Forecasting and Predicting Stochastic Agent-Based Model Data with Biologically-Informed Neural Networks | JoVE Visualize

Area of Science:

Computational Biology
Mathematical Modeling
Artificial Intelligence

Background:

Collective cell migration is crucial for biological processes like wound healing and development.
Agent-based models (ABMs) simulate collective migration but are computationally intensive and difficult to parameterize.
Mean-field differential equation (DE) models offer faster simulations but can inaccurately represent ABM behavior in certain parameter spaces.

Purpose of the Study:

To develop a method using biologically-informed neural networks (BINNs) to create accurate and interpretable DE models for collective migration.
To enable the prediction of ABM behavior in both unseen data and unexplored parameter regions.
To enhance the efficiency of parameter space exploration for ABMs.

Main Methods:

Training biologically-informed neural networks (BINNs) to learn interpretable differential equation (DE) models from ABM data.
Utilizing BINN-guided partial differential equation (PDE) simulations for forecasting future ABM data.
Combining BINN-guided PDE simulations with multivariate interpolation to predict ABM behavior at new parameter values.
Validating the approach with three distinct ABMs of collective migration.

Main Results:

BINN-guided PDE simulations accurately forecast spatial ABM data not encountered during training.
The method successfully predicts ABM behavior in previously unexplored parameter ranges.
A single-compartment BINN-guided PDE accurately captured ABM dynamics where traditional mean-field models were ill-posed or required multiple compartments.
The approach demonstrated efficiency in exploring parameter space.

Conclusions:

Biologically-informed neural networks (BINNs) provide a powerful framework for developing accurate and interpretable DE models from ABMs.
This BINN-guided PDE approach significantly enhances the ability to forecast and predict collective migration dynamics across parameter spaces.
The methodology facilitates efficient exploration of ABM parameter spaces, opening avenues for data-driven tasks like parameter estimation from experimental data.