Synthetic Neurons Are Giving Brain Mapping a Real Speed Boost

The idea of mapping an entire brain has always felt like one of those things that’s technically possible but practically insane. The fruit fly brain map released recently, with 166,000 neurons, took years of AI-assisted work and human experts. A mouse brain is a thousand times bigger. A human brain is a million times bigger than that fruit fly. You start to see the problem.

Google Research has been chipping away at this for over a decade through their Connectomics team. Their latest trick, presented at ICLR 2026, is a model called MoGen (Neuronal Morphology Generation) that creates synthetic neuron shapes. Not real ones, but plausible fakes that look realistic enough to train other AI models on.

Why fake neurons help

Most cells in your body are roughly spherical. Neurons are the weirdos of the cellular world, with long spindly axons that curl and branch, plus dendrites covered in tiny spines. This geometry matters for function, and it’s a nightmare for automated reconstruction.

The current pipeline starts with thin slices of brain tissue imaged at high resolution. Those 2D slices get stacked and aligned, then AI models segment them into 3D shapes. The output gets proofread by humans. That manual proofreading is the bottleneck. It’s slow, expensive, and scales poorly.

MoGen generates synthetic neuron geometries from point clouds, gradually morphing random noise into realistic neural shapes. Training the AI on these synthetic examples alongside real data reduced reconstruction errors by 4.4%. That sounds small, but at the scale of a mouse brain, it translates to 157 person-years of manual proofreading saved. I’d call that meaningful.

The model was trained on mouse neurons, and the animations show the point clouds converging toward recognizable shapes. It’s not generating new biological insights, but it doesn’t need to. It’s making the existing reconstruction process faster and cheaper.

The bigger picture

Google Research’s previous model, PATHFINDER, already improved how neurite segments get combined into full neurons. MoGen adds another layer by enriching the training data. The combination is what moves the needle.

The team has also mapped fragments of zebra finch brain, whole larval zebrafish brain, and a small piece of human brain. They recently started working on a small section of mouse brain. Each step is incremental, but the direction is clear: automated reconstruction at scale.

I’ll be honest, I was skeptical when I first saw the 4.4% figure. But the savings in manual labor make it clear why this matters. Connectomics is one of those fields where small improvements compound dramatically because the datasets are enormous.

What’s next

The obvious next step is scaling MoGen to larger and more diverse neuron types. The current model focuses on mouse neurons, but the approach should generalize. The paper is open access, and the model is available on GitHub, so the community can build on it.

There’s also the question of whether synthetic data can eventually replace real training data entirely. Probably not for the foreseeable future, but as a complement, it’s already proving its worth.

Brain mapping isn’t going to get faster overnight. But tools like MoGen are chipping away at the bottlenecks, one synthetic neuron at a time.