Drug Discovery with Apex Protocol by Nvidia and Numerion Labs (arxiv.org)

Researchers propose APEX (approximate-but-exhaustive search), a new protocol for virtual screening of ultra‑large combinatorial synthesis libraries (CSLs). APEX trains a neural-network surrogate that leverages the structured nature of CSLs to rapidly score every enumerated compound on a consumer GPU in under a minute, producing an “approximate but exhaustive” ranking from which exact top-k hits can be retrieved. The authors build a 10+ million compound benchmark annotated with docking scores for five medically relevant targets and RDKit physicochemical properties to provide ground-truth comparisons, and report that APEX outperforms competing methods on both retrieval accuracy and runtime. This matters because current virtual screening of libraries that reach tens of billions of molecules typically evaluates <0.1% of candidates under realistic compute budgets, so high-scoring compounds are easily missed. APEX both makes full enumeration tractable and amortizes computation when objectives or constraints change during a campaign, enabling fast re‑ranking under new filters. For drug discovery teams, that translates to far cheaper, faster in‑silico prioritization, better hit-finding coverage, and an adaptable workflow for evolving constraints — all without requiring massive cluster resources.