Modeling gene regulatory perturbations via deep learning from high-throughput reporter assays

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Interpreting noncoding genetic variants is challenging for diagnostics.
• High-throughput reporter assays (e.g., STARR-seq) measure regulatory effects but only for variants present in the assay.
• Whole-genome reporter assays generate extensive data suitable for training predictive models.

Purpose of the Study:

• To develop and evaluate BlueSTARR, a retrainable predictive modeling framework for interpreting noncoding variants using whole-genome reporter assay data.
• To identify genomic signatures of selection acting on regulatory variants.
• To assess the model's ability to learn complex regulatory patterns, including drug-dependent transcription factor binding.

Main Methods:

• Training predictive models using the BlueSTARR framework on whole-genome STARR-seq data from multiple cell lines and a drug treatment.
• Analyzing model outputs to detect global signatures of selection across the human genome.
• Testing model performance on synthetic enhancers with known transcription factor binding motifs, using drug perturbation data.

Main Results:

• A global signature of purifying selection against both loss-of-function and gain-of-function regulatory variants was identified.
• A bias consistent with selection against gains of cis-regulatory function in closed chromatin near genes was observed.
• The model, trained on drug perturbation data, successfully learned distance-dependent and treatment-dependent transcription factor binding patterns and their impact on reporter gene activation.
• Modest performance differences were found between various deep-learning architectures on this data modality.

Conclusions:

• Lightweight, retrainable models like BlueSTARR are valuable for uncovering latent signals in novel experimental data.
• The framework demonstrates utility in understanding regulatory variant effects and selection pressures.
• Further improvements in predictive accuracy are possible with larger models and more extensive data, even for state-of-the-art approaches.