Annotation-free discovery of disease-relevant cells in single-cell datasets

Not checked yet

• Complete supplied XML text and captions were read. Original graphics, image-based algorithm content, native PDF, supplemental methods and figures S1–S13, datasets, code, notebooks and cited originals remain unreviewed. Reading algorithm captions does not establish complete pseudocode or native mathematical layout review.
• Annotation-free means expert cell labels are optional; patient disease labels and assumptions about healthy cells are required. Some model parameters are not generally identifiable. The authors acknowledge that systematic batch effects can be indistinguishable from disease effects.
• Leukemia examples include small patient cohorts despite large cell counts. The AML expert-label benchmark outperformed MMIL. In the larger ALL example, only patient-level evaluation was possible because true cell-level residual-disease labels were unavailable. Associations do not establish causal resistant cell lineages or clinical effectiveness.
• The five paired relapse cases reuse the same patients. Diagnostic-to-future-outcome analysis does not establish a registered prospective clinical trial. Tuning and reported comparisons reuse held-out folds; independent nested or external validation was not verified. Synthetic simulations are not additional clinical cohorts.
• Garry P Nolan is one of 12 authors, affiliated with Stanford Pathology. NIH awards U54CA209971 and U19AI100627 are disclosed for him; the first award is printed twice and is one identifier. Original award records were not checked. Public research support is distinct from a secret government approach or classified access.
• The source dates the article August 27, 2025 and its collection August 29. Dataset and software links are availability statements, not project receipt or reproduction. Clinical consent and de-identification are author-reported. No private patient records, controlled datasets, experimental procedures or clinical intervention were collected or performed.