Skip to content
BranchesTags

Repository files navigation

CAR-T Oncology Trials Monitor

Live app: onc-car-t-trial-monitor.streamlit.app DOI

Sister app to the Rheumatology CAR-T Trials Monitor.

An interactive dashboard that tracks CAR-T, CAR-NK, CAAR-T, and CAR-γδ T clinical trials across hematologic and solid tumors, sourced from the public ClinicalTrials.gov registry.

The app provides a filtered trial list, a three-tier disease hierarchy (Branch → Category → Entity), cascading filters, target-antigen classification, cell-therapy modality annotation, approved-product overlays, geographic mapping (global + per-country drilldown), strategic-landscape analyses, publication-ready figures with provenance-tagged CSV exports, and an auto-generated methods section.

Designed as a research and educational resource — not a medical, regulatory, or decision-support tool.

Tab structure

Tab What it does
Overview At-a-glance: branch chips, disease-hierarchy sunburst, by-category / by-target / by-phase / by-year / by-platform panels, top sponsors, recruitment hotspots, snapshot freshness banner
Geography / Map Regional aggregation (China / NA / EU / APAC / MEA / LatAm), country picker → per-country phase + sponsor distribution, world map with site-level scatter, multi-city drilldown table
Data Filterable trial table with column-level search; per-trial drilldown card
Deep Dive Four sub-tabs: By disease (cohort focus), By target (antigen focus + maturity heatmap + temporal trajectory + co-targeting + sponsor commitment + white-space coverage), By product (per-product portfolio + Gantt + activity rate), Strategic landscape (4 cross-cutting analyses)
Publication Figures 11 oncology-tuned figures (Fig 1-10 + Fig 12; Fig 11 is the PRISMA flow in Methods) with provenance-tagged CSV exports
Methods & Appendix Auto-generated methods narrative + ontology + excluded NCT list + curation queue + validation κ — collapsed into a sub-tab row
About Citation, disclaimer, contact

A sidebar Display options expander controls global chart export format (PNG 5× resolution for slides, or SVG vector for journal / Illustrator) and a high-contrast Tableau-20-based palette toggle.

