General Service

Genome Interpretation Report

已验证

Bilingual · 10-step pipeline · All verified and delivered

VCF QC -> ANNOVAR 122-column annotation -> FM tri-model pathogenicity prediction -> ClinVar+dbNSFP ACMG classification -> PharmGKB pharmacogenomics -> PRS polygenic risk -> mitochondrial variants -> NCCN cancer risk assessment -> ACMG carrier screening -> bilingual report. All 10 steps verified and delivered. This page serves three audiences: patients and public wanting to understand genome reports, hospitals and pharma seeking interpretation partners, and investors and peers evaluating technical capabilities.

Patients & Public

Read "What is a Genome Report" and "To Patients" below--you have the right to understand your own genome.

Partners / Hospitals

Focus on "DiVo's Role", "6-Step Pipeline", "Benchmarks"--all 10 steps verified and delivered, with sample reports.

Investors / Peers

Focus on "Differentiation", "Benchmarks", "Glossary"--evaluate the technical barrier of FM tri-model cross-validation.

To Patients and Families

DiVo Gen²AI's genome interpretation service is a business that works in coordination with medical institutions, not a service directly for individual users and patients. The genome interpretation report is for reference only and does not replace medical advice. If you have health concerns, please consult your treating physician or a genetic counselor.

The genome interpretation report provides patient genomic risk stratification for the following flagship services

What is a Genome Interpretation Report

Everyone's genome contains billions of base pairs with millions of variants--most are harmless, a few may cause disease. A genome interpretation report starts from raw sequencing data (VCF file), filtering, annotating, and classifying layer by layer, ultimately telling you which gene loci have clinically significant findings.

A complete interpretation report doesn't just list "which position changed"--it answers "what this variant means"--pathogenic or benign? Which drug metabolism does it affect? What's the polygenic risk score? Which genetic disease mutations are carried? DiVo's report uses three versions for different readers: a plain-language Chinese version for ordinary people, an English professional version for doctors, and a technical appendix for bioinformatics peers.

DiVo uses three genomic foundation models for cross-voting to determine pathogenicity--AlphaGenome (DeepMind), Genos (BGI), Evo2 (Arc Institute)--2/3 consensus determines the result, not a single model run.

Why Tri-Model Cross-Validation

A single foundation model may misjudge due to training data bias. Three models with different principles cross-vote, 2/3 consensus required--covering different populations and modeling methods, higher reliability.

Why a Plain-Language Version

If a genome report only has an English professional version, ordinary people can't understand it. The plain-language Chinese version directly tells you what results mean and what to do--you have the right to know your genome.

DiVo Gen²AI's Role

The genome interpretation report is the core output of DiVo's four-in-one service L1 Source Code dimension, and the input for patient genomic risk stratification in the neoantigen pipeline and CAR-T pipeline. We provide end-to-end computing from VCF to three-format reports--QC, annotation, FM tri-model prediction, ACMG classification, pharmacogenomics, PRS, mitochondrial, cancer risk, carrier screening, report generation, all 10 steps verified and delivered.

We do not do clinical diagnosis (report is for reference only), do not do wet-lab validation of variants, do not do family genetic counseling. We deliver genome interpretation reports that can be directly referenced by physicians.

Core Capability · 6-Step Pipeline

From VCF to bilingual report · All verified and delivered

Step 1VCF QC Standardizationbcftools norm + QC checkStandardized VCF + QC JSON 已验证
Step 2Variant Functional AnnotationANNOVAR (dbNSFP42a)122-column multianno.csv 已验证
Step 3FM Pathogenicity PredictionAlphaGenome · Genos · Evo2 tri-model cross200 variants · 14min · 3/3 models · 2/3 consensus 已验证
Step 4ACMG Auto-ClassificationClinVar VCF + dbNSFP 28 criteria1526 LP / 5206 VUS (6732 variants) 已验证
Step 5Specialized Analysis (4 modules)PharmGKB + PRSice-2 + Mitochondrial + NCCN/ACMG652 drug-gene / PRS / 18 pathogenic sites / cancer+carrier 已验证
Step 6Report Generation (L1/L2/L3)In-house v3 report generator (1200 lines)Bilingual MD/HTML/PDF 已验证

Differentiation

Core differences from single-model/single-format genome reports

3M

Genomic Foundation Model Tri-Model Cross

AlphaGenome + Genos + Evo2 three foundation models vote together, 2/3 consensus determines pathogenicity. Not a single model run, but tri-model cross-validation--covering different modeling principles and population backgrounds.

CN

Plain-Language Chinese Report

Not only an English professional version, but also a plain-language Chinese version for ordinary people--no need to understand algorithms or medical abbreviations, directly tells you what results mean and what to do.

