---
pretty_name: PsychGNN Graph Artifact
license: cc-by-4.0
tags:
- psychiatry
- genomics
- graph
- graph-neural-network
- heterogeneous-graph
- research
---

# PsychGNN Graph Artifact

## Overview

This repository contains the graph data structure used by the PsychGNN models and related research artifacts.

The graph links psychiatric genetic variants to nearby genes and psychiatric disorder groups. It is a structured research object for model training and variant-level analysis, not a patient-level dataset or clinical knowledge graph.

## What is inside

Files currently published:

- `graph_artifact.pt`: serialized graph tensors used by training and the Space
- `metadata.json`: graph-level counts and configuration summary
- `split_counts.json`: train / validation / test positive-edge split counts
- `gene_coordinates_grch37.parquet`: published gene coordinate table used during graph construction

## Graph scope

This release is a GWS-filtered SNP/gene/disorder graph.

Published graph metadata:

- SNP nodes: **18,979**
- Gene nodes: **1,205**
- Disorder nodes: **11**
- SNP→Disorder positive edges: **22,687**
- SNP→Gene edges: **65,634**
- Disorder→Disorder edges: **0**
- GWS threshold: `p < 5e-8`
- SNP→Gene window: `±100 kb`

Disorder nodes in this graph:

- ADHD
- Anxiety
- Autism
- Bipolar disorder
- Borderline personality disorder (`bpd`)
- Eating disorders
- Major depressive disorder
- OCD
- PTSD
- Schizophrenia
- Substance use

## What this graph is and is not

This graph includes:

- GWS psychiatric SNP nodes
- positional SNP→Gene links
- disorder nodes

This graph does not include:

- a full pathway node layer
- gene→gene regulatory edges
- disorder→disorder edges in this release

## How it was derived

High-level construction procedure:

1. Start from the harmonized psychiatric GWAS summary-statistics dataset:
   - [lighteternal/pgc-psychiatric-gwas-harmonized](https://huggingface.co/datasets/lighteternal/pgc-psychiatric-gwas-harmonized)
2. Keep variants reaching genome-wide significance in at least one disorder.
3. Build SNP nodes and disorder nodes.
4. Map SNPs to nearby protein-coding genes using a positional window.
5. Serialize the resulting tensors for model training and Space inference.

## Intended use

This artifact is intended for:

- reproducing the current PsychGNN training run
- powering the public Space
- graph-level research analysis of psychiatric variant-disorder structure

It is not intended for:

- clinical interpretation
- independent causal claims
- general human-genome annotation outside the published graph scope

## Important caveats

- only GWS variants are included, so this is a filtered graph rather than a full summary-statistics graph
- disorder-disorder edges are a correlation proxy, not full LDSC estimates
- biological coverage is intentionally narrower than the long-term PRD
- downstream claims must match the published graph scope, not the aspirational architecture

## Relationship to the rest of the project

This graph artifact is the shared input for the public PsychGNN model repositories.

- built from: [lighteternal/pgc-psychiatric-gwas-harmonized](https://huggingface.co/datasets/lighteternal/pgc-psychiatric-gwas-harmonized)
- consumed by:
  - [lighteternal/psychgnn-masked-edge-imputation-model](https://huggingface.co/lighteternal/psychgnn-masked-edge-imputation-model)
  - [lighteternal/psychgnn-all-tested-pairs-model](https://huggingface.co/lighteternal/psychgnn-all-tested-pairs-model)