Inductive transfer learning for molecular activity prediction: Next-Gen QSAR Models with MolPMoFiT

Area of Science:

• Computational Chemistry
• Machine Learning
• Drug Discovery

Background:

• Deep neural networks (DNNs) excel at predicting molecular properties but typically require large datasets.
• Smaller datasets with challenging endpoints are common in drug discovery and require efficient modeling techniques.
• Leveraging unlabeled chemical data can enhance model performance for specific compound series.

Purpose of the Study:

• To develop an effective transfer learning method for quantitative structure-property/activity relationship (QSPR/QSAR) modeling.
• To improve prediction accuracy for QSPR/QSAR tasks using smaller, specific chemical datasets.
• To utilize large publicly available unlabeled molecular datasets for enhancing model performance.

Main Methods:

• Proposed Molecular Prediction Model Fine-Tuning (MolPMoFiT) approach.
• Employed self-supervised pre-training on one million unlabeled molecules from ChEMBL.
• Task-specific fine-tuning on smaller chemical datasets for QSPR/QSAR tasks.

Main Results:

• MolPMoFiT demonstrated strong predictive performances across four benchmark datasets: lipophilicity, FreeSolv, HIV, and blood-brain barrier penetration.
• Achieved competitive results compared to existing state-of-the-art machine learning techniques.
• Effectively adapted a large pre-trained model to specific QSPR/QSAR tasks with limited data.

Conclusions:

• MolPMoFiT offers an effective transfer learning strategy for QSPR/QSAR modeling, particularly beneficial for smaller datasets.
• Self-supervised pre-training on extensive unlabeled data significantly enhances model generalization and accuracy.
• The approach provides a valuable tool for drug discovery, enabling reliable predictions even with limited specific experimental data.