GRASP: Gene-Relation Adaptive Soft Prompt for Universal Gene Network Inference

Gene networks (GNs) represent the complex molecular interactions that regulate cellular states. Existing sequence and graph based deep learning approaches have advanced GN inference but remain constrained by their reliance on specialized data. Large language models (LLMs) offer a promising alternative by reframing structured biological problems as natural language inference tasks. However, a major bottleneck lies in prompt design, where fixed prompts or task-specific soft prompts cannot adapt to the heterogeneity of gene-gene relationships. Here we introduce GRASP (Gene-Relation Adaptive Soft Prompt), a framework that performs universal GN inference directly from gene symbols. GRASP generates pair-specific adaptive tokens by encoding gene-level representations and relational contrasts into latent vectors, enabling context-aware inference while preserving parameter efficiency. Across large-scale protein-protein interaction (PPI) benchmarks, GRASP consistently outperforms other baselines, achieving ROC-AUC values up to 0.937. GRASP further demonstrates strong cross-species transferability, robust performance in phosphorylation network inference, and discovery potential by recovering previously mislabeled gene interactions.