- Published on
Building a Scaffold-Split Random Forest QSAR Co-Scientist for EGFR Inhibitor Discovery Using ChEMBL, RDKit, SHAP, and BRICS
- Authors
- Name
- DeepHorus
A new artificial intelligence-driven workflow, dubbed an autonomous AI co-scientist, has been developed to accelerate the discovery of next-generation inhibitors for the epidermal growth factor receptor (EGFR), specifically targeting the C797S mutation that confers resistance to osimertinib in non-small cell lung cancer (NSCLC). Demonstrating an end-to-end approach, the system integrates multiple computational toolsâfrom biological target identification to generative molecular designâoffering a scalable framework for drug discovery against challenging resistance mechanisms. Researchers built this pipeline by combining publicly available bioactivity data, cheminformatics libraries, and interpretable machine learning models, as detailed in a recent technical tutorial.
EGFR mutations are among the most common oncogenic drivers in NSCLC, with osimertinib serving as a third-line therapy. However, the emergence of the C797S resistance mutation diminishes osimertinib efficacy, creating an urgent need for novel inhibitors. The new AI co-scientist addresses this challenge by autonomously mining bioactivity data from the ChEMBL database, starting with target intelligence from the EGFR entry (CHEMBL203) and UniProt annotations. This step establishes a foundation for curating high-quality experimental measurements to train predictive models.
The core dataset comprised over 9,000 IC50 measurements for EGFR inhibitors, which were filtered to yield around 4,000 unique molecules with clean pIC50 values. Using RDKit, the workflow standardized structures, removed salts, and aggregated replicate measurements. Molecular representations were generated as Morgan fingerprints (2048 bits) and physicochemical descriptors, ensuring the model learned from chemically meaningful features. Scaffold diversity analysis was performed to guide dataset splitting, aiming for robust generalization across chemotypes unseen during training.
A scaffold-split Random Forest regression model was trained to predict compound potency (pIC50). Model evaluation showed strong performance, with the ability to generalize to novel chemical scaffoldsâa critical requirement for discovering new drug candidates. Interpretability was enhanced using SHAP (SHapley Additive exPlanations) and feature importance analysis, which highlighted structural features most predictive of potency, such as hydrogen bond donors and aromatic ring counts. This allows researchers to identify key molecular determinants of activity.
Beyond prediction, the AI co-scientist ventured into generative design by recombining BRICS fragments from 60 potent drug-like actives. Over 4,000 virtual analogs were generated and scored across gates for potency, drug-likeness (via QED), synthesizability (SA score), novelty relative to known actives, and developability. The shortlisted candidates were cross-referenced against PubChem, providing a filtered list of promising virtual leads. This closed-loop integration of prediction and generation represents a significant step toward autonomous drug design.
The implications are substantial: by systematizing the discovery pipeline, this workflow could dramatically reduce the time and cost of identifying new leads for resistance mutations. Experts note that while the current focus is on EGFR C797S, the architecture is adaptable to other kinase targets and beyond. The combination of open-source tools and transparent modeling fosters reproducibility and broad accessibility for academic and industrial laboratories.
Looking ahead, the team plans to extend the framework to multi-target profiling and incorporate additional data sources, such as ADMET predictions and clinical trial outcomes. As resistance mechanisms continue to challenge precision oncology, autonomous AI co-scientists like this one may become essential in rapid iterative design. This development underscores the growing synergy between cheminformatics, explainable AI, and generative chemistry in modern drug discovery.
