TMB, ICI and aPD1 curation audit

Reviewed 2026-09-16, against repository baseline 52f423e.

The former UCEC_POLE 100% response anchor was a selected exceptional-responder case, not a population response rate. It has been withdrawn. The broader audit found unrelated citations, means entered as medians, mismatched source populations, combined treatment arms, control arms labeled as ICI, and incorrect response denominators. This change corrects confirmed failures and makes unresolved source verification explicit; it does not certify every remaining number.

Issues were filed before their corresponding fixes: #523 (POLE), #524 (complete audit tracker), #525 (regimens/populations/pooling), #526 (TMB statistics/provenance), #527 (incorrect citations), and #528 (anchors/denominators), and #530 (PR review and source recovery), #532 (means and all withdrawn claims), #533 (STK11 subgroup denominators), and #534 (neuroblastoma source recovery), and #535 (retained statistics/cohort review).

Scope and reproducible evidence

Every row in the four numerical reference tables is inventoried, including blanks and contextual estimates: 130 TMB, 110 ICI anchors, 84 aPD1 anchors, and 787 endpoint estimates (1,111 total). No estimate IDs were deleted. The row ledger records each value, citation, locator status and unresolved checks. Run:

python scripts/audit_tmb_ici_apd1.py --output docs/audits/tmb-ici-apd1.csv

The audit checks numeric validity, response/count arithmetic, confidence-interval ordering, drug/regimen consistency, compact-anchor agreement, missing denominators, and provenance. Its output is tested for reproducibility. These checks cannot prove that a source supports the biological population assigned to a row.

The citation inventory includes 191 distinct original and current references. PMID/DOI metadata identifies the actual article behind a citation; resolving an identifier is not numerical source validation. The TMB review table records a disposition for all 130 rows, including rejected legacy values, assay, locator, and review notes.

After source recovery and the best-effort round, 48 numeric TMB rows have direct source checks (41 published medians, two published means, two sample-recomputed medians, and three reported summaries whose statistic is unspecified). Two more rows have checked numerical sources but use explicit population proxies; one is a reviewed capture-size approximation. Another 67 numeric entries still require exact source verification, and 12 explicit gaps remain. All 20 initial withdrawals now have replacement estimates, with these quality distinctions preserved; an existing MPN gap was also filled. The endpoint inventory still flags 338 rows/anchors whose numeric source locator is not verified, 33 without response denominators, and 47 with denominators below 10. These counts overlap; they are review flags, not 338 independently demonstrated errors. Earlier source_verified and locator verified fields remain historical curation metadata, not a fresh scientific certification. #524 remains open for this source work.

POLE: strong biology does not establish a response-rate ranking

Mehnert 2016 describes one patient's partial response. The retained estimate now records a patient response count as context, with no population ORR or binomial CI. Both ICI and aPD1 lookups for UCEC_POLE return an explicit gap. Its unsourced TMB value of 100 mut/Mb is replaced by a separately supported genomic median of 150.8 mut/Mb in 61 patients from Nero 2025, Results and Figure 1B. That study uses a tumor-only TSO500 panel and the 11 pathogenic hotspot definition, including 16 multiple-classifier cases. It is not an ICI response trial.

The supplied critique correctly rejects the 100% population claim, but several of its supporting statements also need qualification:

  • Prospective POLE-selected evidence exists. Rousseau 2022, AcSé nivolumab studied MMR-proficient advanced solid tumors, reporting responses in 5/11 assessable patients with proofreading-deficient POLE variants. This is a pan-cancer result; it does not establish UCEC_POLE as the most responsive indication, and progression occurred in endometrial cancer despite pathogenic variants.
  • GARNET, Table 2, reports 2/5 responses in its exploratory POLE-mutant category. Both responders were dMMR. That small, overlapping biomarker category is not a clean pathogenic-EDM UCEC population estimate.
  • The final avelumab endometrial trial enrolled no POLE-mutated tumors. Its dMMR result cannot be treated as a measured POLE result. The toripalimab basket trial had only three assessable EDM cases; its two EDM responders had colon cancer.
  • The original 2022 rectal dostarlimab report reported 12 patients, not 42. The later 2025 expansion reported 49/49 rectal and 35/54 nonrectal clinical complete responses. Neoadjuvant clinical CR and advanced-disease RECIST ORR are different endpoints and settings. These results cannot establish a universal numerical response ceiling.
  • The POLE pathogenicity study supports variant-level interpretation. The quoted hotspot list is not a complete or accurate whitelist: the recognized endometrial set includes L424I, M444K and D368Y. Exonuclease-domain location alone does not establish pathogenicity, and a non-EDM variant is not automatically proven biologically inert. Multiple-classifier assignment and treatment-response evidence are separate questions.

