scrc-dnai/clinvar-diploid-snv

ClinVar Diploid SNV — mode-of-inheritance-aware genotype benchmark

A zygosity-sensitive relabelling of high-confidence ClinVar SNVs. Each variant is instantiated in both a heterozygous and a homozygous state, and the label is assigned at the genotype level under simplified dominant/recessive inheritance rules. A pathogenic recessive variant is therefore negative as a heterozygous carrier and positive as a homozygote — so variant identity alone is insufficient to predict the label, and a model must represent allele dosage.

Built for the diploid genomic language models at scrc-dnai and run through GFMBench-API.

Labelling rule

ClinVar class	Inheritance	het label	hom label
Benign / Likely benign	any	0	0
Pathogenic / Likely pathogenic	dominant (AD, XL-dom)	1	1
Pathogenic / Likely pathogenic	recessive (AR, XL-rec)	0 (carrier)	1 (affected)

These are benchmark labels, not revisions to ClinVar. A pathogenic recessive variant remains pathogenic in a heterozygous carrier; the label encodes an expected genotype-level clinical outcome under a deliberately simplified model. Penetrance, variable expressivity, hypomorphic alleles, compound heterozygosity and gene-specific exceptions are all omitted.

Composition

170,284 genotype instances = 85,142 unique SNVs x {het, hom}, across 4,121 genes on chr1–22, X and Y. chrX contributes 5,464 instances; chrY contributes exactly 2 (one variant).

That single chrY variant is worth knowing about: its het row is biologically meaningless, since Y is hemizygous. It is a consequence of the construction instantiating every retained variant in both zygosity states without a sex-chromosome exception, and it mirrors the encoding-level limitation that hemizygous calls are folded into homozygous states. Two rows out of 170,284 will not move a metric, but drop chrY if you want the set to be strictly interpretable.

	instances	label 0	label 1
train (all chromosomes except 19–22)	147,894	140,972	6,922
test (chr19, 20, 21, 22)	22,390	21,414	976
total	170,284	162,386	7,898

Split is by chromosome, so no genomic position is shared across train and test.

Positive prevalence is 4.6%. Report AUPRC alongside AUROC; the reference runs used inverse-frequency class rebalancing (weighted random sampler, power 1.0).

Label breakdown by zygosity — note the asymmetry that makes the task zygosity-sensitive:

zygosity	label 0	label 1
het	82,663	2,479
hom	79,723	5,419

The 2,940 extra positives in the homozygous arm are exactly the pathogenic recessive variants, which flip label between the two zygosity states.

Columns

Column	Description
chrom, pos	GRCh38 coordinates (1-based, VCF convention)
ref, alt	single-base alleles
gene	ClinVar gene symbol
zygosity	het or hom — which genotype this row instantiates
label	genotype-level outcome (0/1) per the rule above
genotype_pair, diploid_token	provenance only — see the warning below

Source and filters

ClinVar January 2026 release: variant_summary_2026-01.txt.gz for variants, ClinVarVCVRelease_2026-01.xml.gz for mode of inheritance.

• assembly GRCh38; OriginSimple != somatic; canonical chromosomes
• Type == single nucleotide variant
• clinical significance collapsed to binary; conflicting/uncertain records dropped
• review status >= 2 gold stars (multiple concordant submitters, expert panel, or practice guideline)
• mode of inheritance normalised to dominant/recessive from VCV ModeOfInheritance attributes; variants without a resolvable class dropped (one MOI kept per variant, dominant preferred)
• ref and alt restricted to A/C/G/T

Encoding these genotypes for a diploid model

Each row is a genotype, not a variant: zygosity plus ref/alt fully determine the diploid state to render. To feed the DNT checkpoints, build the two haplotypes and encode them with the shipped token table:

# het -> hap0 = ref, hap1 = alt ; hom -> both haplotypes carry alt
hap0 = ref if row.zygosity == "het" else alt
hap1 = alt

then place that pair at the variant offset inside a reference-derived window and encode with diploid_encode.encode_diploid(...) from any DNT model repo. diploid_tokens.tsv here is the same table (sha256[:16] 9664edd161cc156d) those models were trained against.

Do not use the diploid_token / genotype_pair columns for that. They were written with the phased (ordered-symbol) rows of the token table — e.g. CG -> Ƈ rather than the unphased Ĉ — and the released DNT checkpoints never saw those symbols during pretraining (zero observed count). The benchmark task code ignores these columns and re-derives the encoding from zygosity/ref/alt; do the same. They are retained only for provenance.

Reproduce

create_diploid_snv_parquet.py (included) regenerates this file from the ClinVar releases. The benchmark task expects it at data/clinvar_zygosity/clinvar_heritable_diploid_benchmark.parquet; that exact file is kept under original/ here.

Citation

Manuscript in preparation: A Diploid Genomic Foundation Model, Leib, Zinger, Ofer, Kellerman, Nayshool et al. Please also cite ClinVar (Landrum et al.) and GFMBench-API (Larey et al., 2026).