Stephen D. Pickett

We present a rule-based approach for the generation of new molecules, which derives combination rules between molecular fragments from a user-provided database, and is useful for constrained optimization at defined vectors. The molecules generated are statistically similar to the database provided. Compared to existing machine-learning approaches this affords the user a greater level of control of the molecules generated, which is a common desideratum in drug discovery programs. This is complemented by the use of a build probability score for each generated molecule, which corresponds to the likelihood of its existence in a hypothetical exhaustive enumeration of the source database fragments. The method is applied to the generation of derivatives of a 3,5-dimethyl-4-phenylisoxazole parent structure, a core of current interest for the development of inhibitors of bromodomain-containing protein 4.

Generative design and machine learning are increasingly prevalent in medicinal chemistry. To pilot the comprehensive use of automated molecular design on a project, the BRADSHAW platform was used to optimize a series of inhibitors of Endoplasmic Reticulum Aminopeptidase 1 (ERAP1), an emerging target in cancer immunotherapy and autoimmune diseases. Through four consecutive iterations applying in silico molecular generation, property prediction and filtering, we conducted a multiparameter optimization of potency, physicochemical properties and pharmacokinetics. Continuous refinement of Machine Learning (ML) models led to improved scoring accuracy and compound quality, culminating in identification of in vitro and in vivo tool molecules. We also discuss our reflections on the pilot and integration of automated design into medicinal chemistry projects, including observations of the human factors resulting from increased use of computational design, and recommendations for future projects.

We present SimDMTA, an in silico framework designed to simulate the Design-Make-Test-Analyze (DMTA) cycle used in preclinical drug discovery. Using docking scores as a proxy for biological assays, the simulations allow factors controlling the efficiency of the DMTA cycle to be explored in a manner that would not be feasible using traditional experiments due to time and cost constraints. In this workflow, a machine learning model predicts docking scores, selects compounds using various query strategies, docks selected molecules, and retrains iteratively. Starting from a broad chemical space, the model actively samples molecules derived from a 3,5-dimethyl-4-phenylisoxazole scaffold, an active warhead for the Bromodomain 4 (BRD4) BD1 binding site, to refine its predictions. Our results show that uncertainty-based sampling significantly outperforms greedy and hybrid approaches in both hit discovery and the ability of the model that predicts docking scores to generalize beyond its training set. Notably, by the final iteration, 37 of the top 50 ranked compounds were within the top 1% of the chemical space of all evaluated compounds. Strategies that include some random selection correct systematic biases more rapidly, but are less effective at predicting top-performing molecules. These findings underscore the value of incorporating molecular diversity and uncertainty into design strategies. While such strategies may deprioritize those molecules with the highest absolute predictions in early rounds, they markedly accelerate model refinement, ultimately leading to more effective hit identification in discovery driven by active learning.

To understand the benefits and drawbacks of 3D combinatorial and deep learning generators, a novel benchmark was created focusing on the recreation of important protein-ligand interactions and 3D ligand conformations. Using the BindingMOAD data set with a hold-out blind set, the sequential graph neural network generators, Pocket2Mol and PocketFlow, diffusion models, DiffSBDD and MolSnapper, and combinatorial genetic algorithms, AutoGrow4 and LigBuilderV3, were evaluated. It was discovered that deep learning methods fail to generate structurally valid molecules and 3D conformations, whereas combinatorial methods are slow and generate molecules that are prone to failing 2D MOSES filters. The results from this evaluation guide us toward improving deep learning structure-based generators by placing higher importance on structural validity, 3D ligand conformations, and recreation of important known active site interactions. This benchmark should be used to understand the limitations of future combinatorial and deep learning generators. The package is freely available under an Apache 2.0 license at github.com/gskcheminformatics/SBDD-benchmarking.

Predicting reaction yields in synthetic chemistry remains a significant challenge. This study systematically evaluates the impact of tokenization, molecular representation, pretraining data, and adversarial training on a BERT-based model for yield prediction of Buchwald-Hartwig and Suzuki-Miyaura coupling reactions using publicly available HTE data sets. We demonstrate that molecular representation choice (SMILES, DeepSMILES, SELFIES, Morgan fingerprint-based notation, IUPAC names) has minimal impact on model performance, while typically BPE and SentencePiece tokenization outperform other methods. WordPiece is strongly discouraged for SELFIES and fingerprint-based notation. Furthermore, pretraining with relatively small data sets (<100 K reactions) achieves comparable performance to larger data sets containing millions of examples. The use of artificially generated domain-specific pretraining data is proposed. The artificially generated sets prove to be a good surrogate for the reaction schemes extracted from reaction data sets such as Pistachio or Reaxys. The best performance was observed for hybrid pretraining sets combining the real and the domain-specific, artificial data. Finally, we show that a novel adversarial training approach, perturbing input embeddings dynamically, improves model robustness and generalizability for yield and reaction success prediction. These findings provide valuable insights for developing robust and practical machine learning models for yield prediction in synthetic chemistry. GSK's BERT training code base is made available to the community with this work.

