inconclusive

Do multi-RTK drugs concentrate as the ranker's picks for ZFTA-RELA?

ModeledSpeculative

Question

ZFTA-RELA fusion tumors co-express MERTK + AXL + TYRO3 + MET. The hypothesis predicts multi-RTK drugs (Cabozantinib, Tepotinib, etc.) should dominate the drug ranking for this subgroup, and NOT for PFA.

What we found

No — none of the 5 named drugs pass the subgroup-specificity gate. Cabozantinib is in top-5 for 0% of ZFTA-RELA samples. Tepotinib actually shows up MORE in PFA than ZFTA-RELA. The actual top picks for ZFTA-RELA are CDK4/6, CHK1, and JAK/BTK inhibitors — same as PFA, not subgroup-specific.

Caveat

Our ranker scores drugs in 50-d pathway space. The named RTK-coexpression mechanism is a GENE-level signal that pathway scoring compresses away. A null result here does not refute the biology — it tells us this lens cannot see it.

Next step

Park until raw RNA per sample is materialized into the ETL pipeline. Then re-test with gene-level MERTK/AXL/TYRO3/MET z-scores per sample.

Scroll for the data, methods, and per-sample detail.

For a clinician with a ZFTA-RELA patient

ZFTA-RELA fusion-positive ependymoma is the most common supratentorial subgroup in older children. The standard of care today is maximal safe surgical resection followed by radiation; there is no targeted therapy currently approved for these tumors.

The hypothesis tested here is that ZFTA-RELA tumors co-express multiple receptor tyrosine kinases — MERTK, AXL, and MET — and that a polypharmacology drug hitting several of those kinases at once (cabozantinib, tepotinib) should outperform single-target inhibitors.

What KIRhub's drug-ranker found: none of the 5 hypothesized multi-RTK drugs passed the subgroup-specificity gate. The ranker's top picks for ZFTA-RELA were CDK4/6, CHK1, and JAK/BTK inhibitors — the same families that come up for PFA, not a subgroup-specific signal.

The honest caveat: this is a "we can't see it" result, not a "it isn't there" result. KIRhub ranks drugs against a 50-dimensional pathway-level feature space, and RTK co-expression is fundamentally a gene-level signal that pathway scoring compresses away. A null result at this resolution does not refute the underlying biology. Materializing raw per-sample RNA into the pipeline would let us re-test directly.

Research Use Only — not validated for clinical decisions.

⏸ Inconclusive. This insight is not being iterated on. The pathway-level ranker cannot detect the RTK-coexpression signal the hypothesis predicts; reactivation requires raw RNA materialized in the ETL pipeline (or Parker 2014 ingestion for fusion variant detail). See module hub for the current recommendation.

ZFTA-RELA TAM+MET polypharmacology

Subgroup-specificity check: do the predicted multi-RTK / MET-class drugs concentrate in ZFTA-RELA ependymoma (n=117) relative to PFA (n=123)?

✗Developmental framework: refuted · falsification test

✗ Hypothesis NOT supported by the pathway-reversal ranker

Predicted drugs (Cabozantinib, Tepotinib, Capmatinib, Crizotinib, Gilteritinib) are present in KIRhub's panel but none pass the subgroup-specificity gate. The pathway-level ranker does not surface the multi-RTK polypharmacology signal the hypothesis predicts. Two interpretations: (a) the hypothesis is premature, or (b) the GSVA-based lens misses gene-level RTK-coexpression signal.

Predicted vs actual

ModeledSpeculative

Subgroup gate: top-5 in ≥70% of ZFTA-RELA AND outside top-20 in ≥70% of PFA. Note: Sitravatinib (the alternate) is not in KIRhub's 92-drug panel.

Drug	ZFTA-RELA top-5	ZFTA-RELA median rank	PFA top-5	PFA median rank	Top drivers (ZFTA-RELA)	Passes gate?
Cabozantinibno data (0%)	0.0%	36	0.8%	68	LYN,ABL1,LCK,HCK,FYN	✗ fail
Tepotinibno data (3%)	3.4%	62	13.8%	22	GSK3B,LYN,ABL1,LCK,AXL	✗ fail
Capmatinibno data (0%)	0.0%	51	0.0%	27	EGFR,GSK3B,JAK3,AKT1,CAMK2A	✗ fail
Crizotinibno data (7%)	6.8%	32	0.8%	67	ABL1,LYN,LCK,JAK2,FYN	✗ fail
Gilteritinibno data (9%)	8.5%	45	1.6%	56	EGFR,LYN,ABL1,FYN,LCK	✗ fail

What the ranker picks for ZFTA-RELA — top 25 by top-5 frequency

ModeledSpeculative

Drugs ordered by % of ZFTA-RELA samples (n=117) where the ranker places them in their top-5. PFA columns shown for subgroup-specificity comparison.

