NTAC: AI for typing neurons by connections

NTAC: How AI Classifies Neurons by Connections, Not Shape — A Technical Breakdown for Neuroinformaticians

NTAC (Neuronal Type Assignment from Connectivity) is not another marketing breakthrough in neuroinformatics, but a practically applicable neuron classification algorithm based exclusively on synaptic connection topology. It eliminates the dependency on morphological identification, which becomes a bottleneck when scaling connectomes: in the Drosophila visual lobe, morphologically homogeneous neurons achieve 90% classification accuracy with NTAC versus 50% with NBLAST—all on consumer hardware without GPU acceleration.

Why Morphology is Misleading — and Why This Matters for Connectomics

In traditional neuroanatomy, cell type is defined by dendrite shape, axon structure, localization, and marker expression. However, in densely packed regions—such as the fruit fly's visual lobe—dozens of neuron types exhibit nearly identical morphology. Their differences manifest only in patterns of incoming and outgoing synapses: which neurons activate them, who they project to, at what frequency, and in what order. This information is encoded in connectome graphs, but until NTAC, it was not used systematically for typification.

NTAC proves the hypothesis: synaptic connectivity is the functional "genome" of a neuron. Just as DNA sequence determines protein, the set of input/output edges defines the cell's computational role in the chain. This fundamentally changes the approach to annotating connectomes: instead of manual shape matching, we use automated learning on graph structure.

NTAC Architecture and Operating Modes

The algorithm is implemented as a graph neural network working with directed weighted graphs, where vertices are neurons and edges are synapses with attributes (type: excitatory/inhibitory, contact count, localization within the dendritic tree). Input data is an N×N connectivity matrix (N = number of neurons), normalized by rows and columns to compensate for projection density variations.

NTAC supports two modes:

• Semi-supervised

- Input includes a subset of neurons with known type labels (typically 1–5% of the total)

- The network minimizes cross-entropy between predicted and true labels

- Uses a Graph Convolutional Network (GCN) with two hidden layers and dropout=0.3

- Accuracy: 90.2% on the Drosophila visual lobe connectome (BANC dataset)

• Unsupervised

- No labels: the network applies a modified deep clustering algorithm (DeepCluster v3)

- Graph embeddings are projected into 64-dimensional space, then clustered via k-means with adaptive k selection through silhouette analysis

- Accuracy: 70.4% in the visual lobe, 52.1% on the full fly brain connectome

Both modes run on an Intel i7-11800H CPU in ≤ 4.2 minutes at N=2500—without CUDA or specialized libraries like PyTorch Geometric. Code is written in pure NumPy and SciPy with minimal dependencies.

Comparison with NBLAST and Other Benchmarks

NBLAST—the standard tool for morphological comparison—builds pairwise similarity scores based on matching neuronal tree fragments. Its limitations are obvious:

• Requires full morphological reconstruction (often impossible with low SNR)
• Sensitive to segmentation artifacts and orientation
• Does not scale: calculation time grows as O(N²); at N=1000 it exceeds 12 hours on a single core
• In the visual lobe, it reaches only 48.7% accuracy (F1-score), as 63% of neurons have <5% shape differences

NTAC uses only graph metrics:

• In-degree and out-degree
• Clustering coefficient of the local subgraph
• Average shortest path length to nearest neighbors
• Betweenness centrality in a 50-neuron subnetwork

These 4 metrics are computed in O(N·k), where k is average degree, and serve as input for the classifier. Experiments showed that replacing any of them with random values reduces accuracy by 12–27%.

Key Takeaways

• NTAC proves that synaptic topology contains sufficient information for neuron typification—without morphology, markers, or electrophysiological parameters.
• The algorithm runs on standard laptops: no dependency on GPUs, cloud resources, or HPC clusters.
• Semi-supervised mode requires <5% labeled data—critical for rare cell types difficult to identify manually.
• NTAC is already integrated into the BANC (Brain and Nervous System Connectome) pipeline and used to annotate 12,437 neurons in the current release.
• The method does not replace but complements electrophysiology: predicted types provide hypotheses for subsequent patch-clamp or optogenetics verification.

Application in Large-Scale Projects

The current frontier of connectomics is mouse brain mapping (~70 million neurons). A full connectome would require ~10⁶ hours of manual typification using traditional methods. NTAC reduces this estimate to ~2,000 CPU hours (processing blocks of 5,000 neurons in parallel). Crucially, the algorithm maintains accuracy while scaling—tests on samples sized 10K and 50K showed stable deviation <1.3%.

For the human brain (86 billion neurons), NTAC is not yet applicable in its "full version," but its architecture formed the basis of a hierarchical approach: first clustering at the microcircuit level (cortical columns), then intracolumnar typification. Early prototypes are already being tested on human cortex samples (post-mortem, 3D-EM).

The next stage is integration with single-cell RNA-seq data. Researchers from JAIST and Princeton are developing a hybrid model where NTAC graph embeddings are combined with transcriptomic profiles via cross-modal attention. The goal is to build a unified "structure-function-expression" space where any neuron can be characterized by any of the three modal features.

NTAC is not a final solution but a new foundational layer in the connectome analysis stack. Its value lies not in absolute accuracy, but in reproducibility, scalability, and openness: source code is published under MIT License, documentation includes Jupyter notebooks with live examples on Drosophila and C. elegans.

— Editorial Team