The cHL response anchor remains 71.2% (173/243), now citing the corresponding CheckMate 205 five-year report, rather than the older 69% report. We do not adopt the accompanying assertion that cHL is uniformly low-TMB: the cited Hodgkin genomic cohort uses clinical panels and does not support the former approximate WES median of 2. No substitution-spectrum or netMHCpan binding claim was used to assign clinical response rates.

Confirmed numerical and source corrections

Area Previous claim Corrected treatment
POLE ICI/aPD1 100% ORR from a selected case explicit evidence gap; case response retained as context
COAD, READ, UCEC prevalence-weighted modeled ORRs models retained as context; canonical anchors blank
UCEC_CNH/CNL pMMR/MSS response assigned to molecular subtypes gap: source does not isolate those subtypes
LUAD_STK11 KRAS/STK11 evidence assigned to STK11/KEAP1 union; genomic n=637 used as response n gap; invalid denominator removed
UVM ICI mixed anti-PD-1/PD-L1 cohort labeled PD-1 mixed-agent context retained; first-line PD-1 anchor 11.7% (2/17) from a separate prospective cohort
Ependymoma 1/22 across mono and combo arms nivolumab 1/12 = 8.3%; combination n=10 kept separate
DIPG, medulloblastoma inferred zero ORRs with ordinary anchor confidence formal-ORR gaps; outcome inferences remain context
GBM enrolled n=184 used for ORR response-evaluable n=153; bevacizumab comparator n=156
MDS KEYNOTE-013 enrolled n=28 used for response categories and calculated CIs response-evaluable n=27; enrollment preserved separately
ESCA CheckMate 648 89/325 paired with 28% 90/325, reported rounded 28%; correct source CI
ASPS disease control 91% paired with 47/52 descriptive count-derived 90.4%, explicitly distinguished from a reported DCR endpoint
FL TMB 1.35, described as WES/mean panel median 5.05 mut/Mb, n=119
Pancreatic / midgut NET TMB 1.9 / gap advanced-site genome-wide WGS medians 1.35 / 1.05
BCC / cSCC TMB 65 / 45 with unrelated citations Chalmers panel medians 47.3 (n=92) / 45.2 (n=266)
LCNEC TMB George mean 8.6 entered as median Chalmers panel median 9.9 (n=288), explicitly a different assay/cohort
MDS TMB inferred conversion to 0.3/Mb explicitly reported Chalmers median 0.8/Mb
Penile SCC TMB 4.5 attributed to an erdafitinib trial same value supported by Chalmers Table 1, n=60

Primary sources for these corrections: CheckMate 908, CheckMate 143, KEYNOTE-013 Table 3, CheckMate 648 regulatory report, Figure 2.3.2.2-6, ASPS trial, FL Results 3.2, advanced NEN WGS, and Chalmers Table 1 and Results.

The rejected ADCC and RB values were means, not medians. ADCC now has a replacement panel median and its original mean is retained separately; RB is restored as a mean. The old FL value 1.35 cannot be rescued by relabeling it a mean: the cited paper reports a 409-gene panel median of 5.05, with the observed range starting at 1.69. Other directly checked medians include rectal NET, GIST, DSRCT, BTC, NSCLC and thoracic SMARCA4-deficient tumors; their source populations, assay methods and specimen-versus-patient distinctions remain explicit.

PR review: original sources and recovered medians

An unsupported citation is not evidence that no median exists. The 20-row recovery ledger preserves the original value/citation, what the claimed source actually reports, replacement evidence, and the reason for retaining any gap. Review findings were filed in #530 before the follow-up correction.

P1: retained Chalmers claims disagreed with its actual supplementary table. Additional file 3, Table S1 supports ACC 2.7 (n=204), CHOL 2.5 (liver cohort, n=1456), KIRP 2.7 (n=152), LIHC 3.6 (n=602), PAAD 1.8 (n=2483), THCA 1.8 (n=350), THYM 1.3 (n=108), THYMCA 2.5 (n=168), and UCS 3.6 (n=245). The old numbers did not match. Those rows are corrected, and sample counts are filled for 25 checked Chalmers rows, including an MPN median of 0.8 (n=138). The extracted source rows record exact workbook cell ranges, the download URL and SHA-256. No published medians were averaged.

P2: withdrawal was unnecessarily final for recoverable claims.

