Skip to content

Pediatric Cancer Entries: Manual Content Review (2026-07-25)

Manual domain review of ten pediatric cancer entries in kb/disorders/, reading the entries as an oncologist would: is the biology right, is the nosology current, and is the core disease content present? Findings are ordered by how wrong they are, not by how easy they were to find.

Entries: Acute_Lymphoblastic_Leukemia, Neuroblastoma, Wilms_Tumor, Retinoblastoma, Hepatoblastoma, Medulloblastoma, Ewing_Sarcoma, Osteosarcoma, Alveolar_Rhabdomyosarcoma, Atypical_Teratoid_Rhabdoid_Tumor.

1. Outright factual errors

1.1 Medulloblastoma — the WHO 2021 classification is stated incorrectly

"WHO 2021 stratifies medulloblastoma into four molecular subgroups — WNT-activated, SHH-activated, Group 3, and Group 4"

This is the 2016 WHO / Heidelberg consensus scheme, not WHO CNS5 (2021). WHO 2021 defines the molecularly-defined entities as: MB WNT-activated; MB SHH-activated and TP53-wildtype; MB SHH-activated and TP53-mutant (split into two separate entities); and MB non-WNT/non-SHH, within which Group 3 and Group 4 are provisional subtypes, not top-level groups. WHO 2021 also carries a parallel histologically-defined axis (classic, desmoplastic/nodular, MBEN, large-cell/anaplastic) that this entry omits entirely.

The entry is internally inconsistent about this: the SHH subtype description correctly says "TP53 mutation status further stratifying prognosis under WHO 2021", which only makes sense under the scheme the top-line sentence contradicts.

1.2 Neuroblastoma — Stage 4S is defined with the wrong age cutoff

"A special stage occurring in infants under 18 months with primary tumor and metastases limited to skin, liver, and bone marrow."

INSS Stage 4S is <12 months. The 18-month threshold belongs to the INRG stage MS, a different staging system. The entry uses the INSS label with the INRG cutoff. It also omits the criterion that marrow involvement must be minimal (<10% of nucleated cells) — without that limit, 4S is indistinguishable from stage 4, which is the entire clinical point of the category.

1.3 Neuroblastoma — ALK in familial neuroblastoma is understated

"ALK point mutations occur in 8-10% of sporadic and ~50% of familial neuroblastoma."

The sporadic figure is right. Germline ALK mutations account for the large majority of hereditary neuroblastoma pedigrees — commonly cited at ~75-80% (Mossé et al., Nature 2008). "~50%" materially understates ALK's role as the neuroblastoma predisposition gene.

1.4 Medulloblastoma — the cell-type binding contradicts the node's own text

The single pathophysiology node states, correctly, that "WNT tumors derive from lower-rhombic-lip progenitors and SHH tumors from external-granule-layer granule-neuron precursors" — then binds cell_types to cerebellar granule cell (CL:0001031) alone. That is the SHH lineage only. WNT tumors arise from lower rhombic lip / dorsal brainstem progenitors, and Group 3/4 are now attributed to rhombic-lip-derived progenitors and unipolar brush cells (Hendrikse et al. and Smith et al., Nature 2022). The binding is wrong for three of the four subgroups the node claims to cover.

1.5 Ewing_Sarcoma — "pathognomonic" is too strong, and WHO reclassification is missing

"characterized by the pathognomonic EWS-FLI1 fusion gene, present in approximately 85% of cases"

EWSR1 rearrangement is not pathognomonic — it occurs in desmoplastic small round cell tumor, clear cell sarcoma, myxoid liposarcoma, and others. More importantly, the WHO 2020 soft-tissue classification split CIC-rearranged sarcoma and BCOR-rearranged sarcoma out of the Ewing family as separate entities; the entry does not reflect this. And the ~15% of cases that are not EWS-FLI1 are nowhere modeled — most are EWSR1-ERG (t(21;22), ~10%), with rarer FEV/ETV1/ETV4 partners. For an entry with 16 pathophysiology nodes devoted to fusion biology, omitting the second-most-common fusion is a substantive hole.

1.6 Wilms_Tumor — TP53 frequency contradicts itself

genetic[6] says TP53 mutations occur in "approximately 50-60% of diffuse anaplastic Wilms tumors"; histopathology[4] says "nearly all anaplastic Wilms tumors harbor TP53 mutations". Both cite the same source, which says "nearly all … if one looks hard enough". The 50-60% is the outlier and should go — current understanding is that TP53 alteration is near-universal in diffuse anaplasia when adequately assayed.

