Transformer-based classification and interpretability of NR3C1 expression patterns in OSCC: Metabolic adaptation...

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Oral squamous cell carcinoma (OSCC) survival depends on oncogenic proteins.
• Classifying protein sequences is challenging due to complex relationships.
• Transformer language models show promise for biological sequence analysis.

Purpose of the Study:

• To evaluate transformer models for NR3C1 peptide sequence classification.
• To compare transformer performance against deep learning and traditional machine learning methods.

Main Methods:

• Four transformer models (BERT, RoBERTa, DistilBERT, ALBERT) were assessed.
• Models were trained on UniProt sequences (200-500 aa and >500 aa).
• Sequences were tokenized, padded/truncated to 512 tokens, and data split for training/validation.

Main Results:

• RoBERTa achieved the highest F1-score (0.8574), outperforming deep learning (0.763) and traditional methods (0.693).
• ALBERT showed comparable performance to RoBERTa with significantly fewer parameters.
• Transformer models demonstrated superior accuracy and efficiency.

Conclusions:

• Transformer models, especially RoBERTa, are highly effective for NR3C1 protein sequence classification.
• These models surpass conventional approaches in accuracy and efficiency.
• Advancements in transformer technology enhance computational biology capabilities.