Childhood Cancer / CCDI Alignment Project

Overview

Explore how dismech can contribute to the childhood / pediatric / AYA cancer data ecosystem, with a focus on structured disease mechanism curation for pediatric malignancies and direct alignment with the NCI Childhood Cancer Data Initiative (CCDI).

The core opportunity is to use dismech as an interpretive layer on top of cohort-scale pediatric molecular characterization. CCDI and the Molecular Characterization Initiative (MCI) generate large volumes of WES, fusion, methylation, pathology, and multimodal research data; dismech can explain how those alterations connect to pathways, cell states, histopathology, phenotypes, and treatment response through evidence-backed mechanism graphs.

This is a strong fit for the current repo. dismech already contains multiple pediatric tumors and predisposition syndromes, and the schema already supports genetic, pathophysiology, histopathology, datasets, and clinical_trials. The main question is where to focus curation so it becomes maximally useful for CCDI, MCI, and the upcoming pediatric / AYA rare cancer efforts.

The CCDI Pediatric Cancer Landscape

Existing dismech coverage

Confirmed pediatric / AYA malignancy coverage already in the KB

This is already enough to support a credible pediatric pilot. dismech has coverage of multiple tumors with clear driver events or subgroup biology, including RB1 loss, MYCN amplification, WT1 / CTNNB1 pathways, EWSR1 fusions, PAX3 / PAX7-FOXO1 fusions, BCR-ABL1, and medulloblastoma subgroup mechanisms.

Preliminary coverage gaps to validate

This is a first-pass gap list based on current KB contents. It still needs a formal crosswalk against COG disease categories, the CCDI study inventory, and the Rare Cancer Study target list once those targets are explicit.

MCI Integration Strategy

Schema observations

The current schema already includes the major assay types needed for a first MCI alignment pass, including WES, WGS, METHYLATION, PROTEOMICS, METABOLOMICS, and MULTI_OMICS dataset types. That means Tier 1 work can start immediately without schema changes.

The most likely extension points are:

1. A more explicit molecular classification / assay-result pattern for methylation subgrouping, classifier outputs, and assay provenance.
2. A lightweight way to reference image-derived pathology features or embedding assets from RADIANCE.
3. A repeatable export pattern for graph edges linking genomic calls to mechanisms, phenotypes, and treatments.

Cancer Curation Conventions

This project should use Wilms tumor and issue #1198 as the working pattern for pediatric oncology curation.

1. A dismech entry is the mechanism-graph curation unit, not every ontology subclass. Split into separate disease files only when a subgroup has a genuinely distinct causal program, such as pathway-defined medulloblastoma groups or fusion-defined sarcomas.
2. Keep disease_term MONDO-first whenever a suitable MONDO disease class exists. Add disease-level ncit_mappings routinely for pediatric cancer entries so the same entry is grounded in both MONDO and oncology-native NCIT concepts.
3. Model cancer refinements as flat subtype axes rather than nested lattices or one-file-per-subclass. Common axes include histology, stage, laterality, age group, and predisposition context.
4. Treat subtype_term and subtype mappings as ontology grounding only. They should not imply a separate dismech page or a "Not Yet Curated" badge.
5. Use the most specific NCIT term that provides materially better oncology specificity, especially for histopathology, disease/subtype mappings, biomarkers, and cancer procedures or therapeutic concepts.
6. When a subtype axis is introduced, keep it as explicit and as close to closed as practical. For example, hereditary-predisposition-associated should usually be complemented by sporadic, not somatic.

The current Wilms entry already follows this pattern: MONDO-first disease anchor, disease-level NCIT mapping, flat subtype axes, ontology-grounded subtypes that do not imply separate pages, and NCIT-first oncology treatment terms where they are more precise than the generic NCIT parent.

Priority curation targets

AI-ready knowledge export

1. Use claim extraction work, including the planned extraction pipeline in issue #1100, to turn pediatric cancer literature and reports into structured mechanism claims.
2. Normalize disease entries MONDO-first, add NCIT disease/subtype mappings for cancer-specific grounding, and use HGNC, GENO, GO, CL, and NCIT-first oncology intervention terms so the output is interoperable with CCDI resources.
3. Convert curated mechanism chains into graph edges or triples such as alteration -> gene product -> pathway -> cell state -> phenotype / histopathology / treatment response.
4. Join those graphs to MCI molecular findings and, later, RADIANCE morphology embeddings for multimodal pediatric cancer analysis.
5. Provide exports that are easy to consume in AI pipelines, cohort analysis, and validation workflows.

Rare cancer alignment

The upcoming CCDI Rare Cancer Study is a particularly good fit for dismech. Rare disease mechanism curation is already a core strength of the project, and many pediatric rare cancers are exactly the kinds of entities where a mechanism-first knowledge graph is most useful: small cohorts, heterogeneous biology, limited trial evidence, and high value in connecting a molecular event to a plausible downstream disease process.

A practical alignment strategy:

• Start with rare pediatric tumors that have unusually clear mechanism anchors such as gene fusions, chromatin-remodeling defects, or predisposition syndromes.
• Capture the minimum useful interpretive layer for each entity: disease term, driver event, gene product, pathway node, histopathology context, phenotypes, and any linked datasets or trials.
• Use longitudinal concepts already supported by the schema, especially progression and treatment sections, to represent relapse, resistance, and secondary malignancy risks where relevant.

