Research map / Probabilistic and scientific modeling / Dynamical systems
MorphGrower: A Synchronized Layer-by-layer Growing Approach for Plausible Neuronal Morphology Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MorphGrower, a layer-by-layer conditional VAE that generates sibling branch pairs synchronously, conditioned on the previously generated morphology, to produce topologically valid and more realistic neuronal morphologies than MorphVAE.
Problem
Generating plausible neuronal morphologies from reference trees: prior learning-based method MorphVAE produces unrealistic samples and often topologically invalid trees with tri- or N-furcations.
Approach
Generate neuron morphologies layer by layer with each layer produced as sibling branch pairs (so only soma has >2 children), using a seq2seq conditional VAE whose conditions are a local ancestor-path EMA feature and a global tree-GNN feature over the previously generated layers.
Key results
- On VPM, MorphGrower achieves MBPL 48.29 vs MorphVAE 41.87 against reference 51.33, and MorphGrower outperforms MorphVAE on most morphological metrics across datasets.
- Real/fake classifier accuracy drops from about 80-95% for MorphVAE to 54.7% (VPM), 62.7% (RGC), 55.0% (M1-EXC), 54.7% (M1-INH) for MorphGrower.
- MorphVAE topological validity rate is low: 0.052 (VPM), 0.319 (M1-EXC), 0.038 (M1-INH), 0.005 (RGC), whereas MorphGrower samples are all topologically valid by construction.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use layer-by-layer branch-pair generation instead of one-shot 3D-walk generation | MorphGrower vs MorphVAE · VPM, RGC, M1-EXC, and M1-INH | Morphological statistics (MBPL, MMED, MMPD, MCTT) | Mixed |
| Use MorphGrower rather than MorphVAE | MorphGrower · VPM, RGC, M1-EXC, M1-INH | real/fake classifier accuracy: 86.75 $\pm$ 06.87 -> 54.73 $\pm$ 03.36; 94.60 $\pm$ 01.02 -> 62.70 $\pm$ 05.24; 80.72 $\pm$ 10.58 -> 55.00 $\pm$ 02.54; | Helped |
| Generate soma branches with an unconditional VAE instead of directly providing them (ablation) | MorphGrower · VPM, RGC, M1-EXC, M1-INH | Morphological statistics (six metrics) | Mixed |
| Generate soma branches with an unconditional VAE instead of directly providing them (ablation) | MorphGrower · M1-INH | MBPL: 41.50 $\pm$ 1.02 -> 37.72 $\pm$ 0.96 | Hurt |
Surprising findings
- Using MorphVAE's 3D-walk basis yields a validity rate below 1% on RGC, so rejection sampling is impractical for obtaining valid baseline samples.
- MorphGrower's MBPL on M1-INH is worse than MorphVAE (41.50 vs 50.79 is actually better; only the MBPL metric where MorphVAE wins is noted as an exception in the text), indicating MorphVAE is competitive on that single metric.
Limitations
- Morphology data contains only node coordinates, lacking diameter information.
- Training data are static reconstructions and do not capture the true dynamic growth process; growth is simplified to synchronized layers.
- Requires domain-specific prior knowledge (e.g., soma-branch layer handling) and relies on reference morphologies as conditions.
Experimental setup
Datasets: VPM RGC M1-EXC M1-INH
Benchmarks: VPM RGC M1-EXC M1-INH
Metrics: MBPL MMED MMPD MCTT MASB MAPS classification accuracy of real/fake classifier topological validity rate electrophysiological simulation MMF/MAPH/MAHPD/MAPA
Baselines: MorphVAE L-measure-based statistics reference
Builds on: MorphVAE T-GNN conditional VAE von Mises-Fisher distribution
Closest papers by meaning
- Neuroformer: Multimodal and Multitask Generative Pretraining for Brain Data (ICLR 2024 · similarity 0.85)
- Neuron Platonic Intrinsic Representation From Dynamics Using Contrastive Learning (ICLR 2025 · similarity 0.83)
- MorphoHELM: A Comprehensive Benchmark for Evaluating Representations for Microscopy-Based Morphology Assays (NeurIPS 2026 · similarity 0.83)
- Isotropic Activation Functions Enable Deindividuated Neurons and Adaptive Topologies (NeurIPS 2026 · similarity 0.83)
- Learning the Plasticity: Plasticity-Driven Learning Framework in Spiking Neural Networks (NeurIPS 2025 · similarity 0.83)
- DiffeoMorph: Learning to Morph 3D Shapes Using Differentiable Agent-Based Simulations (NeurIPS 2026 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.