Benchmarking transcription factor binding site prediction models: a comparative analysis on synthetic and biological...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Transcription factors (TFs) regulate gene expression by binding to specific DNA sequences (TFBSs).
• Accurate TFBS identification is vital for understanding cellular dynamics and regulatory mechanisms.
• Traditional methods like position weight matrices (PWMs) have limitations in capturing complex DNA-binding patterns.

Purpose of the Study:

• To systematically benchmark the predictive performance of PWM, SVM, and DL models for TFBS identification.
• To evaluate the impact of training data size, sequence length, and background data on model performance.
• To provide practical guidance for selecting appropriate TFBS prediction models and to offer a database of pretrained SVM models.

Main Methods:

• Benchmarking of PWM, SVM, and DL models using human ChIP-seq data from ENCODE.
• Evaluation of model performance based on varying training dataset sizes, sequence lengths, and kernel functions (for SVMs).
• Assessment of the influence of synthetic versus real biological background data on model training.

Main Results:

• PWMs, SVMs, and DL models exhibit varying strengths and limitations depending on the specific scenario and data characteristics.
• Factors like training dataset size and sequence length significantly impact model predictive performance.
• The choice of background data (synthetic vs. real) affects model training outcomes.

Conclusions:

• SVM- and DL-based models offer advantages over PWMs for TFBS prediction, particularly in capturing complex interactions.
• The study provides crucial insights for optimizing TFBS prediction model selection and application in regulatory genomics.
• A database of pretrained SVM models is introduced to facilitate TFBS detection and advance regulatory genomics research.