Features

  • Live-first data loading — ClinicalTrials.gov v2 fetched on first daily visit and cached 24h; reproducibility-pinning expander lets reviewers freeze to any dated snapshot for paper citations
  • Three-tier ontology — Branch (Heme-onc / Solid-onc / Mixed / Unknown) → Category (20 Tier-2 buckets) → Entity (~70 Tier-3 leaves), with basket-trial handling and post-classification normalisation
  • Cascading sidebar filter — Branch → Category → Entity, plus filters for Phase, Antigen target, Status, Product type, Modality, Country, Age group, Sponsor type, and Classification confidence
  • Sidebar Display options — global PNG/SVG export-format radio (drives every chart's modebar download) and a Tableau-20 high-contrast palette toggle for greyscale-printing / stacked-bar legibility
  • Target classifier covering 23 heme antigens (CD19, BCMA, CD20, CD22, CD5, CD7, CD30, CD33, CD38, CD70, CD123, GPRC5D, FcRH5, SLAMF7, CD79b, Kappa LC, FLT3, CLL1, CD147, CD4, CD1a, IL-5, BAFF-R) and 31 solid antigens (GPC3, Claudin 18.2, Mesothelin, GD2, HER2, EGFR, EGFRvIII, B7-H3, PSMA, PSCA, CEA, EpCAM, MUC1, CLDN6, NKG2D-L, ROR1, L1CAM, CD133, AFP, IL13Rα2, HER3, DLL3, CDH17, GUCY2C, GPNMB, FAP, MET, FGFR4, KRAS, NY-ESO-1, PRAME), plus 7 dual- target combos (CD19/CD22, CD19/CD20, CD19/BCMA, BCMA/GPRC5D, BCMA/CD70, HER2/MUC1, GPC3/MSLN)
  • Ligand-CAR convention — for ligand-based CARs (IL3 → CD123, APRIL → BCMA, BAFF → BAFF-R, NKG2D → NKG2D-L), the classifier records the receptor on the tumour cell, NOT the binding-domain ligand on the construct. Documented in the rater UI's TargetCategory help text
  • Named-product short-circuit — 86 product aliases across 15 antigen groups, spanning all 7 FDA-approved CAR-Ts (tisa-cel, axi-cel, brexu-cel, liso-cel, ide-cel, cilta-cel, obe-cel), 5 NMPA-approved (relma-cel, eque-cel, zevor-cel, inati-cel/Yorwida, renikeolunsai/Hicara, pulkilumab/Pulidekai), 1 India/EU-approved (varni-cel/Qartemi), and clinical-stage products including anitocabtagene autoleucel (anito-cel), BMS-986453, KITE-753, NXC-201, WU-CART-007, GC012F, CT041 (satri-cel), MT027, HBI0101, JY231, Anbal-cel
  • Cell-therapy modality classification — 8 mechanistically distinct buckets (Auto / Allo CAR-T, CAR-NK, CAAR-T, CAR-Treg, CAR-γδ T, In vivo CAR, unclear)
  • Sponsor classification with explicit PI detection — Industry / Academic / Government / Other via 8-step hierarchical heuristic
  • PRISMA-style flow + auto-generated Methods section that live-derives antigen lists, counts, LLM-curation stats, and ontology table from the live config/code (no hand-maintained drift)
  • Publication figures (11 in the dedicated tab; Fig 11 PRISMA in Methods, oncology-tuned, NEJM-flat aesthetic):
    1. Temporal trends by branch with two-panel approval-milestone strip (FDA / EMA / NMPA / CDSCO dots filterable via pill chips, hover tooltip surfaces target + indication per approval)
    2. Phase distribution by branch
    3. Geographic distribution + global site-level scatter overlay
    4. Enrollment landscape — 100%-stacked clinical-size buckets + phase × branch + per-trial dot plot
    5. Branch → Category → Entity sunburst
    6. Heme vs solid antigen target panels
    7. Innovation signals — product type + modality over time (with absolute / % share toggle)
    8. Disease × antigen target heatmap (white-for-zero, label-on-shaded-cells)
    9. Antigen × Branch matrix with phase encoding
    10. Solid-tumour antigen frontier (chronological emergence)
    11. Industry sponsor crowding by antigen
  • Strategic Landscape sub-tab in Deep Dive — 4 cross-cutting analyses (antigen first-in-class timeline, sponsor competition matrix, heme-vs-solid maturity gap, target momentum 24mo windows). All views auto-update with the sidebar filters and degrade gracefully when filters narrow the dataset
  • Geography tab — regional aggregation (China / North America / Europe / Asia-Pacific / Middle East & Africa / Latin America / Other), country picker with per-country phase distribution + top-sponsor leaderboard, single merged world map (country choropleth + open-site dot overlay), per-country drilldown with city scatter and multi-city select
  • CSV exports with #-prefixed provenance headers (snapshot date, filter state, row count, API source, classifier git SHA) — readable via pd.read_csv(path, comment="#")
  • Four-layer validation infrastructure:
    1. Locked regression benchmark (tests/benchmark_set.csv + per-axis F1 floor)
    2. Independent-LLM cross-validation (scripts/validate_independent_llm.py — multi-vendor: Gemini / Groq / OpenAI / Anthropic — Cohen's κ + consensus disagreement bucket)
    3. Snapshot-to-snapshot diff (scripts/snapshot_diff.py — categorises reclassifications as expected / hard-listed / unexplained)
    4. Named-product audit (scripts/audit/named_product_audit.py — 46-product knowledge base; verifies every trial mentioning a known product gets the expected target / branch / modality classification. Achieved 97.2% trial-level accuracy on 181 audited trials in the most recent run)
  • External comparator — Methods-text auto-generated section cross-cites the ASGCT/Citeline Q1 2026 Gene, Cell, & RNA Therapy Landscape Report for cross-validation of antigen ranking, with explicit disclosure of the comparator's gaps (no CAR-T-specific geographic / sponsor-type / autologous- vs-allogeneic stratification — gaps the present analysis fills)
  • Inter-rater κ validation app — separate companion app under validation_study/ for blind two-rater Cohen's κ measurement on a locked 200-trial sample
  • Full Impressum, Datenschutz, and citation block for German academic use

Quick start

Local

git clone https://github.com/ptjeong/ONC-CAR-T-Trials-Monitor.git
cd ONC-CAR-T-Trials-Monitor
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt
streamlit run app.py

The app opens at http://localhost:8501.

