GOLDMARK: Governed Outcome-Linked Diagnostic Model Assessment Reference Kit

Navigate the portal

1. Select a dataset card (TCGA, IMPACT, CPTAC, etc.), then click a run name to drill into its tiling/feature/training progress.
2. Use the “Open Download Center” button to browse sanitized manifests, tiling CSVs, feature stores, training reports, and inference bundles. Every item is cached inside mil_code_beta/download_cache/<slug>, so it travels with the archive.
3. The per-run detail page links to interactive UMAPs, case tables, and CV plots. These visualizations also read from the cached thumbnails/HTML fragments bundled in outputs/jobs/thumbnails and mil_code_beta/download_cache/<slug>/html.
4. Need to jump straight to a cohort later? Bookmark ?dataset=TCGA (or IMPACT/CPTAC) in the URL; the filter bar will open on that subset.

Download & mirror the data

Large artifacts are distributed through the Download Center and reference repositories:

• The on-prem cache includes curated manifests, tiling coordinates, ROC/PR PDFs, cross-validation archives, and inference bundles for offline usage.
• TCGA source data: NCI Genomic Data Commons (GDC).
• When regenerating artifacts, rerun scripts/export_download_catalogs.py to refresh catalog metadata.

Reproduce the pipelines

Everything required to rebuild the cohorts or extend them to new targets is open-source:

• Code, configs, and SLURM helpers: GitHub · GOLDMARK.
• Mutation curation uses OncoKB tiers, reviewed by a board-certified molecular pathologist so only clinically actionable variants become positives.
• Tiling follows transparent heuristics (deterministic tissue masks, QC filters); feature extraction uses ViT-based encoders (Gigapath, UNI, Virchow) with documented checkpoints; aggregation defaults to gated MIL attention (GMA) but you can wire in TransMIL, HIPT, HEAL, or custom heads.
• Re-run the provided notebooks/scripts to rebuild from raw SVS or to create aligned sets for non-TCGA cohorts—every stage writes to the same directory structure consumed by this dashboard.

Why standardization matters

• Comparable tiles: Consistent tissue masking + tile selection means cross-cohort experiments aren’t confounded by variable patch sampling.
• Feature reuse: Pre-extracted embeddings (FP32/FP16) let you benchmark new MIL heads or prognosis models without re-running WSIs.
• Transparent aggregation: The GMA implementation mirrors what’s shipped in production, but the infrastructure exposes logits/labeled predictions so you can audit each decision.
• Clinically curated labels: Passenger mutations stay negative; only OncoKB-vetted alterations become positives, reducing label noise that plagues most TCGA reproductions.
• End-to-end provenance: Every generated PDF, CSV, or thumbnail is stored next to the catalog JSON, so exports are reproducible and portable.

Foundation encoders at a glance

• gigapath_ft – Fine-tuned encoder from the EAGLE study (Nature Medicine 2025), optimized for rare-event detection; pairs with the EAGLE aggregation stack.
• uni – Multi-scale UNI encoder (Nature Medicine 2024) that blends radiology + pathology embeddings for broader context.
• virchow / virchow2 – Internal distilled ViTs derived from Gigapath checkpoints; lighter memory footprint for rapid MIL experimentation.
• prov-gigapath – Province-scale adaptation of Gigapath tuned for heterogeneous scanners and staining protocols.
• h-optimus-0 – Baseline ViT used to benchmark MIL methods; ideal when you want a neutral encoder to compare aggregation functions.

Abbreviations & quick glossary

ACC

Adrenocortical Carcinoma

BLCA

Bladder Cancer

BRCA

Breast Cancer

CESC

Cervical Squamous Cell Carcinoma

COAD

Colon Adenocarcinoma

ESCA

Esophageal Carcinoma

GBM

Glioblastoma

HNSC

Head & Neck Squamous Carcinoma

KIRC/KIRP/KICH

Kidney Renal Clear Cell / Papillary / Chromophobe

LGG

Lower-Grade Glioma

LUAD/LUSC

Lung Adenocarcinoma / Squamous Cell

PAAD

Pancreatic Adenocarcinoma

PRAD

Prostate Adenocarcinoma

READ

Rectal Adenocarcinoma

SKCM

Skin Cutaneous Melanoma

STAD

Stomach Adenocarcinoma

TGCT

Testicular Germ Cell Tumor

THCA

Thyroid Carcinoma

UCEC/UCS

Uterine Corpus Endometrial / Carcinosarcoma

External tooling & references

• NCI Genomic Data Commons (GDC) for primary TCGA source access.
• GOLDMARK GitHub repository for pipeline implementation details.
• OncoKB for clinically curated mutation tiers.
• Hugging Face Hub docs for CLI/SDK usage.
• SLURM (job orchestration), conda/pip for environment management.