pharm_variants.csv

One row is (variant, drug, genotype) → response, at a PharmGKB evidence level. It is a distinct table rather than columns on variants.csv so that a module carrying no drug annotations carries no drug columns — one CSV, one concern.

genotype is part of the identity, not decoration. PharmGKB publishes a clinical annotation per genotype: a summary names variant and drug, and the child rows give one annotation per call, usually three. They are not variations on one finding — they are distinct and sometimes opposed ones, so collapsing them to the variant loses the direction of the effect.

evidence_level is PharmGKB's axis and recommendation_strength is CPIC's. They are not two spellings of confidence and must not be filled from one another.

Identity

Row model	PharmVariantRow (just_dna_format.pgx)
Becomes	pharm_variants.parquet — lead parquet: pharm_variants.parquet
Authored or derived	authored — a person writes it (a drafter may append rows)
Draftable	yes — draft can append rows
Drafted by	clinpgx (clinpgx_draft, projection, matched on rsid, chrom, start, ref, drug, genotype, phenotype_category, annotation_id)
Natural key	variant_key, drug, genotype, phenotype_category, annotation_id
Fact signature	no
In the attestation binding	yes — manifest.inputs[]
Requires at least one of	rsid or chrom+start

Columns

Column	Type	Required	Values	Meaning
rsid	str | None	optional		dbSNP id of the variant, e.g. rs9923231
chrom	str | None	optional		Chromosome (position-only variants)
start	int | None	optional		1-based position, VCF POS convention (position-only) — CPIC/PharmVar publish this convention and it is stored as-is; do not subtract one
ref	str | None	optional		Reference allele (position-only)
gene	str | None	optional		Gene symbol, e.g. VKORC1
requires_callable	bool | None	optional		True when a consumer must prove this position was callable before concluding that the sample carries this row's genotype — the reference-homozygote row is the case, since a variant-only callset emits no record for it and absence is not the call. False records that no such proof is needed, which is the ordinary case for a genotype carrying an alternate allele. Empty says nothing either way, and is not False.
genotype	str | None	optional		Genotype the response applies to, canonical sorted form, e.g. C/T, or a single allele where the contig is hemizygous or haploid. An allele is bases, or a symbolic/structural allele carrying its length (/C). A pipe (C|T) is also accepted and records that the call was phased; with no phase-set column the order names no homolog, so it is phase recorded but unaddressable.
drug	str	required		Drug the response annotation is about, e.g. warfarin
phenotype_category	str | None	optional	one of: dosage, efficacy, metabolism_pk, other, pd, toxicity	What kind of effect this is (VALID_PHENOTYPE_CATEGORIES; multi-valued via [,;|]). Part of the identity — one variant+drug carries separate efficacy, toxicity and metabolism annotations.
annotation_id	str | None	optional		The source's own accession for this annotation, e.g. a PharmGKB clinical-annotation id. Optional, and the tie-break of last resort in the duplicate key — like PgsRow.pgs_id, a source accession is a legitimate identity for a curated record.
response	str | None	optional		Drug response / phenotype, free-form (e.g. 'reduced dose requirement')
evidence_level	str | None	optional	one of: 1A, 1B, 2A, 2B, 3, 4	PharmGKB clinical-annotation evidence level (1A..4)
pmid	str | None	optional		Optional PubMed id grounding THIS row's claim — the literature behind this drug/genotype response, which studies.csv cannot express because a study row attaches to the whole variant. Free-form like StudyRow.pmid (9545397, [PMID: 9545397], a ;-joined list). The pharm row cites; studies.csv and literature.csv describe. Separate from evidence_level, which grades evidence rather than pointing at it.
trait_efo_id	str | None	optional		Optional trait ontology id(s), for cross-module join
conclusion	str	required		Human-readable interpretation

Generated from the row model at build time — reference.authoring_reference(), the same answer describe_table gives an authoring tool. Nothing on this page is hand-kept.