Drug	ZFTA top-5	ZFTA top-20	ZFTA median rank	PFA top-5	PFA top-20	Top drivers
Rabusertibno data (0%)	36.8%	48.7%	26	18.7%	46.3%	CAMK2A,CAMK2D,GSK3B,LYN,RAF1
Zanubrutinibno data (0%)	35.9%	44.4%	36	24.4%	28.5%	EGFR,LCK,LYN,FYN,CSK
Palbociclibno data (0%)	32.5%	41.0%	79	56.1%	65.0%	CAMK2A,PLK1,DYRK1B,GSK3B,EIF2AK2
Abemaciclibno data (0%)	31.6%	43.6%	59	36.6%	57.7%	GSK3B,GSK3A,CAMK2A,PIM1,EGFR
Tofacitinibno data (0%)	31.6%	41.0%	62	26.8%	37.4%	JAK1,JAK2,TYK2,LYN,EGFR
Encorafenibno data (0%)	29.9%	36.8%	84	58.5%	63.4%	GSK3B,GSK3A,NLK,LYN,RAF1
Ibrutinibno data (0%)	21.4%	42.7%	31	12.2%	27.6%	EGFR,FYN,LYN,LCK,CSK
Netarsudilno data (0%)	20.5%	47.0%	30	8.9%	43.9%	PRKX,AKT1,ROCK1,LRRK2,AKT3
Quizartinibcalibrated (100%)	18.8%	46.2%	35	8.1%	39.0%	GSK3B,FYN,PDGFRB,FLT3,RET
Entrectinibno data (0%)	17.9%	43.6%	32	11.4%	29.3%	ABL1,FYN,LYN,LCK,JAK2
Tucatinibno data (0%)	15.4%	30.8%	63	18.7%	33.3%	JAK1,JAK2,LYN,GSK3B,LCK
Alpelisibno data (0%)	12.0%	30.8%	52	8.9%	27.6%	EGFR,JAK2,JAK1,ABL1,GSK3B
Mobocertinibno data (0%)	12.0%	33.3%	51	14.6%	30.9%	EGFR,LYN,ABL1,GSK3B,FYN
Futibatinibno data (0%)	10.3%	23.9%	49	8.9%	17.1%	JAK2,JAK1,FGR,FGFR1,RET
Umbralisibno data (0%)	10.3%	38.5%	52	9.8%	27.6%	LYN,JAK1,JAK2,LCK,ABL1
Repotrectinibno data (0%)	9.4%	39.3%	30	0.0%	13.8%	ABL1,LYN,FYN,LCK,JAK2
Tivozanibno data (0%)	8.5%	47.0%	26	0.8%	21.1%	ABL1,LYN,LCK,FYN,JAK2
Gilteritinibno data (0%)	8.5%	22.2%	45	1.6%	6.5%	EGFR,LYN,ABL1,FYN,LCK
Brigatinibno data (0%)	8.5%	21.4%	43	0.8%	7.3%	EGFR,ABL1,LYN,FYN,LCK
Mitapivatno data (0%)	7.7%	25.6%	44	0.0%	18.7%	LYN,ABL1,LCK,PLK1,AKT1
Darovasertibno data (0%)	7.7%	34.2%	68	22.0%	61.0%	GSK3B,NLK,GSK3A,LYN,PIM1
Crizotinibno data (0%)	6.8%	40.2%	32	0.8%	22.0%	ABL1,LYN,LCK,JAK2,FYN
Lenvatinibno data (0%)	6.8%	24.8%	48	1.6%	15.4%	ABL1,LYN,LCK,EGFR,FGFR1
Bosutinibno data (0%)	6.8%	27.4%	41	0.8%	4.1%	EGFR,ABL1,FYN,LYN,LCK
Gedatolisibno data (0%)	6.0%	29.9%	71	17.9%	61.8%	GSK3B,EGFR,FYN,LCK,PRKX

E1 vs E2 stratification (Holland-specific)

ModeledSpeculative

ZFTA-RELA samples bucketed by the e1_e2 clinical attribute (n=117 with reversal-score coverage). Median rank for the 5 named drugs is shown per subgroup — lower is better.

