Ranking compounds within the same assay beats absolute activity prediction on pooled ChEMBL-style data
- Public bioactivity data clusters by assay origin, so simply guessing a training assay's average activity already performs surprisingly well for compounds from that same assay.
- The authors argue models should prioritize ranking compounds within an assay rather than predicting absolute values on aggregated data.
- On three ChEMBL-derived datasets, rankers trained on same-assay sets outperformed IC50 regression and rankers trained on arbitrary compound sets.
💡 Points Based on the abstract
If same-assay mean is a strong baseline, then random-split benchmarks on pooled ChEMBL mostly reward assay-identity leakage, so I'd distrust any reported R² there. Before adopting this, check how it holds on a new assay or a new series, where no same-assay anchor exists. In practice it fits lead optimization: a few measured compounds per assay, then rank the rest.
Assay-Based Machine Learning: Rethinking Evaluation in Drug Discovery ↗