ChEMBL Metalloproteinase Bioactivity Pipeline

SQL-based extraction and cleaning of bioactivity data for metalloproteinase targets from a local ChEMBL SQLite dump, producing a clean SMILES + activity-class dataset ready for QSAR modeling. Same target family (MMPs) as an earlier published virtual screening study of mine (Ramezani & Shamsara, Mol. Divers. 2018), though built independently against a current ChEMBL release, not the pipeline used for that paper.

What it does

1. Queries the activities/assays/compound_structures/target_dictionary tables directly via sqlite3, joining on molregno/tid, filtered to direct single-protein-target binding assays (confidence_score = 9, assay_type = 'B').
2. Converts standard_value (nM) to pActivity (-log10(M), the same convention as ChEMBL's own pchembl_value) and classifies each compound active/inactive at a configurable pIC50-equivalent threshold (6.5 by default).
3. Drops censored measurements (>/< relations) whose reported bound can't actually resolve which side of the threshold the true value falls on.
4. Deduplicates by (molregno, Target_ChEMBL_ID): rows that disagree on activity class, or whose pActivity values differ by more than 3 log units, are dropped outright as unresolvable; genuinely consistent duplicates are then collapsed to one row.
5. Writes the clean SMILES + activity_cat dataset per target.

Requirements

• A local ChEMBL SQLite dump (chembl_<version>.db), downloaded from ChEMBL's downloads page -- not included here (multi-GB file).
• Python 3, pandas, numpy.

Files

ChEMBL_GitHub.ipynb          the full pipeline
MMP_0.csv .. MMP_4_6.5.csv   intermediate/output CSVs from an MMP run, kept for reference

To run against a different target, edit the t.pref_name LIKE '%...%' clause in the query cell, and the target/path variables at the top.