MTTFsite: cross-cell type TF binding site prediction by using multi-task learning.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Transcription factor binding site (TFBS) prediction is vital for understanding gene expression.
• Supervised learning methods for TFBS prediction require substantial labeled data, which is often unavailable for specific cell types.

Purpose of the Study:

• To develop a multi-task learning framework (MTTFsite) to address the challenge of limited labeled data for TFBS prediction.
• To leverage cross-cell type data to improve TFBS prediction accuracy in data-scarce scenarios.

Main Methods:

• Proposed MTTFsite framework utilizing a shared Convolutional Neural Network (CNN) for common features and private CNNs for cell-type-specific features.
• Evaluation on 241 cell type TF pairs, comparing MTTFsite against baseline and fully shared multi-task models.
• Development of TFChrome, a gene expression prediction method integrating MTTFsite predictions with histone modification data.

Main Results:

• MTTFsite significantly outperformed baseline and fully shared models for cell types with insufficient labeled data (in >89% of pairs).
• MTTFsite showed substantial improvement over baseline and fully shared models for cell types with no labeled data (in >80% and >93% of pairs, respectively).
• Combining MTTFsite predictions with histone modification features resulted in a 5.7% performance improvement for gene expression prediction.

Conclusions:

• MTTFsite effectively addresses the data scarcity problem in TFBS prediction by leveraging multi-task learning.
• The proposed framework enhances TFBS prediction accuracy across various cell types, particularly those with limited training data.
• Integrating MTTFsite with epigenetic features offers a promising approach for accurate gene expression prediction.