MZSGO-DA: Multimodal Zero-Shot Protein Function Prediction Based on Domain-Aware Adapter Pretraining

Protein function prediction is crucial for understanding biological mechanisms and advancing drug discovery. Traditional deep learning methods are constrained by a closed label space and reliance on explicit domain semantics during inference. To overcome these limitations, we propose MZSGO-DA, a method that injects lightweight adapters into ESM 2 and employs dual-channel contrastive learning, which uses sequence–domain and sequence–function channels to encode functional semantics from InterPro domains and GO definitions directly within the sequence encoder. The pretrained adapters are frozen, and a second set of adapters is introduced for task-specific tuning, enabling end-to-end inference without requiring explicit domain embeddings. Experiments across three GO ontologies show that MZSGO-DA achieves strong overall performance and competitive zero-shot generalization, excelling on the molecular function ontology. With only 7.3% trainable parameters, MZSGO-DA offers an efficient and deployable solution for protein function prediction.