Evaluation of deep learning approaches for modeling transcription factor sequence specificity.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Transcription factors (TFs) are crucial regulators of gene expression.
• Accurate identification of TF binding sites and prediction of binding affinities are essential for understanding gene regulation.
• Deep learning (DL) models show promise for predicting DNA-TF binding, but architectural comparisons are lacking.

Purpose of the Study:

• To systematically compare the performance of different deep learning architectures for predicting TF binding specificity.
• To identify the most effective DL architecture for analyzing SELEX-seq and HT-SELEX data.

Main Methods:

• Applied four distinct deep learning architectures to SELEX-seq and HT-SELEX datasets.
• Utilized data from three species and 35 TF families.
• Evaluated model performance using 10-fold cross-validation.

Main Results:

• The hybrid Convolutional Neural Network (CNN) + Deep Neural Network (DNN) architecture achieved the highest performance.
• Demonstrated the relative strengths of different DL models in predicting TF binding specificity.

Conclusions:

• The hybrid CNN + DNN model is highly effective for predicting TF binding specificity.
• This study provides a benchmark for DL model selection in TF binding prediction.
• Findings are applicable to future modeling efforts with increasing high-throughput TF binding data.