A New AI Tool Called MARRVEL-MCP Is Changing the Slowest Part of Rare Disease Diagnosis

The Slowest Part of Rare Disease Diagnosis Is Not the Test

If you have spent any time in the rare disease world, you know the cliche about diagnostic timelines. Average rare disease diagnosis takes 5 to 7 years and several misdiagnoses. The reason people assume the bottleneck is the test itself, that getting whole exome or whole genome sequencing must be the hard part. It is not. Sequencing turnaround has collapsed from months to weeks to days. A full whole genome sequence now runs about $500 to $1,500 and comes back in 2 to 6 weeks in most clinical workflows. The hard part is what happens after the sequence comes back.

Variant interpretation is the bottleneck. A typical whole genome sequence finds 3 to 5 million genetic variants, of which a few thousand might fall in protein-coding regions and a few dozen might plausibly relate to the patient's symptoms. Sorting through those candidates requires a clinical molecular geneticist to consult multiple databases (gnomAD for population frequency, ClinVar for prior reports, OMIM for disease associations, ortholog databases for animal-model evidence, expression atlases like GTEx, and the published literature) and combine the evidence into an interpretation. Even for experts, this takes hours per variant. The bottleneck is not technology. It is human expertise, and there are only a few thousand board-certified clinical molecular geneticists in the United States.

That bottleneck is why patients sit on reports that say variant of uncertain significance (VUS) for years. A 2025 Italian study of 69 VUSes in hypertrophic cardiomyopathy reclassified 17.4% of them using updated gene-specific rules, an average of 5.7 years after the original report, and most moved toward benign (Caroselli et al., Human Mutation, 2025). A German study of arrhythmia genes reclassified 32% of 53 VUSes. Of those, 13% moved to likely benign, 13% to benign and 6% up to likely pathogenic (Martin et al., BMC Cardiovascular Disorders, 2024). In a neurodevelopmental registry, 25.4% of single-gene VUSes were upgraded to likely pathogenic or pathogenic (Kowanda et al., Genetics in Medicine Open, 2024). The implication is uncomfortable. Many patients holding a VUS report from 2020 are sitting on a result that today's evidence would call something different. Nobody has the time to recheck every patient's report every year.

On May 21, 2026, a team at Baylor College of Medicine and Texas Children's Hospital published a paper in The American Journal of Human Genetics that describes a new system called MARRVEL-MCP (Everton et al., 2026, doi 10.1016/j.ajhg.2026.04.012). The work builds on the same team's earlier MARRVEL tool, an aggregator of rare disease genetic databases that launched in 2017 and had more than 43,000 users worldwide in 2025 alone, according to Baylor. The new version adds an artificial intelligence interface that lets you ask questions in plain English.

Mechanically, MARRVEL-MCP is a server that exposes 44 different tools to a large language model. The tools cover gene and variant utilities (HGVS validation, transcript lookups), pathogenicity databases (gnomAD, ClinVar, CADD, REVEL, SpliceAI, AlphaMissense), phenotype resources (HPO, OMIM), expression atlases (GTEx, BrainSpan), ortholog and model organism data (yeast, fruit fly, zebrafish, mouse), and the published literature. When you ask the AI a question like "Is this BRCA1 mutation linked to cancer?", the AI figures out which of the 44 tools it needs to call, queries them in the right sequence, and combines the results into a plain-English answer with the supporting evidence shown.

The natural-language layer matters enormously. The original MARRVEL was powerful but only usable by people who already understood HGVS notation, knew what to type into gnomAD, and could synthesize across 30 web pages worth of results. MARRVEL-MCP removes that expertise tax. A genetic counselor, a curious caregiver, a primary care doctor who is not a geneticist, or a researcher in an adjacent field can now ask questions and get evidence-grounded answers. The team made it freely available at chat.marrvel.org with no installation required.

