Results from the second CACHE challenge are now published in the Journal of Chemical Information and Modeling. I was again a contributing computational team.

This challenge targeted the RNA-binding site of the SARS-CoV-2 Nsp13 helicase — a highly conserved COVID-19 target. Twenty-three teams used protein structure and fragment-screening data, combined with computational and machine learning methods, to each predict up to 100 inhibitory ligands. In total, 1957 compounds were procured and tested.

The overall hit rate was 0.7% by SPR, reflecting the difficulty of the site. The six best-performing workflows employed fragment growing, active learning, or conventional virtual screening augmented with deep-learning scoring functions. Follow-up functional assays identified two compound scaffolds with K_d values below 10 μM that also inhibited in vitro helicase activity.

A clear pattern is emerging across the first two CACHE challenges: workflows that link or grow docked/crystallized fragments, or that use docking of small diverse libraries to train ultrafast machine-learning surrogates, are recurrently among the top performers. CACHE #2 demonstrates that crowd-sourced, critically assessed benchmarking can yield genuinely novel lead matter even for challenging viral targets.