Origin's AI designs custom DNA "switches" that tell a therapy gene exactly when, where, and how loudly to turn on—like writing a personalized instruction manual for every cell type in the body.

Origin's Axis platform uses a transformer-based generative model trained on the ENCODE V4 Registry of cis-regulatory elements to both generate and predict the function of synthetic regulatory DNA sequences. Engineers prompt the model with specific cell types and transcription factor contexts, and Axis outputs novel enhancer and promoter sequences optimized for precise gene expression in the target tissue. The model achieves 6.7% higher accuracy than Google DeepMind's AlphaGenome in regulatory element activity prediction and produces sequences where 72% are entirely unique (no 20bp match to known genomes via BLAT). Generated sequences show up to 9x higher motif enrichment for high-affinity transcription factor binding sites compared to baseline, and in vitro validation confirms highest transcriptional activity in the prompted cell type. This capability fundamentally changes gene therapy vector design from trial-and-error cloning to AI-guided precision engineering, dramatically compressing R&D cycles and improving therapeutic safety profiles.

It's like having an AI architect who can design a house that only unlocks its doors for the right family member in the right room at the right time—except the house is a gene therapy and the rooms are different cell types.

Origin runs a self-reinforcing loop where its AI designs DNA sequences, lab experiments test them, and the results make the AI smarter—like a student who writes their own exam questions and learns from every answer.

Origin has architected a closed-loop data flywheel that is central to its long-term competitive strategy. The Axis model generates millions of candidate regulatory DNA sequences with predicted cell-type-specific activity profiles. A curated subset of these candidates is synthesized and tested in high-throughput in vitro transcriptional assays, producing ground-truth functional data on synthetic sequences that do not exist in any public database. This proprietary experimental data is then fed back into the model as additional training signal, improving prediction accuracy and generation quality with each cycle. Because the sequences are AI-designed and experimentally novel, competitors relying solely on public genomic databases (like ENCODE) cannot replicate this dataset. The flywheel accelerates over time: better models generate more informative candidates, which yield higher-quality training data, which produce even better models. This strategy mirrors successful data-moat approaches in other AI domains (e.g., Tesla's autonomous driving data loop) and positions Origin to compound its technical advantage with every experimental cycle.

It's like a chef who invents new recipes, taste-tests them all, and uses the results to invent even better recipes—except no other chef has access to the tasting notes.

Origin's AI predicts whether a gene therapy might accidentally turn on in the wrong cells and redesigns it to avoid that—like a spell-checker for DNA that catches dangerous typos before they reach a patient.

A critical bottleneck in cell and gene therapy development is ensuring that therapeutic transgenes are expressed only in intended target cells and not in off-target tissues, where uncontrolled expression can cause toxicity, immune reactions, or oncogenesis. Origin's Axis platform addresses this by modeling regulatory element activity across hundreds of cell types simultaneously, enabling engineers to identify and eliminate sequences with predicted off-target activity profiles before any in vivo testing. The model's cell-type-specific transcription factor prompt system allows designers to specify not only where a regulatory element should be active but also where it must be silent. Axis-generated sequences are validated against independent predictive models (e.g., Broad Institute's Malinois) and cross-referenced with ENCODE chromatin accessibility data to confirm tissue specificity. This computational pre-screening dramatically reduces the number of candidates that need expensive and time-consuming animal safety studies, compresses the preclinical development timeline, and provides pharma partners with a quantitative safety profile for each regulatory element—a critical differentiator in regulatory submissions to the FDA. The approach transforms safety from a late-stage gate into an early-stage design parameter.

It's like having a GPS that not only finds the fastest route to your destination but also automatically avoids every neighborhood where your car might break down—before you even start driving.