GraphLD Agent Context

GraphLD implements graphREML and graphREML-ST, LDGM-backed LD matrix operations, enrichment score tests, simulation, LD clumping, BLUP weights, and surrogate marker utilities for GWAS summary statistics.

Project Map

• src/graphld/: main Python package. Core modules include precision.py, io.py, heritability.py, cli.py, blup.py, clumping.py, simulate.py, and surrogates.py.
• src/score_test/: enrichment score-test package and CLI.
• tests/: pytest suite. Prefer small synthetic tests when tracked data fixtures are dirty.
• data/test/: intended small fixtures for LDGMs, sumstats, annotations, VCF, parquet, and score-test HDF5 files.
• data/ldgms/: default location for downloaded LDGMs. The usual metadata file is data/ldgms/metadata.csv; rows point to edgelist and snplist files.

Agent Skills

Use the project-local skills for operational detail:

• graphld-setup: installation, dependency, data-download, and O2 reference guidance.
• graphld-usage: CLI/API usage patterns.

CLI Surface

graphld subcommands are reml, blup, clump, simulate, and surrogates. estest handles enrichment score-test workflows with test, show, add-meta, rm, mv, and convert.

Terminology

• LDGM: linkage disequilibrium graphical model, stored as metadata plus edgelist/snplist files.
• graphREML: REML method for heritability partitioning and enrichment estimation.
• Score test: enrichment test using precomputed derivative statistics in HDF5.
• Summary statistics: GWAS variant-level association data, commonly LDSC .sumstats, GWAS-VCF, or parquet.
• Annotation: variant, region, or gene-set feature used for enrichment analysis.
• Surrogates: marker substitutions used when GWAS variants are missing from LDGMs.

Gotchas

Do not overwrite user data or generated results. This repo often has dirty data/, results/, output/, scorestats/, and fixture files during active analysis. Check file sizes and git status before treating local fixtures as a reliable baseline.