Learning partial differential equations for biological transport models from noisy spatio-temporal data

Abstract

<jats:p>We investigate methods for learning partial differential equation (PDE) models from spatio-temporal data under biologically realistic levels and forms of noise. Recent progress in learning PDEs from data have used sparse regression to select candidate terms from a denoised set of data, including approximated partial derivatives. We analyse the performance in using previous methods to denoise data for the task of discovering the governing system of PDEs. We also develop a novel methodology that uses artificial neural networks (ANNs) to denoise data and approximate partial derivatives. We test the methodology on three PDE models for biological transport, i.e. the advection–diffusion, classical Fisher–Kolmogorov–Petrovsky–Piskunov (Fisher–KPP) and nonlinear Fisher–KPP equations. We show that the ANN methodology outperforms previous denoising methods, including finite differences and both local and global polynomial regression splines, in the ability to accurately approximate partial derivatives and learn the correct PDE model.</jats:p>

Biologically-informed neural networks guide mechanistic modeling from sparse experimental data

“…Note that this weighting factor makes BINNs sensitive to the random choice of training/validation split, since some data points in the initial condition may be more informative than others for equation learning and ultimate model generalizability. This observation was also noted in a recent equation learning study in which the random split of training and validation sets was found to influence the structure of the learned equation [26]. Adopting a strategy similar to this previous study, BINNs were trained 20 times for each data set (using different random training/validation splits).…”

“…is used with proportionality constant γ = 0.2. Note that γ was tuned numerically following the methodology suggested in [26] (see Methods Section for more details). The next term L PDE ensures u MLP satisfies the solution of the governing PDE.…”

“…The data relating u t to all possible model terms inside the library are formulated as a linear regression problem in which sparsity promoting techniques are used to select a small subset of library terms that produce the most parsimonious model. While the sparse regression framework has been successfully demonstrated to circumvent searching through a combinatorially large space of possible candidate models, it can require large amounts of training data and the numerical methods used for denoising and differentiation are not robust to biologically realistic forms and levels of noise, leading to inaccuracies in both the constructed library and learned equations [ 26 ]. Further, the method assumes the unknown function in Eq (2) can be written as a linear combination of nonlinear candidate terms, which may not be true in practice.…”

Development of a hybrid model for a partially known intracellular signaling pathway through correction term estimation and neural network modeling

“…First, a de-noise technique can be implemented prior to the hybrid model development so that the noise in the measurements will become less influential. In the literature, various methods such as finite difference with polynomial spline [67], spectral transformation [68], sparse Bayesian regression [69], and neural networks [70] have been proposed. Second, alternative ANN training mechanisms can be implemented to improve the performance of an ANN.…”

Learning Everywhere: A Taxonomy for the Integration of Machine Learning and Simulations

“…Use ANN's to discover the PDE form of biological transport equations from noisy data. [21] C. MLaroundHPC: Learning Model Details -ML for Data Assimilation (predictor-corrector approach) Data assimilation involves continuous integration of time dependent simulations with observations to correct the model with a suitable combined data plus simulation model. This is for example common practice in weather prediction field.…”