The human genome is about three billion base pairs long. Most AI models built to read it don't bother with the whole thing. They work on short, pre-selected snippets, like trying to understand a novel by only reading the chapter titles. Genomic Intelligence AI is betting that this is the fundamental mistake holding back the field. Its proposed fix is to treat the entire genome as a single, ultra-long prompt for a foundation model, aiming to predict disease risk, explain heritability, and even design genetic edits from the complete sequence [Genomic Intelligence, September 2026].
The bet on context
The company's technical premise is that biology's most important signals are encoded in patterns that span vast stretches of DNA, or in the complex interplay between an individual's genome and other 'omics' data like proteomics or metabolomics. By applying the same long-context transformer architectures that allow large language models to process entire books, the team believes it can capture these distant, non-linear relationships. The goal is a model that doesn't just flag a known risky variant, but can interpret the unique symphony of an individual's entire genetic code [Genomic Intelligence, September 2026]. The platform is designed to integrate these models into what it describes as agentic workflows, combining specialized AI with established bioinformatics tools to guide researchers from raw data to actionable insights [Veniamin Fishman - AIRI | LinkedIn, 2026].
A team built for scale
While the company's commercial structure and funding are not yet public, its disclosed team suggests a focus on the foundational compute and scientific validation required for such an ambitious project. The core group blends AI architecture with wet-lab expertise.
| Role Focus | Named Individual | Noted Background / Affiliation |
|---|---|---|
| Product Vision | Julia Kiseleva, PhD | Experience in mining click logs to train matching systems [Julia Kiseleva - Stealth |
| Genomic Model Science | Benjamin Fishman, PhD | AI consultant with a background from Technion and Weizmann [Ben Fishman - Bar-Ilan University |
| Long-Context Models | Mikhail Burtsev, PhD | Focus on ultra-long-context models and memory architectures [Genomic Intelligence, September 2026]. |
| AI Infrastructure | Natalia Vassilieva, PhD | VP and Field CTO for ML at Cerebras Systems [Genomic Intelligence, September 2026]. |
| AI-for-Health Strategy | Eric Horvitz, MD, PhD | Chief Scientific Officer at Microsoft [Genomic Intelligence, September 2026]. |
The inclusion of senior advisors from Cerebras, a company building some of the world's most powerful AI chips, and from Microsoft's office of the chief scientific officer, points to a recognition that this problem is as much about compute scale as it is about biological insight.
The incumbent to beat
The most direct counter-bet is that we don't need whole-genome models to get most of the value. A large part of the existing genomic analysis industry is built on polygenic risk scores (PRS), which aggregate the effects of thousands of individual variants identified through massive population studies. These scores are statistically powerful, relatively straightforward to compute, and are already being deployed in clinical and consumer settings. The efficiency argument is strong: why process three billion base pairs when a few hundred thousand tagged SNPs might give you 90% of the predictive power for a fraction of the computational cost?
Genomic Intelligence AI's rebuttal would be that the last 10% is where the real medicine happens,the ability to explain the 'missing heritability' for complex diseases, to design precise genetic interventions, or to make accurate predictions for individuals from ancestries underrepresented in existing biobanks. It's a high-risk, high-reward calculation. The computational cost of training and inferring with whole-genome models is not trivial. A back-of-the-envelope estimate: if processing one whole genome with a sophisticated model costs $10 in cloud compute (a speculative but plausible figure for a complex inference), then screening a million individuals becomes a $10 million line item just for inference, before any science is done. For this bet to pay off, the accuracy and novel insights gained must justify an order-of-magnitude increase in processing cost over incumbent methods.
To succeed, Genomic Intelligence AI must prove its models don't just add context, but that this context translates into clinical or research outcomes that are impossible with today's variant-centric tools. It must beat the entrenched, efficient logic of the polygenic risk score.
Sources
- [Genomic Intelligence, September 2026] Genomic Intelligence, Genome-scale AI for biology | https://genomicintelligence.ai/
- [Genomic Intelligence, September 2026] Blog | https://genomicintelligence.ai/blog/
- [Veniamin Fishman - AIRI | LinkedIn, 2026] LinkedIn profile | https://www.linkedin.com/in/veniamin-fishman/
- [Julia Kiseleva - Stealth | LinkedIn, 2026] LinkedIn profile | https://www.linkedin.com/in/julia-kiseleva-24842710/
- [Ben Fishman - Bar-Ilan University | LinkedIn, 2026] LinkedIn profile | https://www.linkedin.com/in/ben--fishman/