Neuronal parts list and wiring diagram for a visual system

Reconstruction accuracy and completeness

The overall quality of our Drosophila brain reconstruction has been evaluated elsewhere^24,31 (a summary of the current status is shown in Extended Data Table 3). Here we describe a few additional checks that are specific to the optic lobe. A small percentage of cells have eluded proofreading efforts. The worst cases are some types with visible ‘bald spots’ in the mid posterior side of the right optic lobe (Supplementary Data 2). In this region, we observed a narrowing and discontinuation of neuronal tracks. Many of these tracks appear to terminate within glial cells, suggesting a potential engulfment of neurons by glia. For most types, under-recovery is hardly visible (Supplementary Data 2).

For a quantitative estimate of under-recovery, we can rely on the ‘modular’ types²⁷, defined as cell types that are in one-to-one correspondence with columns. A previous reconstruction of seven medulla columns identified 20 modular types²⁸. These largely correspond to the cell types that contain from 720 to 800 cells in our reconstruction (Fig. 1d). The top end (800) of this range is probably the true number of columns in this optic lobe. The lower end of this range is 720, suggesting that under-recovery is 10% at most, and typically less than that.

The inner photoreceptors R7 and R8 are about 650 cells each, and the outer photoreceptors R1–6 total about 3,400 in version 783 of the FlyWire connectome. These numbers are not inconsistent with modularity because photoreceptors are especially challenging to proofread in this dataset and under-recovery is higher than typical.

In the left optic lobe, we have proofread around 38,500 intrinsic neurons, as well as 3,700 VPNs, 250 VCNs, 150 heterolateral neurons and 5,000 photoreceptor cells. Tables comparing precise left/right counts by superclass as well as by type are available for download (see the ‘Data availability’ section).

Tm21 (also known as Tm6), Dm2, TmY5a, Tm27 and Mi15 are substantially less numerous than 800, so we agree with the seven column reconstruction²⁸ that they are not modular. On the other hand, some of our types (T2a, Tm3, T4c and T3) contain more than 800 proofread cells (Fig. 1d), which violates the definition of modularity. This partially agrees with the seven column reconstruction²⁸, which regarded T3 and T2a as modular, and T4 and Tm3 as not modular. T4 is an unusual case, as T4c is above 800 while the other T4 types are below 800. It should be noted that all of the above cell numbers could still creep upward with further proofreading.

A genuine analysis of modularity requires going beyond simple cell counts, and analysing locations to check the idea of one-to-one correspondence. Such an analysis is left for future work. Here we apply the term ‘numerous’ to those types containing 720 or more cells, as well as photoreceptor types, and do not commit to whether these types are truly modular.

The seven column reconstruction²⁸ provided a matrix of connections between their modular types. This shows good agreement with our data (Methods and Extended Data Fig. 9), providing a check on the accuracy of our reconstruction in the optic lobe. This validation complements the estimates of reconstruction accuracy in the central brain that are provided in the flagship paper²⁴.

The major limitation of our reconstruction in the optic lobe concerns the automatically detected synapses⁷⁷. Although accuracy is high overall, outgoing photoreceptor synapses are markedly underdetected. This may be because dark cytoplasm (characteristic of photoreceptors) is not well represented in the example synapse images that were used to train the automated synapse detector. Example images of photoreceptor synapses have been included in the training set of an improved automated synapse detector, but the results were not ready in time for this publication, and will be made available in a future release. The classification of inner photoreceptors as yellow and pale is postponed until the future release. In the present paper, the connectivity from photoreceptors to other cell types in this paper is only qualitative and not quantitative. Furthermore, underdetection of photoreceptor synapses could affect the input fractions of other connections due to normalization.

Another cautionary note is that weaker connections in the type–type connectivity matrix (Extended Data Fig. 4) could be artifactual, due to false positives of automated synapse detection. There are some heuristics for guessing whether a connection is artifactual, short of manually inspecting the original EM images. For example, one might distrust weak connections between cells, that is, those with less than some threshold number of synapses. The choice of the threshold value depends on the context⁹. For example, the flagship paper²⁴ discarded connections with less than five synapses, a convention followed by the FlyWire Codex. The predicates of the present work apply a threshold of two synapses rather than five. The different thresholds were chosen because the central brain and optic lobes are very different contexts, as we now explain.

In the central brain, most cell types have cardinality 2 (cell and its mirror twin in the opposite hemisphere; Extended Data Fig. 1e). In the hemibrain, the cardinality is typically reduced to one. Therefore, whether there is a connection between cell type A and cell type B must be decided based on only two or three examples of the ordered pair (A, B) in all the connectomic data that is so far available. Given the small sample size, it makes sense to set the threshold to a relatively high value, if false positives are to be avoided.

On the other hand, in the optic lobe, there are often many examples of the ordered pair (A, B), because so many cell types have high cardinality. Therefore, if a connection is consistently found from type A to type B, one can have reasonable confidence even if the average number of synapses in the connection is not so high. That is why we set the threshold to a relatively low value in the optic lobe predicates. In particular, we have found that certain inhibitory types consistently make connections that involve relatively few synapses, and these connections seem real.

