Large Language Models Enable Semantic Alignment for Cold-Start Compound-Protein Interaction Prediction

Cold-start compound-protein interaction (CPI) prediction remains a major challenge in drug discovery, as models must generalize to unseen compounds, targets, or their combinations beyond the training distribution. Most existing methods rely on shallow feature fusion and primarily fit observed interaction patterns, failing to capture semantic correspondence between heterogeneous molecular entities and thus generalize poorly under cold-start conditions. We propose SACS-CPI, a multimodal framework for cold-start CPI prediction based on large language model (LLM)-driven semantic alignment. By mapping compounds and proteins into a shared semantic space, SACS-CPI strengthens semantic consistency between interacting entities, reduces dependence on observed interaction pairs, and improves prediction for unseen compound-protein combinations. The framework further integrates general semantic knowledge learned by large language models with task-specific features through dynamic gating, while bidirectional cross-attention and bilinear pooling are used to capture high-order interaction patterns. Extensive evaluations across multiple benchmark datasets demonstrate that SACS-CPI consistently outperforms state-of-the-art methods, with robust gains under compound cold-start, protein cold-start, and blind-start settings. These results indicate that semantic alignment provides an effective inductive bias for cold-start CPI prediction. The code and data are available at: https://github.com/yangkun021224/SACS-CPI/.