What a synapse-resolution wiring diagram can and cannot tell you, and how to turn a broad brain question into a measurable connectomics study.
Key community resources for this unit:
| Time | ~90 min to work through, plus a 60 min lab |
| Prerequisites | None |
| You need | Paper or a text editor. No data or code. |
| You finish with | A one-page study brief for a connectomics question of your own |
This unit is deliberately not about microscopes. It is about the reasoning step that happens before anyone buys a microscope, and that most failed connectomics projects skip: deciding what claim the data will be asked to support.
By the end of this unit you should be able to:
Item 3 is the one that matters most. It is also the one most often failed.
Connectomics exists because of a physical mismatch between the size of the things that carry information in the brain and the resolution of the instruments that are cheap and fast.
| Structure | Approximate size | Resolvable by light microscopy (~200–250 nm limit)? |
|---|---|---|
| Cortical neuron soma | 10–25 µm | Yes, easily |
| Dendritic shaft | 0.5–3 µm | Yes |
| Myelinated axon | 0.5–2 µm | Usually |
| Dendritic spine head | 300–800 nm | Marginally |
| Unmyelinated axon in neuropil | 80–300 nm | No |
| Dendritic spine neck | 50–200 nm | No |
| Synaptic cleft | ~20 nm | No |
| Synaptic vesicle | ~40 nm | No |
| Postsynaptic density | 30–50 nm thick | No |
The consequence is specific, not general. Light microscopy can tell you that two neurons’ arbors overlap in space. It cannot tell you whether they are connected, because at 250 nm two membranes 20 nm apart are one blur. In dense cortical neuropil, a 1 µm³ box contains on the order of a dozen distinct neurites; below the diffraction limit they are indistinguishable.
Electron microscopy at 4 nm × 4 nm × 40 nm resolves all of the rows above. That is the entire reason the field tolerates the cost described in the next section.
A caution that belongs here and not later. Resolving a synapse is not the same as knowing its sign, strength, or whether it is active in a behaving animal. EM gives you an anatomical synapse: a presynaptic vesicle pool, a cleft, and a postsynaptic specialization. Sign (excitatory/inhibitory) is usually inferred from morphology (see Unit 05) or from a separate molecular measurement, not read off directly.
Sparse labeling plus diffraction-limited optics gives potential contact (arbor overlap), not synaptic connection. Contact-based predictions of connectivity — “Peters’ rule” style inference — are known to be poor predictors of actual synaptic connectivity in cortex: neurons that touch frequently often do not connect, and connection probability varies strongly by cell type. You would be measuring a proxy whose relationship to the quantity of interest is itself an open research question.
A secondary objection: sparse labeling means you cannot see the unlabeled partner, so even a true synapse has an anonymous other side.
You should be able to do this arithmetic in your head, because it is the fastest way to sanity-check any proposal.
Raw voxel count = (volume) ÷ (voxel volume).
Work an example. Take 1 mm³ of cortex at 4 nm × 4 nm × 40 nm:
x: 1 mm / 4 nm = 1,000,000 nm / 4 nm = 250,000 voxels
y: 1 mm / 4 nm = 250,000 voxels
z: 1 mm / 40 nm = 1,000,000 nm / 40 nm = 25,000 voxels
total = 250,000 x 250,000 x 25,000 = 1.56 x 10^15 voxels
At 8-bit grayscale that is ~1.56 petabytes of raw image data for one cubic millimeter — before alignment, before segmentation, before meshes, before any derived product. Published petascale volumes land in this range: the H01 human temporal cortex sample and the MICrONS mouse visual cortex volume are both roughly one cubic millimeter and both are reported in the 1.4–2 PB range depending on what is counted and how it is compressed.
Now extrapolate, and notice where it breaks:
| Target | Volume | Raw data at 4×4×40 nm | Status |
|---|---|---|---|
| C. elegans nervous system | ~0.00005 mm³ | ~0.1 TB | Done, repeatedly, since 1986 |
| Adult Drosophila brain | ~0.02–0.03 mm³ | ~40 TB | Done (FAFB / FlyWire, hemibrain) |
| Mouse cortical column / 1 mm³ | 1 mm³ | ~1.6 PB | Done (MICrONS) |
| Whole mouse brain | ~500 mm³ | ~800 PB | The current grand challenge |
| Whole human brain | ~1.2 × 10⁶ mm³ | ~10²¹ bytes | Not a plan; a horizon |
Teaching point. The jump from fly to mouse cortex is ~40×. The jump from 1 mm³ to whole mouse brain is ~500×, and it is not a jump in imaging alone — it is a jump in sectioning reliability, storage economics, alignment robustness, segmentation accuracy, and above all proofreading labor. Programs such as NIH BRAIN CONNECTS exist because that 500× is an engineering problem, not a microscopy problem.
