Abstract
Predicting protein‒protein interactions (PPIs) plays a crucial role in understanding biological processes. Although biological experimental methods can identify PPIs, they are costly, time-consuming, labor-intensive, and often lack stability. In contrast, computational approaches for PPI prediction, particularly deep learning methods, can efficiently learn representations from protein sequences. However, the generalizability, robustness, and stability of computational PPI prediction models still need improvement, especially for species with limited verified PPI data. Protein embeddings generated by protein language models can extract features from protein sequences and reflect hierarchical biological structures, making them suitable for predicting PPIs. Therefore, in this study, we propose a novel protein sequence-based PPI prediction framework designed for generalized PPI assessment by integrating two protein language models (PLMs) and an enhanced deep neural network (MPIDNN-GPPI). Specifically, the sequences are embedded using two protein language models, Ankh and ESM-2. A deep neural network is then used to learn representations from the feature vectors produced by PLMs. Subsequently, a multi-head attention mechanism is introduced to capture long-range dependencies and fuse them with DNN-derived representations. Finally, a deep neural network is applied to assess the probability of interaction between two proteins. To evaluate the performance of MPIDNN-GPPI, nine PPI datasets were collected from the STRING database, covering a diverse set of species: five datasets from mammals (D. melanogaster, C. elegans, S. cerevisiae, H. sapiens, and M. musculus), and four datasets from plants (O. sativa, A. thaliana, G. max, and Z. mays). When trained on H. sapiens, MPIDNN-GPPI achieved AUC values of 0.959, 0.966, 0.954, and 0.916 on independent test sets for M. musculus, D. melanogaster, C. elegans, and S. cerevisiae, respectively. These results represent the best performance among all PPI models compared in this study. Similarly, when trained on O. sativa, the model achieved AUC values of 0.96, 0.95, and 0.913 on independent datasets for A. thaliana, G. max, and Z. mays, respectively. Ablation experiments demonstrated that models combining Ankh and ESM-2 outperformed those relying on a single protein language model. Furthermore, MPIDNN-GPPI, which incorporates multi-head attention and deep neural networks (DNN), achieved superior performance compared to models using DNN alone. These findings indicate that MPIDNN-GPPI possesses strong generalization capability for cross-species PPI prediction. The proposed model, trained on one species, can be effectively applied to accurately predict PPIs in other species.