MoGen: Using AI-Generated Fake Neurons to Make Brain Mapping Less Painful

Google Research just dropped a new paper at ICLR 2026, and honestly, it’s one of those rare pieces of work that makes you think “why didn’t anyone do this sooner?” They’re calling it MoGen — short for Neuronal Morphology Generation — and it’s essentially a model that creates fake neurons to help AI get better at recognizing real ones.

Let me back up a bit.

Brain mapping, or connectomics as the cool kids call it, is the process of reconstructing every neuron in a brain to create a complete wiring diagram. The fruit fly brain they released recently? 166,000 neurons. That took years of AI-assisted work plus human experts hunched over screens proofreading. A mouse brain is about a thousand times bigger. A human brain is a thousand times bigger than that. You can see where this is going.

The bottleneck isn’t the imaging anymore — we’ve got electron microscopes that can slice and scan brain tissue at insane resolutions. The bottleneck is segmentation: turning those 2D image stacks into accurate 3D neuron shapes. Google’s PATHFINDER model does this by identifying neurite segments and stitching them together, but it still needs a ton of human proofreading to fix errors.

Here’s where MoGen comes in. Instead of relying only on real neuron shapes for training data — which are scarce and expensive to produce — MoGen generates synthetic neuron geometries. It uses something called point cloud flow matching to gradually morph random point clouds into realistic neural shapes. The results are convincing enough that training PATHFINDER on a mix of real and synthetic data reduces reconstruction errors by 4.4%.

I know what you’re thinking: 4.4% sounds like a rounding error. But scale matters. For a complete mouse brain, that 4.4% translates to 157 person-years of manual proofreading saved. That’s not nothing. That’s a whole lab’s worth of grad students freed up to do actual science instead of staring at neuron reconstructions all day.

The synthetic neurons aren’t just random blobs either. Real neurons come in a dizzying variety of shapes — long spindly axons, branching dendrites with little spines, all that intricate geometry that relates to function. MoGen captures enough of that diversity to be useful without needing to perfectly replicate every rare morphology.

Now, I’ve been around long enough to be skeptical of synthetic data claims. Sometimes you get models that just memorize the training set and poop out near-copies. But the MoGen team seems to have avoided that trap by using a flow-matching approach that generates novel geometries rather than interpolating between existing ones. The paper shows examples that look genuinely new, not just Frankenstein’d from real neurons.

Is this going to unlock whole-brain mapping overnight? No. But it’s the kind of incremental improvement that compounds. Better training data → better reconstruction models → less manual proofreading → faster mapping cycles. Each 4% improvement brings the mouse brain map years closer.

Google’s Connectomics team has been at this for over a decade, working with partners on fragments of zebra finch brain, whole larval zebrafish, and even a tiny slice of human brain. They’re now going after a small section of mouse brain. Tools like MoGen are exactly what they need to keep pushing the scale.

I’m curious to see how this generalizes. Can you train MoGen on one species and use it to improve reconstructions of another? The paper focuses on mouse neurons, but the approach seems species-agnostic. If it works across species, that’s a huge win for the field.

Either way, this is solid work. No hype, no overpromising. Just a clever solution to a real bottleneck, with numbers to back it up. That’s the kind of research I can get behind.