MPCD: A Multitask Graph Transformer for Molecular Property Prediction by Integrating Common and Domain Knowledge

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Deep learning for molecular property prediction often uses masked atom prediction for pretraining.
• This pretraining objective can differ from downstream tasks, leading to suboptimal performance, especially with limited data.

Purpose of the Study:

• To enhance pretraining transferability in molecular property prediction.
• To address the limitations of masked atom prediction by aligning pretraining and fine-tuning objectives with domain knowledge.

Main Methods:

• Proposed MPCD (Molecular Property Prediction with Contrastive Domain knowledge alignment).
• Employed a relation-aware self-attention mechanism for comprehensive structure capture.
• Utilized multitask learning to improve data efficiency and model robustness.

Main Results:

• MPCD demonstrated superior performance compared to state-of-the-art methods.
• Achieved significant improvements in predicting absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties.
• Showcased effectiveness across various data sizes for physicochemical property prediction.

Conclusions:

• Aligning pretraining objectives with domain knowledge is crucial for effective molecular property prediction.
• MPCD offers a robust and data-efficient framework for enhancing deep learning models in cheminformatics.
• The proposed method shows promise for accelerating drug discovery and materials science.