Nona

Nona is a multimodal genomics foundation model from Genentech's BRAID (Biology Research | AI Development) group, introduced in a November 2025 bioRxiv preprint. It addresses a persistent fragmentation in computational genomics: sequence-to-function prediction, DNA language modeling, and generative design of regulatory elements have historically been tackled by separate, specialized models, each with its own architecture, training regime, and assumptions. Nona instead pursues a single pretrained model that performs all three within one framework.

The model's central idea is a multimodal masked-modeling objective that jointly trains on raw DNA sequence and base-resolution functional genomics measurements. By treating both the underlying sequence and its measured functional readouts as modalities to be reconstructed from masked inputs, Nona learns representations that couple genomic sequence to regulatory activity at single-nucleotide resolution. The same pretrained backbone supports discriminative tasks (predicting function from sequence) and generative tasks (designing new sequences with desired regulatory properties), the latter cast as masked discrete diffusion over the sequence modality.

A defining claim of the work is generality without retraining: the authors demonstrate three distinct downstream applications drawn directly from the single pretrained architecture, rather than fine-tuning bespoke models per task. This positions Nona within the broader movement — alongside genomic foundation models such as Enformer-style sequence-to-function predictors and DNA language models like the Nucleotide Transformer and Evo lineage — toward unified, reusable backbones for regulatory genomics.

#Key Features

• Joint sequence and functional-genomics modeling: Nona is trained on both DNA sequence and base-resolution functional genomics signals simultaneously, learning a shared representation that links genomic sequence to measured regulatory activity rather than treating sequence in isolation.
• Unified discriminative and generative backbone: A single pretrained architecture supports sequence-to-function prediction, DNA language modeling, and regulatory element generation, reducing the need for task-specific model families.
• Masked discrete diffusion for design: Generative regulatory-element design is implemented as masked discrete diffusion over the sequence modality, enabling de novo design of regulatory DNA with targeted properties.
• Zero-retraining downstream transfer: The preprint reports three downstream applications derived from the pretrained model without per-task retraining, illustrating the framework's reusability.
• Base-resolution functional grounding: By incorporating base-resolution functional measurements during pretraining, the model aims to capture fine-grained regulatory structure that coarser, sequence-only objectives can miss.

#Technical Details

Nona is a multimodal masked-modeling framework that unifies three capabilities — sequence-to-function prediction, DNA language modeling, and generative regulatory element design — in one pretrained model. Pretraining jointly uses DNA sequence and base-resolution functional genomics data under a masked-reconstruction objective; generation is performed via masked discrete diffusion, a class of discrete generative methods that progressively unmask tokens and is well suited to sequence design. The work is described in the bioRxiv preprint posted on 6 November 2025 (version 2, 18 November 2025; DOI 10.1101/2025.11.06.687036). Detailed architectural hyperparameters such as parameter count, context length, and the specific functional-genomics assays used for training are reported in the preprint; precise figures are not restated here pending confirmation from the primary source.

#Applications

Nona targets researchers in regulatory and functional genomics who need both predictive and generative capabilities from a common model. Sequence-to-function prediction supports interpreting how genomic sequence — including variants — shapes regulatory activity, relevant to prioritizing noncoding variants and dissecting gene regulation. The DNA language modeling capability provides general-purpose sequence representations for downstream analysis, while the masked-diffusion generative mode enables de novo design of regulatory elements such as promoters or enhancers with intended properties, of interest to synthetic biology and therapeutic discovery teams. Because the three applications derive from one pretrained backbone without retraining, the model is positioned to streamline workflows that would otherwise require assembling several specialized tools.

#Impact

Nona contributes to the consolidation of regulatory genomics around unified foundation models by showing that prediction, language modeling, and generative design can share a single multimodal masked-modeling backbone grounded in base-resolution functional data. As a Genentech BRAID release, it reflects industry investment in reusable genomic foundation models for therapeutic discovery. As of June 2026, no public code or model weights were located — the previously expected genentech/nona GitHub and Hugging Face repositories return 404 — and the preprint is released under CC BY-NC 4.0, with the license for any future model weights unspecified. These openness gaps, together with the preprint (not yet peer-reviewed) status, are the primary caveats for prospective users; independent benchmarking and reproducibility await the release of artifacts or peer review.