CHEMBL2ML

An interesting web app that fetches ChEMBL bioactivity data for a target (via UniProt ID), computes molecular descriptors, and trains a simple predictive model (regression, with an optional classification fallback). A great example of using the ChEMBL API and RDKit .

Code is on GitHub https://github.com/agiani99/CHEMBL2ML

What it does

• Input: a UniProt ID (e.g. P00533).
• Fetches:
  • HGNC gene symbol (via genenames.org)
  • ChEMBL target, assays, and activities (via ChEMBL API)
• Builds a dataset with:
  • pchembl_value as the main label
  • RDKit physicochemical descriptors (e.g. MW, LogP, HBD/HBA, TPSA, rings, rotatable bonds)
  • RDKit fragment descriptors (fr_*)
  • Optional ErG-style features using FixedPharmacophoreAnalyzer from erg_calc_fragments_topo.py
• Trains:
  • Regression: Random Forest (default) or XGBoost (if installed)
  • Optional fallback: binary classification (Active if pChEMBL ≥ 6.5) with optional SMOTE
• Exports:
  • Pickled model + preprocessing objects
  • Predictions CSV
  • Top-features JSON
  • Full dataset CSV