HCMI model-tumour DNA methylation beta matrices, restricted to the flagship epigenetic-concordance probe sets

What this is. Per-cancer-type DNA methylation beta-value matrices for Human Cancer Models Initiative (HCMI) patient-derived cancer models and their matched parental tumours, restricted to the probe sets used by the HCMI flagship paper's epigenetic-concordance analysis. 23 Parquet files, one per cancer type, 42.3 MB total. Each file is probes (rows) x samples (columns), Float32, with a leading probe column of Illumina EPIC CpG identifiers. Column names are 5-field HCMI sample barcodes, where dash-field 5 codes the arm: 85/86 = cancer model, everything else = parental tumour tissue. Why it exists. The underlying GDC beta files are 746 files totalling 16.9 GB, so recomputing methylation concordance requires a 16.9 GB download before a single number appears. This derived subset keeps only the probes the published method actually uses, which reduces the working set to 42.3 MB and makes the analysis interactive. It exists nowhere upstream: the GDC serves whole per-sample files only, and cBioPortal's HM450 profile is a gene-level collapsed matrix on a different platform. How it was built. Following the recipe in the authors' own rendered analysis notebook (analyses/notebooks/Epigenetic_Concordance_Fig.3a.html in the NCICCGPO/HCMI-Flagship-Manuscript repository, pinned at commit 8752eba0bdeae944392d71725870b479d6a351a0): Samples with cpg_probe_success_rate > 0.7 and dna_analysis_low_purity_exclusion == FALSE. Probes restricted to the 144,571 normal-tissue-unmethylated CpGs shipped in Supplementary Table 11. Per cancer type, probes retained where beta >= 0.3 in at least 2 samples of that cohort. Beta values read from GDC open-access Methylation Beta Value files, Data Release 45.0. Verification. Pearson correlations computed from these matrices reproduce the methylation_correlation column of the flagship paper's Supplementary Table 3 for 150 of 185 evaluable model-tumour pairs to seven decimal places, with a median absolute difference of 3.1e-09. A trap worth recording. Supplementary Table 11's sheet labelled "53,204 cancer-associated CpGs" is the UMAP probe set behind Figure 4c, not the probe set behind the methylation concordance analysis. Using it produces a plausible-looking near-miss - correlation-of-correlations 0.93 against the published values, with a systematic +0.045 offset and zero exact matches. Provenance and limits. Derived from open-access NCI Genomic Data Commons project HCMI-CMDC (Data Release 45.0) and from CC BY 4.0 supplementary material of ElHarouni et al., Nature 2026. Reproducible denominator is 193 of the paper's 201 evaluated pairs: 8 name a sample absent from the authors' own 824-row methylation sample table. Two samples carry duplicate open beta files in the GDC release and the choice between them is arbitrary; those pairs show the largest residuals. Use. Built by and consumed by the hcx vignette catalog notebook nb10_epigenetic_fidelity.py. Files are addressed individually as <base>/<CANCER_TYPE>.parquet.