Unknown causative features in drug activity prediction

Identify currently unknown causative molecular features that determine small-molecule activity against their cognate targets and incorporate these features into the representations used by AI models to improve predictive performance.

Background

In the discussion on feature relevance, the authors note that predictive features can correlate with but not cause the property of interest, and express concern that some causative features may be missing from models either because they are unknown or cannot be effectively computed. They illustrate related known but hard-to-quantify factors (e.g., ligand and pocket desolvation, induced-fit effects) that limit the accuracy of scoring functions, underscoring the need to discover and include missing causative determinants.

References

A more concerning possibility is that some of the causative features are not modelled. This could happen because we do not know about them.

— Data-centric challenges with the application and adoption of artificial intelligence for drug discovery  (2407.05150 - Ghislat et al., 2024) in Section 2.4 Irrelevant data

Yet the cliff error still sits well above the overall error, as it does for every method, indicating that the benchmark is not solved. Predicting the sharp, local changes in activity that define cliffs remains an open problem.

— Monroe: A Molecular Foundation Model for In-Context Probabilistic Inference  (2608.18982 - Banaszewski et al., 19 Aug 2026) in Section 5, Discussion

Other selections have no obvious chemical explanation. Liver-injury portfolios most often pick the same cardiotoxicity classifier (3 of 6 cells) and an electronic-spectra block (2 of 6), and blood--brain-barrier portfolios also select antimicrobial models (3 of 6). We do not know why these models help; the method makes such choices visible without guaranteeing a coherent chemical story.

— Interpretable-by-Design Descriptor Portfolios Match a 2048-Dimensional Foundation Embedding on Low-Data Molecular Assays  (2609.30789 - Yao et al., 25 Sep 2026) in Section 3.3, “What the selector actually draws on” (paragraph discussing chemically unexplained selections)