Temporal-filter enhanced prediction of whole-brain neural activity using the physiology-aligned latent variable model

Chong Li, Zhongyu Chen, Jianfeng Feng, Xiangyang Xue, Yuguo Yu

Abstract: Understanding the neural mechanisms behind intelligent behaviors in nematodes requires a comprehensive framework that integrates connectome data with recordings of all neuronal activities. However, obtaining full membrane potentials for all labeled neurons is challenging, and even in C. elegans, which has only 302 neurons, current calcium imaging techniques typically capture just over half of these labeled neurons in vivo. Existing advances primarily focus on using spatial correlation across the whole brain to causally predict unknown neuronal activities by using measured activities of half-numbered neurons, but they struggle to achieve high prediction accuracy. By introducing a quasi-independent temporal coding property of populational neurons in living brains, e.g., C. elegans, we establish the Temporal-Filter and Physiology-Aligned Latent Variable Model (TF-PALVM), a new algorithmic leap that synergizes the spatial connectome structure with temporal coding functions of individual neurons to infer the electrical activities of the entire neural ensemble. This model employs an autoencoder network with temporal kernels, rather than relying on spatial correlations, refined to reflect individual neuronal temporal coding functions and predict their future temporal activities, while embedding experimentally derived synaptic weights into a biologically coherent framework. When tested, our model demonstrates unprecedented reconstruction accuracy, surpassing existing models by approximately 75 % in the worm holdout evaluation and 51 % in neuron holdout performance. Moreover, it precisely predicts synaptic polarities, with 75 % of them to be excitatory, matching experimental excitatory synapse data. It is also able to identify the top neuron pairs with the most influence on behavioral correlations, consistent with previous experimental research. Our TF-PALVM stands as a transformative tool for neuroscientific exploration, capable of predicting missing neuronal activity with high fidelity. The success of the model confirms that the temporal response history of individual neurons contains more valuable information than the population network in predicting their present and future responses to sensory inputs. It offers a scalable approach to potentially unravel the complexities of larger, more intricate brains.