Streamlit Community Cloud

  1. Fork this repository to your GitHub account.
  2. On share.streamlit.io, create a new app from the repo with app.py as the entry point.
  3. (Optional, for the validation harness only) add one or more LLM API keys under Secrets so scripts/validate_independent_llm.py can run from the deployed instance:

Data source

All trial data comes from ClinicalTrials.gov API v2. The query combines CAR-based cell-therapy terms with oncology condition-search terms (leukemia, lymphoma, myeloma, solid tumor, glioma, hepatocellular, pancreatic, gastric, colorectal, ovarian, breast, prostate, sarcoma, melanoma, neuroblastoma, mesothelioma, carcinoma). Trials whose only indication is autoimmune / rheumatologic are excluded via a curated keyword list (the inverse of the sister rheum app).

TCR-T products (NY-ESO-1 TCRs, MAGE-A4, afami-cel) are out of scope for v1 — they are not strictly CAR-T constructs. NY-ESO-1 is included as a target in the solid-antigen taxonomy because some CAR-T variants exist, and the comparator analyses (ASGCT) report it as a top oncology target.

Classification strategy

The pipeline uses a hybrid four-layer classifier: deterministic rules for the bulk, a calibrated default for inference gaps, named-product short-circuits, and a multi-round LLM validation loop for the residual ambiguous cases.

  1. Rule-based keyword layer — config.py term tables: ENTITY_TERMS, CATEGORY_FALLBACK_TERMS, HEME_TARGET_TERMS, SOLID_TARGET_TERMS, DUAL_TARGET_LABELS, exclusions. Word-boundary regex for all term lengths prevents prefix collisions (EGFR vs EGFRvIII, hodgkin vs non-hodgkin).
  2. Named-product lookup — NAMED_PRODUCT_TARGETS / NAMED_PRODUCT_TYPES map known products (86 aliases across 15 antigen groups) to their disclosed antigen and manufacturing type. Includes ligand-CAR mappings (IL3 → CD123, APRIL → BCMA, BAFF → BAFF-R, NKG2D → NKG2D-L) so the classifier records the tumor receptor, not the construct binding-domain ligand.
  3. Calibrated Autologous default — if a trial is confirmed as CAR-T but no product-type marker surfaces, the pipeline defaults to Autologous, reflecting the dominant modality in the current landscape (~85 % of approvals and ongoing trials). Each assignment records a ProductTypeSource tag so inferred labels are distinguishable from explicit ones.
  4. LLM validation loop — validate.py (Claude Opus) plus a structured batched workflow processes every low-confidence trial:
    • Export curation_loop.csv from the Methods & Appendix tab.
    • Split into batches of ~130 trials.
    • Launch parallel Claude agents, each receiving the batch CSV and an allowed_values.json listing every valid label. Each agent returns a strict-schema JSON array (nct_id, branch, disease_category, disease_entity, target_category, product_type, exclude, exclude_reason, confidence, notes).
    • Merge into llm_overrides.json; pipeline loads:
      • _LLM_OVERRIDES — per-trial reclassification entries (high/medium confidence, not excluded).
      • _LLM_EXCLUDED_NCT_IDS — trials the LLM flagged as off-scope (PRO studies, follow-up registries, bispecifics/mAbs, device trials, out-of-scope indications). Dropped at the PRISMA hard-exclusion stage, alongside the manual list.
    • Precedence rule: LLM overrides win for specific labels; for "punted" labels (Other_or_unknown / CAR-T_unspecified) the named-product lookup wins if it finds a known product. Prevents stale LLM punts from blocking newly-added named products.
  5. Hard-exclusion list — HARD_EXCLUDED_NCT_IDS in config.py for manually curated exceptions.

Every trial carries a ClassificationConfidence label (high / medium / low) combining rule strength, ProductTypeSource, and LLM-override status. Surfaced in the Data tab and the Methods tab's Classifier-confidence chart so users can filter analyses to high-confidence rows only.

