From variant to clinical research report in seconds: AI-Assisted genomic variant classification with Amazon Quick

The problem: Variant classification is slow and manual

You have a list of genomic variants from a sequencing run. Each one requires a multitude of steps:

1. Looking up the functional consequence in Ensembl VEP.

2. Checking the population frequency in gnomAD.

3. Pulling gene constraint scores (pLI, LOEUF).

4. Manually evaluating each ACMG/AMP criterion.

5. Applying combining rules to reach a 5-tier classification.

6. Writing up the interpretation with supporting evidence.

For a single variant, the process can take 15–30 minutes. For a panel of 50 variants, it's a full day of manual curation—and that's before you write the clinical report.

What if you could type one sentence and get the variant classification back in seconds?

The result: Ask a question, get a clinical report

To view the complete open-source solution, see Variant Classification API on the GitHub website.

With this solution deployed, you open Quick and type:

"Classify variant 1-55039870-G-A"

Seconds later, Quick returns a structured clinical report. No manual lookups. No spreadsheets. No copying and pasting between gnomAD, VEP, and your notes. The entire ACMG evaluation, clinical interpretation, and literature search happens automatically.

Example report:

🟡 VUS — PCSK9 | 1-55039870-G-A

Metric	Value	Interpretation
gnomAD r4 AF	7.09 × 10⁻⁷	Ultra-rare
CADD Phred	36.0	Highly deleterious
SpliceAI Max	0.0	No splice impact
pLI	~0.0	Gene NOT constrained
LOEUF	1.144	Tolerant of LoF

.

gnomAD r4 AF	7.09 × 10⁻⁷	Ultra-rare
CADD Phred	36.0	Highly deleterious
SpliceAI Max	0.0	No splice impact
pLI	~0.0	Gene NOT constrained
LOEUF	1.144	Tolerant of LoF

.

CADD Phred	36.0	Highly deleterious
SpliceAI Max	0.0	No splice impact
pLI	~0.0	Gene NOT constrained
LOEUF	1.144	Tolerant of LoF

.

SpliceAI Max	0.0	No splice impact
pLI	~0.0	Gene NOT constrained
LOEUF	1.144	Tolerant of LoF

.

pLI	~0.0	Gene NOT constrained
LOEUF	1.144	Tolerant of LoF

.

LOEUF	1.144	Tolerant of LoF

ACMG Criteria Evaluated:

Criterion	Strength	Met?	Reason
PVS1	Very Strong	❌	Gene not constrained (pLI≈0, LOEUF=1.14)
PM2	Moderate	✅	Ultra-rare: AF=7.09×10⁻⁷
PP3	Supporting	✅	CADD phred=36.0 (≥20)
BA1	Stand-alone	❌	AF does not exceed 5%
BS1	Strong	❌	AF does not meet threshold
...			(all 10 criteria shown)

.

PVS1	Very Strong	❌	Gene not constrained (pLI≈0, LOEUF=1.14)
PM2	Moderate	✅	Ultra-rare: AF=7.09×10⁻⁷
PP3	Supporting	✅	CADD phred=36.0 (≥20)
BA1	Stand-alone	❌	AF does not exceed 5%
BS1	Strong	❌	AF does not meet threshold
...			(all 10 criteria shown)

.

PVS1	Very Strong	❌	Gene not constrained (pLI≈0, LOEUF=1.14)
PM2	Moderate	✅	Ultra-rare: AF=7.09×10⁻⁷
PP3	Supporting	✅	CADD phred=36.0 (≥20)
BA1	Stand-alone	❌	AF does not exceed 5%
BS1	Strong	❌	AF does not meet threshold
...			(all 10 criteria shown)

.

PM2	Moderate	✅	Ultra-rare: AF=7.09×10⁻⁷
PP3	Supporting	✅	CADD phred=36.0 (≥20)
BA1	Stand-alone	❌	AF does not exceed 5%
BS1	Strong	❌	AF does not meet threshold
...			(all 10 criteria shown)

.

PP3	Supporting	✅	CADD phred=36.0 (≥20)
BA1	Stand-alone	❌	AF does not exceed 5%
BS1	Strong	❌	AF does not meet threshold
...			(all 10 criteria shown)

.

BA1	Stand-alone	❌	AF does not exceed 5%
BS1	Strong	❌	AF does not meet threshold
...			(all 10 criteria shown)

.

BS1	Strong	❌	AF does not meet threshold
...			(all 10 criteria shown)

.

...			(all 10 criteria shown)

Why VUS? This is a stop-gained (nonsense) variant—normally strong pathogenic evidence. But PCSK9 is not constrained against loss-of-function (pLI≈0), so PVS1 cannot be applied. With only PM2 (Moderate) + PP3 (Supporting), the evidence is insufficient for "Likely Pathogenic."

Clinical Context: PCSK9 loss-of-function variants are well-established as protective against cardiovascular disease, associated with reduced LDL cholesterol and up to 88% reduction in coronary events (Cohen et al., 2006).

