Task-specific pre-training for molecular property prediction.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Deep learning models are crucial for molecular property prediction but suffer from overfitting due to limited labeled data.
• Existing methods struggle with generalization in scenarios with scarce labeled molecular datasets.

Purpose of the Study:

• To propose TasProp, a task-specific pre-training strategy to enhance molecular property prediction, especially with small labeled datasets.
• To improve the learning of robust molecular representations for better generalization.

Main Methods:

• TasProp projects labeled and unlabeled data into a unified latent space.
• A task-specific contrastive loss is introduced to align representations with prediction tasks.
• A novel data augmentation technique is proposed to address labeled data scarcity.

Main Results:

• TasProp significantly outperforms state-of-the-art methods on multiple molecular property prediction tasks.
• The strategy shows improved performance on publicly available and curated anesthesiology datasets.
• The method effectively mitigates overfitting and enhances model generalization.

Conclusions:

• TasProp offers an effective solution for molecular property prediction with limited labeled data.
• The task-specific pre-training and data augmentation improve model robustness and predictive accuracy.
• An interactive web resource is available for easy application and exploration of the model.