NNVA: Neural Network Assisted Visual Analysis of Yeast Cell Polarization Simulation.

Area of Science:

• Computational biology
• Scientific visualization
• Machine learning

Background:

• Complex computational models require extensive parameter calibration, which is computationally expensive.
• Analyzing high-dimensional parameter spaces for simulations is challenging.
• Existing methods for parameter analysis are often time-consuming and resource-intensive.

Purpose of the Study:

• To develop an interactive visual analysis system for high-dimensional parameter spaces in complex simulations.
• To enable computational biologists to visually calibrate simulation input parameters efficiently.
• To reduce the computational cost associated with parameter analysis and model calibration.

Main Methods:

• Development of a visual analysis system driven by a neural network-based surrogate model.
• Integration of uncertainty quantification, interpretability, and explainability techniques for neural networks.
• Utilizing activation maximization for parameter sensitivity analysis and optimal configuration recommendations.
• Performing detailed analysis of the trained neural network to extract simulation insights.

Main Results:

• The system facilitates interactive, real-time visualization of simulation outcomes based on parameter modifications.
• Identified multiple novel parameter configurations that yield high cell polarization in the yeast model.
• Demonstrated the effectiveness of neural networks as surrogate models for visual analysis and parameter calibration.
• Validated findings against original simulation results and expert analysis.

Conclusions:

• The proposed visual analysis system significantly accelerates the parameter calibration process for complex simulations.
• Neural network surrogate models, enhanced with UQ and XAI, offer a powerful framework for interactive scientific exploration.
• The system aids in discovering new insights and optimizing parameters for biological simulations, as evidenced by cell polarization studies.