Understanding the intricate workings of the brain is one of the most significant challenges in modern science. A crucial step towards this understanding lies in mapping the neuronal wiring diagram, the complex network of connections that dictate brain function. Recent advancements in connectomics, the field dedicated to mapping these neural connections, have provided unprecedented insights into brain structure and function. This article delves into the reconstruction of the Drosophila brain’s neuronal wiring diagram, focusing on the optic lobe and the methodologies employed to achieve remarkable accuracy and completeness in this complex endeavor.
The Drosophila brain, despite its smaller scale compared to mammalian brains, presents a significant challenge for reconstruction due to its dense neuropil and the sheer number of neurons and synapses. Evaluations of the overall reconstruction quality of the Drosophila brain have been detailed previously, highlighting the current progress in this field. Here, we focus on specific validation checks conducted within the optic lobe, a region crucial for visual processing in Drosophila. While the vast majority of cells have been meticulously proofread, a small percentage still require refinement. In certain instances, particularly in the mid-posterior region of the right optic lobe, some cell types exhibit minor “bald spots,” indicating areas where neuronal tracks become narrow or discontinuous. These tracks sometimes appear to terminate within glial cells, raising the possibility of neuronal engulfment by glia. However, it’s important to note that for most cell types, under-recovery is minimal and often imperceptible.
Alt text: Electron microscopy image of Drosophila optic lobe, highlighting neuronal tracks and potential areas of under-recovery in wiring diagram reconstruction.
To quantitatively assess the extent of under-recovery in our neuronal wiring diagram, we can utilize ‘modular’ cell types. These cell types exhibit a one-to-one correspondence with columns, structural units within the optic lobe. Previous research reconstructing seven medulla columns identified 20 modular types. These types largely correspond to cells counted in our reconstruction, ranging from 720 to 800 cells per type. The upper end of this range, approximately 800 cells, likely represents the true number of columns in this optic lobe. The lower end, around 720 cells, suggests that under-recovery is at most 10%, and typically even less. This quantitative analysis reinforces the high degree of completeness achieved in our Drosophila brain wiring diagram reconstruction.
Photoreceptor neurons, responsible for light detection, present unique challenges in wiring diagram reconstruction. Inner photoreceptors R7 and R8 each number around 650 cells, while outer photoreceptors R1–6 total approximately 3,400 in version 783 of the FlyWire connectome, a publicly accessible resource for connectomic data. These numbers are consistent with modularity, although photoreceptors are more difficult to proofread in this dataset, leading to higher under-recovery rates compared to other neuron types. In the left optic lobe, we have proofread approximately 38,500 intrinsic neurons, along with 3,700 VPNs, 250 VCNs, 150 heterolateral neurons, and 5,000 photoreceptor cells. Detailed tables comparing cell counts between the left and right hemispheres, categorized by superclass and type, are available for download, providing further validation of the wiring diagram’s accuracy across brain hemispheres.
Cell counts alone are insufficient for a comprehensive modularity analysis. Cell types such as Tm21, Dm2, TmY5a, Tm27, and Mi15 are considerably less numerous than 800, supporting prior findings that they are not modular. Conversely, some of our identified types, including T2a, Tm3, T4c, and T3, exceed 800 proofread cells, challenging the strict definition of modularity based solely on cell counts. This aligns partially with previous reconstructions that considered T3 and T2a modular, but T4 and Tm3 as non-modular. The T4 type presents an interesting case, with T4c exceeding 800 cells while other T4 subtypes fall below this threshold. It is important to acknowledge that these cell counts are subject to minor upward revisions with ongoing proofreading efforts, highlighting the iterative nature of wiring diagram refinement.
A complete assessment of modularity necessitates examining cell locations and verifying the one-to-one correspondence with columns, an analysis reserved for future investigations. In this study, we use the term ‘numerous’ to categorize types containing 720 or more cells, as well as photoreceptor types, without definitively classifying them as modular. Prior work provided a connection matrix for modular types, and our data shows strong agreement, offering an independent validation of our optic lobe wiring diagram reconstruction. This validation complements accuracy assessments performed in the central brain, reinforcing the reliability of the complete brain-wide neuronal wiring diagram.