Predicting reaction yields in synthetic chemistry remains a significant challenge. This study systematically evaluates the impact of tokenization, molecular representation, pre-training data, and adversarial training on a BERT-based model for yield prediction of Buchwald-Hartwig and Suzuki-Miyaura coupling reactions using publicly available HTE datasets. We demonstrate that molecular representation choice (SMILES, DeepSMILES, SELFIES, Morgan fingerprint-based notation, IUPAC names) has minimal impact on model performance, while typically BPE and SentencePiece tokenization outperform other methods. WordPiece is strongly discouraged for SELFIES and fingerprint-based notation. Furthermore, pre-training with relatively small data sets (<100K reactions) achieves comparable performance to larger data sets containing millions of examples. A use of artificially generated domain-specific pre-training data is proposed. The artificially generated sets prove to be a good surrogate to the reaction schemes extracted from reaction data sets such as Pistachio or Reaxys. The best performance was observed for hybrid pre-training sets combining the real and the domain-specific, artificial data. Finally, we show that a novel adversarial training approach, perturbing input embeddings dynamically, improves model robustness and generalisability for yield and reaction success prediction. These findings provide valuable insights for developing robust and practical machine learning models for yield prediction in synthetic chemistry. GSK’s BERT training code base is made available to the community with this work.

The design of compounds during hit-to-lead often seeks to explore a vector from a core scaffold to form additional interactions with the target protein. A rational approach to this is to probe the region of a protein accessed by a vector with a systematic placement of pharmacophore features in 3D, particularly when bound structures are not available. Herein, we present bbSelect, an open-source tool built to map the placements of pharmacophore features in 3D Euclidean space from a library of R-groups, employing partitioning to drive a diverse and systematic selection to a user-defined size. An evaluation of bbSelect against established methods exemplified the superiority of bbSelect in its ability to perform diverse selections, achieving high levels of pharmacophore feature placement coverage with selection sizes of a fraction of the total set and without the introduction of excess complexity. bbSelect also reports visualizations and rationale to enable users to understand and interrogate results. This provides a tool for the drug discovery community to guide their hit-to-lead activities.

Federated multipartner machine learning has been touted as an appealing and efficient method to increase the effective training data volume and thereby the predictivity of models, particularly when the generation of training data is resource-intensive. In the landmark MELLODDY project, indeed, each of ten pharmaceutical companies realized aggregated improvements on its own classification or regression models through federated learning. To this end, they leveraged a novel implementation extending multitask learning across partners, on a platform audited for privacy and security. The experiments involved an unprecedented cross-pharma data set of 2.6+ billion confidential experimental activity data points, documenting 21+ million physical small molecules and 40+ thousand assays in on-target and secondary pharmacodynamics and pharmacokinetics. Appropriate complementary metrics were developed to evaluate the predictive performance in the federated setting. In addition to predictive performance increases in labeled space, the results point toward an extended applicability domain in federated learning. Increases in collective training data volume, including by means of auxiliary data resulting from single concentration high-throughput and imaging assays, continued to boost predictive performance, albeit with a saturating return. Markedly higher improvements were observed for the pharmacokinetics and safety panel assay-based task subsets.

Accurate methods to predict solubility from molecular structure are highly sought after in the chemical sciences. To assess the state of the art, the American Chemical Society organized a "Second Solubility Challenge" in 2019, in which competitors were invited to submit blinded predictions of the solubilities of 132 drug-like molecules. In the first part of this article, we describe the development of two models that were submitted to the Blind Challenge in 2019 but which have not previously been reported. These models were based on computationally inexpensive molecular descriptors and traditional machine learning algorithms and were trained on a relatively small data set of 300 molecules. In the second part of the article, to test the hypothesis that predictions would improve with more advanced algorithms and higher volumes of training data, we compare these original predictions with those made after the deadline using deep learning models trained on larger solubility data sets consisting of 2999 and 5697 molecules. The results show that there are several algorithms that are able to obtain near state-of-the-art performance on the solubility challenge data sets, with the best model, a graph convolutional neural network, resulting in an RMSE of 0.86 log units. Critical analysis of the models reveals systematic differences between the performance of models using certain feature sets and training data sets. The results suggest that careful selection of high quality training data from relevant regions of chemical space is critical for prediction accuracy but that other methodological issues remain problematic for machine learning solubility models, such as the difficulty in modeling complex chemical spaces from sparse training data sets.

The molecule generation process in de novo design can produce many hundreds of thousands of compounds from an input pool of one or more known compounds. Property filtering and other fast scoring methods can reduce this set in accordance with project goals. However, there may still be many tens of thousands of compounds to select from. The question we pose here is how to identify the most relevant compounds from a set for further analysis and selection. We define relevant as the compounds most suitable for exploring the SAR. Whilst similarity methods can be used to identify the most similar compounds to the input set we show that they are not well suited to this more general task. We introduce the Chemical Annotation Score as a novel method for calculating chemical distance. We show the superiority to fingerprint based similarity methods for identifying the most relevant set of compounds from a large pool for further exploration given an input query or queries.