analiza aiqbench
This research introduces VANDAM, a framework that enhances genomic foundation models by integrating DNA molecular priors to better capture biochemical and structural properties beyond simple sequence tokenization. By explicitly training on these physical attributes, the method improves performance across diverse genomic tasks without requiring an independent modality. This advancement represents a significant step toward more biologically realistic AI models capable of leveraging established biophysical knowledge for downstream applications.
VANDAM: Viewing a nucleotide sequence with DNA molecular priors
arXiv:2610.00411v1 Announce Type: new 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.
Authors:Jeremy Levy, Ariel Larey, Yury Nahshan, Raizy Kellerman, Elay Dahan, Amit Bleiweiss, Guy Leib, Omri Nayshool, Dan Ofer, Tal Zinger, Dan Dominissini, Gideon Rechavi, Marissa Wirth, Simon Lee, Dung Hoang, Noam D. Beckmann, Shane O'Connell, Nicole Bussola, Alexander W. Charney, Yoli Shavit, Nati Daniel
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.
Machine Learning (cs.LG); Machine Learning (stat.ML)
arXiv:2610.00411 [cs.LG]
(or arXiv:2610.00411v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2610.00411 arXiv-issued DOI via DataCite (pending registration)
Text integral sub licența deschisă a publicației · 211 cuvinte
Subiecte conexe
AlphaGenome Atlas: A predictive map of every possible DNA letter change in the human genome
Google DeepMind · 8 sept. 2026
Deep and Probabilistic Models for Gene Regulatory Network Inference
arXiv · 20 iul. 2026
[AINews] Opus 5.5 is good at explainer videos
Latent Space · 29 sept. 2026
[AINews] The Future of Latent Space
Latent Space · 25 sept. 2026