The 41 to 94 Percent Accuracy Result That Has Geneticists Paying Attention

The headline finding from the MARRVEL-MCP paper is that a small open-source language model (gpt-oss-20b, a 20-billion-parameter model that runs locally on consumer hardware) went from 41% accuracy on Mendelian disease variant interpretation tasks to 94% accuracy when given access to the MARRVEL-MCP toolset. Without the tool, the same model got barely two in five answers right. With the tool, it got more than nine in ten right. This is not a small marginal improvement. This is the difference between a useful research aide and an unreliable chatbot.

“By giving smaller models access to the right curated tools and structured context, we can make them smarter for specialized tasks. For example, gpt-oss-20b, a model that can be installed locally, improved to 94% with MARRVEL-MCP from 41% accuracy without MARRVEL-MCP.”

Dr. Hyun-Hwan Jeong, co-corresponding author, Baylor College of Medicine

The result matters because it suggests something important about the path forward for medical AI. The dominant story for the past two years has been about ever-larger frontier models (GPT-4, Claude, Gemini) trained on more data with more compute. The Baylor team's result suggests that for specialized domains like rare disease genetics, small models with the right curated tools matched or beat larger models working without tools, and gpt-oss-20b with the tools came close to the best proprietary models. That has practical consequences. A 20-billion-parameter model can run on a single workstation, which means the patient's genetic data never has to leave the hospital network or be sent to a third-party API. For pediatric genetic medicine where data sensitivity is paramount, that privacy story is as important as the accuracy story.

An important caveat. The 94% number is the pass rate on 100 expert-curated questions set up by the researchers, not a guarantee that any individual patient's variant will be correctly interpreted. Real clinical interpretation requires correlating with the patient's specific phenotype, family history, and clinical context that the test alone cannot supply. The right way to read the result is that MARRVEL-MCP raises the floor on what a non-expert plus a tool can do, not that it ceilings what an expert can do or replaces clinical judgment.

Why the Model Context Protocol Behind MARRVEL-MCP Matters

The MCP in MARRVEL-MCP stands for Model Context Protocol, an open standard that Anthropic released in November 2024 and donated to the Linux Foundation in December 2025. MCP solves a specific problem in AI engineering called the M-by-N problem, where M different AI models each need to talk to N different external tools or data sources. Before MCP, every model-tool pairing required a custom integration. Connecting 10 models to 100 tools meant 1,000 custom integrations. MCP collapses this to a single standardized protocol, the same way USB collapsed dozens of incompatible cable standards into one.

For rare disease and medicine more broadly, the architectural choice matters in a way that goes beyond engineering convenience. A general-purpose language model asked a medical question will sometimes hallucinate plausible-sounding answers that are wrong. The standard fix is called retrieval-augmented generation (RAG), where the model is given relevant documents to ground its answer. MCP is the next step in that direction, giving the model the ability to actually use tools instead of merely looking things up, perform queries in the right order, and combine the results. The output is anchored in the actual database responses, not in the model's parametric memory. For a rare disease patient, this means the chatbot answering your question is calling gnomAD, ClinVar, and the literature databases directly, and what you see is what those databases actually say.

Adoption of MCP across the AI industry has been rapid. By December 2025, Anthropic reported more than 10,000 active public MCP servers, and products including ChatGPT, Gemini and Microsoft Copilot had adopted the protocol. MARRVEL-MCP is one of the first major academic-medical applications of the protocol, but it almost certainly will not be the last. Expect MCP-style integrations for other rare disease workflows, drug interaction checking, eligibility screening for clinical trials, and dozens of other patient-facing tools over the next 12 to 24 months.

What This Means for Your Variant of Uncertain Significance

The most practical implication for patients sitting on a VUS report today is that re-checking the classification just got dramatically easier. The standard clinical practice is to reanalyze VUSes every 1 to 2 years, but in real-world practice many patients never get a reanalysis. Genetic counselors are scarce, lab fees apply, and most reports come with no built-in mechanism to flag a patient when the same variant gets reclassified in a public database years later.

