ESMDNN-PPI: a new protein–protein interaction prediction model developed with protein language model of ESM2 and deep neural network

Abstract

Abstract Protein–protein interaction (PPI) plays an important role in the biological process. While, there are limitations of long spend time and high labor cost in traditional lab based PPIs detection approaches. Although many computation-based methods have been proposed for prediction of PPIs, achieving high predictive performance and overcoming low generalization performance remain challenging issues. This study developed a novel PPIs prediction method by combining protein language model (PLM) of ESM2 and deep neural network, which show high predictive and generalization performance. Specifically, after protein-coding of protein sequence, the PLM of ESM2 is embedded. Then pre-training weight which trained on human dataset was transferred to other species dataset. The generalization of the model we established is tested on four independent datasets. The results show that values of area under precision–recall curve (AUPR) and area under the receiver operating characteristic (ROC) curve achieved 93.06% and 98.69% on human dataset respectively. AUPR values achieved 87.54%, 84.95%, 81.99%, and 66.23% on datasets of Mus musculus, Drosophila melanogaster, Caenorhabditis elegans and Saccharomyces cerevisiae, which are higher of 13.61%–78.78%, 11.35%–75.71%, 12.76%–73.1% and 11.77%–56.94% than multilayer perceptron + convolutional neural network (MLP + CNN), MLP + gated recurrent unit (GRU) and MLP + CNN + GRU based models on these four dataset respectively. The results indicate that this PPIs prediction method we developed can extract features that better characterize the nature of PPIs with protein sequence, and achieving a high predictive and generalization performance for predicting PPIs.