AML Combo Validation

A mechanism-aware, patient-specific combination therapy predictor for acute myeloid leukemia.

Does mechanism-informed combination prediction outperform best-predicted single drug for AML patients? This project answers that question, finds the precision-medicine subpopulation where combinations win, and ships an end-to-end clinical kit pipeline for new patients.

---

Headline findings

Metric	Value
Single-drug baseline (per-patient Spearman)	ρ = 0.704 — gate ≥ 0.40 passed
FLT3-mutant AML (n=179) combo vs best single	Δ = +16.67 AUC units [95% CI 14.98–18.19]
FLT3-mutant win rate	89.9% of patients
FLT3 wild-type (n=434)	Δ = −14.14 (no combo advantage)
Permutation p-value	< 0.001
TCGA-LAML validation	Top picks match published clinical combos
Clinical drugs annotated	20/20
Tests passing	86/86

Top combination recommendations

For FLT3-mutant AML patients, the model’s most-frequent rank-1 picks are:

1. Gilteritinib + Venetoclax (115 patients) — matches NCT03625505, LACEWING rationale
2. Quizartinib + Venetoclax (100 patients) — matches NCT03441555
3. Trametinib + Venetoclax (82 patients) — parallel-pathway MEK + BCL2 block

All align with ongoing or completed AML clinical trials.

---

Documents

📘 Research

• Pre-registered thesis — one-line hypothesis + four acceptable outcomes
• Manuscript draft — full paper with methods, results, discussion
• Week-4 head-to-head summary — Δ distribution + subgroup stats
• Week-5 TCGA independent-cohort validation

🧪 Clinical kit

• Kit readiness report — 104-dim features, ELN computer, top-3 combo API
• RNA-Seq SOP for upstream pipeline — STAR + featureCounts requirements and OOD-detection layers
• Induction-response validation — honest negative finding vs 7+3 CR

🔬 Research pipelines (3+ drug extensions)

• Path A — Clonal coverage × IDA — N-drug scoring via patient-specific clonal decomposition
• Path B — Set Transformer — arbitrary-arity neural architecture
• Path D — Higher-Order Factorization feasibility — tensor-decomposition benchmark
• Route C — Clinical regimen retrieval — trial-matched evidence layer

📂 Drug knowledge base

• Drug mechanism annotation v2 — 32 drugs × 39 mechanism axes

📅 Development log

• Week 1 — Day 1: scaffold
• Week 1 — Day 2: Baseline A single-drug MLP
• Week 1 — Day 3: DrugComb ETL
• Week 1 — Day 4: TCGA-LAML ETL
• Week 1 — Day 5: integration + Week-3 smoke test

---

Using the kit

For a new AML patient:

from combo_val.clinical.kit_schema import KitInput, MutationCall
from combo_val.clinical.kit_predict import predict_for_patient, pretty_print_kit_output
import pandas as pd

kit = KitInput(
    patient_id="P-2026-0001",
    mutations=[
        MutationCall(gene="FLT3", is_ITD=True, allelic_ratio=0.58),
        MutationCall(gene="NPM1"),
    ],
    karyotype_text="46,XX[20]",
    wbc=90, platelet=35, hemoglobin=8.1, ldh=1100,
    age=48, sex="female",
    blast_pct_bm=72,
    is_initial_diagnosis=True,
)

# gene_symbol → raw count, matching BeatAML pipeline (STAR + featureCounts)
rna_counts = pd.read_csv("patient_counts.tsv", sep="\t").set_index("gene")["count"]

out = predict_for_patient(rna_counts, kit)
print(pretty_print_kit_output(out))

Output: ELN 2017 risk class, top-5 combination recommendations with predicted AUC + mechanism scores, top-5 single-drug picks, driver-specific cautions, RNA-Seq QC confidence notes.

---

Data sources

Source	Role	Size
BeatAML 2.0 (Bottomly et al., Cancer Cell 2022)	Training + internal validation	613 patients × 165 drugs × ~55K AUC measurements
DrugComb v1.5 (Zagidullin et al., NAR 2019)	Combination synergy training	186 strict AML pairs (ALMANAC on HL-60)
TCGA-LAML PanCancer Atlas 2018	Independent cohort	173 patients

---

Citation

Manuscript in preparation. For now, cite the repository:

Tom, E. (2026). AML Combo Validation: mechanism-aware combination therapy prediction for AML.
https://github.com/Smugpigeon/AML-Combo-Validation

---

Contact

Erick Tom — ericktom94720@gmail.com

Code on GitHub →