AnukritiAI
Anukriti AIOpen the platformRequest a briefing
Anukriti AI

Model drug response across every ancestry — before the trial, not after.

Most evidence on how genes shape drug response comes from people of European ancestry. Anukriti turns population structure into an auditable input, so ancestry-specific risk surfaces early in drug development instead of late, when it is expensive to fix.

AWS AIdeasTop 20 — selected from 10,000+ ideas across 115 countries

Drug-response evidence skews European. Everyone else inherits the risk.

Pharmacogenomics is the study of how a person’s genes shape their response to a drug. Because most of that evidence comes from European-ancestry populations, risk in African, East Asian, South Asian, and admixed American groups is underrepresented — and it tends to surface at the clinical-trial stage, where changes are costly.

For the gene CYP2C19, “poor metaboliser” status carries roughly a 5× difference in risk across ancestries.

How we make ancestry an auditable input
AFREUREASSASAMR
Illustrative of published genotype-frequency differences across the 1000 Genomes phase-3 superpopulations (n=2,504). Not patient data.

Deterministic first. AI explains second. Auditable throughout.

Anukriti’s authority is a deterministic engine: the same input always returns the same result, traceable to its evidence. A plain- language AI layer sits on top for people who aren’t geneticists — but it never overrides the engine.

  1. 01

    Genotype in

    A patient or cohort genotype, resolved against the five 1000 Genomes superpopulations — 2,504 real phase-3 samples.

  2. 02

    Deterministic engine

    The authority. Given the same input it returns the same auditable result, grounded in CPIC and PharmVar — never a black box.

  3. 03

    Auditable output

    A trial-ready result with provenance and a call status — plus a plain-language AI explanation layered on top. The engine decides; the AI never does.

Every call carries provenance from CPIC and PharmVar, and resolves to one of three honest states:

  • calledEvidence sufficient — a result is returned.
  • cannot_callThe engine refuses rather than guess, and cites the rule.
  • insufficient_dataNot enough signal to decide — stated plainly.

Change the ancestry. Watch the recommendation change.

The same drug, read against different genetic backgrounds. This is a deterministic, offline illustration of what population-aware interpretation surfaces — pick a workflow and a population.

Workflow
Superpopulation
DrugClopidogrel
High risk
Gene
CYP2C19
Variant
*2/*3
Metaboliser status
Poor metaboliser
Population signal
Poor-metaboliser prevalence is markedly higher in parts of East Asia.

InterpretationSubstantially reduced active metabolite — an alternative may be warranted.

Illustrative output for research and education only. Not patient data, and not medical advice.

Tested against a 5,700-patient warfarin cohort.

On the IWPC warfarin cohort, the engine’s predicted risk tiers line up with the doses patients were actually prescribed — a clean gradient from low to high. That is a consistency check, not an outcome: it shows the tiers track real prescribing, not that any patient was dosed wrongly. We make no claim about patient outcomes from this cohort.

  • 5,700IWPC warfarin cohortPatients from the International Warfarin Pharmacogenetics Consortium used to test dosing across ancestries.IWPC · Klein et al., NEJM 2009
  • 2,5041000 Genomes samplesReal phase-3 reference samples spanning all five superpopulations.1000 Genomes Project, phase 3
  • 3Live PGx workflowsClopidogrel/CYP2C19, warfarin/CYP2C9+VKORC1, and simvastatin/SLCO1B1 — all CPIC level A.
  • 5Superpopulations coveredAFR, EUR, EAS, SAS, AMR — every major ancestry in the output.
Predicted risk tier vs. actual prescribed dose (mg / week)
45.8Low risk33.63Standard21.51High risk
anukriti-pgx-core 0.11.0 vs. IWPC, re-run 16 November 2026 (Klein et al., NEJM 2009; CPIC 2017); n = 827 / 1,267 / 1,904. Unchanged to the decimal on engines 0.7.1, 0.9.0 and 0.11.0. A retrospective consistency check on a public cohort, not a clinical validation — no patient outcome is claimed from it.

Open findingsA companion CPIC-table audit ships with open findings disclosed, including a CYP2C9 table mismatch our own audit found. We publish what doesn’t work yet, not only what does.

You can run the audit yourself.

panelaudit is public, Apache-2.0, and archived with a DOI. It answers one question, before any patient is sequenced: what a given genotyping panel gets wrong in a given population, and which position to add next. It computes from CPIC’s public allele definitions and frequencies. No patient data. No human samples. No sequence data.

Cost of transferring a European-optimal five-locus panel (misclassifications per 100,000 people)

  • CYP2C19Oceanian27,133
  • DPYDSub-Saharan African5,508
  • CYP3A5Latino4,941

Pinned by the repository’s own CI. In 25 of the 33 gene–population pairs with any measurable cost, a population-optimal panel of the same size brings misclassification to exactly zero — so the residual error is removable using variants CPIC already publishes.

