Iterative improvement of deep learning models using synthetic regulatory genomics

Area of Science:

• Genomics
• Computational Biology
• Epigenetics

Background:

• Generative deep learning models can reconstruct epigenetic data from genome sequences.
• Predictive accuracy on sequences diverging from the reference genome, like disease variants, remains unclear.

Purpose of the Study:

• To evaluate the predictive power of the Enformer model on engineered DNA sequences with deletions, inversions, and rearrangements.
• To fine-tune Enformer using experimental data from synthetic constructs to improve its generalizability.

Main Methods:

• Utilized the Enformer model to predict DNA accessibility on engineered sequences.
• Applied synthetic regulatory genomics to create and profile dozens of modified DNase I hypersensitive sites (DHSs).
• Fine-tuned Enformer with the generated data and assessed performance on other epigenetic tracks.

Main Results:

• Enformer showed good correlation between predicted and measured DNA accessibility.
• Model performance decreased for sequences with altered DHS order or orientation compared to simpler modifications.
• Fine-tuning Enformer significantly reduced prediction error and maintained performance on other tracks.

Conclusions:

• Current deep learning models exhibit limitations with novel sequences lacking features present in training data.
• Iterative fine-tuning with synthetic data enhances model generalizability.
• This approach can aid in the functional classification of regulatory variants from population studies.