scrc-dnai/clinvar-diploid-indel

ClinVar Diploid Indel — zero-shot genotype benchmark

High-confidence ClinVar insertions and deletions, each instantiated in a heterozygous and a homozygous state and labelled at the genotype level under the same simplified inheritance rules as the SNV benchmark.

Its purpose is narrower: to test whether a representation that keeps the actual inserted and deleted bases — rather than collapsing every indel into one generic symbol — carries usable variant-effect signal. Evaluated zero-shot, as a single held-out set with no fine-tuning, scored by masked log-likelihood ratio or by cosine distance between variant and reference embeddings.

Composition

3,004 genotype instances = 1,502 unique indels x {het, hom}, across 481 genes on 23 chromosomes.

	label 0	label 1	total
het	857	645	1,502
hom	82	1,420	1,502
total	939	2,065	3,004

This set is strongly pathogenic-enriched. Only 82 of 1,502 variants are benign — ClinVar's curated indels are overwhelmingly pathogenic. The 939 negatives are mostly heterozygous carriers of pathogenic recessive indels (775 of them), not benign variants. Treat the negative class accordingly: it is largely "pathogenic allele, carrier dosage", not "harmless".

Property	Distribution
indel_type	2,898 deletion / 106 insertion
dominant_flag	1,610 recessive / 1,394 dominant (instances)
quality_stars	2,172 two-star / 832 three-star
delta_len	median -1, IQR -4 to -1, min -7,636, max +32

Length change is heavily skewed toward short deletions, with a long tail of large ones — the -7,636 bp outlier is a multi-kilobase deletion, so windowing must tolerate alleles far longer than the model's context.

Columns

Column	Description
chrom, pos	GRCh38 coordinates (1-based, VCF convention)
ref, alt	allele strings; explicit A/C/G/T only, sharing a common left-anchor base
gene	ClinVar gene symbol
zygosity	het or hom
label	genotype-level outcome (0/1)
indel_type	insertion or deletion, from the sign of the length change
ref_len, alt_len, delta_len	allele lengths and alt_len - ref_len
dominant_flag	1 = dominant, 0 = recessive
quality_stars	ClinVar review status as gold stars (>= 2 here)

Source and filters

Same ClinVar January 2026 release, assembly, germline and 2-gold-star thresholds and mode-of-inheritance relabelling as the SNV benchmark, but retaining insertion/deletion/indel types. Additionally: symbolic, missing and star alleles rejected; ref and alt required to differ in length and share a common left-anchor base, consistent with left-aligned VCF representation.

Indel genotypes are rendered through local biallelic alignment: the shorter allele is padded with gaps inside a brace-delimited span, and each aligned base pair becomes one diploid token, so allele content survives instead of collapsing into a generic event token.

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.

A note on homozygous deletions

Worth knowing when scoring this set: in the v6 encoding a homozygous deletion leaves no trace at all — both haplotypes lose the same bases, the position is dropped, and no marker or gap token is emitted. Heterozygous deletions are represented (as gap-heterozygous states inside {...}). The homozygous arm of a deletion row is therefore encoded as plain shortened reference.

Reproduce

create_diploid_indels_parquet.py (included) regenerates this file. The benchmark task expects it at data/diploid_clean/clinvar_diploid_indel.parquet.

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).