MVA Rare Disease Hackathon · 2026

Finding one meaningful clue in an entire genome.

A genome is a three-billion-letter instruction book. We built an AI-assisted system that filters millions of spelling differences, connects them to a child’s symptoms, and ranks the explanations a specialist should inspect first.

See how it works Read the paper Explore the repository

By Mame Diarra Toure and Bonaventure F. P. Dossou

From noise to a testable lead
3 billionDNA letters
Millionsof differences
A shortlistworth reviewing
#1 BUB1Baccepted variant pair
5.08Mvariants screened
#1accepted pair rank
313,171temporal test records
275,739full-background trials
Research paper 2026

Evaluating the Ranking Layer of Rare-Disease Decision Support

Bayesian Uncertainty Under Temporal and Patient Shift

GenAI for Health Workshop @ NeurIPS 2026 · Sydney
Bonaventure F. P. Dossou
Mame Diarra Toure
The complete study

The project is more than the winning answer.

Finding the accepted variant pair in one child was the starting point. The paper asks a more demanding question: which parts of a rare-disease ranking system can be trusted when the data, the patient, or the search space changes?

In the paper, we separate candidate generation, molecular ranking, phenotype evidence, and uncertainty. We compare deterministic baselines, deep ensembles, singular Bayesian neural networks, Bayes-by-Backprop, wider architectures, and a predictor-augmented pairwise SBNN.

3distribution shifts
9model families
Read the paper Download PDF · 2.9 MB
The challenge, without the jargon

Think of it as a medical detective story.

The child’s DNA contains many differences that are completely harmless. The difficult part is deciding which few deserve attention—and doing so in a way that is fast, traceable and honest about uncertainty.

01

Read the enormous book

Sequencing turns DNA into data. We check its quality and line it up against a reference genome so each difference has an address.

02

Remove ordinary spelling changes

Most variants are common or biologically unlikely to explain the disease. Evidence-based filters remove this background noise.

03

Match DNA clues to symptoms

The remaining candidates are compared with the child’s clinical features, then several models rank the most coherent explanations.

The challenge outcome

The accepted BUB1B pair rose to the very top.

The system ranked two changes in the same gene as the strongest candidate pair. This is the type of pattern known as compound heterozygosity: two different variants may jointly disrupt the two copies of one gene.

This is a research ranking result—not a clinical diagnosis. Laboratory and family-based confirmation remain essential.

100Rank
Points
Rank #1
F-max 1.000
The leading biological hypothesis

Two variants in BUB1B, interpreted together.

BUB1B encodes BUBR1, a protein central to the checkpoint that helps cells distribute chromosomes correctly during division.

Pathogenic anchor

p.Leu737Ter

A stop-gain variant with strong clinical evidence, high predicted impact, and no observed homozygotes in the population reference used.

Key uncertainty

p.Asn1002Lys

An ultra-rare missense variant with computational support, but without independent proof of pathogenicity or confirmed phase with the first allele.

Together they form the strongest causal hypothesis for the MVA phenotype. The paper is explicit that parental or long-range phasing and functional validation are still needed.

Our approach

A chain of evidence, not a black box.

Each stage has one job. A candidate can be traced from raw genome data to its final position in the ranked list.

1

Map

Place each DNA read at the right location in the reference genome.

2

Filter

Keep rare, high-quality changes with plausible biological effects.

3

Connect

Form candidate pairs and compare their genes with the child’s symptoms.

4

Question

Ask multiple AI models—and measure when their answers are uncertain.

5

Rank

Deliver a short, auditable list for expert review.

Six-stage system diagram showing candidate generation, clinical filtering, molecular ranking, out-of-distribution warning, patient reasoning, and the Track 1 decision.
The complete system boundary. Deterministic validity checks remain separate from learned pathogenicity scores and uncertainty estimates.
Why nine models?

We treated the models like a panel of specialists.

One model can be confidently wrong. We therefore built different kinds of learners and tested them under the same rules. Some are fast pattern finders, some act like a committee, and Bayesian models keep a distribution of plausible answers instead of pretending one set of parameters is certainly correct.

What we found

The challenge score was perfect. The stronger story is what happened next.

We froze the key choices, moved forward in time, tried independent solved cases and finally searched realistic whole-genome backgrounds. Each test asks a harder question.

Challenge child
1

accepted pair rank

The BUB1B pair was placed first. All eight submission-era learned models agreed.

100 Rank Points · F-max 1.000
Later variants
84.35%

Top-1 accuracy

The selected SBNN found the correct item first when ranking later ClinVar variants among 100 candidates.

PR-AUC 0.9736 · F-max 0.9235
Independent solved cases
719

post-2023 cases

We tested symptom-aware reranking over five independently resampled candidate cohorts.

Includes 64 two-variant cases
Full-genome backgrounds
86.63%

Top-100 recall

The augmented SBNN recovered the causal variant in its first 100 suggestions across 275,739 trials.

