Integrating deep learning 3D tracking and biophysical EOD modeling for precise, noninvasive computational...

Recording the movement of freely swimming, weakly electric fish during naturalistic behavior and social interactions poses persistent challenges to neuroethologists. Most species are nocturnal; consequently, illumination for video recording often disrupts their natural behavior. Substantial work has addressed this issue using shallow-water tanks, infrared illumination, or indirect movement derivation. While inferring position from a fish's own electrical signals has also been explored, it remains a promising yet largely unresolved problem in 3D. We introduce an open-source, multimodal, and data-driven method as a first step toward 3D tracking based on electrical signals. Applied to the pulse-type fish Gymnotus carapo, this approach links 3D posture to self-generated electric field dynamics. This proof-of-concept study utilized a deep tank, a peripheral electrode array, and a multi-camera system to record a freely swimming animal. Using deep-learning algorithms, we reconstructed 3D skeletal trajectories from video recordings, achieving a spatial accuracy of approximately 0.8 mm per coordinate. Subsequently, we simulated the electric potentials generated by three charge distribution models mapped onto the skeletal structure throughout the tracking period. By comparing these simulations with experimental recordings, we evaluated the consistency of the forward modeling framework and the relative performance of each model. The results suggest that a biologically inspired asymmetric charge distribution better reproduces experimental observations than simpler traditional dipole-like models. Furthermore, we developed a tool dedicated to the precise, non-invasive study of charge distribution along the electric organ. It supports the evaluation of biophysical models linking body posture to electric signal generation. These findings suggest that paired electrical and spatial datasets may support the development of video-free 3D tracking methods based on machine learning for dark, deep-water environments. These results could also assist in the design of novel biomimetic electrolocation systems. Looking ahead, this methodology could be extended to multiple animals and adapted to some other pulse- or wave-type electric fish, thereby offering a framework for integrating anatomy, locomotion, electrosense, and electrogenesis in neuroethological studies.