Radical Numerics: AI-Driven Genomics and Biosecurity Strategy
Radical Numerics CEO Eric Gwynn discusses the Omni model, which outperforms specialized tools in human genomics. The company leverages long-context DNA language models for drug discovery, rare earth extraction, and biodefense, addressing the critical gap in AI-based biological safety.
The Rise of Generative Genomics
Radical Numerics is redefining the intersection of artificial intelligence and biology with the release of Omni, a large language model trained on DNA sequences. Unlike previous models that focused on short-context pattern recognition, Omni leverages long-context capabilities to process up to 2 million base pairs. This architectural advantage allows the model to capture complex, long-range regulatory interactions within the human genome, a domain where traditional bioinformatics tools have historically struggled. By moving beyond simple sequence matching, Omni provides a probabilistic framework for understanding how specific mutations influence disease, marking a significant shift from descriptive to predictive genomics.
Strategic Performance and Applications
The company’s latest benchmarking reveals that Omni outperforms specialized, supervised models in variant effect prediction, particularly in non-coding regions. This capability is critical because the majority of disease-causing variants reside in these regulatory areas, which constitute the vast majority of the human genome. Beyond healthcare, Radical Numerics is applying these models to industrial challenges, such as designing proteins for rare earth mineral extraction. By utilizing genomic context to guide protein design, the company addresses strategic supply chain issues, demonstrating the versatility of DNA language models in solving high-value commercial problems.
The Biosecurity Imperative
A core strategic pillar for Radical Numerics is the "dual mandate" of advancing both biological design and defense. The company argues that the defensive side of the AI-bio arms race is significantly lagging behind offensive capabilities. Current biodefense tools rely on static sequence matching, which is easily bypassed by AI-generated, functionally equivalent but genetically distinct pathogens. Omni addresses this by using functional embeddings to detect pathogenicity based on biological capability rather than exact sequence matches. This approach enables more robust surveillance, attribution, and countermeasure development, positioning the company as a key player in national security and public health infrastructure.
Conclusion
Radical Numerics is positioning itself at the forefront of "general biological intelligence," aiming to unify multiple biological modalities under a single AI framework. By combining frontier AI research with practical applications in health and security, the company is creating a new category of biotech enterprise. The focus on long-context processing and functional safety represents a sustainable competitive advantage, offering a scalable solution to the complex challenges of modern biology and global security.
Key insights
-
Long-context DNA models can capture regulatory interactions that short-context models miss, significantly improving disease variant prediction in non-coding regions.
Impact: This capability accelerates the diagnosis of genetic diseases and enables more precise therapeutic targeting, reducing the time and cost of drug discovery.
-
The same AI models capable of generating novel biological sequences are best suited for detecting pathogenic threats, creating a natural synergy between offensive and defensive capabilities.
Impact: This dual mandate allows companies to offer comprehensive safety solutions, differentiating them from pure-play design labs and addressing a critical gap in the market.
-
Traditional biodefense methods based on sequence matching are obsolete against AI-generated pathogens that maintain function while altering genetic spelling.
Impact: There is a urgent demand for function-aware detection systems, presenting a high-value opportunity for AI companies to partner with government and healthcare sectors.
-
Chain-of-thought reasoning can be applied to biological sequence design by training models on progressively optimized fitness scores, enabling in-silico evolution.
Impact: This technique reduces the need for extensive wet-lab experimentation, allowing for faster iteration and discovery of high-performing biological sequences.
-
Genomic context can be leveraged to design proteins for industrial applications, such as selective rare earth mineral extraction, addressing critical supply chain issues.
Impact: Expanding AI genomics beyond healthcare into industrial biotechnology opens new revenue streams and strategic partnerships in sectors like mining and materials science.
Action items
-
Integrate long-context DNA language models into R&D pipelines to improve the prediction of non-coding variant effects.
Impact: This will enhance the accuracy of genetic diagnostics and reduce the false positive rate in variant analysis, leading to more effective personalized medicine.
-
Develop function-aware biodefense tools that detect pathogenicity based on biological capability rather than exact sequence matches.
Impact: This addresses the vulnerability of current sequence-matching systems to AI-generated obfuscated pathogens, strengthening national and corporate biosecurity.
-
Adopt a dual-mandate strategy that pairs offensive biological design capabilities with defensive safety monitoring and attribution tools.
Impact: This holistic approach positions the company as a trusted partner for both innovation and safety, appealing to a broader range of stakeholders including regulators and investors.
-
Apply chain-of-thought training techniques to biological sequence design tasks to enable in-silico optimization of fitness scores.
Impact: This reduces the reliance on physical wet-lab experiments, accelerating the discovery of high-performing sequences for therapeutics and industrial applications.
-
Explore the application of genomic context in designing proteins for industrial challenges, such as rare earth mineral extraction.
Impact: This diversifies the company’s revenue streams and positions it at the intersection of AI, biotechnology, and strategic resource management.
Quotes
“A model that is good at generating, turns out is also very good at discriminating or predicting if a sequence is pathogenic or not.”
“The design side is going to get more capable. The defensive side needs to try its best to get ahead.”
“We felt it was important as a lab that a team that was both building the design capabilities is actually also best suited for building the defense capabilities because they're basically the same models.”