Biophysical Society Thematic Meeting | Canterbury 2023

Towards a More Perfect Union: Multi-Scale Models of Muscle and Their Experimental Validation

Thursday Speaker Abstracts

MACHINE LEARNING FOR BUILDING CLASSIFIERS AND RATE ESTIMATES IN SIMULATED TWITCHES Travis Tune 1,2 ; Anthony Asencio 2,3 ; Sage Malingen 2,3 ; Kristina Kooiker 1 ; Jennifer Davis 3,4 ; Thomas Daniel 2 ; Farid Moussavi-Harami 1,4 ; 1 University of Washington, Medicine/Cardiology, Seattle, WA, USA 2 University of Washington, Biology, Seattle, WA, USA 3 University of Washington, Bioengineering, Seattle, WA, USA 4 University of Washington, Laboratory Medicine and Pathology, Seattle, WA, USA Computational models of the sarcomere allow for understanding mechanisms and modeling disease conditions or drugs. We have used a spatially explicit Monte-Carlo Markov-Chain model to simulate a set of twitches by introducing rate constant perturbations that mimic thin and thick filament activation associated with genetic cardiomyopathies. Harnessing this large, labeled data set, we used machine learning (ML) methods to classify disease states based on a collection of temporal features. These features include the actual time history of simulated twitches, the twitch tension index (TTI, an integral of the twitch), the first two principal components (PCs) based on singular value decomposition and traditional activation/relaxation parameters. Some of these features, notably TTI combined with the raw twitch data and PCs and TTI approach 80% classification accuracy. We extended the model from our earlier analysis of three cross-bridge state transitions to five state transitions along with the super-relaxed state or myosin “off state” – essentially a six-state model modified and then generated even larger data set and varying a larger parameter space. We also implemented improved estimates of state transition probabilities using a newly developed algorithm for computing the matrix exponential of the rate constant matrix. Moreover, the spatially explicit nature of our model also now allows for cooperative activation via nearest neighbor interactions between troponin sites on the thin filament. We have now simulated 500,000 combinations of rate parameters to solve the inverse problem of estimating rate parameters for specified twitches by implementing Bayesian parameter estimation method using a Conditional Variational AutoEncoder to compute the one and two-dimensional marginalized posteriors (probability distributions) that predict the rates. We will apply this technique to identify therapeutic targets for different abnormal twitches.

59