Neural Noise

MorphSys: branch-aware contrastive learning for neuron morphology graphs

2 min read

  • MorphSys is a self-supervised contrastive learning framework combining a branch-aware module and a GNN-based module for neuron morphology representation (Journal of Neural Engineering).1
  • Branch-level representation with inter-branch attention captures relationships among dendrite branches; supports cell typing and structure–function links.1
  • On the BIL (Brain Image Library) dataset, MorphSys achieves a KNN accuracy of 83.99%, surpassing the previous state-of-the-art by 2.99 percentage points in the neuron cell type classification benchmark.2
  • The Branch-Aware module introduces Inter-Branch Attention to explicitly model relationships between dendrite branches — a structural feature missed by conventional tree-graph models that rely solely on node-level message passing.2
  • The GNN-based module uses GatedGraphConv with SumPool as the optimal architecture after ablation study, demonstrating robust performance across multiple GNN architectures for learning local features of the neuron tree graph.2
  • MorphSys is self-supervised: it requires no cell-type labels during training, making it applicable to large-scale unlabeled morphology datasets from resources such as NeuroMorpho.org.2
  • Ablation studies confirmed that both the Branch-Aware module and the GNN module contribute independently to performance; in-depth analysis identified powerful approaches for branch feature extraction beyond standard graph convolution.2
  • Published in Journal of Neural Engineering (IOP Publishing); online ahead of print February 2026 (DOI: 10.1088/1741-2552/ae4381).1
  • Situates within a competitive field of neuron morphology representation methods including GraphDINO (transformer-based), TreeMoCo (LSTM + contrastive), and MorphRep (graph transformer pre-trained on >250K neurons); MorphSys’s branch-level decomposition is a distinctive architectural choice.3
  • Downstream applications: accurate cell typing from morphology alone can support large-scale connectome analysis, optogenetic target identification, and interpretation of morphological changes in disease models.2

Footnotes

  1. https://iopscience.iop.org/article/10.1088/1741-2552/ae4381 ↩ ↩2 ↩3

  2. https://pubmed.ncbi.nlm.nih.gov/41662773/ ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  3. https://arxiv.org/abs/2112.12482 ↩

Graph View