How can AI accelerate advances in physiology?

The commentary explores how artificial intelligence can accelerate progress in physiology, particularly the long-standing challenge of connecting genotype and molecular structure to tissue and organ function. While high-throughput molecular data have expanded rapidly, predicting pathophysiological phenotypes and explaining mechanisms at the organ level remain bottlenecks. The piece frames the central question: how can AI, especially machine learning, complement mechanistic physiological understanding to bridge molecular data and functional phenotypes?

The article situates the discussion within work on cardiomyopathies caused by sarcomeric mutations. It highlights Davis et al. (2016), who showed that the time integral of the isometric twitch tension (tension–time index) predicts phenotype: larger, longer twitches associate with hypertrophic cardiomyopathy, while smaller, shorter twitches associate with dilated cardiomyopathy. It also references Asencio et al. (2023), who used machine learning with a mechanistic, stochastic, spatially explicit multi-filament half-sarcomere model to classify cardiomyopathy-associated mutations, illustrating how ML can extract informative features from physiological signals when coupled with biophysical models.

Although this is a perspective, it details the approach of Asencio et al. (2023): (1) simulate isometric twitch dynamics using a stochastic, spatially explicit, multi-filament half-sarcomere model driven by an experimentally derived calcium transient; (2) vary kinetic parameters related to Ca2+ binding to troponin C and crossbridge formation to emulate six combinations of sarcomeric gene mutations and dosages associated with distinct cardiomyopathic phenotypes; (3) generate a training set of twitch tension time courses; (4) apply unsupervised and supervised machine learning to identify features of twitch morphology that discriminate among mutations and phenotypes; (5) evaluate classification accuracy and compare feature sets, including full twitch time courses, tension–time index, and principal modes of twitch variation.

Machine learning, when paired with mechanistic models of sarcomere dynamics, can classify cardiomyopathy-associated sarcomeric mutations with nearly 80% accuracy based on twitch morphology. The tension–time index combined with two principal modes of twitch variation slightly outperformed using the entire twitch time course for classification, suggesting compact, physiologically meaningful features can be more robust than high-dimensional signals. The work underscores the rich mechanistic information embedded in classical physiological signals like isometric twitch tension.

The findings illustrate a practical path for AI to accelerate physiology: leveraging well-validated mechanistic models to generate informative synthetic datasets that reveal genotype–phenotype relationships and guide feature discovery. This approach addresses the scarcity and variability of clinical/experimental datasets and mitigates the black-box nature of ML by rooting features in biophysical mechanisms. The identification of concise, mechanistically interpretable features (e.g., tension–time index and principal modes) provides better generalization potential and clearer links to underlying molecular processes, advancing the translation from molecular perturbations to tissue- and organ-level function.

Coupling machine learning with predictive, mechanistic models offers a promising strategy to discover physiologically meaningful genotype–phenotype relationships from classical signals such as isometric twitch tension. The commentary emphasizes that such hybrid approaches can accelerate hypothesis generation and mechanistic insight. Future work should expand training to include many more mutations and rigorously test classifiers on measured twitches from human muscle with unknown mutations, aiming to validate robustness in real-world, variable datasets.

The commentary notes several constraints: machine learning approaches need large, high-quality training data and can function as black boxes lacking mechanistic explanations. The synthetic-data-trained classifiers must be validated on experimentally measured twitches, which are confounded by variability. Broader coverage of mutations and dosages is required to ensure generalizability and clinical relevance.