DMPE: A Dual‐Branch Molecular Property Encapsulation Framework With Kolmogorov‐Arnold Networks

Accurate molecular representation learning is critical for accelerating property prediction in drug discovery. Here, we propose the Dual‐branch Molecular Property Encapsulation (DMPE) framework. Its core component, the Refined Interactive Graph Attention Framework (RIGAF), captures both intramolecular topology and intermolecular structural similarity within the broader chemical space. Subsequently, a Kolmogorov‐Arnold Network (KAN)‐based Embedding and Fusion (KAEF) module uses spline‐based transformations to integrate these graph features with complementary molecular fingerprints, enhancing model interpretability while maintaining strong generalizability. DMPE performs competitively, with superior ROC‐AUC scores on scaffold‐based BBBP (0.927) and multi‐label SIDER (0.691) benchmarks, alongside promising accuracy across nine of 14 breast cancer cell lines. Key components of our architecture are validated through ablation studies. Model reliability is bolstered by Monte Carlo dropout for uncertainty estimation, which also serves as an ensemble strategy to enhance accuracy. The framework was applied to the identification of potential hematopoietic progenitor kinase 1 (HPK1) inhibitors for hepatocellular carcinoma.