prs-sample-scores

Published individual PRS and catalog evidence for public genomes, used by just-prs. Dataset: just-dna-seq/prs-sample-scores.

Published at: 2026-08-19T13:20:54.904786+00:00 Parent runtime/evidence revision: 1248dc9223f131c953af8d90a08e2472685d8fec

Current coverage: 7 public genomes, 5337 PGS IDs, 62622 analysis-eligible runtime rows, 11816 trait summaries.

How to read this dataset

What to look at: ANALYSIS.md is the analysis hint — trait cards, model disagreement, restoration deltas, five-population percentiles, and family concordance. Start there if you want queries, not schemas.

Then:

1. data/manifest.json — pinned revisions, hashes, grains, and exclusion rules
2. data/trait_summaries.parquet — ancestry-selected usable-scope medians
3. data/actionability.parquet and data/trait_contexts.parquet — evidence, not scores
4. data/model_analysis.parquet — drill-down for one sample × PGS × profile

Raw runtime scores remain in data/runtime_results.parquet. Do not compare raw scores across PGS IDs. A percentile is relative to one 1000G superpopulation, not a disease probability.

File map

Runtime-owned (do not overwrite from evidence or integration jobs):

• data/samples.parquet
• data/sample_ancestry.parquet
• data/runtime_results.parquet
• data/runtime_manifest.json

Evidence-owned:

• data/traits.parquet
• data/score_trait_links.parquet
• data/papers.parquet
• data/score_paper_links.parquet
• data/guidelines.parquet
• data/guideline_trait_links.parquet
• data/actionability.parquet
• data/trait_contexts.parquet
• data/record_search_terms.parquet
• data/evidence_manifest.json

Integration-owned:

• data/model_analysis.parquet (schema 1)
• data/trait_summaries.parquet (schema 1)
• data/manifest.json (schema 1)
• README.md (this file)
• AGENTS.md
• ANALYSIS.md (what to look at)

Scoring profiles

Two published profiles, aggregated independently:

• grch38-wgs-pass-unrestored-v1 — observed variants plus only safely resolvable hom-ref absences
• grch38-wgs-pass-restored-v1 — additionally classifies absent WGS loci as homozygous reference using a pinned reference-allele universe

Restoration is not genotype imputation and does not infer alternate alleles from LD. Compare the two profiles only for the same sample and PGS ID after each side has been summarized on its own.

Ancestry and percentiles

Selected percentile population is the sample's broad 1000G superpopulation from sample_ancestry.parquet. In this release every validated public genome selects EUR at confidence 1.0; that is derived per sample, not hardcoded.

CEU and IBS are nearest 1000 Genomes reference cohorts, not ethnicity, nationality, or separate percentile populations. Display CEU as “Northern/Western European” and IBS as “Iberian/Spanish” with the IGSR cohort definition. Fine-cohort labels never drive percentiles, coherence, or risk.

model_analysis.population_metrics always has exactly five entries (AFR, AMR, EAS, EUR, SAS). Unavailable entries are null plus a reason, never a fabricated zero.

PGS IDs without a trustworthy selected-population distribution stay in model_analysis with analysis_eligible=false and never enter summaries. Examples: PGS000011, PGS000012, PGS000013, PGS000014, PGS000015, PGS000016, PGS000017, PGS000018, PGS000019, PGS000021, PGS000022, PGS000024.

Quality, risk, and heritability

• Match rate is the fraction of scoring variants used. Weight-mass coverage is the fraction of absolute model weight. Neither is ancestry.
• Usable models require ≥50% match (just_prs.trait_summary.is_usable_model).
• Absolute risk uses just_prs.absolute_risk.estimate_absolute_risk from a selected-population z-score plus pinned prevalence and OR/AUROC. Missing science is N/A / null, not zero. A risk ratio of 1 is the population average.
• h² is population-level SNP heritability: the fraction of trait variation statistically associated with genetic differences in a studied population. It is not an individual's “percent genetic,” causal fraction, or disease probability.
• Trait joins use ontology IDs (just_prs.ontology.normalize_trait_id plus pinned aliases). Never join prevalence or h² by label.
• not_assessed means evidence was not established, not “not actionable.”

Licenses and privacy

The combined dataset license is CC-BY-4.0. Named public genomes (Anton, Livia) have their own source licenses on samples.parquet. Anonymized o-* family samples are publication-allowed under the recorded consent basis only. Private ingest aliases are local-only and must not appear here. Unknown genomes are never uploaded.

Family rows are not independent. Parent/child concordance is expected from shared inheritance. This seven-person dataset cannot estimate heritability, penetrance, segregation, or clinical transmission. Do not rank a family winner.

Load with Polars

import polars as pl

manifest = pl.read_json("data/manifest.json")  # or json.load
analysis = pl.read_parquet("data/model_analysis.parquet")
summaries = pl.read_parquet("data/trait_summaries.parquet")
eligible = analysis.filter(pl.col("analysis_eligible"))
eur = (
    eligible.drop("source_revision")
    .explode("population_metrics")
    .unnest("population_metrics")
    .filter(pl.col("superpopulation") == "EUR")
)

Load with DuckDB

SELECT sample_id, trait_id, score_profile_id, median_pct, n_usable
FROM read_parquet('data/trait_summaries.parquet')
WHERE percentile_population = 'EUR'
  AND n_usable > 0
ORDER BY sample_id, trait_id;

Always filter analysis_eligible before treating a percentile or risk value as a lookup result. Quarantined PGS IDs remain in model_analysis for audit.

