Unsupervised and supervised AI on molecular dynamics simulations reveals complex characteristics of HLA-A2-peptide immunogenicity.

Immunologic recognition of peptide antigens bound to class I major histocompatibility complex (MHC) molecules is essential to both novel immunotherapeutic development and human health at large. Current methods for predicting antigen peptide immunogenicity rely primarily on simple sequence representations, which allow for some understanding of immunogenic features but provide inadequate consideration of the full scale of molecular mechanisms tied to peptide recognition. We here characterize contributions that unsupervised and supervised artificial intelligence (AI) methods can make toward understanding and predicting MHC(HLA-A2)-peptide complex immunogenicity when applied to large ensembles of molecular dynamics simulations. We first show that an unsupervised AI method allows us to identify subtle features that drive immunogenicity differences between a cancer neoantigen and its wild-type peptide counterpart. Next, we demonstrate that a supervised AI method for class I MHC(HLA-A2)-peptide complex classification significantly outperforms a sequence model on small datasets corrected for trivial sequence correlations. Furthermore, we show that both unsupervised and supervised approaches reveal determinants of immunogenicity based on time-dependent molecular fluctuations and anchor position dynamics outside the MHC binding groove. We discuss implications of these structural and dynamic immunogenicity correlates for the induction of T cell responses and therapeutic T cell receptor design.

On the Choice of Active Site Sequences for Kinase-Ligand Affinity Prediction.

Recent work showed that active site rather than full-protein-sequence information improves predictive performance in kinase-ligand binding affinity prediction. To refine the notion of an "active site", we here propose and compare multiple definitions. We report significant evidence that our novel definition is superior to previous definitions and better models of ATP-noncompetitive inhibitors. Moreover, we leverage the discontiguity of the active site sequence to motivate novel protein-sequence augmentation strategies and find that combining them further improves performance.

CACHE (Critical Assessment of Computational Hit-finding Experiments): A public-private partnership benchmarking initiative to enable the development of computational methods for hit-finding.

One aspirational goal of computational chemistry is to predict potent and drug-like binders for any protein, such that only those that bind are synthesized. In this Roadmap, we describe the launch of Critical Assessment of Computational Hit-finding Experiments (CACHE), a public benchmarking project to compare and improve small molecule hit-finding algorithms through cycles of prediction and experimental testing. Participants will predict small molecule binders for new and biologically relevant protein targets representing different prediction scenarios. Predicted compounds will be tested rigorously in an experimental hub, and all predicted binders as well as all experimental screening data, including the chemical structures of experimentally tested compounds, will be made publicly available, and not subject to any intellectual property restrictions. The ability of a range of computational approaches to find novel binders will be evaluated, compared, and openly published. CACHE will launch 3 new benchmarking exercises every year. The outcomes will be better prediction methods, new small molecule binders for target proteins of importance for fundamental biology or drug discovery, and a major technological step towards achieving the goal of Target 2035, a global initiative to identify pharmacological probes for all human proteins.

Analysis of Training and Seed Bias in Small Molecules Generated with a Conditional Graph-Based Variational Autoencoder─Insights for Practical AI-Driven Molecule Generation.

The application of deep learning to generative molecule design has shown early promise for accelerating lead series development. However, questions remain concerning how factors like training, data set, and seed bias impact the technology's utility to medicinal and computational chemists. In this work, we analyze the impact of seed and training bias on the output of an activity-conditioned graph-based variational autoencoder (VAE). Leveraging a massive, labeled data set corresponding to the dopamine D2 receptor, our graph-based generative model is shown to excel in producing desired conditioned activities and favorable unconditioned physical properties in generated molecules. We implement an activity-swapping method that allows for the activation, deactivation, or retention of activity of molecular seeds, and we apply independent deep learning classifiers to verify the generative results. Overall, we uncover relationships between noise, molecular seeds, and training set selection across a range of latent-space sampling procedures, providing important insights for practical AI-driven molecule generation.

Simplified, interpretable graph convolutional neural networks for small molecule activity prediction.