MARRVEL-MCP changes the economics of this. You can take a variant from your genetic test report (the gene name and the HGVS notation, both should be on your printed report) and ask the chatbot whether the current evidence supports the original classification. The system will pull the latest ClinVar entries, check the gnomAD population frequency, run pathogenicity predictions, look up any new published case reports, and tell you what the aggregated current evidence says. This is not a clinical reclassification (you still need your geneticist to formally rerun the ACMG criteria for that), but it is a strong signal that points you toward whether re-analysis is worth requesting from your clinical lab.

For patients with rare disease genetic test reports older than 2 years, an annual MARRVEL-MCP check is now a reasonable practice. Reclassification rates of 17 to 32% across various disease groups mean that real value is sitting in those old reports, and the cost of checking is now your evening rather than your $400 reanalysis fee plus a 6-month wait. If the check surfaces new evidence pointing in a particular direction, that gives you a defensible reason to ask your genetics team for a formal reanalysis.

How Patients and Caregivers Can Actually Use MARRVEL-MCP Today

The public hosted version of MARRVEL-MCP is at chat.marrvel.org. No account required. To get useful answers, the question matters. Vague questions like "is this bad" produce vague responses. Specific questions that include the gene symbol and the variant in HGVS notation (the format that looks like c.123A>G or p.Arg45Cys, which appears on every modern genetic report) produce answers grounded in the actual database evidence.

Good question patterns that work well include the following. "What does the current evidence say about the pathogenicity of BRCA2 c.5946delT?" "Has anyone reported the SCN1A R1648H variant in patients with Dravet syndrome?" "What is the population frequency of this variant in gnomAD?" "Are there published case reports of patients with both this gene mutation and these symptoms?" The chatbot will return the underlying data plus its synthesis, so you can see the actual ClinVar entries, the actual frequency numbers, and the actual citations it relied on. Bring the answer to your appointment, not as a substitute for your geneticist's judgment, but as a starting point for the conversation.

Important safety practice. The chatbot is a research tool, not a clinical decision support system. Do not stop, start, or change any medical treatment based on what it tells you. The output is not part of your medical record. The reasoning chain that the LLM uses is mostly transparent (the tool calls are shown), but the synthesis layer can still occasionally misinterpret a database result. Treat the output as the genetic-literature equivalent of a thorough PubMed search summarized by a careful research assistant, not as a verdict from a board-certified clinician.

What MARRVEL-MCP Cannot Do (And Why Your Geneticist Still Matters)

The tool is genuinely good at what it does. It is also limited in specific ways that any patient relying on it should understand.

First, MARRVEL-MCP cannot do clinical correlation. It does not know your phenotype, your family history, your symptom timeline, or your prior labs. Variant interpretation in ACMG-AMP standards requires combining variant-level evidence with patient-level evidence. A variant that is "likely pathogenic" in someone with a matching phenotype may be a coincidental finding in someone whose symptoms do not fit. Your clinical geneticist or genetic counselor brings that correlation, and no current AI can replace it.

Second, MARRVEL-MCP is not FDA-cleared as a diagnostic device. The MARRVEL team is explicit about this. The tool is intended for research use and patient education, not for diagnostic decision-making. Insurance coverage decisions, clinical care plans, and treatment choices need to flow through your medical team, not through a chatbot.

Third, the databases the tool queries have known coverage gaps. gnomAD underrepresents non-European populations. ClinVar entries vary in quality and depth depending on which lab submitted them. The published literature has its own biases (more papers on the genes that already get studied, fewer on the genes that do not). A MARRVEL-MCP query is only as comprehensive as the databases it pulls from. For patients from underrepresented backgrounds, the limitations of the source data may be larger than the limitations of the AI layer.

