Neuronal wiring diagram of an adult brain

Although afferent and efferent neurons make up a numerically small proportion of the brain (estimated 13.9% and 1.1%, respectively), they are important because they connect the brain to the outside world. Examining connections of these neurons is useful when attempting to predict the functions of intrinsic neurons from the connectome. For example, one might try to identify the shortest path in the connectome from an afferent (input) neuron that leads to a given intrinsic neuron. The sensory modality of the afferent neuron could provide a clue regarding the function of the intrinsic neuron. This approach, although intuitive, ignores connection strengths and multiplicities of parallel pathways. We therefore use a probabilistic model to estimate information flow in the connectome, starting from a set of seed neurons (Fig. 6a and Methods).

The likelihood of a neuron being traversed increases with the fraction of inputs from already traversed neurons up to an input fraction of 30%, after which traversal is guaranteed (Fig. 6a). We ran the traversal model for every subset of afferent neurons as seeds (n = 12 input modalities to the central brain (Fig. 2e and Supplementary Fig. 8; full list in Methods)). We then measured the flow distance from these starting neurons to all intrinsic and efferent neurons of the central brain. For instance, the neurons reached early from gustatory neurons (Fig. 6b) match second-order projection neurons identified by Snell et al. using trans-Tango.

To visualize information flow for neurons with inputs in the central brain in a common space, we treated the traversal distances starting from each seed population as a neuron embedding and built a uniform manifold approximation and projection (UMAP) from all of these embeddings (Fig. 6c). Within the map, we found that neurons of the same cell class (for example, two groups of Kenyon cells, all mushroom body output neurons, all antennal lobe local neurons and all central complex neurons) cluster, indicating that cell types can in part be defined by their proximity to different input neurons. Next, we displayed traversal order on top of the UMAP plot to compare traversal orders starting from different modalities (Fig. 6c,d). We find that almost every neuron in the central brain can be reached by starting from any modality—this ‘small world’ property of the network is covered in more detail by Lin et al. Comparing orders revealed that almost all neurons in the central brain are reached early starting from some modality, with the exception of neurons in the central complex (Fig. 6c,d and Extended Data Fig. 9), highlighting that the central complex is dominated by internal computations. Kenyon cells were contained in two clusters—one of which is targeted very early from olfactory receptor neurons and the other is targeted early by VPNs.

We then ranked all neurons by their traversal distance from each set of starting neurons and normalized the order to percentiles. For instance, a neuron at the 20th percentile had a lower rank than 80% of neurons. This enabled us to determine how early information from each afferent modality reached various targets, including the descending neurons, endocrine neurons, motor neurons and VCNs (Fig. 6e and Extended Data Fig. 8a). Endocrine neurons are closest to the gustatory sensory neurons, whereas motor and descending neurons were reached early for mechanosensory and visual afferents (Extended Data Fig. 8a).

We next tested whether the afferent cell classes target inhibitory neurons early or late. We found that putative inhibitory neurons (neurons predicted to express GABA and glutamate) were overrepresented in the set of early neurons (Fig. 6f). Surprisingly, we identified a sequence of GABAergic and glutamatergic peaks in the sequence of targeted neurons that was replicated for almost all afferent modalities (Extended Data Fig. 8b).

Our information flow analysis provides a compressed representation of the connectome, but currently ignores signs of connections (neurotransmitter identity) and the biophysics of neurons and synapses, and therefore terms such as ‘early’ and ‘late’ should not be interpreted as true latencies to sensory stimulation. Shiu et al. and Pospisil et al. use the connectome to model Drosophila brain dynamics and include connection weights (number of synapses) and putative connection signs (excitatory or inhibitory).