Start from a research question

18,004 patients profiled on a clinical panel give tight confidence intervals on alteration frequencies, including for alterations too rare to study in a thousand-patient research cohort. Frequency and co-occurrence need only the variant calls and the tumor type, both of which are present.

Clinical sequencing populations are enriched for advanced and pretreated disease relative to the largely treatment-naive primary tumors in research cohorts. With 18,004 cases the comparison is well powered, and the discrepancy is itself a caution about generalising research-cohort findings to clinical practice.

Scale plus a clinically-deployed panel makes this the right cohort for estimating what fraction of patients would be eligible for a given targeted therapy, a question that research cohorts systematically answer wrong.

2,281 RNA-sequenced paediatric cases can be contrasted directly with adult AML cohorts, and the marker paper established that the differences are substantial and age-dependent. This remains the definitive paediatric comparison set.

Vital status informative for 86.6% of 2,492 cases with median follow-up of 41.7 months gives a well-powered survival cohort, and therapeutic agents are recorded for 35% so treatment can be partially controlled.

251 Black or African American patients alongside 1,551 white patients is one of the better-powered settings in this corpus for a race-stratified comparison in a paediatric cancer.

This is the question CPTAC exists to answer, and it needs both halves joined. Genomics on over 1,800 cases here pairs with proteomic studies in the PDC on the same participants, allowing copy number and mutation effects to be traced to protein and phosphosite abundance.

Progression or recurrence is informative for 77.9% of 1,866 cases with a median follow-up of 22.8 months - unusual completeness for a recurrence endpoint, and the basis for a genuinely pan-cancer recurrence analysis.

Six cancer types collected prospectively under a common protocol make cross-tissue subtype comparison less confounded by collection differences than a retrospective pan-cancer assembly.

Treatment outcome and regimen are populated for 100% of 1,510 cases, and regimen strings name the actual protocol arm, so response can be modeled within a treatment context rather than across an uncontrolled mixture of therapies. Whole genome sequencing on 1,507 cases supports genome-wide discovery at this sample size.

333 of 1,510 patients are recorded as Hispanic or Latino - both a substantial absolute number and a substantial share, which is unusual. With treatment held closer to constant by the trial protocol, the ethnicity contrast is less confounded by care access than in an observational cohort.

Vital status is informative for 84% of cases and median follow-up is 100.2 months, so late relapse and long-term survival are genuinely observable rather than censored away. Very few genomic cohorts of this size reach this follow-up duration.

Structural variant and copy number calls exist for 1,504 cases, and the regimen field records the risk-stratified arm each patient received, allowing molecular subtype to be crossed with the therapy actually delivered.

Expression, methylation, copy number, microRNA and protein array data on largely the same 1,098 tumors, with subtype assignments established across hundreds of prior papers. This is the standard benchmark, and the density of prior work is the point.

Therapeutic agents are recorded for 71% of cases - cyclophosphamide, tamoxifen, anastrozole, paclitaxel and doxorubicin are the commonest - and a treatment outcome field is populated for 64%. Combined with expression on 1,095 cases this supports agent-stratified response analysis.

Imaging for these same patients is held in the Imaging Data Commons and is openly downloadable, while the molecular data sit in the GDC. Joining them gives a multimodal cohort most users never realise is available.

Vital status is informative for 99.9% of cases and follow-up time is computable, giving a clean overall survival endpoint on a large cohort.

Models and their source tumors are both characterized, with near-complete whole genome coverage, which makes fidelity measurable rather than assumed. The answer determines how much any model-based result should be trusted.

Coverage spans many tumor types including rare ones. For a researcher working on a cancer with no established cell line, this is often the only route to any model at all, and it is the cohort's most practically valuable use.

Prior treatment is informative for 88.9% of donors and treatment given is recorded for all, so models derived from treated and untreated patients can be compared - a question with direct bearing on how model-based drug screens should be interpreted.

Radiology on 607 cases and pathology slides on 389, paired with expression on 293, methylation on 423 and copy number on 599 for the same patients. Glioma is the disease where imaging is most clinically central, and this pairing is openly available.

