TOWARDS A FOUNDATION MODEL FOR PHOTOSTIMULATION

In vivo two-photon holographic optogenetic photostimulation provides a powerful framework for simultaneously observing and perturbing neural activity, enabling causal interrogation of neural circuits. By combining targeted stimulation with large-scale recordings, these datasets offer unique opportunities to probe neuronal computation and causality by linking stimulation protocols to underlying neural dynamics. Recent progress in transformer-based models has shown their effectiveness in capturing long-range dependencies in complex sequential data. Applied to neural recordings, transformers can capture spatiotemporal dependencies present in neuronal populations. Recent work also suggests that attention mechanisms can be exploited to gain insight into neuronal functionality, by allowing directed connectivity matrices to be approximated. In parallel, proto-foundation models of neuronal population, trained on large, heterogeneous datasets, have shown the ability to generalize across animals, tasks and trials, capturing the shared statistical and dynamical structure of population activity, often interpretable as low-dimensional latent dynamics. Thus, recent work points at the possibility to develop a photostimulation foundation model capable of learning causal relationships not only between neurons, but also between neurons and photostimulation inputs, allowing for counterfactual inference of unobserved or unrealized perturbations. Here, we benchmark different transformer-based approaches on simulated and real calcium imaging data, explicitly incorporating perturbation information into the model, comparing their performance across tasks, such as roll-out trace prediction or within-neuron trace inference. Overall, our study lays the groundwork for the development of photostimulation foundation models that can integrate perturbation-informed transformer architectures to uncover causal relationships within neural circuits.