ESMR4FBP: A pLM-based regression prediction model for specific properties of food-derived peptides optimized multiple bionic metaheuristic algorithms.

Due to the growing emphasis on food safety, peptide research is increasingly focusing on food sources. Traditional methods for determining peptide properties are expensive. While artificial intelligence (AI) models can reduce cost, existing peptide models often lack accuracy. This study aimed to develop a regression model capable of predicting peptide properties. We integrated the ESM-2 model with the LSTM architecture and optimized the model structure using three metaheuristic algorithms, including WOA, SSA, and HHO. Using an antioxidant tripeptide dataset, our model achieved an R of 0.9458 and RMSE of 0.3135, outperforming the state-of-the-art (SOTA) model by 11.66 % and 50.00 %, respectively. The developed model was further applied to the bitter peptide dataset, resulting in R of 0.8385 and RMSE of 0.4414, respectively. These results suggest that our model has the potential to accurately predict the properties of various types of peptides.