A primary limitation in our optic lobe wiring diagram reconstruction pertains to automatically detected synapses. While overall synapse detection accuracy is high, outgoing photoreceptor synapses are notably underdetected. This may be due to the dark cytoplasm characteristic of photoreceptors, which might not be adequately represented in the synapse example images used to train the automated synapse detector. Ongoing efforts are addressing this issue by incorporating photoreceptor synapse examples into the training set of an improved detector. However, these improved results were not available for this publication. Consequently, our analysis of connectivity from photoreceptors to other cell types is currently qualitative rather than fully quantitative. Furthermore, underdetection of photoreceptor synapses could introduce biases in input fraction calculations for other connections due to normalization effects.
Another important consideration is the potential for artifactual weak connections in the type–type connectivity matrix. These could arise from false positives in automated synapse detection. Heuristics can help identify potentially artifactual connections without manual EM image inspection. For instance, weak connections, defined as those below a certain synapse number threshold, might be suspect. The appropriate threshold value depends on the context. Prior work in the central brain discarded connections with fewer than five synapses. However, for the optic lobe, we applied a lower threshold of two synapses. This difference arises because of the distinct contexts of the central brain and optic lobe.
Alt text: Diagram comparing cell cardinality in central brain versus optic lobe, highlighting differences in neuron type repetition.
In the central brain, most cell types have a cardinality of two, representing a cell and its mirror twin in the opposite hemisphere. In the hemibrain, this cardinality typically reduces to one. Therefore, determining connections between cell types in the central brain relies on limited examples, necessitating a higher synapse threshold to minimize false positives. Conversely, the optic lobe often contains numerous examples of ordered cell type pairs due to high cell cardinality. Consistent connections between types in the optic lobe can be considered reliable even with a lower average synapse count. This justification led to our choice of a lower synapse threshold in the optic lobe wiring diagram analysis. Notably, we observed that certain inhibitory cell types consistently form connections with relatively few synapses, yet these connections appear to be genuine and functionally relevant.
Another heuristic involves examining asymmetry in the connection matrix. Significant differences in synapse counts between connections from type A to B versus B to A might indicate spurious connections. A strong A-to-B connection implies a large contact area, increasing the chance of false-positive B-to-A synapses. Finally, existing knowledge about the absence of certain connections can guide artifact identification. For example, T1 cells are known to lack output synapses. Therefore, any outgoing T1 synapses detected in our data were considered false positives and excluded from the wiring diagram.
Morphological Cell Typing in Wiring Diagram Construction
Our approach to cell typing within the connectomic wiring diagram framework begins with an initial set of types to define feature vectors for cells. This initial seeding relies on morphological cell typing, a time-honored method, often augmented by computational tools that analyze connectivity. While ‘morphology’ strictly refers to shape, orientation and position are arguably more fundamental properties influencing neuropil layer stratification. Thus, ‘single-cell anatomy’ might be a more accurate term, although morphology remains the conventional term in the field.
Stage 1: Crowdsourced Annotation of Known Cell Types
Initial annotation of optic lobe neurons was crowdsourced, leveraging the expertise of volunteers from Drosophila laboratories and later expanding to include citizen scientists. This phase primarily focused on labeling cells of known types, especially the most abundant ones.
Drosophila Lab Annotators
Experts like E.K. and D.G. proofread and annotated medulla neurons upstream of the anterior visual pathway, encompassing many medulla and lamina neurons crucial to this study. Annotated neurons primarily included Dm2, Mi15, R7, and R8, as well as various L, Dm, Mi, Tm, C, and Sm cells. Known neuron types were identified based on morphology and, to some extent, connectivity patterns. Annotators also meticulously identified all Mi1 neurons in both hemispheres to map every medulla column. These Mi1 neurons were instrumental in creating a medulla layer map based on Mi1 stratification, which subsequently aided citizen scientists in identifying medulla cell types.
Citizen Scientists
Top-performing players from Eyewire, a citizen science platform for connectomics, were invited to contribute to proofreading in FlyWire. After three months of proofreading the right optic lobe, these citizen scientists were encouraged to label neurons when confident in their identification. Most citizen scientists engaged in both annotation and proofreading, sometimes labeling cells post-proofreading and at other times searching for specific cell types to proofread.
