Abstract:Successful development from a single cell to a complex, multicellular organism requires that cells differentiate in a coordinated and organized manner in response to a number of chemical morphogens. While the molecular underpinnings may be complex, the resulting phenomenon, in which a cell decides between one fate or another, is relatively simple. A body of work, rooted in dynamical systems theory, has formalized this notion of cellular decision making as flow in a Waddington-like landscape, in which cells evolve according to gradient-like dynamics within a potential that changes shape in response to a number of signals. We present a framework leveraging neural networks as universal function approximators to infer such a parameterized landscape from gene expression data. Inspired by the success of physics-informed machine learning in data-limited contexts, we enforce principled constraints motivated not by physical laws, but by this phenomenological understanding of differentiation. Our data-driven approach infers a governing landscape atop a manifold situated within expression space, thereby describing the dynamics of interest in a biologically meaningful context. The resulting system provides an intuitive, visualizable, and interpretable model of cellular differentiation dynamics.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to understand the dynamic mechanisms in the process of cell differentiation, especially how to infer a parameterized landscape model from gene expression data, which can describe how cells make fate decisions in response to external signals during development. Specifically, the authors proposed a framework that uses neural networks as universal function approximators to infer such a parameterized landscape model from single - cell RNA sequencing data. This model can not only provide an intuitive and visual model of cell differentiation dynamics, but also reveal how chemical signals affect this process, thus providing a theoretical basis for synthetic biology and medical applications. ### Core Problems of the Paper 1. **Dynamic Model of Cell Fate Determination**: - How to extract the dynamic information of cell differentiation from high - dimensional gene expression data? - How to construct a model that can describe how cells make fate decisions in response to external signals during development? 2. **Parameterized Landscape Model**: - How to use neural networks to learn a parameterized landscape model from data? - How to incorporate the influence of signals into the model so that the model can reflect the influence of signals on cell fate determination? 3. **Biological Significance**: - How to ensure that the model has biological significance in low - dimensional space, that is, the landscape and signal mappings in the model can correspond to the actual gene expression data? - How to predict cell fate changes under different signal conditions through the model? ### Method Overview - **Potential Module**: - Use a neural network to represent a static potential function \(\tilde{\phi}(x)\), which is defined on a \(d\)-dimensional phase space. - To ensure the smoothness and non - escape of the trajectory, a regularization term \(\Phi_0(x) = C_{\text{conf}}\|x\|^4\) is introduced. - **Signal Mapping Module**: - Define a transformation \(\Psi\) that maps the signal vector \(s\in\mathbb{R}^d\) to a tilt vector \(\tau\in\mathbb{R}^d\) that affects the landscape. - Implement this mapping through a neural network so that the model can learn how signals affect the shape of the landscape. - **Noise Kernel**: - Capture the randomness of cell movement in the landscape and ensure that the model can describe the stochastic nature of cell differentiation. ### Mathematical Expressions - **Potential Function**: \[ \Phi_\theta(x)=\Phi_0(x)+\Phi_{\text{nn},\theta}(x) \] where \(\Phi_0(x) = C_{\text{conf}}\|x\|^4\) is the regularization term, and \(\Phi_{\text{nn},\theta}(x)\) is the nonlinear potential function represented by the neural network. - **Parameterized Landscape**: \[ \phi(x; s)=\tilde{\phi}(x)+x^T\psi(s) \] where \(\psi(s)\) is the mapping from the signal vector \(s\) to the tilt vector \(\tau\). - **Gradient Dynamics**: \[ F(x; s)=-\nabla_x\phi(x; s) \] ### Application Examples - **Synthetic Data Testing**: - Use the binary choice and binary flip landscape models with completely specified synthetic data generation processes to verify the effectiveness of the model. - **Real - Data Application**: - Apply the model to the in - vitro experimental data of the early mouse embryonic ectoderm to demonstrate its application potential in actual biological systems. Through these methods, the paper aims to provide a model that can extract cell differentiation dynamic information from high - dimensional gene expression data and predict cell fate changes under different signal conditions. This provides a basis for understanding and developing the regulatory mechanisms of cell fate determination.

Learning geometric models for developmental dynamics