Genomics & variant interpretation

Sequencing a patient genome is now cheap; deciding which of the millions of differences from the reference actually explain a disease is not. AI improves two steps: accurate variant calling from raw reads, and prioritising and classifying variants (protein and non-coding effect predictors, phenotype matching, literature mining). LLM agents draft the evidence summary, but classification and the clinical report stay with a qualified human geneticist.

Problem & value

• Today: a typical exome or genome yields tens of thousands to millions of variants. Analysts filter by population frequency, inheritance model and gene panels, then manually check ClinVar, literature and predictors for each candidate. Many cases stay unsolved, and variants of uncertain significance (VUS) pile up and need periodic re-analysis as evidence changes.
• AI value: better calls in difficult regions, ranking of missense and non-coding variants by predicted effect, phenotype-driven prioritisation, faster literature triage, and consistent structured evidence for each ACMG/AMP criterion. This shortens analyst time per case and makes re-analysis of old VUS affordable.
• Evidence base: the ACMG/AMP guideline defines the standard five-tier classification (pathogenic to benign) that tools map their output onto [1]; ClinGen has calibrated how computational predictors may be used as evidence strength within it [2]. Published benchmarks show modern predictors separate known pathogenic from benign missense variants well, but they remain supporting evidence, not a diagnosis [3][4].

Variants

• Clinical genetics / diagnostic lab: rare-disease and cancer germline/somatic reporting under ISO 15189 and IVDR.
• Hospital precision oncology: tumour-normal calling, molecular tumour board preparation.
• Pharma/biotech R&D: variant-to-function screens, target validation, cohort analysis (see SCI-017).
• Agri-food: see SCI-020.

How it works

1. sample
2. sequencing (short/long reads)
3. alignment + variant calling (SNV, indel, structural variants, repeats)
4. annotation (consequence, population frequency, ClinVar, gene constraint)
5. AI effect scoring (missense, splice, non-coding/regulatory)
6. phenotype matching (HPO terms, inheritance, pedigree)
7. LLM/RAG evidence assembly (literature, case reports, ClinGen curations)
8. ACMG/AMP classification proposal
9. HUMAN: variant scientist review + clinical geneticist sign-off
10. report + periodic re-analysis

• Variant calling: deep-learning callers (DeepVariant-style CNNs, DRAGEN's ML-assisted pipeline, Clair3 for long reads) outperform pure statistical callers on accuracy, especially for indels and long-read data [5][6].
• Effect prediction: protein language models and structure-aware models (AlphaMissense, ESM-based, PrimateAI-3D) score missense variants; sequence-to-function models (AlphaGenome, Enformer-class) predict regulatory and splice effects of non-coding variants; SpliceAI-style models score splicing [3][4][7].
• Phenotype matching: HPO-based ranking (Exomiser, LIRICAL), increasingly with LLMs extracting HPO terms from clinical notes.
• LLM layer: retrieval over PubMed, ClinVar and gene-disease curations; the model summarises evidence per criterion with citations. Every cited statement must link to a retrievable source.
• Human-in-the-loop: mandatory at classification and reporting. Predictors may only be applied at the evidence strength calibrated by ClinGen [2].

Data preparation

Typical inputs are raw reads (FASTQ/BAM/CRAM), VCF files, phenotype descriptions and pedigrees. They usually sit on sequencer output shares, LIMS or clinical information systems, and in free-text findings or scanned referral forms. Typical gaps: inconsistent sample and patient identifiers, phenotypes only as free text instead of HPO terms, mixed reference genome builds, missing consent flags for secondary use, and knowledge scattered over past reports and PDF literature.

Preparation steps: build a data inventory and map each data source to the use case (PREP-001, PREP-002); harmonise identifiers and genome build; code phenotypes as HPO terms; pseudonymise exports before they leave the clinical zone (PREP-022); harvest labels from past signed-off classifications as gold-standard examples (PREP-020, PREP-018); digitise paper referral forms (PREP-006); and check purpose and legal basis for reuse of genetic data (PREP-025). Tools: standard bioinformatics workflow managers (Nextflow, Snakemake), OCR for forms, a metadata catalogue.

Expect roughly half of the project effort in data preparation, validation and consent clean-up rather than in the AI components. See GUIDE-003.

Implementation paths

Which path for whom

• Enterprise (clinical labs, biotech): Path A. Certified, IVDR-aware interpretation platforms with audit trails exist and save the validation effort a lab would otherwise carry itself.
• Developer: Path C. Open-source callers and annotation pipelines (e.g. DeepVariant, VEP) run on own GPU nodes and keep full control over pipeline versions.
• Public (university hospitals, research institutes): Path C. Patient genomes should stay on own or national infrastructure, which procurement and data-protection rules favour.

