecular graphs that circumvents the need for expensive graph matching procedures or node ordering heuristics of previous likelihood-based methods. Our method adapts generative adversarial networks (GANs) to operate directly on graph-structured data. We combine our approach with a reinforcement learning objective to encourage the generation of molecules with specific desired chemical properties. In experiments on the QM9 chemical database, we demonstrate that our model is capable of generating close to 100% valid compounds. MolGAN compares favorably both to recent proposals that use string-based (SMILES) representations of molecules and to a likelihood-based method that directly generates graphs, albeit being susceptible to mode collapse.

The identification of novel drug-target (DT) interactions is a substantial part of the drug discovery process. Most of the computational methods that have been proposed to predict DT interactions have focused on binary classification, where the goal is to determine whether a DT pair interacts or not. However, protein-ligand interactions assume a continuum of binding strength values, also called binding affinity and predicting this value still remains a challenge. The increase in the affinity data available in DT knowledge-bases allows the use of advanced learning techniques such as deep learning architectures in the prediction of binding affinities. In this study, we propose a deep-learning based model that uses only sequence information of both targets and drugs to predict DT interaction binding affinities. The few studies that focus on DT binding affinity prediction use either 3D structures of protein-ligand complexes or 2D features of compounds. One novel approach used in this work is the modeling of protein sequences and compound 1D representations with convolutional neural networks (CNNs). The results show that the proposed deep learning based model that uses the 1D representations of targets and drugs is an effective approach for drug target binding affinity prediction. The model in which high-level representations of a drug and a target are constructed via CNNs achieved the best Concordance Index (CI) performance in one of our larger benchmark data sets, outperforming the KronRLS algorithm and SimBoost, a state-of-the-art method for DT binding affinity prediction.

Background
Previous studies have suggested deep learning to be a highly effective approach for screening lead compounds for new drugs. Several deep learning models have been developed by addressing the use of various kinds of fingerprints and graph convolution architectures. However, these methods are either advantageous or disadvantageous depending on whether they (1) can distinguish structural differences including chirality of compounds, and (2) can automatically discover effective features.

Results
We developed another deep learning model for compound classification. In this method, we constructed a distributed representation of compounds based on the SMILES notation, which linearly represents a compound structure, and applied the SMILES-based representation to a convolutional neural network (CNN). The use of SMILES allows us to process all types of compounds while incorporating a broad range of structure information, and representation learning by CNN automatically acquires a low-dimensional representation of input features. In a benchmark experiment using the TOX 21 dataset, our method outperformed conventional fingerprint methods, and performed comparably against the winning model of the TOX 21 Challenge. Multivariate analysis confirmed that the chemical space consisting of the features learned by SMILES-based representation learning adequately expressed a richer feature space that enabled the accurate discrimination of compounds. Using motif detection with the learned filters, not only important known structures (motifs) such as protein-binding sites but also structures of unknown functional groups were detected.

The discovery of new materials can bring enormous societal and technological progress. In this context, exploring completely the large space of potential materials is computationally intractable. Here, we review methods for achieving inverse design, which aims to discover tailored materials from the starting point of a particular desired functionality. Recent advances from the rapidly growing field of artificial intelligence, mostly from the subfield of machine learning, have resulted in a fertile exchange of ideas, where approaches to inverse molecular design are being proposed and employed at a rapid pace. Among these, deep generative models have been applied to numerous classes of materials: rational design of prospective drugs, synthetic routes to organic compounds, and optimization of photovoltaics and redox flow batteries, as well as a variety of other solid-state materials.

Simplified Molecular Input Line Entry System (SMILES) is a single line text representation of a unique molecule. One molecule can however have multiple SMILES strings, which is a reason that canonical SMILES have been defined, which ensures a one to one correspondence between SMILES string and molecule. Here the fact that multiple SMILES represent the same molecule is explored as a technique for data augmentation of a molecular QSAR dataset modeled by a long short term memory (LSTM) cell based neural network. The augmented dataset was 130 times bigger than the original. The network trained with the augmented dataset shows better performance on a test set when compared to a model built with only one canonical SMILES string per molecule. The correlation coefficient R2 on the test set was improved from 0.56 to 0.66 when using SMILES enumeration, and the root mean square error (RMS) likewise fell from 0.62 to 0.55. The technique also works in the prediction phase. By taking the average per molecule of the predictions for the enumerated SMILES a further improvement to a correlation coefficient of 0.68 and a RMS of 0.52 was found.