Citizen scientists were provided with visual guides to optic lobe cells drawn from existing literature. FlyWire incorporated a 3D mesh overlay delineating the four primary optic lobe neuropils. Visual identification heavily relied on single-cell anatomy. Initially, labeling type families (Dm, Tm, Mi, etc.) was encouraged, particularly for novice annotators. Annotation of specific types (Dm3, Tm2, etc.) evolved over time. Canonical naming conventions were reinforced by software tools that facilitated easy selection and submission of preformatted type names, ensuring consistency in the growing wiring diagram database.
Alt text: FlyWire interface screenshot showcasing tools for 3D neuron visualization and annotation in Drosophila brain wiring diagram project.
Additional community resources, including discussion forums, blogs, shared drives, chat platforms, dedicated email, and Twitch livestreams, fostered a collaborative environment for knowledge exchange among citizen scientists, community managers, and researchers. Community managers addressed queries, provided resources such as visual guides, shared project updates, offered troubleshooting support, and managed overall community activities. Daily annotation statistics were shared to track project progress. Weekly Twitch livestreams facilitated live interaction, demonstrations, and communal problem-solving, led by community managers. This rich ecosystem enabled citizen scientists to self-organize, leading to community-driven information sharing, programmatic tool development, and the creation of ‘cell farms.’
Community-Driven Information Sharing
Citizen scientists collaboratively developed comprehensive guides with text and screenshots that expanded upon the initial visual guides. They actively sought and studied publicly available scientific literature and resources related to the optic lobe. Findings were shared on discuss.flywire.ai, which accumulated over 2,500 posts by October 10, 2023, demonstrating the vibrant community engagement in building the Drosophila brain wiring diagram. Community managers further facilitated information flow by sharing insights from scientific literature, consulting Drosophila specialists on FlyWire, and providing feedback to citizen scientists.
Programmatic Tools
Programmatic tools emerged to aid in searching for cells of the same type, streamlining the wiring diagram reconstruction process. One key script traced ‘partners-of-partners,’ identifying cells sharing common synaptic partners. This was based on the assumption that cells of the same type would likely synapse with similar target cells, a principle that often held true. The tool could search for partners-of-all-partners or partners-of-any-partners. Resulting cell lists, often extensive, were filtered to exclude already identified cells or segments with small sizes or low IDs, which were likely unproofread. Another tool, developed using lobula plate tangential cells (HS, VS, H1), assisted in defining layers within the lobula plate, facilitating identification of cell types, particularly T4 and T5, crucial for motion detection circuits within the wiring diagram.
Cell Farms
Citizen scientists created ‘farms’ within FlyWire or Neuroglancer, visualizing all identified cells of a given type collectively. These farms visually indicated areas where cells of a specific type remained to be found. In ‘bald spot’ regions, a common strategy to locate missing cells involved systematically moving the 2D plane and adding segments to the farm in search of matching cell types. Farms also proved valuable in identifying cells near neuropil edges, where neuron morphology could be distorted. Having a comprehensive view of all cells of the same type enabled extrapolation of the expected morphology of cells located at neuropil boundaries, enhancing accuracy in wiring diagram reconstruction.
Stage 2: Centralized Annotation and Discovery of New Cell Types
A dedicated team of image analysts at Princeton finalized annotation of remaining cells within known types and embarked on discovering novel cell types, expanding the known diversity of the Drosophila optic lobe wiring diagram. Initial community annotations were rigorously compared against existing literature to ensure accuracy. Once validated, these cells served as queries for Codex search tools, which identified previously unannotated cells exhibiting similar connectivity profiles. Search query results were evaluated based on morphology and stratification to confirm matches with the target cell type. In cases where cell type distinctions remained ambiguous, predicted neurotransmitter profiles were used as supplementary guidance. This iterative process enabled the creation of a preliminary clustering of both previously known and newly discovered cell types, providing a more complete picture of the neuronal wiring diagram.
Connectomic Cell Typing: Moving Beyond Morphology in Wiring Diagrams
Morphological cell typing eventually reached its limitations, hindering further progress in refining the Drosophila brain wiring diagram. Expert annotators encountered difficulties classifying Tm5 cells into the three known subtypes, unaware of the eventual discovery of six distinct Tm5 types. This bottleneck necessitated a transition to connectomic cell typing, leveraging connectivity patterns for cell type classification. In retrospect, this transition could have been implemented earlier. Connectomic cell typing requires an initial set of seed types, but this initial seeding did not need to be as exhaustive as it ultimately became. Future research will explore extending the connectomic approach to be applicable from the outset of wiring diagram reconstruction.
