MarkTechPost showed how to build an AI co-researcher QSAR for finding EGFR inhibitors

MarkTechPost on July 6, 2026, published a practical tutorial on creating an autonomous AI research assistant for discovering EGFR C797S mutation inhibitors — one of the reasons for drug resistance in lung cancer.

How the System Collects Data

The pipeline begins with defining the target protein: through ChEMBL and UniProt databases, the authors find structural and functional information about EGFR and its therapy-resistant C797S form. Records of experimental IC50 activity are extracted from ChEMBL — the concentration of a substance at which half the target's activity is inhibited. These values are converted to pIC50, a standard logarithmic scale of binding strength, convenient for training a machine learning model.

• Data sources — ChEMBL (chemical activity) and UniProt (protein target) databases
• Target mutation — EGFR C797S, one of the reasons for tumor resistance to targeted drugs
• Activity measure — IC50, converted to pIC50
• Molecular structure encoding — Morgan fingerprints via the RDKit library
• Model — Random Forest with scaffold-split data partitioning

How the QSAR Model is Trained

After data collection and cleaning, RDKit standardizes chemical structures and calculates Morgan fingerprints — numeric vectors encoding the environment of each atom in a molecule. A Random Forest model is trained on these features, learning to predict pIC50 values for new, previously unseen compounds.

A key methodological detail is scaffold-split. Instead of randomly dividing data into training and test sets, the authors group molecules by chemical scaffold and ensure that compounds with the same skeleton do not end up in both sets simultaneously. This approach more honestly tests the model's ability to generalize knowledge to structurally new classes of substances, rather than simply memorizing variants of already familiar molecules.

Why SHAP and Candidate Generation Matter

To understand which specific molecular fragments increase predicted activity, the authors apply SHAP — a method for interpreting machine learning models, showing the contribution of each feature to the final prediction. This transforms Random Forest from a black box into a tool that tells chemists which structural elements are worth preserving or strengthening in new candidates.

Next, the BRICS algorithm is included in the pipeline — it breaks known active molecules into fragments at typical bond-breaking points during chemical synthesis. By reassembling these fragments in new combinations, the system generates a library of candidate structures and uses the trained QSAR model to rank them by predicted binding strength to EGFR C797S, selecting the most promising variants for further testing.

What This Means

The tutorial shows how a combination of open databases, RDKit, classical machine learning, and explainable AI transforms into a working tool for early-stage drug development — without proprietary platforms and with transparent logic for every prediction that a chemist can verify.

Frequently Asked Questions

What is scaffold-split in QSAR model training?

This is a way of dividing data into training and test sets by chemical scaffold of molecules, rather than randomly — this tests whether the model can predict activity of truly new structures, rather than simply recognizing those similar to ones already seen.

Why are BRICS fragments needed in the pipeline?

BRICS breaks known active molecules into fragments at typical bond-breaking points during synthesis, and then allows reassembling these fragments into new candidate structures that the model ranks by predicted binding strength to EGFR C797S.