Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster B in Poster Session B: Tuesday, August 4, 2:00 – 3:45 pm, Kimmel Center, Shorin & Rosenthal Rooms

Connectome-Constrained Spiking Networks Reproduce Emergent Computations in Larval Olfactory Pathway

Jordan Watts¹, Barbara Webb¹; ¹University of Edinburgh

Presenter: Jordan Watts

The complete synaptic connectome of the larval Drosophila melanogaster brain offers an opportunity to build structurally faithful neural circuit models. However, the connectome provides only the wiring diagram, not the biophysical dynamics or functional logic underlying neural computation; one approach to bridging this gap is supervised training on a biologically relevant task, with the resulting neural activity evaluated against existing data. We present a spiking neural network model of the larval olfactory pathway, from odor receptors (ORs) via the antennal lobe to Kenyon cell (KC) activation in the mushroom body (MB). All connectivity is fixed by the connectome, complemented by biologically plausible neuron dynamics including spiking and non-spiking neurons, two-compartment KCs, gap junctions, synaptic depression, and axo-axonic connections. Despite this richness, the model requires only 449 free biological parameters, which are trained on a 28-odor classification task using electrophysiological OR response data as input. We use two-phase artificial-to-spiking neural network transfer learning: a rate-based teacher is trained to convergence, then its weights are transferred to the spiking network and fine-tuned with surrogate gradient descent under full biophysical constraints for accuracy and realistic KC activation sparsity. The trained model spontaneously reproduces decorrelation localized to the MB, anterior paired lateral neuron suppression of KC activity by ~50%, concentration-invariant gain control across a 17-fold concentration range, and sub-additive KC responses to odor mixtures matching experimental observations. An ensemble of independently trained models converges to consistent KC representations and parameters, demonstrating that the connectome dominates the learned solutions. Additionally, systematically perturbing circuit components reveals their individual contributions and generates testable hypotheses. We demonstrate that the connectome, combined with biophysical constraints and a simple learning objective, produces circuit-level computations matching independent biological measurements. This establishes a generalizable methodology beyond larval olfaction for translating connectome data into testable spiking models to determine microcircuit functional logic and the roles of constituent neurons.

Topic Area: Methods, Tools, Theory & Neural Coding