Validation infrastructure

Four independent layers of validation ship with the repo:

1. Locked regression benchmark — tests/benchmark_set.csv plus tests/test_benchmark.py. Pivotal CAR-T trials with hand-curated ground truth across every classification axis. F1 floor enforced per axis; CI fails on regression.

python -m pytest tests/test_benchmark.py -v -s

2. Independent-LLM cross-validation — scripts/validate_independent_llm.py. Stratified sample of N trials sent to a non-Claude LLM (Gemini / Groq / OpenAI) for blind re-classification, with Cohen's κ + consensus-disagreement bucket. Breaks the Claude-curates-Claude agreement bias of the original curation tool. Free-tier friendly (gemini-2.5-flash-lite, llama-3.1-8b-instant).

export GEMINI_API_KEY=...    # https://aistudio.google.com/apikey (free tier)
export GROQ_API_KEY=...      # https://console.groq.com           (free tier)
pip install google-genai groq

python scripts/validate_independent_llm.py --n 100             # both providers if both keys set
python scripts/validate_independent_llm.py --n 50 --providers groq

Output goes to reports/independent_llm_validation.md (gitignored) — per-axis κ, plus a Consensus disagreements section listing trials where every reviewer agrees on a label different from the pipeline. That section is the highest-signal triage list.

3. Snapshot-to-snapshot diff — scripts/snapshot_diff.py. Categorises every reclassification between two snapshots as expected (LLM override) / hard-listed / unexplained. The unexplained bucket catches pipeline / config edits with wider blast radius than intended.

python scripts/snapshot_diff.py snapshots/2026-04-24 snapshots/<new-date>

4. Named-product audit — scripts/audit/named_product_audit.py. Maintains a 46-product knowledge base (scripts/audit/known_products.py) covering FDA/NMPA/EMA/CDSCO-approved CAR-Ts plus clinical-stage codenames. For every product, checks that all trials mentioning it in the dataset are classified to the expected target / branch / product type. Achieved 97.2% trial-level accuracy on 181 audited trials in the most recent run; mismatches surface as a structured PARTIAL / MISS report for triage.

python scripts/audit/named_product_audit.py /path/to/trials_view.csv

5. Inter-rater κ companion app — validation_study/app.py is a separate Streamlit app for blind two-rater Cohen's κ measurement on a locked 200-trial pre-registered sample (validation_study/sample_v1.json, sha256-anchored). 8 axes (Branch, DiseaseCategory, DiseaseEntity, TargetCategory, ProductType, SponsorType, Platform, TrialDesign). Token- gated rater URLs; durable submission storage with append-only audit log.

Legacy single-vendor curation (validate.py)

validate.py (Claude-only) generated the entries currently in llm_overrides.json. Kept for historical reproducibility but deprecated — use the multi-vendor harness above for new curation.

External comparator

The Methods tab auto-generated narrative cross-cites the ASGCT/Citeline Q1 2026 Gene, Cell, & RNA Therapy Landscape Report (April 2026). Top oncology antigen targets reported by ASGCT (CD19 154 programs, BCMA 68, CD20 29, CD22 27, HER2 25, claudin 18 24, GPC3 23, KRAS 20, mesothelin 17, CD70 16, NY-ESO-1 15, CD7 15, EGFR 14, CD33 14, B7-H3 14, PRAME 13, MUC1 13, GPRC5D 11, CLEC12A 11) align with the antigens captured by HEME_TARGET_TERMS / SOLID_TARGET_TERMS; KRAS, NY-ESO-1, and PRAME were added in May 2026 to match.

ASGCT does NOT separate CAR-T from broader gene-therapy aggregates and provides no geographic, sponsor-type, or autologous/allogeneic stratification — gaps the present analysis fills via direct CT.gov queries with curated CAR-T classification.

Snapshots

The app defaults to live data with a 24h cache. For paper-citable reproducibility, the sidebar exposes a Reproducibility — pin a frozen dataset expander that:

  1. Saves the current live data as a dated snapshot (trials.csv, sites.csv, prisma.json, metadata.json written to snapshots/<YYYY-MM-DD>/).
  2. Pins any previously-saved snapshot for the rest of the session (sidebar status flips to "Pinned to … · n trials"). Click Unpin to return to live data.

