MOTIVATION: Although deep learning has significantly advanced the field of protein function prediction, current approaches are limited by their reliance on a narrow set of modalities. Specifically, they primarily rely on sequence patterns and treat protein domain data and functional labels merely as categorical tags. Consequently, they fail to capitalize on the semantic richness embedded within their textual definitions. These constraints hinder their ability to generalize to novel labels. To tackle this issue, we present MZSGO, a multimodal zero-shot framework that fuses evolutionary signals from protein language models with semantic features derived from large language models (LLMs). By employing an adaptive gated fusion mechanism, MZSGO effectively aligns sequence-based and text-based modalities to enable robust predictions for unseen labels. RESULTS: By unifying protein representations and functional annotations, we bridge the semantic gap that limits current approaches. Results indicate that while our model remains competitive on supervised benchmarks, it demonstrates a marked advantage over existing methods in zero-shot tasks. It specifically excels at recognizing previously unseen long-tail and novel Gene Ontology (GO) terms. AVAILABILITY AND IMPLEMENTATION: The source code and datasets are available at https://github.com/toxic-byte/MZSGO.