Borrowing a Cohort: Cell-Line-Anchored Reasoning for Single Cancer Cases

A single patient with advanced cancer sometimes responds astonishingly well to a drug — a 94% tumor shrinkage on low-dose 5-FU and leucovorin in one esophageal cancer case, or a complete, year-long remission on the immunotherapy sintilimab in a bladder cancer patient who could not tolerate platinum chemotherapy. The clinical literature is full of such reports, and they are epistemically stuck: with one patient, you cannot tell whether the outcome reflects a real, reproducible biological sensitivity or a one-off confluence of untracked factors. At the other end of the data spectrum sit the high-throughput cancer cell line screens — hundreds of molecularly profiled cell lines tested against hundreds of drugs — which are abundant but notorious for translating poorly to patients, partly because models trained on them learn tissue-of-origin artifacts and generic drug potency rather than person-specific sensitivity. And in between lies a third problem the other two ignore: the clinical question itself is not fixed. In both cases above, complications (radiation-induced mediastinitis, renal dysfunction) forced an abrupt reformulation of what "successful treatment" even meant, mid-course.

This proposal builds a system that connects all three problems into one loop. The core move is to treat the cell line panel as a borrowed cohort for a patient who has none: given a single patient's molecular profile, retrieve the subpopulation of cell lines that is mechanistically analogous to that patient, use that subpanel — not a patient cohort — as the reference population for judging whether the patient's predicted drug sensitivity is genuinely personalized or just generic potency, and let each new clinical event re-specify which subpanel and which candidate drugs are in play. The design is grounded in the literature on ill-defined, co-evolving problem solving, where partial solutions reveal constraints that force the problem statement itself to change, and in recent evidence that drug-response metrics, not just model architectures, are what block personalization.

Core Claim

The claim has three parts, each addressing a documented gap. First, the personalization signal in drug response comes from relative sensitivity — how a specific sample responds compared to the population — not from raw potency. Overcoming limitations in current measures of drug response to enable AI-driven precision oncology showed this directly: models trained on raw $IC_{50}$ or AUC learn drug-wide potency and ignore omics entirely, and only z-scoring responses across samples forces models to learn sample-specific sensitivity. But at $N=1$, z-scoring across patients is mathematically impossible — there is no patient population. We claim the cell line panel can serve as the normalization population instead. Concretely, for a patient profile $p$ and drug $d$, with $\hat{r}_d(p)$ the predicted response, we define an anchored score over a retrieved subpanel of cell lines $\mathcal{C}_p$:

$$z_d(p) \;=\; \frac{\hat{r}_d(p) - \mu_d(\mathcal{C}_p)}{\sigma_d(\mathcal{C}_p)}$$

where $\mu_d$ and $\sigma_d$ are the mean and standard deviation of response to drug $d$ across the subpanel. A high $z_d$ means "this patient is predicted to be unusually sensitive for a tumor that looks like theirs" — the exact signal the cohort z-score was meant to capture, computed without a cohort.

Second, the retrieval itself must respect cell line specificity. QSAR Studies of PC-3 cell line inhibition activity of TSA and SAHA-like hydroxamic acids shows drug-activity rules are tied to the host baseline, and Enhancing drug and cell line representations via contrastive learning for improved anti-cancer drug prioritization shows that contrastively learned, mechanism-of-action-aware representations reduce cell line memorization. So the analogy engine retrieves cell lines using pathway-level and perturbation-response features rather than raw expression, to avoid retrieving on tissue-of-origin artifacts.

Third, the loop is co-evolutionary. Following A Cooperative Problem Solving System for User Interface Design, where constructing a partial solution makes the situation "talk back" and reveal new constraints, each clinical event (a toxicity, a contraindication, an organ failure) enters the system as a hard or soft constraint that prunes the drug set and re-restricts the retrieved subpanel, and the re-scored candidate list feeds back into what the clinician monitors next. This is what distinguishes the system from every transfer-learning model in the corpus: DeepSADR and GANDALF: Generative AttentioN based Data Augmentation and predictive modeLing Framework for personalized cancer treatment both treat prediction as a static, well-defined problem — fixed gene panel, fixed drug list, fixed label — fine-tuned or generated once, with no mechanism for the specification to change mid-course.

