DETECTING MIXED SELECTIVITY BEHIND NEURAL DYNAMICS FOR MULTISENSORY INTEGRATION IN <EM>DROSOPHILA</EM>

The capacity of animal brains to synthesize distinct sensory streams into a unified perceptual experience is a fundamental feat of neural computation. In Drosophila, multisensory integration underpins critical survival behaviors, yet circuit mechanisms that transform unisensory inputs into integrated, behaviorally relevant outputs remain poorly understood. While static connectomes provide a structural scaffold, they do not directly reveal the functional logic of dynamic information processing. Here, we investigated the neural dynamics of multisensory integration by constructing a large-scale spiking neural network constrained by the complete Drosophila brain connectome. The activation of thermosensory and hygrosensory receptor neurons accurately predicts neural responses along the multisensory pathways, including antennal lobe (AL), lateral horn (LH), and mushroom body (MB). Our analysis reveals that multimodal convergence begins early in the circuit, where specific populations of AL projection neurons and lateral horn neurons exhibit nonlinear mixed selectivity. Using dimensionality reduction on population state vectors, we demonstrate that these high-dimensional representations can effectively untangle sensory conditions, providing a robust substrate for stimulus classification. Furthermore, we applied Convergent Cross Mapping (CCM) to detect effective connectivity from the simulated spike trains, recovering functional information flow that cannot be fully predicted by the static synaptic weights. These findings suggest that the Drosophila brain employs a hierarchical strategy of progressive mixing, where early nonlinearities in the AL and LH facilitate the representation of complex environmental features. Our work bridges the gap between structural connectomics and functional neural dynamics, offering a computational framework to decode the circuit principles of multisensory perception.