The setup
The challenge and the data
Identify pairs of cell-surface markers that discriminate malignant prostate cells from healthy human cell types, and make the result checkable by requiring the method to recover a pair whose answer is already known.
Background
A major limitation of antigen-directed therapy in solid tumors is on-target, off-tumor activity. A single tumor antigen is rarely absent from every healthy tissue, so CAR-T, T-cell engagers, and antibody-drug conjugates directed at one antigen carry a normal-tissue liability. PSMA, for example, is a well-established prostate antigen and is also expressed in the duodenum.
Combinatorial, logic-gated targeting requires two conditions before a cell is engaged and can recover the specificity a single antigen cannot provide. Prostate adenocarcinoma is a suitable test bed because a combinatorial benchmark exists: PSMA and PSCA are a preclinically validated split-signal AND-gate CAR pair, so a systematic scan can be checked against a known answer before it nominates anything new.
The question
Across a curated panel of 29 cell-surface markers, which pairs separate malignant prostate cells from every healthy human cell type, under AND logic or NOT logic, when scored per patient rather than on pooled cells? And does the scan recover PSMA-PSCA before it proposes anything new?
The data
| Source | Cells | Donors | What it contributes |
|---|---|---|---|
| Localised prostate tumor cohort hormone therapy naive adenocarcinoma | 68,322 | 24 | 3,590 malignant cells after doublet removal, plus 16,249 adjacent-benign prostate cells kept as a matched same-patient control. Malignant identity is the authors' copy-number and signature call. |
| Tabula Sapiens 2.0 via CELLxGENE Census, release 2025-01-30 | 1,025,717 | — | The safety denominator. 173 cell types across 65 tissue labels, forming 442 donor-replicated tissue-by-cell-type populations. Restricted to the 10x 3' v3 assay to match the tumor chemistry, because a raw-count positivity threshold is not comparable across assays. |
| Human Protein Atlas | — | — | Subcellular localization and normal-tissue RNA, used to qualify surface accessibility and cross-check the normal-tissue liabilities found in the single-cell data. |
How a pair is scored
- Positivity is per cell, not per sample. A cell counts as positive for a marker at a raw count of at least 1. The threshold is swept for sensitivity rather than asserted. Nothing is averaged into a pseudo-bulk value first.
- Coverage is computed within each patient, never pooled. For markers A and B, AND coverage is the fraction of that patient's malignant cells positive for both. The 18 per-patient scores are then summarized by their median and by their lower decile, the coverage floor. Pooling would weight a pair by cell yield rather than by patient, so a donor contributing thousands of cells could determine the score.
- Each pair is evaluated against three references. Malignant coverage; the matched benign prostate from the same cohort, so that a pair expressed across the whole gland is penalised rather than rewarded; and the worst-case co-positive fraction across the healthy reference.
- The normal-tissue reference is donor-robust. Within each tissue-by-cell-type population the fraction is computed per donor and summarized by the median across donors, retaining only populations with at least two donors. A single donor, or a population of a few cells, therefore cannot determine a worst case. Each reported liability carries its supporting donor and cell counts.
- Label-confounded pairs are excluded. The cohort's malignant label is derived partly from expression signatures containing HPN and EPCAM. A pair containing either gene is scored against a label that gene helped define, so it is flagged as confounded, held off the frontier, and not nominated.
- Ranking is a Pareto frontier, not a composite score. The two axes are the per-patient coverage floor and the worst-case extra-prostatic liability. No weighting between them is imposed, because the clinically acceptable trade-off is not established by this data. Only surface-accessible markers are nominatable.
Why liabilities exclude the prostate
Both statistics are computed and both are retained in the tables. The figures report the extra-prostatic statistic. Normal prostate expression may be clinically tolerable in selected advanced-disease contexts where the gland is already treated, whereas every other organ is dose-limiting. The all-organ statistic is the stricter of the two and is given alongside it.
The falsification condition
The positive control is a pass/fail gate set before the scan is run. If PSMA-PSCA had not been recovered, the scoring would be considered wrong and no nomination from it would be trustworthy. Recovery does not, however, establish the safety of any pair: co-detected transcript is not co-expressed protein. The limitations section states what the control does and does not support.