1.7 Same drug, two different chemical entities

Neuroblastoma binds cyclophosphamide to CHEBI:4026 — whose label, recorded in the file itself, is "cyclophosphamide hydrate" — while Medulloblastoma binds CHEBI:4027 "cyclophosphamide". The monohydrate is a different chemical entity from the drug substance. Both labels are real, so this is invisible to term validation; it should be CHEBI:4027 in both.

2. Nosology / classification problems

2.1 Wilms_Tumor — the subtype list conflates four orthogonal axes

has_subtypes lists as siblings: histology (Favorable, Anaplastic, Blastemal, Epithelial, Stromal, Mixed), laterality (Bilateral, Unilateral), etiology (Hereditary, Sporadic), and age (Childhood, Adult).

"Unilateral", "Sporadic", and "Childhood" are not subtypes — they are the default case, and their descriptions say as much ("Most Wilms tumors present as unilateral renal masses"). Worse, the histology entries mix two incompatible classification systems: "Favorable/Anaplastic" is the COG system applied at upfront nephrectomy, while "Blastemal/Epithelial/Stromal-predominant" is the SIOP system applied after preoperative chemotherapy. The entry's own treatment section correctly explains that COG and SIOP differ in exactly this way, then the subtype list flattens both into one namespace. Since has_subtypes[].name is the foreign-key target for subtype: references elsewhere, this makes the axis unusable.

2.2 Atypical_Teratoid_Rhabdoid_Tumor — subgroups named but not modeled

TYR, SHH, and MYC are correctly identified as the three consensus methylation subgroups, but they carry no mechanism. The distinguishing biology — TYR (melanosomal/tyrosinase program, infratentorial, youngest), SHH (SHH/NOTCH signaling, supra- and infratentorial), MYC (MYC/HOX, supratentorial) — is absent, and the SHH subgroup is not linked to any SHH mechanism module despite the KB having the machinery for it. The pathograph terminates in a generic "Aggressive Tumor Cell Proliferation" node that could belong to any cancer.

3. Missing core disease content

3.1 Osteosarcoma — no histopathology, no biochemistry, and the key prognostic factor is absent

histopathology: [] and biochemical: [] are both empty. For osteosarcoma this is the most serious content gap in the ten entries:

  • The defining diagnostic criterion — production of malignant osteoid by neoplastic cells — is never stated as a pathology finding.
  • The histologic subtypes (osteoblastic ~50%, chondroblastic ~25%, fibroblastic ~25%) are absent.
  • Percent tumor necrosis after neoadjuvant chemotherapy (Huvos grade; ≥90% = good responder) — the single strongest prognostic factor in localized osteosarcoma, and the entire reason chemotherapy is given before surgery — is absent. The entry describes neoadjuvant MAP without ever explaining what it is for.
  • Alkaline phosphatase and LDH, standard prognostic labs, are absent.
  • Radiographic hallmarks (Codman triangle, sunburst periosteal reaction) are absent.

Subtypes list only Conventional High-Grade, Telangiectatic, and Small Cell — omitting parosteal and periosteal surface osteosarcomas (low/intermediate grade, managed very differently) and secondary osteosarcoma (Paget disease, prior radiation). The entry's own cited source lists "low grade central, telangiectatic, small-cell, surface and intracortical", so these were available and dropped.

Predisposition syndromes (Li-Fraumeni, hereditary retinoblastoma, Rothmund-Thomson, Werner, Bloom, Diamond-Blackfan) appear only inside a quoted snippet, never as structured content — even though the Retinoblastoma entry in this same KB names osteosarcoma as its principal second malignancy. The cross-reference is one-directional.

3.2 Neuroblastoma — the last decade of neuroblastoma biology is missing

The pathograph has three nodes, and ALK Signaling Activation is an orphan with no downstream edge at all — it connects to nothing. Absent entirely:

  • Telomere maintenance, the axis that now organizes neuroblastoma risk biology: MYCN amplification, TERT rearrangement, and ATRX mutation/ALT (the latter defining the indolent adolescent/young-adult subtype). Neither TERT nor ATRX appears anywhere.
  • PHOX2B — a genuine germline predisposition gene (with Hirschsprung disease and congenital central hypoventilation) — appears only inside a quoted snippet, not as a curated gene.
  • Adrenergic vs mesenchymal cell-state plasticity, the dominant conceptual advance in the field.
  • Spontaneous regression — called a defining feature in the entry's own description, then never modeled, despite being the mechanistically interesting thing about neuroblastoma (TrkA/NGF-dependent apoptosis, telomere maintenance failure).
  • Opsoclonus-myoclonus syndrome, the classic paraneoplastic presentation; also Horner syndrome and dumbbell-tumor spinal cord compression. VIP diarrhea is present, so the paraneoplastic category was considered.
  • MIBG therapy and ALK inhibitors as treatments — crizotinib is named in prose under the ALK node, and lorlatinib is now in frontline COG trials, yet no ALK-directed treatment is curated even though ALK is modeled as a mechanism.
  • International Neuroblastoma Pathology Classification (Shimada) — favorable vs unfavorable histology by MKI, differentiation, and stromal content. The entire histopathology block is one content-free node reading "Neuroblastoma is a malignant tumor of neural crest origin."

