sigmutselcovs covariate matrices: TCGA-UCS

Pre-built per-gene covariate matrices for TCGA-UCS (Uterine carcinosarcoma (malignant mixed Mullerian tumor): a biphasic tumor of the uterine corpus whose sarcomatous component arises from the carcinoma by epithelial-mesenchymal transition, so its cell of origin is endometrial epithelium, the same as UCEC's (TCGA's marker paper, Cherniack et al. 2017, finds it closest to serous endometrial carcinoma). Every tissue source is therefore UCEC's, re-verified 2026-10-04. GTEx maps to Uterus (the existing UCS key; representative gtex_uterus). Expression is tumor-only: all 57 TCGA-UCS STAR-count files are primary tumors, no normals. No TCGA ATAC-seq for UCS (not in Corces et al. 2018; atac: null). roadmap null: none of Roadmap's 127 epigenomes is uterine (E097 ovary and the GI smooth-muscle epigenomes are the wrong organ or cell). encode_chromatin: UCEC's adult uterus tissue tracks -- H3K9me3, H3K36me3, H3K27ac, H3K27me3 from one 51y donor (ENCDO271OUW), H3K4me1 and DNase from a 53y donor (ENCDO793LXB) -- plus H3K4me3 from the 51y donor (ENCSR743GCE, combined replicates), which UCEC's row records as unreleased but has been released since 2018; ENCODE has no uterus H3K9ac. Cell line: Ishikawa DNase-seq, the only endometrial cancer line in ENCODE (endometrioid adenocarcinoma, so a weaker match for carcinosarcoma than for UCEC), taken from the DMSO vehicle-control experiment (ENCSR000EIP) rather than UCEC's 4-hydroxytamoxifen-treated one. No carcinosarcoma or uterine sarcoma line exists in ENCODE (MES-SA, SK-UT-1, SK-LMS-1, HEC-1, AN3-CA, KLE, RL95-2 checked). ENCODE's adult uterus ATAC-seq (46y, 59y) and endometrium FAIRE-seq are left out, as ATAC is not an encode_chromatin label here. repliseq null: endometrium is a settled replication-timing negative (no source in ENCODE, GSE137764 or GEO, Ishikawa and HEC-1 included), and the cell-of-origin question leads back to endometrial epithelium; fibroblast or smooth-muscle profiles would be an organ-level guess for the sarcomatous component and are declined. Assemblies: encode_chromatin GRCh38.), built with sigmutselcovs: GTEx tissue-matched gene expression, TCGA per-sample gene expression, replication timing (mean replication time and CLR-transformed S-phase fractions), Roadmap Epigenomics histone marks, and TCGA ATAC-seq accessibility, indexed by Ensembl gene ID. Files: cov_matrix_full.parquet — the full matrix (all covariate columns) cov_matrix_pca.parquet — PCA-reduced to 910 components, the smallest number reaching 99% cumulative explained variance (see pca_manifest.json) cov_matrix_columns.csv — a data dictionary: one row per column of the full matrix, with source, description, assembly, and units build_manifest.json, pca_manifest.json — build provenance and PCA parameters Full source citations and methodology: SOURCES.md. Release of 2026-10-26. Every chromatin and ATAC column is one track, summarized as log1p of the mean signal over the gene's CDS span (first to last coding base); replication timing covers chrX. cov_matrix_pca.parquet now reduces the combined matrix: this cohort's own covariates plus the tissue-agnostic GENERIC pool (GENERIC record), with the pool's columns for chromatin tracks this cohort already has left out. That is the matrix the sigmutsel covariate model is fitted on; it predicts held-out mutation counts better than the cohort's own matrix alone in the cohorts tested, so the first components of this artifact are the ones a model should use. Versions before 2026-10-26 reduced the own matrix alone. cov_matrix_full.parquet is still the cohort's own matrix; GENERIC's is in its record. pca_manifest.json records which matrix was reduced (matrix) and the column counts. To rebuild: sigmutselcovs pca-artifact UCS --data-dir DIR --generic-dir GENERIC_DIR.