KD_MultiSucc: incorporating multi-teacher knowledge distillation and word embeddings for cross-species prediction of protein succinylation sites

Abstract

Protein succinylation is a vital post-translational modification (PTM) that involves the covalent attachment of a succinyl group (-CO-CH2-CH2-CO-) to the lysine residue of a protein molecule. The mechanism underlying the succinylation process plays a critical role in regulating protein structure, stability, and function, contributing to various biological processes, including metabolism, gene expression, and signal transduction. Succinylation has also been associated with numerous diseases, such as cancer, neurodegenerative disorders, and metabolic syndromes. Due to its important roles, the accurate prediction of succinylation sites is essential for a comprehensive understanding of the mechanisms underlying succinylation. Although research on the identification of protein succinylation sites has been increasing, experimental methods remain time-consuming and costly, underscoring the need for efficient computational approaches. In this study, we present KD_MultiSucc, a model for cross-species prediction of succinylation sites using Multi-Teacher Knowledge Distillation and Word Embedding. The proposed method leverages the strengths of both Knowledge Distillation and Word Embedding techniques to reduce computational complexity while maintaining high accuracy in predicting protein succinylation sites across species. Experimental results demonstrate that the proposed predictor outperforms existing predictors, providing a valuable contribution to PTM research and biomedical applications. To assist readers and researchers, the codes and resources related to this work have been made freely accessible on GitHub at https://github.com/nuinvtnu/KD_MultiSucc/.

Keywords: bi-direction long short-term memory (Bi-LSTM); convolutional neural network (CNN); knowledge distillation; natural language processing (NLP); succinylation; word embedding.

Figure 1.

System architecture of KD_MultiSucc: A hybrid Student model combining CNN1D and Bi-LSTM, trained under multi-teacher knowledge distillation with CNN1D and Bi-LSTM Teachers

Figure 2.

Two sample logo of generic and multiple-species

Figure 3.

t-SNE visualization of the model before training by different of models on Generic dataset: (a) before training Teacher 1 model, (b) after training teahcer 1 model, (c) before training Teacher 2 model, (d) after training Teahcer 2 model, (e) before training Student model, (f) after training Student model using Knowlege distillation

Figure 4.

t-SNE visualization of the model before training by different of models on H.sapiens dataset. (a) before training Teacher 1 model, (b) after training Teahcer 1 model, (c) before training Teacher 2 model, (d) after training teahcer 2 model, (e) before training Student model, (f) after training Student model using Knowlege distillation