BERT-Kcr

BERT-Kcr predicts lysine crotonylation (Kcr) sites on proteins using transfer learning with pre-trained Bidirectional Encoder Representations from Transformers (BERT) models to capture sequence context for post-translational modification site identification.

Key Features:

• Predicts lysine crotonylation (Kcr) sites: Focuses specifically on identifying Kcr, a post-translational modification implicated in cell differentiation and metabolism.
• Transfer learning with pre-trained BERT models: Employs Bidirectional Encoder Representations from Transformers (BERT) models adapted from NLP to encode protein sequence context.
• Amino acid tokenization as NLP tokens: Treats each amino acid as analogous to a word to capture complex patterns and dependencies within sequences.
• BiLSTM downstream model: Processes BERT-encoded features through a Bidirectional Long Short-Term Memory (BiLSTM) network to capture sequential information in both forward and backward directions.
• Performance (10-fold cross-validation AUROC 0.983): Achieved an AUROC of 0.983 in 10-fold cross-validation.
• Independent-test comparison to Deep-Kcr (~5% AUROC improvement): Outperformed the state-of-the-art model Deep-Kcr by approximately 5% AUROC on independent datasets.

Scientific Applications:

• Kcr site identification: Enables computational prediction of lysine crotonylation sites to support proteome annotation of PTMs.
• Functional studies of cell differentiation and metabolism: Facilitates investigation of Kcr roles in biological processes such as cell differentiation and metabolism.
• Complement to experimental PTM mapping: Serves as a computational alternative to complement experimental identification of Kcr sites.

Methodology:

Protein sequences are encoded using pre-trained BERT models with amino acids treated as tokens, and the encoded features are processed through a Bidirectional Long Short-Term Memory (BiLSTM) network as part of a transfer learning approach.