Exploiting deep transfer learning for the prediction of functional non-coding variants using genomic sequence.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Genome-wide association studies (GWAS) identify numerous variants, but many non-coding variants lack functional validation.
• High-throughput assays validate functional non-coding variants (NCVs), yet their scarcity due to cost and technical challenges limits predictive modeling.
• Small sample sizes of validated NCVs hinder the development of reliable supervised machine learning models for genome-wide prediction.

Purpose of the Study:

• To develop a deep transfer learning model for improved prediction of functional non-coding variants (NCVs).
• To address the challenge of limited validated functional NCVs by leveraging large datasets for feature learning.
• To enhance the accuracy of predicting causal non-coding variants across the genome.

Main Methods:

• Utilized a deep transfer learning model based on convolutional neural networks.
• Employed large-scale generic functional NCVs for low-level feature learning and context-specific NCVs for high-level feature learning.
• Leveraged transfer learning to mitigate issues associated with small sample sizes of validated functional variants.

Main Results:

• The deep transfer learning model demonstrated superior performance compared to standard deep learning models without pretraining or retraining.
• Evaluated on three MPRA datasets and 16 GWAS datasets, the model significantly outperformed 18 existing computational methods.
• Achieved enhanced prediction accuracy for functional non-coding variants in both MPRA and GWAS contexts.

Conclusions:

• Deep transfer learning offers a robust solution for predicting functional non-coding variants despite data limitations.
• The proposed model provides a significant advancement in identifying causal non-coding variants, crucial for understanding complex traits.
• The developed model represents a powerful tool for genomic research, enhancing the interpretation of non-coding regions.