Another heuristic is to look for extreme asymmetry in the matrix. If the number of synapses from A to B is much larger than from B to A, the latter connection might be spurious. The reason is that the strong connection from A to B means the contact area between A and B is large, which means more opportunity for false-positive synapses from B to A. False-positive rates for synapses are estimated in the flagship paper²⁴.

Finally, it may be known from other studies that a connection does not exist. For example, T1 cells lack output synapses^26,78. Therefore, in our analyses, we typically regarded the few outgoing T1 synapses in our data as false positives and discarded them.

Morphological cell typing

Our connectomic cell approach to typing is initially seeded with some set of types, to define the feature vectors for cells (Fig. 2a), after which the types are refined by computational methods. For the initial seeding, we relied on the time-honoured approach of morphological cell typing, sometimes assisted by computational tools that analysed connectivity. It is worth noting that ‘morphology’ is a misnomer, because it refers to shape only, strictly speaking. Orientation and position are actually more fundamental properties because of their influence on stratification in neuropil layers. Thus, ‘single-cell anatomy’ would be more accurate than morphology, although the latter is the standard term.

Stage 1: crowdsourced annotation of known types

Annotations of optic lobe neurons were initially crowdsourced. The first annotators were volunteers from Drosophila laboratories. They were later joined by citizen scientists. At this stage, the annotation effort was mainly devoted to labelling cells of known types, especially the most numerous types.

Drosophila lab annotators

E.K. and D.G. proofread and annotated medulla neurons that were upstream of the anterior visual pathway. These included many of the medulla and lamina neurons discussed in this study. The annotated neurons were primarily Dm2, Mi15, R7, and R8, but also comprised various L, Dm, Mi, Tm, C and Sm cells. Previously known neuron types were identified primarily by morphology and partially by connectivity. Annotators additionally found all Mi1 neurons in both hemispheres to find every medulla column. These Mi1 neurons were used to create a map of medulla layers based on Mi1 stratification⁶, which later aided citizen scientists to identify medulla cell types.

Citizen scientists

The top 100 players from Eyewire⁷⁹ had been invited to proofread in FlyWire²⁴. After 3 months of proofreading in the right optic lobe, they were encouraged to also label neurons when they felt confident. Most citizen scientists did a mixture of annotation and proofreading. Sometimes they annotated cells after proofreading, and other times searched for cells of a particular type to proofread.

Citizen scientists were provided with a visual guide to optic lobe cells sourced from the literature^6,80. FlyWire made available a 3D mesh overlay indicating the four main optic lobe neuropils. Visual identification was primarily based on single-cell anatomy. Initially, labelling of type families (that is, Dm, Tm, Mi and so on) was encouraged, especially for novices. Annotation of specific types (such as Dm3, Tm2) developed over time. The use of canonical names was further enforced by a software tool that enabled easy selection and submission of preformatted type names.

Additional community resources (discussion board/forum, blog, shared Google drive, chat, dedicated email and Twitch livestream) fostered an environment for sharing ideas and information between community members (citizen scientists, community managers and researchers). Community managers answered questions, provided resources such as the visual guide, shared updates, performed troubleshooting and general organization of community activity. Daily stats including number of annotations submitted per individual were shared on the discussion board/forum to provide project progress. Live interaction, demonstrations and communal problem solving occurred during weekly Twitch video livestreams led by a community manager. The environment created by these resources allowed citizen scientists to self-organize in several ways: community driven information sharing, programmatic tools and ‘farms’.

Community-driven information sharing

Citizen scientists created a comprehensive guide with text and screenshots that expanded on the visual guide. They also found and studied any publicly available scientific literature or resources regarding the optic lobe. They shared findings at discuss.flywire.ai, which as of 10 October 2023 had over 2,500 posts. Community managers interacted with citizen scientists by sharing findings from the scientific literature, consulting Drosophila specialists on FlyWire and providing feedback.

Programmatic Tools

Programmatic tools were created to help with searching for cells of the same type. One important script traced partners-of-partners, that is, source cell→downstream partners→their upstream partners, or source cell→upstream partners→their downstream partners. This was based on the assumption that cells of the same type will probably synapse with the same target cells, which often turned out to be true. The tool could either look for partners-of-all-partners or partners-of-any-partners. The resulting lists of cells could be very long, and were filtered by excluding cells that had already been identified, or excluding segments with small sizes or low ID numbers (which had probably not yet been proofread). Another tool created from lobula plate tangential cells (for example, HS, VS, H1) aided definition of layers in the lobula plate. This facilitated identification of various cell types, especially T4 and T5.

Cell farms

Citizen scientists created farms in FlyWire or Neuroglancer with all the found cells of a given type visible. Farms showed visually where cells still remained to be found. If they found a bald spot, a popular method to find missing cells was to move the 2D plane in that place and add segments to the farm one after another in search of cells of the correct type. Farms also helped with identifying cells near to the edges of neuropils, where neurons are usually deformed. Having a view of all other cells of the same type made it possible to extrapolate to how a cell at the edge should look.

Stage 2: centralized annotation and discovery of new types

A team of image analysts at Princeton finished the annotation of the remaining cells in known types, and also discovered new types. Community annotations were initially compared with existing literature to confirm accuracy. Once validated, these cells were used to query various Codex search tools that returned previously unannotated cells exhibiting connectivity similar to that of the cell in the query. The hits from the search query were evaluated by morphology and stratification to confirm match with the target cell type. In some cases in which cell type distinctions were uncertain, predicted neurotransmitters⁴⁵ were used for additional guidance. This process enabled us to create a preliminary clustering of all previously known and new types.