Fourth, hallucination risk is reduced but not eliminated. The MCP architecture grounds the LLM in actual tool outputs, which is a huge improvement over a vanilla chatbot. The synthesis layer where the model summarizes the tool outputs into a plain-English answer is still a language model output, and it can occasionally misstate or oversummarize what the underlying data says. The transparent tool-call chain in MARRVEL-MCP lets you check this, but checking is on the user.

Why a Small Open-Source Model Coming Close to Frontier Models Is the Real Story

Most rare disease AI coverage will focus on the 94% accuracy number. That is the headline, and it deserves attention. The deeper story is the model size. The Baylor team got 94% accuracy from gpt-oss-20b, a 20-billion-parameter open-source model that runs locally on a single high-end workstation with a consumer GPU. Frontier models are far larger, run in data centers, and usually require sending your data to a third-party API.

For pediatric genetics, the privacy and equity implications are enormous. A children's hospital can run gpt-oss-20b plus MARRVEL-MCP locally on its own hardware, never letting patient genetic data leave the institutional network. That eliminates a class of HIPAA, IRB, and data-governance concerns that have slowed AI adoption in medical genetics for years. It also means that resource-limited settings (community hospitals, international clinics, rural genetic services) can run a research-grade variant interpretation pipeline without paying frontier API fees that scale with usage.

More broadly, the result is a data point in an ongoing debate about whether the future of medical AI is dominated by a few large API-based foundation models or by smaller specialized local models running domain-specific tools. The MARRVEL-MCP result is a strong vote for the second path. For diseases where the relevant knowledge is structured and curatable (which describes most of Mendelian genetics), giving a small model the right tools may matter more than scale.

What Is Coming Next for AI in Rare Disease Diagnosis

Dr. Liu, co-corresponding author on the MARRVEL-MCP paper, told the Baylor press team that the next step is to revamp the main MARRVEL platform with agentic AI features. In practice, that means moving from "the user asks, the system answers" toward "the system takes independent actions on the user's behalf." Imagine a service that monitors your VUS in the background, checks the relevant databases every week, and notifies you when new evidence shifts the classification. That is the agentic direction.

Other groups are working on adjacent pieces. Several research teams have been building AI tools for clinical trial eligibility matching, including some that use MCP-style architectures to query ClinicalTrials.gov directly. Phenotype-to-gene matching tools (think Face2Gene for facial features, plus newer tools that match symptom clusters to candidate disease genes) are getting LLM upgrades. Multi-modal models that can read both genetic data and clinical notes are entering early clinical pilots. The pace of change suggests that the patient-facing rare disease tooling of 2028 will look qualitatively different from 2026.

For patients on the diagnostic odyssey right now, the practical takeaway is that the tools you are working with this year are likely the worst they will ever be. The technology is improving fast. Stay connected to your patient advocacy organization, ask your genetics team about reanalysis on a 1 to 2 year cycle, and check MARRVEL-MCP yourself periodically. The answer to your question may already exist and just not have reached the front of your medical record yet.

Frequently Asked Questions About MARRVEL-MCP and AI Variant Interpretation

Find Your Way Forward in the Diagnostic Journey

If you are early in the diagnostic odyssey and have not yet had whole exome or whole genome sequencing, our guide on how whole genome sequencing can end a diagnostic odyssey covers the basics of the test, insurance coverage, and what to expect from the results. If you are further along and sitting on a VUS or an inconclusive report, MARRVEL-MCP gives you a new tool to add to your toolkit. If you suspect a specific rare disease but have not yet pursued testing, our rare disease pages list active trials by condition and link to genetic counseling resources.

Trial Friend lists active clinical trials for hundreds of rare diseases in patient-friendly language, including which antibody subtypes, gene variants, or clinical criteria each trial enrolls. If MARRVEL-MCP helps you identify a likely diagnosis, the next question is often what trials are open for that condition. We try to make finding the answer take minutes rather than the weeks it would take to do the same search manually. The diagnostic odyssey is still long, but it is getting shorter every year. The tools are improving. So is the community supporting you through it.