Stage 3: Connectivity-Based Refinement: Splitting and Merging Types and Auto-Correction
Computational methods were employed to refine cell type classifications by splitting types that proved challenging to separate in Stage 2. Candidates for splitting, such as Tm5, were suggested by image analysts. Other candidates were identified based on unusually high cell counts or large type radii. Hierarchical clustering with average linkage was applied, and splits were accepted if they did not violate the tiling principle related to spatial coverage.
Conversely, computational methods also addressed improper type splitting from Stage 2 by merging types. Candidates for merging included types with limited spatial coverage or those exhibiting close proximity in cell type dendrograms. Merge decisions were guided by hierarchical clustering of cells from candidate types and validated if merging improved spatial coverage, ensuring a more accurate and biologically relevant wiring diagram classification.
Upon reaching a final list of cell types, the ‘center’ of each type was estimated using the element-wise trimmed mean. Then, for every cell, the nearest type center was computed using Jaccard distance. For 98% of cells, the nearest type center aligned with the assigned type. Discrepancies were manually reviewed. In most cases, the algorithm was correct, and human annotators had made errors, often due to inattention. Remaining discrepancies primarily stemmed from proofreading errors. In some instances, type centers were contaminated by misassigned cells, leading to further misassignments by the algorithm. After addressing these issues, automatic corrections were applied to all but 0.1% of cells, which were rejected based on distance thresholds, ensuring a high degree of accuracy in the final wiring diagram and cell type classification.
Validation of Cell Typing Accuracy in Wiring Diagram
Based on the auto-correction procedure, we estimate cell type assignment accuracy to be between 98% and 99.9%. Another measure of cell typing quality is the ‘radius’ of each type, representing the average distance from its cells to its center. The center is computed by approximately minimizing the sum of Jaccard distances from each cell in the type to the center. A large type radius can indicate that a type encompasses dissimilar cells and should be further split. For our final types, radii vary but almost all remain below 0.6. Lat, however, exhibits an exceptionally high type radius, suggesting it warrants splitting in future work. Type radii remain consistent whether boundary types are included in the feature vector or not.
Discrimination with Logical Predicates in Wiring Diagrams
Given the high dimensionality of feature vectors, simpler insights into type-defining characteristics are valuable. One approach involves identifying sets of simple logical predicates based on connectivity that accurately predict type membership. For a given cell, we define attributes like ‘is connected to input type t‘ (receiving at least one connection from type t) and ‘is connected to output type t‘ (making at least one connection to type t).
For each type, an optimal predicate is constructed, consisting of input and output type tuples (maximum size 5), optimized for the F-score of type prediction. F-score is the harmonic mean of precision and recall. Recall measures a predicate’s ability to identify all positive instances of a type, while precision measures the proportion of true positive predictions among all positive predictions.
The predicate computation process is exhaustive, exploring all combinations of input and output type tuples and calculating precision, recall, and F-score for each type. Optimization techniques are used to accelerate computation. For example, the 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 almost all identified types, logical predicates with five or fewer input/output attributes predict type membership with an average F-score of 0.93. Predicate attributes are distinctive partners, not necessarily the most connected ones. Predicates for each type and family are detailed in supplementary materials.
Experiments involving random shuffling of a small fraction of types demonstrated a substantial drop in predicate quality, indicating that we are not overfitting the data, as expected due to the brevity of the predicates. Excluding boundary types has a minimal impact on predicate quality, similar to clustering metrics.
Discrimination with Two-Dimensional Projections of Wiring Diagrams
Another approach to interpretability involves examining low-dimensional projections of the high-dimensional feature vector. For each cell type, a small subset of dimensions sufficient for accurate discrimination is selected. Feature vectors are normalized to represent ‘fraction of input synapses from type t‘ or ‘fraction of output synapses to type t.’
For instance, Pm family cells can be visualized in a 2D space defined by C3 input fraction and TmY3 output fraction. Pm04 cells are well-separated in this space and can be discriminated with 100% accuracy using the rule ‘C3 input fraction greater than 0.01 and TmY3 output fraction greater than 0.01.’ This two-feature discriminator is more accurate than either feature alone.