Connectomic cell typing

Eventually morphology became insufficient for further progress. Expert annotators, for example, struggled to classify Tm5 cells into the three known types, not knowing that there would turn out to be six Tm5 types. At this point, we were forced to transition to connectomic cell typing. In retrospect, this transition could have been made much earlier. As mentioned above, connectomic cell typing must be seeded with an initial set of types, but the seeding did not have to be as thorough as it ended up. We leave for future work the challenge of extending the connectomic approach so it can be used from start to finish.

Stage 3: connectivity-based splitting and merging of types and auto-correction

We used computational methods to split types that could not be properly split in stage 2. Some candidates for splitting (such as Tm5) were suggested by the image analysts. Some candidates were suspicious because they contained so many cells. Finally, some candidates were scrutinized because their type radii were large. We applied hierarchical clustering with average linkage, and accepted the splits if they did not violate the tiling principle as described in the ‘Spatial coverage’ section.

We also applied computational methods to merge types that had been improperly split in stage 2. Here the candidates were types with low spatial coverage of the visual field, or types that were suspiciously close in the dendrogram of cell types (Fig. 2c). Merge decisions were made by hierarchical clustering of cells from types that were candidates for merging, and validated if they improved spatial coverage.

Once we arrived at the final list of types, we estimated the ‘centre’ of each type using the element-wise trimmed mean. Then, for every cell, we computed the nearest type centre by Jaccard distance. For 98% of the cells, the nearest type centre coincided with the assigned type. We sampled some disagreements and reviewed them manually. In the majority of cases, the algorithm was correct, and the human annotators had made errors, usually of inattention. The remaining cases were mostly attributable to proofreading errors. There were also cases in which type centres had been contaminated by human-misassigned cells (see the ‘Morphological variation’ section), which in turn led to more misassignment by the algorithm. After addressing these issues, we applied the automatic corrections to all but 0.1% of cells, which were rejected using distance thresholds.

Validation

On the basis of the auto-correction procedure, we estimate that our cell type assignments are between 98% and 99.9% accurate. For another measure of the quality of our cell typing, we computed the ‘radius’ of each type, defined as the average distance from its cells to its centre. Here we computed the centre by approximately minimizing the sum of Jaccard distances from each cell in the type to the centre (see the ‘Computational concepts’ section). A large type radius can be a sign that the type contains dissimilar cells, and should be split. For our final types, the radii vary, but almost all lie below 0.6 (Extended Data Fig. 3a). Lat has an exceptionally high type radius, and deserves to be split (see the ‘Cross-neuropil tangential and amacrine’ section). The type radii are essentially the same, whether or not boundary types are included in the feature vector (data not shown).

Discrimination with logical predicates

Because the feature vector is rather high dimensional, it would be helpful to have simpler insights into what makes a type. One approach is to find a set of simple logical predicates based on connectivity that predict type membership with high accuracy. For a given cell, we define the attribute ‘is connected to input type t’ as meaning that the cell receives at least one connection from some cell of type t. Similarly, the attribute ‘is connected to output type t’ means that the cell makes at least one connection onto some cell of type t.

An optimal predicate is constructed for each type that consists of 2 tuples: input types and output types. Both tuples are limited to size 5 at most, and they are optimal with respect to the F-score of their prediction of the subject type, defined as follows:

• Recall of a predicate for type T is the ratio of true positive predictions (cells matching the predicate) to the total number of true positives (cells of type T). It measures the predicate’s ability to identify all positive instances of a given type.

• Precision is the ratio of true positive predictions (predictions that are indeed of type T) to the total number of positive predictions made by the logical predicate.

• F-score is the harmonic mean of precision and recall—a single metric that combines both precision and recall into one value.

On a high level, the process for computing the predicates is exhaustive—for each type, we look for all possible combinations of input type tuples and output type tuples and compute their precision, recall and F-score. A few optimization techniques are used to speed up this computation, by calculating minimum precision and recall thresholds from the current best candidate predicate and pruning many tuples early.

For example, the logical predicate ‘is connected to input type Tm9 and output type Am1 and output type LPi15’ predicts T5b cells with 99% precision and 99% recall. For all but three of the identified types, we found a logical predicate with 5 or fewer input/output attributes that predicts type membership with an average F-score of 0.93, weighted by the number of cells in type (Extended Data Fig. 4 and Supplementary Data 1). Some of the attributes in a predicate are the top most connected partner types, but this is not necessarily the case. The attributes are distinctive partners, which are not always the most connected partners. The predicate for each type is shown on its card in Supplementary Data 2. For each family, the predicates for all types can be shown together in a single graph containing all of the relevant attributes (Supplementary Data 3).

We experimented with searching for predicates after randomly shuffling a small fraction of types (namely, swapping types for 5% of randomly picked pairs of neurons). We found that precision and recall of the best predicates dropped substantially, suggesting that we are not overfitting. This was expected because the predicates are short.

We also measured the drop in the quality of predicates if excluding boundary types (where the predicates are allowed to contain intrinsic types only). As is the case with the clustering metrics, the impact on predicates is marginal (weighted mean F-score drops from 0.93 to 0.92).

