MPIDNN-GPPI: multi-protein language model with an improved deep neural network for generalized protein‒protein...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Predicting protein-protein interactions (PPIs) is vital for understanding biological processes.
• Experimental PPI identification is costly and time-consuming.
• Computational methods, especially deep learning, offer efficient alternatives but struggle with generalizability, robustness, and stability, particularly for species with limited data.

Purpose of the Study:

• To develop a novel, sequence-based protein-protein interaction (PPI) prediction framework (MPIDNN-GPPI) with enhanced generalizability and robustness.
• To leverage protein language models (PLMs) for feature extraction and deep neural networks (DNNs) for interaction prediction.
• To improve the accuracy and stability of computational PPI prediction, especially across different species.

Main Methods:

• Utilized two protein language models, Ankh and ESM-2, to generate protein sequence embeddings.
• Employed a deep neural network (DNN) to learn representations from PLM-generated feature vectors.
• Integrated a multi-head attention mechanism to capture long-range dependencies and fuse them with DNN representations for interaction probability assessment.

Main Results:

• MPIDNN-GPPI achieved high AUC values across diverse species, demonstrating strong cross-species prediction performance (e.g., 0.959 AUC when trained on H. sapiens and tested on M. musculus).
• Models combining Ankh and ESM-2 embeddings outperformed those using a single PLM.
• The inclusion of multi-head attention significantly improved performance compared to DNN alone, confirming the model's generalization capability.

Conclusions:

• MPIDNN-GPPI exhibits significant generalization capability for cross-species PPI prediction.
• The proposed framework effectively predicts PPIs even when trained on data from a single species.
• This approach offers a more efficient and accurate solution for PPI prediction in diverse biological contexts.