Wiring Patterns Alone Can Identify Neuron Types With Over 90% Accuracy

Based on research from the Japan Advanced Institute of Science and Technology (JAIST), Princeton University, the University of Edinburgh, and the Polytechnic University of Catalonia, published in Natu

Every neuron in the brain is defined not just by its shape but by its conversations—the thousands of synaptic connections it forms with other cells. Yet for decades, neuroscientists have classified neurons primarily by how they look under a microscope, a painstaking process that requires expert annotation and scales poorly to the massive connectome datasets now emerging from modern electron microscopy. What if the wiring itself—stripped of all morphological detail—could reveal what kind of neuron you are looking at?

An international team spanning four institutions and three continents has shown that it can. Their algorithm, called NTAC (Neuron Type Analysis from Connectivity), identifies neuron types from synaptic connectivity patterns alone, achieving over 90% accuracy while requiring only a small fraction of labeled training data. The results, published in Nature Communications in January 2026, suggest that the connectome carries far more cell-identity information than previously appreciated—and that we have been underusing the richest data source available in modern brain mapping.

Key Discovery

The core insight behind NTAC is deceptively simple: neurons of the same type tend to connect to the same partners. Rather than examining dendritic branching patterns or soma size, the algorithm builds a profile of each neuron based entirely on whom it talks to—its pre- and postsynaptic partners and the strengths of those connections.

Working with dense connectome reconstructions of the Drosophila (fruit fly) brain, the researchers tested NTAC against established morphology-based classification methods. The results were striking. NTAC achieved classification accuracy above 90% across multiple brain regions, while morphology-based approaches often fell below 10% accuracy in comparable tasks. Even more remarkably, when run in a fully unsupervised mode—with no labeled examples at all—NTAC still recovered approximately 70% of known neuron types correctly.

The algorithm works in two stages. First, it constructs a connectivity fingerprint for each neuron, encoding the pattern and weight of its synaptic inputs and outputs. Second, it applies a semi-supervised clustering approach that can leverage even a handful of labeled neurons to dramatically boost classification performance. This means that expert annotations, which are expensive and slow to produce, can be used sparingly rather than exhaustively.

Why This Matters

The timing of this work is not accidental. Neuroscience is in the middle of a connectomics revolution. In 2023, researchers published the first complete wiring diagram of an adult Drosophila brain, containing roughly 140,000 neurons and tens of millions of synaptic connections. Similar efforts are underway for portions of the mouse and even human brain, generating datasets of staggering complexity.

The bottleneck is no longer imaging—it is interpretation. Electron microscopy can now produce nanometer-resolution volumes faster than scientists can annotate them. Every neuron must be segmented, traced, and ultimately classified into a type, a process that has traditionally required trained neuroanatomists spending months or years with each dataset.

NTAC offers a way to shortcut that pipeline. By extracting cell-type information directly from the connectivity graph, it turns the connectome’s greatest asset—its comprehensive synapse-level wiring data—into a classification tool. This could dramatically accelerate the annotation of new connectome datasets and reduce dependence on subjective morphological criteria that can vary between labs and species.

The approach also aligns with a growing recognition in neuroscience that connectivity may be a more fundamental marker of neuronal identity than morphology. Two neurons can look similar under the microscope yet serve entirely different circuit functions based on their wiring. Conversely, neurons with different shapes may play equivalent roles if they share the same connectivity logic.

The Bigger Picture

NTAC sits at the intersection of three powerful trends reshaping brain science.

The first is the connectomics race. Following the completion of the Drosophila connectome, multiple groups are now pushing toward whole-brain wiring diagrams in larger organisms. The Allen Institute, Google, and academic consortia are generating petabyte-scale datasets of mouse cortex. Automated analysis tools like NTAC will be essential infrastructure for making sense of these maps.

The second trend is the application of artificial intelligence to neuroscience. Machine learning has already transformed connectome reconstruction, with deep networks performing neuron segmentation and synapse detection at near-human accuracy. NTAC extends this paradigm to the higher-level problem of cell-type classification, showing that graph-based computational approaches can extract biological meaning that escapes traditional analysis.

The third is the legacy of the Human Brain Project and related large-scale brain initiatives. These programs invested heavily in building shared data standards, atlases, and computational platforms. NTAC’s ability to work across brain regions with minimal supervision makes it a natural fit for these infrastructure ecosystems, where scalable and reproducible classification methods are urgently needed.

The international character of the collaboration—spanning JAIST in Japan, Princeton in the United States, Edinburgh in the United Kingdom, and the Polytechnic University of Catalonia in Spain—itself reflects the increasingly global and interdisciplinary nature of connectomics research.

Limitations and What Comes Next

The authors are forthright about the boundaries of their current work. The most significant limitation is that NTAC has been validated exclusively in the fruit fly brain. Drosophila offers a uniquely favorable testing ground: its brain is small enough to reconstruct completely, and its neuron types are relatively well characterized through decades of genetic and behavioral studies.

Mammalian brains present a different challenge. Cortical neurons are far more numerous, their connectivity patterns are sparser and more variable, and complete connectomes do not yet exist for any mammalian brain region above a cubic millimeter. Whether connectivity fingerprints will be as diagnostic in mouse or human cortex—where a single neuron type may exhibit substantial wiring variability across individuals—remains an open question.

There are also conceptual questions. Neuron types are not always cleanly separable; some populations grade continuously into one another, and type boundaries can depend on which features you prioritize. NTAC’s performance in such ambiguous cases has not been fully explored.

Looking ahead, the researchers suggest several extensions. Combining connectivity-based classification with morphological and transcriptomic data could yield even more robust typologies. Applying NTAC to partial connectomes—where only a fraction of neurons have been reconstructed—would test its practical utility in the incomplete datasets that are far more common than whole-brain wiring diagrams. And adapting the algorithm for the larger, noisier graphs characteristic of vertebrate brains will be a critical next step.

At a Glance

• What: NTAC, a new algorithm that classifies neuron types using only synaptic connectivity data, without relying on cell morphology.
• Accuracy: Over 90% in supervised mode; approximately 70% in fully unsupervised mode with no labeled training examples.
• Comparison: Morphology-based classification methods fell below 10% accuracy on equivalent tasks.
• Organism: Validated on dense connectome reconstructions of the Drosophila (fruit fly) brain.
• Implication: Synaptic wiring patterns carry rich cell-identity information that could accelerate brain mapping efforts worldwide.
• Collaboration: Researchers from JAIST (Japan), Princeton (USA), Edinburgh (UK), and the Polytechnic University of Catalonia (Spain).