Notably, the entry's own bottom-of-file references list cites papers on noradrenergic/mesenchymal identity transitions, SWI/SNF and cell plasticity, and telomere-maintenance copy-number dosage — all with findings: []. The entry has collected the literature for the biology it is missing and never modeled it.

3.3 Retinoblastoma — the exceptions to the two-hit model are absent

The description asserts flatly that "Biallelic loss of RB1 function is required for tumorigenesis". This is contradicted by the recognized MYCN-amplified, RB1-wildtype retinoblastoma (~2% of cases; unilateral, very early onset, aggressive histology — Rushlow et al., Lancet Oncol 2013). An entry built entirely around two-hit sufficiency should carry its principal counterexample.

Also missing:

  • Trilateral retinoblastoma (intracranial pineal/suprasellar tumor in germline carriers) — clinically critical, and dismech already has a Pineoblastoma entry to link to.
  • The cell of origin is bound to retinal progenitor cell; current evidence favors the maturing cone precursor.
  • That RB1 loss alone is insufficient in humans — progression requires additional events (MYCN, MDM4 gain, BCOR, 1q/6p gain).
  • Intravitreal chemotherapy for vitreous seeds, now standard alongside the intra-arterial route the entry does describe.
  • 13q14 deletion syndrome.

The entry also cites PMID:41567907 — a paper specifically about adjuvant chemotherapy for high-risk histopathologic features after enucleation — solely for the throwaway line "retinoblastoma is the most common intraocular malignancy", while the paper's actual subject (postlaminar optic nerve invasion and massive choroidal invasion driving adjuvant therapy) goes unused. Those high-risk features are themselves absent from the entry.

3.4 Alveolar_Rhabdomyosarcoma — the alveolar architecture is never described

The single histopathology node reads, in full: "Rhabdomyosarcoma is a malignant tumor of mesenchymal origin." The disease is named for its histologic pattern — discohesive cells lining fibrovascular septa in an alveolar-like arrangement, with wreath-like multinucleated giant cells — and that pattern appears nowhere. Also missing: primary site (parameningeal, orbit, genitourinary), which is among the strongest prognostic variables in rhabdomyosarcoma, and IRS grouping/stage.

3.5 Medulloblastoma — no genetics block, no histology, no dissemination

  • No genetic: section at all, for a tumor the entry itself describes as molecularly defined. Predisposition is entirely absent: Gorlin (PTCH1), Li-Fraumeni (TP53), Turcot/FAP (APC), Fanconi anemia (BRCA2/PALB2), and ELP1 — the most common medulloblastoma predisposition gene, ~14% of SHH-MB.
  • No histopathology: — classic / desmoplastic-nodular / MBEN / large-cell-anaplastic are WHO entities with real prognostic weight (MBEN favorable in infants; LCA adverse).
  • Leptomeningeal dissemination / M-stage is the dominant prognostic factor and the reason craniospinal irradiation exists. It appears only as a justification inside the CSI treatment description — no phenotype, no mechanism node.
  • SHH-pathway inhibitors (vismodegib, sonidegib) are named in the description as an active area but not curated as a treatment, despite being the flagship targeted therapy in this disease.

3.6 Hepatoblastoma — strong entry, epidemiologic and clinical gaps

The mechanism graph is the best in the set (see §5). What is missing is clinical:

  • Very low birth weight / extreme prematurity is the strongest established risk factor for hepatoblastoma — absent. Trisomy 18 likewise.
  • AFP interpretation in infancy: physiologic AFP is very high in neonates and declines over the first 6-8 months. Without that, "elevated AFP" as a biomarker is not interpretable in the exact age group this tumor affects.
  • PRETEXT staging appears only inside a treatment description.
  • Only two phenotypes (abdominal mass, hepatomegaly). Missing precocious puberty from β-hCG-secreting tumors, thrombocytosis, anemia, failure to thrive.

