AI-Assisted Design of a New Chemical Space of Potential Small Molecules Targeting ROR1 in Triple-Negative Breast Cancer

Introduction. Triple-negative breast cancer is characterized by the absence of ER, PR, and HER2 expression, which necessitates the search for alternative molecular targets. One such target is ROR1, specifically its pseudokinase domain.

Objective. Compare generative AI models based on their ability to generate structurally diverse small molecules with ROR1-targeting potential, and conduct a preliminary computational evaluation of these molecules.

Methods. AI models GPT-5.6 Luna, Gemini 3.5 Flash-Lite, and Claude Sonnet 5.5 were given the same structured prompt based on 130 compounds from four ROR1 studies. Each model generated 10 SMILES strings, which were validated by conversion to molecular graphs. Structural diversity was assessed using the Tanimoto coefficient. Docking was performed in AutoDock Vina v1.2.5 against ROR1 (PDB 6TU9), using ponatinib as the reference ligand.

Results. In-group Tanimoto distance was 0.77±0.13 for ChatGPT, 0.68±0.17 for Claude, and 0.68±0.11 for Gemini-generated sets. The mean Tanimoto similarity to the reference set was 0.23±0.12, 0.29±0.10, and 0.30±0.09, respectively. The average docking scores were −10.38±1.04, −10.27±1.07, and −10.66±0.93 kcal/mol, respectively. Gemini-ROR1-05 exhibited the best docking score (−12.03 kcal/mol, versus −12.44 kcal/mol for ponatinib). Claude-generated molecules outperformed ChatGPT, Gemini, and reference sets based on QED (0.51±0.17; range 0.25–0.77) and SA (2.57±0.17) scores.

Conclusions. All models generated valid and structurally diverse molecules. ChatGPT demonstrated the greatest in-group diversity, while Gemini achieved the best average docking score. These metrics did not reveal a clear advantage for any single model. Some compounds had docking performance close to that of ponatinib, which supports their further validation.