In collaboration with Taiwanese colleagues and Google Research and Google DeepMind, researchers at Stanford University School of Medicine have, for the first time, developed an AI pipeline that facilitates genetic discoveries.

With an estimated 350 million people worldwide affected by rare genetic diseases, determining the genes responsible could expedite treatment development.

“We demonstrated that large language models (LLMs) could facilitate genetic discovery in mice and could generate genetic diagnoses for hearing loss and rare genetic diseases in humans,” says Gary Peltz, Professor of Anaesthesiology, Perioperative and Pain Medicine at Stanford University school of medicine and author of the study.

Genetic variants and disease

The genome is a full set of DNA instructions in a cell. These instructions code for RNA and the proteins needed for cellular function. DNA consists of four base pairs, and the order of these determines the instructions. However, this order can vary, and any individual may have thousands of variants in their DNA sequences, known as ‘variants of unknown significance’. We do not know if these variants cause disease or not.

Hearing loss affects a third of adults over the age of 61 and 80% of those over 85 years old. Half of those affected have a genetic cause, but not all genetic factors are identified and the significance of genetic variants in those with hearing loss is unknown.

“Identifying the genetic basis for hearing loss has assumed greater urgency since restorative therapies for other genetic causes could soon be available,” notes Peltz.

The current method for discovering genetic factors causing disease is a genome-wide association study (GWAS). This method looks for statistical association between the genetic factor (genotype) and the presence of a trait or condition (phenotype). The problem is that, while GWAS can correctly identify disease-causing variants, it also identifies variants based on false positive associations.

We need ways to separate the true disease-causing variants from the false positives to better understand genetic diseases.

Furthermore, Peltz explains that currently sorting through vast amounts of data to determine disease-causing genetic variants is expensive, time consuming and requires the expertise of clinical geneticists.

“AI-based methods for genomic analysis could rapidly accelerate the analysis and reduce costs. More broadly, AI-based analysis of genomic sequences could improve healthcare for billions of individuals by providing access to precision genomic health.”

Large language models for genetic diagnosis

To overcome these problems, the researchers designed AI pipelines capable of processing a list of candidate genes to identify those most likely to cause disease.

The researchers used two general-purpose LLMs, namely, Med-PaLM 2 (a specialist in the medical domain) and Gemini 2.5 Pro (a reasoning model for complex problem solving). Gemini 2.5 Pro is a newer generation model and, unlike Med-PaLM 2, is not medically trained or fine-tuned.

First, Med-PaLM 2 was asked to analyse genes from mouse GWAS studies. It not only correctly identified genes with verified causative factors but also discovered a new genetic factor causing spontaneous hearing loss. This finding was validated experimentally.

The team then moved on to using Gemini 2.5 Pro in two human studies: a hearing loss study of 20 patients and a rare genetic disease study of 6 patients.

Tao Tu, research Scientist at Google DeepMind and co-first author of the study, describes the AI pipeline they developed and how it works: “The pipeline worked by taking a filtered list of patient genes, asking the model to research those genes against medical literature and specific patient symptoms, and then requiring it to reason its ranking with cited evidence. The AI’s conclusions were compared against diagnoses made by a clinical geneticist and an otolaryngologist.”

Gemini 2.5 Pro successfully identified genetic factors causing hearing loss without any medical finetuning or training. Because of the complexity of the rare genetic diseases study, the team modified the pipeline to consider the multiple complex symptoms experienced by these patients. Again, Gemini 2.5 Pro successfully identified causative genetic variants aiding genetic diagnosis for rare diseases.

“These results indicate that AI-based methods could subsequently be used to efficiently determine the genetic basis for the estimated ∼350 M people globally with suspected genetic diseases or rare, undiagnosed syndromes,”wrote the researchers in their paper, published in the journal Advanced Science.

The AI future of genetic discovery

Moving forward, the researchers are integrating the AI pipeline into agentic frameworks for autonomous AI agent incorporation. This would enable plug-ins to other models containing data about gene mutations and proteins and improve the pipeline’s capabilities.

In a fast-moving field, the researchers are ambitious in their plans to advance the large language models. Peltz hopes that soon they will have a model capable of reading patients’ electronic records and analysing their genome in a fully automated way.

“These advanced LLMs will have a dramatic impact on biomedical research. They represent a new type of knowledge generating AI that has reasoning and improved agentic capabilities. This will enable those LLMs to produce novel hypotheses that can catalyze biomedical discoveries, accelerate finding new treatments, and improve healthcare,” he notes.

However, there remains a need for human evaluation and interpretation of AI output, and clinicians and geneticists will remain an important part of the process.

In addition to the improvements and integration of AI capabilities for better accuracy, Peltz acknowledges wider challenges in the field: “A major challenge for genetic research is that our analysis of allelic effects focuses on the parts of the genome that encode mRNAs and proteins (~2% of the genome). However, many genetic changes also affect the other 98% of the genome, and we need to develop better methods for interpreting the effect of those changes.”

Peltz believes the recent release of alpha genome (a deep learning model that can interpret genome sequence variations) could do exactly that and unlock information in this other part of the genome. 

In the future, genetics and AI are likely to have a prominent role in medicine, as Peltz explains: “Enabling individuals to better understand the impact of their genetic determinants could have a far-reaching and transformative impact on public health: medical practice could shift from its current focus on disease treatment toward disease prevention; customized plans for disease prevention could be developed based upon genetic factors.”

Reference: T. Tu et al., Genetic Diagnosis and Discovery Enabled by Large Language Models, Advanced Science (2026), DOI: 10.1002/advs.202518656

Featured Image Credit: Gerd Altmann via Pixabay