In spring, many animals become more active and prepare for reproduction. Additionally, some animals migrate in autumn, and in winter, they hibernate or enter dormancy to survive the harsh season. As these examples demonstrate, animals can appropriately sense seasonal changes and flexibly alter their physiological functions and behaviors. Many animals adapt to seasons by relying on information from day length and temperature, but it is also known that they actively adapt to seasonal changes by utilizing an internal biological clock that maintains a rhythm with an approximately one-year cycle. However, many aspects of the molecular mechanisms underlying seasonal adaptation in animals remain unclear.

Medaka (Oryzias latipes) is an excellent model for understanding seasonal adaptation mechanisms in animals, not only because it exhibits sophisticated seasonal responses, but also because high-precision genomic information and genome editing technologies are available. We have developed data-driven research by applying omics analyses and genome editing technologies to this excellent model. In this presentation, I will introduce the mechanisms of seasonal adaptation in animals that have been revealed through these approaches.

Modern biology and medicine increasingly rely on high-dimensional spatio-temporal data such as live-cell imaging and whole-brain medical images. While machine learning methods provide powerful tools for predictive analysis, they do not necessarily provide insight into the underlying biological mechanisms. In this talk, I present a unified framework that combines biophysics modeling with machine learning methods to extract hidden physical and biochemical information from image data. Such an approach has long been studied in engineering under the name of inverse problem/analysis. However, in biomedical sciences, mathematical models tend to be far more uncertain and complex, which poses challenges in modeling, simulation and learning [1,2].

First, I introduce a fortunate case in which the material properties of cultured epithelia can be inferred by using textbook machine learning methods [3]. By coupling cellular velocity and force measurements with a growing elastic sheet model, we estimate stresses and material parameters without applying invasive external forces, revealing an unusual mechanical behavior consistent with epithelial function.

Second, I turn to a more challenging case in which effective biochemical heterogeneity governing the progression of Alzheimer's disease is inferred from human brain PET scans. A core difficulty is that we need to estimate millions of parameter values to fully characterize biochemical heterogeneity across the whole brain. To address this issue, we formulate a reaction-diffusion model on realistic brain geometry and implement it in a differentiable simulation framework, which enables efficient learning via error backpropagation through time. We then estimate the spatially heterogeneous amplification rate of pathological tau, a quantity that is otherwise inaccessible to direct measurement.

Finally, I will discuss open problems and future directions, with a particular focus on applications to spatial omics data.