We here present a streamlined, explainable graph convolutional neural network (gCNN) architecture for small molecule activity prediction. We first conduct a hyperparameter optimization across nearly 800 protein targets that produces a simplified gCNN QSAR architecture, and we observe that such a model can yield performance improvements over both standard gCNN and RF methods on difficult-to-classify test sets. Additionally, we discuss how reductions in convolutional layer dimensions potentially speak to the "anatomical" needs of gCNNs with respect to radial coarse graining of molecular substructure. We augment this simplified architecture with saliency map technology that highlights molecular substructures relevant to activity, and we perform saliency analysis on nearly 100 data-rich protein targets. We show that resultant substructural clusters are useful visualization tools for understanding substructure-activity relationships. We go on to highlight connections between our models' saliency predictions and observations made in the medicinal chemistry literature, focusing on four case studies of past lead finding and lead optimization campaigns.

Active Site Sequence Representations of Human Kinases Outperform Full Sequence Representations for Affinity Prediction and Inhibitor Generation: 3D Effects in a 1D Model.

Recent advances in deep learning have enabled the development of large-scale multimodal models for virtual screening and de novo molecular design. The human kinome with its abundant sequence and inhibitor data presents an attractive opportunity to develop proteochemometric models that exploit the size and internal diversity of this family of targets. Here, we challenge a standard practice in sequence-based affinity prediction models: instead of leveraging the full primary structure of proteins, each target is represented by a sequence of 29 discontiguous residues defining the ATP binding site. In kinase-ligand binding affinity prediction, our results show that the reduced active site sequence representation is not only computationally more efficient but consistently yields significantly higher performance than the full primary structure. This trend persists across different models, data sets, and performance metrics and holds true when predicting pIC50 for both unseen ligands and kinases. Our interpretability analysis reveals a potential explanation for the superiority of the active site models: whereas only mild statistical effects about the extraction of three-dimensional (3D) interaction sites take place in the full sequence models, the active site models are equipped with an implicit but strong inductive bias about the 3D structure stemming from the discontiguity of the active sites. Moreover, in direct comparisons, our models perform similarly or better than previous state-of-the-art approaches in affinity prediction. We then investigate a de novo molecular design task and find that the active site provides benefits in the computational efficiency, but otherwise, both kinase representations yield similar optimized affinities (for both SMILES- and SELFIES-based molecular generators). Our work challenges the assumption that the full primary structure is indispensable for modeling human kinases.

Combining Docking Pose Rank and Structure with Deep Learning Improves Protein-Ligand Binding Mode Prediction over a Baseline Docking Approach.

We present a simple, modular graph-based convolutional neural network that takes structural information from protein-ligand complexes as input to generate models for activity and binding mode prediction. Complex structures are generated by a standard docking procedure and fed into a dual-graph architecture that includes separate subnetworks for the ligand bonded topology and the ligand-protein contact map. Recent work has indicated that data set bias drives many past promising results derived from combining deep learning and docking. Our dual-graph network allows contributions from ligand identity that give rise to such biases to be distinguished from effects of protein-ligand interactions on classification. We show that our neural network is capable of learning from protein structural information when, as in the case of binding mode prediction, an unbiased data set is constructed. We next develop a deep learning model for binding mode prediction that uses docking ranking as input in combination with docking structures. This strategy mirrors past consensus models and outperforms a baseline docking program (AutoDock Vina) in a variety of tests, including on cross-docking data sets that mimic real-world docking use cases. Furthermore, the magnitudes of network predictions serve as reliable measures of model confidence.

Role of chronic toxicology studies in revealing new toxicities.

Chronic (>3 months) preclinical toxicology studies are conducted to support the safe conduct of clinical trials exceeding 3 months in duration. We have conducted a review of 32 chronic toxicology studies in non-rodents (22 studies in dogs and 10 in non-human primates) and 27 chronic toxicology studies in rats dosed with Merck compounds to determine the frequency at which additional target organ toxicities are observed in chronic toxicology studies as compared to subchronic studies of 3 months in duration. Our review shows that majority of the findings are observed in the subchronic studies since additional target organs were not observed in 24 chronic non rodent studies and in 21 chronic rodent studies. However, 6 studies in non rodents and 6 studies in rodents yielded new findings that were not seen in studies of 3-month or shorter duration. For 3 compounds the new safety findings did contribute to termination of clinical development plans. Although the incidence of compound termination associated with chronic toxicology study observations is low (∼10%), the observations made in these studies can be important for evaluating human safety risk.

