Application of adaptive-network-based fuzzy inference systems to the parameter optimization of a biochemical rule-based model

Abstract

Our main contribution is an efficient machine learning approach to fitting parameters of a biological model. We study the binding of the shrimp protein Pen a 1 with antibody-receptor complexes because this process is important in understanding the allergic response. Previously, we developed a BioNetGen model that simulates this process. We previously developed a method for encoding steric effects via the optimization of two parameters: the cutoff distance and the rule rate. We optimized these two parameters by fitting the output to that generated by a 3D robotics-inspired Monte Carlo simulation that explicitly represents molecular geometry.In this work, we aim to optimize the parameters for our BioNetGen model using an efficient method: an adaptive-network-based fuzzy inference system implemented in MAT-LAB. We want to develop fuzzy systems that can accurately predict the rule binding rate and cutoff distance given a residual-sum-of-squares value or a probability distribution. We construct the fuzzy systems using fuzzy c-means clustering with existing data from BioNetGen model parameter scans as the training data. We create and test fuzzy systems with various input data and number of clusters, and analyze their performance with regard to the effective optimization of our rule-based model. We find that the fuzzy system that uses a residual-sum-of-squares value as the input value performs acceptably well. However, the performance of the fuzzy systems that use probabilities as their input values perform inconsistently in our tests.The results of this study suggest that the system that uses a residual-sum-of-squares value as the input value could potentially be used to find an adequate fit for our biochemical model. However, the systems that use probabilities as their input values need further development to improve the consistency and reliability of their output. Testing more values for other clustering parameters other than the number of clusters may accomplish this. Further research could also include similar studies using other training or clustering algorithms. This methodology could be modified for use with fitting other biological models.