#11 · Epigenomics
Deep learning models predict regulatory states from DNA sequence and chromatin context
The dominant thread is deep learning applied to DNA sequence to predict epigenomic and regulatory states, spanning DNA methylation (DMRU, iDNA-DAPHA), transcription factor binding (QTFPred, TRAFICA), and enhancer-enhancer interactions (ETNet). Several models use generative or pretrained architectures: DMRU is an explainable encoder-decoder generative model, iDNA-DAPHA uses domain-adaptive pretraining with hierarchical attention, and TRAFICA is an open chromatin language model pretrained on ATAC-seq and fine-tuned with LoRA. The newer entries add context beyond the sequence alone: CHROME incorporates polymer-ensemble-filtered Hi-C structure through graph attention, VariantFormer conditions gene expression prediction on personalized diploid genomes, and Clair-Mosaic uses phased haplotype context for mosaic variant calling. Enformer fine-tuning on synthetic DHS rearrangements reduces accessibility error for sequences diverging from reference. Clair-Mosaic is a variant caller rather than a regulatory predictor, so the shared thread is strongest across the regulatory-prediction items.
Published papers
DMRU: generative deep learning to unravel condition-specific cytosine methylation in plants.
DMRU, an explainable deep encoder-decoder generative model, predicts condition-specific cytosine methylation in plants from transcriptome status, with reported…
QTFPred: robust high-performance quantum machine learning modeling that predicts main and cooperative transcription factor bindings with base resolution.
QTFPred, a quantum-classical hybrid neural network, predicts transcription factor binding at base resolution, reporting state-of-the-art accuracy on 49 ENCODE…
ETNet: an interpretable transformer framework for enhancer-enhancer interaction prediction with cross-context transferability.
ETNet, a CNN-Transformer model that predicts enhancer-enhancer interactions from DNA sequence, shows cross-cell-line and enhancer-promoter transferability with…
iDNA-DAPHA: a generic framework for methylation prediction via domain-adaptive pretraining and hierarchical attention.
iDNA-DAPHA is a two-stage deep learning framework for DNA methylation prediction, using domain-adaptive pretraining across species and hierarchical attention,…
TRAFICA: an open chromatin language model to improve transcription factor binding affinity prediction.
TRAFICA, an open chromatin language model pretrained on ATAC-seq sequences and fine-tuned with LoRA on PBM and HT-SELEX data, improves TF-DNA binding affinity…
Iterative improvement of deep learning models using synthetic regulatory genomics.
Fine-tunes Enformer on synthetic DHS rearrangement data, reducing accessibility prediction error for sequences diverging from reference.
Preprints
Clair-Mosaic: A deep-learning method for long-read mosaic small variant calling
Clair-Mosaic is a deep-learning caller for low-VAF mosaic small variants from ONT and PacBio long reads, using haplotype phasing and a Bayesian mosaic-germline…
VariantFormer: A hierarchical transformer integrating DNA sequences with genetic variations and regulatory landscapes for personalized gene expression prediction
VariantFormer, a 1.2B-parameter transformer, predicts gene-level RNA abundance from personalized diploid genomes, improving eQTL effect estimates and in silico…
A Chromatin-Structure-Guided Framework for Predictive and Interpretable Regulatory Genomics
CHROME: a polymer-ensemble-filtered Hi-C graph attention framework predicting cell line-specific ChIP-seq, eQTL and ClinVar variant effects with interpretable…
Topic history
| Date | Rank |
|---|---|
| #11 | |
| #9 | |
| #10 | |
| #6 | |
| #13 |