We introduce a convolutional neural network that operates directly on graphs. These networks allow end-to-end learning of prediction pipelines whose inputs are graphs of arbitrary size and shape. The architecture we present generalizes standard molecular feature extraction methods based on circular fingerprints. We show that these data-driven features are more interpretable, and have better predictive performance on a variety of tasks.

One of the most successful methods for predicting the properties of chemical compounds is the quantitative structure–activity relationship (QSAR) methods. The prediction accuracy of QSAR models has recently been greatly improved by employing deep learning technology. Especially, newly developed molecular featurizers based on graph convolution operations on molecular graphs significantly outperform the conventional extended connectivity fingerprints (ECFP) feature in both classification and regression tasks, indicating that it is critical to develop more effective new featurizers to fully realize the power of deep learning techniques. Motivated by the fact that there is a clear analogy between chemical compounds and natural languages, this work develops a new molecular featurizer, FP2VEC, which represents a chemical compound as a set of trainable embedding vectors.

We seek to automate the design of molecules based on specific chemical properties. In computational terms, this task involves continuous embedding and generation of molecular graphs. Our primary contribution is the direct realization of molecular graphs, a task previously approached by generating linear SMILES strings instead of graphs. Our junction tree variational autoencoder generates molecular graphs in two phases, by first generating a tree-structured scaffold over chemical substructures, and then combining them into a molecule with a graph message passing network. This approach allows us to incrementally expand molecules while maintaining chemical validity at every step. We evaluate our model on multiple tasks ranging from molecular generation to optimization. Across these tasks, our model outperforms previous state-of-the-art baselines by a significant margin.

In the last few years, we have seen the transformative impact of deep learning in many applications, particularly in speech recognition and computer vision. Inspired by Google’s Inception-ResNet deep convolutional neural network (CNN) for image classification, we have developed “Chemception”, a deep CNN for the prediction of chemical properties, using just the images of 2D drawings of molecules. We develop Chemception without providing any additional explicit chemistry knowledge, such as basic concepts like periodicity, or advanced features like molecular descriptors and fingerprints. We then show how Chemception can serve as a general-purpose neural network architecture for predicting toxicity, activity, and solvation properties when trained on a modest database of 600 to 40,000 compounds. When compared to multi-layer perceptron (MLP) deep neural networks trained with ECFP fingerprints, Chemception slightly outperforms in activity and solvation prediction and slightly underperforms in toxicity prediction. Having matched the performance of expert-developed QSAR/QSPR deep learning models, our work demonstrates the plausibility of using deep neural networks to assist in computational chemistry research, where the feature engineering process is performed primarily by a deep learning algorithm.

Shallow machine learning methods have been applied to chemoinformatics problems with some success. As more data becomes available and more complex problems are tackled, deep machine learning methods may also become useful. Here, we present a brief overview of deep learning methods and show in particular how recursive neural network approaches can be applied to the problem of predicting molecular properties. However, molecules are typically described by undirected cyclic graphs, while recursive approaches typically use directed acyclic graphs. Thus, we develop methods to address this discrepancy, essentially by considering an ensemble of recursive neural networks associated with all possible vertex-centered acyclic orientations of the molecular graph. One advantage of this approach is that it relies only minimally on the identification of suitable molecular descriptors because suitable representations are learned automatically from the data. Several variants of this approach are applied to the problem of redicting aqueous solubility and tested on four benchmark data sets. Experimental results show that the performance of the deep learning methods matches or exceeds the performance of other state-of-the-art methods according to several evaluation metrics and expose the fundamental limitations arising from training sets that are too small or too noisy. A Web-based predictor, AquaSol, is available online through the ChemDB portal (cdb.ics.uci.edu) together with additional material.