3.7 Acute_Lymphoblastic_Leukemia — excellent, with clinical omissions

  • Down syndrome-associated ALL — a major epidemiologic subgroup with distinct biology (CRLF2, JAK2) and markedly different treatment toxicity.
  • Infant KMT2A-rearranged ALL — KMT2A-r exists as a subtype but the infant context (dismal prognosis, distinct biology) is prose only.
  • Tumor lysis syndrome and hyperleukocytosis — the defining acute complications at presentation.
  • TPMT / NUDT15 pharmacogenomics — the flagship pediatric-oncology pharmacogenetic, governing 6-mercaptopurine dosing through two-plus years of maintenance.
  • CNS status (CNS1/2/3), which drives the intrathecal therapy the entry does curate.

3.8 Atypical_Teratoid_Rhabdoid_Tumor — predisposition syndrome unstructured

Rhabdoid tumor predisposition syndrome (RTPS1/RTPS2) — germline SMARCB1 or SMARCA4 alterations in roughly 25-35% of AT/RT, driving genetic counseling and surveillance — appears only in a SMARCB1 gene note and in two unused deep-research reference stubs. No link to the sibling Rhabdoid_Tumor entry for synchronous/metachronous renal disease.

4. Claim–evidence mismatches found by reading snippets against their claims

These are cases where a real, correctly-quoted source does not support the statement it is attached to.

Frequency bands contradicted by their own cited numbers:

  • Alveolar_Rhabdomyosarcoma — Metastatic Disease is FREQUENT (30-79%); the cited snippet reads "Seventeen (13.3%) patients had metastatic disease at diagnosis". That is the OCCASIONAL band.
  • Ewing_Sarcoma — Metastatic Disease is FREQUENT, supported by three snippets that all report survival in metastatic disease ("five-year survival rate to 20% to 30%", "10-30% 5-year event-free survival", "3-year EFS 37.4%"). None reports how often metastasis occurs. Survival percentages were read as frequency percentages.
  • Wilms_Tumor — one snippet ("up to 35% of patients can present with either hematuria, hypertension, fever, or flank pain") is read as FREQUENT (30-79%) for hematuria and hypertension but OCCASIONAL (5-29%) for fever and flank pain. A ≤35% ceiling on the union of four symptoms cannot give 30-79% to two of them.

Numbers with no numeric source:

  • Wilms_Tumor — CTNNB1 "~15%" and WTX/AMER1 "15-20%" both rest on the aggregate "WT1, β-catenin, and WTX together account for about one-third of Wilms tumor cases". IGF2 loss of imprinting "approximately 70%" rests on "closely associated with … many Wilms tumors".
  • Neuroblastoma — MYCN "~20%" and ALK "8-10% / ~50%" rest on snippets containing no percentage at all.

Evidence that argues the opposite of its explanation:

  • Wilms_Tumor genetic[5] — the explanation reads "Confirms frequency and prognostic relevance"; the snippet reads "Combined LOH 1p and 16q has limited impact as a prognostic marker".
  • Wilms_Tumor genetic[7] asserts sensitization to ferroptosis via GPX4 inhibition; the cited snippet mentions neither ferroptosis nor GPX4.

Model-system evidence carrying human clinical claims:

  • Neuroblastoma — PMID:41560679 (human iPSC-derived neural crest cells transplanted into immunodeficient mice; the abstract calls it "an in vitro model") is the sole support for five claims, including MYCN's ~20% frequency, its status as "the strongest independent adverse prognostic factor", and 1p/11q/17q risk stratification. For that last one the snippet reports NF1 loss and 17q gain acquired in the cell model — it says nothing about 1p, 11q, or prognosis. Two of the five carry no evidence_source, so they default to HUMAN_CLINICAL for a cell-culture paper.

Unusable snippets in Ewing_Sarcoma: a large share of snippets are clipped mid-clause — "It is associated in 85% of cases with the", "recruited by the EWS-FLI1 fusion protein to tumor-specific enhancers and", "had longer alleles (>135", "depletion promoted a pro-metastatic phenotype" (no subject named). I checked several against the source abstracts: the underlying science is correct — this is a quoting defect, not fabrication. But a reader cannot confirm support without re-fetching the paper, which is what the snippet exists to prevent.