Generally, cell type discriminators are based on thresholding input and output fractions and taking the conjunction of results. The search for a discriminator identifies relevant dimensions and threshold values. To simplify the search, discrimination is performed only against other types within the same neuropil family. Under these conditions, usually just two dimensions of the normalized feature vector suffice. Discriminators for all types in multi-type families are provided in supplementary materials. Many, though not all, discriminations are highly accurate. Both intrinsic and boundary types are used as discriminative features in these wiring diagram analyses.
Computational Concepts Underlying Wiring Diagram Analysis
Connectivity Measures: Cell-to-Cell, Type-to-Cell, Cell-to-Type, and Type-to-Type Wiring Diagrams
We define a weighted cell-to-cell connectivity matrix *w*ij, representing the number of synapses from neuron i to neuron j. Weighted out-degree and in-degree of neuron i are:
$$begin{array}{cc}{d}_{i}^{+}=sum _{j}{w}_{{ij}} & {d}_{i}^{-}=sum _{j}{w}_{{ji}}end{array}$$
Where sums are over all brain neurons. For optic lobe intrinsic cells, sums are restricted to intrinsic and boundary neurons of that optic lobe.
Let *A*it be a 0–1 matrix assigning neuron i to type t. Column and row sums satisfy:
$$begin{array}{cc}{n}_{t}=sum _{i}{A}_{{it}} & 1=sum _{t}{A}_{{it}}end{array}$$
Where *n*t is the number of cells of type t.
The cell-to-type connectivity matrix *O*it is output synapses from neuron i to type t neurons:
$${O}_{{it}}=sum _{j}{w}_{{ij}}{A}_{{jt}}$$
For fixed i, *Oit is the output feature vector. Similarly, the type-to-cell connectivity matrix Itj is input synapses from type t neurons to neuron j*:
$${I}_{{tj}}=sum _{j}{A}_{{it}}{w}_{{ij}}$$
For fixed j, *I*tj is the input feature vector. Concatenating the ith row and column of these matrices forms the full feature vector for cell i. Input and output feature vectors can be normalized by degree to yield input and output fractions.
The type-to-type connectivity matrix *W*st is synapses from type s neurons to type t neurons, forming a high-level wiring diagram:
$${W}_{{st}}=sum _{{ij}}{A}_{{is}}{w}_{{ij}}{A}_{{jt}}$$
Weighted degree of type t is the sum of weighted degrees of cells in type t:
$$begin{array}{cc}{D}_{t}^{+}=sum _{i}{A}_{{it}}{d}_{i}^{+} & {D}_{t}^{-}=sum _{i}{A}_{{it}}{d}_{i}^{-}end{array}$$
Normalizing by degree yields output fractions of type s, *Wst/Ds+, and input fractions of type t, Wst/Dt−*. Feature vectors can also be based on connection number (>=2 synapses) instead of synapse number, yielding similar results and suppressing noise. Feature dimensions can include only intrinsic types or both intrinsic and boundary types, with similar outcomes. For hierarchical clustering, the feature vector for each cell type is formed by concatenating input and output fractions.
Similarity and Distance Measures in Wiring Diagram Analysis
Weighted Jaccard similarity between feature vectors x and y is:
$$Jleft({bf{x}},{bf{y}}right)=frac{{sum }_{t}min left({x}_{t},{y}_{t}right)}{{sum }_{{t}^{{prime} }}max left({x}_{{t}^{{prime} }},{y}_{{t}^{{prime} }}right)}$$
Weighted Jaccard distance d(x,y) is 1 – J(x,y). Jaccard similarity empirically performs better than cosine similarity for sparse feature vectors in cell typing.
Type Centers in Wiring Diagram Space
Given feature vectors x**a, the center c** minimizes:
$$sum _{a}dleft({{bf{x}}}^{a},{bf{c}}right)$$
This convex cost function has a unique minimum, approximated using various methods. For auto-correction, the element-wise trimmed mean is used. For type radii, a coordinate descent approach minimizes the cost function.
