MyVariant.info Queries - Usage Guide

Overview

Query the BioThings myvariant.info aggregator for ClinVar, gnomAD, dbSNP, dbNSFP (which contains AlphaMissense, REVEL, CADD, SpliceAI, BayesDel, and 40+ predictors), COSMIC, CIViC, and other clinical variant annotations in a single batched API call. The skill addresses dbNSFP version drift (the dominant staleness vector for in-silico predictors), Pejaver 2022 calibrated PP3/BP4 thresholds, and the reproducibility pattern of recording _meta.src versions alongside results.

Prerequisites

pip install myvariant requests pandas

Quick Start

Tell the agent what to do: - "Annotate these 500 variants with ClinVar + gnomAD grpmax FAF95 + AlphaMissense + REVEL + SpliceAI in one batch" - "Search myvariant for all ClinVar Pathogenic variants in BRCA1 with star >= 2" - "Record _meta.src versions for each source alongside my results for reproducibility" - "Find variants in chr17:43000000-44000000 with CADD phred > 25 and ClinVar P/LP" - "Batch convert these 200 HGVS strings to canonical IDs and pull aggregated annotations"

Example Prompts

Batch Annotation

"For these 100 HGVS-g variants, pull ClinVar significance, ClinVar review status, gnomAD v4 grpmax FAF95, AlphaMissense score, REVEL score, CADD phred, SpliceAI DS_max, and COSMIC IDs. Return as a pandas DataFrame and record source versions."

"Annotate my 50k-variant cohort with ClinVar + gnomAD in batches of 1000 with rate-limit-safe sleeping; merge results into the input table."

Lucene Search

"Search myvariant for all variants in TP53 with ClinVar Pathogenic AND CADD phred > 25; return top 500."

"Find variants in chr17:43044294-43125483 (BRCA1 region) with AlphaMissense score > 0.564 OR REVEL > 0.7."

"Search for ClinVar VCEP-curated variants (review_status:reviewed_by_expert_panel) in cardiac sarcomere genes."

Reproducibility

"Run my annotation pipeline and record per-source versions in the output JSON sidecar: dbnsfp.version, clinvar.version, gnomad_exome.version, cadd.version."

"Compare my last analysis (dbNSFP v4.5) to today's results; report which AlphaMissense scores changed due to dbNSFP version drift."

Single-Source Deep Dive

"Pull all dbNSFP fields for this variant: REVEL, BayesDel, AlphaMissense, EVE, ESM1b, PrimateAI-3D, MetaRNN, VEST4, MutPred. Apply Pejaver 2022 PP3/BP4 calibration to REVEL." (Defer Pejaver calibration logic to acmg-classification.)

"For 100 rare missense variants, batch-query myvariant for all in-silico predictors and report the top-3 most concordant tools."

What the Agent Will Do

1. Initialize MyVariantInfo() client; canonicalize input IDs (HGVS preferred over rsID for unambiguous matching).
2. For < 1000 variants: single getvariants(list, fields=...) call with explicit field selection.
3. For > 1000: chunk to 1000 + sleep 0.5s between chunks.
4. Always include _meta in fields for version tracking.
5. Use Lucene query for structured search (gene + significance + region + score thresholds).
6. Parse nested response defensively (.get('source', {}) rather than direct key access).
7. Return DataFrame + per-source version dict; emit reproducibility sidecar.
8. Defer ACMG classification logic to clinical-databases/acmg-classification.

Tips

• Use HGVS-g (e.g., chr7:g.140453136A>T) for unambiguous matching; rsID is multi-allelic for ~6-8% of variants.
• _meta.src.<source>.version is the canonical version field; record alongside results for reproducibility.
• dbNSFP version lags primary tool releases by 6-18 months; AlphaMissense entered dbNSFP 4.4 (~2024). For cutting-edge predictions query the source API directly.
• AlphaMissense developer threshold (0.564) is NOT the Pejaver 2022 PP3-calibrated threshold; do not auto-apply as PP3 evidence. ClinGen has not endorsed AlphaMissense PP3 strength as of May 2026.
• Do not stack REVEL + BayesDel + VEST4 as independent PP3 evidence; they share training data per Pejaver 2022.
• Batch limit is 1000 IDs per POST; exceeding silently truncates.
• Field path errors return None, not errors; check structure with print(mv.getvariant(test_id)) first.
• myvariant does NOT produce ACMG calls; pair with acmg-classification for classification.
• For PHI-sensitive workflows, switch to OpenCRAVAT (local install); myvariant requires HTTP.
• Sample overlap exists across aggregated sources; ClinVar variants often derive from gnomAD individuals; treat as not statistically independent.

Related Skills

• clinical-databases/clinvar-lookup - Source-level ClinVar deep queries
• clinical-databases/gnomad-frequencies - Source-level gnomAD deep queries
• clinical-databases/dbsnp-queries - Source-level rsID resolution and merge-chain
• clinical-databases/acmg-classification - ACMG with Pejaver calibrated thresholds
• clinical-databases/variant-prioritization - Pipeline using aggregated annotations