Code Recovered median, mut/Mb n Source and limit
UVM 0.34 80 Wu 2019 Table 1; old number was correct, DOI was wrong
NUTM 1.0 71 Kim 2025; known-TMB subset of 116 registry patients; heterogeneous panels
UCEC_POLE 150.8 61 Nero 2025; pathogenic hotspot cohort including multiple classifiers
MCL 3.3 75 Chalmers Table S1, lymph-node mantle-cell cohort
MTC 1.8 96 Chalmers Table S1, medullary thyroid
ACINIC 1.8 81 Chalmers Table S1, salivary acinic-cell cohort
ADCC 1.8 184 Chalmers Table S1, salivary adenoid cystic cohort
VSCC 5.2 72 Chalmers Table S1, vulvar SCC
ANSC 5.4 232 Chalmers Table S1, anal SCC
SARC_CHON 1.7 93 Chalmers Table S1, bone cohort; not pooled with soft-tissue chondrosarcoma

All Chalmers replacements above come from the actual published supplement. Its coding-panel definition includes synonymous substitutions and indels, then filters drivers and germline variants. Wu uses nonsynonymous coding mutations divided by 38 Mb. These medians cannot be interpreted as assay-harmonized estimates.

For UCEC_MSI, Leon-Castillo 2020 Table 2 provides a median of 21.5 in 127 MSI-H, POLE-wild-type tumors. This replaces 18, which the old TCGA citation did not establish as a median. The reanalysis includes synonymous coding variants and uses a 38 Mb denominator.

At that stage, seven original population-median claims remained unresolved: HL, CTCL, HCL, BRCA_Normal, LUAD_EGFR, UCEC_CNL and UCEC_CNH. The best-effort round below now supplies explicit estimates for each without reinstating the unsupported claims. The initial review established, for example:

  • The originally cited salivary paper reports alterations per tumor and fractions above a TMB threshold, not the claimed ACINIC median of 2.6.
  • The neuroblastoma paper does mention an overall approximately 0.6 exonic mut/Mb median in its introduction, citing earlier work. It does not establish separate MYCN subgroup medians.
  • EGFR evidence is available: Hastings 2019 reports 3.8 in 383 EGFR-mutant lung cancers, and Offin 2019 reports 3.77 in 153 metastatic exon19del/L858R cases. Hastings is now selected as an explicitly labeled broader-cohort proxy; its population fit is not presented as exact.
  • The primary Lawrence 2013 text establishes the median statistic and reports AML 0.37/Mb, but the precise remaining per-type claims require its sample supplement. Download attempts returned access challenges, not the data. These remain explicitly unverified; reading the abstract does not validate the numbers.

Runtime review found no additional failures in the current regimen-fallback, overlap-exclusion and gap-resolution cases. Passing software tests does not close the remaining scientific verification work in #524.

Canonical response anchors now agree numerically with their selected endpoint rows. This includes 5.3% for pretreated TNBC, 14.6% for PD-L1-positive cervical cancer, 19.3% for response-evaluable ESCC, 7.8% for GBM, 41.6% for intermediate/poor risk RCC on nivolumab/ipilimumab, 8.5% for mesothelioma, 25.9% for PD-L1-positive NPC, 5% for the selected PD-L1-positive prostate cohort, 43.7% for melanoma, and 11.6% for gastric/GEJ cancer. These are selected source settings, not universal unselected rates for each ontology label. Prior anchors that rounded or blended these values are preserved in git history and described in row notes.

Interpretation and safeguards

  • Default pooling selects one regimen and primary evidence only. Opting into alternates retains source rows, but shared citation/NCT/trial identities block a numerical pool because overlap has not been excluded. Invalid count/rate pairs also block pooling. Independent trials can still differ clinically; this guard is not a meta-analysis or a proof of cohort independence.
  • Drug labels now separate monotherapy, PD-1/CTLA-4, PD-L1 combinations, chemotherapy-containing arms, CTLA-4 alone, non-ICI controls, and mixed-agent cohorts. Controls, inferred outcomes, modeled blends and out-of-population context do not contribute to pools. The forest plot explicitly requests alternate trial points while retaining a primary-only numerical summary.
  • DART high-grade pan-NEN evidence is not LCNEC-specific; differentiated thyroid context is not ATC evidence; overall soft-tissue sarcoma is not synovial sarcoma. These rows remain inspectable, but cannot supply those subtype pools.
  • The legacy apd1 table includes explicitly tagged fallback regimens. Filter drug_target == 'PD-1' for monotherapy analyses. Inspect endpoint population, denominator, biomarker selection and evidence inheritance alongside every value.
  • TMB published_median provenance requires a checked source. Other reviewed types distinguish means, recalculated statistics, unspecified summaries, population proxies and approximations. Legacy needs_source_review values still await verification, even if their old confidence label was high. For the directly source-checked subset:
from oncoref import tmb

reviewed = tmb.cancer_tmb_df().query("source_review_status == 'source_checked'")
# Keep tmb_assay, source_scope, source_review_notes and n_samples in comparisons.