5. What is genuinely well done

  • Hepatoblastoma has the best mechanism graph: CTNNB1 → β-catenin/YAP1 coactivation → fetal hepatic progenitor state → oncofetal program → NFE2L2 stress adaptation, with the C2 molecular-risk signature and the low-AFP / SMARCB1-rhabdoid diagnostic boundary handled explicitly. It also splits a single paper's evidence into human, mouse, and in-vitro items according to which experiment supports which claim — the correct reading of evidence_source.
  • Acute_Lymphoblastic_Leukemia is the most epistemically careful entry in the KB. Nodes scope their own claims ("does not generalize kinase dependence to every B-ALL subtype"), indirect causal edges are marked PARTIAL, and percentages are deliberately left inside quotations rather than promoted to fields — the exact discipline whose absence causes the §4 errors elsewhere.
  • Ewing_Sarcoma has the deepest mechanistic content in the KB — EWS-FLI1 dosage-sensitive hubs, GGAA microsatellite enhancer creation, germline repeat architecture as a susceptibility modifier, ETV6 counter-regulation, NuRD/CHD4 repression, DHX9/SLFN11 replication-fork biology, STAG2 modification. This is research-grade curation; the defects are at the edges (fusion diversity, snippet quoting), not the core.
  • Alveolar_Rhabdomyosarcoma models drug mechanisms properly, with target_mechanisms linking each agent to the node it acts on, plus discussions carrying open questions and proposed experiments.

6. Remediation status

The seven priority items were worked in order. All seven are now applied on branch claude/pediatric-cancers-kb-review-we40oa, across seven entries.

# Item Status
1 WHO 2021 medulloblastoma classification (§1.1) and cell-type binding (§1.4) Done
2 Stage 4S age cutoff and familial ALK (§1.2, §1.3) Done
3 Osteosarcoma histopathology + biochemical blocks (§3.1) Done
4 Wilms TP53 self-contradiction (§1.6) Done
4b Wilms subtype axis untangling (§2.1) Not started
5 Retinoblastoma two-hit exceptions (§3.3); Neuroblastoma telomere axis (§3.2) Done
6 Three frequency bands contradicted by their own evidence (§4) Done
7 Clipped Ewing_Sarcoma snippets (§4) Partial — the metastasis snippets re-quoted at sentence boundaries; the fusion-biology snippets not yet

Also applied beyond the priority list: the cyclophosphamide CHEBI mismatch (§1.7); the Ewing fusion-diversity and WHO 2020 reclassification gap (§1.5); the Wilms_Tumor evidence items whose explanations contradicted their snippets (§4); the Neuroblastoma orphan ALK node, ALK-directed treatment, and PHOX2B (§3.2); Retinoblastoma trilateral disease (§3.3); and the Osteosarcoma germline-predisposition block, which makes the Retinoblastoma → osteosarcoma second-malignancy cross-reference bidirectional (§3.1).

Remaining, in rough priority order

  1. Wilms_Tumor subtype axes (§2.1) — the COG/SIOP histology systems are still flattened into one namespace alongside laterality, etiology, and age. This is a foreign-key-bearing field, so it needs a migration of every subtype: reference, not just an edit to has_subtypes.
  2. Medulloblastoma (§3.5) — no genetic: block at all (ELP1, PTCH1, TP53, APC, BRCA2/PALB2), no histopathology:, no leptomeningeal dissemination node, no SHH-pathway inhibitor treatment.
  3. Alveolar_Rhabdomyosarcoma (§3.4) — the alveolar architecture the disease is named for is still absent from its one-sentence histopathology node.
  4. Neuroblastoma residual (§3.2) — adrenergic/mesenchymal plasticity, spontaneous regression mechanism, INPC/Shimada histopathology, MIBG therapy, opsoclonus-myoclonus.
  5. Atypical_Teratoid_Rhabdoid_Tumor (§2.2, §3.8) — TYR/SHH/MYC subgroup mechanisms; RTPS1/RTPS2 as structured content.
  6. Retinoblastoma residual (§3.3) — intravitreal chemotherapy, high-risk histopathologic features after enucleation, 13q14 deletion syndrome.
  7. Hepatoblastoma (§3.6) and Acute_Lymphoblastic_Leukemia (§3.7) clinical gaps — both are strong entries with well-defined, additive holes.
  8. Remaining clipped Ewing_Sarcoma fusion-biology snippets (§4).

Method

Every entry was read in full and assessed against current pediatric-oncology knowledge (WHO CNS5 2021, WHO soft tissue 2020, INSS/INRG staging, COG and SIOP protocols). Each evidence item was read together with the claim it is attached to. Cached abstracts were consulted where a snippet's meaning was unclear.

The review itself modified no KB files; the fixes recorded in §6 were applied afterwards on the same branch, each entry re-validated with linkml-validate, linkml-term-validator and linkml-reference-validator, and each carrying a history record under history/disorders/.