Better models have not produced faster programs; I argue the problem is what we spend experiments on. Three quantities should govern the choice: surprise (how far a result moves us), benefit (what changes if the answer is good) and regret (what being wrong costs, and when). This needs emulators, to ask counterfactuals without running them, and difficulty-adjusted outcomes, borrowed from sport statistics. I will set out both, and what they'd change in drug discovery.

Structural models derived from deep learning approaches, such as AlphaFold3 and Boltz-2, are increasingly being used for structure-based design efforts.  We show that these methods can be used for generating conformational ensembles leading to the identification of potentially druggable cryptic pockets. In addition, molecular docking to these structural models is compared to co-folding. Finally, the accuracy of ligand poses by co-folding is evaluated for pairs of similar known binders and non-binders to illustrate challenges that remain.

Predicted protein pocket co-engagement prioritizes documented drug–drug interactions without clinical-label training. Across 21,945 pairs of 210 drugs, composite protein multiple-ligand (COLIG) rankings achieved 49.3% precision among the top 5%, versus 29.6% background. Kinase COLIGs reached 67.0% precision among the top 100 pairs, versus 39.9% for noncontact same-pocket assignments. Jackknifing supported robustness, while explicit networks provided no consistent incremental benefit. These findings support pocket-based screening but do not establish causal mechanisms.

This presentation highlights a powerful computational framework that combines quantitative high-throughput screening (qHTS) data with machine learning to expand hit diversity, followed by AI-guided scaffold-centric exploration and reaction-based QSAR modeling strategies, showcasing real-world success stories from massive virtual library screenings. Key takeaways include the core methodologies behind these advanced, open-source workflows, providing actionable strategic insights to efficiently discover novel nanomolar inhibitors in similar projects.