Event

Understanding how neurons interact at the network level is a central goal in systems neuroscience. Recent advances in two-photon calcium imaging combined with optogenetic stimulation provide unprecedented opportunities to probe neuronal populations at scale. However, extracting causal relationships from such data remains fundamentally challenging due to three key issues: (i) neural spiking activity is not directly observed but must be inferred from noisy and temporally blurred calcium signals, (ii) experimental measurements are subject to substantial noise and model mismatch, and (iii) unobserved confounding can induce spurious dependencies that invalidate standard correlation-based or regression-based approaches.

In this work, we propose a unified, iterative framework that jointly addresses latent spike inference and causal network estimation through a generative modeling approach that accounts for noisy observations, optogenetic perturbations, and hidden confounders. By casting spike recovery and causal discovery as a coupled inverse problem, we explicitly model how uncertainty propagates between these components. We develop complementary inference strategies and show that their differences are most evident in downstream causal tasks. To estimate causal interactions, we leverage an instrumental variable framework tailored to optogenetic experiments, enabling identification under confounding.

Through simulations and real experimental data, we demonstrate improved recovery of network structure compared to standard approaches, particularly in confounded settings. An iterative refinement procedure further enhances stability and accuracy. Finally, we outline extensions toward quasi-experimental designs and closed-loop experimentation, providing a scalable and principled framework for causal network identification from indirect and noisy neural data.