MarkTechPost shows how to build an AI co-scientist with a QSAR model to hunt for EGFR inhibitors

On July 6, 2026, MarkTechPost published a hands-on tutorial on building an autonomous AI co-scientist to search for inhibitors of the EGFR C797S mutation, one of the causes of drug resistance in lung cancer.

How the system collects data

The pipeline starts by defining the target protein: using the ChEMBL and UniProt databases, the authors find structural and functional information on EGFR and its therapy-resistant C797S form. From ChEMBL they extract experimental IC50 activity records, the concentration of a substance at which half of the target's activity is inhibited. These values are converted to pIC50, a standard logarithmic scale of binding strength that suits training a machine learning model.

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

How the QSAR model is trained

After the data is collected and cleaned, RDKit standardizes the chemical structures and computes Morgan fingerprints, numerical vectors that encode the environment of each atom in a molecule. A Random Forest model is trained on these features to predict pIC50 for new compounds it has not seen before.

The key methodological detail is the scaffold-split. Instead of splitting data randomly into training and test sets, the authors group molecules by chemical scaffold and make sure compounds sharing the same core do not land in both sets. This tests more honestly whether the model can generalize to structurally new classes of compounds, rather than just memorizing variants of molecules it already knows.

Why SHAP and candidate generation matter

To see which parts of a molecule raise the predicted activity, the authors use SHAP, a machine learning interpretability method that shows each feature's contribution to the final prediction. This turns the Random Forest from a black box into a tool that tells chemists which structural elements to keep or strengthen in new candidates.

Next, the pipeline applies the BRICS algorithm, which breaks known active molecules into fragments at the bond-cleavage points typical of 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 ones for further testing.

What this means

The tutorial shows how open databases, RDKit, classical machine learning and explainable AI combine into a working tool for early-stage drug discovery, without proprietary platforms and with transparent logic behind each prediction that a chemist can check.

Frequently asked questions

What is a scaffold-split in QSAR model training?

It is a way of dividing data into training and test sets by the chemical scaffold of the molecules rather than at random. This checks whether the model can predict the activity of genuinely new structures, rather than just recognizing ones similar to those it has already seen.

Why does the pipeline need BRICS fragments?

BRICS breaks known active molecules into fragments at the bond-cleavage points typical of synthesis, then lets the pipeline reassemble them into new candidate structures, which the model ranks by predicted binding strength to EGFR C797S.