Glossary

• PRS — polygenic score for one PGS model; scale is model-specific
• PGS ID — PGS Catalog score identifier
• z-score — (score - reference mean) / std for one superpopulation
• percentile — Phi(z) × 100; relative standing, not probability
• match rate — fraction of scoring variants used
• weight coverage — fraction of absolute model weight used
• restoration — hom-ref fill of absent WGS loci, not imputation
• broad ancestry — AFR/AMR/EAS/EUR/SAS reference similarity
• nearest 1000G cohort — CEU/IBS-style fine label inside that panel
• absolute risk — lifetime/population probability from z + prevalence + effect
• risk ratio — absolute risk / population prevalence; 1 = average
• h² — population SNP heritability, not an individual genetic percent
• ontology mapping — EFO/MONDO/OBA/HP identifiers for the same or related trait
• not_assessed — evidence not established

Medical and privacy limits

These tables are research/education artifacts for citizen scientists. They are not a diagnosis, prescription, or clinical recommendation. Pharmacogenomic context is not prescribing advice. Cite the PGS Catalog page for each PGS ID (https://www.pgscatalog.org/score/{pgs_id}/).

model_analysis.parquet

Primary key: sample_id, pgs_id, scoring_build, score_profile_id, scoring_fingerprint

Column	Type	Notes
sample_id	str
pgs_id	str
scoring_build	str
score_profile_id	str
scoring_fingerprint	str
status	str
error	`str	None`
score	`float	None`
variants_matched	`int	None`
variants_total	`int	None`
match_rate	`float	None`
weight_mass_coverage	`float	None`
sample_genotype_sha256	str
source_revision	str
analysis_eligible	bool
exclusion_reasons	list
publication_allowed	bool
selected_superpopulation	`str	None`
selected_superpopulation_confidence	`float	None`
closest_cohort	`str	None`
closest_cohort_confidence	`float	None`
closest_cohort_label	`str	None`
catalog_name	`str	None`
trait_reported	`str	None`
quality_label	`str	None`
quality_key	`str	None`
is_harmonized	`bool	None`
auroc	`float	None`
or_per_sd	`float	None`
population_metrics	list
traits	list
papers	list

trait_summaries.parquet

Primary key: sample_id, trait_id, score_profile_id, percentile_population

Column	Type	Notes
sample_id	str
trait_id	str
trait_label	`str	None`
score_profile_id	str
percentile_population	str
model_scope	str
percentile_source	str
n_total	int
n_eligible	int
n_usable	int
n_percentile_available	int
n_quarantined	int
n_failed	int
n_excluded	int
median_pct	`float	None`
mean_pct	`float	None`
std_pct	`float	None`
min_pct	`float	None`
max_pct	`float	None`
spread	`float	None`
outliers	list
most_reliable_pgs_id	`str	None`
most_reliable_pct	`float	None`
absolute_risk	`str	None`
population_average	`str	None`
risk_vs_average	`str	None`
n_risk_models	int
heritability_text	`str	None`
heritability_detail	`str	None`
prs_actionability_status	str
condition_actionability_status	str
context_resolution_status	str
context_classes	list
paired_profile_id	`str	None`
paired_n_usable	`int	None`
paired_median_pct	`float	None`
delta_median_pct	`float	None`
caveats	list

Evidence tables

traits.parquet

Primary key: trait_id

Column	Type	Notes
trait_id	str
label	str
definition	`str	None`
synonyms_json	str
category	`str	None`
ontology_prefix	`str	None`
canonical_url	`str	None`
aliases_json	str
icd10_codes_json	str
mapping_source	str
mapping_confidence	str
mapping_status	str
retrieved_at	`str	None`
source_revision	`str	None`
Primary key: pgs_id, trait_id

Column	Type	Notes
pgs_id	str
trait_id	str
relationship_source	str
trait_reported	`str	None`
Primary key: paper_id

Column	Type	Notes
paper_id	str
pmid	`str	None`
pmcid	`str	None`
doi	`str	None`
pgp_id	`str	None`
title	`str	None`
authors	`str	None`
first_author	`str	None`
journal	`str	None`
date_publication	`str	None`
citation_text	`str	None`
pubmed_url	`str	None`
europepmc_url	`str	None`
doi_url	`str	None`
is_open_access	`bool	None`
content_license	`str	None`
abstract_text	`str	None`
abstract_status	str
resolution_status	str
Primary key: pgs_id, paper_id, relationship_type

Column	Type	Notes
pgs_id	str
paper_id	str
relationship_type	str
pgp_id	`str	None`
ppm_id	`str	None`
Primary key: guideline_id

Column	Type	Notes
guideline_id	str
organization	str
jurisdiction	`str	None`
title	str
url	str
version	`str	None`
publication_date	`str	None`
update_date	`str	None`
withdrawal_date	`str	None`
recommendation_direction	str
grade	`str	None`
action_type	`str	None`
target_condition	`str	None`
target_population	`str	None`
eligibility_criteria	`str	None`
required_context	`str	None`
recommendation_text	`str	None`
prs_specific	bool
source_license	`str	None`
obtainability	str
retrieval_url	str
retrieval_fingerprint	str
retrieved_at	str
is_current	bool
Primary key: guideline_id, trait_id

Column	Type	Notes
guideline_id	str
trait_id	str
mapping_basis	str
Primary key: trait_id, guideline_id

Column	Type	Notes
trait_id	str
guideline_id	str
prs_actionability_status	str
condition_actionability_status	str
context_resolution_status	str
supporting_guideline_ids_json	str
notes	`str	None`
Primary key: trait_id, context_class

Column	Type	Notes
trait_id	str
context_class	str
basis	str
why_interesting	str
high_percentile_means	`str	None`
research_only	bool
Primary key: trait_id, track, term_kind, term

Column	Type	Notes
trait_id	str
track	str
context_class	`str	None`
term	str
term_kind	str
provenance	str