Machine Learning-Based Multi-Modal and Multi-Granularity Feature Fusion Framework for Accurate Prediction of Molecular Properties

The accurate prediction of molecular properties is a pivotal component in advancing chemical engineering technology. However, the efficacy of many machine learning-based quantitative structure–property relationship (QSPR) models is highly constrained by their reliance on specific types of molecular representations. To address this issue, a multi-modal and multi-granularity feature fusion framework has been designed to fully explore diverse information sources and improve the prediction accuracy of molecular properties. Initially, a self-supervised pretrained sequence feature encoder is developed, utilizing molecular fingerprints and atomic-level information to capture the intricate features of molecules. Meanwhile, the atom-level graph knowledge is integrated with the motif-level graph knowledge by developing a hierarchical graph feature encoder, and thereby enhancing the capacity to learn molecular topological information. Subsequently, several strategies including low-rank multimodal fusion are employed to synthesize the learned features. Comprehensive evaluations across four molecular data sets demonstrate that the proposed framework achieves superior accuracy and reliability. Through the analysis of the distribution of different features and comprehensive ablation studies, the ability of the proposed framework to capture multimodal features and extract additional potential information has been demonstrated. By systematically leveraging this information, the constructed framework enhances predictive performance and stability, thereby expanding the application prospects of machine learning-based QSPR in advancing intelligent chemical processes.