Voxel volume = 8 × 8 × 8 = 512 nm³. 0.05 mm³ = 5 × 10⁷ µm³ = 5 × 10¹⁶ nm³. 5 × 10¹⁶ / 512 ≈ 1 × 10¹⁴ voxels ≈ 100 TB.
What changes: isotropy. At 4 × 4 × 40 nm the z-axis is 10× coarser than xy, so thin processes running in-plane are easy and processes crossing z steeply are hard — this anisotropy is the single largest driver of automated segmentation errors (Unit 08). Isotropic 8 nm removes that asymmetry, at the cost of throughput and of a hard limit on how large a volume FIB-SEM can practically mill.
This is the section to argue about. Sort every claim you plan to make into one of three bins.
Rule: any Bin B claim must name its assumption in the same sentence, or in the figure caption. “Putatively excitatory (asymmetric morphology)” costs four words.
A wiring diagram is a constraint on the space of possible dynamics, not a simulation. The most common failure of connectomics writing is a Bin C sentence in the abstract supported by a Bin A measurement in the results.
Version 0 (as originally proposed): “We will use connectomics to understand how the cortex implements predictive coding.”
This is Bin C, it has no measurable endpoint, and no result could falsify it. Repair it in four moves.
Move 1 — find the structural signature. Predictive-coding architectures generally require some form of asymmetry between feedforward and feedback pathways, and a population that receives both. In mouse visual cortex, a candidate structural signature is: feedback axons from higher visual areas preferentially target a different dendritic compartment of layer 2/3 pyramidal cells than feedforward thalamocortical and layer 4 inputs do.
Move 2 — define the measurement, with units. For each proofread L2/3 pyramidal cell, compute the fraction of input synapses by presynaptic source class, binned by path distance from soma (0–50, 50–150, 150–300, >300 µm) and by compartment (spine head, spine neck, shaft, soma, AIS). Report as synapses per source class per compartment per cell, n cells.
Move 3 — state the null. Sources are distributed across compartments in proportion to available postsynaptic surface area in each compartment. This null is important: dendrites have more distal surface than proximal, so “more distal synapses” is the expected result and is not evidence of anything.
Move 4 — write the non-claim. “These data constrain where feedback input arrives on the dendritic tree. They do not establish that these synapses carry prediction error signals, and they do not establish sign; sign is inferred from synaptic morphology and from the identified source cell type.”
Version 1 (usable): “In mouse visual cortex, do inter-areal feedback axons target a systematically more distal dendritic compartment of L2/3 pyramidal cells than L4 feedforward axons do, relative to a surface-area-proportional null, measured across ≥50 proofread cells?”
Version 1 can be answered, can be wrong, and can be scoped in a grant.
(i) Bin A, provided “random” is defined against a stated null model and the claim is restricted to the reconstructed population. It becomes Bin B if it is extended to “input is unstructured”, since unstructured with respect to one labeling scheme may be structured with respect to another.
(ii) Bin B, verging on C. “Is GABAergic and forms symmetric synapses onto the pyramidal cell’s soma and proximal dendrite” is Bin A/B. “Inhibits firing” is a functional claim; a GABAergic synapse can be depolarizing depending on chloride reversal potential and developmental stage. State it as “putatively inhibitory”.
(iii) Bin A, and note that it is a ratio, which quietly controls for a lot of reconstruction bias. Ratios between comparably reconstructed populations are more robust than absolute counts, which are sensitive to completeness. This is a habit worth acquiring early.
A structural claim is only as good as the completeness of the reconstruction behind it. Match the claim type to the reconstruction state you actually have:
| Claim type | Minimum reconstruction state | Typical failure if you skip this |
|---|---|---|
| “Cell X synapses onto cell Y” | Both partners proofread through the synapse; synapse manually verified | A merge error invents the connection |
| “Cell X has n inputs” | Full dendritic arbor proofread and closed | Split errors truncate the arbor; n is an undercount of unknown size |
| “Type A prefers type B over type C” | Both target populations proofread to comparable completeness | Differential completeness masquerades as biological preference |
| “Motif M is enriched” | A closed subgraph with quantified edge precision/recall | False edges from merges inflate dense motifs superlinearly |
| “This projection is absent” | Stated detection sensitivity and searched volume | Absence of evidence reported as evidence of absence |
The fourth row deserves emphasis and reappears in Unit 09: merge errors do not add noise symmetrically to motif counts. A single merge fuses two neurons’ partner lists, which manufactures triangles and reciprocal pairs at a rate far above the error rate itself. Motif analysis on unproofread segmentation is not conservative; it is biased in a specific and predictable direction.
Concrete anchors, so that “connectomics is useful” is a claim you can defend with examples rather than enthusiasm.
