Abstract
High-dimensional biological and physiology-adjacent time series are often noisy, partially observed, and heterogeneous across subjects, space, and time, making it difficult to learn continuous-time models that are both stable to integrate and useful for intervention analysis. We propose a reusable latent-dynamics backbone that couples a mask-aware variational autoencoder with a latent neural ordinary differential equation, and returns the learned dynamics to observation space through decoded rollouts or decoder pushforward. In this work, we use the term foundation model in a dynamical sense: a shared transferable backbone for temporal modeling across datasets and downstream tasks. We evaluate the framework on three datasets: 64-channel electroencephalography motor movement and imagery, the AirQualityUCI multi-variate environmental time series with meteorological and calendar controls, and a registered Drosophila blastoderm gene-expression atlas. Across electroencephalography and air quality, the same backbone improves open-loop forecasting relative to classical baselines and supports controlled counterfactual rollouts. In electroencephalography, it also enables data-efficient subject adaptation. In air quality, it supports interpretable in silico intervention screening over exogenous drivers. In the Drosophila atlas, the black-box backbone supervises a sparse editable Hill-type mechanistic student model, yielding candidate regulatory structure, equation-level perturbation hypotheses, and an interpretable artificial-intelligence workflow for mechanistic analysis. These results show that a shared latent-dynamics backbone can unify forecasting, adaptation, counterfactual analysis, interpretable AI, and mechanistic distillation across heterogeneous biological time-series settings.