Neuronal wiring diagram of an adult brain

We imported the automatically predicted synapses from Buhmann et al., which we combined with the synapse segmentations by Heinrich et al. to assign scores to all synapses to improve precision. Buhmann et al. introduced a machine learning model to predict for each voxel whether it is part of a postsynaptic site. For voxels classified as postsynaptic a vector to the presynaptic site is predicted which is then used to created synaptic connections. Hence, synaptic partners predicted by Buhmann et al. are represented by a connector between a postsynaptic and a presynaptic location without further annotation about the size of the synapse. Heinrich et al. on the other hand segmented the synaptic clefts. Buhmann et al. suggested using the probability maps from Heinrich et al. to improve performance by locating the highest probability score along their predicted connectors (called score in the next paragraph).

The synapse classifier by Buhmann et al. was trained on ground truth from the CREMI challenge (https://cremi.org). The three CREMI datasets contain three 5 × 5 × 5 µm cubes from the calyx in FAFB14 with 1,965 synapses. While the classifier from Buhmann et al. was trained and validated on only this dataset, they evaluated its performance on multiple regions (calyx, lateral horn, ellipsoid body and protocerebral bridge). It should be noted that performance varies by region.

The dataset published by Buhmann et al. contained ~244 million synapses. We removed synapses from the imported list if they fulfilled any of the following criteria: (1) either the pre- or postsynaptic location remained unassigned to a segment (proofread or unproofread); (2) It had a score ≤50. Additionally, we removed duplicate synapse annotations between the same pre- and postsynaptic partners, defined as those within a distance of 100 nm from another synapse annotation according to their presynaptic coordinate. After filtering, we were left with ~130 million synapses.

Eckstein et al. created a machine learning model to predict neurotransmitter identities for all synapses from Buhmann et al. based on the electron microscopy imagery alone. Each synapse was assigned a probability for one of six neurotransmitters: acetylcholine, glutamate, GABA, serotonin, dopamine and octopamine. They used neurotransmitter identities published for individual neuronal cell types and built a dataset with 3,025 neurons with known transmitter type assuming Dale’s law applies. Eckstein et al. reported a per-synapse accuracy of 87% and a per neuron (majority vote) accuracy of 94%.

The methods described in this section used the FAFB14 version of the electron microscopy stack. We applied a transformation to all synapses to map them into the FlyWire FAFB14.1 space. The vector field for the transformation had a resolution of 64 × 64 × 40 nm.

Connection threshold

For all the analyses presented in this paper, save for synapse distributions, we employed a consistent threshold of >4. Our decision to use a synapse threshold on connections was due partly to the fact that synapses in the FlyWire’s brain dataset were not manually proofread. For these analyses, many of which demonstrate the high interconnectivity of the fly brain, we chose a conservative threshold to ensure that considered connections are real. Use of a threshold is also in keeping with previous work analysing wiring diagrams in Drosophila. Thus, we are probably undercounting the number of true connections. The distribution of synapse counts (Fig. 3f) does not display any bimodality that could be used to set the threshold. Therefore, the choice of 5 synapses per connection is a reasonable but arbitrary one. By analysing the network properties of the FlyWire brain connectome, Lin et al. found that statistical properties of the whole-brain network, such as reciprocity and clustering coefficient, are robust to our choice of threshold. The FlyWire data are available without an imposed threshold, so users can choose their own appropriate threshold for their specific use case.

Neuropil projectome construction

Under the simplifying assumptions that information flow through the neuron can be approximated by the fraction of synapses in a given region, and that inputs and outputs can be treated independently, we can construct a matrix representing the projections of a single neuron between neuropils. The fractional inputs of a given neuron are a 1 × N vector containing the fraction of incoming synapses the neuron has in each of the N neuropils, and the fractional outputs are a similar vector containing the fraction of outgoing synapses in each of the N neuropils. We multiply these vectors against each other to generate the N × N matrix of the neuron’s fractional weights. Summing these matrices across all intrinsic neurons produces a matrix of neuropil-to-neuropil connectivity (Fig. 4a). In this projectome, all neurons contribute an equal total weight of one.

Dominant input side

We assigned neuropils to the left and right hemispheres or the centre if the neuropil has no homologue. We then counted how many postsynapses each neuron had in each of these three regions and assigned it to the one with the largest count.

Contralateral and bilateral neuron analysis

For each neuron, we calculated the fraction of presynapses in the left and right hemisphere. The hemisphere opposite its dominant input side was named the contralateral hemisphere. We excluded neurons that had either most of their presynapses or most of their postsynapses in the centre region.

Rank analysis and information flow

We used the information flow algorithm implemented by Schlegel et al. (https://github.com/navis-org/navis) to calculate a rank for each neuron starting with a set of seed neurons. The algorithm traverses the synapse graph of neurons probabilistically. The likelihood of a neuron being added to the traversed set increased linearly with the fraction of synapses it receives from already traversed neurons up to 30% and was guaranteed above this threshold. We repeated the rank calculation for all sets of afferent neurons as seed as well as the whole set of sensory neurons. The groups we used are: olfactory receptor neurons, gustatory receptor neurons, mechanosensory Johnston’s organ neurons, head and neck bristle mechanosensory neurons, mechanosensory taste peg neurons, thermosensory neurons, hygrosensory neurons, VPNs, visual photoreceptors, ocellar photoreceptors and ascending neurons.

Additionally, we created input seeds by combining all listed modalities, all sensory modalities, and all listed modalities with visual sensory groups excluded.

For each modality we performed 10,000 runs, which were averaged. We then ordered the neurons according to their rank and assigned them a percentile based on their location in the order. To compute a reduced dimensionality, we treated the vector of all ranks (one for each modality) as neuron embedding and calculated two dimensional embeddings using UMAP with the following parameters: n_components=2, min_dist=0.35, metric = “cosine”, n_neighbors=50, learning_rate =.1, n_epochs=1000.