Methylation on 423 cases with subgroup definitions established across a large literature makes this the standard validation set for glioma classification.

Vital status is informative for 96.3% of 617 cases, giving a clean overall survival endpoint on a well-characterized cohort.

Whole genome sequencing on 395 cases with structural variant and copy number calls on 394 supports genome-wide signature analysis, including the fusion-driven and small-deletion patterns the marker paper associated with radiation. This is the cohort's central strength and the reason it was assembled.

RNA sequencing on 438 cases and microRNA on 437 gives a well-powered expression cohort. TCGA-THCA provides a 507-case sporadic comparison group in the same repository with the same harmonization, which makes a cross-cohort contrast unusually tractable.

The cohort's defining feature is exposure during a developmental window. Age variables together with the genomic data allow the question to be posed, and it is one of the few human settings where exposure timing is externally fixed by a single dated event rather than self-reported.

microRNA data exist on 437 of 449 cases, an unusually complete layer, and microRNA assays are far more tractable in resource-limited screening settings than sequencing.

Methylation arrays are available on 120 cases and RNA sequencing on 123, on the same tumors as whole genome sequencing that can be used to type and subclassify the integrated HPV. This is the design the marker paper used, and the cohort is large enough to repeat it with different clade groupings or to test subtype assignments derived from TCGA-CESC against an independent, differently constituted population.

207 of 212 cases are recorded as Black or African American, making this the only cohort in this corpus where a non-white group constitutes more than half the participants. Expression and methylation data allow subtype classifiers trained on TCGA to be applied directly, and classifier failure is itself the finding.

ECOG performance status is informative on 98.6% of cases, which is exceptional - most GDC projects record it for none. Treatment type is populated for every case, therapeutic agents for 29.7% (cisplatin, paclitaxel and carboplatin are the commonest), and a treatment outcome field is populated for 67.9%. Together these support a response analysis that is impossible in most cohorts of this size.

Vital status is informative for 100% of cases and follow-up time is computable, giving a usable overall survival endpoint. The short follow-up is itself scientifically interesting because it reflects real-world attrition in the setting where the cohort was recruited.

Histopathology and radiology are both available for 211 of 212 cases, alongside the molecular data. Imaging models trained largely on North American and European cohorts have documented performance gaps on under-represented populations, and this is one of the few resources that permits an external check.

This is the defining radiogenomics question and it needs paired data on the same patients, which this collection provides for the subset with matched microarrays. Imaging is available for all 211.

Both modalities were acquired for the same patients pre-surgically, so feature stability across modality and acquisition can be assessed directly - a prerequisite for any radiomics claim and rarely testable.

Expert segmentations and semantic annotations are provided, and the Imaging Data Commons holds several other lung collections that can serve as external validation within one access framework.

Phosphorylation, acetylation, glycosylation and ubiquitylation are measured on largely the same 184 to 193 tumors. Comparing regulation across four modification types on one cohort is an analysis this dataset uniquely enables; almost every other resource offers at most two.

Ubiquitylome data on 184 cases is among the largest such datasets anywhere. Degradation is a major regulatory axis that expression data cannot see, and a cohort this size makes subgroup discovery feasible rather than anecdotal.

Metabolomics on 187 cases is available alongside the phosphoproteome, so metabolic state can be related to signaling directly instead of being inferred from enzyme transcript levels - which is what most metabolic analyses in cancer actually do.

Interaction data on 153 cases is exceptionally rare in a tumor cohort, and provides a measured reference against which inferred co-expression networks can be scored.

Lipidomic profiles on 98 primary AML samples, alongside proteome on all 164, allow lipid class abundance to be related to protein markers of myeloid differentiation on the same patients. Measured lipidomics on primary AML at this scale is essentially unavailable elsewhere.

Metabolomics and proteomics on the same 98 patients allow a direct test of an assumption that a large body of computational metabolic modeling depends on and rarely gets to check.

Acetylome on 164 cases and metabolome on 98 overlap substantially. Acetylation depends on acetyl-CoA availability, so this cohort can test a mechanistic link that is usually assumed rather than measured.

