Transfer learning identifies sequence determinants of cell-type specific regulatory element accessibility.

Area of Science:

• Genomics
• Computational Biology
• Epigenetics

Background:

• Genetic variants in regulatory elements contribute to disease pathogenesis.
• Understanding DNA's regulatory code is crucial for deciphering disease etiology.
• Deep learning models for DNA sequence analysis require substantial training data.

Purpose of the Study:

• To develop a transfer learning method for modeling DNA regulatory activity.
• To enable accurate prediction of cell-type specific chromatin accessibility from DNA sequence.
• To overcome data limitations in deep learning for genomic sequence analysis.

Main Methods:

• Developed ChromTransfer, a transfer learning approach using a pre-trained, cell-type agnostic model.
• Fine-tuned the model on specific regulatory sequences for cell-type specific predictions.
• Evaluated performance against models without pre-training.

Main Results:

• ChromTransfer achieved superior performance in learning cell-type specific chromatin accessibility.
• The method demonstrated effectiveness even with small input datasets, maintaining high accuracy.
• Identified sequence features recognized by key transcription factors as important for predictions.

Conclusions:

• ChromTransfer is a powerful tool for deciphering the DNA regulatory code.
• The method facilitates accurate modeling of gene regulation with limited data.
• Enables deeper understanding of disease mechanisms driven by regulatory element dysfunction.