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

Molecular property prediction with
deep learning often employs
self-supervised learning techniques to learn common knowledge through
masked atom prediction. However, the common knowledge gained by masked
atom prediction dramatically differs from the graph-level optimization
objective of downstream tasks, which results in suboptimal problems.
Particularly for properties with limited data, the failure to consider
domain knowledge results in a direct search in an immense common space,
rendering it infeasible to identify the global optimum. To address
this, we propose MPCD, which enhances pretraining transferability
by aligning the optimization objectives between pretraining and fine-tuning
with domain knowledge. MPCD also leverages multitask learning to improve
data utilization and model robustness. Technically, MPCD employs a
relation-aware self-attention mechanism to capture molecules’
local and global structures comprehensively. Extensive validation
demonstrates that MPCD outperforms state-of-the-art methods for absorption,
distribution, metabolism, excretion, and toxicity (ADMET) and physicochemical
prediction across various data sizes.