Glycoproteome on all 164 cases; cell-surface glycosylation is directly relevant to immunotherapy target selection in myeloid disease.

Multiplexed tissue imaging with single-cell RNA sequencing on the same specimens, in an atlas designed around this question, with therapy records available as a clinical component to define the treatment context.

Neighborhood definitions derived here can be tested against other HTAN atlases that share the multiplexed imaging component, giving cross-atlas comparison within a single access framework and a common data model.

This atlas carries the extended clinical data tier that most atlases lack, alongside therapy and follow-up. That is what allows spatial features to be related to what was actually given and what happened next.

Ubiquitylome on 80 cases alongside proteome on all 115 allows the two regulatory modes to be separated on the same tumors - the direct test of an assumption that expression studies make implicitly.

Both modification layers are measured on all 115 cases, allowing cross-modification analysis on a common set of tumors rather than across separate cohorts.

Direct measurement of ubiquitylation state in primary tumors is the substrate for target nomination in a therapeutic area where target selection is usually based on inference.

Whole genome sequencing on all 101 metastatic biopsies pairs naturally with TCGA-PRAD's 500 primary tumors in the same repository under the same harmonization. The contrast is the cohort's clearest use and does not depend on any outcome field.

RNA sequencing on 99 metastatic biopsies is the right substrate for AR-signaling and neuroendocrine scoring, and metastatic tissue is where these phenotypes actually occur.

Structural variant calls exist for all 101 cases from whole genome rather than exome data, which is what makes rearrangement analysis credible here.

This is the question most people want the cohort for, and it is the one the harmonized records cannot support. It is listed here to be explicit that it requires outcome data obtained separately.

Exome on 69 and expression on 49 organoids, with measured drug responses reported in the source publication. Because response is measured experimentally rather than inferred from clinical records, this avoids the treatment confounding that defeats observational response analysis.

A model-fidelity question that determines how much weight any organoid result deserves, and one that can be posed here because both the organoid profiles and the source tumor stage information are available.

Expression on 49 organoids can be scored against the basal-like and classical subtype definitions derived from bulk tumor cohorts, testing whether those subtypes survive in culture.

Whole genome sequencing on 66 cases from an HIV-enriched population, contrastable with the many DLBCL cohorts derived from largely HIV-negative populations. Whole genome rather than exome makes structural and viral-integration analysis possible.

RNA sequencing on 55 cases with ECOG performance status on 71% allows transcriptional programs to be related to a clinical proxy for host status.

Vital status is informative for 76% of cases, follow-up is derivable for 57 of 70 with a median of 11.3 months extending to 99.6, and 24% carry a named regimen.

This is a predictive-biomarker question, and it requires randomized treatment assignment to answer properly. RNA sequencing on 44 cases supports cell-of-origin and subtype classification, and the trial provides the randomization.

Classifier performance usually degrades when treatment heterogeneity is removed, because some apparent prognostic signal is really treatment-selection signal. A uniformly treated, randomized cohort is the right place to detect that.

Vital status is informative for all 45 cases with 60.4 months of median follow-up and cause of death recorded for 35.6% - unusually complete for a cohort this small.

Whole genome sequencing on all 39 cases, contrastable with the large TCGA lung cohorts in the same repository under the same harmonization. Whole genome rather than exome allows mutational signature and structural analysis.

Alongside the cervical and lymphoma cohorts in the same program, it allows a hypothesis about HIV-associated tumor biology to be tested across three tumor types collected under a common protocol.

Vital status is informative for 89.7% of cases, with follow-up derivable for 36 of 39 and extending to 104 months despite a short median of 6.4.

Serial biopsies with multiplexed imaging and spatial transcriptomics on the same patients across treatment, with therapy and follow-up clinical components present. This longitudinal spatial design is essentially unique in public data.

Single-cell ATAC and single-cell RNA on the same serial samples allow regulatory and transcriptional state to be ordered in time within a patient.

Electron microscopy appears in no other dataset in this corpus. Linking ultrastructure to spatial and single-cell molecular data on the same specimens is possible only here.

Both assays are present on shared specimens, which makes cross-platform concordance measurable - a methodological question the spatial field needs answered and rarely can.