Multi-task molecular representation learning based on soft prompting of the important subgraph.

Understanding the intricate relationship between molecular structure and biological activity is fundamental to modern drug discovery and genomic research. However, capturing the full geometric and functional complexity of molecules remains a significant challenge. To address this, we introduce a Multi-task Molecular Representation Learning framework based on Soft Prompting of the Important Subgraph (MMRL). In molecular generation, retrosynthetic prediction, and molecular property prediction tasks, the model utilizes the important subgraph information to generate soft prompts, which help these tasks obtain the better molecular representations through the shared encoding. Additionally, the model addresses the issue that graph neural networks mainly focus on neighboring atoms and connectivity information and do not effectively utilize holistic molecular geometry structure for molecular encoding. The model achieves significant performance improvements compared to many classical baseline systems on ZINC250K and QM9 dataset in molecular generation task, USPTO-50K dataset (unknown and known reaction classes) in retrosynthetic prediction task, and eight benchmark datasets in molecular property prediction. These results demonstrate that MMRL not only achieves superior computational performance but also provides more biologically meaningful representations, offering a promising tool for accelerating the identification and optimization of novel therapeutic candidates in complex biological systems.