A self-supervised DNA foundation model with collapse-resistant multimodal fusion

In plain English

The model first learns from DNA sequence, then adds two views of which genomic regions are open in cells. A normalization step is designed to keep the accessibility signal from collapsing into an uninformative shortcut.

What they found

• The preprint reports a 4.6-fold AUPRC improvement over a DNA-only baseline for peak detection.
• The authors report transfer across variant and regulatory-association datasets.

The catch

• The work is a preprint and has not been peer reviewed.
• The sole author discloses being founder and CEO of ModelsLive.

Methods at a glance

Study type

Computational-method preprint

Biological system

Human genomic regulation

Organism

Human

Dataset

DNA sequence, chromatin-accessibility data, ClinVar, GTEx eQTL, and PBMC caQTL benchmarks

Sample size

Not reported

Model architecture

Self-supervised DNA foundation model with collapse-resistant multimodal fusion

Training objective

Learn regulatory genomic representations while preserving complementary modalities

Code

No public code repository is identified in the abstract.

Data

The abstract names public external validation datasets.

Tasks and metrics

Tasks: Peak detection; Variant and regulatory-association transfer

Metrics: AUPRC

Baselines: DNA-only foundation-model baseline

Principal findings

• The authors report a 4.6-fold AUPRC improvement over the DNA-only baseline for peak detection.

What remains unresolved

External validation: ClinVar, GTEx eQTL, and PBMC caQTL datasets were used for external transfer evaluation.

• Preprint evidence only.
• The abstract does not establish prospective experimental utility.
• The author reports a company conflict of interest.