Three Formats · Two Languages

Plain-Language Chinese

已验证

Easy-to-understand health action plan, no algorithm knowledge needed

MD / HTML / PDF

English Professional

已验证

VCF interpretation report for medical professionals

MD / HTML / PDF

Technical Appendix

已验证

Detailed explanation of ACMG/PRS/pharmacogenomics metrics

MD / PDF

ACMG Classification Evidence

ClinVar VCF direct parsing + dbNSFP42a 28-criteria auto-evaluation, evidence-backed

1,526
Likely Pathogenic
PP3+PS2 dual evidence
5,206
VUS
Insufficient evidence below threshold
3,308
PP3 Pathogenic Evidence
CADD≥25 / multiple support
3,834
PS2 ClinVar Pathogenic
VCF direct parsing 100% match
Samples:22:17097818:G:A IL17RA · Immunodeficiency 51 · ClinVar=Pathogenic · PP3(CADD=34)+PS2 -> Likely Pathogenic | 22:17182647:C:T ADA2 · Deficiency of adenosine deaminase 2 · ClinVar=Pathogenic · PP3+PS2 -> Likely Pathogenic

PharmGKB Pharmacogenomics

Directly parsed from PharmGKB relationships.tsv, 111,254 drug-gene associations

652
Matched variants
9
Genes involved
43
Associated drugs
Samples:IL17RA -> TNF-alpha inhibitors (PD, ambiguous)  |  ADA2 -> associated drug-gene pairs (PD)

Tri-Model GPU Inference Results

AutoDL RTX 4080 · 2026-07-02 · chr22 200 variants

3/3
Models verified
Genos + AG + Evo2
14min
Total inference time
4.2s/variant avg
16.1GB
VRAM peak
三模型常驻显存
2/3
Consensus voting
多数同意方判定
Key fixes:Evo2-1b->7b_base · vortex SDPA fallback · AlphaGenome attention bias dimension fix

Genomic Foundation Model Tri-Model Cross

AlphaGenome + Genos + Evo2 vote together, 2/3 consensus determines pathogenicity

✓ GPU

AlphaGenome

DeepMind PyTorch port · 878MB权重

8192bp context · Regulatory region variant expertise

AutoDL GPU inference verified ✓

✓ GPU

Genos-1.2B

BGI · Mixtral MoE · 4.7GB权重

8K bp context · Chinese population optimization

AutoDL GPU inference verified ✓

✓ GPU

Evo2-7b_base

Arc Institute · BF16 · 13GB权重

Long-context DNA modeling · SDPA fallback

AutoDL GPU inference verified ✓

Benchmarks

All from actual run outputs, not estimates

MetricValueNote
Run batches15+Bilingual MD/HTML/PDF all formats
Variant coverage6,732Single chromosome (chr22) complete report
ANNOVAR columns122dbNSFP42a (48GB) + refGene + ensGene
ClinVar match rate100%6,732/6,732 variants all matched ClinVar records
ACMG classification1,526 LPLikely Pathogenic accounts for 22.7% (PP3+PS2 dual evidence)
PharmGKB matches6529 genes × 43 drugs association entries
dbNSFP CADD3,387Among them 3,064 with CADD≥25 strong pathogenic signal
FM models3/3Genos + AlphaGenome + Evo2-7b all GPU inference verified (AutoDL RTX 4080)
Specialized analysis modules4/45b PRSice(BC 249/CAD 140K SNP); 5c mitochondrial 18 pathogenic sites(ClinVar chrM); 5d NCCN 15 genes; 5e ACMG 11 genes(11/11 detected)
Report formats3MD + HTML + PDF
Language中英双语Plain-language + Professional
Report versionv3In-house report generator ~1200 lines Python

Sample Reports Available

Genomic health report + technical appendix, PDF bilingual version. All 10 pipeline steps verified and delivered.

Honest Boundaries

What we can and cannot do, clearly stated

What We Can Do

VCF end-to-end interpretation (QC->annotation->classification->report)
FM tri-model pathogenicity cross-prediction (AlphaGenome + Genos + Evo2)
ACMG auto-classification + ClinVar matching
Pharmacogenomics (PharmGKB 652 entries) + PRS + mitochondrial + cancer risk
Bilingual three-format reports (plain-language + professional + technical appendix)

What We Don't Do

No clinical diagnosis (report is for reference only, does not replace medical advice)
No whole-genome WGS interpretation (currently verified to single chromosome chr22)
No Sanger validation or other wet-lab confirmation of variants
No family genetic counseling and pedigree analysis
No guarantee that all VUS will receive definitive classification

Glossary

5 core terms in genome interpretation

Abbr.Full NameTranslationExplanation
VCFVariant Call FormatVariant Call FormatStandard file format for genomic variants, recording mutation information at each position
ACMGAmerican College of Medical GeneticsAmerican College of Medical GeneticsAuthoritative body establishing genetic variant classification standards, 5-tier: pathogenic/likely pathogenic/VUS/likely benign/benign
PRSPolygenic Risk ScorePolygenic Risk ScoreAggregates small effects from multiple gene loci to assess overall genetic risk for a disease
FMFoundation ModelFoundation ModelLarge AI model pre-trained on massive data, adaptable to various downstream tasks
ClinVarClinVar DatabaseClinVar DatabasePublic database maintained by NCBI, recording relationships between genetic variants and diseases