Google Research has released MoGen, a generative AI model that produces synthetic 3D neuron geometries for use as training data in brain-mapping AI pipelines. The paper, “MoGen: Detailed neuronal morphology generation via point cloud flow matching,” is being presented at ICLR 2026. According to the research post, using a PATHFINDER reconstruction model trained with 10% MoGen-generated synthetic data reduced reconstruction errors by 4.4%, primarily through a reduction in merge errors. At the scale of a complete mouse brain, the post estimates this improvement is equivalent to saving 157 person-years of manual proofreading.
The background context makes that number easier to read. The fruit fly brain — with 166,000 neurons — required years of work by AI systems and human experts to map completely. A mouse brain is a thousand times larger. A human brain is a thousand times larger still. The proofreading bottleneck is the main obstacle to scaling connectomics to more complex organisms, and reducing error rates directly reduces how much human correction is required.
What connectomics reconstruction looks like
The reconstruction pipeline the post describes begins by imaging thin slices of brain tissue, then stacking, aligning, and segmenting the 2D images into 3D neuron shapes. Google Research’s most recent AI reconstruction model, PATHFINDER, works by first identifying neurite segments — subsections of a neuron — and then combining them into a complete neuron.
Two categories of error dominate the work for proofreaders. Split errors occur when two neurite segments that should be connected are instead separated by the model. Merge errors occur when two unrelated neurites are incorrectly combined. Both require manual correction by researchers, graduate students, or technical specialists. The 4.4% error reduction MoGen enables is driven primarily by fewer merge errors.
How MoGen generates neurons
Neurons are unusual training subjects. Unlike most cells, which are roughly spherical, neurons have highly varied spindly geometries — long, thin axons that can curl, twist, and branch; networks of dendrites with shorter protrusions called dendritic spines; and many synapses where signals pass between cells. That complexity is part of what makes connectomics reconstruction hard and what makes diverse synthetic training data potentially valuable.
MoGen uses the PointInfinity point cloud flow matching model, trained to gradually transform a random cloud of 3D points into realistic 3D neuronal shapes. The training data was 1,795 axons from mouse cortex tissue reconstructions that had been previously verified by humans. Points were sampled from the surfaces of those reconstructed neurons to create the training set.
To validate that MoGen’s output was realistic, human experts were asked to classify a set of neurites that was a mix of real and simulated. The post reports that the simulated neurons passed that validation step. Millions of those synthetic shapes were then added to the PATHFINDER training pipeline.
The post also notes that MoGen can be directed to generate specific types of neurons by tuning parameters such as length, spatial extent, and branching. The initial study used a random assortment of shapes, but the post suggests that future work could focus on geometries that are especially prone to reconstruction errors — a more targeted use of synthetic data.
Species coverage and open release
Because neuron shapes vary across species, the team also trained species-specific versions of MoGen on zebra finch and fruit fly neurons, in addition to the primary mouse model. That work came from prior collaboration with Google Research partners on a zebra finch brain fragment, a complete larval zebrafish brain, and fragments of human brain tissue.
MoGen has been released as an open-source model, including the species-specific trained variants, as a resource for the broader connectomics research community.
The post also describes an ongoing effort to use simulated neurons to create synthetic electron microscope images, which would provide additional training data earlier in the reconstruction pipeline — before the 3D reconstruction stage MoGen currently targets. That would extend the synthetic data approach to the raw imaging inputs rather than just the geometric outputs.
The Google Research Connectomics team frames this as one of several foundational tools built over more than a decade of collaborative brain science research. The immediate result — a measurable error reduction in the current state-of-the-art reconstruction model — is described as the first time a modern generative AI approach has advanced the best available connectomics method. The scale translation from a 4.4% metric improvement to 157 person-years saved is the post’s clearest statement of why that matters.