Warping Animal Biomarker Trajectories onto Human Time

Animal studies measure how a biomarker changes over weeks. Human trials care about how the same biomarker changes over years. Today these two time axes are joined by guesswork, typically a fixed lifespan ratio such as "one mouse month equals three human years." This proposal replaces that guess with a learned, monotone mapping from animal age to human-equivalent age, fit jointly with a human trajectory model. The animal data then serves as a prior that lets us forecast individual human trajectories from fewer human visits. The uncertainty bands on those forecasts are calibrated only on held-out humans, so their coverage guarantee survives even though the animals are a different population.

The payoff is concrete. Longitudinal human cohorts are sparse, with missed visits, dropout, and a handful of measurements per person. Animal cohorts are denser in time but live on a compressed and distorted clock. If the distortion can be estimated and undone, animal data can make up for thin human data. If the distortion turns out to be a simple shift and stretch, that is also worth knowing, because it would justify cheap landmark-based translation. The project is built to tell these cases apart rather than to assume the complex answer.

Core Claim

Cross-species longitudinal transfer should be posed as estimating a time reparameterization together with an amplitude map, not as aligning feature distributions. Concretely, let $f(t)$ be the population-average human trajectory of a biomarker as a function of human age $t$. We model an animal measurement at animal age $u$ as

$$ y_{\text{animal}}(u) = a + b\, f(\tau(u)) + \varepsilon, $$

Here $\tau$ is a monotone increasing warp from animal age to human-equivalent age, $a$ and $b$ absorb differences in measurement scale between species (for example, intracortical recordings versus scalp EEG), and $\varepsilon$ is noise. The warp $\tau$ is parameterized as a monotone spline in log-age, and it contains the affine map $\tau(u) = c + d\log u$ as a special case. We fit $f$, $\tau$, $a$, and $b$ jointly, together with per-subject random effects for humans.

The prediction intervals are built in a way that avoids a known trap. We produce split-conformal bands whose calibration set contains only human subjects. Conformal prediction, a method that turns any point forecaster into intervals with guaranteed coverage when calibration and test cases are exchangeable, keeps its guarantee here because the animal data changes only the point forecaster, never the calibration population.

Two things would be new. The first is an explicit, uncertainty-quantified cross-species developmental or disease clock learned from trajectory shape. The second is a clean test of whether that clock needs to be nonlinear at all.

Where This Sits

Three strands of prior work motivate the gap.

Cross-species evidence for the refinement of intrinsic neural timescales supporting executive system maturation through adolescence measures the autocorrelation window of neural signals, roughly how long activity "remembers" itself. It does this in 161 humans with up to three EEG visits, 14 human intracranial patients, and 8 adolescent macaques. It fits separate smooth age curves per species and finds the same shortening with age. The comparison across species is qualitative, however, and no operator places macaque age on the human axis. That missing operator is exactly what we build.

Canine EEG Helps Human: Cross-Species and Cross-Modality Epileptic Seizure Detection via Multi-Space Alignment shows that dog intracranial EEG can substitute for most labeled human data. With 5% labeled human data plus canine data, it matches training on 20% human data. It achieves this by aligning static feature distributions over short windows, with no notion of time along a disease course. We borrow its framing of animal data as a data-efficiency lever, but we replace distribution matching with time warping.

Cross-species neuroscience: closing the explanatory gap argues conceptually for shared analytical coordinates across species but gives no algorithm.

Three methods supply our tools. DTWNet: a Dynamic Time Warping Network makes dynamic time warping differentiable. We use it as a competitor rather than as our warp, because unconstrained per-pair warping is not a species-level clock. Uncertainty-Calibrated Prediction of Randomly-Timed Biomarker Trajectories with Conformal Bands gives the simultaneous conformal bands for irregularly timed trajectories that we reuse. It normalizes errors by a predicted time-varying standard deviation and derives a rate-of-change bound that flagged 17.5% more mild-cognitive-impairment-to-Alzheimer's progressors than mean slopes. Its guarantee assumes exchangeability within the human population, which is why we never calibrate on animals. Longitudinal Flow Matching for Trajectory Modeling and Learning Multi-Modal Biomarker Representations via Globally Aligned Longitudinal Enrichments are strong single-species trajectory models. Both assume one shared time axis, which is the assumption we relax.

