HALO: Hybrid Attention Model for Subcellular Localization

Subcellular localization prediction is critical for understanding protein functions and interactions, providing insights into cellular mechanisms and potential therapeutic targets. We propose HALO (Hybrid Attention model for subcellular LOcalization), a framework that integrates semantic embeddings from fine-tuned protein language models (e.g., ESM) with structural information derived from AlphaFold. HALO uses a graph attention network (GAT) to incorporate biochemical, structural, and sequence-derived features into a unified representation, while dynamically balancing their contributions. Crucially, the design allows HALO to operate in two modes: (i) a sequence-only mode, where predictions are made from the fine-tuned protein language model (PLM) when structural data are unavailable, and (ii) a hybrid mode, where structural adjacency and biochemical features complement PLM predictions, especially in low-confidence regions. We evaluate HALO on multiple datasets with minimal homology between training and test sets, where it achieves competitive performance across key metrics. By flexibly combining sequence-based and structure-informed predictions, HALO addresses the limitations of relying on a single modality and offers an adaptable framework for accurate and generalizable subcellular localization.