Anchored Historical Controls for Aged Muscle Regeneration Studies

Experiments on aged-muscle repair are small, slow, and expensive. A mouse must be housed for about two years before it counts as "aged," and many labs end up testing a rejuvenating intervention (exercise, a drug, a gene knockdown) with only a handful of animals per arm. Some test it with no concurrent untreated aged arm at all, and compare instead against young mice, against pre-treatment values, or against controls from an earlier experiment. The trouble is that injured muscle heals on its own. A cardiotoxin injection, a snake-venom injury model, sets off a predictable cascade of immune infiltration, stem-cell expansion, and new fiber formation. Without a matched control, an "improvement" may simply be the normal recovery curve, a lab-to-lab measurement offset, or a whole-body side effect of the intervention that has nothing to do with repair.

We propose a way to borrow untreated controls from past experiments safely. The method rests on two anchors. The first is a small concurrent control arm, which ties the borrowed controls to the current experiment. The second is the animal's own uninjured contralateral muscle, the matching muscle in the opposite leg, used as a "negative control outcome." That leg goes through the same lab, batch, staining, and systemic intervention but receives no injury. It therefore reveals measurement drift, and it also reveals any part of the treatment effect that is not specific to regeneration. The claim is that these two anchors together let a study with two or three concurrent aged controls recover the effect estimate of a fully controlled study, with honest error bars. The claim can be tested directly by hiding real controls and checking.

Core Claim

In aged-muscle regeneration, the main threat to a historical comparison is not unadjusted covariates. Mouse covariates are few: age, sex, strain, injury dose, days post-injury. The real threats are between-lab and between-batch outcome drift, and treatment effects that are systemic rather than regenerative. Both can be measured on tissue the experiment already collects. An estimator that models the natural recovery curve from a pooled bank of historical untreated aged controls, removes each study's offset using the uninjured contralateral muscle, and borrows a historical arm only when it agrees with a small concurrent anchor should do three things. It should keep false-positive rates near nominal. It should cut the number of aged control animals needed. And it should separate "regenerates better" from "grew bigger muscles everywhere."

Why This Gap, and What Already Exists

The aging biology motivates the problem. Aerobic exercise suppresses CCN2 secretion from senescent muscle stem cells and boosts muscle regeneration in aged mice shows that aged stem cells secrete the pro-fibrotic factor CCN2. It also shows that treadmill training or the AMPK activator AICAR (a drug that switches on a cellular energy-sensing enzyme) in 20–25-month-old mice raises myofiber cross-sectional area and the number of newly formed fibers, which carry the embryonic myosin marker (eMHC+), after injury. The authors themselves leave the contribution of other cell types and of systemic signaling unresolved. Exercise also enlarges uninjured fibers. A cross-sectional-area gain in injured muscle is therefore ambiguous unless something measures what happened in uninjured muscle of the same animals, and that is exactly what our contralateral anchor does.

A tempting idea is to use the large public single-cell maps of aged muscle as external controls. These are Human skeletal muscle aging atlas, which profiled 90,902 cells and 92,259 nuclei from 17 human donors and 8 mice, and the macrophage-fibrosis atlas in Single-cell and spatial transcriptomics identify a macrophage population associated with skeletal muscle fibrosis. Those resources constrain rather than supply the design. They profile uninjured or dystrophic tissue, not tissue at a given day of repair, so they cannot stand in for the recovery curve. The human atlas also carries a confounder its authors flag: intensive-care donors were mechanically ventilated. We use the atlases only for secondary checks of cell-type composition. They are not counterfactuals.

The methods come from the external-control literature. How to improve the quality of comparisons using external control cohorts in single-arm clinical trials? sets out the estimand-cohort-adjustment workflow (weighting, matching on aggregates). It warns that "unanchored" comparisons, those without a shared reference arm, rest on the untestable assumption that all confounders are observed, and that era and measurement drift break them. Our design answers that warning by building in two anchors. Doubly Protected Estimation for Survival Outcomes Utilizing External Controls for Randomized Clinical Trials supplies the selective-borrowing idea: a penalty that shrinks incompatible external controls away using doubly robust pseudo-outcomes. It targets censored survival times and does not use negative-control outcomes. We adapt it to continuous, time-indexed histology outcomes and add contralateral calibration. Robust Synthetic Control denoises a matrix of donor series before weighting them, and serves as a competing way to combine historical arms.

