Week 1 Day 5 — End-to-end integration + Week 3 smoke test

What ran

  1. combo_val.integration_check — 20 data-layer health checks across all Week 1 canonical tables. Report in week1_day5_integration_report.md. All checks green.

  2. combo_val.combo.combo_predictor — first end-to-end Week 3 run on the Week 1 data. Trains a 19-drug pair-residual MLP on 186 ALMANAC-HL-60 pairs, applies it to all 13,530 unique drug pairs for 613 BeatAML patients, saves the 613 × 165 × 165 combo-AUC tensor. (Formally Week-3 work; included here as the last smoke-test gate that the Week 1 data is complete and coherent.)

Integration check summary

Check Result
Canonical files exist 13/13 ✓
BeatAML / TCGA feature-schema match 80 cols identical ✓
Drug vocab consistency 0 orphans in strict pairs ✓
Patient IDs unique per cohort 613 BeatAML, 173 TCGA ✓
AUC distribution [0, 286], median 207.5 ✓
Mutations binary 0/1 only across 25 cols ✓
ELN ordinal set {0.0, 0.5, 1.0, 1.5, 2.0} ✓
Week-3 readiness 186 pairs, 19 drugs, synergy std 14.7 ✓
Week-5 readiness TCGA 145/173 patients with ≥1 mutation, deceased 65.9% ✓

Combo predictor first run

CV performance on 186 strict pairs (patient-agnostic; 5-fold random split):

Metric Value
Mean fold val RMSE 12.68 (vs null-baseline RMSE ≈ 14.70)
CV pooled Pearson 0.481
CV pooled Spearman 0.384
Per-fold RMSE range 9.90 – 15.72

The model beats the null-baseline (predicting the cohort mean synergy for every pair) by ~14% in RMSE. Pearson 0.48 is a credible signal given only 186 training pairs; the Spearman-vs-Pearson gap (0.38 vs 0.48) suggests the model captures ranking coarsely but not tail synergy extremes — expected with this little data.

Top rank-1 combo picks (sanity check, not Week 4’s real analysis yet)

Pair Times as rank-1 across 613 patients
Elesclomol + Panobinostat 229
Elesclomol + SNS-032 170
Elesclomol + Venetoclax 63
Elesclomol + Foretinib 57
Elesclomol + Trametinib 16

Elesclomol dominates because Baseline A predicts it has a low single-drug AUC (≈ 85-100) for many patients, and the additive 0.5 * (AUC_d1 + AUC_d2) term picks whichever partner best complements it. Venetoclax and Trametinib are in the top-5 list — these are clinically meaningful AML targets (BCL2 / MEK), so the model isn’t completely ignoring biology.

Week 3 proper will address:

  1. Calibrate synergy_to_auc_scale so that predicted synergies are actually comparable to raw AUC units (not blindly added at scale 1.0).
  2. Add a mechanism-prior term for patients whose driver profile matches specific drug mechanisms (FLT3+BCL2 for FLT3-mutant patients, IDH2+HMA for IDH2-mutant patients, etc.).
  3. Drop patient-agnostic synergy for a patient-modulated residual once we have more than 186 training pairs (or use ablation study to confirm patient features add value for combo adjustment).

Tests

51/51 pass (was 46 before Day 5; added 5 combo-predictor unit tests covering SynergyMLP symmetry, gradient flow, output shape, synthetic data loader, and synthetic end-to-end training).

Week 1 recap

The research-grade data layer for the AML combo-validation project is complete in 5 days:

Day Deliverable Core metric
1 Project scaffold, scratch repo, dependencies —
2 BeatAML ETL + Baseline A (single-drug MLP) per-patient ρ = 0.704 (gate ≥0.40)
3 DrugComb v1.5 AML subset ETL 186 strict combo pairs + 1,603 mono screens
4 TCGA-LAML cBioPortal ETL 173 patients, 145 with mutations, 80-dim schema
5 Integration smoke test + combo predictor end-to-end 20/20 checks ✓; combo RMSE 12.68

Ready to begin Week 2 (single-drug model hardening & expansion) and Week 3 (combo model proper training + calibration).