Modeling binding specificities of transcription factor pairs with random forests

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Transcription factors (TFs) regulate gene expression by binding to specific DNA regions.
• Cooperative TF-TF binding involves two TFs binding DNA as a pair.
• Existing methods using position weight matrices (PWMs) may not fully capture the complexity of pairwise TF-DNA interactions.

Purpose of the Study:

• To develop novel computational methods for predicting joint transcription factor-transcription factor (TF-TF) binding sites.
• To improve the accuracy of predicting TF-TF binding site interactions beyond current state-of-the-art methods.

Main Methods:

• Proposed two random forest (RF) based methods: ComBind and JointRF.
• Trained models using large-scale CAP-SELEX DNA libraries.
• JointRF utilizes pre-defined pairwise PWMs, while ComBind considers multiple orientations and spacings simultaneously.

Main Results:

• JointRF achieved an AUROC of 0.75, outperforming pairwise PWMs (AUROC 0.59).
• ComBind achieved an AUROC of 0.78, significantly outperforming both PWMs and JointRF (p<0.00195).
• ComBind effectively predicts TF-TF binding sites without prior knowledge of binding preferences.

Conclusions:

• Random forest models are suitable for modeling TF-TF-DNA binding specificities.
• ComBind represents a significant advancement in predicting pairwise TF-TF binding site accuracy.
• Further research is needed to evaluate ComBind for practical applications.