Rapid expansion of high-resolution data in 3D and molecular activity space led to new methods and a large variety of 3D/ML/AI predictive models of thousands of activities. A pipeline of target definition, defining its chemical and 3D state, search of synthesizable chemicals in giga-/tera-spaces, and re-ranking the top compounds by a complex profile is presented. Several projects illustrate the process that led to drug candidates in clinical trials.

Make-on-demand libraries are now so large that brute-force docking can’t keep up. ChemSTEP is a simple, transparent way to explore huge libraries so you only dock what’s worth docking, recovering most top virtual hits at a fraction of the compute. Instead of black-box AI/ML, ChemSTEP uses familiar chemical-similarity logic, yet delivers over 2,000-fold acceleration and equivalent-to-superior performance compared to AI/ML. We’ll show results from docking a 1-trillion library against model targets.

Intrinsically disordered proteins (IDPs) challenge structure-based drug discovery by existing as dynamic ensembles rather than fixed structures. We argue this is fundamentally a representation problem. This talk reframes IDP modeling as learning over conformational distributions, reviews current computational approaches, and highlights how probabilistic machine learning enables ensemble-based reasoning, opening new directions for AI-driven discovery of disordered targets.

Third-generation co-folding models struggle with predicting structures of macrocycles which contain non-canonical amino acids, thus their application is limited to hit-finding. Vilya goes beyond this limitation to realize the power of such models in a much broader chemical space. I will describe how we build such models, describe their capabilities in structure- and property-prediction, and show vignettes of how they can be used throughout the discovery process.

In this presentation, we introduce how Koliber’s machine-learning technology, integrated with Robust Diagnostics' peptide-array technology, overcomes these limitations. We demonstrate that large libraries are unnecessary, as Koliber’s machine learning can optimize initial hits to achieve improved binding affinity. We also present visualization techniques for detecting binding modes, offering new insights into peptide-array applications for therapeutic peptide discovery.

Cyclic peptides have long been considered attractive as a drug modality due to their medium size and combining the advantages of small molecules and biologics. However, membrane permeability remains a significant challenge. Recently, physics-based Generative AI has emerged as a promising technology to design cyclic peptides with specific properties in mind. Here we focus on applying this method to design de novo cyclic peptides with drug-like oral bioavailability.