Relevant Literature:

1. Cohen et al. — Sequence variations in PCSK9, low LDL, and protection against CHD (PubMed)
2. PCSK9: From Nature's Loss to Patient's Gain (2024) (AHA Journals)

⚠️ This classification is computational and based on 10 of 28 ACMG-AMP criteria. Clinical interpretation requires review by a qualified geneticist.

How it works: the science

Behind the scenes, this solution implements the ACMG/AMP 2015 guidelines (Richards et al., Genetics in Medicine), the standard framework that's used worldwide for clinical variant interpretation. The framework defines 28 criteria. The solution automates the 10 criteria that can be computationally evaluated from two public data sources.

Data sources

Source	What it provides	What it’s used for
Ensembl VEP (European Bioinformatics Institute)	Functional consequence, CADD score, SpliceAI splice predictions, exon position, NMD flags	PVS1, PM4, PP2, PP3, BP1, BP4, BP7
gnomAD r4 (Broad Institute)	Population allele frequency across 730k+ individuals, gene constraint scores (pLI, LOEUF, oe_mis, mis_z)	PVS1, PM2, BA1, BS1

The 10 automated ACMG criteria

Criterion	Strength	What it asks	How it's answered
PVS1	Very Strong (Pathogenic)	Is this a null variant in a gene intolerant to loss-of-function (LoF)?	Stop-gained/frameshift + pLI ≥ 0.9 or LOEUF ≤ 0.35
PM2	Moderate (Pathogenic)	Is this variant absent or ultra-rare in the population?	gnomAD AF < 0.01%
PM4	Moderate (Pathogenic)	Does this change protein length?	In-frame indel or stop-loss
PP2	Supporting (Pathogenic)	Is this a missense in a gene that doesn't tolerate missense?	mis_z ≥ 3.09 or oe_mis ≤ 0.8
PP3	Supporting (Pathogenic)	Do computational tools predict damage?	CADD phred ≥ 20 or SpliceAI ≥ 0.2
BA1	Stand-alone (Benign)	Is this too common to cause disease?	gnomAD AF > 5%
BS1	Strong (Benign)	Is the frequency higher than expected?	gnomAD AF between 0.01% and 1%
BP1	Supporting (Benign)	Is this a missense in a gene where only LoF causes disease?	Missense + high pLI + tolerant oe_mis
BP4	Supporting (Benign)	Do computational tools predict benign?	CADD phred < 10
BP7	Supporting (Benign)	Is this a synonymous variant with no splice impact?	Synonymous + SpliceAI < 0.1

Combining rules

After the criteria are evaluated, the ACMG combining rules determine the final classification:

Classification	Required Evidence
Pathogenic	1 Very Strong + ≥1 Strong, or 1 Very Strong + ≥2 Moderate, or ≥2 Strong
Likely Pathogenic	1 Very Strong + 1 Moderate, or 1 Strong + 1–2 Moderate, or ≥3 Moderate
Benign	1 Stand-alone, or ≥2 Strong benign
Likely Benign	1 Strong + 1 Supporting, or ≥2 Supporting benign
VUS	Everything else — insufficient evidence in either direction

What's NOT automated and why

The remaining 18 ACMG criteria require data can't be computationally derived for the following reasons:

• Functional studies (PS3/BS3) require wet-lab experimental results.

• Segregation (PP1/BS4) requires family pedigree data.

• De novo status (PS2/PM6) requires parental sequencing.

• Literature evidence (PP4) requires expert curation.

• ClinVar cross-references (PS1/PM5/PP5/BP6) are possible to automate in future versions.

The tool is designed for triage and prioritization. It identifies which variants need human expert review, not which variants are definitively pathogenic.

How it works: the technology

Architecture

What each layer does

Layer	Role	Key point
Amazon Quick	Provides natural-language interface, orchestrates workflow, searches literature, formats report	The LLM never makes biological predictions. It only interprets structured results.
AWS Lambda	Hosts the classification API, fetches data, evaluates ACMG criteria deterministically	Same variant always produces same classification. It's fully reproducible.
Ensembl VEP	Predicts functional consequence of a variant on genes and proteins	Public API, no credentials needed.
gnomAD r4	Provides population allele frequency and gene constraint metrics	Public API, most current release.

The primary design principle

All classification logic is deterministic. The ACMG criteria engine uses fixed thresholds and combining rules with no AI inference. The role of Quick is purely interpretive, translating between human language and structured API calls. This results in the following outcomes:

• The same variant always gets the same classification.

• Results are auditable. Every criterion shows exactly why it was or wasn't met.

• The AI can't hallucinate a classification. It can only report what the engine computed.

Deploy the solution yourself

Before you deploy the solution, make sure that your environment meets the solution’s prerequisites. For a list of prerequisites, see Setup on the GitHub website.

