VANDAM: Viewing a nucleotide sequence with DNA molecular priors
Abstract
Contemporary Genomic Foundation Models (GFMs) rely on a DNA-as-a-string paradigm that employs masked token prediction objectives for pretraining. However, this abstraction does not explicitly model the biochemical, structural, and physical properties essential to biological function. Many molecular properties can be estimated from sequence using established biophysical models, so their utility lies not in providing an independent modality, but in introducing priors that training objectives can explicitly exploit. We introduce VANDAM, a framework that extends the training of GFMs with DNA molecular priors. In self-supervised training, VANDAM predicts regional molecular properties from pooled representations. When functional labels are available and can reward retaining molecular priors, local features are additionally injected at the input. VANDAM consistently improves downstream performance across four architecture families and nine held-out genomic tasks by complementing token-based objectives. Probing experiments further demonstrate that the use of molecular priors generalizes to other unseen molecular properties.
Related
- Geodesics with Unified Tangent-constrained Priors and Curvature Regularization
- Inconsistency between Frequentist and Bayesian inference, approximation of improper priors and permissible priors
- Exact signal recovery from generative priors with sigmoid output activation
- Sieve-free ReLU Bayesian priors with near-minimax posterior contraction
- From priors to performance: enhancing statistical efficiency with Bayesian dynamic borrowing
BibTeX
@misc{levy2026vandam,
title = {{VANDAM: Viewing a nucleotide sequence with DNA molecular priors}},
author = {Jeremy Levy and Ariel Larey and Yury Nahshan and Raizy Kellerman and Elay Dahan and Amit Bleiweiss and Guy Leib and Omri Nayshool and Dan Ofer and Tal Zinger and Dan Dominissini and Gideon Rechavi and Marissa Wirth and Simon Lee and Dung Hoang and Noam D. Beckmann and Shane O'Connell and Nicole Bussola and Alexander W. Charney and Yoli Shavit and Nati Daniel},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.00411},
eprint = {2610.00411},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.00411},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - VANDAM: Viewing a nucleotide sequence with DNA molecular priors
AU - Jeremy Levy
AU - Ariel Larey
AU - Yury Nahshan
AU - Raizy Kellerman
AU - Elay Dahan
AU - Amit Bleiweiss
AU - Guy Leib
AU - Omri Nayshool
AU - Dan Ofer
AU - Tal Zinger
AU - Dan Dominissini
AU - Gideon Rechavi
AU - Marissa Wirth
AU - Simon Lee
AU - Dung Hoang
AU - Noam D. Beckmann
AU - Shane O'Connell
AU - Nicole Bussola
AU - Alexander W. Charney
AU - Yoli Shavit
AU - Nati Daniel
PY - 2026
DA - 2026/10/01
PB - arXiv
DO - 10.48550/arxiv.2610.00411
UR - https://arxiv.org/abs/2610.00411
AB - Contemporary Genomic Foundation Models (GFMs) rely on a DNA-as-a-string paradigm that employs masked token prediction objectives for pretraining. However, this abstraction does not explicitly model the biochemical, structural, and physical properties essential to biological function. Many molecular properties can be estimated from sequence using established biophysical models, so their utility lies not in providing an independent modality, but in introducing priors that training objectives can explicitly exploit. We introduce VANDAM, a framework that extends the training of GFMs with DNA molecular priors. In self-supervised training, VANDAM predicts regional molecular properties from pooled representations. When functional labels are available and can reward retaining molecular priors, local features are additionally injected at the input. VANDAM consistently improves downstream performance across four architecture families and nine held-out genomic tasks by complementing token-based objectives. Probing experiments further demonstrate that the use of molecular priors generalizes to other unseen molecular properties.
ER -