Open source work under #524 includes reproducing exact Lawrence supplementary medians, resolving unmatched Chalmers populations, validating pediatric and molecular-subtype proxies, resolving review-derived citations, and checking all remaining ICI endpoint populations and denominators against their original tables. No inference from TMB, mutation mechanism, binding predictions, or a response-rate ranking fills those gaps.

Typed means and continued verification (#532)

A supported mean is valid evidence when it is labeled as a mean. The TMB API now provides tmb_mut_mb and tmb_statistic; the scalar lookup prefers a median and falls back to a mean, then an explicitly typed estimate. median_tmb_mut_mb stays blank for mean-only evidence, and mean_tmb_mut_mb stays blank when a mean from that source has not been curated. Assays and populations still require inspection before comparing numbers.

Retinoblastoma WGS Results 2.2 supports mean 0.085 genome-wide substitutions/Mb, n=21. This restores the original value with published_mean provenance; it is not coding TMB or an SNV-plus-indel rate. Ho 2013 supports mean approximately 0.31 nonsilent mutations/Mb, n=60 for ADCC. That independent estimate is retained in the alternative-source ledger, without attributing it to the cohort supplying the selected panel median.

Second verification round: every withdrawn response anchor

The 10-code response recovery ledger accounts for every original canonical response withdrawal in both ICI and aPD1. It distinguishes observed outcomes, modeled prevalence blends, inferred zeros, regimen mismatches, and population mismatches. Means cannot remedy those issues.

Rossi 2019 restores a UVM PD-1 anchor: 11.7%, 2/17, with first-line setting, prospective observational design, RECIST 1.1 endpoint, and low confidence explicit. The old mixed-agent 3.6% estimate remains separately available as context.

Further review of Skoulidis 2018 found subgroup denominator errors (#533). SU2C KL is 4/54, with PD-1 alone or combined with CTLA-4, rather than 7.4% over all 174 patients. CheckMate-057 KL is 0/6, not 0/24; its spurious computed CI is removed. Both remain contextual for the broader STK11/KEAP1 ontology code.

The alternative TMB ledger now preserves verified EGFR-mutant lung mean/median, MF and Sezary estimates, and a newer Hodgkin panel mean/median with their actual cohort limits. The original Hodgkin table contains 37 numeric entries while its text reports 34; the transcribed table column preserves patient row numbers and the source-image hash. Its tabular median is 8.8, but the mismatch prevents labeling that as the paper's n=34 cohort median. TCGA endometrial supplements were downloaded and inspected: their selected-gene SMG tables cannot be summed into whole-exome per-patient TMB. These source limits, and checks of classic HCL, normal-like breast, and the original high-risk neuroblastoma report, are recorded for all remaining TMB gaps.

Third round: recover MYCN subgroup data and recheck retained NBL

Lee 2020 reports a 0.66 mut/Mb median in 58 MYCN-nonamplified tumors, after excluding synonymous variants. It spans 26 high-risk and 32 non-high-risk East Asian patients. This restores NBL_MYCNnonamp using a directly relevant cohort.

Pugh 2013 supplies actual per-sample total-exonic and nonsilent rates in Supplementary Table 1. Reanalysis of its 77 MYCN-amplified high-risk cases yields a nonsilent median of 0.4330966 mut/Mb, curated as 0.43 and labeled sample_recomputed_median. The separate nonamplified high-risk group has n=158; five unknown-status cases are excluded only from subgroup calculations. The complete 240-case extraction reproduces the paper's rounded overall medians, 0.60 total exonic and 0.48 nonsilent. The retained parent NBL value 0.60 now cites this direct primary source with high-risk scope and synonymous inclusion explicit.

Extracted sample rates and recomputed means/medians with source hash make the calculation reviewable. Reproduce the extraction from the original workbook:

python scripts/recompute_neuroblastoma_tmb.py --workbook Pugh-TableS1.xlsx --output-dir docs/audits

The alternative ledger preserves both statistics for both mutation definitions. The default amplified and nonamplified values come from different risk/ancestry cohorts and cannot establish an effect of MYCN amplification on TMB. Use the within-Pugh subgroup summaries for a like-assay comparison, with high-risk scope.

