This repository contains K-Dense's reproducible code, compact derived results, ranked submission, and methods report for Track 1 of the Rare Disease, Real Kid: MVA Hackathon 2026.
The reproducible workflow behind K-Dense's Track 1 submission: the versioned ranking configuration, the scripts that build and validate the submission CSV from it, and compact derived evidence snapshots from the analysis.
The ranked candidate list is results/k-dense_aubcake_phenotype-guided-curation-v1.csv.
EPCR values are manually assigned ranking confidences used to order candidates; they are not
empirically calibrated probabilities. docs/methods.md documents the
approach, evidence, and limitations in full.
This records a research challenge analysis, not a clinical diagnosis. The top-ranked candidates are unphased and include a variant of uncertain significance; nothing here should be described as proven compound heterozygosity.
Install uv, clone this repository, and synchronize the locked environment:
uv sync --frozenObtain the gated challenge VCF and index under the official terms, then place them at:
data/WGS_EX2312012_HGWCNDSX7.vcf.gz
data/WGS_EX2312012_HGWCNDSX7.vcf.gz.tbi
Run the deterministic checks, annotation, submission build, and validation:
bash scripts/run_reproduction.shThe script verifies the two primary alleles against the VCF, regenerates callset reconnaissance metrics, queries current Ensembl VEP annotations, rebuilds the CSV from the versioned ranking configuration, and validates the submission schema. To reproduce the genome-wide ClinVar screen, also place a GRCh38 ClinVar VCF at data/clinvar.vcf.gz and rerun the same command. Database outputs may drift as live public resources change; the observed 2026-08-27 snapshots are preserved under results/.
config/— exact primary variants and manually curated ranking configuration.scripts/— VCF checks, ClinVar intersection, VEP annotation, CSV builder/validator, and orchestration.results/— submitted ranking and compact observed evidence snapshots.docs/methods.md— complete Track 1 methods, rationale, limitations, and references.supporting/— filled methods-description workbooks.
Raw FASTQ files, the raw VCF and index, and the phenotype document are gated challenge data and are intentionally excluded. Do not publish them. The workflow expects authorized users to acquire these inputs directly from the official dataset. See docs/DATA_ACCESS.md.
Python dependencies are pinned in uv.lock; use uv sync --frozen rather than installing unpinned packages. The final ranking combines automated extraction/screening with documented manual review. No successful Exomiser output was obtained, and no conclusion relies on Exomiser. The approximate ACMG secondary-findings gene-window enumeration from exploratory work was not a valid ACMG SF interpretation and is excluded from this public reproducibility bundle.
This repository is released under CC BY 4.0. It records a research challenge analysis, not a clinical diagnosis or treatment recommendation. Both candidate variants require orthogonal confirmation, phase requires parental or validated long-range testing, and the missense allele requires functional evaluation.