Abstract
BACKGROUND: Accurate prediction of microbial protein functions is essential for understanding microbial physiology, discovering novel probiotics, and driving biotechnological innovation. However, protein function prediction remains challenging due to the hierarchical and class-imbalanced nature of functional labels, particularly in large-scale annotations such as Enzyme Commission (EC) numbers and Gene Ontology (GO) terms. Most existing deep learning approaches fail to adequately address the long-tail distribution problem. METHODS: We propose a Hierarchical Cascaded Context Network (HCCN) that explicitly models functional hierarchies and emphasizes prediction of low-frequency (long-tail) labels. For EC classification, we design a coarse-to-fine network that captures parent–child dependencies among hierarchical labels. For GO prediction, we construct a semantically grounded hierarchical structure using ontology embedding and clustering, and develop an attention-based multi-level cascade predictor to exploit structured dependencies across Biological Process (BPO), Molecular Function (MFO), and Cellular Component (CCO). To mitigate label imbalance, we introduce a dynamic resampling strategy and a hierarchical loss weighting mechanism, which enforce inter-level regularization and enhance sensitivity to rare functions. RESULTS: Experimental results show that HCCN consistently outperforms traditional sequence-alignment methods (e.g., DIAMOND, BLAST) and baseline neural networks (MLP and DeepGOPlus) across all major functional categories. On the full test set, HCCN achieves AUPR gains of up to 5.5% (EC), 6.5% (BPO), 4.9% (MFO), and 5.3% (CCO) over the best baseline. For low-frequency labels, HCCN demonstrates strong few-shot generalization, with improvements of + 11.2% (EC low) , + 6.7% (BPO low) , + 9.2% (MFO low) and + 4.6% (CCO low) in mAUPR. CONCLUSIONS: The proposed HCCN framework provides an effective solution to hierarchical and imbalanced protein function prediction, significantly improving performance on long-tail functional labels. Code and data are publicly available at: https://github.com/YangLab-BUPT/HCCN. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s12859-025-06326-7.