Reproducibility & Statistical Rigor
Every result is deterministic, every claim is verifiable. All analyses use seed=42, Bonferroni + BH-FDR correction, and 10,000 bootstrap iterations.
Deterministic Pipeline
The entire pipeline is bitwise-reproducible. Same image → same 16 features, every time, on any machine.
42
Random seed
Fixed
1,407
Images analyzed
Zenodo 10301912
64
Organoids (N)
4 cell lines
0.788
Median IoU
Segmentation
Effect Size Forest Plot
All 27 morphological comparisons (9 features × 3 disease clones vs. wildtype) with 95% bootstrap confidence intervals (10,000 iterations). Faded rows are non-significant (BH-FDR q ≥ 0.05).
Rank-biserial effect size (r) with 95% bootstrap CI
Effect size: rank-biserial correlation (r). Positive r = disease clone > wildtype; negative r = disease clone < wildtype. CIs via percentile bootstrap with 10,000 resamples (seed=42).
Jackknife Stability Analysis
Leave-one-out jackknife on all 23 significant comparisons. Every effect survives removal of any single organoid — zero sign changes across 736 jackknife subsamples.
23/23
Highly stable
std < 0.05, no sign changes
0
Sign changes
Across 736 jackknife subsamples
<0.04
Max jackknife std
All comparisons highly stable
Effect size stability per comparison (jackknife range)
Batch Effects: Negligible
ICC(1) analysis across 2 independent lab preparations (LabA, LabB) shows negligible lab-to-lab variability. Average ICC = 0.012 (max: 0.026).
ICC(1) per Feature
% of variance attributable to lab. All < 5% (negligible). 0/9 features show significant lab effects.
1.2%
Average lab variance (ICC)
98.8% of total variance is biological signal, not batch artifact. Clone effects are 7–14× larger than lab effects in mixed-effects models.
Mixed-Effects Model
Formula: area ~ clone + (1|lab)
A1A effect: +264,309 μm² (p < 10&sup-16;)
B2A effect: -155,646 μm² (p = 2.1×10&sup-6;)
TH2 effect: +518,528 μm² (p < 10&sup-55;)
QC Threshold Robustness
Results are invariant to quality control threshold. Even under strict QC (retaining only 14% of images), all significant comparisons persist.
Standard QC
Sharpness ≥ 1.0, SNR ≥ 1.0
Strict QC
Sharpness ≥ 3.0
Strict QC removes 86% of images but retains all 64 organoids and all 27 significant comparisons. This demonstrates that results are driven by true biological signal, not image quality artifacts.
Feature Independence & Correlation Structure
Spearman rank correlation matrix (N=64 organoids) reveals 3 independent feature clusters matching the 3 PCA components that capture 99.4% of variance.
Size
Area, Perim., Eq.Dia., Major, Minor
→ PC1-Size (71.9%)
Roundness
Circ., Solid.
→ PC3-Roundness (8.2%)
Asymmetry
Ecc., Elong.
→ PC2-Asymmetry (19.4%)
Hierarchical clustering (complete linkage, distance = 1 - |ρ|, cut at 0.2). Features within clusters are highly correlated (|ρ| > 0.8); features across clusters are independent. This validates the PCA-based multi-modal integration strategy on the Multi-Modal page.
FAIR Data Principles
All data supporting ConductScreen is Findable, Accessible, Interoperable, and Reusable.
Source Data
Zenodo 10301912 (public, DOI-persistent)
Live Verification
screen.conductscience.com (all analyses live)
CSV Export
1,407 images × 16 features downloadable
Docker Container
Exact pipeline reproduction, seed=42