Route A — Clinical kit readiness (new-patient pipeline)

What changed

Added a clinical-kit layer that lets a new AML patient’s raw data flow through the full prediction pipeline. Prior to this, the pipeline only worked on the 613 BeatAML patients whose 80-dim feature vectors were already computed and cached.

New src/combo_val/clinical/ package

File Purpose
kit_schema.py KitInput / KitOutput / MutationCall dataclasses — the contract a new patient’s data must satisfy
karyotype_parser.py Parse ISCN karyotype strings → structured flags (complex, -5/-7, del(17p), t(8;21), inv(16), t(15;17), t(9;11))
eln_computer.py Compute ELN 2017 risk category from karyotype + mutation panel (no external dependency)
feature_builder.py build_patient_features_from_raw(rna_counts, kit) → 104-dim vector
kit_predict.py End-to-end: features → MLP predictions → combo scoring → KitOutput
demo_kit_run.py Runnable demo on two synthetic patients

ETL changes (src/combo_val/data/beataml_etl.py)

  1. Persist RNA preprocessor (data/canonical/beataml_rna_preprocessor.joblib):
    • 5000 variance-selected gene symbols
    • 50 × 5000 PCA component matrix (float32)
    • 5000-dim PCA mean vector
    • Clinical column medians for imputation
    • Full feature_columns list (for schema-version handshake)
  2. Extended clinical features: 5 → 29 columns (total feature vector 80 → 104):
    • Demographics: sex, is_relapse
    • Labs (log where skewed): WBC, platelet, Hb, LDH, ALT, AST, albumin
    • FLT3 detail: ITD flag, TKD flag, allelic ratio (was: single mut_FLT3)
    • CEBPA: biallelic flag (was: bundled into mut_CEBPA)
    • Fusions: PML-RARA, KMT2A-r, CBFB-MYH11, RUNX1-RUNX1T1, other (was: unused)
    • Karyotype: complex, monosomy 5 or 7, del(17p) (was: unused)
    • Disease state: prior MDS, prior chemo, is initial diagnosis (was: bundled)

    29-column coverage (of 613 feature patients):

    Group Coverage after BeatAML-median imputation
    Demographics 100% (sex via consensus_sex)
    CBC core (WBC/plt/Hb) 70–80% observed, rest imputed
    LDH / liver enzymes 40–60% observed, rest imputed
    FLT3-ITD flag 99.7% observed
    Fusions 17.8% non-empty → one-hot 5 classes
    Karyotype flags 89.4% parseable

Retraining outcomes

Baseline A retrained on 104-dim features:

Version n_features per-patient ρ MAE Notes
v1 80 0.7036 ± 0.018 35.04 original Week 2
v2 104 0.7005 ± 0.018 35.10 current primary model

Extended features did not improve predictive accuracy — the labs have ~40–60% missingness and median imputation adds noise. The gain is in kit enablement, not in prediction quality. Baseline A remains well above the pre-registered ≥ 0.40 gate.

Week 4 head-to-head re-runs with the v2 model preserve the qualitative finding:

  v1 80-dim v2 104-dim
FLT3-mut Δ +16.67 [14.98, 18.19] +11.60 [9.27, 13.85]
FLT3-mut pct combo wins 89.9% 80.4%
FLT3-wt Δ −14.14 [−15.27, −13.05] −17.55 [−19.34, −15.72]
FLT3-wt pct combo wins 6.2% 10.8%

Effect size for FLT3-mut shrinks slightly under v2. Still highly significant and in the same direction. Top-ranked combos change: v2’s #1 and #2 picks are now Gilteritinib + Venetoclax (115 patients) and Quizartinib + Venetoclax (100 patients) — a cleaner clinical match than v1’s ranking.