Application of an automated natural language processing (NLP) workflow to enable federated search of external biomedical content in drug discovery and development.

External content sources such as MEDLINE(®), National Institutes of Health (NIH) grants and conference websites provide access to the latest breaking biomedical information, which can inform pharmaceutical and biotechnology company pipeline decisions. The value of the sites for industry, however, is limited by the use of the public internet, the limited synonyms, the rarity of batch searching capability and the disconnected nature of the sites. Fortunately, many sites now offer their content for download and we have developed an automated internal workflow that uses text mining and tailored ontologies for programmatic search and knowledge extraction. We believe such an efficient and secure approach provides a competitive advantage to companies needing access to the latest information for a range of use cases and complements manually curated commercial sources.

Developing timely insights into comparative effectiveness research with a text-mining pipeline.

Comparative effectiveness research (CER) provides evidence for the relative effectiveness and risks of different treatment options and informs decisions made by healthcare providers, payers, and pharmaceutical companies. CER data come from retrospective analyses as well as prospective clinical trials. Here, we describe the development of a text-mining pipeline based on natural language processing (NLP) that extracts key information from three different trial data sources: NIH ClinicalTrials.gov, WHO International Clinical Trials Registry Platform (ICTRP), and Citeline Trialtrove. The pipeline leverages tailored terminologies to produce an integrated and structured output, capturing any trials in which pharmaceutical products of interest are compared with another therapy. The timely information alerts generated by this system provide the earliest and most complete picture of emerging clinical research.

QSAR Prediction of Passive Permeability in the LLC-PK1 Cell Line: Trends in Molecular Properties and Cross-Prediction of Caco-2 Permeabilities.

A QSAR model for predicting passive permeability (Papp ) was derived from Papp values measured in the LLC-PK1 cell line. The QSAR method and descriptor set that performed best in terms of cross-validation was random forest with a combination of AP, DP, and MOE_2D descriptors. The QSAR model was used to predict the Caco-2 cell permeability for 313 compounds described in the literature with good success. We find that passive permeability for different cell lines can be predicted with similar molecular properties and descriptors. It is shown that the variation in experimental measurements of Papp is smaller than the error in QSAR predictions indicating that predictions are not quantitatively perfect, although qualitatively useful. We get better predictions if the training set is large and diverse, rather than smaller and more internally consistent. This is because prediction accuracy falls off quickly with decreasing similarity to the training set and it is therefore better to have as large a training set as possible. While single physical parameters are not as good as a full QSAR model in predicting Papp , logD seems the most important parameter. Intermediate values of logD are associated with higher Papp .

Drug-like density: a method of quantifying the "bindability" of a protein target based on a very large set of pockets and drug-like ligands from the Protein Data Bank.

One approach to estimating the "chemical tractability" of a candidate protein target where we know the atomic resolution structure is to examine the physical properties of potential binding sites. A number of other workers have addressed this issue. We characterize ~290,000 "pockets" from ~42,000 protein crystal structures in terms of a three parameter "pocket space": volume, buriedness, and hydrophobicity. A metric DLID (drug-like density) measures how likely a pocket is to bind a drug-like molecule. This is calculated from the count of other pockets in its local neighborhood in pocket space that contain drug-like cocrystallized ligands and the count of total pockets in the neighborhood. Surprisingly, despite being defined locally, a global trend in DLID can be predicted by a simple linear regression on log(volume), buriedness, and hydrophobicity. Two levels of simplification are necessary to relate the DLID of individual pockets to "targets": taking the best DLID per Protein Data Bank (PDB) entry (because any given crystal structure can have many pockets), and taking the median DLID over all PDB entries for the same target (because different crystal structures of the same protein can vary because of artifacts and real conformational changes). We can show that median DLIDs for targets that are detectably homologous in sequence are reasonably similar and that median DLIDs correlate with the "druggability" estimate of Cheng et al. (Nature Biotechnology 2007, 25, 71-75).

QSAR models for predicting the similarity in binding profiles for pairs of protein kinases and the variation of models between experimental data sets.

