NeuroGT: Biophysically grounded graph transformers for self-supervised representation learning of neuronal morphology

The intricate shape of a neuron is a fundamental determinant of its computational role, and its quantitative analysis is crucial for deciphering brain function. The advent of high-throughput imaging has produced neuronal reconstructions at unprecedented scales, posing a central challenge: how to develop representations that are both computationally efficient and biologically faithful. Existing approaches often lack biophysically grounded inductive biases, overlooking physical constraints such as electrotonic signal attenuation. Furthermore, self-supervised pretraining schemes based on a single objective fail to reconcile local geometry with global morphology, limiting generalization in downstream tasks requiring a holistic structure-function understanding. To overcome these limitations, we introduce NeuroGT, a Graph Transformer framework that integrates biophysical inductive biases into a hybrid self-supervised learning paradigm. We propose two novel encodings: an Electrotonic Positional Encoding (EPE), derived from the cable equation to model voltage diffusion and decay along dendritic arbors, and a k-means Shortest Path Encoding (k-SPE), which captures the hierarchical topology of neuronal trees. Together, these encodings provide the Transformer with domain-specific structural knowledge, enabling it to learn biologically faithful and interpretable representations. Complementing this design, NeuroGT employs a multi-task objective that couples graph-level contrastive learning with node-level coordinate denoising, thereby enforcing both global morphological invariance and local geometric fidelity. Our framework is rigorously validated on multiple large-scale collections of rodent neuronal reconstructions. NeuroGT achieves state-of-the-art performance on benchmarks, including cell-type classification and neuron retrieval. The resulting embeddings align closely with anatomical brain-region organization and reveal salient morphology-function relationships, bridging the gap between data-driven representation learning and neurobiological discovery. The code is available at https://github.com/big-rain/NeuroGT.