When a patient's own cells are reprogrammed into stem cells and grown into the cell type that fails in their disease, researchers get a living biopsy of an otherwise inaccessible organ — a way to watch that patient's neurons fire or their heart cells contract, and to test drugs directly on their genotype rather than on a mouse's. But almost everything these platforms measure is, statistically speaking, a case report. The most cited mechanistic result in this literature — a pro-arrhythmic phenotype traced to a TTN missense mutation in a single 47-year-old patient's cardiomyocytes, with a 51.64% drop in peak sodium current and a 71.56% drop in L-type calcium current — was established entirely on cells from one donor, and the authors themselves concede it is impossible to separate the variant's effect from that patient's unmeasured genetic background (A pathogenic titin missense mutation in hiPSC-derived cardiomyocytes predisposes to ventricular fibrillation in acute ST-segment elevation myocardial infarction). The same platforms are also stuck in an open-loop workflow: stimulation is pre-programmed, and analysis happens afterward by hand (All-optical electrophysiology for high-throughput functional characterization of human iPSC-derived motor neuron model of ALS).
We propose to fix both problems with one system. The core idea: run a patient's cell line as a sequential randomized N-of-1 trial in the dish, executed by a layered control architecture — a nested stack where a fast reactive loop handles millisecond-scale electrical safety, a mid-level loop schedules randomized treatment-versus-control blocks, and a slow deliberative layer, built from AI diagnostic models, updates the experiment plan — with every "stop the experiment, this drug works on this patient" decision gated by anytime-valid confidence sequences, a statistical device that keeps false-positive rates at 5% no matter how often the accumulating data is inspected. In return, the dish relaxes exactly the constraints that hobble N-of-1 trials in humans: randomization is trivial, compounds can be washed out between blocks, and sister wells give parallel arms, so the open problems of irreversible interventions and missing washout noted in the human N-of-1 literature largely disappear in vitro (Anytime-valid inference in N-of-1 trials).
Core Claim
No existing system closes the loop on patient-derived cells with valid sequential statistics. The all-optical ALS platform (All-optical electrophysiology for human iPSC-derived motor neuron model of ALS) characterizes 1,771 motor neurons at 34 cells per hour but delivers pre-set light pulses and analyzes firing phenotypes offline; the machine-learning cardiac work (Rethinking multiscale cardiac electrophysiology with machine learning and predictive modelling) is purely a passive predictor on rodent cultures, never an actuator; and the anytime-valid machinery (Anytime-valid inference in N-of-1 trials) has never touched a cellular model or a control loop. Our claim is that these three gaps are one gap: without a multi-rate control stack you cannot react to cellular events in real time at all; and once you can react, you are continuously inspecting data, which is precisely the regime where naive statistics collapse — repeated two-sample testing after every treatment block inflates false positives to 60% at a nominal 5% threshold, as reported in the N-of-1 simulations (Anytime-valid inference in N-of-1 trials). The statistical contract and the control architecture must therefore be designed together, and we claim the result is a platform that declares patient-specific drug effects online, with valid error control, using fewer cells than a fixed-size experiment.
The layered architecture is not decoration. The formal theory of multi-rate control shows that splitting deliberation from reactive execution is what lets slow, compute-heavy planners coexist with fast, rigid safety enforcement: on an ANYmal quadruped crossing stepping stones, enforcing safety barriers at both the planning and feedback layers produced zero task failures, while enforcing them at only one layer produced 5 to 6 failures (Towards a Theory of Control Architecture: A quantitative framework for layered multi-rate control). We transplant that decomposition onto an optical electrophysiology rig, where the "plant" is a dish of a patient's neurons.
Research Program
The platform has three layers. The low layer runs at the millisecond scale dictated by the recording hardware (optical voltage imaging resolves about 1–2 ms): it clamps stimulation intensity, enforces hard invariants such as "do not drive the cell into depolarization block," and rejects disturbances like baseline fluorescence drift. The mid-level layer schedules blocks: it randomizes the order of drug and vehicle exposure across sister wells, manages washout intervals, and balances conditions across differentiation batches, addressing the inter-line and batch heterogeneity that the neurodegeneration phenotyping literature flags as the main reason single-patient screens fail to generalize (Phenotyping Neurodegeneration in Human iPSCs). The high-level layer runs per experiment (hours to days): it maintains a population prior calibrated on large hiPSC cohorts such as the 1,047-participant Answer ALS resource (Answer ALS, a large-scale resource for sporadic and familial ALS combining clinical and multi-omics data from induced pluripotent cell lines), uses AI quantification to put the current line's phenotype on the same scale as the prior, and — crucially — is the only layer allowed to change the experiment's objective, so that its inference latency (hundreds of milliseconds per deep-model call) never breaches the low layer's deadlines.
