Learning system parameters from turing patterns.

Area of Science:

• Computational Biology
• Mathematical Biology
• Developmental Biology

Background:

• Turing mechanisms explain spatial pattern formation in biological development via reaction-diffusion processes.
• Identifying specific Turing mechanisms and their parameters in biological systems is a significant challenge.
• Existing methods struggle with the variability of patterns arising from unknown initial conditions.

Purpose of the Study:

• To develop an approach for predicting Turing parameter values from observed Turing patterns.
• To enable the identification of reaction-diffusion models that generate observed biological patterns.
• To provide a method that requires only a single pattern for parameter prediction.

Main Methods:

• Introduced a novel invariant pattern representation using resistance distance histograms.
• Employed Wasserstein kernels to handle pattern variability and unknown initial conditions.
• Utilized numerical model evaluation with random initial data for training and prediction.

Main Results:

• Demonstrated accurate prediction of single Turing parameter values from a single pattern.
• Achieved reasonably accurate joint prediction of all four parameters of the Gierer-Meinhardt model.
• Showcased that classical methods outperform neural networks on small datasets, while neural networks excel on large datasets.

Conclusions:

• The developed approach effectively predicts Turing model parameters from limited pattern data.
• This method significantly advances the ability to identify and characterize Turing mechanisms in biological systems.
• The reliance on a single pattern input makes this approach highly practical for real-world biological pattern analysis.