Note what is common to the useful results: each one is a census or a comparison, not an assertion about computation. That is the shape of a defensible connectomics result today.
These four slides are framing devices, not evidence. Use each one to rehearse the sort in §3: for whatever claim the slide invites, ask whether structure alone could establish it, whether it needs a declared assumption, or whether it needs physiology the data does not contain.
Techtalk S10: The motivating question, asked before any instrument is chosen. Use it to practice the move this unit is built on — name the measurable structural endpoint, the null model, and the explicit non-claim that would have to replace it before the question is fundable.
Techtalk S11: Framing for what counts as brain data. Read it against the resolution table in §1: for whatever measurement the slide treats as data, ask which rows of that table it resolves, and therefore whether it can distinguish arbor overlap from an actual connection.
Techtalk S12: The reverse-engineering analogy, and where it stops. Check it against the bin boundaries in §3 — a wiring diagram is a constraint on the space of possible dynamics, not a simulation, so any “this circuit computes X” reading of the analogy is a Bin C claim wearing Bin A clothing.
Module12 L1 S04: Scope framing for the course. Compare whatever ambition it states against the extrapolation table in §2, and ask where on the fly-to-mouse-to-human ladder it lands — and whether the limiting factor at that rung is microscopy or proofreading labor.
Attribution: neuroAI and outreach source decks (historical/context visuals).
Deliverable: one page. Produce it in this order; do not skip to the method.
| Not yet | Proficient | Strong | |
|---|---|---|---|
| Falsifiability | No stated outcome would disconfirm the hypothesis | A disconfirming outcome is named | Disconfirming outcome named and distinguished from “underpowered” |
| Measurement | Named qualitatively (“look at connectivity”) | Three measurements with units | Measurements chosen at the right scale for the hypothesis (local motif claims use local statistics, not whole-graph summaries) |
| Null model | Absent, or “compare to random” | A named null | Null preserves the nuisance structure that matters (degree, distance, cell-type composition) |
| Scope discipline | Bin C claims present | Non-claim stated | Bin B assumptions named inline where they occur |
| Feasibility | No data estimate | Volume and voxel size given | Data estimate computed, and proofreading effort acknowledged as the binding constraint |
Self-grading tip: the single best predictor of a weak brief is that step 7 is empty or vague. If you cannot name a sentence you refuse to write, you have not yet found the boundary of your evidence.
“We’ll figure out the analysis once we have the data.” Recover by writing the figure caption for your main result now, with fake numbers. If you cannot write the caption, the experiment is not specified.
Using whole-graph summary statistics to test a local hypothesis. Mean path length and global clustering coefficient are nearly useless for claims about a specific microcircuit, and they are highly sensitive to reconstruction errors. If the hypothesis is about a three-cell motif, measure the three-cell motif.
Treating a reconstruction boundary as a biological boundary. Neurons cut by the volume edge have truncated arbors. Any per-cell count is biased downward, and the bias is worse for cells near the edge and for cell types with large arbors. Either restrict analysis to cells whose relevant arbor is fully contained, or model the truncation explicitly.
Comparing populations that were proofread differently. If population A was proofread to completion and population B was not, every difference you find is confounded with effort. Match proofreading protocol before comparing, and report proofreading state per cell.
Some of what this unit teaches is technique. Some of it is professional norm — the things experienced people do without being asked, and which nobody states out loud because they assume you already know. Those are worth naming, because they are distributed unequally by background rather than by ability.
From this unit:
Name the assumption in the same sentence as the claim. “Putatively excitatory (asymmetric morphology)” costs four words. Writing “excitatory” instead is not a shorthand; it is a different claim.
Write down the sentence you refuse to write. Every study brief should carry an explicit non-claim. Reviewers read it as confidence, not weakness — but nobody tells you that, so most people leave it blank.
Prefer ratios to absolute counts when comparing. A ratio between comparably reconstructed populations quietly controls for a great deal of reconstruction bias. Experienced people reach for it automatically.
The collected set, and why making these explicit is a fairness intervention rather than etiquette, is in the hidden curriculum.
Imaging physics (Unit 03), segmentation methods (Unit 08), and the statistics of motif testing (Unit 09). It also does not cover non-EM connectomics — barcoding approaches such as MAPseq/BARseq, and diffusion MRI tractography — which answer different questions at different scales; see Unit 02 for how these fit together.
This unit is anchored to canonical papers and datasets used in connectomics practice. Use these as required preparation before activities.
Capability target: Frame a biologically meaningful connectomics question as a testable structural hypothesis with explicit evidence limits.
From Question to Test - Convert an aspirational idea into a measurable connectomics plan. The unit's own lab above is the graded version of this exercise; do that one.
Translate broad brain questions into testable structural hypotheses with clear evidence boundaries.