Add TCGA-LUAD (585 patients) + full GDC dictionary-derived schema + TcgaHfPatient pydantic reference impl + dictionary version capture

README.md

@@ -17,6 +17,10 @@ configs:
     data_files:
       - split: train
         path: TCGA-DLBC/*.parquet
+  - config_name: TCGA-LUAD
+    data_files:
+      - split: train
+        path: TCGA-LUAD/*.parquet
 ---
 
 # TCGA Patients (Open Access)
@@ -24,9 +28,13 @@ configs:
 **Open-access** TCGA patient data from the NCI Genomic Data Commons (GDC).
 One HuggingFace subset per TCGA project; one row per patient.
 
-- **Projects included:** `TCGA-CHOL`, `TCGA-DLBC`
-- **Generated:** 2026-05-04 03:21:55 UTC
+- **Projects included:** `TCGA-CHOL`, `TCGA-DLBC`, `TCGA-LUAD`
+- **Generated:** 2026-05-04 05:19:32 UTC
 - **Source:** NCI GDC `/cases` endpoint, open-access tier only.
+- **Schema:** derived from the [GDC Data Dictionary][gdc-dict]; the live
+  dictionary snapshot the GDC was serving when the data was fetched is
+  hashed into each project's `gdc_status.json` (and stored alongside the raw
+  data on the producer side, not shipped with the parquet).
 - **GDC data release:** Data Release 45.0 - December 04, 2025
 
 ## GDC data model primer
@@ -46,36 +54,30 @@ case          one patient (TCGA-XX-1234)
             └── aliquot    a vial of that analyte handed off for sequencing
 ```
 
-In this dataset we **flatten portions and analytes away**, since they're
-mostly physical-prep grouping. Each row's `samples` field is `list<sample>`,
-and each sample carries an `aliquots` list directly. The DNA/RNA distinction
-is preserved on each aliquot via the `analyte_type` field (hoisted from the
-parent analyte).
-
-**Tumor vs normal.** Somatic variant calling (e.g. open-access MAFs) compares
-a **tumor sample** against a matched **normal sample** from the same patient
-to identify mutations acquired by the tumor. The normal usually comes from
-blood (`Blood Derived Normal`) or adjacent healthy tissue (`Solid Tissue
-Normal`). Each variant carries the aliquot UUIDs of both members of the pair.
-
-**Sample types** you will see in this dataset:
-- `Primary Tumor` — the index tumor at diagnosis
-- `Solid Tissue Normal` — adjacent histologically normal tissue
-- `Blood Derived Normal` — the patient's matched blood
-- `Bone Marrow Normal`, `Recurrent Tumor`, `Metastatic`, ... (less common in TCGA)
-
-**Analyte types** drive what assays an aliquot supports:
-- `DNA` aliquots → WXS / WGS / methylation array / SNP array
-- `RNA` aliquots → RNA-Seq / miRNA-Seq
-
-**GDC file taxonomy** (used to identify what's in each `samples_*` column):
-- `data_category`: the broad bucket (e.g. *Simple Nucleotide Variation*,
-  *Transcriptome Profiling*).
-- `data_type`: the specific data product (e.g. *Masked Somatic Mutation*,
-  *Gene Expression Quantification*). **Our top-level molecular columns are
-  named `samples_<data_type_snake_case>`** so the source product is
-  unambiguous.
-- `experimental_strategy`: the underlying assay (e.g. *WXS*, *RNA-Seq*).
+We preserve this hierarchy verbatim — each `samples[i]` carries
+`portions[j].analytes[k].aliquots[m]` exactly as the GDC returns it. No
+flattening, no field hoisting; portion- and analyte-level fields like
+`is_ffpe`, `a260_a280_ratio`, and `normal_tumor_genotype_snp_match` are
+preserved. The `TcgaHfPatient` reference implementation (below) provides
+flat-aliquot convenience views over this tree.
+
+**Top-level molecular columns** are named `samples_<gdc_data_type_snake_case>`
+(e.g. `samples_masked_somatic_mutation`, `samples_gene_expression_quantification`).
+Each entry carries FK fields back to the patient's `samples[].portions[].analytes[].aliquots[]`
+so cross-modality joins are local to the row.
+
+**Timeline anchor is uniform per the GDC dictionary.** Every `days_to_*` field
+is documented as days from the case's `index_date` (a top-level field on each
+row; for TCGA usually `"Diagnosis"`). `TcgaHfPatient.timeline()` returns every
+dated event for the patient on this single anchor — clinical (consent →
+diagnosis → treatments → follow-ups → lost-to-follow-up → death) plus
+biospecimen (`sample_procurement` from `days_to_sample_procurement`,
+`bcr_receipt` from `days_to_collection`).
+
+For some TCGA cases `days_to_collection` exceeds `days_to_death`. We don't
+attempt to reinterpret these — we surface the count via
+`TcgaHfPatient.consistency_check()` as `bcr_receipts_after_death` and leave
+interpretation to the consumer.
 
 **Reference Python implementation:** the `tcga2hf` package on GitHub ships a
 fully-typed pydantic `TcgaHfPatient` model that mirrors this schema and adds

TCGA-CHOL/train.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b261c92f2a1d02c1412cbb1b9fdc381c8a9c9e0e4967ee93f1547f1b8702da1f
-size 51303188
+oid sha256:7aabe3f746dfc5d79f895443e98ad06bad6955119ed8ea3c6608f5ffd749f7a1
+size 51390111

TCGA-DLBC/train.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c08c0c1eaae44aa1fc0de45a6c6002b1f99a0ebdfe321081176eea16afc92002
-size 57772197
+oid sha256:99dc3160cd608ebfb06b266d95be670301e9fd13cc461a7c5fc75d22b777e67d
+size 57869184

TCGA-LUAD/train.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:84c5a93057d09405d95a11a442b4ba00ba444a6c1b22e68db1bdb65fe2fd6e46
+size 788972385