Discrimination with two-dimensional projections

Another approach to interpretability is to look at low-dimensional projections of the 2T-dimensional feature vector. For each cell type, we select a small subset of dimensions that suffice to accurately discriminate that type from other types (Extended Data Fig. 3c). Here we normalize the feature vector so that its elements represent the ‘fraction of input synapses received from type t’ or ‘fraction of output synapses sent to type t’. In these normalized quantities, the denominator is the total number of all input or output synapses, not just the synapses with other neurons intrinsic to the optic lobe.

For example, we can visualize all cells in the Pm family in the two-dimensional space of C3 input fraction and TmY3 output fraction (Extended Data Fig. 3c). In this space, Pm04 cells are well-separated from other Pm cells, and can be discriminated with 100% accuracy by ‘C3 input fraction greater than 0.01 and TmY3 output fraction greater than 0.01’. This conjunction of two features is a more accurate discriminator than either feature by itself.

More generally, a cell type discriminator is based on thresholding a set of input and output fractions, and taking the conjunction of the result. The search for a discriminator finds a set of dimensions, along with threshold values for the dimensions. To simplify the search, we require that the cell type be discriminated only from other types in the same neuropil family, rather than from all other types. Under these conditions, it almost always suffices to use just two dimensions of the normalized feature vector.

Discriminators for all types in all families containing more than one type are provided in Supplementary Data 4. Many although not all discriminations are highly accurate. Both intrinsic and boundary types are included as discriminative features.

Computational concepts

Connectivity: cell-to-cell, type-to-cell, cell-to-type and type-to-type

Define a (weighted) cell-to-cell connectivity matrix w_ij, as the number of synapses from neuron i to neuron j. The weighted out-degree and in-degree of neuron i are:

The sums are over all neurons in the brain. If neuron i is a cell intrinsic to one optic lobe, the only nonvanishing terms in the sums are due to the intrinsic and boundary neurons for that optic lobe.

Let A_it be the 0–1 matrix that assigns neuron i to type t. The column and row sums of the assignment matrix satisfy

where n_t is the number of cells assigned to type t.

The cell-to-type connectivity matrix O_it is the number of output synapses from neuron i to neurons of type t,

For fixed i, O_it is known as the output feature vector of cell i. Similarly, the type-to-cell connectivity matrix I_tj is the number of input synapses from neurons of type t onto neuron j,

For fixed j, I_tj is known as the input feature vector of cell j. The ith row and ith column of these matrices are concatenated to form the full feature vector for cell i (Fig. 2a).

The input and output feature vectors can be normalized by degree to yield input and output fractions of cell i, O_it/d_i⁺ and I_ti/d_i⁻. Elements of these matrices are used for the discriminating 2D projections (Extended Data Fig. 3c).

The type-to-type connectivity matrix is the number of synapses from neurons of type s to neurons of type t,

The weighted degree of type t is the sum of the weighted degrees of the cells in type t,

The sums are over all neurons in the brain, similar to equation (1). Normalizing by degree yields the output fractions of type s, W_st/D_s⁺, where t runs from 1 to T. The input fractions of type t are similarly given by W_st/D_t⁻, where s runs from 1 to T. Selected output and input fractions of types are shown in Supplementary Data 5.

Alternatively, the feature vectors can be based on connection number rather than synapse number, where a connection is defined as two or more synapses from one neuron to another. Then, weighted degree is replaced by unweighted degree in the above definitions. The threshold of two synapses is intended to suppress noise due to false positives in the automated synapse detection. Synapse number and connection number give similar results, and we use both in our analyses.

We found that it was sufficient for feature dimensions to include only intrinsic types (T = 227). Alternatively, feature dimensions can be defined as including both intrinsic and boundary types (T > 700), and this yields similar results (data not shown).

For the hierarchical clustering of cell types (Fig. 2c), the feature vector for each cell type is obtained by concatenating the vectors of input and output fractions for that cell type.

Similarity and distance measures

The weighted Jaccard similarity between feature vectors x and y is defined by

and the weighted Jaccard distance d(x,y) is defined as one minus the weighted Jaccard similarity. These quantities are bounded between zero and one since our feature vectors are nonnegative. In our cell typing efforts, we have found empirically that Jaccard similarity works better than cosine similarity when feature vectors are sparse.

Type centres

Given a set of feature vectors x^a, the centre c can be defined as the vector minimizing

This cost function is convex, as d is a metric satisfying the triangle inequality. Therefore, the cost function has a unique minimum. We used various approximate methods to minimize the cost function.

For auto-correction of type assignments, we used the element-wise trimmed mean. We found empirically that this gave good robustness to noise from false synapse detections. For the type radii, we used a coordinate descent approach, minimizing the cost function with respect to each c_i in turn. The loop included every i for which some x_i was non-zero. This converged within a few iterations of the loop.

Hierarchical clustering of cell types

The type-to-type connectivity matrix of equation (5) was the starting point for clustering the cell types. For each cell type, the corresponding row and column of the matrix were normalized to become input and output fractions, as described in the text following equation (6), and then concatenated (this is yet another way of computing type centres). Feature vectors included only dimensions corresponding to cell types intrinsic to the optic lobe. Then, average linkage hierarchical clustering was applied to yield a dendrogram (Fig. 2c). The dendrogram was thresholded to produce a flat clustering (Fig. 2c).