Hierarchical Clustering of Cell Types in Wiring Diagrams
The type-to-type connectivity matrix is the basis for cell type clustering. Rows and columns are normalized to input and output fractions and concatenated to form feature vectors. Average linkage hierarchical clustering yields a dendrogram. Dendrogram thresholding produces flat clusters. Cluster memberships should be interpreted cautiously as they depend on the clustering algorithm. Clusters contain core groups of highly similar types that merge early in agglomeration. Later merges are less certain. Cluster mixing across neuropil families can be attributed to both ‘noise’ and genuine biological interactions between subsystems in the Drosophila brain wiring diagram.
Wiring Diagrams: Visualizing Neural Connections
Reduction for Wiring Diagram Readability
To enhance wiring diagram readability, only top type-to-type connections are displayed. For each cell type, top input and output cell types are selected by ranking partners by total synapse count. Ties are resolved by including runner-ups within 5% of the winner. Figure 3 shows top connections between all optic lobe intrinsic types. Subsequent figures focus on subsystems, including top input/output connections with the rest of the network and boundary types. Extended Data Figures 5 and 6 show top input and output connections separately for improved clarity in visualizing the wiring diagram.
Colors, Shapes, and Layout in Wiring Diagrams
Nodes (cell types) are colored by clusters. Node size reflects connection count, highlighting highly connected types. Node shapes encode type numerosities (hexagon to ellipse for most to least numerous). Lines indicate connections; color encodes input/output direction, and width represents synapse number. Arrowheads indicate neurotransmitter predictions (excitatory/inhibitory). Cytoscape software was used for wiring diagram layout, employing organic or hierarchical layouts with manual adjustments to minimize obstructions and improve the visual clarity of the wiring diagram.
Intrinsic Versus Boundary Neurons in the Optic Lobe Wiring Diagram
The optic lobes are divided into five neuropils. Non-photoreceptor cells with synapses in these regions are categorized as optic lobe intrinsic neurons or boundary neurons. Optic lobe intrinsic neurons are almost entirely confined to one optic lobe (>=95% of synapses within optic lobe regions). Boundary neurons have 5–95% of synapses in optic lobe regions and are visual projection, centrifugal, or heterolateral neurons, connecting the optic lobe wiring diagram to the rest of the brain.
Axon Versus Dendrite: Defining Neuron Polarity in Wiring Diagrams
An axon is defined as a neuronal portion with a high presynapse to postsynapse ratio, either in absolute terms or relative to other neuronal regions (dendrites). Axons are typically not purely output elements, possessing some postsynapses. Axon identification can often be visually aided by the presence of varicosities (presynaptic boutons). Dendrites have a high postsynapse to presynapse ratio. Amacrine cells lack a clear axon-dendrite distinction, with intermingled presynapses and postsynapses throughout their neurites.
Columnar Neurons: Building Blocks of the Optic Lobe Wiring Diagram
Columnar families, based on neuropils, include L, C, T1, T2, T3, T4, and T5 families. These families consist of ‘numerous’ types (~800 cells). Lawf1 and Lawf2 types are grouped with T1 despite neuropil and connectivity differences. T1 is separated due to lacking output synapses, despite sharing neuropils with Lawf1. Distal and proximal medulla are considered separate neuropils in the wiring diagram structure.
Mi Interneurons in the Wiring Diagram
Mi interneurons project from distal to proximal medulla, encompassing both numerous and less numerous types. We identified Mi1, 2, 4, 9, 10, 13, 14, and 15. Mi1, Mi4, and Mi9 align with the classical definition, while Mi13 projects in the opposite direction. Other Mi types are less polarized, potentially better classified as “narrow-field amacrine” rather than strictly columnar in the wiring diagram context.
Tm Transmedullary Neurons in the Wiring Diagram
Tm cells project from distal medulla to lobula. We identified Tm1, 2, 3, 4, 7, 9, 16, 20, 21, 25, and 27. Tm5 was split into six types, Tm8 into two, and Tm6 and Tm21 were merged into Tm21. Tm1a, Tm4a, and Tm27Y were merged into Tm1, Tm4, and Tm27, respectively. TmY5 was merged into TmY5a. New types Tm31–Tm37, projecting from serpentine medulla to lobula, were added. Tm23 and Tm24 were reclassified into the Li family due to their synapse locations, despite soma location and neurite path. Approximately half of the ~26 Tm family types are newly identified in this detailed wiring diagram.
