Synthetic Neurons Are Making Brain Mapping Faster, and That Matters

The Connectomics team at Google Research has been quietly chipping away at one of the hardest problems in neuroscience: mapping entire brains. They released the full fruit fly brain last year—166,000 neurons, years of work. Now they’re coming after mammals, and that’s a whole different beast.

A mouse brain is a thousand times larger than a fruit fly’s. A human brain is another thousand times beyond that. At those scales, every efficiency gain matters, and the team’s latest trick is generating synthetic neurons to train their AI models.

The bottleneck is manual proofreading

Connectomics starts with slicing brain tissue into impossibly thin sections, imaging them, then stacking and aligning the 2D images to reconstruct 3D neurons. AI models handle the initial segmentation, but human experts still have to proofread and correct the output. That manual step is the slowest part of the pipeline, and it’s what keeps us from mapping bigger brains faster.

The team’s existing reconstruction model, PATHFINDER, identifies neurite segments—the branches and axons that make up a neuron—and stitches them together. But real neurons come in wildly diverse shapes, and the model struggles with edge cases it hasn’t seen enough of.

MoGen: synthetic neurons on demand

The new paper, “MoGen: Detailed neuronal morphology generation via point cloud flow matching,” presented at ICLR 2026, tackles this by generating synthetic neuron geometries. The model starts with random point clouds and gradually morphs them into realistic neural shapes. Think of it as a generative model for neuron anatomy.

Training PATHFINDER on a mix of real and MoGen-generated synthetic neurons reduced reconstruction errors by 4.4%. That sounds small, but at the scale of a complete mouse brain, it translates to 157 person-years of manual proofreading saved. That’s not just a number—it’s the difference between a project taking a decade versus maybe seven years.

I’ve seen similar synthetic data approaches in other domains, like generating synthetic medical images to train diagnostic models. The idea is always the same: if you can’t collect enough real examples of rare or difficult cases, make them yourself. MoGen is doing exactly that for neuron morphology.

What makes neurons hard

Most cells in the body are roughly spherical. Neurons are not. They send signals along long, thin axons that curl and branch unpredictably. They receive signals through dendrites, which sprout short protrusions called dendritic spines. Each neuron forms thousands of synapses, specialized junctions where signals leap between cells.

This intricate geometry is directly tied to function, and it’s a nightmare for reconstruction algorithms. A model trained mostly on common shapes will miss the rare, weird neurons that might be critical for understanding how a circuit works. MoGen can generate those rare shapes on demand, filling the gaps in the training data.

The bigger picture

Google Research has been in the connectomics game for over a decade. They’ve mapped fragments of zebra finch brain, whole larval zebrafish brain, a small piece of human cortex, and they recently started on a mouse brain section. Each project pushes the tools further.

MoGen adds to a growing stack of foundational tools: segmentation models, proofreading interfaces, and now synthetic data generators. The 4.4% error reduction is modest, but it’s directional. The team suggests further improvements are possible, and I’d expect to see more generative approaches in future iterations.

What I find interesting is that they’re not trying to skip the manual proofreading step entirely. They’re making it faster, which is a more realistic goal. Full automation for brain mapping is still a long way off, but shaving years off the timeline for a mouse brain is a concrete win.

The fruit fly map was a milestone. The mouse brain will be a landmark. MoGen is one more tool that makes that landmark reachable.