lighteternal/pgc-psychiatric-gwas-harmonized

PGC Psychiatric GWAS Harmonized Dataset

Overview

This repository contains a harmonized release of public psychiatric GWAS summary statistics assembled from the OpenMed / Psychiatric Genomics Consortium collection.

It is intended as a research substrate for:

• heterogeneous graph construction
• cross-disorder variant modeling
• psychiatric association imputation
• temporal and provenance-aware evaluation
• secondary Mendelian-randomization analyses

The raw OpenMed/PGC collection is very large and heterogeneous. It contains multiple study waves, sub-analyses, and repeated source-specific representations of the same loci. This repository does not preserve that raw structure verbatim. Instead, it provides a reproducible harmonized release in two layers:

• data/: a primary modeling view
• analysis/: a broader provenance-preserving view

This is a research dataset. It is not intended for clinical interpretation or patient-level reporting.

Source datasets

Primary disorder inputs used in the main data/ view:

Disorder	Source dataset	Config
ADHD	OpenMed/pgc-adhd	adhd2022
Anxiety	OpenMed/pgc-anxiety	anx2026
Autism	OpenMed/pgc-autism	asd2019
Bipolar disorder	OpenMed/pgc-bipolar	bip2021_noUKBB
Borderline personality disorder	OpenMed/pgc-other	bpd2025
Eating disorders	OpenMed/pgc-eating-disorders	an2017
Major depressive disorder	OpenMed/pgc-mdd	mdd2018
Obsessive-compulsive disorder	OpenMed/pgc-ocd-tourette	ocd2025
Post-traumatic stress disorder	OpenMed/pgc-ptsd	ptsd2019
Schizophrenia	OpenMed/pgc-schizophrenia	scz2014
Substance use	OpenMed/pgc-substance-use	SUD2023

Additional retained sources in the analysis/ layer:

Disorder	Source dataset	Config	Role
ADHD	OpenMed/pgc-adhd	adhd2019	temporal
ADHD	OpenMed/pgc-adhd	adhd2018_SexSpecific	temporal
ADHD	OpenMed/pgc-adhd	adhd2010	temporal
Cross-disorder	OpenMed/pgc-cross-disorder	cdg2019	exploratory

Dataset structure

data/

Primary modeling view.

• one harmonized row per variant_key, disorder
• intended for graph construction, model training, and disorder-level statistical analysis
• includes 11 primary psychiatric disorder groups:
  • adhd
  • anxiety
  • autism
  • bipolar
  • bpd
  • eating_disorders
  • mdd
  • ocd
  • ptsd
  • schizophrenia
  • substance_use

Current aggregate counts:

• total rows: 102,795,695
• global unique variant_id: 28,039,986

analysis/

Provenance-preserving analysis view.

• retains multiple usable waves or releases where available
• preserves source_wave, source_config, and role
• includes exploratory resources such as cross_disorder
• intended for temporal validation, provenance auditing, and wave-to-wave replication analysis

Harmonization design

The release applies the following transformations:

1. Resolve source columns by observed non-null coverage rather than alias order alone.
2. Normalize effect-size fields to either beta or log_or.
3. Preserve source alleles, genomic location, p-values, and available quality fields.
4. Build an allele-aware canonical key for each variant.
5. Flip signed effects when necessary so rows sharing the same canonical variant key are aligned consistently.
6. Derive standard errors when the source does not report them directly.
7. Publish:
   • a primary deduplicated view in data/
   • a provenance-preserving multi-wave view in analysis/

This release uses an internally consistent allele-ordering strategy. It does not assert full external-reference REF/ALT resolution.

Column reference

Variant identity and genomic position

• variant_id
  • final pipeline identifier used downstream
  • equal to variant_key in this release
• variant_key
  • allele-aware canonical key
  • format: chr:pos:allele1:allele2
• snp_id
  • source SNP identifier when present
• raw_snp_id
  • source-study SNP identifier before any back-filling
• chr
  • chromosome label after harmonization
• pos
  • genomic coordinate after numeric coercion and positive-position filtering

Alleles and orientation

• a1
  • source-reported first allele column
• a2
  • source-reported second allele column
• effect_allele
  • canonical first allele after ordered representation
• other_allele
  • canonical second allele after ordered representation
• was_flipped
  • true if the sign of effect_size was inverted to align the row with the canonical allele ordering
• reference_resolved
  • whether alleles were resolved against an external reference genome
  • currently false in this release
• alignment_method
  • alignment strategy used for the row
  • currently allele_ordered

Association statistics

• effect_size
  • harmonized association effect
• effect_kind
  • beta for regression-style effects
  • log_or for odds ratios converted to log-odds
• se
  • standard error of the effect estimate
• se_source
  • reported if carried from source
  • derived if computed from effect_size and pval
• pval
  • source GWAS p-value
• sig_tier
  • coarse significance tier used downstream
  • gws if p < 5e-8
  • suggestive if 1e-5 > p >= 5e-8
  • null otherwise

Optional quality fields

• maf
  • allele-frequency field when available
• info
  • imputation or variant-quality metric when available

Provenance and intended use

• disorder
  • canonical disorder label used across the release
• source_repo
  • source Hugging Face dataset repository
• source_config
  • exact source study/config identifier
• source_wave
  • retained study wave or release label
• role
  • intended use of the row
  • primary, temporal, or exploratory
• source_file
  • compact provenance tag of the form source_repo:source_config
• tested_in_disorder
  • indicates that the variant-disorder pair was observed in the source study

Downstream graph suitability flags

• is_palindromic
  • true for allele pairs such as A/T or C/G
• is_biallelic_acgt
  • true for clean biallelic A/C/G/T rows
• usable_for_graph
  • graph-safety filter used by downstream graph construction

Temporal structure

Temporal and release-level structure is preserved in analysis/ whenever the upstream source is usable.

This supports:

• wave-to-wave replication checks
• temporal validation of downstream predictive models
• provenance-aware auditing

The data/ view remains compact and modeling-ready.

Not every historical release is guaranteed usable. If an upstream source does not expose a valid effect field in its current schema, that release is recorded as unavailable rather than synthesized into the harmonized table.

Quality control artifacts

The repository publishes:

• stats.json
• qc_manifest.json
• source_vs_harmonized_audit.json
• reports/harmonize_full_summary.json

These support:

• per-disorder row-count auditing
• duplicate classification
• missingness inspection
• effect-scale inspection
• allele-quality inspection
• source-vs-harmonized reconciliation

Intended use

This dataset is designed to support:

• cross-disorder psychiatric variant modeling
• heterogeneous graph construction over SNP, gene, and disorder nodes
• imputation of missing disorder association profiles
• temporal and provenance-aware evaluation
• Mendelian-randomization-style secondary analyses

This dataset is not designed for:

• patient-level prediction
• diagnostic use
• treatment recommendations

Practical guidance

• Use data/ for graph construction, model training, or primary disorder-level analysis.
• Use analysis/ when you need provenance, temporal comparisons, retained waves, or exploratory labels.
• Check se_source before assuming standard errors were reported directly.
• Check usable_for_graph before feeding rows into graph construction code.
• Treat cross_disorder as distinct from the 11 primary psychiatric disorder labels.

Limitations

• The canonical variant key is internally consistent but not a full external-reference REF/ALT assertion.
• Missingness in maf and info is largely source-driven.
• Some historical releases remain unusable if the upstream schema does not expose a valid effect field.
• The raw OpenMed/PGC collection is much larger than this harmonized release because it contains multiple releases, sub-analyses, and technical redundancies that are not all appropriate for a primary modeling table.