Calibrating agent-based models to tumor images using representation learning.

Area of Science:

• Computational Biology
• Cancer Research
• Biophysics

Background:

• Agent-based models (ABMs) are valuable for studying tumor development and therapeutic responses, offering insights into spatiotemporal evolution.
• Current ABM parameterization relies on disparate data sources, lacking rigorous estimation methods.
• Existing methods struggle to quantitatively compare ABM simulations with spatial tumor image data.

Purpose of the Study:

• To develop a novel quantitative method for comparing agent-based model (ABM) simulations with experimental tumor images.
• To enable rigorous parameter estimation for ABMs by minimizing the distance between simulated and experimental image representations.
• To demonstrate the application of this method for parameter fitting in distinct ABM scenarios.

Main Methods:

• Utilized neural networks to generate low-dimensional representations (embeddings) of both tumor images and ABM simulations.
• Defined the distance between these low-dimensional points as a quantitative measure of difference.
• Employed standard parameter-fitting algorithms to minimize the discrepancy between simulated and experimental image embeddings.

Main Results:

• Successfully created a quantitative framework for comparing ABM simulations and tumor images.
• Demonstrated the method's efficacy in parameter estimation for two different ABMs.
• Established a robust approach for enhancing the accuracy and reliability of ABM parameterization.

Conclusions:

• The presented neural network-based approach provides a robust and quantitative method for estimating parameters in agent-based models (ABMs).
• This technique overcomes limitations in comparing spatial data from ABM simulations and tumor images.
• Facilitates more accurate modeling of tumor development and therapeutic responses through improved parameter fitting.