Smiles-based bioactivity prediction through molecular encoder selection and data augmentation.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Quantitative structure-activity relationship (QSAR) modeling

Background:

• Quantitative prediction of inhibitor potency accelerates drug discovery.
• Data-driven approaches, including AI, are increasingly vital in pharmaceutical research.
• Benchmarking challenges drive innovation in AI drug discovery.

Purpose of the Study:

• To develop a machine learning methodology for predicting IC50 values of ASK1 inhibitors from SMILES.
• To identify the optimal combination of molecular encoders and regression models for bioactivity prediction.
• To establish a reproducible benchmark for single-target bioactivity prediction.

Main Methods:

• Developed a SMILES-based machine learning workflow.
• Systematically compared sequence- and graph-based molecular representations (ChemBERTa-2, MolCLR).
• Utilized pre-trained encoders, regression models, data augmentation (embedding-level mix-up), and hyperparameter tuning (SVR).

Main Results:

• Achieved the highest overall predictive performance in the "Jump AI(.py) 2025" competition.
• ChemBERTa-77M-MLM embeddings with Support Vector Regression (SVR) showed superior predictive power.
• Embedding-level mix-up augmentation and SVR hyperparameter tuning further enhanced performance.

Conclusions:

• Careful SMILES preprocessing and encoder selection are critical for accurate IC50 prediction.
• The proposed methodology improves model generalizability, especially with limited or imbalanced data.
• This approach offers a valuable tool for efficient drug discovery and can be extended to other kinase inhibitors.