Can Large Language Models Solve More Rare Disease Genomes?
LLM-assisted rare disease re-assessment improved genomic analysis, increasing diagnostic yield and supporting clinician-led precision medicine and possibly drug development.
LLM-assisted rare disease assessment is emerging as one of the most promising applications of artificial intelligence in genomic medicine, helping clinical geneticists reanalyse previously unsolved rare disease genomes, uncovering diagnoses that conventional workflows had missed.
Why So Many Rare Disease Cases Remain Unsolved
Whole-genome and whole-exome sequencing have transformed rare disease diagnostics. Yet obtaining a molecular diagnosis remains difficult.
Current evidence suggests that more than half of patients with suspected Mendelian disorders remain undiagnosed following their initial genomic analysis, even after specialist review.
New disease genes continue to be discovered, variant classifications evolve, and genotype-phenotype relationships become better understood over time.
As a result, genomes that were once considered “negative” can become diagnosable years later through systematic reanalysis.
The challenge is scale.
Thousands of previously sequenced genomes require periodic review, but manual reanalysis demands significant clinical genetics expertise and extensive literature searches.
This creates an opportunity for AI systems capable of rapidly integrating multiple sources of biomedical evidence.
The findings in a new study published in NEJM AI demonstrate that large language models (LLMs) can be valuable decision-support tools in rare disease genomics, helping address one of precision medicine’s most persistent challenges: The growing backlog of patients who remain undiagnosed despite undergoing comprehensive genome sequencing.
Rather than replacing clinicians, the LLM-assisted technology functions as an evidence-based reasoning tool, synthesising genomic data, phenotype information, and biomedical literature to generate hypotheses for expert review.
How a Rare Disease Genome LLM-assisted workflow Works
Unlike conventional AI diagnostic systems that attempt to predict a diagnosis directly, the workflow described in the NEJM AI study was designed to support clinical reasoning.
Researchers combined:
- clinician notes
- Human Phenotype Ontology (HPO) terms
- annotated variant call format (VCF) files
- current biomedical literature
Using OpenAI's o3 Deep Research model, the system generated explanation-rich diagnostic hypotheses linking candidate variants with the patient's phenotype and supporting scientific evidence.
Importantly, the LLM did not classify variants or make clinical decisions.
Every potential diagnosis underwent expert review using established ACMG/AMP variant interpretation guidelines before confirmation.
That is critical. The model functions as a reasoning and prioritisation engine, helping clinical geneticists identify promising leads rather than replacing established diagnostic processes.
AI in Pharma: Why the Future of Healthcare Starts With Patients, Not Tech
Kate O’Reilly, President & Chair of the Healthcare Businesswomen’s Association (HBA) Dublin-Ireland Chapter & Healthcare Transformation Partner at Roche, discusses AI in pharma, patient engagement, and the future of healthcare innovation.
The LLM Rare Disease Genome Study’s Results
The retrospective study evaluated 376 previously unsolved rare disease cases drawn from multiple clinical cohorts.
The results were encouraging.
The LLM-assisted workflow identified:
|
Outcome |
Result |
|
Previously unsolved cases analysed |
376 |
|
New confirmed local diagnoses |
18 |
|
Additional diagnostic yield |
4.8% |
|
Previously established diagnoses rediscovered |
7 |
Beyond the confirmed diagnoses, the system also generated biologically plausible hypotheses that may warrant future investigation as scientific knowledge continues to evolve.
Although a 4.8% increase may appear modest, its significance becomes clearer when viewed in context. These patients had already undergone extensive genomic analysis and expert interpretation.
Identifying 18 additional diagnoses from a cohort previously considered unsolved represents a meaningful improvement for families who may have spent years searching for answers.
Why The LLM Rare Disease Genome Findings Matter for Drug Development
The applications of any potential LLM tool to help with rare disease genome identification could go well beyond clinical diagnosis.
Rare disease genomics has become an increasingly important source of target discovery for the pharmaceutical industry.
Many successful therapies have originated from understanding the genetic basis of uncommon disorders before expanding into broader disease indications.
Improving diagnostic yield creates several downstream benefits for therapeutic development.
Expanding genetically defined patient populations
More accurate diagnoses help researchers identify additional patients with specific molecular disorders, supporting natural history studies and improving recruitment for precision medicine trials.
