Methods & data
How the numbers on this site are made — and their limits. Last reviewed 2026-07-11.
How we find the blind spots
D2I2 catalogues 1,648 diseases that matter in India and asks a simple question of each one: is this a place where a genomics study could actually move the needle for Indian patients, and is anyone looking?
We answer in two tiers. Tier 1 is the tractable frontier: 22 diseases where there is a real genomic handle in our data. A handle means one of three concrete things - a European-built risk score whose mis-reading of South Asians we can measure, a gene set with pathogenic variants actually seen in South Asians, or a single variant that is far more common here than in Europe. These get a number. Tier 2 is 276 high-burden diseases with no verified India incidence and no genomic handle yet. We do not dress these up as ready studies; we mark them honestly as data gaps.
The Tier 1 number weights three things: how much the disease burdens India, how big the South-Asian blind spot is, and how tractable a study would be. The gap and the burden carry most of the weight because a blind spot only matters where the disease is common.
Technical detail
Tier-1 whitespace score = 0.45 x IndiaBurden + 0.45 x SouthAsianGap + 0.10 x Tractability, each term on a 0-1 scale.
IndiaBurden reflects India disease burden; SouthAsianGap is the size of the measured South-Asian genomic blind spot (PRS mis-stratification magnitude, or the count of South-Asian-observed pathogenic variants relative to the strongest gene set); Tractability reflects how runnable a validation study is once an Indian genotyped-plus-phenotyped sample is in hand. Burden and gap are weighted equally and dominate, because a transferability gap is only worth chasing where the disease is common. Tractability is a light thumb on the scale, not a driver. A disease qualifies for Tier 1 only if it carries at least one of the three handle types (prs, common_variant, rare_variant); everything else that is high-burden but handle-less falls to Tier 2 as a descriptive gap. The score deliberately concentrates on the tractable frontier because you can only run a genomics study where a handle exists.
How a risk score can misread an ancestry
A polygenic risk score adds up many small-effect DNA variants into one number that estimates your genetic risk for a disease. The catch: almost every score in wide use was built and calibrated on European data. Its cut-off for 'high risk' is set to the European top 10 percent.
Think of it as a ruler set for the wrong crowd. If you calibrate a height ruler on one population and then hold it up against another whose average is different, everyone in the second group reads too tall or too short - not because they changed, but because the ruler's zero is in the wrong place. That is what a European risk score does to South Asians.
We measure the error two ways. 'Mis-stratification percent' is the share of a population that lands above the European high-risk cut-off. Europeans sit at 10 percent by definition. When a South-Asian group reads 30 percent, the score is flagging three times as many people as it was designed to - not because they are sicker, but because the ruler is off. 'Mean-shift in SD' says how far that group's average score sits from the European average, measured in standard deviations. A shift of +0.80 SD means the whole distribution has slid up by most of a standard deviation.
Worked example - the Type 2 diabetes over-flag. We ran the European-built T2D score PGS000033 (ten variants) across the 1000 Genomes populations. Telugu Indians land at 30.6 percent above the European high-risk cut-off, versus the 10 percent the score was designed for - a 3.1x over-flag, driven by a +0.80 SD upward mean-shift. One reason is the MTNR1B risk variant rs10830963, carried at 42.6 percent in South Asians against 28.8 percent in Europeans. The score is not wrong about the DNA; its zero point is simply set for the wrong crowd. Note this cuts both ways: for coronary artery disease the same method shows the European score placing only 0.7 percent of Sri Lankan Tamils above the cut-off, badly UNDER-warning a group with well-documented excess heart disease.
Technical detail
Method: analytical Hardy-Weinberg mean/variance from 1000 Genomes phase-3 allele frequencies. For a PRS summed over independent variants, the population mean is sum(2 x p_i x beta_i) and the variance is sum(2 x p_i x (1 - p_i) x beta_i^2), where p_i is the effect-allele frequency in that population and beta_i the reported effect size. Swapping the EUR frequency vector for a target population's vector shifts the mean; we express that shift in EUR standard-deviation units (shift_vs_eur_sd). Mis-stratification percent is then the mass of a Normal(shifted mean, population variance) lying above the EUR 90th-percentile threshold; EUR is 10.0 percent by construction. For PGS000033 the ordered shifts run EUR 0.0 SD (10.0%), through PJL +0.544 SD (22.4%), GIH +0.587 (22.6%), BEB +0.641 (24.5%), STU +0.789 (29.9%), to ITU +0.796 SD (30.6%). This model captures the allele-frequency mean-shift ONLY. It does not model linkage-disequilibrium differences or effect-size (beta) heterogeneity between ancestries, both of which can further degrade transferability. It is an analytical screen for where mis-calibration is likely and large, not an individual-level validated risk estimate.
Our data sources & versions
Every number on D2I2 traces to a public, citable source. We do not generate genotypes; we combine reference datasets to spotlight where an India-facing study is missing.
Allele frequencies come from the 1000 Genomes Project phase 3, using its five South-Asian sub-populations: Punjabi from Lahore (PJL), Gujarati (GIH), Bengali (BEB), Indian Telugu (ITU) and Sri Lankan Tamil (STU), each around 86 to 103 samples, roughly 489 South Asians in total. Polygenic scores are pulled by their stable PGS Catalog IDs (for example PGS000033 for type 2 diabetes). Variant pathogenicity predictions come from AlphaMissense, accessed through dbNSFP and MyVariant.info. Where-is-it-seen and how-often counts come from gnomAD. Plain-language disease summaries are drawn from Wikipedia and then simplified.