Path A — Buy (recommended for most labs)

• Clinical interpretation platforms: QIAGEN Clinical Insight / QCI Interpret (QIAGEN is EU-headquartered in the Netherlands/Germany), Illumina Connected Software / DRAGEN and Emedgene, Franklin by Genoox, VarSome Clinical (Saphetor, Switzerland), Nexus/Alissa Interpret (Agilent), SOPHiA GENETICS (Swiss/EU, DDM platform), Fabric Genomics, Velsera (formerly Seven Bridges/Pierian).
• Sequencer-integrated secondary analysis: Illumina DRAGEN (on-instrument, cloud or on-prem FPGA), NVIDIA Parabricks (GPU), Oxford Nanopore with Dorado/EPI2ME.
• Choose A when: you need a validated, supportable workflow for diagnostics, with audit trail and database curation included, and no in-house bioinformatics team.

Path B — Build on cloud APIs

• Predictor APIs: AlphaGenome API (Google DeepMind, non-commercial research use at launch; check current terms), AlphaMissense precomputed scores (CC BY), Ensembl VEP REST service for annotation [7][8].
• LLM APIs: reasoning models (Claude, Gemini, GPT, Mistral as EU provider) for literature synthesis and report drafting, with structured outputs and retrieval over licensed or open corpora (PubMed, ClinVar, PMC OA).
• Cloud genomics: AWS HealthOmics, Google Cloud Batch/Life Sciences successors, Azure batch workflows, in EU regions; Nextflow/nf-core pipelines on top. Verify DPA, EU residency and processor sub-contractors, since genome data is special-category data.
• Choose B when: research setting, de-identified or consented data, and want flexible custom logic.

Path C — Self-host (open source)

• Calling: DeepVariant, Clair3, GATK, Sentieon (commercial, self-hosted), nf-core/sarek and nf-core/raredisease pipelines [5][9].
• Annotation/prioritisation: Ensembl VEP (offline cache), ANNOVAR, Exomiser, SpliceAI, open ESM models, AlphaGenome open weights where licence permits, OpenCRAVAT.
• LLM layer: open-weight models (Llama, Qwen, Mistral, domain models such as BioMistral) via vLLM with a local PubMed/ClinVar index (vector DB plus BM25).
• Hardware: HW3 (one 48-141 GB GPU) for GPU-accelerated calling and local LLM; the pipeline is mostly CPU and storage-bound. Raw data reaches hundreds of GB per genome, so storage and backup planning dominate.
• Choose C when: data may not leave the hospital or institute, and a bioinformatics team exists.

Path D — Custom training

Fine-tune callers on lab-specific chemistry or ancestries, train variant-effect models on assay data (MAVEs, saturation mutagenesis) for a specific gene panel, or calibrate predictors on in-house curated variants. Needs well-curated, ancestry-diverse labelled variants; predictor bias towards European-ancestry data is a documented issue.

State of the art (2026)

• Calling: deep-learning callers reach very high accuracy on SNVs and indels against GIAB truth sets; the frontier is structural variants, tandem repeats and long-read pangenome-based calling [5][6].
• Missense effect: AlphaMissense released predictions for all possible human missense variants and classified a majority as likely benign or likely pathogenic, with the remainder ambiguous [3]. Protein language models and structure-aware models are the standard components of ensembles.
• Non-coding: AlphaGenome (DeepMind, 2025) predicts many regulatory modalities from up to one megabase of DNA context, a step towards scoring non-coding variants; clinical validation is still limited [4].
• Agents: LLM agents increasingly orchestrate annotation tools and literature search for variant curation. They help triage but hallucinated citations and over-confident classification are known failure modes, so outputs need citation verification.
• Limits: predictors are trained on ClinVar-like labels (circularity, ancestry bias); they cannot replace functional assays, segregation or de novo evidence; VUS rates are higher in under-represented ancestries; gene-disease validity must be checked separately (ClinGen).

Shared resources & combinations

• Scientific foundation models (scientific-fm) on GPU: SCI-001 protein structure models and SCI-021 sequence-design models share the same protein/DNA model serving stack (HW3).
• Multi-omics and targets: SCI-017 reuses annotated variant and expression data for target/biomarker work.
• Clinical context: HLTH-004 clinical decision support, HLTH-028 terminology mapping (HPO, SNOMED, FHIR Genomics), KNOW-010 guideline assistant for literature RAG.

