ACMG Variant Classification#

Use this skill when a user wants a structured ACMG/AMP-style interpretation workflow for a germline SNV/indel.

Interaction mode#

Default to a guided interview workflow.

When using this skill with a live user:

1. Ask for one block of information at a time
2. Wait for the user's answer before moving on
3. Do not request all evidence at once unless the user asks for a bulk template
4. Explicitly track what is known, unknown, and still needed
5. Treat phenotype, family history, segregation data, and parental genotypes as user-supplied inputs that may arrive incrementally

Recommended guided sequence:

1. Variant identity: gene, transcript, build, c.HGVS, p.HGVS, variant type
2. Clinical phenotype / suspected disease
3. Inheritance model and family structure
4. Parental genotype status and de novo / segregation details
5. Population / database / literature evidence
6. Functional and computational evidence
7. Criteria assignment and final review

At each step, summarize back in one compact block:

• confirmed facts
• missing facts
• provisional ACMG implications

Safety / scope#

Always say clearly:

• This is decision support, not a final clinical diagnosis.
• Gene/disease-specific ClinGen guidance overrides generic ACMG rules where applicable.
• Final classification requires expert manual review.

Inputs you should collect#

Use templates/intake.md and ask for or normalize these fields:

• Gene
• Transcript
• Genome build
• c.HGVS
• p.HGVS
• Variant type
• Zygosity
• Inheritance model
• Phenotype / disease context
• Population frequency evidence
• Functional evidence
• Segregation / de novo evidence
• Database assertions
• Literature evidence

If transcript, genome build, or HGVS is unclear, stop and ask for clarification before classification.

Standard workflow#

Step 1: Confirm scope#

Proceed only if all are true:

1. Variant is a germline small variant (SNV/indel)
2. Naming/build/transcript are defined
3. User understands output is review-only
4. Any gene-specific ACMG framework has been checked

Step 2: Normalize the record#

Create a clean variant record using templates/intake.md.

Step 3: Gather evidence by ACMG bucket#

Pathogenic side:

• PVS1
• PS1, PS2, PS3, PS4
• PM1, PM2, PM3, PM4, PM5, PM6
• PP1, PP2, PP3, PP4

Benign side:

• BA1
• BS1, BS2, BS3, BS4
• BP1, BP2, BP3, BP4, BP5, BP7

Step 4: Assign criteria carefully#

Use templates/evidence-table.md. For each criterion, record:

• code
• strength
• triggered yes/no
• reason
• source
• caveat / limitation

Do not double count overlapping evidence.

Step 5: Evaluate conflicts#

If both pathogenic and benign evidence exist:

1. Check whether evidence is truly independent
2. Downgrade/remove misapplied criteria if needed
3. If conflict remains unresolved, prefer VUS over forced certainty
4. State what additional data could resolve the conflict

Step 6: Apply combination logic#

Use scripts/classifier.py or reproduce its logic manually.

Pathogenic if any:

• 1 Very Strong + >=1 Strong

• 1 Very Strong + >=2 Moderate

• 1 Very Strong + 1 Moderate + 1 Supporting

• 1 Very Strong + >=2 Supporting

• =2 Strong

• 1 Strong + >=3 Moderate

• 1 Strong + 2 Moderate + >=2 Supporting

• 1 Strong + 1 Moderate + >=4 Supporting

• =3 Moderate + >=3 Supporting

Likely Pathogenic if any:

• 1 Very Strong + 1 Moderate

• 1 Strong + 1 to 2 Moderate

• 1 Strong + >=2 Supporting

• =3 Moderate

• 2 Moderate + >=2 Supporting

• 1 Moderate + >=4 Supporting

Benign if any:

• BA1

• =2 Strong benign criteria

Likely Benign if any:

• 1 Strong benign + 1 Supporting benign

• =2 Supporting benign

Else: VUS

Guided questioning pattern#

Use short, sequential prompts:

• Step A: ask only for variant identity fields
• Step B: ask only for phenotype and suspected diagnosis
• Step C: ask only for pedigree / family history / inheritance
• Step D: ask only for parental genotypes and segregation/de novo details
• Step E: ask only for outside evidence such as ClinVar, literature, frequency, and functional assays
• Step F: summarize triggered or candidate ACMG criteria before giving a provisional class

Included files#

• templates/intake.md
• templates/evidence-table.md
• references/sop.md
• references/test_cases.json
• scripts/classifier.py