Technical detail
1000 Genomes Project phase 3 (2015 release, GRCh37/GRCh38-mapped allele frequencies) - South-Asian panel PJL (n~96), GIH (n~103), ITU (n~102), STU (n~102), BEB (n~86); ~489 SAS individuals; note GIH/ITU/STU are diaspora-sampled (Houston/UK). PGS Catalog - scores referenced by immutable PGS IDs (PGS000033, PGS000010, PGS000047, PGS000061, PGS000067, PGS000350, etc.). AlphaMissense (Cheng et al., Science 2023; ~71M proteome-wide missense predictions) served via dbNSFP and the MyVariant.info API. gnomAD (v4-era joint exome+genome frequencies) for South-Asian observed counts and European-absence flags. Wikipedia for lay summaries, cleaned of stray replacement characters before display. All version pins are advisory: the pipeline that regenerates these JSON files is the source of truth for the exact build vintage.
What we do NOT claim (limitations)
D2I2 is a map of where to look, not a diagnosis and not a finished result. Three honest limits.
First, the PRS mis-stratification is an analytical estimate from allele frequencies. It has NOT been validated against real health outcomes in Indian individuals. It tells you a European score is probably mis-calibrated here and roughly how badly; it does not tell you any one person's risk.
Second, the reference panels are small. A few hundred South Asians in 1000 Genomes, mostly sampled from the diaspora, cannot capture the genetic depth of a subcontinent with thousands of endogamous communities. A signal absent from the panel is not proven absent from India.
Third, our model captures the allele-frequency shift and nothing more. Real transferability also breaks on linkage-disequilibrium differences and on effect sizes that differ between ancestries. Those can make the true gap larger or smaller than our screen suggests. Every Tier 1 entry is a fundable hypothesis, not a conclusion.
Technical detail
The mean-shift model assumes Hardy-Weinberg equilibrium, variant independence (no LD modelling), and ancestry-invariant effect sizes - all approximations. Consequences: (1) shift_vs_eur_sd and misstrat_pct are screening estimates, unvalidated against individual Indian phenotype data; (2) small, partly-diaspora reference panels under-sample India's endogamous structure, so European-absence and SAS-observation flags carry sampling error; (3) LD and beta heterogeneity are unmodelled, so true PRS transferability loss may exceed or fall short of the allele-frequency component alone. Tier-1 items are validation hypotheses conditional on obtaining a genotyped-plus-phenotyped Indian cohort; they are not claims about deployed clinical performance.
Data provenance & the GenomeIndia gap
Here is the most important caveat, stated plainly: our allele frequencies rest on a few hundred South Asians in the 1000 Genomes reference panel. That is thin ground for a country of India's genetic depth, and it is the single biggest thing standing between D2I2's screening estimates and validated, India-specific numbers.
The fix now exists. In January 2025, the GenomeIndia project released whole-genome data for 10,000 Indians spanning 83 population groups, archived at the Indian Biological Data Centre (IBDC) and accessible to researchers under the BIOTECH-PRIDE guidelines. It is the foundation for a genuine Indian reference genome. Two earlier resources also help: CSIR's IndiGen programme sequenced 1,029 Indian genomes (IndiGenomes database, CSIR-IGIB), reporting that roughly a third of the variants it found were not in global databases; and GenomeAsia 100K published a pan-Asian pilot of 1,739 individuals across 219 population groups in Nature in 2019.
Our plan is to swap the thin 1000 Genomes South-Asian panel for GenomeIndia and IndiGenomes allele frequencies as access is secured, recompute every mis-stratification estimate on India-native reference data, and version the result. That single upgrade turns a screening signal built on ~489 South Asians into one grounded in tens of thousands of Indian genomes.
Technical detail
Current allele-frequency substrate: 1000 Genomes phase-3 SAS (~489 individuals). Target substrate: GenomeIndia (10,000 whole genomes, 83 communities; released Jan 2025 via IBDC under FeED / BIOTECH-PRIDE governance) and IndiGenomes (1,029 genomes, CSIR-IGIB; ~32% India-unique variants), with GenomeAsia 100K (1,739-individual pilot, Nature 2019) as a pan-Asian cross-check. Migration path: (1) obtain access under the respective data-access protocols; (2) recompute HWE mean/variance mis-stratification and SAS-observed pathogenic-variant counts on India-native frequencies; (3) publish a versioned, DOI-tagged dataset so downstream users can pin to a build. Until that lands, D2I2 numbers should be read as 1000G-referenced screening estimates, not GenomeIndia-grounded facts.
Sources
How to cite
D2I2: Decoding Disease in India. India Whitespace Score and genomic transferability atlas. d2i2.org (accessed 2026-07-11).
A versioned, DOI-tagged dataset release is planned. Once the pipeline output is pinned to GenomeIndia / IndiGenomes reference frequencies, each build will carry a stable version identifier and a citable DOI, so analyses can reference the exact data vintage rather than a live-site snapshot.