Benchmarking deep learning models for predicting anticancer drug potency (IC50) with insights for medicinal chemists

The study provides unique insights for medicinal chemists by analyzing the influence of compounds' physicochemical properties (e.g., molecular weight, lipophilicity, hydrogen bond donors/acceptors) and cell line tissue types on prediction accuracy. It was found that prediction errors exhibited weak correlations (r ≈ 0.3) with molecular weight (MW) and hydrogen bond acceptors (HBA), and negligible correlations with other properties like clogP, RotB, and Fsp³.

This suggests that the models' potency prediction accuracy remains largely consistent across compounds with diverse sizes, polarities, flexibility, and solubilities, a significant advantage over traditional molecular docking programs which often show decreased accuracy with increasing molecular complexity. Furthermore, the tissue type of cell lines did not significantly influence drug potency prediction accuracy, with a log10 error variation of 0.4, substantially smaller than the experimental variability.

A critical limitation identified is the sharply reduced accuracy of DL models for unseen compounds and out-of-distribution (OOD) samples. This highlights the need for improved generalization strategies, such as domain generalization and meta-learning, and better chemical feature representation. The inherent variability in experimental IC50 measurements (up to 400% for the same drug-cell line pair across different protocols) also poses a challenge, emphasizing the utility of the newly proposed EVAPA metric.

Future work will focus on applying advanced deep learning architectures, incorporating OOD learning strategies, and leveraging more diverse training data to enhance IC50 prediction for novel compounds. Collaboration with industry and academia is crucial to assess the real-world utility of these models in drug discovery pipelines, with a strong emphasis on enhancing model explainability to guide lead optimization effectively.