The challenges of generalizability in artificial intelligence for ADME/Tox endpoint and activity prediction

INTRODUCTION

Artificial intelligence (AI) has seen a massive resurgence in recent years with wide successes in computer vision, natural language processing, and games. The similar creation of robust and accurate AI models for ADME/Tox endpoint and activity prediction would be revolutionary to drug discovery pipelines. There have been numerous demonstrations of successful applications, but a key challenge remains: how generalizable are these predictive models?

AREAS COVERED

The authors present a summary of current promising components of AI models in the context of early drug discovery where ADME/Tox endpoint and activity prediction is the main driver of the iterative drug design process. Following that is a review of applicability domains and dataset construction considerations which determine generalizability bottlenecks for AI deployment. Further reviewed is the role of promising learning frameworks - multitask, transfer, and meta learning - which leverage auxiliary data to overcome issues of generalizability.

EXPERT OPINION

The authors conclude that the most promising direction toward integrating reliable and informative AI models into the drug discovery pipeline is a conjunction of learned feature representations, deep learning, and novel learning frameworks. Such a solution would address the sparse and incomplete datasets that are available for key endpoints related to drug discovery.

Pyroglutamate and O-linked glycan determine functional production of anti-IL17A and anti-IL22 peptide-antibody bispecific genetic fusions

Protein biosynthesis and extracellular secretion are essential biological processes for therapeutic protein production in mammalian cells, which offer the capacity for correct folding and proper post-translational modifications. In this study, we have generated bispecific therapeutic fusion proteins in mammalian cells by combining a peptide and an antibody into a single open reading frame. A neutralizing peptide directed against interleukin-17A (IL17A) was genetically fused to the N termini of an anti-IL22 antibody, through either the light chain, the heavy chain, or both chains. Although the resulting fusion proteins bound and inhibited IL22 with the same affinity and potency as the unmodified anti-IL22 antibody, the peptide modality in the fusion scaffold was not active in the cell-based assay due to the N-terminal degradation. When a glutamine residue was introduced at the N terminus, which can be cyclized to form pyroglutamate in mammalian cells, the IL17A neutralization activity of the fusion protein was restored. Interestingly, the mass spectroscopic analysis of the purified fusion protein revealed an unexpected O-linked glycosylation modification at threonine 5 of the anti-IL17A peptide. The subsequent removal of this post-translational modification by site-directed mutagenesis drastically enhanced the IL17A binding affinity and neutralization potency for the resulting fusion protein. These results provide direct experimental evidence that post-translational modifications during protein biosynthesis along secretory pathways play critical roles in determining the structure and function of therapeutic proteins produced by mammalian cells. The newly engineered peptide-antibody genetic fusion is promising for therapeutically targeting multiple antigens in a single antibody-like molecule.

Comparison of several molecular docking programs: pose prediction and virtual screening accuracy

Molecular docking programs are widely used modeling tools for predicting ligand binding modes and structure based virtual screening. In this study, six molecular docking programs (DOCK, FlexX, GLIDE, ICM, PhDOCK, and Surflex) were evaluated using metrics intended to assess docking pose and virtual screening accuracy. Cognate ligand docking to 68 diverse, high-resolution X-ray complexes revealed that ICM, GLIDE, and Surflex generated ligand poses close to the X-ray conformation more often than the other docking programs. GLIDE and Surflex also outperformed the other docking programs when used for virtual screening, based on mean ROC AUC and ROC enrichment values obtained for the 40 protein targets in the Directory of Useful Decoys (DUD). Further analysis uncovered general trends in accuracy that are specific for particular protein families. Modifying basic parameters in the software was shown to have a significant effect on docking and virtual screening results, suggesting that expert knowledge is critical for optimizing the accuracy of these methods.

Lead optimization via high-throughput molecular docking

Structure-based lead optimization approaches are increasingly playing a role in the drug-discovery process. Recent advances in 'high-throughput' molecular docking methods and examples of their successful use in lead optimization are reviewed. Measures of docking accuracy, scoring function comparisons, and consensus approaches are discussed. Differences in docking protocols typically used for lead optimization versus lead generation are highlighted; this section includes a discussion of the latest methods for the incorporation of protein flexibility. New approaches developed specifically for the design of combinatorial libraries as well as those designed or used for 'fragment' versus lead optimization are presented. Finally, potential future improvements to the technology are outlined.

Novel 2D fingerprints for ligand-based virtual screening

This paper describes the development of a set of new 2D fingerprints for the purposes of virtual screening in a pharmaceutical environment. The new fingerprints are based on established ones: the changes in their design included the introduction of overlapping pharmacophore feature types, feature counts for pharmacophore and structural fingerprints, as well as changes in the resolution in property description for property fingerprints. The effects of each of these changes on virtual screening performance were monitored using two types of training sets, emulating different stages in the drug discovery process. The results demonstrate that these changes all lead to an improvement in virtual screening performance.

The Use of Consensus Scoring in Ligand-Based Virtual Screening

A new consensus approach has been developed for ligand-based virtual screening. It involves combining highly disparate properties in order to improve performance in virtual screening. The properties include structural, 2D pharmacophore and property-based fingerprints, scores derived using BCUT descriptors, and 3D pharmacophore approaches. Different approaches for the combination of all or some of these methods have been tested. Logistic regression and sum ranks were found to be the most advantageous in different pharmaceutical applications. The three major reasons consensus scoring appears to enrich data sets better than single scoring functions are (1) using multiple scoring functions is similar to repeated samplings, in which case the mean is closer to the true value than any single value, (2) due to the better clustering of actives, multiple sampling will recover more actives than inactives, and (3) different methods seem to agree more on the ranking of the actives than on the inactives. Furthermore, consensus results are not only better but are also more consistent across receptor systems.

Predicting synthetic accessibility: application in drug discovery and development

The estimation and use of synthetic accessibility in the drug discovery process is discussed. A distinction is drawn between synthetic feasibility and accessibility and the practical uses of an accessibility score are examined. Various techniques used in the estimation of accessibility are described and their utility and potential accuracy compared.