Two further papers constrain the design. Treatment of acute cerebral ischemia using animal models: a meta-analysis finds heterogeneity above 80% across 145 animal stroke studies, driven partly by occlusion duration. Timing choices in animals therefore move effect sizes a lot, which warns against any fixed time conversion. Pancreatic α- and β-cellular clocks have distinct molecular properties and impact on islet hormone secretion and gene expression shows a roughly 4-hour phase offset between cell types within one mouse organ. So a single scalar warp can be wrong even inside one species, and our semi-synthetic benchmark includes phase-type distortions for this reason.

Research Program

The work has three parts.

The first part is a controlled benchmark where the true warp is known. We take human trajectories from ADNI and BLSA, the Alzheimer's and aging cohorts harmonized in the conformal-bands paper. From them we generate synthetic "animal" cohorts by pushing each trajectory through known distortions: an affine log-age warp, a smooth nonlinear warp with an inflection, an amplitude rescaling, a periodic phase offset, denser but shorter sampling windows, and added noise. This tells us whether the estimator recovers $\tau$, and whether the nonlinear family helps only when the truth is nonlinear. That second check guards against a flexible model winning by overfitting.

The second part is a real cross-species test using the intrinsic-timescale data. Human scalp EEG autocorrelation windows define $f$. Macaque LFP autocorrelation windows are the animal source, and $a$ and $b$ absorb the recording modality gap. The human intracranial cohort is a within-species modality check: fitting it should yield a near-identity warp with nonzero $a$ and $b$. This is a useful negative control, because a warp estimator that "finds" large time distortion between two human modalities is confusing amplitude with timing. The macaque data are cross-sectional across ages 3.2 to 5.7, and there are only 8 animals. The warp is therefore identified from population curve shape, not from within-animal change. We will report posterior or bootstrap width on $\tau$ honestly and expect it to be wide. Whether these data can be obtained from the authors is a practical risk. If they cannot, the semi-synthetic benchmark and a second real pairing, sought during the project, carry the claims.

The third part is calibrated forecasting. For each held-out human, we forecast the next-visit biomarker value from earlier visits. We produce 90% simultaneous conformal bands using normalized nonconformity scores computed on human calibration subjects only, and we compare against pooling animals into calibration.

Baselines

The naive reference point is a human-only model: a generalized additive mixed model of the kind used in the intrinsic-timescale paper, fit only to human visits, ignoring animals entirely. Every cross-species method must beat it in the sparse-human regime to justify existing. It gets the same human training subjects, the same random-effect structure, and the same smoothness tuning by nested cross-validation.

The practitioner's default is fixed lifespan-ratio scaling. Animal ages are multiplied by a constant ratio and pooled with human data under the same model, with a species intercept. It is run with the ratio taken from the literature, not tuned on test data. Beating it means lower next-visit error at equal human data.

The strongest simple competitor, and the real threat, is affine log-age registration. Here $\tau$ is restricted to a shift and stretch in log-age, estimated by matching a landmark such as the age at which the curve flattens, or by least squares. It has two parameters and is hard to overfit. The proposal counts as beating it only if the nonlinear warp lowers held-out human error by more than the bootstrap spread of the difference, and only on data where the semi-synthetic results say nonlinearity is detectable at that sample size.

The strongest existing learned method for cross-species transfer is distribution alignment in the style of Canine EEG Helps Human: Cross-Species and Cross-Modality Epileptic Seizure Detection via Multi-Space Alignment. We adapt it to regression by training a trajectory forecaster with a maximum-mean-discrepancy penalty between animal and human features pooled over time slices. It gets the same backbone, compute, and hyperparameter search budget as our model. Its reported number is over 90% AUC for seizure detection, and 5% labeled human data plus canine data matched 20% human data alone. That is a classification result on short windows, and no reported result found on Lacuna for this method on longitudinal trajectory forecasting.