Representation-learning estimators are the closest "intersection" work. Multi-Task Adversarial Learning for Treatment Effect Estimation in Basket Trials makes groups indistinguishable in a learned feature space to correct selection bias in single-arm cohorts. Dynamic Inter-treatment Information Sharing for Individualized Treatment Effects Estimation generates arm-specific networks from a shared hypernetwork. Both address high-dimensional covariate imbalance with no anchor. We argue that problem is not the binding one here. Finally, Really Doing Great at Estimating CATE? A Critical Look at ML Benchmarking Practices in Treatment Effect Estimation shows that semi-synthetic benchmarks like IHDP favor particular estimator families through their simulation design. That result constrains our evaluation: we validate only on real hidden controls, never on simulated outcomes.

Research Program

The first component is a control bank. It pools untreated aged (and young) animals that received cardiotoxin injury, drawn from archived vehicle arms in participating labs and from published experiments with recoverable per-animal data. Each animal record includes study, batch, age, sex, strain, injury dose, harvest day, and paired measurements on the injured and contralateral muscle. Those measurements are fiber cross-sectional-area distribution, eMHC+ fiber fraction, centrally nucleated fiber fraction, and collagen-positive area. Where available, the record also includes cell-type proportions. Assembling this bank is the main cost and risk of the project. If paired contralateral data are rare in archives, we will collect them prospectively in a multi-lab arm.

The second component is the estimator. For animal $i$ in study $s$, let $Y_i$ be the injured-muscle outcome and $N_i$ the same measurement on the contralateral muscle. Let $m(t, a, x)$ be the natural recovery curve learned from the bank, as a function of days post-injury $t$, age $a$, and covariates $x$. The effect on treated aged animals is estimated as

$$\hat\tau = \frac{1}{n_1}\sum_{i \in \text{treated}} \Big[\, Y_i - m(t_i, a_i, x_i) - \hat b_s - \gamma\,(N_i - \bar N^{\text{ctrl}}_s) \Big],$$

Here $\hat b_s$ is the study's offset from the bank, estimated from its concurrent anchor controls and contralateral residuals. $\bar N^{\text{ctrl}}_s$ is the anchor controls' mean contralateral value. $\gamma$ is the share of the uninjured-muscle shift that carries over to injured muscle; it is fitted in the bank. Setting $\gamma = 0$ estimates the total effect. Fitting $\gamma$ estimates the regeneration-specific effect, and the uninjured term alone estimates the systemic effect. Historical studies enter $m$ through selective borrowing in the style of the doubly protected estimator: a study whose drift-corrected residuals disagree with the anchor is penalized toward zero weight. Confidence intervals come from a study-level bootstrap. The bootstrap resamples whole studies rather than animals, so that between-study drift is counted as uncertainty.

The third component is a validation protocol built entirely from real data. Every study in the bank that has a full concurrent control arm becomes a test case. We hide all but $k$ of its controls ($k$ = 0, 2, 3, 5), estimate the effect with each method, and compare against the full-control estimate. For placebo tests, we split a control arm into fake "treated" and "control" halves, where the true effect is zero.

Baselines

The naive reference point is what single-arm and young-versus-aged designs implicitly do. Treated aged animals are compared to pre-injury or young values, or to an unadjusted pooled historical mean at the same harvest day. It is run on exactly the same held-out studies. We expect it to fail placebo tests badly, and any method that merely matches it has no value.

The concurrent-only estimator uses just the $k$ retained concurrent controls. It is the honest small-sample practice a careful lab would follow. It is unbiased, so beating it means equal bias with narrower intervals at the same nominal coverage, or equivalently matching its precision with fewer aged animals.

Covariate-adjusted borrowing fits regression or inverse probability weighting on age, sex, strain, dose, and harvest day over the whole bank. This is the external-control workflow of How to improve the quality of comparisons using external control cohorts in single-arm clinical trials? applied with the same recovery-curve model class and the same tuning budget as ours. It is the most likely competitor, because mouse covariates are few. Beating it means lower placebo false-positive rates and better coverage when between-lab drift is present.