Strengthening genotype-phenotype knowledge
Every confirmed diagnosis contributes to a growing evidence base linking genetic variants with disease mechanisms. This improves target validation and biomarker discovery across therapeutic research.
Accelerating rare disease research
By reducing the time required to identify candidate variants, LLM-assisted workflows could allow clinical researchers to focus more resources on functional validation and translational studies.
In this way, AI is supporting not only diagnosis but also the discovery ecosystem that underpins future therapeutic innovation.
AI Should Support Clinicians, Not Replace Them
One of the study’s most important conclusions is its emphasis on clinical governance.
The authors acknowledge several limitations. The research was retrospective, involved relatively modest numbers of newly diagnosed cases, and depended on the quality of clinical documentation and phenotype data available. Structural variants, repeat expansions, mosaicism, and non-coding variants also require further evaluation in larger prospective studies.
The researchers also note that LLMs may misinterpret clinical context or overgeneralise published evidence. To reduce this risk, every result underwent multiple rounds of expert review and formal variant classification before a diagnosis was accepted.
These findings reinforce an emerging consensus across healthcare AI: LLMs are most effective when embedded within validated, clinician-supervised workflows.
They are well suited to:
- synthesising biomedical literature
- prioritising candidate variants
- generating evidence-based diagnostic hypotheses
- supporting multidisciplinary genomic review
They are not a replacement for clinical judgement, laboratory validation, or established genomic interpretation standards.
LLMs in the Next Stage of Genomic Medicine
The authors of the study argue that the logical next step is prospective multicentre evaluation, measuring not only diagnostic yield but also time to diagnosis, clinician effort, downstream patient management, and reproducibility across different electronic health record systems and genomic pipelines.
They also highlight future opportunities to integrate long-read sequencing, transcriptomics, richer phenotyping, audit trails, and version-controlled prompts to strengthen governance and reproducibility.
If those studies confirm the current findings, LLM-assisted reanalysis could become a routine component of genomic medicine.
Instead of replacing clinical geneticists, these systems may help ensure that rapidly expanding biomedical knowledge is translated into diagnoses more quickly and consistently than manual review alone can achieve.
For pharmaceutical organisations investing in precision medicine and rare disease therapeutics, this highlights how explainable, clinician-supervised AI can strengthen the entire translational research pathway, from gene discovery and patient stratification to biomarker development and target validation.
At Pharmatica, we examine the technologies reshaping therapeutic drug discovery and precision medicine. By translating emerging scientific evidence into strategic Insights, we help life sciences leaders understand where advances in artificial intelligence can deliver measurable value across research, development, and patient care.
Pharmatica: Insight. Connection. Impact.
Frequently Asked Questions
What is LLM-assisted rare disease diagnosis?
LLM-assisted rare disease diagnosis uses large language models to analyse clinical notes, phenotype information, genomic variants, and biomedical literature to generate evidence-based diagnostic hypotheses for clinician review. The technology supports, rather than replaces, clinical geneticists.
How much did the LLM improve diagnostic yield?
In the NEJM AI study, the LLM-assisted workflow analysed 376 previously unsolved rare disease genomes and identified 18 additional confirmed diagnoses, increasing overall diagnostic yield by 4.8% after expert clinical validation.
Can large language models diagnose rare diseases independently?
No. The study demonstrated that LLMs should be used as clinician-supervised decision-support tools. Every diagnostic hypothesis required expert review and formal variant interpretation using established ACMG/AMP guidelines before a diagnosis was confirmed.
Why are rare disease genomes difficult to diagnose?
Many rare diseases involve complex genotype-phenotype relationships, newly discovered disease genes, evolving scientific literature, and variants of uncertain significance. These factors mean that many genomes require periodic reanalysis as biomedical knowledge advances.
What could LLMs mean for precision medicine?
LLMs may help clinicians identify missed diagnoses, prioritise candidate variants, and synthesise rapidly expanding biomedical literature. As part of validated clinical workflows, they have the potential to improve rare disease diagnosis while supporting precision medicine and future therapeutic research.
Did you enjoy the content?
Why not support Nicole Dale by giving this content a like
Comments (0)