Variant impact

DeepVRegulome

A DNABERT-based framework that predicts how short genomic variants change transcription factor binding, histone marks and splicing, explains each call with motif-level evidence, and links high-impact variants to patient outcomes.

Bioinformatics Advances, under revision NIH R01LM013722

Paper Code Models PyPI Web app

458transcription factor binding models

4histone mark models, plus splice-site models

9,837binding-site-altering mutations found in glioblastoma

1,352mutations linked to patient survival

The problem

Most disease-associated variants fall outside protein-coding genes, in the regulatory DNA that controls when and where genes switch on. Predicting whether a single base change disrupts a binding site, a histone mark or a splice site, and whether that matters for patients, is still hard.

How it works

Fine-tuned models. DNABERT models fine-tuned on ENCODE and GENCODE data for transcription factor binding, histone marks and splice sites.
Score. Each variant is scored by comparing reference and alternate alleles with log-odds ratios.
Explain. Attention-based motif analysis shows which sequence pattern a variant disrupts.
Connect to outcomes. Kaplan-Meier and Cox survival analyses link high-impact variants and regulatory regions to clinical outcomes.

Key results

Benchmarked against SNP-SELEX allele-specific binding data and compared with four established variant effect predictors.
In glioblastoma, found 572 splice-disrupting mutations and 9,837 transcription factor binding site-altering mutations, each present in more than 10% of samples.
Linked 1,352 mutations and 563 disrupted regulatory regions to patient survival, allowing patients to be stratified by noncoding mutation signatures.

Use it

pip install deepvregulome

Fine-tuned models are on Hugging Face, and results can be explored in the interactive web app.

Papers

Team

Pratik Dutta (lead), Matthew Obusan, Rekha Sathian, Max Chao, Pallavi Surana, Nimisha Papineni, Yanrong Ji, Zhihan Zhou, Han Liu, Alisa Yurovsky, Ramana V. Davuluri

Support

DeepVRegulome is supported by NIH R01LM013722.

Cite

@article{dutta2025deepvregulome,
  title   = {DeepVRegulome: DNABERT-based Deep-Learning Framework for Predicting
             the Functional Impact of Short Genomic Variants on the Human Regulome},
  author  = {Dutta, Pratik and Obusan, Matthew and Sathian, Rekha and Chao, Max and
             Surana, Pallavi and Papineni, Nimisha and Ji, Yanrong and Zhou, Zhihan and
             Liu, Han and Yurovsky, Alisa and Davuluri, Ramana V.},
  journal = {arXiv preprint arXiv:2511.09026},
  year    = {2025}
}