NeuroTrailblazers banner

Overview

Identifying neuron types is essential for interpreting a connectome. A wiring diagram of undifferentiated nodes has far less explanatory power than one where each node carries a cell-type label. Cell-type classification determines which connectivity patterns are “expected” (excitatory neurons connecting to nearby neurons) versus “surprising” (rare long-range inhibitory connections). In EM connectomics, cell types are inferred from morphology, connectivity patterns, and — when available — correlative functional or molecular data.

Instructor script: the cell-type classification challenge

What defines a cell type?

This is one of the most debated questions in neuroscience. At a practical level, cell types are groups of neurons that share morphological, physiological, and molecular properties. The challenge for EM connectomics is that we have access to morphology and connectivity but typically not to molecular markers or electrophysiology (with exceptions like the MICrONS dataset, which combines EM with calcium imaging).

Historical classification systems (Cajal, Lorente de Nó, Markram et al.) relied on morphology from Golgi staining and intracellular fills. EM connectomics provides a different view — more complete morphology (every branch, every spine) but without the staining selectivity that highlights individual cells against a blank background.

The major division: excitatory vs inhibitory

In mammalian cortex, the most fundamental classification is:

Excitatory neurons (~80% of cortical neurons):

Inhibitory interneurons (~20% of cortical neurons):

EM classification rule of thumb:

This works for the majority of cortical neurons but has exceptions (some interneurons have sparse spines; spiny stellate cells in layer 4 look different from pyramidal cells).

Morphological classification in EM

Pyramidal neurons

The most common excitatory neuron in cortex (layers 2-6):

Identification cues:

Subtype classification by layer and projection: | Subtype | Layer | Projection target | EM distinguishing features | |———|——-|——————-|—————————| | Layer 2/3 pyramidal | 2/3 | Other cortical areas (callosal, associational) | Medium soma, prominent apical reaching L1 | | Layer 4 spiny stellate | 4 | Local (within column) | Stellate dendrites (no clear apical), heavily spiny | | Layer 5 thick-tufted (ET) | 5 | Subcortical (thalamus, brainstem, spinal cord) | Large soma, thick apical with prominent L1 tuft, thick axon | | Layer 5 thin-tufted (IT) | 5 | Other cortical areas | Smaller soma, thinner apical, less prominent tuft | | Layer 6 corticothalamic | 6 | Thalamus | Apical reaches L4 (not L1), distinctive morphology |

Inhibitory interneuron types

Inhibitory neurons are far more morphologically diverse. The major types identifiable in EM:

Basket cells (PV+):

Chandelier cells (PV+):

Martinotti cells (SST+):

Bipolar/VIP+ cells:

Neurogliaform cells:

Connectivity-based classification

The connectivity fingerprint

Even without morphological reconstruction, neurons can be classified by their connectivity pattern alone:

Neurons of the same type tend to have similar connectivity fingerprints. This allows clustering-based classification using the connectome graph directly.

Methods

  1. Feature engineering: For each neuron, compute: in-degree, out-degree, fraction of input from excitatory vs inhibitory sources, fraction of output onto soma vs dendrites, laminar distribution of inputs/outputs.
  2. Dimensionality reduction: PCA, UMAP, or t-SNE on the feature vectors.
  3. Clustering: K-means, hierarchical clustering, or Gaussian mixture models on the reduced representation.
  4. Validation: Compare to morphological types (where known) or molecular markers (if available from correlative data).

Example from FlyWire

In the FlyWire whole-brain connectome (Dorkenwald et al. 2024, Schlegel et al. 2024), cell types were assigned by combining:

This hybrid approach identified ~8,000 cell types in the adult Drosophila brain.

Example from MICrONS

Turner et al. (2022) classified neurons in mouse visual cortex using:

Worked example: classifying a neuron in layer 2/3

Given: A fully reconstructed neuron in layer 2/3 of mouse visual cortex.

Step 1: Excitatory or inhibitory?

Step 2: Morphological subtype

Step 3: Connectivity check

Step 4: Functional data (if available)

Classification: Layer 2/3 pyramidal neuron, likely callosal-projecting (based on axon trajectory toward white matter). Confidence: high.

Challenges and limitations

Incomplete reconstructions

Most neurons in a connectomics volume are not fully reconstructed — their axons or dendrites extend beyond the imaged volume. Classification based on partial morphology is less reliable. Solution: weight classification by the available evidence and flag incompleteness.

Continuous variation

Cell types are not always discrete categories — there can be continuous variation within types (e.g., L5 pyramidal cells show a spectrum from thick-tufted to thin-tufted). Whether to split one type into two or treat it as one type with variation is a judgment call that depends on the analysis question.

Species differences

Cell-type taxonomies developed in mouse may not transfer directly to human or fly. The same morphological features may indicate different types in different species. Cross-species comparison requires careful homology mapping.

Common misconceptions

Misconception Reality Teaching note
“Cell types are discrete and obvious” Many neurons fall on continua between types; classification depends on the criteria used Report classification confidence and criteria
“Morphology alone is sufficient” Molecular markers and physiology can distinguish types that look similar in EM Use all available evidence; flag morphology-only classifications
“The same types exist in all species” Cell-type diversity varies across species and regions Don’t assume mouse taxonomy applies to fly or human
“More types = better classification” Over-splitting creates types with too few members for statistical analysis Balance granularity with statistical power

References