Graph Transformer Networks for Predicting Anti-HIV Active Molecules
Zhaoyao Yan, Cangfeng Ding, Lerong Ma, Lu Cao · 2024
Predicting active anti-HIV molecules is a fundamental task in the field of drug development. It drives progress in anti-HIV drug discovery, including aspects such as design, synthesis and activity evaluation. However, existing methods employ atom-level labelling techniques that ignore the relationship between substructural features such as functional groups and pharmacophores in the molecule. In addition, these methods have difficulty taking full account of the structural information of molecules, which limits their scalability in the drug space. To address these issues, this paper proposes an end-to-end neural network-based approach, called MolGTPool, which utilizes Graph Transformer and Graph Pooling to predict potential antiHIV active molecules. This proposed approach employs multilevel union modules to concatenate the outputs from diverse molecular embedding layers, thus enabling the final layer to amalgamate information from its precursors. Experiments show that MolGTPool is able to predict the structure of molecules with anti-HIV activity more accurately. In addition, MolGTPool shows excellent scalability on the MoleculeNet.