Stanford researchers did something new. With help from Taiwanese partners, Google Research, and Google DeepMind they built an AI pipeline for genetic discovery. It actually works.
“We demonstrated that large language models (LLms) could facilitate genetic discovery in mice… and generate genetic diagnoses… in humans,” says Gary Peltze
Gary Peltz is a professor of Anesthesia at Stanford. He led the study. Why does it matter? Around 350 million people live with rare genetic diseases. If we can pin down the responsible genes treatment gets faster. Much faster.
The problem with DNA noise
Think about the genome. It is a set of DNA instructions for making RNA and proteins. Four base pairs. Order matters. But the order shifts. Every person carries thousands of DNA sequence changes called ‘variants of unknown significance.’
Do these variants cause sickness? We do not know.
Hearing loss illustrates the blur. One third of adults over 61 deal with it. That number hits 80 percent past 85. Half those cases stem from genetics. But we rarely know which parts. Peltz points out the clock is ticking. Restorative therapies are coming soon for other genetic conditions. Identifying hearing loss genes suddenly feels urgent.
Why old methods fall short
Scientists usually run a genome-wide association study, or GWAS. They look for statistical links between a genotype and a trait.
The trick? GWAS finds real culprits. It also flags false positives. Lots of them. You need to separate the wheat from the chaff to truly understand disease. Current ways do that manually. They burn cash. They waste time. You need expensive clinical geneticist expertise to sift through the data.
AI changes that equation. Fast. Cheaper. Accurate? Peltz thinks so. He argues genomic analysis by AI could improve health care for billions.
Med-PaLM meets Gemini
The team built pipelines to sort candidate genes. They used two distinct LLMs. Med-PaLM 2 specializes in medical knowledge. Gemini 2.5 Pro handles complex reasoning but lacks medical fine-tuning.
First up Med-PaLM 2 analyzed mouse GWAS genes. It spotted known causal genes correctly. Then it found something new. A genetic factor for spontaneous hearing loss in mice. Lab experiments later validated that finding.
Then came the human trials. Using Gemini 2.5 Pro the team tackled two groups. Twenty patients with hearing loss. Six with rare genetic syndromes.
“The pipeline… required [it] to reason its ranking with cited evidence.” — Tao Tu, Google DeepMind
Tao Tu is a research scientist at DeepMind and a co-author. The model reviewed patient genes against medical literature. It cross-referenced symptoms. Then it ranked likely causes. Its output competed head-to-head with diagnoses from actual doctors. An otolaryngologist. A clinical geneticist.
Gemini won without any prior medical training. It nailed the hearing loss cases. The rare disease cases were tougher. Complex symptoms everywhere. The team tweaked the pipeline to account for that messiness. Gemini still found the causative variants. It helped diagnose the rare diseases.
The study landed in Advanced Science. The authors argue AI can efficiently pinpoint the genetic roots of suspected diseases in hundreds of millions of people worldwide.
Agents and the future of genomes
What happens next? The researchers plan to integrate these pipelines into autonomous AI agent frameworks. Think plugins for protein or mutation data. It gets more capable.
Peltz has big dreams. A model reading electronic health records. Analyzing the genome. Fully automated. He sees a dramatic shift. These new reasoning models produce hypotheses. They spark discovery. They accelerate treatment searches.
Does that mean we ditch doctors?
No way. Humans must interpret the AI. Geneticists and clinicians stay central.
There is another gap too. Our analysis ignores most of the genome. Only two percent of DNA codes for mRNA or proteins. We analyze that part. We miss the other ninety-eight percent. Genetic changes happen there. We just don’t understand them well yet.
Peltz points to AlphaGenome as a potential fix. This deep learning model might interpret variations across that ignored 98%. It could unlock hidden information.
If AI keeps advancing genetics changes from treatment to prevention. Custom plans based on your code. It sounds radical.
Peltz sees a shift coming. From fixing disease to preventing it entirely.