Audits seven genes

  • DPYD
  • TPMT
  • NUDT15
  • CYP2B6
  • CYP2C19
  • CYP2C9
  • CYP3A5

Refuses eight more, by name

CYP2D6
Phenotype depends on copy number, which no set of rsIDs can express.
G6PD
X-linked.
NAT2
The engine’s table records a phenotype term in its function column.
SLCO1B1, VKORC1, CYP1A2, CYP3A4
The tables needed do not exist.
UGT1A1
Three alleles share one position and do not share a function, so a set of rsIDs cannot tell them apart.

Refusing by name beats producing a confidently wrong report.

Reproduce it

python -m panelaudit --gene DPYD --panel cpic4 --population "Central/South Asian"

No third-party runtime dependencies.

Read this number carefullyA misclassification is a property of a panel, not a patient. These counts are not counts of clinical harm, and we do not present them as any.

Why population-aware PGx matters.

Published evidence shows genotype-frequency differences materially change drug safety and efficacy. Four examples, each with the population that carries the highest documented risk.

  • Clopidogrel

    CYP2C19
    *2*3

    Poor metabolisers form too little active drug, weakening platelet inhibition and raising cardiovascular event risk.

    Highest risk: EAS

    Poor-metaboliser prevalence is markedly higher in parts of East Asia than in many European cohorts.

    NCBI Medical Genetics Summary

  • Warfarin

    CYP2C9 + VKORC1
    *2*3−1639G>A

    Combined variant status drives the correct weekly dose; getting it wrong risks bleeding or clotting.

    Highest risk: AFR

    VKORC1 haplotype distribution diverges most from the European reference used by most dosing algorithms.

    NIH/PMC population dosing study

  • Simvastatin

    SLCO1B1
    *5 (c.521T>C)

    Reduced hepatic uptake raises plasma statin levels, increasing myopathy and rhabdomyolysis risk.

    Highest risk: SAS

    *5 allele frequency is notably higher in South Asian populations (~16%) than in some others (~2%).

    NCBI Medical Genetics Summary

  • Carbamazepine

    HLA-B
    *15:02

    Carriers face sharply elevated risk of Stevens–Johnson Syndrome, a life-threatening skin reaction.

    Highest risk: EAS

    HLA-B*15:02 is strongly associated with severe reactions and clusters in several Asian cohorts.

    FDA recommendation background

From research prototype to clinical-grade validation.

Five phases, staged honestly. The first three are complete and live; the fourth is new and early; the fifth is in progress.

  1. 01Completed · Live

    Deterministic trial wedge

    Three shipped workflows with auditable call statuses and CPIC + PharmVar provenance.

    • Clopidogrel / CYP2C19
    • Warfarin / CYP2C9 + VKORC1 + CYP4F2
    • Simvastatin / SLCO1B1
  2. 02Completed · Live

    Reviewer trust layer

    Benchmarks, confidence calibration, reproducibility packs, evidence mapping, and reviewer reports.

    • Evidence Sufficiency Layer with named refusals
    • 12 R-rules · 10 V-rules · 9 U-rules
    • 3 bias-detection kinds
  3. 03Completed · Live

    Pilot operations

    The production API and a public, zero-auth demo surface are running, with real 1000 Genomes sample mode.

    • Production API on Azure Container Apps + MongoDB Atlas
    • Deployed engine: anukriti-pgx-core 0.11.0, reported by the API itself
    • Public zero-auth demo, 1000 Genomes sample mode
    • Per-cohort outcome distributions
    • Shareable run permalinks
  4. 04New · Early

    Trial-planning surface

    A cohort-level question for a trial sponsor: given the genetic architecture of two cohorts for one enzyme, at what fraction metabolised does a drug cross into a dosing liability. It takes no patient, no genotype and no VCF, and returns no dose.

    • Trial-planning tool at /tools/fm-star, over four API routes
    • Four named refusals: U-FM, U-PREV, U-CV, U-MODULE
    • Of three audited probes, only S-warfarin (CYP2C9) passes the extraction audit
    • Omeprazole and metoprolol therefore give upper bounds, not estimates
    • Early work — not validated, and not offered as a clinical conclusion
  5. 05In progress

    Clinical-grade validation

    Deeper concordance benchmarking and an ancestry-specific evidence pack.

    • PharmCAT concordance benchmark
    • BCHE / Vysya founder-community wedge (Kerdoncuff 2025)
    • CYP2D6 CNV calling via Cyrius
    • Full AFR-specific evidence pack
    • Partnership outreach: PharmVar, LASI-DAD, GenomeIndia

Recognised early. Built on the standards the field trusts.

Named an AWS AIdeas winner — one of 20 chosen from more than 10,000 ideas across 115 countries, and featured on the AWS Builder Center.

The AWS AIdeas program surfaces early-stage ideas building with AWS AI. Anukriti was recognised for its population-aware pharmacogenomics approach.

Built on
  • 1000 Genomes
  • ChEMBL
  • RDKit
  • CPIC
  • PharmVar

See where population-aware PGx changes your decisions.

Anukriti is a research and education platform for exploring population-specific drug response — not a clinical diagnostic. If you’re weighing the science, the market, or a partnership, we’ll walk you through what’s live and what’s next.

  • InvestorsA structural gap in a high-value market, with early validation.
  • ProgramsA small team already shipping and validating real workflows.
  • GrantsAuditable, reproducible, and built around health equity.