We propose a direct QSAR methodology to predict how similar the inhibitor-binding profiles of two protein kinases are likely to be, based on the properties of the residues surrounding the ATP-binding site. We produce a random forest model for each of five data sets (one in-house, four from the literature) where multiple compounds are tested on many kinases. Each model is self-consistent by cross-validation, and all models point to only a few residues in the active site controlling the binding profiles. While all models include the "gatekeeper" as one of the important residues, consistent with previous literature, some models suggest other residues as being more important. We apply each model to predict the similarity in binding profile to all pairs in a set of 411 kinases from the human genome and get very different predictions from each model. This turns out not to be an issue with model-building but with the fact that the experimental data sets disagree about which kinases are similar to which others. It is possible to build a model combining all the data from the five data sets that is reasonably self-consistent but not surprisingly, given the disagreement between data sets, less self-consistent than the individual models.

Discovery of potent, selective, and orally active carboxylic acid based inhibitors of matrix metalloproteinase-13.

The matrix metalloproteinase enzyme MMP-13 plays a key role in the degradation of type II collagen in cartilage and bone in osteoarthritis (OA). An effective MMP-13 inhibitor would therefore be a novel disease modifying therapy for the treatment of arthritis. Our efforts have resulted in the discovery of a series of carboxylic acid inhibitors of MMP-13 that do not significantly inhibit the related MMP-1 (collagenase-1) or tumor necrosis factor-alpha (TNF-alpha) converting enzyme (TACE). It has previously been suggested (but not proven) that inhibition of the latter two enzymes could lead to side effects. A promising carboxylic acid lead 9 was identified and a convergent synthesis developed. This paper describes the optimization of 9 and the identification of a compound 24f for further development. Compound 24f is a subnanomolar inhibitor of MMP-13 (IC(50) value 0.5 nM and K(i) of 0.19 nM) having no activity against MMP-1 or TACE (IC(50) of >10000 nM). Furthermore, in a rat model of MMP-13-induced cartilage degradation, 24f significantly reduced proteoglycan release following oral dosing at 30 mg/kg (75% inhibition, p < 0.05) and at 10 mg/kg (40% inhibition, p < 0.05).

Multiple protein structures and multiple ligands: effects on the apparent goodness of virtual screening results.

As an extension to a previous published study (McGaughey et al., J Chem Inf Model 47:1504-1519, 2007) comparing 2D and 3D similarity methods to docking, we apply a subset of those virtual screening methods (TOPOSIM, SQW, ROCS-color, and Glide) to a set of protein/ligand pairs where the protein is the target for docking and the cocrystallized ligand is the target for the similarity methods. Each protein is represented by a maximum of five crystal structures. We search a diverse subset of the MDDR as well as a diverse small subset of the MCIDB, Merck's proprietary database. It is seen that the relative effectiveness of virtual screening methods, as measured by the enrichment factor, is highly dependent on the particular crystal structure or ligand, and on the database being searched. 2D similarity methods appear very good for the MDDR, but poor for the MCIDB. However, ROCS-color (a 3D similarity method) does well for both databases.

Comparison of topological, shape, and docking methods in virtual screening.

Virtual screening benchmarking studies were carried out on 11 targets to evaluate the performance of three commonly used approaches: 2D ligand similarity (Daylight, TOPOSIM), 3D ligand similarity (SQW, ROCS), and protein structure-based docking (FLOG, FRED, Glide). Active and decoy compound sets were assembled from both the MDDR and the Merck compound databases. Averaged over multiple targets, ligand-based methods outperformed docking algorithms. This was true for 3D ligand-based methods only when chemical typing was included. Using mean enrichment factor as a performance metric, Glide appears to be the best docking method among the three with FRED a close second. Results for all virtual screening methods are database dependent and can vary greatly for particular targets.

Inhibitors of human histone deacetylase: synthesis and enzyme and cellular activity of straight chain hydroxamates.

Inhibitors of histone deacetylase (HDAC) have been shown to induce terminal differentiation of human tumor cell lines and to have antitumor effects in vivo. We have prepared analogues of suberoylanilide hydroxamic acid (SAHA) and trichostatin A and have evaluated them in a human HDAC enzyme inhibition assay, a p21(waf1) (p21) promoter assay, and in monolayer growth inhibition assays. One compound, 4-(dimethylamino)-N-[7-(hydroxyamino)-7-oxoheptyl]-benzamide, was found to affect the growth of a panel of eight human tumor cell lines differentially.