Built with Claude: Life Sciences hackathon
Research track · 2026
Per-patient discovery of combinatorial cell-surface targets in prostate cancer
A curated panel of 29 cell-surface markers, scanned in single-cell data for pairs that discriminate malignant prostate cells from every healthy human cell type, under AND and NOT logic. Pairs are scored within each patient and then summarized across 18 of them, against a matched benign-prostate control and an assay-matched, donor-robust healthy reference of 1,025,717 cells. The preclinically validated PSMA-PSCA pair is recovered as a positive control before any new pair is nominated.
These are transcript-level, hypothesis-generating rankings, not evidence of therapeutic safety. Establishing a target would require protein-level, same-cell measurement in the tumor and in its worst normal tissues, by CITE-seq, multiplex immunohistochemistry, or dual-colour flow cytometry, together with surface proteomics to confirm epitope accessibility. None of that was performed here.
Challenge & data
The background, the single-cell atlases, and how a candidate pair is scored.
MethodHow I used Claude
The instruction file, the session summary, and how the Claude Science tasks are organised.
ScienceReport
Abstract, methods, results, robustness, and limitations.
ToolExplore with Claude
Colour cells by marker combination and query the committed tables.
Background
Antigen-directed therapy in solid tumors is limited by on-target, off-tumor activity: few tumor antigens are absent from every healthy tissue, so single-target CAR-T, T-cell engagers, and antibody-drug conjugates carry a normal-tissue liability. Logic-gated targeting requires two conditions before a cell is engaged, which can recover the missing specificity.
An AND gate engages a cell only when both markers are present. A NOT gate engages a cell positive for an activator unless a blocker is also present. NOT gates are directed: the activator and the blocker are not interchangeable, so each ordered combination is a distinct hypothesis and is scored separately.
Positive control
PSMA (FOLH1) and PSCA are a preclinically validated split-signal AND-gate CAR pair, so the scan can be checked against a known answer. Each antigen alone is broadly co-detected outside the prostate. Requiring both reduces the worst-case extra-prostatic co-detection to below either single marker.
| Marker requirement | Malignant cells positive median across patients | Normal cells positive highest extra-prostatic population | Cell type |
|---|---|---|---|
| PSMA alone | 0.74 | 0.85 | enterocyte of duodenum |
| PSCA alone | 0.16 | 0.95 | bladder urothelial cell |
| PSMA AND PSCA | 0.10 | 0.15 | enteroendocrine cell of small intestine |
The benchmark is recovered on the reduction in worst-case extra-prostatic positivity, not on coverage. Malignant co-detection of PSMA and PSCA is low (0.10), consistent with limited PSCA transcript detection in this dataset. The panel contained the known pair and the analysis was not preregistered, so recovery is not a blind rediscovery. Against 200 random surface pairs, the benchmark scores higher on extra-prostatic selectivity than 57% of them. That comparison is a sanity check rather than evidence of specificity, since a pair can also score well by covering very few malignant cells.
Nominated pair
Pairs are ranked on a Pareto frontier of the per-patient coverage floor against the worst-case extra-prostatic liability. 9 clean surface-accessible pairs are non-dominated. The pair carried forward is PSMA × STEAP1, an AND gate. It is neither the highest-coverage nor the lowest-liability pair on the frontier: it is the frontier pair with the lowest liability among those with substantial coverage, and it was selected because both antigens have clinical-grade binders. That selection is a translational judgement applied on top of the frontier, not the output of an optimizer.
| What | Value | Read it as |
|---|---|---|
| Malignant coverage, median across patients | 0.69 | 6.7x the benchmark's median co-detection, and only that. Not 6.7x better |
| Coverage floor, lower decile across patients | 0.45 | indicative, not precise: the bootstrap interval over patients is wide |
| Worst healthy tissue outside the prostate | 0.17 | mucus secreting cell. Comparable to the benchmark's 0.15, and marginally worse, not better |
| Malignant minus matched benign prostate | 0.60 | discriminates malignant from benign prostate epithelium in the same patients |
The single markers carry substantially higher liabilities: PSMA reaches 0.85 and STEAP1 0.50 in their respective worst extra-prostatic populations. A median coverage of 0.69 is not obviously adequate for monotherapy, and whether the residual liability is clinically acceptable cannot be determined from transcript data.
Convergence with an independent clinical programme
The scan places PSMA x STEAP1 on the Pareto frontier from the single-cell data alone, with no knowledge of the clinical pipeline. A dual PSMA/STEAP1 antibody-drug conjugate (ABBV-969) independently entered phase 1. That an unsupervised ranking of 29 markers converges on a combination a developer selected independently is corroboration that the scoring identifies translationally credible pairs, alongside the recovery of the PSMA-PSCA benchmark.
Two boundaries on that claim. ABBV-969 binds either antigen, so it supports the targetability of PSMA and STEAP1 individually and does not test AND-gate logic. And the panel is curated from antigens with existing clinical interest, so the candidate pool was already enriched for credible targets. The pair is therefore not a novel target proposal; the contribution is the AND framing and the scoring beneath it.
Limitations
These are transcript-level, hypothesis-generating rankings, not evidence of therapeutic safety. Single-cell RNA measures mRNA, not surface antigen density, extracellular epitope availability, binding affinity, or shedding. Same-cell surface co-expression requires CITE-seq, multiplex IHC, or dual-colour flow to confirm; none was performed here.
The NOT gate is a negative result. No activator-blocker combination in the panel achieved both high malignant coverage and low predicted normal-tissue escape. The blocker-negative call additionally depends on single-cell dropout, which is not proof of absence.
Coverage does not generalize to advanced disease. In an independent cohort, tumor specificity replicates but the nominated pair's coverage is AR-lineage dependent, falling sharply in castration-resistant and neuroendocrine states.
The discovery cohort is a single cohort of localised, hormone-naive disease, and its malignant label is the authors' own signature and copy-number call, adopted rather than re-derived. The report states the full set of limitations.