VANDAM: Viewing a nucleotide sequence with DNA molecular priors

2026-09-30

AI Summary

Existing genomic foundation models rely exclusively on string-based paradigms, neglecting the biochemical and biophysical properties of DNA. To address this limitation, this work proposes VANDAM, a framework that innovatively incorporates molecular attributes estimated by biophysical models as explicit training priors into self-supervised learning, rather than treating them as independent modalities. Specifically, VANDAM constructs a joint optimization objective through masked token prediction, regional molecular property prediction, and local feature injection at the input level, guiding the model to learn physically meaningful DNA representations. Experimental results demonstrate that VANDAM achieves significant performance improvements across four architectures and nine downstream tasks. Furthermore, the introduced molecular priors exhibit strong generalization capabilities to unseen properties, highlighting the effectiveness of integrating biophysical knowledge into genomic representation learning.

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.

Problem

• Genomic Foundation Models
• DNA molecular priors
• biophysical properties
• self-supervised pretraining

Innovation

• Genomic Foundation Models
• DNA molecular priors
• Self-supervised learning
• Biophysical properties
• Representation learning