Research Program

We build three components. The analogy engine maps a patient's molecular profile (mutations plus expression, as available) into the feature space of a drug-screening panel such as DepMap or GDSC, and retrieves a mechanistically matched subpanel of cell lines. Retrieval features will use pathway-level aggregations in the spirit of DeepSADR's decomposition of genes into pathways, and contrastively trained drug/cell representations in the spirit of SiamCDR, explicitly to prevent the tissue-of-origin shortcuts documented for models like DeepMeth and UniD. The anchored scorer computes $z_d$ for every admissible drug. The constraint loop ingests the clinical narrative as a stream of events, converts each into a formal restriction on the admissible drug set and the retrieved subpanel (e.g., "renal dysfunction" excludes platinum agents and re-weights toward profiles with intact renal-handling pathways), and re-runs retrieval and scoring after each event — the computational analogue of the reformulation that both the esophageal and bladder cases demanded.

A second research product is the evaluation protocol itself. There is no benchmark for "did this N-of-1 reasoning generalize," just as MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineering observed that curated benchmarks miss the iterative objective-discovery structure of real work. We construct pseudo-N-of-1 protocols from existing cohorts by strict leave-one-out masking, plus a narrative benchmark built from the published case reports, so that the field has a repeatable answer to "does an isolated case reflect pattern or outlier."

Baselines

The naive reference point is the drug-potency predictor: rank drugs by their mean response across all screened cell lines, ignoring the patient profile entirely. This is not a strawman — Overcoming limitations in current measures of drug response reports that such a mean baseline matched or beat pan-drug deep networks on raw $IC_{50}$, which is precisely the failure mode our anchored z-score is designed to escape. Any method claiming personalization must fail to be beaten by this baseline on the anchored metric while also being honest that raw-metric comparisons favor it by construction.

The strongest existing system is DeepSADR, Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction, which transfers from 966 DepMap cell lines to TCGA patients via BRICS drug fragments, pathway-level gene decomposition, and a Set Transformer readout. It reports an AUC of 0.957 on the clinical drug Sorafenib in TCGA versus 0.811 for the next best baseline, an average AUC of 0.856, and a drop to 0.662 when its Adaptive Readout is ablated — numbers that make it the system to beat. The fair comparison requires running DeepSADR in two modes under our protocol: its intended cohort fine-tuning mode (an upper bound that uses the very patient cohorts our N-of-1 setting forbids), and a frozen mode where its cell-line-trained weights are applied to a single held-out patient with no access to other patients, with the same input features, and the same tuning budget as our system. Beating it means exceeding the frozen-mode AUC on the leave-one-patient-out TCGA protocol; approaching its fine-tuned 0.957 on Sorafenib without touching a patient cohort would be the strong result.

The strongest baseline on the personalization metric itself is the drug-specific linear regression with gene selection from Overcoming limitations in current measures of drug response, which achieved Pearson correlations up to roughly 0.93 predicting z-scored AUC in the 43-organoid pancreatic PDO cohort. We will run it in an N-of-1 mode — genes selected and coefficients fit on cell lines and previously seen organoids only, applied to one held-out organoid at a time — and our method must beat it on that footing to justify the retrieval machinery. Finally, the key ablation replaces the anchored z-score with a raw predicted response (no subpanel normalization), and a second abation freezes the specification so no clinical event can re-restrict the drug set; the first isolates whether normalization across the borrowed cohort is doing the work, the second whether the co-evolutionary loop is.

Why not just…?

