When Validity Isn't Enough: Reliability Gaps in Molecular Generation and KRAS Case Study

ICLR 2026 Conference Submission24912 Authors

Keywords: Molecule Generation, Generative Models, KRAS, Benchmark

Abstract: Molecule generation remains a core challenge in computational chemistry. Practical use of generative models is complicated by strict chemical, structural, and biological constraints: candidate compounds must satisfy physicochemical bounds, avoid reactive or toxic substructures, be synthesizable, and plausibly bind a target. We are the first to perform such comprehensive analysis of modern molecule generators via the Five-Stage Filtering Pipeline, a target-agnostic, practice-oriented benchmark for evaluating de novo generators using the following stages: (i) physicochemical descriptors; (ii) structural alerts; (iii) synthesis feasibility; (iv) docking and binding affinity estimation; and (v) blind medicinal chemist review. We compare 18 generators across three families (unconditional, ligand-based, and protein-based), and to make it practically relevant, apply the pipeline to KRAS G12D switch-II pocket for conditional design case study. Less than 1% of molecules pass all stages, exposing a gap between high scores on standard generative metrics and practical medicinal chemistry usage. We release our benchmark, and code to enable reproducible evaluation and to focus future model development on practically useful chemical space.

Primary Area: datasets and benchmarks

Submission Number: 24912