Open-source pipeline mass-produces thousands of target-specific QSAR models from ChEMBL
- An automated, freely available workflow turns ChEMBL bioactivity records into target-specific QSAR models, covering both regression and classification, and yielded over two thousand models.
- Compounds are encoded with ErG and RDKit descriptors, fitted with tree-based learners tuned by Optuna, and class imbalance is handled by oversampling; data are split by assay and measurement type.
- Only models that clear fixed quality cut-offs are kept, and external checks use the DELTA database; an interactive visualization tool is included for exploring descriptor patterns.
💡 Points Based on the abstract
Mass-producing models is easy; trusting them is not. Splitting by assay and measurement type is sensible, but the abstract shows no comparison against existing baselines and no prospective test. Before any of these models goes near a live project, I'd check how close the DELTA external sets sit to the training chemistry.