The precise memberships in the clusters warrant cautious interpretation, as the clusters are the outcome of just one clustering algorithm (average linkage), and differ if another clustering algorithm is used. Each cluster contains core groups of types that are highly similar to each other, that is, types that merge early during agglomeration (closer to the circumference of the dendrogram). These are more certain to have similar visual functions, and tend to be grouped together by any clustering algorithm. Types that are merged late (closer to the origin of the dendrogram) are less similar, and their cluster membership is more arbitrary. Some degree of arbitrariness is inevitable when one divides the visual system into separate subsystems, because subsystems interact with each other, and types that mediate such interactions are borderline cases.

Each cluster is generally a mixture of types from multiple neuropil families. Sceptics might regard such mixing as arising from the ‘noisiness’ in the clustering noted above at the largest distances. Indeed, the nearest types, those that merge in the dendrogram farther from the centre (Fig. 2c), tend to be from the same neuropil family. But plenty of dendrogram merges between types of different families happen at intermediate distances rather than the largest distances. Thus, some of the mixing of types from different neuropil families seems genuinely rooted in biology.

Wiring diagrams

Reduction

To make the wiring diagrams readable, we display only the top type-to-type connections, which are defined as follows. For every cell type, the top input cell type and top output cell type are selected by ranking connected partners by the total number of synapses in the connection. If cell types are nearly tied, any runner up within 5% of the winner is also displayed. Figure 3 shows the top connections between all optic lobe intrinsic types. Figures 4–7 each focus on one or a few subsystems, but also include the top input/output connections they participate in with the rest of the network as well as top output connections to boundary types (for example, in Fig. 4, Dm2 is selected because it belongs to cluster 5, luminance channel, but then also other types outside of ON, OFF, and luminance channels are included because either Dm2 is their top input/output type or the other way around). Extended Data Figs. 5 and 6 show the top input and top output connections separately, for improved readability. For the top output connections we also include boundary types (VPNs).

Colours and shapes

Nodes, representing cell types, are coloured by clusters. Node size encodes the number of drawn connections, so that types that are top input/output of many other types look larger. Node shapes encode type numerosities (number of cells of that type), from most numerous (hexagon) to least (ellipse) (see the figure legends). The lines indicate connections between cell types. The line colour encodes the relationship (top input or top output) and the line width is proportional to the number of synapses connecting the respective types. The line arrowheads encode neurotransmitter predictions (excitatory/cholinergic or inhibitory/GABAergic/glutamatergic).

Layout

We used Cytoscape⁸¹ to draw the wiring diagrams. Organic layout was used for Figs. 3 and 7c, and hierarchical layout was used for the others. The hierarchical layout tries to make arrows point downwards. After Cytoscape automatically generated a diagram, nodes were manually shifted by small displacements to minimize the number of obstructions.

Intrinsic versus boundary

The optic lobes are divided into five regions (neuropils): lamina of the compound eye (LA); medulla (ME); accessory medulla (AME); lobula (LO); lobula plate (LOP). All non-photoreceptor cells with synapses in these regions are split into two groups: optic lobe intrinsic neurons and boundary neurons.

Optic lobe intrinsic neurons are almost entirely contained in one of the optic lobes (left or right), more precisely, 95% or more of their synapses are assigned to the five optic lobe regions listed above.

Boundary neurons are those with at least 5% (and less than 95%) of synapses in the optic lobe regions, and are either visual projection, visual centrifugal or heterolateral neurons.

Axon versus dendrite

In the main text (in the ‘Class, family and type’ section), we used the term ‘axon’. An axon is defined as some portion of the neuron with a high ratio of presynapses to postsynapses. This ratio might be high in an absolute sense. Or the ratio in the axon might only be high relative to the ratio elsewhere in the neuron (the dendrite). In either case, the axon is typically not a pure output element, but has some postsynapses as well as presynapses. For many types it is obvious whether there is an axon, but for a few types we have made judgement calls. Even without examining synapses, the axon can often be recognized from the presence of varicosities, which are presynaptic boutons. The opposite of an axon is a dendrite, which has a high ratio of postsynapses to presynapses.

An amacrine cell is defined as one for which the axon–dendrite distinction does not hold, and presynapses and postsynapses are intermingled in roughly the same ratio throughout. The branches of an amacrine cell are often called dendrites, but the neutral term ‘neurite’ is perhaps better for avoiding confusion.

Columnar neurons

Fischbach and Dittrich⁶ defined 13 columnar families based on neuropils (Fig. 1a). Families consisting exclusively of ‘numerous’ (∼800 cells) types include L (lamina to medulla), C (medulla to lamina), T1 (distal medulla to lamina), T2 (distal and proximal medulla to lobula), T3 (proximal medulla to lobula), T4 (proximal medulla to lobula plate) and T5 (lobula to lobula plate). We follow the convention of grouping the less numerous Lawf1 (distal medulla to lamina) and Lawf2 (proximal and distal medulla to lamina) types in the same family, despite the differences between their neuropils and connectivity. Although T1 shares the same neuropils with Lawf1, T1 lacks output synapses^26,78, so it is an outlier and deserves to be a separate family. Distal and proximal medulla are regarded as two separate neuropils⁶.

