TransBrain

Mouse models carry most of the mechanistic weight in neuroscience — optogenetics, viral tracing, targeted mutations — while human studies are largely observational and imaging-based. Translating between the two is normally done by hand, region by region, using homology assignments that are confident for subcortex and contentious for cortex. There has been no quantitative apparatus for asking whether a whole-brain map measured in mice corresponds to anything in humans.

TransBrain, from the Institute of Biophysics at the Chinese Academy of Sciences with University of Chinese Academy of Sciences and Beijing Normal University, builds that apparatus. It embeds human and mouse brain regions into a single latent space, so a phenotype defined on one species' atlas can be projected onto the other's by a linear operation in that space. Translation runs in both directions.

Two components do the work. The first is a deep neural network trained on integrated multimodal human transcriptomics to produce region-specific representations, which raises cortical human–mouse correspondence accuracy by 89.5% over the raw transcriptomic profiles and exposes two evolutionarily conserved cortical gradients accounting for more than half of cortical organisational variance. The second is a random-walk graph representation learning step that folds anatomical hierarchy and structural connectivity — viral tracer data in mice, diffusion tractography in humans — into a unified cross-species space.

#Key Features

• Bidirectional whole-brain mapping: A regional phenotype vector — functional gradients, structural measures, group difference maps — is translated between species by projecting through the shared embedding rather than by manual homology lookup.
• Cortical and subcortical detachment: The region-specific network is trained separately for cortex and subcortex, addressing the fact that cortical transcriptomic correspondence is far weaker than subcortical and would otherwise be swamped.
• Multiple human parcellations: Precomputed embeddings ship for the Brainnetome, Desikan-Killiany, and AAL atlases, so the framework drops into existing neuroimaging pipelines without reparcellation.
• Graph-based fusion of anatomy and connectivity: Random walks over a graph combining anatomical hierarchy with species-specific structural connectomes produce region embeddings that respect both taxonomy and wiring.
• Three worked cross-species analyses: Tutorials reproduce resting-state fMRI gradient conservation, Neurosynth annotation of mouse optogenetic circuits, and translation of autism gene-mutation effects to individual-level human imaging deviations.

#Technical Details

The region classifier is a feed-forward network with three hidden ReLU layers and dropout, trained to predict brain region identity from transcriptomic profiles; the penultimate layer supplies the region embedding used downstream. Its training data integrates microarray transcriptomics with large-scale single-nucleus RNA sequencing, and the integrated datasets are deposited on Zenodo in four parts. Graph embeddings are generated by random walks with return and in-out parameters set to 0.01 and 0.1, and the resulting pickled embeddings for the three human atlases ship inside the installed package.

The Python package is installed with pip install transbrain and supports Python 3.8–3.11 on Linux, Windows, and macOS, depending on nibabel, nilearn, scikit-learn, scipy, and anndata. Its runtime API covers atlas fetching, phenotype mapping in either direction, and visualisation; the full construction pipeline — transcriptomic preprocessing, data integration, network training, and graph walks — is published separately in the repository for groups wanting to rebuild the space on their own atlases or data. A unit test confirms installation, and a code-free web interface accepts phenotype uploads for users who do not want to write Python.

#Applications

The framework targets the translational gap in psychiatric and neurodevelopmental research. A mouse model carrying an autism-associated mutation produces a whole-brain phenotype that can now be projected onto human imaging space and compared against patient-level deviation maps, turning a qualitative claim of face validity into a quantitative one. Circuit-level manipulations only possible in rodents — optogenetic stimulation of a defined projection — can be annotated with the human cognitive terms associated with the corresponding human regions. In the other direction, a human case-control imaging finding can be projected onto the mouse atlas to identify where a mechanistic follow-up experiment should be aimed.

#Impact

TransBrain gives cross-species neuroscience a reusable, installable tool where the practice had been bespoke per-study homology mapping, and its publication in Nature Methods reflects that positioning as infrastructure rather than a single result. The most transferable finding is methodological: transcriptomic similarity alone is a weak basis for cortical homology, and learning region-discriminative representations before matching substantially improves it. The code is released under Apache 2.0, though the Nature Methods article itself sits behind a paywall rather than being open access. Its scope is set by the atlases and modalities it was built on — regional, parcellation-level phenotypes in human and mouse — and extending it to other species or to voxel-level analyses requires rerunning the construction pipeline rather than reusing the shipped embeddings.