Predicting which genes will respond to transcription factor perturbations

Area of Science:

• Systems biology
• Genomics
• Computational biology

Background:

• Understanding transcriptional regulatory networks is crucial for predicting gene expression changes.
• Previous machine learning models predicted gene expression using data from the same biological samples.
• Predicting gene responses to transcription factor perturbation without data from perturbed cells presents a significant challenge.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting gene responses to transcription factor perturbation.
• To assess the utility of various genomic features, including transcription factor binding, histone marks, gene expression levels, and expression variation, for this prediction task.
• To identify key molecular features that influence a gene's propensity to respond to transcription factor perturbations.

Main Methods:

• Trained machine learning models to predict gene responses to transcription factor perturbation using pre-perturbation data.
• Evaluated the predictive power of transcription factor location data (ChIP-seq).
• Assessed the contribution of gene expression level, expression variation, histone marks (H3K4me1, H3K4me3), and sequence-based/epigenetic properties.

Main Results:

• Transcription factor location data (ChIP-seq) showed minimal utility in predicting gene responses.
• Pre-perturbation gene expression level and expression variation were highly predictive of responses to any transcription factor perturbation.
• Certain histone marks (H3K4me1, H3K4me3) showed some predictive power when located downstream of the transcription start site, but less than expression-based features.
• Gene expression level and variation effectively summarize sequence-based and epigenetic properties influencing response.

Conclusions:

• Gene expression level and variation are critical determinants of responsiveness to transcription factor perturbation.
• The difficulty in predicting responses from binding locations is partly due to these underlying molecular features being reflected in expression patterns.
• Some genes are inherently poised to respond, while others are resistant, highlighting a systems-level property rather than solely relying on direct binding information.