MPCD: A Multitask\nGraph Transformer for Molecular\nProperty Prediction by Integrating Common and Domain Knowledge

Molecular property prediction with\ndeep learning often employs\nself-supervised learning techniques to learn common knowledge through\nmasked atom prediction. However, the common knowledge gained by masked\natom prediction dramatically differs from the graph-level optimization\nobjective of downstream tasks, which results in suboptimal problems.\nParticularly for properties with limited data, the failure to consider\ndomain knowledge results in a direct search in an immense common space,\nrendering it infeasible to identify the global optimum. To address\nthis, we propose MPCD, which enhances pretraining transferability\nby aligning the optimization objectives between pretraining and fine-tuning\nwith domain knowledge. MPCD also leverages multitask learning to improve\ndata utilization and model robustness. Technically, MPCD employs a\nrelation-aware self-attention mechanism to capture molecules’\nlocal and global structures comprehensively. Extensive validation\ndemonstrates that MPCD outperforms state-of-the-art methods for absorption,\ndistribution, metabolism, excretion, and toxicity (ADMET) and physicochemical\nprediction across various data sizes.