Pre-trained protein language model sheds new light on the prediction of Arabidopsis protein–protein interactions

Abstract

Protein–protein interactions (PPIs) are heavily involved in many biological processes. Consequently, the identification of PPIs in the model plant Arabidopsis is of great significance to deeply understand plant growth and development, and then to promote the basic research of crop improvement. Although many experimental Arabidopsis PPIs have been determined currently, the known interactomic data of Arabidopsis is far from complete. In this context, developing effective machine learning models from existing PPI data to predict unknown Arabidopsis PPIs conveniently and rapidly is still urgently needed. We used a large-scale pre-trained protein language model (pLM) called ESM-1b to convert protein sequences into high-dimensional vectors and then used them as the input of multilayer perceptron (MLP). To avoid the performance overestimation frequently occurring in PPI prediction, we employed stringent datasets to train and evaluate the predictive model. The results showed that the combination of ESM-1b and MLP (i.e., ESMAraPPI) achieved more accurate performance than the predictive models inferred from other pLMs or baseline sequence encoding schemes. In particular, the proposed ESMAraPPI yielded an AUPR value of 0.810 when tested on an independent test set where both proteins in each protein pair are unseen in the training dataset, suggesting its strong generalization and extrapolating ability. Moreover, the proposed ESMAraPPI model performed better than several state-of-the-art generic or plant-specific PPI predictors. Protein sequence embeddings from the pre-trained model ESM-1b contain rich protein semantic information. By combining with the MLP algorithm, ESM-1b revealed excellent performance in predicting Arabidopsis PPIs. We anticipate that the proposed predictive model (ESMAraPPI) can serve as a very competitive tool to accelerate the identification of Arabidopsis interactome.