NyxBind: enhancing deep neural representations for transcription factor binding site prediction via contrastive...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Pretrained genomic language models excel at DNA sequence patterns but struggle with subtle transcription factor binding site (TFBS) motif discrimination.
• Contrastive learning shows promise in improving embedding discriminative power by modeling instance similarities and differences.

Purpose of the Study:

• Introduce NyxBind, a TFBS prediction model utilizing contrastive learning across multiple TFBS types.
• Enhance regulatory sequence representations for improved TFBS prediction accuracy.

Main Methods:

• Applied contrastive learning across diverse TFBS types within the NyxBind model.
• Evaluated NyxBind on 159 TFBS prediction tasks.
• Enabled parameter-efficient fine-tuning and motif visualization.

Main Results:

• NyxBind achieved superior performance across all evaluation metrics on 159 TFBS prediction tasks.
• Demonstrated a 4.71 percentage point improvement in Matthews Correlation Coefficient over DNABERT2.
• Achieved accurate motif visualization, closely aligning with experimental data.

Conclusions:

• NyxBind represents a significant advancement in TFBS prediction through contrastive learning.
• The model offers enhanced accuracy, flexibility in fine-tuning, and interpretable motif visualizations.
• NyxBind effectively captures subtle differences in regulatory sequences crucial for gene regulation studies.