Both data paths (load_frozen for pinned, load_live for default) include mtime-based cache invalidation: edits to config.py / pipeline.py / llm_overrides.json (or to the snapshot CSV directly) are detected on the next render and the Streamlit cache rebuilds automatically.

Publication CSVs include a #-prefixed header block with snapshot date, filter state, row count, source URL, and classifier git SHA — readable via pd.read_csv(path, comment="#"). The classifier-SHA tag means a reviewer downloading the same snapshot through a future code revision can detect classification drift even if the trial set is identical.

Repository layout

Path Purpose
app.py Streamlit UI, filters, tabs, figures, exports (~9,300 LOC)
pipeline.py API fetch, tri-level classifier, PRISMA, snapshot I/O (~1,750 LOC)
config.py Disease ontology, target antigens, named products, exclusions (~780 LOC)
llm_overrides.json Per-trial classification overrides from LLM curation
tests/ Unit tests (test_classifier.py), regression benchmark (benchmark_set.csv + test_benchmark.py), Methods-text drift guards (test_methods_text.py), filter NaN regression (test_filter_completeness.py)
scripts/ validate_independent_llm.py (multi-vendor LLM cross-validation), snapshot_diff.py, backfill_site_geo.py, generate_validation_sample.py
scripts/audit/ Named-product classification audit — known_products.py (46-product knowledge base) + named_product_audit.py
validation_study/ Companion inter-rater κ Streamlit app + locked 200-trial sample (sample_v1.json) + per-rater response-state files
snapshots/<YYYY-MM-DD>/ Reproducible frozen datasets (trials.csv, sites.csv, prisma.json, metadata.json)
reports/ Validation-loop output (gitignored)
docs/internal/ Cross-app sync briefs, port prompts, audit prompts (e.g. LIGAND_CAR_TARGET_FIX_BRIEF_RHEUM.md, ASGCT_Q1_2026_PORT_BRIEF_RHEUM.md, NAMED_PRODUCT_AUDIT_PROMPT.md)
.github/workflows/ CI: pytest matrix on Python 3.11 + 3.12
.github/ISSUE_TEMPLATE/ Quarterly approvals review, bug report, classification correction
validate.py Deprecated — single-vendor Claude curation that produced llm_overrides.json. Kept for historical reproducibility
requirements.txt Pinned Python dependencies (no kaleido — image export uses Plotly's browser-side modebar, configured globally via the sidebar Display options)
LICENSE MIT
CITATION.cff Citation metadata (Zenodo DOI, version, author)
SECURITY.md Vulnerability reporting policy

Citation

Jeong P. CAR-T Oncology Trials Monitor [Internet]. Klinik I für Innere Medizin, Hämatologie und Onkologie, Klinische Immunologie und Rheumatologie, Universitätsklinikum Köln; 2026. DOI: 10.5281/zenodo.19738097. Available from: https://onc-car-t-trial-monitor.streamlit.app Source: ClinicalTrials.gov API v2.

The live app surfaces an auto-populated citation block under the About tab.

License

MIT. Copyright (c) 2026 Peter Jeong, Universitätsklinikum Köln.

Contact

Peter Jeong Universitätsklinikum Köln Klinik I für Innere Medizin — Klinische Immunologie und Rheumatologie ✉ peter.jeong@uk-koeln.de

About

Interactive dashboard tracking CAR-T, CAR-NK, CAAR-T, and CAR-γδ T clinical trials across heme- and solid-onc from ClinicalTrials.gov v2. Sister app to Rheumatology-CAR-T-Trials-Monitor.

onc-car-t-trial-monitor.streamlit.app

Topics

bioinformaticscar-tcell-therapyclinical-trialsclinicaltrials-govdahsboarddata-visualizationhematologyoncologystreamlit

Resources

Readme
MIT license

Security policy

Security policy
Activity

Stars

1 star

Watchers

0 watching

Forks

0 forks
Report repository

Releases

Packages

Contributors

Languages