DeepVRegulome: DNABERT Framework Predicts Non-Coding Variant Impacts
A team of researchers has introduced DeepVRegulome, a deep-learning system that utilizes 464 specialized DNABERT models to assess the functional implications of short genomic variants on the human regulome. These models, which were trained using ENCODE and GENCODE datasets, consist of 458 transcription factor models, 4 histone mark models, and 2 splice site models. The framework incorporates analytical methods, including log-odds ratios for quantitative variant scoring, attention-based motif analysis, and survival analysis through Kaplan-Meier and Cox proportional hazards models to connect high-impact variants with clinical results. This innovation tackles the complexities of interpreting non-coding variants identified through whole-genome sequencing.
Key facts
- DeepVRegulome integrates 464 fine-tuned DNABERT models.
- Models include 458 transcription factor, 4 histone mark, and 2 splice site models.
- Trained on ENCODE and GENCODE datasets.
- Uses log-odds ratios for quantitative variant scoring.
- Employs attention-based motif analysis for disrupted sequence patterns.
- Survival analysis via Kaplan-Meier and Cox models links variants to clinical outcomes.
- Focuses on non-coding short variants from whole-genome sequencing.
- Aims to predict functional impact on the human regulome.
Entities
Institutions
- ENCODE
- GENCODE