MetaPLI uses retrieved assay context to predict binding affinity without retraining
- Retrieval-augmented, in-context regression model that predicts absolute binding affinity from protein–ligand complex structures plus a few known activity measurements.
- New assay results are added to the retrieved context, informing later predictions and compound selection in a lead-optimization loop without retraining.
- Benchmarks, validation design (retrospective vs prospective) and comparisons with existing methods were not available in the accessible record.
💡 Points Based on the abstract
Before this replaces anything on your project, ask how the retrieval is split: if the retrieved context shares a chemical series with the query compounds, you're measuring local QSAR with extra steps, and activity cliffs will hurt exactly where you need it. Run a series-held-out test, then compare against a simple local model and FEP on your own last two rounds.