Chemical structure extraction from documents remains a hard problem due to both false positive identification of structures during segmentation and errors in the predicted structures. Current approaches rely on handcrafted rules and subroutines that perform reasonably well generally, but still routinely encounter situations where recognition rates are not yet satisfactory and systematic improvement is challenging. Complications impacting performance of current approaches include the diversity in visual styles used by various software to render structures, the frequent use of ad hoc annotations, and other challenges related to image quality, including resolution and noise. We here present end-to-end deep learning solutions for both segmenting molecular structures from documents and for predicting chemical structures from these segmented images. This deep learning-based approach does not require any handcrafted features, is learned directly from data, and is robust against variations in image quality and style. Using the deep-learning approach described herein we show that it is possible to perform well on both segmentation and prediction of low resolution images containing moderately sized molecules found in journal articles and patents.

Materials exhibiting higher mobilities than conventional organic semiconducting materials such as fullerenes and fused thiophenes are in high demand for applications in printed electronics. To discover new molecules in the heteroacene family that might show improved hole mobility, three de novo design methods were applied. Machine learning (ML) models were generated based on previously calculated hole reorganization energies of a quarter million examples of heteroacenes, where the energies were calculated by applying density functional theory (DFT) and a massive cloud computing environment. The three generative methods applied were (1) the continuous space method, where molecular structures are converted into continuous variables by applying the variational autoencoder/decoder technique; (2) the method based on reinforcement learning of SMILES strings (the REINVENT method); and (3) the junction tree variational autoencoder method that directly generates molecular graphs. Among the three methods, the second and third methods succeeded in obtaining chemical structures whose DFT-calculated hole reorganization energy was lower than the lowest energy in the training dataset. This suggests that an extrapolative materials design protocol can be developed by applying generative modeling to a quantitative structure−property relationship (QSPR) utility function.

The hit identification process usually involves the profiling of millions to more recently billions of compounds either via traditional experimental high-throughput screens (HTS) or computational virtual high-throughput screens (vHTS). We have previously demonstrated that, by coupling reaction-based enumeration, active learning, and free energy calculations, a similarly largescale exploration of chemical space can be extended to the hit-tolead process. In this work, we augment that approach by coupling large scale enumeration and cloud-based free energy perturbation (FEP) profiling with goal-directed generative machine learning, which results in a higher enrichment of potent ideas compared to large scale enumeration alone, while simultaneously staying within the bounds of predefined drug-like property space. We can achieve this by building the molecular distribution for generative machine learning from the PathFinder rules-based enumeration and optimizing for a weighted sum QSAR-based multiparameter optimization function. We examine the utility of this combined approach by designing potent inhibitors of cyclin-dependent kinase 2 (CDK2) and demonstrate a coupled workflow that can (1) provide a 6.4-fold enrichment improvement in identifying <10 nM compounds over random selection and a 1.5-fold enrichment in identifying <10 nM compounds over our previous method, (2) rapidly explore relevant chemical space outside the bounds of commercial reagents, (3) use generative ML approaches to "learn" the SAR from large scale in silico enumerations and generate novel idea molecules for a flexible receptor site that are both potent and within relevant physicochemical space, and (4) produce over 3 000 000 idea molecules and run 1935 FEP simulations, identifying 69 ideas with a predicted IC 50 < 10 nM and 358 ideas with a predicted IC 50 < 100 nM. The reported data suggest combining both reaction-based and generative machine learning for ideation results in a higher enrichment of potent compounds over previously described approaches and has the potential to rapidly accelerate the discovery of novel chemical matter within a predefined potency and property space.

The hit identification process usually involves the profiling of millions to more recently billions of compounds either via traditional experimental high throughput screens (HTS) or computational virtual high throughput screens (vHTS). We have previously demonstrated that by coupling reaction-based enumeration, active learning and free energy calculations, a similarly large-scale exploration of chemical space can be extended to the hit-to-lead process. In this work, we augment that approach by coupling large scale enumeration and cloud-based FEP profiling with goal-directed generative machine learning, which results in a higher enrichment of potent ideas compared to large scale enumeration alone, while simultaneously staying within the bounds of a predefined drug-like property space. We are able to achieve this by building the molecular distribution for generative machine learning from the PathFinder rules-based enumeration and optimizing for a weighted sum QSAR based multi-parameter optimization function. We examine the utility of this combined approach by designing potent inhibitors of cyclin-dependent kinase 2 (CDK2) and demonstrate a coupled workflow that can: (1) provide a 6.4 fold enrichment improvement in identifying < 10nM compounds over random selection, and a 1.5 fold enrichment in identifying < 10nM compounds over our previous method (2) rapidly explore relevant chemical space outside the bounds of commercial reagents, (3) use generative ML approaches to “learn” the SAR from large scale in silico enumerations and generate novel idea molecules for a flexible receptor site that are both potent and within relevant physicochemical space and (4) produce over 3,000,000 idea molecules and run 2153 FEP simulations, identifying 69 ideas with a predicted IC₅₀ < 10nM and 358 ideas with a predicted IC₅₀ <100 nM. The reported data suggest combining both reaction-based and generative machine learning for ideation results in a higher enrichment of potent compounds over previously described approaches, and can rapidly accelerate the discovery of novel chemical matter within a predefined potency and property space.<br>

The past few years have seen a significantly increased interest in applying contemporary machine learning methods to drug discovery, materials science, and other applications in chemistry. Recent advances in deep learning, coupled with the ever-expanding volume of publicly available data, have enabled a breadth of new directions to explore, both in accelerating commercial applications and in enabling new research directions. Many machine learning methods cannot utilize molecule data stored in common formats, e.g., SMILES or connection table, and first require molecules to be descriptorized and processed into representations amenable to machine learning. Historically, molecular featurization has been performed through non-learned transformations that are usually coarse-grained and highly lossy, such as molecular fingerprints that encounter bit collisions and discard the overall molecular topology. By contrast, learned featurization may provide richer, more descriptive representations of molecules, leading to more powerful and accurate models. We compare common non-learned featurization methods with those that are learned and explore the different families of deep neural architectures used to obtain learned representations. We also discuss recent work that explores the addition of constraints to models that induce stronger physical priors in deep neural network architectures. Imposing physical constraints in neural models can lead to more robust featurizations and improved transfer learning.