TmY Neurons in the Wiring Diagram
TmY cells project from distal medulla to both lobula and lobula plate. We identified TmY3, TmY4, TmY5a, TmY10, TmY11, TmY14, TmY15, TmY16, and TmY20. TmY9 was divided into two types. A new type, TmY31, was added.
Y Neurons in the Wiring Diagram
Y cells project from proximal medulla to lobula and lobula plate, similar to TmY but traversing only the proximal medulla. We identified Y1, Y3, Y4, Y11, and Y12. Y1, Y11, and Y12 are motion subsystem neurons with lobula plate synapses. Y3 and Y4, with fewer lobula plate synapses, are object subsystem neurons. Y3 is more numerous and predicted cholinergic.
Tlp Neurons in the Wiring Diagram
Tlp neurons project from lobula plate to lobula. We identified Tlp1, Tlp4, Tlp5, and Tlp14. Tlp11, Tlp12, and Tlp13 names are proposed to be retired, as they correspond to Tlp5, Tlp1, and Tlp4, respectively, simplifying the Tlp nomenclature in the wiring diagram.
Interneurons: Local Circuitry in the Wiring Diagram
Local interneurons are confined to a single neuropil. They constitute the majority of types but a minority of cells. Lai is the only lamina interneuron. Dm and Pm interneurons stratify in distal and proximal medulla, respectively. We significantly increased Pm type count and slightly increased Dm types. The Sm family, almost entirely new, is the largest family. Li and LPi interneurons stratify in lobula and lobula plate. Interneurons are typically amacrine, presumed inhibitory, but some are tangential or cholinergic, and vary in field width.
Dm Interneurons in the Wiring Diagram
Dm1–Dm8, Dm9–10, and Dm11–20 were previously defined. We do not observe Dm5 and Dm7. Most types are predicted glutamatergic or GABAergic, with a few cholinergic types. Dm3p, Dm3q, and a new type, Dm3v, are identified. Dm8 was split into Dm8a and Dm8b.
DmDRA Interneurons in the Wiring Diagram
DRA photoreceptors target different medulla layers and output cell types compared to non-DRA photoreceptors. DRA-R7 connects to DmDRA1, and DRA-R8 to DmDRA2. DmDRA1 and DmDRA2 exhibit arched coverage in M6 layer of dorsal medulla. R7-DRA and R8-DRA annotation is incomplete and will be improved in future releases. DmDRA1 receives R7 input but is located in M7. It could be classified as Sm, but we retain DmDRA for historical reasons in the wiring diagram nomenclature.
Pm Interneurons in the Wiring Diagram
Pm1, Pm1a, and Pm2 were each split into two types. Pm3 and Pm4 remain unchanged. Six new Pm types were identified, totaling 14 Pm types (Pm01–Pm14). All are predicted GABAergic. Pm1 was split into Pm06 and Pm04, Pm1a into Pm02 and Pm01, and Pm2 into Pm03 and Pm08.
Sm Interneurons in the Wiring Diagram
Sm interneurons, a large new family, stratify in the serpentine layer (M7) of the medulla, bridging Dm and Pm types. Named Sm01–Sm43, they include medulla tangential intrinsic types. Sm stratification in M7 facilitates communication with Mt cells and inner photoreceptor terminals, implicating Sm types in chromatic stimuli processing and color subsystem function in the wiring diagram. The Sm family significantly expands medulla interneuron diversity and may relate to the M6-LN class, although Sm primarily stratifies in M7, not M6.
Li Interneurons in the Wiring Diagram
After initial Li1 and Li2 types, 12 more (Li11–20, mALC1, mALC2) were identified. We confirmed Li2, Li12, Li16, mALC1, and mALC2 and identified 21 new Li types, totaling 33 (Li01–Li33). Tm23 and Tm24 were reclassified into the Li family. Hemibrain Li12, Li16, and Li11 were reclassified as Li27, Li28, and split into Li25 and Li19, respectively. Hemibrain Li18 was split into Li08, Li04, and Li07, highlighting refined classification within the wiring diagram.