The strongest trajectory-alignment competitor is a DTWNet-style warp, per DTWNet: a Dynamic Time Warping Network, which aligns animal and human mean curves by differentiable time warping without a monotone smooth species-level constraint.

For reference points on the human-only side, Uncertainty-Calibrated Prediction of Randomly-Timed Biomarker Trajectories with Conformal Bands reports nominal 90% simultaneous coverage and the 17.5% progressor-detection gain on ADNI and BLSA. Longitudinal Flow Matching for Trajectory Modeling reports downstream classifier accuracy rising from 71.7% to 80.8% on ADNI1. For the intrinsic-timescale data, no reported numeric forecasting result found on Lacuna.

Three ablations isolate the key ingredients. The first removes the warp while keeping the amplitude map and species intercept. The second removes the amplitude map, forcing $a=0$ and $b=1$. The third calibrates the conformal bands on pooled animals plus humans instead of humans alone. The third ablation should visibly break coverage. If it does not, our argument for in-target calibration is weaker than we claim.

Why not just…?

Why not just use a fixed lifespan ratio? It is free and standard. However, the stroke meta-analysis shows that timing choices in animals drive large heterogeneity in outcomes, and the islet clock paper shows timing offsets that no single ratio captures. Neither result proves the ratio fails for a slow developmental biomarker. The fixed-ratio baseline is included precisely so the data decide. If it matches the learned warp, the conclusion is that the ratio is adequate for that biomarker.

Why not just fit a two-parameter affine registration? This is the alternative most likely to work as well, and we treat it as the null model rather than as a foil. Developmental curves such as autocorrelation-window shortening may well differ across species by a shift and stretch in log-age. With 8 macaques, a nonlinear warp may be statistically indistinguishable from affine. The semi-synthetic benchmark tells us the minimum sample size at which a known nonlinearity of given size is detectable. The real-data comparison is judged against that detection limit. If affine wins or ties, the deliverable becomes a calibrated affine cross-species clock with uncertainty, plus evidence about which data would be needed to detect nonlinearity. That is still a useful and publishable result, and the conformal forecasting layer is unchanged.

Why not just pool animal and human data with feature-distribution alignment, as in the canine seizure work? Distribution alignment makes animal and human features look alike but does not say which animal age corresponds to which human age. For forecasting a person's next visit, that correspondence is the whole point. The pooled-alignment baseline tests this directly. If it matches our model on next-visit error, the explicit clock adds interpretability but not accuracy, and we will say so.

Why not just skip animals and use a stronger human-only trajectory model? Human-only models like flow matching or globally aligned enrichments are strong when human data are plentiful. The claim here concerns the sparse-human regime. We therefore subsample human training subjects to 10%, 25%, 50%, and 100%. We expect the animal prior's advantage to shrink toward zero at full data, and a flat advantage curve would itself be suspicious.

Decisive Checks

Three results would settle the proposal.

On the semi-synthetic ADNI/BLSA benchmark, the learned warp should recover the true $\tau$, with low integrated error in human-equivalent years. It should beat affine registration only on the nonlinear-warp conditions and tie on the affine ones. Winning everywhere would indicate overfitting.

On the real intrinsic-timescale data, the human-intracranial-versus-scalp fit should give a near-identity warp. For macaques, the question is whether animal-informed forecasts cut next-visit root-mean-squared error at 10% and 25% human data relative to the human-only mixed model. We summarize this as the human sample size needed to reach full-data error, the trajectory analogue of the 5%-versus-20% result in the canine seizure work.

Across both settings, human-only calibration should hold 90% simultaneous coverage within sampling error, while pooled calibration should drift. Band width is reported alongside, since narrow but invalid bands are the failure we are guarding against.

If the macaque prior helps and the nonlinear warp beats affine beyond the detection limit, the paper's claim is that trajectory-shaped cross-species clocks are needed. If only the prior helps, the claim is that an affine clock plus in-target calibration suffices. If neither holds, the negative result bounds how much sparse cross-sectional animal data can contribute to human longitudinal forecasting.