BERT-Kgly: A Bidirectional Encoder Representations From Transformers (BERT)-Based Model for Predicting Lysine...

Area of Science:

• Biochemistry
• Computational Biology
• Proteomics

Background:

• Protein lysine glycation is a crucial posttranslational modification (PTM) that alters protein function and can lead to disease.
• Identifying glycation sites is vital for understanding disease mechanisms and developing treatments.
• Experimental methods for identifying glycation sites are costly and time-consuming.

Purpose of the Study:

• To develop an efficient and accurate computational predictor for protein lysine glycation sites.
• To leverage advanced machine learning techniques, specifically Bidirectional Encoder Representations from Transformers (BERT), for this prediction task.

Main Methods:

• Utilized embedding features from pretrained BERT models to represent protein segments.
• Explored three different pretrained BERT models for optimal feature extraction.
• Employed three distinct deep neural networks as downstream models for prediction.
• Trained and validated models using protein sequence data.

Main Results:

• The predictor, BERT-Kgly, demonstrated superior performance in identifying protein lysine glycation sites.
• The model utilizing embeddings from a BERT model pretrained on a large UniProt dataset (556,603 sequences) achieved the best results.
• Independent testing confirmed BERT-Kgly's superiority over existing computational methods.

Conclusions:

• BERT-Kgly provides a highly efficient and accurate computational approach for predicting protein lysine glycation sites.
• This tool can serve as a valuable supplement to experimental methods, accelerating research in glycation-related diseases.
• The findings highlight the potential of large-scale pretrained language models in advancing PTM site prediction.