MorphSys: a branch-aware contrastive learning framework for neuron morphology graphs.

Abstract

<i>Objectives.</i>Neuron morphology critically defines cellular identity and function, informing cell typing, soma localization, and neurological disorder diagnosis. However, two main challenges hinder progress: the difficulty of learning meaningful representations from complex, tree-like structures, and the limited availability of expert annotations at scale.<i>Approach.</i>To address these challenges, we propose MorphSys, a self-supervised contrastive learning framework that complements a branch-aware module and a graph neural network (GNN)-based module. We present a branch-level representation of neuron morphology by introducing an inter-branch attention, which captures inter-branch relationships among dendrite branches that are overlooked by conventional graph models relying on node-level message passing. In parallel, the GNN-based module robustly encodes local morphological patterns across different GNN architectures, providing complementary fine-grained structural features.<i>Main results.</i>Comprehensive experiments on five datasets demonstrate the superior performance of MorphSys in downstream tasks. In self-supervised neuron morphological classification, MorphSys achieves a k-nearest neighbor accuracy (KNN-Acc) of 83.52% on the N7 dataset, outperforming the previous state-of-the-art by 3.77%. For soma location prediction, MorphSys attains the highest KNN-Acc of 83.99%, 67.32%, and 75.33% on the Brain Image Library, ACT, and Janelia MouseLight datasets compared to baseline methods. The repeated trials and statistical analysis also confirm that the observed performance gains are highly significant.<i>Significance.</i>These results highlight that MorphSys serves as an effective tool for learning robust representations of neuron morphology and performing morphology-driven neuronal analysis. The code is available athttps://github.com/YuuYuuYuuY/MorphSys.