Iterative improvement of deep learning models using synthetic regulatory genomics

Area of Science:

• Genomics
• Computational Biology
• Epigenetics

Background:

• Deep learning models can predict genome-wide epigenetic patterns from reference sequences.
• The predictive power of these models on non-reference sequences, like disease variants, remains unclear.

Purpose of the Study:

• To evaluate the performance of the Enformer model on engineered DNA sequences with varying degrees of divergence from the reference genome.
• To improve the generalizability of deep learning models for epigenetic prediction using synthetic regulatory genomics data.

Main Methods:

• Utilized the Enformer model to predict DNA accessibility and RNA transcription across engineered sequences.
• Employed synthetic regulatory genomics to create and test dozens of deletions, inversions, and rearrangements of DNase I hypersensitive sites (DHSs).
• Fine-tuned the Enformer model using experimental data from engineered sequences.

Main Results:

• Enformer showed good correlation between predicted and experimental DNA accessibility, with performance decreasing for sequences with altered DHS order or orientation.
• Model performance was best for sequences closely resembling the reference genome (e.g., single deletions).
• Fine-tuning significantly reduced prediction error and maintained strong predictive performance for other epigenetic tracks.

Conclusions:

• Current deep learning models exhibit limitations in predicting epigenetic patterns for novel sequences with critical feature divergences.
• An iterative approach, incorporating profiling of synthetic constructs, enhances model generalizability.
• This improved generalizability is crucial for the functional classification of regulatory variants identified in population studies.