Cross-species regulatory sequence activity prediction
Problem
Regulatory sequence activity models were trained mostly on human data, leaving model organism genomes underused for gene regulation analysis.
Why it matters
Better regulatory prediction aids analysis of human genetic variants associated with molecular phenotypes and disease.
What the AI did
Deep convolutional neural networks were trained simultaneously on human and mouse genomes, plus a transfer learning approach applying mouse models to human variant analysis.
Results
Training on both genomes improved prediction accuracy on held-out sequences; numeric values not stated in source
Impact
Enables leveraging thousands of non-human epigenomic and transcriptional datasets for human analysis.
Record
| Venue | PLoS Computational Biology |
|---|---|
| Published | 2020 |
| Work type | article |
| Authors | David R. Kelley |
| Citations | 228 (OpenAlex, at import time) |
| References | 71 |
| Model or method | Deep convolutional neural networks with multi-genome training and transfer learning |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1371/journal.pcbi.1008050 |
| Open access | gold |
| Keywords | CNN, regulatory sequences, transfer learning, genomics, cross-species |
Importance rating
significant — thousands to millions affected
Gene regulation affects human disease per abstract; leveraging model organism data aids variant analysis, a broad genomics advance, but no scale stated.