Wave 3 · KHIDI RFP 11-1 [6-3] (p.61–64) · TRL 3 → 5
S-PUB
✓ PASS
Solo Reached: 85/100 ·
+CRO/Clinical: 93/100
Cross-Cohort IIP/IPF Lung Transcriptome Classifier (GSE32537 → GSE24206) & Privacy-Preserving Synth Engine
20-gene ECM/alveolar lung transcriptome classifier trained strictly on GSE32537 (n=130, GPL6244) reaches 0.953 combined holdout AUROC (0.941 zero-shot on external GSE24206, GPL570) and 0.900 DP-Synth TSTR AUROC at ε=2.0 (MIA <= 0.513).
2. Solver-Free Recomputed Benchmark Metrics
Target: Combined & External GSE24206 AUROC >= 0.90 (> MUC5B + 0.05), DP-Synth MIA <= 0.55Cross-Cohort Locked & External Lung Transcriptome AUROC (n=77) (n = 77)
0.9529 AUROC
95% CI [0.8983, 0.9925]
Baseline Comparison (Single-Locus MUC5B Transcript Expression Baseline)
0.8701 AUROC
IIP / IPF Multi-Omics & DP-Synth
Zero-Shot External Cohort GSE24206 AUROC (GPL570, n=23)
0.9412 AUROC
GSE32537 Locked Subject Holdout AUROC (n=54)
0.9464 AUROC
DP-Synth (ε = 2.0) Train-on-Synthetic Test-on-Real AUROC
0.9002 AUROC
Worst-Case Membership Inference Attack (MIA) AUROC (ε=1..8)
0.5134 AUROC
Cross-Platform ECM/Alveolar Marker Directional Concordance
1
Interactive Domain Simulation & Cohort Evidence Canvas
PASSagerLog2Fc
-1.2451
clinicalAction
CONFIRMED_FIBROTIC_IIP_ANTIFIBROTIC_ELIGIBILITY_REVIEW
externalGse24206Auroc
0.9412
mmp7Log2Fc
2.55
predictedIpfProb
0.9306
Mathematical / Biophysical Methodology & Upgrade vs APOSTOLUS v1
Maps verified NCBI GEO probes across Affymetrix HuGene-1_0-st and HG-U133 Plus 2.0 via label-free cohort Z-score harmonization, fits L2-regularized logistic classification on discovery biopsies only, and synthesizes ε-differentially private lung cohorts audited against nearest-neighbor MIA and DCR.
Advancement Over Legacy APOSTOLUS v1
Upgrades APOSTOLUS synth8 onto the EXACT target disease of RFP 11-1 [6-3]: 217 real LTRC IIP/IPF lung transcriptomes (GSE32537) externally validated on independent IPF cohort GSE24206 + DP-TSTR audit.