LPi Interneurons in the Wiring Diagram
LPi names were initially based on lobula plate layer stratification. We added nine new types, totaling 15 LPi types (LPi01–LPi15). Stratification-based naming is no longer sufficient. New names (LPi01–LPi15) are based on increasing cell volume. Stratification-based name correspondences are detailed in Codex, reflecting the evolving complexity of the lobula plate circuitry within the wiring diagram.
Cross-Neuropil Tangential and Amacrine Neurons in the Wiring Diagram
Most multi-neuropil types are columnar. Lat is a known tangential type spanning medulla to lamina. We introduce new tangential (MLt1–8, LMt1–4) and amacrine (LMa1–5) cross-neuropil type families. With PDt, CT1, and Am1, there are 21 non-columnar cross-neuropil types. New types (except PDt) contain 10–100 cells. Tangential types connect neuropils within one optic lobe, focusing on axonal orientation, not wide-field projection out of the optic lobe.
PDt Tangential Neurons in the Wiring Diagram
We identified one tangential type projecting from proximal to distal medulla.
MLt Tangential Neurons in the Wiring Diagram
MLt1, projecting from medulla to lobula, was previously identified. We discovered MLt2–MLt8. MLt1–MLt3 dendrites span distal and proximal medulla (L input). MLt4 dendrites are in proximal medulla. MLt5–MLt8 arbors overlap with serpentine layer M7 (Sm connections). MLt7 and MLt8 are dorsal/dorsal rim restricted.
LMt Tangential Neurons in the Wiring Diagram
We identified LMt1–LMt4 projecting from lobula to medulla. Axonal arbors are in proximal medulla near M7 (Pm targets). Only LMt4 exhibits partial coverage within the wiring diagram.
LLPt Tangential Neurons in the Wiring Diagram
We discovered one tangential type, LLPt, projecting from lobula to lobula plate.
LMa Amacrine Neurons in the Wiring Diagram
We discovered LMa1–LMa4 amacrine types spanning lobula and medulla. LMa1–LMa4 are coupled with T2, T2a, and T3, and LMa4 and LMa3 synapse onto T4 and T5. LMa family may include CT1, but new LMa types are smaller, covering fractions of the visual field, unlike wide-field CT1 in the wiring diagram.
MLLPa Amacrine Neurons in the Wiring Diagram
Am1 is a wide-field amacrine cell spanning medulla, lobula, and lobula plate. No other amacrine types with such extended reach were found in this Drosophila brain wiring diagram reconstruction.
Correspondences with Molecular–Morphological Types: Validating Wiring Diagram Accuracy
Tm5 Subtypes in the Wiring Diagram
Tm5a, Tm5b, and Tm5c were previously defined by anatomy and Ort expression. Tm5a is cholinergic, with dendrites from M6 to M3, and often a lobula axon ‘hook.’ Tm5b is cholinergic, with multiple dendrites from M6 to M3. Tm5c is glutamatergic, with dendrites to distal medulla surface. Three of our types align with these descriptions and receive direct R7/R8 input, validating molecular-morphological correspondences within the wiring diagram.
Dm8 Subtypes in the Wiring Diagram
Molecular studies divided Dm8 into yDm8 and pDm8 (DIPγ expression). yDm8 and pDm8 exhibit different spectral sensitivities. yDm8 and pDm8 dendrites connect to R7 in yellow and pale columns, respectively. Our Dm8a, strongly coupled with Tm5a, may correspond to yDm8. Dm8a and Dm8b synapse onto Tm5a and Tm5b, respectively. However, Tm5a/Tm5b and yellow/pale columns are not in one-to-one correspondence. Dm8a/Dm8b and yDm8/pDm8 correspondence remains speculative and requires future investigation with improved photoreceptor synapse data, further refining the wiring diagram’s accuracy.
Additional Wiring Diagram Validation
HHMI Janelia released a preprint detailing cell types in the right optic lobe of a male Drosophila brain. Their intrinsic cell type list is almost identical to ours, with minor naming differences for new types. Our left optic lobe typing results match the right optic lobe data. These replications across hemispheres and individuals further validate our findings and the robustness of the Drosophila brain wiring diagram reconstruction.
Reporting Summary
Further research design details are available in the Nature Portfolio Reporting Summary linked to this article, providing complete transparency and methodological rigor for the Drosophila brain wiring diagram project.
