Enhancing robustness in protein function prediction via missing modality imputation and adaptive multimodal fusion

Protein function prediction is a fundamental task in the post-genomic era and has made significant progress through the integration of multiple biological modalities. However, the main challenges in this field, such as incomplete modality information and suboptimal fusion strategies, continue to hinder predictive performance. To address these limitations, we propose ProMIAF, a novel framework that combines Missing modality Imputation with Adaptive multimodal Fusion for robust Protein function prediction. Specifically, ProMIAF employs advanced protein generation techniques to effectively recover missing structural and textual modalities, mitigating the impact of incomplete data. Modality-specific encoders are then used to extract intrinsic features of different modalities, producing enriched protein representations. These features are integrated via a gated attention mechanism that dynamically reweights modality contributions, enabling effective fusion of diverse biological evidences. Additionally, ProMIAF incorporates a network propagation module to exploit topological structures from sequence homology and protein-protein interaction networks, further enhancing predictive accuracy. Experimental results demonstrate that ProMIAF outperforms state-of-the-art methods in predicting novel functional annotations. On the biological process branch, ProMIAF achieves improvements of approximately 8.21%, 8.90% and 2.61% over ProtGO in terms of AUPR, S_min and F_max, respectively. Even in the absence of structural and textual modalities, ProMIAF can effectively leverage cross-modal complementarity to improve the robustness and accuracy of protein function prediction.