Why not just prompt an LLM with the case report and a retrieval index of published cases? The corpus contains direct evidence against this. In Case-Based Reasoning Enhances the Predictive Power of LLMs in Drug-Drug Interaction, base LLMs performed poorly on pharmacological prediction, and grounding them in retrieved structured cases produced gains as large as 463% for models like DeepSeek-V3 — gains that came from structured case mechanisms, not free text, and those cases were cataloged drug pairs with known ground truth, not unstructured single-patient narratives. DrugBERT adds that response criteria extracted from clinical text do not transfer across cancer types, and Fine-tuning Zero-shot Large Language Models for Patient-reported Outcomes shows zero-shot clinical reading needs dedicated fine-tuning. We will still run an LLM-plus-retrieval baseline on the narrative benchmark, because if it matches the constraint loop, the loop is over-engineered — but the evidence predicts it will retrieve textually similar cases rather than mechanistically analogous cell line baselines.

Why not just pool all cell lines and patients into one multi-task network? A cross-study analysis of drug response prediction in cancer cell lines is the direct negative result: cross-study $R^2$ rarely exceeded 0.65 even with identical viability assays and fell to roughly 0.31 across different assays, and the multi-task deep network UnoMT improved cross-study $R^2$ by only 0.11 over LightGBM. DeepSADR's own ablations show naive pooling fails to transfer. Pooling addresses neither the normalization problem nor the specification drift; it assumes both away.

Why not just use the simple linear regression the metric paper already validated? This is the real threat, and we treat it as such: that model hit Pearson correlations up to ~0.93 on z-scored AUC in the PDO cohort, and if a version of it trained on cell lines alone matches our anchored scorer on the leave-one-organoid-out protocol, the retrieval machinery and the co-evolutionary loop lose their justification and we should say so. What the linear model lacks by construction is an N-of-1 normalization target — its z-score is computed across the 43 organoids, which the leave-one-out condition removes — and a mechanism for the specification to change when a contraindication arrives. Experiment 2 below is designed precisely to settle the first question, and if the linear model wins it, we will report the anchored protocol as the contribution and drop the model complexity.

Why not just screen more cell lines to close the specificity gap? The cross-study analysis addressed this head on: models trained on a small fraction of cell lines but all drugs performed nearly as well as full-coverage models, while limited drug diversity caused severe degradation — drug diversity, not cell line count, drives generalizability. More screening does not produce the missing object, which is a normalization population matched to one patient.

Proposed Experiments and Decisive Checks

Experiment 1 runs the pseudo-N-of-1 protocol on TCGA (the 555-patient, 5-drug clinical setting used by DeepSADR): each patient is masked from all training and adaptation, evaluated alone, and scored by AUC on binary response. The decisive comparison is our anchored score versus DeepSADR-frozen; the sanity floor is the drug-potency mean predictor. Experiment 2 uses the 43-organoid, 26-drug pancreatic PDO cohort, evaluating prediction of z-scored AUC under leave-one-organoid-out with Pearson correlation as the metric, against the drug-specific linear regression in the same N-of-1 footing. Because the cohort z-score can be computed when the organoid is left in, we additionally test whether our cell-line-anchored $z_d$ correlates with the cohort-computed z-score across the full panel — a direct check that the borrowed cohort reproduces the signal the metric paper says matters. Experiment 3 builds the specification-drift benchmark from the two published case narratives: the system receives the case incrementally, and we measure whether the actually-administered successful drug rises in the ranking after each constraining event and whether contraindicated drugs fall, with the static-specification ablation as the comparator; we are explicit that these are retrospective, few-patient proofs of mechanism, not power-calibrated validation. Experiment 4 addresses the outlier-versus-pattern question quantitatively: for a synthetic patient constructed by injecting a known pathway-sensitivity signal into a held-out profile, the corroboration test asks whether the retrieved subpanel itself shows elevated sensitivity to the pathway-matched drug — the system's answer to "is this patient's response a mechanism or a fluke," which is the question every N-of-1 report in the corpus leaves open.

The decisive checks, in order: whether the anchored score beats DeepSADR-frozen on TCGA at strict N-of-1; whether it beats the linear regression when the cohort is removed; whether it tracks the cohort z-score where both are computable; and whether the ranking genuinely re-forms when the situation talks back.