Demo output (synthetic patient walk-through)

python -m combo_val.clinical.demo_kit_run produces two reports:

Patient 1 — 45y female, NPM1-mut + FLT3-ITD (AR 0.62):

Predicted ELN 2017: Intermediate   Fitness: fit_for_intensive
Driver flags: FLT3_ITD, NPM1
TOP COMBOS
 1. Gilteritinib      + Venetoclax           predicted AUC = 224.1  (mech +1.03)
 2. Quizartinib       + Venetoclax           predicted AUC = 226.2  (mech +1.03)
 3. Midostaurin       + Venetoclax           predicted AUC = 242.8  (mech +1.03)
CAUTIONS
 ⚠ High LDH (1240 U/L) — elevated TLS risk
 ⚠ FLT3-ITD positive — monitor QTc on quizartinib/gilteritinib

Patient 2 — 72y male, TP53 + complex karyotype (-7, del(5q), +8, t(3;3), del(17p)):

Predicted ELN 2017: Adverse        Fitness: unfit
Driver flags: TP53
CAUTIONS
 ⚠ Low platelets (25 ×10⁹/L) — bleeding precautions
 ⚠ TP53 mutation — conventional induction poorly effective; consider trial enrollment

ELN inference is correct in both cases. The combo picks and cautions are biologically sensible.

Remaining limitations (what the kit still CAN’T do)

  1. Absolute AUC values are not calibrated for out-of-distribution inputs. The synthetic lognormal RNA counts used in the demo produce AUC estimates near 220–290 — much higher than BeatAML patients (60–150 range). For real deployment, RNA-Seq must be processed through the same count / TPM pipeline BeatAML used (STAR + featureCounts → raw counts), not a generic pipeline.

  2. Ex-vivo prediction ≠ clinical response. Route B’s negative finding still stands: even the actual observed ex-vivo cytarabine AUC doesn’t predict 7+3 CR in BeatAML (ROC-AUC 0.53, CI crosses 0.5). The kit predicts ex-vivo cell kill, not clinical remission.

  3. Clinical kit must be validated prospectively on organoid / primary sample measurements, not retrospectively against outcomes patients received under legacy therapy.

  4. Only 8 of the 20 clinically-relevant drugs have mechanism-axis annotation. Midostaurin, Quizartinib, Gilteritinib, Venetoclax, Azacitidine, Cytarabine, Ivosidenib, Enasidenib — good. Trametinib, Selumetinib, Ruxolitinib, Sorafenib, Dasatinib, etc. are NOT annotated and the mech prior contributes zero for them. This is why Trametinib- containing pairs sometimes win via baseline AUC alone.

How a clinician / operator uses the kit

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

# 1. Gather inputs from NGS report, cytogenetics report, CBC, chemistry panel
kit = KitInput(
    patient_id="P-2026-04-23-0001",
    mutations=[
        MutationCall(gene="FLT3", is_ITD=True, allelic_ratio=0.58, vaf=0.43),
        MutationCall(gene="NPM1", variant_type="missense", vaf=0.41),
    ],
    karyotype_text="46,XX[20]",
    wbc=90.0, platelet=35.0, hemoglobin=8.1, ldh=1100.0,
    alt=30.0, ast=38.0, albumin=3.4, creatinine=0.9,
    blast_pct_bm=72.0, blast_pct_pb=60.0,
    age=48, sex="female",
    is_relapse=False, prior_mds=False,
    is_initial_diagnosis=True,
)

# 2. Pass RNA-Seq counts (gene_symbol → count)
rna_counts = pd.read_csv("patient_counts.tsv", sep="\t").set_index("gene")["count"]

# 3. Predict
out = predict_for_patient(rna_counts, kit)

# 4. Format for clinical report
print(pretty_print_kit_output(out))

Tests

80 pass (was 59 before Route A; 21 new clinical-kit tests cover):

Files committed

src/combo_val/clinical/
├── __init__.py
├── kit_schema.py
├── karyotype_parser.py
├── eln_computer.py
├── feature_builder.py
├── kit_predict.py
└── demo_kit_run.py

data/canonical/
└── beataml_rna_preprocessor.joblib     (~1 MB: genes + PCA + medians)

docs/
└── kit_readiness.md                    (this file)

tests/
└── test_clinical_kit.py                (21 new tests)

Next (if user wants more)