Biologically relevant transfer learning improves transcription factor binding prediction.

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Deep learning excels at transcription factor (TF) binding prediction but demands extensive training data.
• Transfer learning offers a solution by reducing data requirements and boosting performance compared to training individual models.

Purpose of the Study:

• To evaluate a transfer learning strategy for TF binding prediction.
• To assess the impact of pre-training with biologically relevant TFs on model performance.

Main Methods:

• A two-step transfer learning approach: multi-task pre-training followed by single-task fine-tuning.
• Initializing single-task models with weights from a pre-trained multi-task model.
• Training single-task models at a lower learning rate.

Main Results:

• Transfer learning significantly improves TF binding prediction accuracy, especially for TFs with limited ChIP-seq data (~500 peak regions).
• Pre-training with biologically relevant TFs enhances model performance.
• Model interpretation reveals that pre-training refines learned features to match TF binding motifs and capture additional biological information.

Conclusions:

• Transfer learning is a powerful and effective technique for TF binding prediction.
• This strategy improves model performance and data efficiency, particularly in scenarios with limited biological data.