The statistical engine sits between the mid and high layers. Let $Y_i^{(T)}$ and $Y_i^{(C)}$ denote the measured phenotype (say, spontaneous firing rate) in the $i$-th treatment and control block. The running effect estimate and stopping rule are
$$\hat\Delta_t \;=\; \frac{1}{t}\sum_{i \le t}\big(Y_i^{(T)} - Y_i^{(C)}\big), \qquad \text{stop and declare an effect when } |\hat\Delta_t| > w_t,\quad w_t = O\!\Big(\sqrt{\tfrac{\log \log t}{t}}\Big)\,,$$
where the boundary width $w_t$ follows from treating the estimation error as a martingale and applying Ville's inequality, exactly the construction validated in the N-of-1 setting (Anytime-valid inference in N-of-1 trials); the $\sqrt{\log\log t/t}$ form is the law-of-the-iterated-logarithm price of being allowed to look continuously. The same sequences, run against a population prior instead of an in-dish control arm, answer the single-case attribution problem: they bound whether a new donor's phenotype sits outside population noise — the formal test that the TTN study lacked.
The AI-assisted diagnosis literature contributes a specific, evidence-backed component: machine quantification standardizes noisy human and site-level measurement. When 31 pathologists across 10 hospitals scored PD-L1 staining, AI assistance raised inter-observer concordance from an intraclass correlation of 0.674–0.736 to 0.950, with juniors reaching 0.973 (How can artificial intelligence models assist PD-L1 expression scoring in breast cancer: results of multi-institutional ring studies). We use the high-level AI model in the same role — not to replace the assay, but to map phenotype measurements from different batches and labs onto one scale, so that an N-of-1 comparison against a multi-site population prior is apples-to-apples.
Baselines
The naive reference point is what automated platforms do today by default: continuously inspect accumulating data with an ordinary two-sample test and alert when it crosses a threshold. This baseline is directly characterized in the N-of-1 simulations, where checking after every block inflated the false-positive rate to 60% against a nominal $\alpha = 0.05$ (Anytime-valid inference in N-of-1 trials); we will re-measure this rate in our rig's null condition. Beating it means holding the empirical false-decision rate at or below 5% under the null.
The strongest existing competitor is the pre-registered, fixed-size, open-loop experiment, instantiated by the all-optical ALS platform: pre-set optogenetic stimulation, a fixed number of cells per condition, and post-hoc analysis, which is the protocol that established the SOD1 A4V phenotypes (basal firing 1.50 Hz in mutants versus 0.98 Hz in genome-corrected controls, and a 32% higher likelihood of depolarization block under strong drive) at a throughput of 34 cells per hour (All-optical electrophysiology for human iPSC-derived motor neuron model of ALS). We will run this baseline on the same rig, the same donor lines, the same isogenic controls, the same microscope hours and the same model-tuning budget, with the sample size chosen by standard power calculation. Beating it means matching its error control and power while reducing cells-to-decision and calendar time per declared effect, and additionally reducing safety events (depolarization block, photobleaching dropouts) that open-loop protocols cannot prevent at all.
Two ablations isolate our key ingredients. Replacing the martingale boundary $w_t$ with a naive sequential test, while keeping the full control stack, isolates the statistical ingredient. Collapsing the three layers into a single flat controller that runs perception, planning, and actuation in one loop isolates the architectural ingredient; the robotics evidence predicts this collapse matters under hard deadlines (zero versus 5–6 failures with layered versus single-layer safety enforcement on ANYmal, Towards a Theory of Control Architecture), though we must verify it on a biological plant, and the hierarchical-learning literature warns that on simple tasks hierarchy adds only overhead, performing on par with flat baselines (Feudal Graph Reinforcement Learning) — so the ablation must be run at stimulation regimes where latency genuinely binds.
Why not just…?
Why not just pre-register the sample size and use classical statistics? This is the serious objection, and it is partially correct: a fixed-n design with no interim looks is valid. But it is incompatible with any automated platform that monitors continuously — and monitoring continuously is the entire point of closing the loop — and it pays in cells and calendar time, since n must be sized for the worst-case effect. Our null-calibration experiment makes the tradeoff quantitative: same rig, same lines, three stopping rules, and we report false-positive rates alongside cells-to-decision, so the efficiency cost of fixed-n (and the invalidity of peeking) is measured rather than asserted. If fixed-n matches our confidence sequences on cells-to-decision, our claim of practical benefit is wrong and we will say so; the attribution benefit against population priors would survive regardless.
