Week 5 — TCGA-LAML independent-cohort validation
What we tested
- Recommendation reproducibility: Do the mechanism-prior top combos for TCGA FLT3-mut patients match the mechanism classes recommended for BeatAML FLT3-mut patients (Week 4)?
- Biological validation via OS: Are TCGA patients flagged “driver- positive” (= combo-benefiting) by the mech prior more clinically adverse (shorter OS) than driver-negative patients? If yes, the mech prior is flagging real clinical need even in a cohort that never received the recommended combos.
Scoring approach for TCGA
We apply the mechanism-prior scorer (see combo/mechanism_prior.py) directly
to the 173 TCGA-LAML patients. No Baseline A single-drug MLP transfer — the
two cohorts fit independent PCAs so direct transfer isn’t meaningful. The
mech prior uses mutation features (FLT3, IDH1, IDH2, NPM1, KMT2A), which are
schema-identical across cohorts, plus the curated drug-mechanism matrix.
We attempted to ALSO transfer the BeatAML-learned pair synergy residual, but the median-residual heuristic was dominated by a single ALMANAC pair (Cytarabine+Ruxolitinib) whose very-negative synergy over-rode the mechanism signal for every TCGA patient. Week 5 therefore scores by mech prior only, and reports top-3 recommendation sets per patient rather than a single argsort-dependent winner.
Mutation prevalence (TCGA vs BeatAML)
| Gene | TCGA | BeatAML | Literature expected |
|---|---|---|---|
| FLT3 | 30.1% | 35.2% | ~30% |
| NPM1 | 27.7% | 32.0% | ~30% |
| IDH1 | 9.2% | ~4% | ~8% |
| IDH2 | 10.4% | ~10% | ~12% |
| DNMT3A | 24.9% | 17.5% | ~22% |
Prevalences reproduce well across cohorts. This sanity-checks the TCGA ETL — we’re working with the right patient population.
Top combo recommendations for TCGA (by top-3 vote share)
| Pair | Patients with this in top-3 | Clinical status |
|---|---|---|
| Enasidenib + Quizartinib | 112 | Dual-target (IDH2 + FLT3); combination research ongoing |
| Cytarabine + Enasidenib | 108 | Chemo + IDH2i |
| Enasidenib + Midostaurin | 96 | IDH2i + FLT3i |
| Azacytidine + Midostaurin | 47 | HMA + FLT3i — matches AZA-Midostaurin trials (ASP2215-ACE) |
| Azacytidine + Gilteritinib | 47 | HMA + FLT3i — matches LACEWING trial |
| Midostaurin + Venetoclax | 47 | FLT3i + BCL2i — mechanism match with BeatAML’s Quiz+Ven |
| Enasidenib + Venetoclax | 14 | IDH2i + BCL2i — matches ENAVEN trial (published 2022) |
| Cytarabine + Ivosidenib | 13 | Chemo + IDH1i |
| Azacytidine + Ivosidenib | 11 | FDA-APPROVED combination (2022) |
| Ivosidenib + Venetoclax | 11 | IDH1i + BCL2i — matches clinical trial |
Reproducibility at the mechanism-class level: BeatAML’s top pick (Quizartinib + Venetoclax = FLT3i + BCL2i) shows up in TCGA as Midostaurin + Venetoclax (same mechanism class). Several TCGA top picks match published AML clinical-trial combinations (AZA+Ivosidenib is an FDA-approved combo for newly-diagnosed IDH1-mut AML; Aza+Gilteritinib is in LACEWING; Enasidenib+Venetoclax in ENAVEN). This is strong biological face validity.
OS analysis
Overall survival comparison between mech-prior-flagged and mech-agnostic groups:
| Group | n | Median OS (months) | Deceased |
|---|---|---|---|
| Driver-positive (any of FLT3/IDH1/IDH2/NPM1/KMT2A mut) | 86 | 9.50 | 67.4% |
| Driver-negative | 75 | 12.03 | 60.0% |
| FLT3-mutant | 47 | 8.05 | 66.0% |
| FLT3-wild-type | 114 | 13.53 | 63.2% |
Mann-Whitney p-value for driver+ vs driver- OS distributions: p=0.065.
Not a formal log-rank test (no censoring correction), but the direction is clinically expected: driver-mutated AML patients have shorter survival than driver-negative patients. The mechanism prior is flagging exactly the population that has highest clinical need for more aggressive therapy.
What this doesn’t prove (and why)
Week 5 shows that our recommendation system identifies the clinically right subgroup (driver-mutated AML, with shorter OS) and suggests biologically reasonable combos (matching published AML clinical-trial combinations).
Week 5 does NOT prove our combos would ACTUALLY extend OS if given — because TCGA patients received conventional chemotherapy, not the modern targeted combos we recommend (Quizartinib approved 2023, Gilteritinib 2018; TCGA cohort collected 2002-2012). To show combo-recommendation → longer OS would require a prospective trial (or retrospective data where patients DID receive the recommended combos).
Verdict
Week 5 validates the consistency and plausibility of the combo- recommendation logic in an independent cohort. Combined with Week 4’s FLT3-mut Δ = +17 AUC finding, the overall evidence supports the pre-registered “partial yes” outcome:
Mechanism-aware combination prediction for AML converges on clinically- rational combinations and flags the correct high-risk patient population (driver-positive), matching both literature-validated trial designs and FDA-approved combination therapies.
Outputs
runs/tcga_validation/
├── tcga_top_combos.csv # per-patient top-3 mech-prior combos
└── tcga_validation_summary.json # summary stats + OS breakdown