Computational prediction and characterization of cell-type-specific and shared binding sites.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Cell-type-specific gene expression relies on transcription factor (TF) binding.
• TF binding is influenced by sequence preferences, co-factors, and chromatin accessibility.
• Predicting cell-type-specific and shared TF binding sites computationally remains challenging.

Purpose of the Study:

• To develop and validate computational approaches for predicting and characterizing cell-type-specific and shared TF binding sites.
• To integrate diverse features for enhanced prediction accuracy.
• To compare the performance of XGBoost and convolutional neural network (CNN) models.

Main Methods:

• Proposed two computational models: one XGBoost-based and one CNN-based.
• Utilized ChIP-seq datasets from GM12878 and K562 human hematopoietic cell lines.
• Integrated multiple feature types, including DNase signals, to train the models.

Main Results:

• Both proposed computational approaches significantly outperformed competing methods.
• Identified key feature contributions for cell-type-specific and shared binding sites using SHAP values.
• Demonstrated the CNN model's ability to predict binding sites with and without DNase signals.
• Validated the generalization ability of the models across different binding factors.

Conclusions:

• The developed XGBoost and CNN models offer robust computational tools for predicting cell-type-specific and shared TF binding sites.
• Feature analysis provides insights into the determinants of TF binding specificity.
• These methods advance the understanding of gene regulation in different cell types.