Reconstructing Circuit Connectivity from in vivo Spike Trains Using Deep Domain-Adaptive Matching
Kaiwen Sheng,Shanghang Zhang,Maxime Beau,Peng Qu,Le Yang,Xiaofei Liu,Liuyuan He,Youhui Zhang,Lei Ma,Kai Du
DOI: https://doi.org/10.1101/2022.10.03.510694
2024-12-07
Abstract:Inferring the connectivity of neural circuits from in vivo experimental data is essential for understanding the neural architectures that underpin behavior and cognition. However, advanced machine learning approaches in this domain, particularly deep learning, are significantly constrained by incomplete in vivo observations. This limitation restricts the ability to assign 'ground-truth' labels to in vivo data, compelling researchers to rely on synthetic data generated by biophysical neural models for initial training. Yet, this strategy introduces a widely perceived insurmountable challenge: 'model mismatch', where synthetic model dynamics fail to accurately match true neural dynamics. To tackle these interrelated challenges, we introduce DeepDAM (Deep Domain-Adaptive Matching), a robust, adaptive, and versatile framework that fundamentally transforms the training of deep neural networks (DNNs) for inference tasks using both synthetic data and unlabeled in vivo data. DeepDAM fine-tunes DNNs on a fusion of synthetic and unlabeled in vivo data, adaptively aligning the DNN's feature space with in vivo neural dynamics, thereby effectively mitigating the model mismatch challenge. Impressively, we validated the framework with extracellular recordings in the hippocampal CA1 region of freely behaving mice, which achieves an exemplary Matthews correlation coefficient of 0.97-1.0 in connectivity inference accuracy, significantly outperforming existing methods (~0.6-0.7). Importantly, our framework can well adapt to diverse experimental conditions and a broad spectrum of neural properties and scales, demonstrating its high generalizability across various in vivo and ex vivo scenarios. This marks a significant step towards the accurate and comprehensive reconstruction of functional mammalian brains using data-driven methodologies.
Neuroscience