Mi

Fischbach and Dittrich⁶ defined Mi as projecting from distal to proximal medulla. Mi contains both numerous and less numerous types. We identified five (Mi1, 2, 4, 9, 10) of the dozen Mi types originally defined⁶, and three (Mi13, 14, 15) types uncovered by EM reconstruction²⁷. Mi1, Mi4, and Mi9 are consistent with the classical definition, but Mi13 projects from proximal to distal medulla. Other Mi types are less polarized, and the term “narrow-field amacrine” might be more accurate than “columnar”. Nevertheless we will adhere to the convention that they are columnar. Narrow-field amacrine cells are also found in the Sm family, and exist in the mammalian retina⁸².

Tm transmedullary

As classically defined⁶, Tm cells project from the distal medulla to the lobula. Tm1 through Tm26 and Tm28 were defined⁶, and Tm27/Tm27Y was reported later⁸³. We were able to identify Tm1, 2, 3, 4, 7, 9, 16, 20, 21, 25 and 27. We split Tm5 into six types, and Tm8 into two types. We merged Tm6 and Tm21 into a single type Tm21. We prefer the latter name because the cells more closely match the Tm21 stratification as drawn by Fischbach and Dittrich⁶. Tm1a and Tm4a were defined as morphological variants⁶, but we have found that they do not differ in connectivity and are not common, so we have merged them into Tm1 and Tm4, respectively. We merged Tm27Y into Tm27⁸³. TmY5 was merged into TmY5a^6,84, the name that has appeared more often in the literature. These morphological distinctions originally arose because the projection into the lobula plate, the differentiator between Tm and TmY, can vary across cells in a type. We added new types Tm31 to Tm37, which project from the serpentine medulla to the lobula. We moved Tm23 and Tm24 to the Li family. They were originally classified as Tm because their cell bodies are in the distal rind of the medulla, and they send a neurite along the columnar axis of the medulla to reach the lobula⁶. However, they do not form synapses in the medulla, so we regard them as Li neurons despite their soma locations. Overall, around half of the 26 types in the Tm family are new.

TmY

TmY cells project from the distal medulla to the lobula and lobula plate. The Y refers to the divergence of branches to the lobula and lobula plate. Previous definitions include TmY1 to TmY13⁶; TmY5a^6,84; TmY14²⁷; TmY15²⁹; and TmY16, TmY18 and TmY20³⁰. We identified TmY3, TmY4, TmY5a, TmY10, TmY11, TmY14, TmY15, TmY16 and TmY20. We divided TmY9 into two types, as discussed in a companion paper⁶⁰. We added a new type, TmY31.

Y

Y cells project from the proximal medulla to the lobula and lobula plate. They are similar to TmY cells, but the latter traverse both the distal and proximal medulla⁶. Previous definitions were Y1 and Y3 to Y6⁶; and Y11 and Y12¹⁰. We have identified Y1, Y3, Y4, Y11 and Y12 in our reconstruction, and have not found any new Y types. Y1, Y11 and Y12 have the majority of their synapses in the lobula plate, and are assigned to the motion subsystem. Y3 and Y4 have few synapses in the lobula plate, and are assigned to the object subsystem (Fig. 2). Y3 is more numerous (∼300 cells) than Y4, and is the only Y type that is predicted cholinergic.

Tlp

A Tlp neuron projects from the lobula plate to the lobula. Tlp1 to Tlp5 were defined first⁶, and Tlp11 to Tlp14 were defined later on¹⁰. We have identified Tlp1, Tlp4, Tlp5 and Tlp14. We propose that the names Tlp11, Tlp12 and Tlp13 should be retired¹⁰, as these types can now be unambiguously identified with Tlp5, Tlp1 and Tlp4, respectively.

Interneurons

A local interneuron is defined as being completely confined to a single neuropil (Fig. 1b). Interneurons make up the majority of types, but a minority of cells (Fig. 1e). Lai is the only lamina interneuron. Dm and Pm interneurons⁶ stratify in the distal or proximal medulla, respectively. We have more than doubled the number of Pm types, and slightly increased the number of Dm types. We introduce the Sm family, which is almost completely new and contains more types than any other family (Fig. 1f). Li and LPi interneurons stratify in the lobula or lobula plate, respectively. Interneurons are usually amacrine and presumed inhibitory (GABA or glutamate), but some are tangential or cholinergic. Interneurons are often wide field but some are narrow field.

Dm

Dm1 to Dm8⁶; Dm9 and 10²⁷; and Dm11 to Dm20⁸⁵ were previously defined. We do not observe Dm5 and Dm7, consistent with a previous study⁸⁵. Most types are predicted to secrete glutamate or GABA, but there are also a few cholinergic types (Supplementary Data 1). To Dm3p and Dm3q^61,62,85, we added a third type, Dm3v (Supplementary Data 2). We split Dm8 into Dm8a and Dm8b (see the ‘Correspondences with molecular–morphological types’ section).

