Yesterday, HuggingFace – in collaboration with the MVA Society, SageBionetworks, BEACON, Anthropic, and Amazon Web Services (AWS) – announced a hackathon aimed at finding answers for a child living with Mosaic Variegated Aneuploidy (MVA). This ultra-rare genetic condition affects fewer than 50 people worldwide. The child’s family has graciously opened his case to the research community in the hope of finding answers.
This initiative aligns directly with the mission at Gamow Labs: to build the AI-native approach that turns genomes and phenotypes into auditable clinical insight.
Within an hour of downloading the contest files, we had submitted the first correct submission on the leaderboard. We achieved a perfect score, exactly detecting the compound heterozygous pathogenic variants in the BUB1B gene. We accomplished this using “George,” our agentic harness for variant interpretation, specifically designed for tasks like this.
Our analysis of the genome
The hackathon gave participants access to the child’s whole genome sequence and clinical information. The first question was straightforward: does the genome contain variants that explain his MVA?
We started by inspecting the genes known to cause Mosaic Variegated Aneuploidy: BUB1B, CEP57, and TRIP13. The strongest signal was in BUB1B, where we found two rare variants in the gene:
BUB1B c.2210T>G, resulting inp.(Leu737Ter), a nonsense variantBUB1B c.3006T>G, resulting inp.(Asn1002Lys), a missense variant
Both variants were heterozygous and could affect opposite BUB1B copies, but the available singleton data did not establish phase. This is exactly the kind of pattern we would expect for an autosomal-recessive condition such as MVA1, where affected individuals need damaging variants affecting both copies of the gene.
Additionally, BUB1B is not a speculative disease gene. ClinGen classifies its relationship with MVA1 as “Definitive”, and loss of BUBR1 function is a known mechanism for the chromosome-segregation problems that characterize the condition.
So based on that, BUB1B became our leading candidate.
Variant 1: p.(Leu737Ter)
Diving into the first variant, p.(Leu737Ter) is a premature stop in the BUB1B protein. It occurs well before the final exon-exon junction, making it likely to trigger nonsense-mediated decay and eliminate the resulting transcript. Loss of function is an established mechanism for MVA1. The variant is also extremely rare in population databases, with no observed homozygotes.
But probably most importantly, this isn’t a novel interpretation from our team. ClinVar contains multiple submissions classifying the variant as Pathogenic/Likely Pathogenic for MVA1, without conflicting interpretations.
Using ACMG/AMP criteria with ClinGen refinements, we classified it as Likely Pathogenic.
Variant 2: p.(Asn1002Lys)
The second BUB1B variant, p.(Asn1002Lys), is much harder to interpret.
It is extraordinarily rare: only one allele appears in the gnomAD v4 exome dataset, and it is absent from the available genome data. It also sits inside the BUBR1 pseudokinase domain, a region of the protein that is important for its role in the spindle assembly checkpoint.
Several of the computational predictors have split calibrations. When we are faced with something like this, the honest move is to treat the computational evidence as prioritization only and keep it out of the classification.
So our classification for p.(Asn1002Lys) is VUS.
Our submission
We are left with an established, likely pathogenic BUB1B variant and a second, essentially private variant in the same gene that could plausibly affect the protein.
Putting those pieces together, we think the most likely explanation for this child’s disease is MVA1 caused by compound heterozygous variants in BUB1B. Our estimated probability of a causal relationship is 0.85.
Our method
The point of this hackathon, we believe, was beyond having AI produce a gene name and declare victory. The goal is to find answers for this family, using systems that can help researchers and clinicians move from a genome to a candidate explanation, show the evidence behind that explanation, expose uncertainty, and identify the next experiment that could resolve it.
Our analysis used Gamow Labs’s internal system alongside targeted analysis and manual ACMG curation. We inspected the VCF, narrowed our analysis to the known MVA genes, validated the BUB1B transcript model, reviewed population frequencies and ClinVar records, checked gene-disease validity, examined protein domains and literature, and evaluated the candidate variants using ACMG/AMP criteria with ClinGen refinements.
We also independently reproduced the BUB1B finding using multiple methods. The compound-heterozygous finding was recovered from the VCF itself, and we were able to reproduce the signal using commercial LLMs prompted to investigate the known MVA genes.
Gaps we discovered
In this domain, there are a few things that still need to be investigated.
Most importantly, p.(Asn1002Lys) is widely undocumented. There is currently no published literature describing this variant, and it is not present in ClinVar. A functional assay looking at the spindle assembly checkpoint, chromosome mis-segregation, or BUBR1 in patient cells could provide the evidence needed to move it beyond a VUS.
That aside, we know that the two variants are in the same gene. We don’t yet know whether they are on opposite copies of the gene (in trans), which is what we would expect if both variants are contributing to a recessive disease. The variants are 10,911 bp apart and were not placed in a shared read-backed phase block. Parental segregation is the most direct next test; targeted long-range or long-read sequencing could also resolve phase.
Next steps
We are so grateful for the opportunity this hackathon provided to help this family. We are actively working on our submission to Track 2 to further this investigative research.
Our submission gives a concrete, testable hypothesis about what is causing this child’s disease, along with a clear map of the evidence and the remaining uncertainty.
These are exactly the kind of problems we want Gamow Labs to help solve. If you are similarly motivated, we are hiring.