Median causal rank: 14
Why this matters

Success on one solved child cannot select the best model.

Every compatible learned adapter placed the accepted pair first once the biologically constrained six-pair shortlist existed. We therefore evaluated the ranking layer on separate data where models could meaningfully disagree. That distinction—between producing valid candidates and ranking them—is central to the paper.

Beyond one child

Would the idea survive outside the hackathon?

A perfect score on one known case is encouraging, but it is not enough. We deliberately tested whether the models could face newer records, different patients and far larger search spaces.

A
Time shift

Train on the past, test on the future

Models learned from the December 2023 ClinVar release and were tested only on later, unseen identifiers.

B
Patient shift

Move from variants to solved cases

Phenotype information improved ranking across every evaluated model in repeated post-2023 case cohorts.

C
Scale shift

Put the signal back into genome-scale noise

2,577 causal variants were crossed with 107 real backgrounds to create 275,739 locked trials.

01
Temporal shift

Can a model recognize evidence that appeared later?

Models learned from a December 2023 ClinVar snapshot. The untouched evaluation contains 313,171 odd-ID records absent in that snapshot and present by August 2026.

Best Top-1
84.35%
Best F-max
0.9235
Best unseen-gene OOD AUROC
0.7485

Takeaway: the wide rank-50 SBNN ranked best, while the deep ensemble detected unfamiliar genes best. Ranking and uncertainty are different capabilities.

Three-panel comparison of Top-1 ranking, F-max, and unseen-gene out-of-distribution AUROC across seven model families.
Temporal ranking and unseen-gene detection select different models.
02
Patient and phenotype shift

Do molecular scores improve when symptoms are added?

We evaluated 719 deduplicated solved Phenopacket cases whose causal variants were absent from ClinVar 2023, including 64 two-variant cases closest to the hackathon setting.

Molecular-only Top-1
29.2–34.2%
Gain from HPO
+23.1–26.6 pts
Best overall Top-1
58.9 ± 1.0%

Takeaway: phenotype evidence helped strongly when historical gene–phenotype knowledge existed, but provided almost no aggregate improvement for unseen genes.

Molecular-only and molecular plus HPO Top-1 accuracy across eight models for 719 solved cases and 64 two-variant cases.
Adding HPO information improves patient-level ranking across every evaluated model.
03
Search-space shift

Can the method work against realistic genome backgrounds?

VarPB crossed 2,577 causal variants with exonic candidates from 107 real genomes, creating 275,739 locked spike-in trials with complete score-file coverage.

Augmented SBNN Top-20
65.70%
Augmented SBNN Top-100
86.63%
Median causal rank
14

Takeaway: the augmented SBNN matched REVEL within 0.20 point at Top-20 and exceeded it by 1.81 points at Top-100; REVEL remained stronger at Top-1 and MRR.

Median causal rank and Top-20 recall across original models, the augmented pairwise SBNN, REVEL, and AlphaMissense.
Compact-shortlist success does not automatically transfer to full-background prioritization.
The hardest stress test

Competitive with widely used protein-effect predictors.

On the full-background VarPB benchmark, our predictor-augmented SBNN nearly matched REVEL at Top-20 and surpassed it at Top-100. It also substantially outperformed AlphaMissense at both shortlist sizes.

Higher is better. “Top-100” asks whether the causal variant appears anywhere in the first 100 candidates sent for review.

Why Bayesian models?

A useful system should know when it might be wrong.

Traditional models often return one confident-looking score. Our Bayesian models can generate many plausible predictions, revealing where the answer is stable and where more evidence or human review is needed. That uncertainty is part of the output—not an afterthought.

The central result

One confidence score cannot represent every kind of uncertainty.

01

Molecular ranking

How strongly do the observed variant features resemble pathogenic evidence?

02

Model uncertainty

How stable is the prediction across posterior samples or independently trained networks?

03

Phenotype coverage

Does the reference knowledge actually contain the candidate gene–disease relationship?

04

Biological validity

Are the variants real, rare, inheritance-compatible, correctly represented, and plausibly in trans?

The paper argues that decision-support systems should report these signals separately so that a fluent explanation cannot hide an unsupported ranking.

Scientific transparency

Strong evidence, clearly bounded.

The repository contains the models, ranked outputs, evaluation tables, integrity hashes, tests, the paper, and supporting technical material. It intentionally excludes the private 85 GB patient bundle and all personally sensitive source data.

Browse the complete code Read the paper
What this work does not claim
  • It is a research prioritization system, not a clinical diagnostic device.
  • Constructed case cohorts do not replace prospective hospital validation.
  • The two child variants still require direct phase and functional confirmation.
  • REVEL and AlphaMissense are inputs to the augmented model, so that model is an evidence integrator—not an independent replacement.

Next: broader clinical cohorts, additional ancestries and sequencing platforms, long-read or family phasing, and prospective calibration studies.

The team

Built across machine learning and computational genomics.

GitHub repository Read the paper Mame’s research homepage Bonaventure’s research homepage