Synthetic Neurons Are Making Brain Mapping Less Excruciating

I’ve been following the connectomics space for a while, and the bottleneck has always been the same: you can image a brain at incredible resolution, but turning those images into accurate 3D models of individual neurons is agonizingly slow. Google Research just dropped a new approach that might actually move the needle.

Their new paper, “MoGen: Detailed neuronal morphology generation via point cloud flow matching,” is set to appear at ICLR 2026. The core idea is elegant: instead of relying solely on real neuron scans for training data, they generate synthetic neurons that look and behave like the real thing. These fake neurons then help train the AI models that do the actual reconstruction work.

The result? A 4.4% reduction in reconstruction errors. That doesn’t sound like much until you do the math. For a complete mouse brain — roughly a thousand times larger than the fruit fly brain they just mapped — that 4.4% translates to 157 person-years of manual proofreading saved. Let that sink in. That’s not just a tweak; that’s the difference between a project being feasible and being a pipe dream.

The synthetic neuron factory

MoGen works by learning the statistical patterns of real neuron shapes from a training set, then generating new variations. The model uses a point cloud representation — essentially a cloud of 3D points that defines the neuron’s structure — and applies a flow matching technique to gradually morph random noise into realistic morphologies. The animation in the paper shows this process in action: blobs of points slowly twisting into the spindly, branching shapes that characterize real neurons.

This is clever because real neuron data is scarce and expensive to produce. Each neuron requires hours of expert annotation. By generating synthetic examples, MoGen effectively multiplies the available training data without the cost. The team trained it on mouse neurons, but the approach should generalize to other species.

Why shape matters

Most cells in the body are roughly spherical. Neurons are not. They have long, thin axons that can curl and branch, and dendrites with tiny spines that receive signals. This complex geometry is directly tied to function — how a neuron connects, how fast it fires, what it responds to. Getting the shape wrong means getting the wiring wrong.

Google’s current reconstruction model, PATHFINDER, works by identifying neurite segments (the branches) and stitching them together into complete neurons. Training it on synthetic data from MoGen helps it handle the weird edge cases — the neurons that don’t look like the textbook examples — without needing humans to manually flag every anomaly.

The bigger picture

This isn’t Google’s first rodeo in connectomics. They’ve been at this for over a decade, mapping fragments of zebra finch brain, larval zebrafish brain, and even a small chunk of human brain. They recently launched an effort to map a section of mouse brain. The fruit fly brain they released earlier this year, with 166,000 neurons, was a milestone, but it also highlighted how far we are from mammalian-scale maps.

A mouse brain is a thousand times larger than a fruit fly brain. A human brain is a million times larger. At current rates, mapping a full human brain would take centuries. Techniques like MoGen are necessary to compress that timeline into something a human researcher might see in their lifetime.

My take

The 4.4% improvement is real, but I’m more interested in the direction this points. Synthetic data generation has been a game-changer in other areas of computer vision — why not here? The paper acknowledges that the current model still has limitations: it doesn’t capture all the fine details of dendritic spines or synaptic connections. But as a proof of concept, it’s solid.

I also appreciate that Google published the model and the paper openly. Too much of this kind of work gets locked behind proprietary walls. The Connectomics team has a history of releasing tools and datasets, which is how the field advances.

One thing I’d like to see: a direct comparison of how much synthetic data is needed to achieve that 4.4% gain. Is it a 10x increase in training examples? 100x? The paper is light on that detail. Also, I wonder how well the synthetic neurons capture the variability across different brain regions. A cortical neuron looks very different from a Purkinje cell in the cerebellum. If the training data doesn’t cover that diversity, the model might not generalize.

Still, this is a smart, practical step. Brain mapping is one of those problems where every incremental improvement compounds. A 4.4% error reduction today might be a 20% reduction in a few years as the models get better. And 157 person-years saved? That’s not incremental. That’s transformative.