DmDRA

The DRA differs from the rest of the retina in its organization of inner photoreceptors. Photoreceptors in non-DRA and DRA differ in their axonal target layers and output cell types^54,86. Specifically, DRA-R7 connects with DmDRA1, whereas DRA-R8 connects to DmDRA2^54,87. These distinctive connectivity patterns result in DmDRA1 and DmDRA2 types exhibiting an arched coverage primarily in the M6 layer of the dorsal medulla (Fig. 9b). R7-DRA and R8-DRA are incompletely annotated at present, and this will be rectified in a future release. DmDRA1 receives R7 input, but sits squarely in M7. This could be regarded as an Sm type, but we have chosen not to change the name for historical reasons.

Pm

Pm1, 1a and 2⁶ were each split into two types. Pm3 and 4 remain as previously defined⁸⁵. We additionally identified six new Pm types, for a total of 14 Pm types, numbered Pm01 to Pm14 in order of increasing average cell volume. The new names can be distinguished from the old ones by the presence of leading zeros. All are predicted GABAergic. Pm1 was split into Pm06 and Pm04, Pm1a into Pm02 and Pm01, and Pm2 into Pm03 and Pm08.

Sm

Dm and Pm interneurons are defined⁶ to stratify on the distal or proximal side, respectively, of the serpentine layer (M7) of the medulla. Many interneuron types turn out to have significant stratification in the serpentine layer, and these borderline cases constitute a large new Sm family of interneurons, almost all new. They have been named Sm01 to Sm43, mostly in order of increasing average cell volume. The Sm family includes types recently named medulla tangential intrinsic⁴². We avoid using this term indiscriminately because some Sm types are tangential while others are amacrine. Some Sm types spill over from M7 into the distal or proximal medulla, and a few reach from M7 to more distant medulla layers.

Sm stratification in M7 has functional implications. First, Sm types are positioned to communicate with the medulla tangential (Mt) cells and other boundary types that are important conduits of information in and out of the optic lobe (Supplementary Data 5). Second, Sm types are positioned to communicate with the inner photoreceptor terminals, which are in M6 or at the edge of M7. Consequently many Sm types are involved in the processing of chromatic stimuli, and end up being assigned to the colour subsystem.

The Sm family more than doubles the number of medulla interneuron types, relative to the old scheme with only Pm and Dm. The Sm family might be related to the M6-LN class of neuron previously defined⁸⁸. The correspondence is unclear because M6-LN neurons are defined to stratify in M6, while Sm mainly stratifies in M7. But some Sm types stratify at the border between M6 and M7, and therefore could be compatible with the M6-LN description.

Li

After two lobula intrinsic types (Li1 and Li2) were initially defined⁶, 12 more (Li11 to 20 and mALC1 and mALC2) were identified by the hemibrain reconstruction⁹. Of these, we have confirmed Li2, Li12, Li16, mALC1 and mALC2. We identified 21 additional Li types, but have not been able to make conclusive correspondences with previously identified types. As mentioned earlier, we transfer Tm23 and Tm24⁶ from the Tm to the Li family. This amounts to a total of 33 Li types, which have been named Li01 to Li33 in order of increasing average cell volume.

Collisions with Li1 and Li2⁶ are avoided by the presence of leading zeros in our new names. The hemibrain names Li11 to Li20 and mALC1 and mALC2⁹ have been used by few or no publications, so there is little cost associated with name changes. In any case, we were only able to establish conclusive correspondences for a minority of the hemibrain Li11 to Li20 types, which are detailed in Supplementary Data 1. Hemibrain Li12 is now Li27 (jigsaw pair), and hemibrain Li16 is now Li28 (pair of full-field cells). Hemibrain Li11 was split into Li25 and Li19 (see the ‘Morphological variation’ section). Hemibrain Li18 was split into three types: (1) Li08 covers the whole visual field. (2) Li04 covers a dorsal region except for the dorsal rim. It is tangentially polarized, with the axon more dorsal than the dendrites. Both axon and dendrite point in the posterior direction, perpendicular to the direction of polarization. The dendrites are more thickly stratified than the axon. (3) Li07 has ventral coverage only. The axons are in one layer, and extend over a larger area than the dendrites, which hook around into another layer and are mostly near the ventral rim. We considered merging Li04 and Li07, but their connectivity is quite different. Furthermore, in a hierarchical agglomerative clustering, Li07 would merge with Li08 before Li04.

LPi

LPi names were originally based on stratification in layers 1 to 4 of the lobula plate, including LPi1-2 and 2-1¹⁰; LPi3-4 and 4-3⁸; and LPi2b and LPi34-12¹⁰ (we are not counting fragments for which correspondences are not easy to establish). We have added nine new types, for a total of 15 LPi types.

Now that LPi types have multiplied, stratification is no longer sufficient for naming. The naming system could be salvaged by adding letters to distinguish between cells of different sizes. For example, LPi15 and LPi05 could be called LPi2-1f and LPi2-1s, where ‘f’ means full-field and ‘s’ means small. For simplicity and brevity, we instead chose the names LPi01 to LPi15, in order of increasing average cell volume. Correspondences with old stratification-based names are detailed in Codex.

Cross-neuropil tangential and amacrine

Most types that span multiple neuropils are columnar. One tangential type that spans multiple neuropils inside the optic lobe was previously described: Lat has a tangential axon that projects from the medulla to the lamina⁶. There is some heterogeneity in the Lat population, as reflected in the large type radius (Extended Data Fig. 3a). We have decided to leave splitting for future work, as Lat has many dense core vesicles that are presently unannotated.

