KnoMol: A Knowledge-Enhanced\nGraph Transformer for\nMolecular Property Prediction

Molecular\nproperty prediction (MPP) techniques are pivotal in reducing\ndrug development costs by preemptively predicting bioactivity and\nADMET properties. Despite the application of numerous deep learning\napproaches, enhancing the representational capacity of these models\nremains a significant challenge. This paper presents a novel knowledge-based\nTransformer framework, KnoMol, designed to improve the understanding\nof molecular structures. KnoMol integrates expert chemical knowledge\ninto the Transformer, emulating the analytical methods of medicinal\nchemists. Additionally, the multiperspective attention mechanism provides\na more precise way to represent ring systems. In the evaluation experiments,\nKnoMol achieved state-of-the-art performance on both MoleculeNet and\nsmall-scale data sets, surpassing existing models in terms of accuracy\nand generalization. Further research indicated that the incorporation\nof knowledge significantly reduces KnoMol’s reliance on data\nvolumes, offering a solution to the challenge of data scarcity. Moreover,\nKnoMol identified several new inhibitors of HER2 in a case study,\ndemonstrating its value in real-world applications. Overall, this\nresearch not only provides a powerful tool for MPP but also serves\nas a successful precedent for embedding knowledge into Transformers,\nwith positive implications for computer-aided drug discovery and the\ndevelopment of MPP algorithms.