"drug protein interaction dataset"

"Accurate drug-target interaction (DTI) prediction is essential for accelerating computational drug discovery. However, current graph-based methods primarily rely on 1D sequences or basic 2D topologies, failing to effectively capture the fine-grained 3D geometry and intrinsic topological stability required for physically plausible predictions. Here, we propose the Heterogeneous Equivariant and Topological Network (HET-Net), a novel multi-modal framework that comprehensively models spatial and topological constraints. To robustly represent ligands, HET-Net features a multi-scale spatial drug perception module that integrates 3D-aware graph attention networks with persistent homology, simultaneously capturing local atomic arrangements and stable global topologies (e.g., persistent rings). Concurrently, target proteins are modeled via a geometric equivariant structural encoder mathematically guaranteeing 3D rotational equivariance while preserving critical spatial directionality. This structural encoding is further augmented by evolutionary context derived from protein language models. Finally, a hybrid architecture synergistically integrates these geometric, topological, and evolutionary descriptors. Extensive benchmarking across seven datasets demonstrates that HET-Net significantly outperforms state-of-the-art methods. Notably, HET-Net exhibits exceptional robustness in cold-start scenarios, highlighting its capability to generalize across novel chemical and biological spaces for real-world drug discovery."