Why not just use CRISPR isogenic controls? Where a single causal variant exists, isogenic correction is the gold standard, and we will use the genome-corrected controls already available in the SOD1 A4V system as one of our sanity arms (All-optical electrophysiology for human iPSC-derived motor neuron model of ALS). But isogenic editing answers a different question than ours: it attributes an effect to one edited variant, while we bound whether a whole donor's response — polygenic background included, as in sporadic ALS or the polygenic risk profiles of Brugada syndrome — sits outside population noise. It is also slow and expensive per line, and it provides no stopping rule for continuous monitoring. The TTN study explicitly lacked isogenic rescue and could not bound population risk at all (A pathogenic titin missense mutation in hiPSC-derived cardiomyocytes); our method is aimed exactly at those cases.
Why not just train a population model on Answer ALS and flag outlier donors? A population model gives a point prediction, not a valid sequential test, and continuous outlier flagging inherits the optional-stopping problem it is meant to solve: an automated monitor that alerts when a donor's predicted trajectory deviates will fire far too often under the null, exactly as the naive sequential test does (60% at nominal 5%, Anytime-valid inference in N-of-1 trials). It is also fragile where models are known to be weak — the ALS digital severity models track speech and limb function (voice AUC 0.86, limb median AUC 0.73) but fail on respiratory decline (A machine-learning based objective measure for ALS disease severity) — and batch effects confound outlier scores across labs (Phenotyping Neurodegeneration in Human iPSCs). Our null experiment will measure the outlier monitor's false-flag rate head-to-head with the confidence-sequence rule on identical data.
Proposed Experiments
Experiment one is null calibration. We run vehicle-versus-vehicle blocks (and mutant-corrected isogenic pairs) across at least six independent differentiations, mirroring the six-batch design of the optical ALS study, and compare empirical false-decision rates for the naive sequential test, the fixed-n test, and the confidence sequences, all at nominal 5%. The decisive prediction, taken from the reported 60% inflation in the N-of-1 simulations (Anytime-valid inference in N-of-1 trials), is that naive continuous testing collapses, while confidence sequences stay at or below 5% — the paper reports they held Type-I error strictly below 0.05 across all time points in its own simulations, and we test whether that transfers to real optical electrophysiology noise.
Experiment two is a real effect with a matched mechanism. The all-optical study traced the SOD1 A4V phenotype computationally to a deficit in slowly activating potassium ($K_V7$) conductance (All-optical electrophysiology), so we test a $K_V7$ channel opener as the treatment arm. Success is declaring the patient-specific correction of hyperexcitability (toward the 0.98 Hz control baseline) with valid error control, using fewer cells than the fixed-n baseline requires at matched power, at the platform's 34 cells-per-hour throughput. We treat this as beating the strongest existing method only if cells-to-decision drops by at least half at equal false-positive rates.
Experiment three is the architecture test: under strong stimulation that drives mutants toward depolarization block, we compare safety events (block rate, photobleaching dropouts) between the layered stack and the flat controller. The robotics precedent is zero versus 5–6 failures (Towards a Theory of Control Architecture); our hypothesis is that the low layer's millisecond clamps prevent what the flat controller's inference latency cannot, but this is the least-cited-in-dish claim and the experiment exists to settle it, not to confirm it.
Experiment four is the N-of-1 attribution endpoint, our answer to the TTN problem: a new donor line with an uncharacterized background is screened against a population prior built from cohort-scale resources (Answer ALS), and the confidence sequences must be able to declare either "patient-specific effect, outside population noise" or "within noise" as data accumulates — the formal boundary between proof of existence and proof of pattern that single-donor studies currently cannot draw.
The known risks are real but bounded: optical recording's 1–2 ms resolution and photobleaching limit how fast and how long the low layer can act (All-optical electrophysiology); drug washout kinetics vary by compound, so block lengths must be validated per drug; and hiPSC derivatives remain developmentally immature (Phenotyping Neurodegeneration in Human iPSCs), which limits disease relevance but not the validity of the within-line statistical comparison. None of these threaten the central testable claim, which is that a layered controller with a martingale stopping rule turns a case-report-in-a-dish into a valid, self-terminating single-subject trial.
References
- A pathogenic titin missense mutation in hiPSC-derived cardiomyocytes predisposes to ventricular fibrillation in acute ST-segment elevation myocardial infarction
- All-optical electrophysiology for high-throughput functional characterization of human iPSC-derived motor neuron model of ALS
- Anytime-valid inference in N-of-1 trials
- Rethinking multiscale cardiac electrophysiology with machine learning and predictive modelling
- Towards a Theory of Control Architecture: A quantitative framework for layered multi-rate control
- Phenotyping Neurodegeneration in Human iPSCs
- Answer ALS, a large-scale resource for sporadic and familial ALS combining clinical and multi-omics data from induced pluripotent cell lines
- How can artificial intelligence models assist PD-L1 expression scoring in breast cancer: results of multi-institutional ring studies
- Feudal Graph Reinforcement Learning
- A machine-learning based objective measure for ALS disease severity