非臨床の問い

The primary hypothesis, H1, is that the integrated model more accurately predicts the normal–disease response difference, at the level of unseen donors or cell lines, than morphology-only or lineage-only models. H0 is that there is no incremental performance, and H2 is that information remains after adjusting for state, dose, and culture conditions within the same tissue or lineage.

It will be explicitly stated that disease-label classification, causal mechanisms, transdifferentiation, therapeutic effects, STAP, and proof of new physics are outside the study’s scope.

The priority order is: normal and disease-related cells derived from the same donor; then multiple donors paired under the same tissue and medium conditions; then multiple normal and disease cell lines of the same cell type. Include sham exposure, a gas-only/non-activated device, matched normal–disease controls, positive and negative QC for every assay, and QC for day, operator, and instrument.

Device modality, distance, duration, gas, voltage, and other conditions will be fixed in advance within approved SOPs and safety limits. Temperature, pH, evaporation, and, where needed, a RONS proxy measure will be recorded as dose-, day-, and instrument-level covariates.

The primary endpoint will be the sham-normalized normal–disease response difference for each biological unit, derived from a pre-specified exposure-intensity–response curve. The primary metric will be the normal–disease difference in survival-curve AUC (or a pre-specified equivalent single metric); the IC50 ratio will be limited to supplementary reporting.

Before exposure, morphology will be imaged with magnification, illumination, segmentation, and exclusion rules fixed in advance. Features will include the finite-scale-range $\hat D_F$, lacunarity, boundary roughness, area, circularity, density, and related measures. Image fields or wells will not be counted as independent samples.

Lineage-related features will be limited to provenance and establishment history, clonal composition within the same donor, quality-confirmed somatic variants, and phylogenetic uncertainty. Cell cycle, stress, differentiation, metabolism, density, and transcriptional programs will be handled as a separate block of state features.

Because scRNA-seq is destructive, the study will not claim to have directly predicted pre-exposure lineage and post-exposure survival in the same cell. Lots with the same origin will be split, and the feature-measurement group and response-measurement group will be matched at corresponding clonal or hierarchical units.

For RNA-seq-derived somatic variants, RNA editing, expression bias, allelic dropout, low coverage, mapping errors, and batch effects will be evaluated in advance. Confirmatory lineage features will require agreement with targeted DNA sequencing, WGS/WES, or an appropriate DNA method.

If features are discordant or QC-ineligible, RNA-derived features will be demoted to exploratory annotations. Variant distances between independently established cell lines will not be interpreted as direct ancestor–descendant relationships.

M0 will contain only pre-specified confounders, such as dose, day, instrument, medium, density, and passage number. M1 will be M0 plus morphology; M2 will be M0 plus lineage-related features; and M3 will be M0 plus morphology plus lineage-related features. Adding state features, stratifying by state, and restricting to DNA-validated features will be sensitivity analyses.

The models will be compared using the same outcome, preprocessing, outer splits, and complexity control. Feature selection, missing-value imputation, normalization, and adjustment will be performed only within inner folds. Feature importance will be exploratory and will not be read as a causal effect.

Donors or cell lines will be the minimum unit of splitting; wells, images, time points, doses, RNA/DNA, clones, and frozen vials from the same origin will be fixed in the same fold. Development will use group-level nested cross-validation, followed by temporal internal evaluation in a held-out group collected later.

After the protocol, feature extraction, models, thresholds, and analysis code are frozen, external validation will be performed only once in an unseen group from another time period and, where possible, another operator, instrument, or independent facility. The external data will not be used for retraining.

Primary performance will be the MAE of the within-pair response difference in unseen donors or cell lines. Report the group-specific MAE differences for M3 versus M1 and for M3 versus M2 (negative values indicate improvement), 95% cluster-bootstrap CIs, calibration intercept and slope, and prediction-interval coverage. The minimum practically useful improvement magnitude will be defined in advance.

Independent donors or cell lines will be distinguished from technical replicates of culture days, wells, and fields; culture days will not be counted as donor N. Sample size will be determined by hierarchical simulation using variance, ICC, missingness, and dose response obtained at the preliminary stage; no arbitrary fixed N will be stated.

If device dose or reference specimens are QC-ineligible, biological comparisons will stop and the measurement system will be corrected. If lineage features cannot be DNA-validated, the hypothesis will remain unverified and those features will be excluded from confirmatory analysis. If the difference from sham or the normal–disease difference is below measurement sensitivity, the primary endpoint will not be changed post hoc.

If, in the frozen evaluation, M3 does not outperform the stand-alone models, the CI does not support the minimum practically useful improvement, or calibration is poor, the integrated hypothesis will be concluded to be unsupported and the study will end. Negative results will also be retained and reported.

For human-derived specimens, confirm the scope of consent, secondary use, de-identification, MTA, and required ethics review. CAP devices must meet facility review and SOP requirements for biosafety, electricity, UV, ozone/NOx, gas, exhaust, fire, and waste. No patient exposure or experiments with live HIV virus will be performed.

Deliverables will proceed in sequence: design lock (protocol, SAP, and data dictionary); feasibility QC and a DNA report; a frozen pipeline and development report; a held-out-group report; and external validation with a reproducibility package. If a gate is ineligible, the work will not proceed to the next stage.