Shared cell ancestry inflates apparent skill in a cell-response model
Source check: Source checked. Not independent scientific confirmation.
Finding: In a retrospective audit of STATE, mean correlation on expected-null cases in one donor fell from 0.260 with reference cells sharing ancestry to 0.005 with independently sampled references, against 0.081 for a context-mean baseline. The model also fell below baseline with independent references in three additional donors. The finding concerns apparent predictive skill under the tested evaluation design.
Technology: STATE cell-response transfer model
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
- AI contribution
- The trained STATE transfer model generated cell-response estimates whose accuracy could be compared against overlapping and independent reference samples.
- Test and result
- The audit compared reference overlap conditions, with equal reference sample sizes and the same learned predictions. The primary quantitative comparison uses one donor, with follow-up STATE comparisons across four available donors.
- Limitation / still unknown
- This retrospective audit characterizes the tested data and model, with expected-null cases defined operationally. It does not measure biological benefit or imply all transfer estimates are inflated.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.32962v1 · https://arxiv.org/html/2610.32962v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 21 Nov 2026. Review scope: arXiv v1 front matter, section 2.7, Figure 7, relative baseline comparison and non-operational evaluation of reported model outcomes.
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.