MoGen: How Google Is Using Synthetic Neurons to Speed Up Brain Mapping

You’ve probably seen those stunning wiring diagrams of the fruit fly brain—166,000 neurons mapped out in glorious detail. That was a massive achievement, but it took years of AI-assisted work and human experts squinting at data. Now imagine doing that for a mouse brain, which is a thousand times larger, or a human brain, which is another thousand times larger on top of that. We’re not there yet, and brute force won’t get us there.

Google Research’s connectomics team has been chipping away at this problem for over a decade. Their latest trick, published as a paper called “MoGen: Detailed neuronal morphology generation via point cloud flow matching” and accepted at ICLR 2026, takes a clever shortcut: instead of relying solely on real neural data to train their AI, they generate synthetic neurons that look and behave like the real thing.

The idea is straightforward. AI models that reconstruct 3D neurons from microscope images need lots of training examples. Real examples are expensive to produce and annotate. So why not generate fake ones that are good enough to teach the model? That’s what MoGen does. It starts with random point clouds and gradually morphs them into realistic neural shapes—axons, dendrites, spines, the whole spindly mess.

What’s the payoff? A 4.4% reduction in reconstruction errors. That doesn’t sound earth-shattering until you run the numbers. At the scale of a complete mouse brain, a 4.4% error drop translates to roughly 157 person-years of manual proofreading saved. That’s real time, real money, and real progress toward mapping something that actually matters for understanding mammalian brains.

I’ll be honest: I was skeptical when I first saw the number. Four-point-four percent feels like a rounding error in some contexts. But in connectomics, where every misidentified branch or missed synapse compounds across millions of neurons, even small improvements in the base model cascade into massive savings downstream. The bottleneck isn’t the AI—it’s the humans who have to verify everything. Cut their workload by even a fraction, and you unblock the whole pipeline.

MoGen isn’t some standalone tool either. It’s part of a broader ecosystem Google has been building: the PATHFINDER reconstruction model, the Connectomics website, and partnerships that have already produced maps of larval zebrafish, zebra finch, and small chunks of human brain. They recently started on a mouse brain section. The synthetic neuron approach slots right in as a data augmentation layer—more training data, better models, fewer errors.

Of course, there’s a limit to how far synthetic data can take you. Neurons are wildly diverse. Some are long and spindly, others are compact and bushy. Their shapes encode function. If your generative model doesn’t capture that diversity, you risk training your AI on a cartoon version of reality. MoGen seems to handle this reasonably well based on the error reduction, but I’d want to see how it performs on edge cases—rare neuron types or pathological tissue—before declaring victory.

Still, this is the kind of incremental, practical work that moves the needle in computational neuroscience. No hype about AGI or brain-computer interfaces. Just better tools for a hard problem. And if it helps us get to a full mouse connectome before I retire, I’ll take it.