To deploy the solution, complete the following steps:

1. Run the following command to clone the repository:
   git clone https://github.com/aws-samples/sample-genomic-variant-classification-serverless.git
cd sample-genomic-variant-classification-serverless
2. Run the following command to set up the Python environment:
   python3 -m venv .venv && source .venv/bin/activate
pip install -r lambda/requirements.txt
3. Run the following command to deploy to AWS:
   cp samconfig.example.toml samconfig.toml
# edit region if needed
sam build
sam deploy

The entire solution deploys in under 5 minutes. AWS Serverless Application Model (AWS SAM) creates the Lambda function, Amazon API Gateway resource, and AWS Identity and Access Management (IAM) role. The API is live and ready to classify variants.

To immediately test the solution, run the following commands:

API_URL=\$(aws cloudformation describe-stacks \  
  --stack-name variant-classification-v2 --region us-east-1 \  
  --query "Stacks\[0\].Outputs\[?OutputKey=='ApiUrl'\].OutputValue" --output text)  
    
curl -s -X POST "\$API_URL/classify" \  
  -H "Content-Type: application/json" \  
  -d '{"variant_id": "1-55039870-G-A"}' \| python3 -m json.tool

For more information, see Setup on the GitHub website.

For information about costs, see AWS Lambda pricing and Amazon API Gateway pricing.

What Quick brings to this solution

Quick transforms a JSON API into a tool that researchers actually want to use. Without Quick, you'd need to write curl commands or build a custom frontend.

With Quick, you get the following capabilities:

Capability	What it does
Custom Action Connectors	Registers the classification API so that the AI agent can call it. You upload the OpenAPI spec and set the base URL.
Quick Flows	Parses the input, calls the API, searches the literature, and generates report.
Chat Agents	Interprets the 10-criteria ACMG evaluation into plain English, explains why criteria were or weren't met, and adds clinical context.
Web Search	Automatically finds relevant PubMed papers about the gene/variant and includes links in the report.
Formatted Reports	Produces publication-ready output with tables, evidence summaries, and literature citations.

Set up the Quick integration

After you deploy the solution, complete the following steps to connect the API to Quick:

1. Register the connector: Upload the openapi.json from the repository as a custom action connector. Set the base URL to your deployed API Gateway endpoint.
2. Create a Quick Flow: The repository README includes a complete flow definition that you can enter directly into the Quick console. It sets up the full pipeline: from input, route, and classify to literature search, report, and follow-up options.
3. Run the API: Type a variant ID and get back a full clinical report with ACMG criteria, evidence, and literature in seconds.

Batch classification

For larger panels, run the following command to classify up to 20 variants in a single request:

curl -s -X POST "\$API_URL/classify/batch" \  
  -H "Content-Type: application/json" \  
  -d '{"variants": [
    {"variant_id": "1-55039870-G-A"},
    {"variant_id": "17-43045682-T-C"},
    {"variant_id": "1-55039774-C-T"}
  ]}' | python3 -m json.tool

Note: Replace the values for variant_id with the variant IDs.

Or, enter the following prompt in Quick: "Classify these three variants: 1-55039870-G-A, 17-43045682-T-C, 1-55039774-C-T". The Quick flow automatically uses the batch endpoint and generates a summary table with variant-level detail blocks.

Limitations

Limitation	Implication
10 of 28 ACMG criteria automated	Functional studies, segregation, de novo status, and ClinVar require manual review.
Research-grade only	Not validated for clinical diagnostic reporting because it requires geneticist review.
Public API rate limits	VEP and gnomAD might throttle under heavy load. The tool retries automatically, but might time out on large batches.
GRCh38 default	For GRCh37 coordinates, specify dataset: "gnomad_r2_1".

What's next

The following features are next for the tool:

• ClinVar integration: Automatically cross-reference known pathogenic or benign classifications.

• VCF file upload: Accept a VCF, classify all variants, return annotated results.

• Local VEP deployment: Deploy a containerized VEP for high throughput without rate limits.

• AWS Step Functions: Orchestrate classification of thousands of variants in parallel.

Conclusion

Variant classification doesn't have to be a manual, time-consuming process. By combining a deterministic ACMG criteria engine with the natural-language interface of Quick, researchers can go from a variant ID to a fully interpreted clinical report—with evidence, literature, and next steps—in seconds rather than minutes.

The classification logic is transparent, auditable, and reproducible. The AI adds interpretation and context without making biological predictions. And it all runs serverless on AWS.

Get started: Variant classifications API on the GitHub website.

References

• Richards S, et al. "Standards and guidelines for the interpretation of sequence variants." Genetics in Medicine. 2015;17(5):405-424. doi:10.1038/gim.2015.30

• Kircher M, et al. "A general framework for estimating the relative pathogenicity of human genetic variants." Nature Genetics. 2014;46(3):310-315.

• Jaganathan K, et al. "Predicting splicing from primary sequence with deep learning." Cell. 2019;176(3):535-548.

• Karczewski KJ, et al. "The mutational constraint spectrum quantified from variation in 141,456 humans." Nature. 2020;581:434-443.

• Cohen JC, et al. "Sequence variations in PCSK9, low LDL, and protection against coronary heart disease." New England Journal of Medicine. 2006;354(12):1264-1272.

• Amazon Quick User Guide.