Tasks

Tier 1: Low Effort (audits, mappings, pilot design)

• [ ] Build a coverage matrix comparing current pediatric / AYA entries against COG disease groups
• [ ] Cross-reference current coverage against CCDI ecosystem studies and portals
• [ ] Add a rare-cancer target crosswalk once the Rare Cancer Study list is explicit
• [ ] Draft an MCI-to-dismech field mapping for WES, fusions, methylation, and pathology
• [ ] Select 5 pilot pediatric cancers for deep mechanism curation
• [ ] Audit current pediatric entries for MONDO-first disease anchors, disease-level NCIT mappings, and flat subtype axes
• [ ] Audit current pediatric entries for datasets, clinical_trials, and molecular subgroup completeness

Tier 2: Medium Effort (new entries and backfill)

• [ ] Create Hepatoblastoma
• [ ] Create Medulloblastoma Group 3
• [ ] Create Medulloblastoma Group 4
• [ ] Decide which ependymoma molecular groups merit separate files versus subtype axes
• [ ] Create ATRT as a distinct CNS entry if the existing Rhabdoid Tumor entry is too broad
• [ ] Create at least one pediatric rare leukemia entry (JMML, KMT2A-rearranged infant leukemia, or Ph-like ALL)
• [ ] Backfill existing pediatric entries with MCI-relevant datasets and histopathology
• [ ] Add explicit fusion-to-pathway chains for existing fusion-driven tumors

Tier 3: High Effort (schema, export, multimodal)

• [ ] Decide whether a dedicated molecular classification / assay-result class is needed
• [ ] Prototype AI-ready export of pediatric mechanism graphs as triples or edge tables
• [ ] Link claim extraction output to pediatric cancer graph construction
• [ ] Define how RADIANCE image features or embeddings should be represented or referenced
• [ ] Add cohort-level or case-level interpretation examples showing genomic call -> mechanism graph -> clinical relevance

Tier 4: Aspirational (external alignment)

• [ ] Align curated pediatric mechanism graphs with CCDI Hub discovery concepts and study metadata
• [ ] Evaluate a contribution pathway into CCDI AI / harmonization workflows
• [ ] Package a rare pediatric cancer starter set for the upcoming Rare Cancer Study
• [ ] Explore privacy-preserving or federated knowledge integration patterns for institutional pediatric data

Cross-project synergies

STATUS

Scoping

• [x] Create project file
• [ ] Inventory current pediatric / AYA cancer coverage in the KB
• [ ] Crosswalk current coverage to COG disease categories
• [ ] Crosswalk current coverage to CCDI ecosystem studies
• [ ] Crosswalk current coverage to Rare Cancer Study target entities

Integration design

• [ ] Draft MCI mapping for WES / fusion / methylation / pathology inputs
• [ ] Record schema extension decision for molecular classification and assay provenance
• [x] Record Wilms / #1198 cancer curation conventions for pediatric oncology entries
• [ ] Draft AI-ready export plan for pediatric mechanism graphs
• [ ] Draft RADIANCE integration concept

Execution

• [ ] Select pilot pediatric curation set
• [ ] Curate first missing high-priority pediatric cancer
• [ ] Backfill existing pediatric entries with MCI-aligned datasets and molecular subgroup context

Notes

2026-04-12 (Project Creation)

Key observations:

1. dismech already has a stronger pediatric cancer foothold than the generic cancer project alone suggests, including multiple pediatric solid tumors, sarcomas, CNS tumors, and predisposition syndromes.
2. The current schema is already adequate for a first-pass MCI alignment because it has dedicated genetic, pathophysiology, histopathology, datasets, and clinical_trials sections plus pediatric-relevant dataset types such as WES, WGS, and METHYLATION.
3. The biggest near-term modeling challenge is not fusion biology; it is representing methylation-based subgrouping, assay provenance, and pathology-derived AI features in a way that stays clean and reusable.
4. The Rare Cancer Study is unusually well aligned with dismech because rare, mechanism-rich tumors are where evidence-backed curation is most helpful.
5. The highest-leverage AI contribution is likely a pipeline from claim extraction -> normalized mechanism graph -> pediatric molecular data integration, rather than a standalone model effort.
6. The gap list in this file is intentionally preliminary and should be validated against official CCDI / COG disease inventories before curation priorities are frozen.

2026-04-12 (Cancer Curation Conventions from Wilms / #1198)

Modeling decisions carried forward into this project:

1. Use one disease file per coherent mechanism graph, not one file per NCIT or MONDO subclass.
2. Keep disease_term MONDO-first and add disease-level NCIT mappings routinely for pediatric cancer entries.
3. Model cancer refinements as flat subtype axes such as histology, stage, laterality, age group, and predisposition context.
4. Treat subtype_term and subtype mappings as ontology grounding only rather than signals that a separate dismech page should exist.
5. Use the most specific NCIT term where it gives meaningfully better oncology specificity, especially for cancer procedures, histopathology, and disease/subtype mapping.
6. Use Wilms tumor as the worked example for future childhood-cancer curation, especially when deciding whether a subgroup should be a separate file or a facet within one entry.