Robust synthetic control (Robust Synthetic Control) treats each historical study's control time course as a donor series. It denoises the series and weights them to match the anchor. Its paper reports lower error under noise and missingness, but the file gives no decimal values: no reported result found on Lacuna for synthetic control on muscle regeneration data.

The strongest existing method is doubly protected selective borrowing (Doubly Protected Estimation for Survival Outcomes Utilizing External Controls for Randomized Clinical Trials), re-expressed for continuous outcomes, with the same anchor arm but without contralateral calibration. Its reported benchmark is that reaching a 60% probability of study success required 100 trial patients versus 150 for trial-only analysis. This is a sample-size saving on a migraine trial, not a muscle result. For muscle regeneration endpoints, no reported result found on Lacuna. Removing the contralateral term from our estimator yields exactly this baseline, so the comparison doubles as the ablation that isolates our key ingredient. Further ablations drop the recovery-curve model (borrowing raw same-day means) and drop selective borrowing (equal weights). Beating the doubly protected baseline means lower placebo false-positive rate and bias specifically in studies with known lab or batch shifts, at no loss of power elsewhere.

Why not just…?

Why not just run properly sized concurrent controls? When it is affordable, it should be done, and our validation treats full concurrent arms as ground truth. The constraint is that aged mice take roughly two years to produce, as in the 20–25-month cohorts of Aerobic exercise suppresses CCN2 secretion from senescent muscle stem cells and boosts muscle regeneration in aged mice, so arms stay small. The proposal does not replace controls. It measures how many can be safely replaced, and the hidden-control experiment reports this as an effective-sample-size curve. If that curve shows no saving, the honest conclusion is "run the controls," and we will report it.

Why not just regression-adjust historical controls on the recorded covariates? This is a real threat. With only five or so covariates, adjustment might capture most of the confounding. Our bet is that between-lab drift is the dominant unmeasured factor: staining protocols, cardiotoxin lot, image-analysis thresholds. This is the "era effect" that the single-arm external-control literature identifies as fatal to unanchored comparisons. The placebo tests settle the question directly. If covariate adjustment achieves nominal false-positive rates across labs, the contralateral anchor is unnecessary and we will say so.

Why not just compare injured to contralateral muscle within treated animals? That difference measures the effect of injury, not of treatment, and it shifts with harvest day as the muscle recovers. It removes drift but leaves the counterfactual undefined. The contralateral muscle is useful only as a calibration and decomposition signal next to a recovery model and an anchor. The ablation that uses only within-animal differences tests this.

Why not just use adversarial or hypernetwork treatment-effect learners? Multi-Task Adversarial Learning for Treatment Effect Estimation in Basket Trials (ATE error 5.12 vs 5.76 on News-16) and Dynamic Inter-treatment Information Sharing for Individualized Treatment Effects Estimation (out-of-sample error in individual effects, measured as PEHE, cut from 1.32 to 1.12 on IHDP) address high-dimensional covariate imbalance. Their gains were shown on simulated outcomes, which Really Doing Great at Estimating CATE? A Critical Look at ML Benchmarking Practices in Treatment Effect Estimation shows can favor such architectures. Our setting has few covariates and an identification problem, not a representation problem. We will include a tuned multi-task learner on the cell-proportion outcomes only, where dimensionality is higher. We do not expect it to fix drift that no covariate records.

Decisive Checks

Three results would decide the project. The first is a placebo test across all held-out control arms. The anchored estimator must keep the false-positive rate near 5% at $k = 3$, while naive and covariate-only borrowing exceed it in cross-lab cases. If it fails, the method is unsafe and we stop. The second is the hidden-control curve: the number of concurrent aged controls our estimator needs to match the precision of a full-control analysis, compared with the doubly protected baseline. The third is a prospective scientific test. We replicate the AICAR arm of the CCN2 study in aged mice with a full concurrent control arm and paired contralateral harvests, analyze it as if only three controls existed, and check agreement with the full analysis. In the same experiment we estimate how much of the cross-sectional-area gain also appears in uninjured muscle. If a large share does, much of the reported "regeneration" benefit is systemic hypertrophy, and that finding would matter for the aging-muscle field whether or not the estimator saves animals.