E1 n=103

Cabozantinib	32
Tepotinib	64
Crizotinib	29
Capmatinib	59
Gilteritinib	45
AXL-in-top5 (median)	0.0

E2 n=14

Cabozantinib	63
Tepotinib	32
Crizotinib	69
Capmatinib	25
Gilteritinib	46
AXL-in-top5 (median)	0.0

unknown n=0

Cabozantinib	—
Tepotinib	—
Crizotinib	—
Capmatinib	—
Gilteritinib	—
AXL-in-top5 (median)	—

TAM ternary — drug coverage of MERTK / AXL / TYRO3

MeasuredDerived

Each point is a drug; position is the relative fraction of inhibition_pct contributed to MERTK / AXL / TYRO3 in drug_target_evidence (S6 top-3 inhibitor list). Highlighted drugs are the 5 named picks. n=3 drugs with non-zero TAM-kinase evidence.

AXL ranking proxy — AXL-in-top-5-drivers per ZFTA-RELA sample

ModeledSpeculative

For each ZFTA-RELA sample, count how many of its top-5 ranker drugs list AXL as a primary driver kinase. Structural proxy for AXL-high — coarser than gene-level expression. Histogram is stacked by E1 / E2 / unknown. E1 histogram is not bimodal by the >=25% peak rule — the AXL-splits-E1 claim is not surfaced by this proxy. (Absence of bimodality does not refute it; the proxy is structural, not transcriptional.)

E1 (n=103) E2 (n=14) unknown (n=0)

How this test runs (caveats from the spec)

Per-sample drug rankings come from the Phase 7 pathway_reversal_scores parquet — ranking against the 92-drug KIRhub panel using selectivity-adjusted score.
ZFTA-RELA cohort = fusion = ZFTA_RELA; PFA cohort = EPN_classification = PF-A OR posterior-fossa EPN without PF-B/PF_SE marker.
Our ranker scores at pathway level (GSVA hallmark/KEGG/reactome). RTK co-expression at the gene level (MERTK/AXL/TYRO3/MET z-scores per sample) is not evaluated here — raw RNA expression per gene is not in the ingested parquets. A negative result is therefore consistent with two stories: the polypharmacology mechanism may be invisible to a pathway-level scorer, or the mechanism is genuinely not concentrated in ZFTA-RELA.
AXL bimodality within E1 requires gene-level AXL expression and is deferred until raw RNA is materialized into a per-sample parquet.

# Platform
Saifudeen, A., et al. (2026). KIRhub: a falsification-first research workbench for translational oncology. Nature Biotechnology. https://doi.org/10.1038/s41587-026-03090-8

# Data source
McFerrin, L. G., et al. (2018). Oncoscape: a tool for interactive cancer genomics data analysis. Nature Genetics.
Arora, S., et al. (2026). A pan-pediatric brain tumor reference map: medulloblastoma and ependymoma in shared UMAP space. Neuro-Oncology.

# Datasets used in this insight
Oncoscape pediatric brain tumor compendium (n=1,358)
117 ZFTA-RELA EPN samples (E1/E2 stratified)

# Provenance
insight_id: zfta-rela

Do multi-RTK drugs concentrate as the ranker's picks for ZFTA-RELA?i

✗ Hypothesis NOT supported by the pathway-reversal ranker

Predicted vs actuali

What the ranker picks for ZFTA-RELA — top 25 by top-5 frequencyi

E1 vs E2 stratification (Holland-specific)i

TAM ternary — drug coverage of MERTK / AXL / TYRO3i

AXL ranking proxy — AXL-in-top-5-drivers per ZFTA-RELA samplei

How this test runs (caveats from the spec)

Do multi-RTK drugs concentrate as the ranker's picks for ZFTA-RELA?

Predicted vs actual

What the ranker picks for ZFTA-RELA — top 25 by top-5 frequency

E1 vs E2 stratification (Holland-specific)

TAM ternary — drug coverage of MERTK / AXL / TYRO3

AXL ranking proxy — AXL-in-top-5-drivers per ZFTA-RELA sample