FusionEP: Adaptive Cross-Modal Deep Fusion Network for Essential Protein Identification via Multi-Omics Integration and Contrastive Learning

Identifying essential proteins lays a foundation for dissecting disease mechanisms and mining therapeutic targets. Existing algorithms mostly adopt unimodal inputs and struggle to capture multi-level biological interactions from genomics, subcellular localization and proteomics. We present FusionEP, which is an adaptive cross-modal deep fusion network built upon contrastive learning to predict essential proteins. FusionEP integrates gene expression profiles, subcellular localization features, and protein interaction map structures. For gene expression information, we design a six-layer Transformer with 12 attention heads and learnable positional encodings for feature extraction. For subcellular features, a multi-scale convolutional neural network (CNN) with four kernel sizes and a channel spatial attention mechanism is designed for feature extraction. A four-layer Graph Attention Network (GAT) with dynamic attention, residual connections, and target node readout functionality is designed to extract feature from protein–protein interaction (PPI) networks. FusionEP employs an adaptive fusion gate for dynamic feature aggregation and supervised contrastive learning for cross-modal alignment. Furthermore, prototype-based classification with focal loss optimization is used to address class imbalance. We evaluated the model’s performance on the yeast biological dataset. The results showed that FusionEP achieved an AUC-ROC of 0.9523 ± 0.003 in five independent experiments. FusionEP provides an end-to-end solution for multimodal protein analysis. This work sets a new method for integrating heterogeneous biological data in precision medicine applications.