Analytical validation

What we measured, on which genomes, with the numbers as they came out.

Variant calling at 30× and the low-pass imputation backbone, validated against public reference genomes: Genome in a Bottle (GIAB) truth sets and ten South Asian genomes held out of the imputation panel. Every figure below is reproduced unchanged from the versioned result tables of 15 October 2026.

What this page covers

The sequencing and bioinformatics layer shared by both tests: short-read variant calling at full depth, and population-reference imputation of low-pass data. Public samples only. Scores are against the Genome in a Bottle (GIAB) v4.2.1 truth sets inside their high-confidence regions, and against the hold-out genomes' own 30× genotypes.

What it does not cover yet

Clinical validation of the assays is in progress with our partner laboratory and is not reported here: the high-depth targeted read of the clinical gene set in Genavli Screen, the bundled SMN1, thalassaemia and fragile X assays, sample collection and transport, and the end-to-end interpretation workflow. Tests become orderable when that validation completes.

1 · Variant calling at 30×

Four GIAB genomes against their truth sets

Full-depth runs (about 40×) of NA12878 and the HG005 Chinese trio through our short-read calling pipeline, filtered calls only, benchmarked with the GA4GH method against GIAB v4.2.1 inside the high-confidence regions. Recall is the share of true variants found; precision the share of calls that are true.

Sample	Type	Recall	Precision	F1
HG001 (NA12878)	SNP	0.99980	0.99950	0.99965
HG001 (NA12878)	INDEL	0.99813	0.99850	0.99832
HG005 (Chinese trio, son)	SNP	0.99933	0.99957	0.99945
HG005 (Chinese trio, son)	INDEL	0.99899	0.99915	0.99907
HG006 (father)	SNP	0.99922	0.99948	0.99935
HG006 (father)	INDEL	0.99863	0.99907	0.99885
HG007 (mother)	SNP	0.99915	0.99951	0.99933
HG007 (mother)	INDEL	0.99841	0.99896	0.99868

2 · The low-pass backbone

What imputation adds at each depth

NA12878 reads downsampled to five depths from the full-depth run (an optimistic proxy for a true low-pass library), then imputed against a 1000 Genomes reference panel from which NA12878's relatives were removed. Raw columns are the calls before imputation. Non-reference discordance (NRD) and genotype r² are scored at every panel site inside the GIAB confident regions.

Depth	Raw SNP recall	Raw SNP precision	Imputed SNP recall	Imputed SNP precision	Imputed INDEL F1	NRD %	Genotype r²
0.5×	0.04993	0.76358	0.95675	0.98336	0.92122	3.933	0.9743
1×	0.18856	0.89413	0.96630	0.98972	0.92968	2.343	0.9848
2×	0.47901	0.96081	0.97118	0.99326	0.93305	1.475	0.9904
4×	0.80230	0.98931	0.97430	0.99553	0.93471	0.905	0.9941
5×	0.87221	0.99282	0.97502	0.99608	0.93492	0.769	0.9950

SNP recall by depth, before and after imputation (NA12878, GIAB confident regions)

Raw calls
After imputation

Keeping only confident genotypes (posterior ≥ 0.9)

Depth	SNP recall	SNP precision
1×	0.94903	0.99659
2×	0.96250	0.99732
4×	0.97010	0.99789
5×	0.97185	0.99811

How to read this

This is why the backbone is used for ancestry, traits and common-variant predisposition and never for rare clinical variants: common sites impute to over 97% recall at 1× and above, but the rare end does not, which the next section shows directly. Clinical genes in Genavli Screen are read at high depth and reported as observed, not imputed. 0.5× is below every product tier and is shown for completeness.

3 · South Asian genomes held out of the panel

The number that speaks for Indian patients

Ten 1000 Genomes South Asian genomes (Bengali, Gujarati, Telugu, Punjabi and Sri Lankan Tamil, two each) were removed from the reference panel, downsampled from their own 30× data and imputed against the remaining samples. Their truth is their own 30× genotypes. Scored genome-wide the hold-outs look worse than NA12878, but their truth includes hard regions and short-read error that GIAB's confident regions exclude; the right-hand columns score the same imputed genomes inside those regions, where the gap nearly closes. Cohort means over ten genomes.

Depth	Genome-wide NRD %	Genome-wide r²	Confident regions NRD %	Confident regions r²	NA12878 NRD %	NA12878 r²
0.5×	4.901	0.9666	3.895	0.9737	3.933	0.9743
1×	3.530	0.9760	2.491	0.9832	2.343	0.9848
2×	2.774	0.9811	1.682	0.9887	1.475	0.9904
4×	2.311	0.9843	1.132	0.9924	0.905	0.9941
5×	2.204	0.9850	0.989	0.9934	0.769	0.9950

Where the remaining gap is: the rare end

Dosage r² by minor-allele frequency among the panel's 504 South Asian founders, inside the confident regions. Common alleles impute as well for the hold-outs as for NA12878; alleles rarer than 1% in South Asians do not, which is why rare clinical variants are read directly rather than imputed.

Dosage r² by South Asian minor-allele frequency, hold-outs by depth, inside the confident regions

Hold-outs 0.5×
Hold-outs 1×
Hold-outs 2×
Hold-outs 4×
Hold-outs 5×
NA12878 5×, for reference

The y axis starts at 0.4 to show the shape of the curves; no value on this page is below 0.49.

Run	<0.1%	0.1–1%	1–5%	5–10%	10–20%	20–50%
Hold-outs 0.5×	0.491	0.690	0.882	0.941	0.955	0.965
Hold-outs 1×	0.548	0.748	0.912	0.960	0.971	0.979
Hold-outs 2×	0.611	0.795	0.931	0.969	0.979	0.985
Hold-outs 4×	0.678	0.835	0.942	0.974	0.983	0.989
Hold-outs 5×	0.702	0.847	0.945	0.976	0.984	0.990
HG001 5×, for reference	0.869	0.928	0.965	0.978	0.983	0.988

Methods and provenance

How these numbers were produced

Samples. GIAB HG001 (NA12878), HG005, HG006, HG007 from the public HiSeq runs; ten 1000 Genomes South Asian samples from the NYGC 30× release. No patient data.
Calling. Short-read germline calling on the public sequencing runs; the low-pass series by downsampling the full-depth reads of the same run.
Imputation. Population-reference imputation against the 1000 Genomes phased panel, autosomes, GRCh38, with the target's relatives and the hold-outs removed from the panel.
Scoring. GA4GH benchmarking against GIAB v4.2.1 high-confidence regions for recall and precision; non-reference discordance and r² against the hold-outs' own genotypes, binned by global or South Asian founder allele frequency.

Known limitations. Downsampled full-depth libraries are an optimistic proxy for true low-pass libraries. Genome-wide hold-out scores include the short-read truth's own error. The confident-genotype subset at 0.5× could not be scored and is omitted.
Versioning. Result tables are generated by the pipeline and committed with it; this page is updated only from those tables. Snapshot date 15 October 2026.
Requests. Full methods, tool versions and per-sample tables are shared with clinicians and laboratories under agreement: support@genavli.in.

Back to the overview.