NEuRT: A Transformer-Based Model for Explainable Neuronal Activity Analysis

The study of neuronal activity is essential for understanding brain function and its alterations in neurodegenerative diseases. Advances in in vivo imaging have enabled real-time observation of neuronal dynamics, but classical statistical methods struggle to capture the complex, time-dependent interactions within neuronal networks. Machine learning offers promising solutions for analyzing high-dimensional neuronal data, yet their application in neuroscience remains limited. Here, we introduce NEuRT, a Bidirectional Encoder Representations from Transformers (BERT)-based model adapted for neuronal activity analysis. NEuRT leverages self-attention mechanisms to interpret complex neuronal interactions, providing insights into patterns that traditional methods may overlook. Pre-trained on the recently introduced large annotated dataset MICrONS for signal reconstruction, NeuRT demonstrates strong generalization, effectively reconstructing activity from both visual cortex two-photon and hippocampal miniature fluorescence microscopy. Built on the BERT architecture, the NEuRT model can be efficiently fine-tuned for a wide range of downstream tasks. We showcase its application in classifying wild-type and transgenic Alzheimer's disease model mice, based on hippocampal activity, revealing group-specific features through attention map analysis. By reducing reliance on extensive labeled data, addressing a critical challenge in neuroscience, NEuRT bridges fundamental neuroscience and disease research, offering a robust framework for AI-driven and explainable neuronal activity analysis.