Origin Bio’s Axis AI Model Outperforms AlphaGenome by 6.7% in Regulatory DNA Prediction
Image Credit: Sangharsh Lohakare | Splash
Biotechnology startup Origin Bio has introduced Axis, an artificial intelligence model aimed at predicting and generating regulatory DNA sequences, advancing the use of AI in tackling gene therapy obstacles.
The model, unveiled on Oct 8, surpasses Google DeepMind's AlphaGenome by an average of 6.7% in forecasting protein binding to regulatory DNA elements, based on benchmarks outlined on Origin Bio's website.
Details of the Axis Model
Axis serves as a versatile tool handling three primary tasks: generating DNA sequences from prompts, predicting functions from DNA, and creating sequences according to specified functions. It employs a shared Transformer architecture, a neural network type prevalent in AI, with distinct heads for nucleotide prediction and assay results.
Incorporating prompts for cell types and transcription factors, the model guides outputs effectively. For generation, it uses an inverse entropy sampling method, focusing on confident positions to construct sequences step by step.
Origin Bio trained Axis using data from the ENCODE V4 Registry, emphasising cis regulatory elements sorted by activity. The firm divided training and test data to prevent overlap, aligning with AlphaGenome's development approach.
Tests revealed that sequences produced with targeted transcription factor prompts exhibited up to nine times more motif occurrences than those without prompts. This feature supports designing regulatory DNA with exact binding traits, such as enhancers or promoters.
Origin demos a web interface on the announcement page and is offering free API access by request through a form, aimed at researchers for gathering feedback.
Background on AI in Genomics
AI's growth in biology arises from the challenge of interpreting the genome's extensive non coding areas, comprising 98% of human DNA and regulating gene expression. Conventional techniques falter with intricate long range interactions and variant impacts.
Google DeepMind's AlphaGenome, released on June 25, established a standard by predicting multimodal genomic attributes like RNA expression, splicing, and chromatin accessibility from sequences up to one million base pairs. It excelled over external models on 22 of 24 single sequence assessments and matched or surpassed them on 24 of 26 variant effect evaluations.
Axis extends this by concentrating on regulatory element binding, vital for grasping how DNA activates or deactivates genes. Origin Bio crafted the model to overcome constraints in single task AI systems, leveraging multitask learning studies to merge generative and predictive functions.
The San Francisco based startup seeks to refine gene and cell therapies amid rising needs. Therapies such as Luxturna for retinal conditions and Zolgensma for spinal muscular atrophy demonstrate potential but encounter problems with effectiveness, immune reactions, and unintended activity.
Reasons Behind the Development
Origin Bio developed Axis to address these therapy barriers via AI directed design. Regulatory DNA dictates gene activation timing and location, and suboptimal design may cause toxicity or subdued outcomes.
Through prompt based generation, Axis permits researchers to designate transcription factors for precise binding, possibly curbing off tissue expression in treatments like CAR T cell therapy or AAV gene delivery.
The multitask method draws from research showing joint training boosts performance, as observed in prior AI publications. This integration captures more comprehensive DNA behaviour representations, filling voids in tools that manage tasks independently.
Potential Impacts
Axis may bolster biotech studies by enabling more secure therapies. For example, it aids in crafting tissue specific promoters to restrict gene expression, reducing adverse effects.
More broadly, the model supports synthetic biology, where custom DNA circuits resembling switches or dials could facilitate sophisticated gene control. Initial tests indicate Axis can integrate multiple motifs for combined interactions.
Yet, laboratory confirmation is forthcoming, with Origin Bio intending additional testing to verify predictions in actual cells.
Future Trends in AI Driven Biology
AI's involvement in genomics is broadening to generative tools that devise new sequences, as shown in recent progress. Stanford's Evo model, launched in December 2024, generates genomic code by learning from DNA sequences, aiding in understanding bacterial and viral genomes.
Yale's collaborations, such as the October 2025 AI model with Google for single cell RNA analysis and the August 2025 RNA editing tools, highlight safer genetic modifications and enhanced tissue insights.
Trends suggest merging modalities past DNA, including RNA and proteins, for cohesive biological models. Journals like Nature emphasise AI's acceleration of medical and agricultural processes, forecasting variant impacts and anti CRISPR proteins for editing control.
Obstacles persist in encompassing ultra long range effects and cell specific patterns, where models like AlphaGenome and Axis display limits. With increasing computational resources, specialists foresee AI unraveling non coding DNA's structure, enhancing disease studies and eco friendly bioengineering.
Origin Bio's efforts fit this evolution, enabling startups to rival major tech firms in making genomic tools accessible.
TheDayAfterAI News
We are a leading AI-focused digital news platform, combining AI-generated reporting with human editorial oversight. By aggregating and synthesizing the latest developments in AI — spanning innovation, technology, ethics, policy and business — we deliver timely, accurate and thought-provoking content.