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

Area of Science:

• Computational chemistry
• Pharmacology
• Bioinformatics

Background:

• Accurate prediction of small molecule potency (IC50) is crucial for developing new anticancer drugs.
• Deep learning (DL) models offer potential for improving IC50 prediction accuracy.

Purpose of the Study:

• To benchmark five DL models (DeepCDR, DrugCell, PaccMann, Precily, tCNN) against a baseline for IC50 prediction.
• To evaluate model performance using diverse error metrics, including a novel Experimental Variability-Aware Prediction Accuracy statistic.
• To assess model generalizability across different datasets, cell lines, and novel compounds.

Main Methods:

• Standardized GDSC datasets and recent anticancer compounds were used for benchmarking.
• Five DL models and a mean-based Baseline were evaluated.
• Performance was assessed using standard error metrics, percentage error, log error, three-sigma limit, and a new accuracy statistic.

Main Results:

• DL models performed well on random splits and unseen cell lines but showed decreased accuracy for unseen compounds.
• DeepCDR, DrugCell, and tCNN demonstrated slightly superior performance across most tests.
• Several DL models did not significantly outperform the Baseline, and prediction error showed weak correlation with compound/cell line properties.

Conclusions:

• While DL models show potential in IC50 prediction, their performance on novel compounds requires further improvement.
• The developed web server provides a practical tool for predicting IC50 values of new compounds against cancer cell lines.
• Further research is needed to understand and improve the correlation between prediction errors and molecular/biological properties.