Here we introduce two new families of cross-neuropil types that are tangential (MLt1-8 and LMt1-4), and one that is amacrine (LMa1-5). Along with two new tangential families (PDt, LLPt) that contain only single types, and the known CT1 and Am1 types, that is a total of 21 cross-neuropil types that are non-columnar (Fig. 1c). Each of the new types (except PDt with 6 cells) contains between 10 and 100 cells.

The tangential types connect neuropils within one optic lobe and do not leave the optic lobe. Our usage of the term ‘tangential’ focuses on axonal orientation only. It should not be misunderstood to imply a wide-field neuron that projects out of the optic lobe, which is the case for the well-known lobula plate tangential cells or lobula tangential cells. The term ‘tangential’ presupposes that we can identify an axonal arbour for the cell (see the ‘Axon versus dendrite’ section).

PDt

We found one tangential type that projects from proximal to distal medulla (Supplementary Data 2).

MLt

ML1 was previously identified⁴² as a tangential neuron projecting from the medulla to lobula. We will refer to this type as MLt1, and have discovered more types of the same family, MLt2 to MLt8. Mlt1 and Mlt2 dendrites span both distal and proximal medulla, and Mlt3 dendrites are in the distal medulla, so MLt1 to MLt3 receive L input (Supplementary Data 2 and 5). Mlt4 dendrites are in the proximal medulla (Supplementary Data 2). Mlt5 to Mlt8 have substantial arbour overlap with the serpentine layer M7 (Supplementary Data 2), and are therefore connected with many Sm types to be discussed later on (Supplementary Data 5). Interaction between MLt types is fairly weak, with the exception of MLt7 to MLt5 (Supplementary Data 5). MLt7 and MLt8 are restricted to the dorsal and dorsal rim areas.

LMt

We identified four tangential types (LMt1 to LMt4) that project from the lobula to medulla. Their axonal arbours are all in the proximal medulla (Supplementary Data 2), thinly stratified near layer M7, so they have many Pm targets (Supplementary Data 5). Only LMt4 exhibits partial coverage.

LLPt

We discovered one tangential type that projected from the lobula to lobula plate, and called it LLPt. This is just a single type, rather than a family.

LMa

We discovered four amacrine types that extend over the lobula and medulla. LMa1 to LMa4 are coupled with T2, T2a and T3, and LMa4 and LMa3 synapse onto T4 and T5 (Supplementary Data 5). The LMa family could be said to include CT1, a known amacrine cell that also extends over both the lobula and medulla. However, the new LMa types consist of smaller cells that each cover a fraction of the visual field, whereas CT1 is a wide-field cell.

MLLPa

Am1 was defined¹⁰ as a wide-field amacrine cell that extends over the medulla, lobula and lobula plate. We found no other amacrine types like Am1 with such an extended reach.

Correspondences with molecular–morphological types

Tm5

Tm5a, Tm5b and Tm5c were originally defined by single-cell anatomy and Ort expression^7,50. Tm5a is cholinergic, the majority of the cells extend one dendrite from M6 to M3, and often has a ‘hook’ at the end of its lobula axon. Tm5b is cholinergic, and most (~80%) cells extend several dendrites from M6 to M3. Tm5c is glutamatergic and extends its dendrites up to the surface of the distal medulla. Three of our types are consistent with these morphological descriptions (Fig. 7a), and receive direct input from inner photoreceptors R7 or R8.

Dm8

Molecular studies previously divided Dm8 cells into two types (yDm8 and pDm8), depending on whether or not they express DIPγ^51,53. Physiological studies demonstrated that yDm8 and pDm8 have differing spectral sensitivities⁸⁹. The main dendrites of yDm8 and pDm8 were found to connect with R7 in yellow and pale columns, respectively. On the basis of its strong coupling with Tm5a, our Dm8a probably has some correspondence with yDm8, which is likewise selectively connected with Tm5a^51,53. It is not yet clear whether there is a true one-to-one correspondence of yDm8 and pDm8 with Dm8a and Dm8b. It is the case that Dm8a and Dm8b strongly prefer to synapse onto Tm5a and Tm5b, respectively. However, Tm5a and Tm5b are not in one-to-one correspondence with yellow and pale columns. Rather, the main dendritic branch of Tm5a is specific to yellow columns, while the main dendritic branches of Tm5b are found in both yellow and pale columns⁵⁰. Furthermore, Dm8a and Dm8b cells are roughly equal in number, while the yDm8:pDm8 ratio is expected to be substantially greater than one^51,53, like the ratio of yellow to pale columns. Thus, the correspondence of Dm8a and Dm8b with yDm8 and pDm8 is still speculative. The yellow/pale issue should be revisited in the future when accurate photoreceptor synapses become available (see the ‘Reconstruction accuracy and completeness’ section).

Additional validation

HHMI Janelia has released a preprint detailing cell types in the right optic lobe of an adult male Drosophila brain⁹⁰. The list of intrinsic cell types is almost identical to ours, apart from naming differences in new types. Since our original submission, we have completed typing of the left optic lobe of our female fly brain reconstruction, and the results match the right optic lobe analysed in the present paper. These replications in another hemisphere of the same brain and in the brain of another individual fly provide additional validation of our findings.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.