Improved ADME Prediction by Multitask Pretraining on Predicted Data: Insights from the ASAP-Polaris-OpenADMET Blind...

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Bioinformatics

Background:

• Absorption, distribution, metabolism, and excretion (ADME) properties are critical for drug development success.
• Limited public data for ADME prediction hinders machine learning model development.
• Industry-released predicted labels offer new opportunities for training data.

Purpose of the Study:

• To develop an improved machine learning approach for predicting molecular ADME profiles.
• To leverage transfer learning and graph neural networks (GNNs) for ADME prediction.
• To utilize publicly available predicted labels for model pretraining and fine-tuning.

Main Methods:

• Adoption of transfer learning using a multitask graph neural network (GNN).
• Rich representation learning and focused fine-tuning on experimental data.
• Exploration of a pretraining strategy integrating experimental and predicted labels.

Main Results:

• Achieved competitive results in the ASAP-Polaris-OpenADMET antiviral ADME challenge 2025.
• Ranked fourth on aggregated mean absolute error (MAE) and tied second on aggregated Pearson R.
• Post-competition optimization surpassed the third-place entry in MAE without proprietary data.

Conclusions:

• Transfer learning with GNNs is effective for ADME prediction, especially with limited experimental data.
• Integrating predicted and experimental labels shows promise for pretraining strategies.
• The study demonstrates a viable approach for utilizing predicted labels in real-world drug discovery tasks.