Abstract
Inferring the monosynaptic connectivity of neural circuits from in vivo experimental data is essential for understanding the neural architecture that underpins behavior and cognition. However, advanced machine learning (ML) methods, especially deep learning, face significant challenges because in vivo observation is limited and incomplete, making it impractical to obtain ground-truth labeling. As a result, researchers typically rely on synthetic data generated by biophysical neural models for initial training. However, this reliance introduces the well-known “model mismatch” problem, whereby synthetic neural dynamics do not accurately reflect the complexity of real neural activity. To address these challenges, we introduce DeepDAM (Deep Domain Adaptive Matching), a flexible and robust framework that combines cutting-edge ML techniques with biophysical modeling. DeepDAM utilizes both synthetic data and unlabeled in vivo recordings to fine-tune deep neural networks (DNNs), so that the feature space of the DNNs is adaptively aligned with real neural dynamics, thus effectively mitigating the model mismatch problem and dramatically improving inference performance. We validated DeepDAM using extracellular recordings in the hippocampal CA1 region of freely behaving mice. Surprisingly, the framework achieved a Matthews correlation coefficient of 0.97–1.0 for monosynaptic connectivity inference, significantly outperforming existing methods (∼0.6–0.7). Additionally, our framework demonstrates robust adaptability to diverse experimental conditions and a broad range of neural properties and scales, including inference of single-neuron biophysics, synaptic dynamics, and microcircuit dynamics in multiple ex vivo scenarios. This work marks a significant step towards the accurate and comprehensive reconstruction of functional mammalian brains by integrating data-driven ML techniques with first-principles insights.
Competing Interest Statement
The authors have declared no competing interest.
Footnotes
Revise the introduction and abstract to clearly outline the work