Fourth round: retained statistics and cohort denominators

Fernandez-Cuesta 2014 explicitly reports lung carcinoid mean 0.4 nonsynonymous mutations/Mb, from 44 independent pairs (29 WGS and 15 WES). NET_LUNG now uses the mean field, with no invented median.

de Bitter 2022 supports GBC median 5.5 in 96 TMB-evaluable samples. TSO500 TMB includes synonymous variants; the separate 54-gene actionable panel is not the TMB assay. The cohort was selected for at least pT2a disease and 30% tumor content. Results define TMB-high as >=10, although the abstract uses >10; the detailed Results support the curated threshold.

The craniopharyngioma source places 0.9/Mb in its mixed 15-case discovery cohort, not three papillary cases. Its exact statistic remains unresolved. George 2015 reports SCLC 8.62 nonsynonymous mutations/Mb in 110 WGS specimens without naming the summary statistic in the prose. Both original workbooks were inspected; sample metadata and variant calls alone do not establish callable denominators for a recalculated per-Mb median. These two values remain needs_source_review, with corrected cohort notes and supplement URLs/hashes.

Fifth round: best-effort estimates for all seven remaining withdrawals (#536)

Means are usable when identified correctly. A median describes the middle sample; a mean is sensitive to a high-TMB tail. Mixing them without labels can change a cross-cancer ranking. Neither summary determines an individual patient's TMB or ICI response. Both are retained when available from the same cohort.

The seven-row estimate ledger records the selected values and limitations; the original rejected claims remain in the recovery ledger.

Code Selected mut/Mb Statistic and source population
HL 7.66 Wienand 2019, reported summary, n=23 newly diagnosed classical HL, flow-sorted HRS-cell WES; exact summary statistic not assigned
CTCL 3.5 Mycosis fungoides 2024, median in 67 specimens from 48 patients; MF proxy for broader CTCL
BRCA_Normal 1.28 Recomputed median of 36 normal-like TCGA PanCancer Atlas cases; same-cohort mean 1.93
LUAD_EGFR 3.8 Hastings 2019, median in 383 EGFR-mutant lung cancers; same-cohort mean 5.6; broader lung-cohort proxy
HCL ~0.2 Approximation from five classic HCL exomes in Bibi 2016, Table 2 and capture methods
UCEC_CNL 2.9 TCGA 2013, reported subgroup rate, n=90; summary statistic unspecified
UCEC_CNH 2.3 Same source, n=60; summary statistic unspecified

For the two endometrial groups, Figure 2 confirms subgroup counts and the mutations-per-Mb scale. The prose inconsistently combines a per-base exponent with per-Mb wording; the curated values use the figure scale. These remain the original genomic classes, not automatic substitutions for IHC surrogate groups. A numeric TMB estimate does not fill either subtype's response-evidence gap.

Normal-like breast cancer is not TNBC. Normal-like is a gene-expression classification, potentially influenced by nonneoplastic tissue. TNBC is defined by ER, PR and HER2 status. Most TNBC is basal-like, but the classifications overlap imperfectly (Prat 2013). The new normal-like estimate joins patient subtype labels to sample nonsynonymous TMB in cBioPortal's pinned TCGA release. All 36 eligible samples represent unique patients. The extracted rates and source hashes and full-precision calculation make it reproducible. These are portal-provided rates; callable bases were not independently reconstructed. The cited PanCancer publication identifies the study, not a published normal-like median. This release differs from the eight normal-like cases excluded in the 2012 breast paper. Confidence remains low because of subtype interpretation, cohort size and purity effects.

python scripts/recompute_normal_like_breast_tmb.py --sample data_clinical_sample.txt --patient data_clinical_patient.txt --sequenced cases_sequenced.txt --output-dir docs/audits

For HCL, unflagged nonsynonymous counts 10, 22, 7, 10 and 4 have median 10. Dividing by the two nominal capture sizes (44.1 and 50 Mb) gives 0.20–0.23, rounded to ~0.2. The count extraction and calculation retain the restrictive filtering: using all candidate nonsynonymous counts gives 0.38–0.43 instead. These ranges show method sensitivity, not confidence intervals. Neither the sample-specific kit assignment nor the callable coding intersection is established. This is a low-confidence approximation; matching the old rounded number does not validate the old source or a population-median claim.

The API preserves blank mean/median fields for HL, HCL and the two endometrial reported rates. Their values live in estimate_tmb_mut_mb, selected only after median and mean, with tmb_statistic set to unspecified or approximate. The audit retains explicit flags for unknown summary statistic, population proxy and approximation. Source checking is separate from fit to the ontology population.