================================================================ HaploGraph -- 1000 Genomes Haploblock Co-occurrence Graph ================================================================ A population-scale graph where nodes are haploblock clusters (distinct genomic background states) and edges record how often two clusters are jointly carried by the same individuals, built from 1000 Genomes phased haplotypes (N = 2,548 individuals). See haplograph_unified.pdf in this directory for the full scientific write-up, methodology, and key findings. This file covers the file layout and schema only. ---------------------------------------------------------------- LAYOUT ---------------------------------------------------------------- 1000G/ README.txt (this file) haplograph_unified.pdf (methodology, findings, figures) phenotypes_real.csv (real 1000G ancestry/sex labels) checksums.md5 chr1/ nodes.csv.gz edges.csv.gz edges_lift_above_threshold.csv.gz top_edges_by_lift.csv.gz islands.csv.gz position_heatmap.csv.gz chr2/ ... ... chr22/ ... Every chromosome follows the same six-file schema described below. There is no global edges.csv file -- edges are same-chromosome-only by design (see PDF, "Scientific Design Decisions"), so a concatenated file would add no information over the per-chromosome files while being far larger. ---------------------------------------------------------------- nodes.csv.gz ---------------------------------------------------------------- Full, unfiltered binary feature matrix -- every cluster ever observed in this chromosome, one row per cluster, one column per individual. id,high_dim_edge,IND1,IND2,IND3,... chr21_14215892-14284114_cluster1,chr21_14215892-14284114,1,0,1,... id __cluster -- unique cluster ID high_dim_edge the haploblock region this cluster belongs to IND* 1/0 -- does this individual carry this cluster ---------------------------------------------------------------- edges.csv.gz ---------------------------------------------------------------- Co-occurrence graph, filtered by a symmetric minor-frequency threshold (min(support, N-support) >= 25 -- see PDF for rationale) and restricted to same-chromosome pairs. source,target,weight chr21_14215892-14284114_cluster1,chr21_14284114-14356253_cluster219,7 weight = number of individuals carrying both source and target. IMPORTANT: raw edge weight is dominated by a population-frequency artifact (the "mega-hub" effect, see PDF) and should not be used directly as a measure of genuine linkage. Normalize by lift before interpreting co-occurrence strength: lift(A,B) = weight(A,B) / (support(A) * support(B) / N) A lift near 1 means "co-occurring at chance level"; lift much greater than 1 means genuine enrichment beyond what marginal frequency explains. ---------------------------------------------------------------- edges_lift_above_threshold.csv.gz ---------------------------------------------------------------- Every edge with lift >= 5, i.e. co-occurring at least 5x more than expected under independence. This is the full, filtered set of edges with a real, non-trivial signal -- pre-computed so you do not need to compute lift over the entire edges.csv yourself. source,target,weight,lift chr21_39655927-39674795_cluster90,chr21_39700970-39725817_cluster37,26,81.79 Same columns as edges.csv, plus lift (see above). Substantially smaller than edges.csv (roughly 1-3% of total edges pass this threshold) -- start here rather than the full edges.csv if you want real co-occurrence signal without the mega-hub artifact. ---------------------------------------------------------------- top_edges_by_lift.csv.gz ---------------------------------------------------------------- The single strongest 200 edges by lift, for this chromosome. Same columns as edges_lift_above_threshold.csv.gz, just truncated to the top 200 -- useful for a quick look without loading the full lift-filtered file. ---------------------------------------------------------------- islands.csv.gz ---------------------------------------------------------------- Candidate extended-haplotype clusters: groups of clusters that are densely, mutually co-occurring (not just chain-connected -- see PDF, "Limitations") and stay within a bounded genomic span. Each row is one candidate multi-block extended haplotype. island_id,chromosome,n_nodes,n_edges,density,span_start,span_end,span_kb,max_lift,mean_lift 1,chr21,4,3,0.5,39665361,39713393,48.0,81.79,55.0 n_nodes number of clusters (nodes) in this island n_edges number of edges among those clusters density n_edges / max-possible-edges for n_nodes -- how mutually interconnected the island's clusters are (near 1.0 = tightly co-inherited as one block) span_kb genomic width of the island, in kb (capped at 500kb during detection; see PDF for why a span cap alone is not sufficient and density filtering is also required) ---------------------------------------------------------------- position_heatmap.csv.gz ---------------------------------------------------------------- Edge weight, binned by genomic position (500kb bins), for visualizing co-occurrence structure across the chromosome as a heatmap (see PDF, Figure 3, for an example). bin_i,bin_j,total_weight,edge_count bin_i, bin_j 500kb bin start coordinates for the two positions total_weight sum of edge weights between clusters in these bins edge_count number of edges contributing to that sum Concentration near the diagonal (bin_i close to bin_j) reflects real local linkage; broad off-diagonal bands reflect the mega-hub artifact, not genuine long-range linkage (see PDF). ---------------------------------------------------------------- phenotypes_real.csv ---------------------------------------------------------------- Long format, one row per (individual, phenotype) pair: individual_id,phenotype,value,source HG00097,ancestry,EUR,1000G_panel HG00097,population,GBR,1000G_panel HG00097,sex,male,1000G_panel 1000 Genomes collected NO phenotype data beyond self-reported ancestry, fine-grained population, and sex (confirmed via IGSR). Any other trait associated with this cohort elsewhere is necessarily simulated (e.g. a polygenic score projected from external GWAS summary statistics), never measured -- do not conflate the two. This file contains only the real, IGSR-sourced fields above. ---------------------------------------------------------------- COVERAGE AND PROVENANCE ---------------------------------------------------------------- Chromosomes Status Individuals Notes chr1-22 (all autosomes) Complete 2,548 MIN_CLUSTER_SUPPORT=25, same-chromosome-only chrX, chrY Not available -- not present in the upstream haploblock clustering data All 22 autosomes use identical filtering thresholds and generation logic, so files are directly comparable across the whole autosomal genome. Sex chromosomes are not included in this release. ---------------------------------------------------------------- CITATION AND PROVENANCE ---------------------------------------------------------------- Method follows the discretization framework of Kubica et al. (2025), "Decoding Complex Genotype-Phenotype Interactions by Discretizing the Genome." Built at the Unit for Genomic Medicine, MDxCORE, Rigshospitalet, September 2026, for the Nordic Biobank x NVIDIA Federated Learning Hackathon. ---------------------------------------------------------------- INTEGRITY ---------------------------------------------------------------- Verify downloads against checksums.md5: md5sum -c checksums.md5 ---------------------------------------------------------------- QUESTIONS ---------------------------------------------------------------- See haplograph_unified.pdf for full methodology, limitations, and known caveats (fixed haploblock-boundary resolution, trans-chromosomal effects excluded by default, symmetric filter's completeness assumption). Contact: Mauricio, Unit for Genomic Medicine, MDxCORE, Rigshospitalet.