Effort, cost & KPIs

• Path A: weeks to months for validation and accreditation; licence per sample or per seat.
• Path B: 3-6 months for a research pipeline; cost is dominated by compute and storage, not LLM calls.
• Path C: 6-12 months with a bioinformatics team; clinical use adds validation under ISO 15189.
• KPIs: analyst minutes per case, diagnostic yield, VUS rate and reclassification rate, concordance against GIAB/in-house truth sets, turnaround time sample to report, share of LLM citations verified correct, re-analysis coverage.

People & change

Involve management and the lab director first (budget, accreditation), then the department lead, geneticists as end users, IT and IT security, and the data protection officer (DPO) from the start: genetic data is a special category under Art. 9 GDPR, so a DPIA (Art. 35 GDPR), processor agreements (Art. 28) and a clear legal basis are required (GUIDE-005). Legal/compliance checks IVDR and accreditation duties.

Works council: depends. Where a works council exists, the system is mainly used on samples, not staff, but classification logs and sign-off statistics can be suited to monitor staff performance; under the Federal Labour Court's reading of §87(1) no. 6 BetrVG objective suitability suffices, and §90(1) no. 3 requires timely information. In public hospitals the staff council applies. See GUIDE-004.

AI literacy: Art. 4 AI Act requires measures that support staff's AI literacy; train geneticists on predictor limits and failure modes (GUIDE-006). Pilot with one gene panel and two experienced geneticists, compare with the current workflow, then widen. Typical concern: deskilling or liability. Address it by keeping every classification a proposal that a qualified geneticist signs off, with evidence shown per criterion (GUIDE-008).

Risks, regulation & compliance

• EU AI Act: depends. Software with a medical purpose that supports diagnosis is typically a medical device (MDR/IVDR); AI-based devices that need notified-body conformity assessment are high-risk under Art. 6(1) and Annex I. Pure research use is largely out of scope (Art. 2(6)); research-only labelling must be genuine. Under the Digital Omnibus on AI (Reg. (EU) 2026/1744), high-risk obligations for Art. 6(1)/Annex I products apply from 2 August 2028 (Annex III systems: 2 December 2027) [10].
• IVDR: in-house tests in health institutions fall under Art. 5(5) conditions; commercial interpretation software generally needs CE marking.
• GDPR: genetic data is special-category data (Art. 9); a genome is inherently identifying and also reveals relatives' information. Needs a legal basis, DPIA (Art. 35), minimisation, EU residency and strict access control; re-identification of "anonymised" genomes is realistic. Germany's Gendiagnostikgesetz (GenDG) requires consent and genetic counselling for diagnostic testing [11].
• Mitigations: human sign-off, versioned databases, locked pipeline versions, audit logs, benchmarks on GIAB, ancestry-stratified evaluation, citation verification for LLM output, no commercial LLM API without a DPA and EU processing.

Getting started

1. Define the use (research vs. diagnostic) because it decides IVDR, GenDG and AI Act obligations.
2. Benchmark a calling pipeline (DRAGEN, DeepVariant via nf-core/sarek) on GIAB reference samples.
3. Add Ensembl VEP annotation with ClinVar, gnomAD and one missense predictor; compare with 50 solved in-house cases.
4. Pilot an LLM evidence summariser over PubMed/ClinVar with enforced citations; measure error rate with geneticists.
5. Complete DPIA, consent and data-residency review before any patient data touches a cloud service.

Sources

1. Richards et al. 2015, ACMG/AMP standards for sequence variant interpretation — https://pubmed.ncbi.nlm.nih.gov/25741868/
2. Pejaver et al. 2022, ClinGen calibration of computational predictors — https://pubmed.ncbi.nlm.nih.gov/36413997/
3. Cheng et al. 2023, AlphaMissense, Science — https://www.science.org/doi/10.1126/science.adg7492
4. Google DeepMind, AlphaGenome — https://deepmind.google/discover/blog/alphagenome-ai-for-better-understanding-the-genome/
5. DeepVariant — https://github.com/google/deepvariant
6. Illumina DRAGEN — https://www.illumina.com/products/by-type/informatics-products/dragen-secondary-analysis.html
7. Ensembl Variant Effect Predictor — https://www.ensembl.org/info/docs/tools/vep/index.html
8. ClinVar (NCBI) — https://www.ncbi.nlm.nih.gov/clinvar/
9. nf-core/sarek — https://nf-co.re/sarek
10. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj ; amended by Regulation (EU) 2026/1744 (Digital Omnibus on AI, OJ 24 Jul 2026) — https://eur-lex.europa.eu/eli/reg/2026/1744/oj
11. Regulation (EU) 2016/679 (GDPR), Art. 4(13), 9 — https://eur-lex.europa.eu/eli/reg/2016/679/oj