Modeling proprioception with physics-informed neural networks

Proprioception is essential for planning and executing precise motor actions. It relies on sensory signals from mechanosensory neurons distributed within muscles, tendons, and joints. Muscle spindles convey information through sensory afferents to the central nervous system. While traditionally viewed as stretch receptors encoding muscle length and velocity, recent insights suggest they may function as adaptable signal-processing devices. Spinal gamma motor neurons are believed to drive this adaptability by modulating the responses of sensory afferent to task demands. However, the mechanisms by which this top-down modulation occurs are still unknown. To elucidate this, we propose a novel model of muscle spindles that merges structural fidelity with computational efficiency, leveraging the power of physics-informed neural networks (PINNs). The unique advantage of PINNs lies in their ability to incorporate physics knowledge, thereby enhancing model interpretability and predictive accuracy. By integrating principles of biomechanics and neural dynamics, our model captures the interplay of biomechanics and transduction processes within muscle spindles while maintaining computational tractability. Through validation across multiple experimental datasets and species, our model demonstrates superior performance in fitting experimental data and provides interpretability of the sources of variability in afferent responses including their top-down control. By bridging the gap between biomechanics and neural dynamics, our model offers a comprehensive understanding of muscle spindle function, shedding light on their role as adaptable signal processors in sensorimotor control. The proposed framework holds promise for advancing our understanding of proprioceptive mechanisms and functions along the sensorimotor pathway.