in silico
Phase 1: In Silico Validation
- Objective
- Determine whether a label-free morphological feature set can (a) reconstruct Cell Painting-equivalent morphological profiles with R²≥0.70, (b) yield an MPS with ICC≥0.75 across batches/donors, and (c) predict 14-day wound closure with ρ≥0.70/AUC≥0.75 using existing public and internal datasets, before any wet-lab spend.
- Estimated cost
- €500-2000 (cloud GPU + software licenses)
- Estimated duration
- 4-6 weeks
- Success criteria
- R² label-free vs Cell Painting MPS · ≥0.70 (95% CI lower bound ≥0.50) · (5-fold CV linear regression on held-out scaffolds)
- ICC across batches/donors · ≥0.75 · (Mixed-effects variance components (lme4))
- Simulated power for ρ=0.70 vs 0.50 · ≥0.80 at n=20 · (Bootstrap simulation (10,000 iterations))
- Simulated power for AUC=0.75 vs 0.50 · ≥0.80 at n=20/group · (pROC + DeLong simulation)
- Partial correlation MPS vs fiber diameter · ≤0.30 after donor control · (Partial Spearman on simulated data)
- Go if
- R²≥0.70 AND ICC≥0.75 AND simulated power≥0.80 for both ρ and AUC endpoints
- No-go if
- R²<0.50 OR ICC<0.50 OR simulated power<0.60 at n=40 (hypothesis not viable with label-free imaging)
- Pivot if
- 0.50≤R²<0.70 OR 0.50≤ICC<0.75: pivot to hybrid label-free + minimal fluorescence (e.g., 1 nuclear dye) or increase feature set / retrain with domain adaptation
- Risks
- Public Cell Painting datasets (BBBC021/022) are from cancer cell lines, not scaffold-seeded MSCs — domain shift may invalidate R² estimateProbability: highMitigation: Use only as proxy; prioritize any internal MSC pilot data; apply domain adaptation (CORAL, adversarial training); plan Phase 2 to confirm on real MSCs
- Label-free channels in public datasets are not true QPI/autofluorescence — simulation may overestimate R²Probability: mediumMitigation: Restrict to datasets with brightfield + fluorescence pairs; validate with a small internal QPI acquisition if available
- GAN translation introduces hallucinated features that inflate R²Probability: mediumMitigation: Use cycle-consistency loss + feature-level (not pixel-level) loss; report both pixel and feature R²; ablate GAN vs direct feature regression
- Mixed-effects model on synthetic data yields unstable ICCProbability: lowMitigation: Use parametric bootstrap for CI; require ≥3 batches × ≥5 donors in simulation