Loading…
Regulatory DNA model that generates enhancer and promoter sequences and predicts their chromatin accessibility and transcription-factor binding.
No providers recorded yet. Browse all providers
A gene therapy fails as often on its promoter as on its payload. Drive a chimeric antigen receptor from a strong constitutive backbone and receptor density on the T cell surface rises high enough to produce tonic signaling and exhaustion; swap in a milder EFS- or MND-class promoter and the same construct persists longer while still killing tumor cells. The tradeoff recurs in AAV, where an overly strong or poorly targeted promoter raises antigen burden and immunogenicity while too weak a one loses potency. Yet a construct designer still picks from a short list of natural promoters and enhancers, largely empirically.
Axis, announced by Origin on 8 October 2025, treats that list as something to be designed rather than picked from. The company frames the gap in existing DNA models as one of direction: a given model natively performs DNA-to-DNA, DNA-to-function, or function-to-DNA, but not all three. Axis unifies them in a single network, so the same weights that score an existing cis-regulatory element for chromatin accessibility and transcription-factor binding can also emit a new element conditioned on a target cell type and a desired binding profile. Origin's argument for that design is representational: exposing one model to generation and prediction over the same domain, it reports, made the prediction side better.
The predictive baseline Axis measures itself against is AlphaGenome. Origin reports outperforming it on regulatory element binding activity by 6.7% on average, and notes that its processing of the underlying experimental files is identical to the AlphaGenome paper's, which is what makes the two comparable. These are company-reported numbers from a blog announcement; no preprint or peer-reviewed paper describing Axis has appeared, and no independent replication exists.
The architecture is a Transformer backbone shared by both tasks, with a nucleotide head and an assay head that diverge after the final Transformer layer, the assay branch carrying additional intermediate processing. Training minimizes a linear combination of a cross-entropy term for nucleotide prediction and a factorized Poisson plus multinomial term capturing both the profile and the intensity of signal over the regulatory region. A later Origin preprint on optimizer choice for regulatory DNA transformers — which studies the same architecture family but never names Axis — reports that its models used muP initialization and layer pre-normalization, that this was not enough to enforce scale invariance, and that a layer post-norm and QK-norm were added to enforce it. Parameter count, context length and training compute are undisclosed.
Training data comes entirely from ENCODE: intervals of independent cis-regulatory elements from the ENCODE V4 Registry as published on SCREEN, ranked by mean activity with a z-score derived for every biosample-and-assay combination. Elements are partitioned into train and test scaffolds with genomic-coordinate overlap verified to be zero, so benchmark intervals are disjoint from pretraining data. Reported benchmarks cover chromatin accessibility and transcription-factor binding, broken out by factor role.
The immediate target is promoter and enhancer engineering for cell and gene therapy, where a designed element can concentrate expression in a target tissue or tune receptor density into a therapeutic window. A second use is perturbation biology: libraries spanning a graded range of transcriptional strength let a gene's expression be titrated continuously rather than knocked out, turning a binary perturbation into a dose-response experiment. Origin has released one such library, Switch — 10,000 Axis-designed proximal enhancer-like sequences for SK-N-SH, HepG2 and K562, each annotated with predicted MPRA activity from the Broad Institute's Malinois model, binding-site calls, sequence quality metrics, and a duplex structure rendered with Protenix.
Axis is a company announcement rather than a publication, and its claims rest on Origin's own evaluations. The model itself is not downloadable: there is no code repository, released checkpoint, model card or datasheet, and access is by request through a form on the announcement page, which Origin describes as a free API for research institutions and individuals. The Switch library is the one exception, exportable as a CSV of its 10,000 sequences and their annotations. Wet-lab validation of generated sequences is stated as forthcoming, and the Switch release supplies computational annotations rather than measured activity. Training covers human ENCODE cis-regulatory elements only, and the extension to RNA, protein and cellular representations is future work. What is concrete is the design point: one promptable model that both writes regulatory DNA and scores it, with the generative objective reported to improve the predictive one.