Tera-MIND

Tera-MIND (Tera-scale Mouse brain simulation via spatial mRNA-guided Diffusion) is a generative framework for synthesizing complete virtual mouse brains in three dimensions, with cellular morphological detail rendered across roughly 0.77 × 10¹² voxels — a teravoxel-scale volume far beyond what conventional image-generation models can produce in one pass. Rather than treating image synthesis and molecular biology as separate problems, Tera-MIND conditions the entire generation process on spatially resolved gene expression, so that the simulated tissue morphology is driven by the underlying transcriptomic state of each region.

The model was developed by Jiqing Wu, Ingrid Berg, Yawei Li, Ender Konukoglu, and Viktor H. Koelzer at the University of Zurich and ETH Zurich, and first released on arXiv in March 2025. It addresses a practical bottleneck in computational neuroscience and spatial omics: real whole-brain, single-cell spatial transcriptomic atlases are enormously expensive to generate and are limited to the specimens that have actually been sectioned and imaged. Tera-MIND offers a way to simulate morphologically coherent brain volumes directly from spatial expression maps, and to interrogate where and how genes co-vary across the tissue.

A key reason the work stands out is that generating a full brain at teravoxel resolution is infeasible for a monolithic 3D model. Tera-MIND instead uses a patch-based, boundary-aware diffusion scheme that generates the volume piece by piece while keeping adjacent patches seamlessly consistent, then assembles them into a coherent whole-brain reconstruction.

#Key Features

• Tera-scale 3D generation: Produces virtual mouse brains spanning roughly 0.77 × 10¹² voxels with cellular-level morphological detail, a scale that requires generating and stitching the volume patch by patch rather than all at once.
• Spatial mRNA conditioning: Spatially resolved gene expression is the conditional input, so the synthesized tissue morphology is guided by the local transcriptomic state rather than generated unconditionally.
• Boundary-aware diffusion: A patch-based diffusion design enforces consistency across patch boundaries, avoiding the seams and discontinuities that naive tiling would create in a large 3D volume.
• 3D gene-gene self-attention: Self-attention over genes lets the model analyze spatial molecular interactions for key transcriptomic programs, including glutamatergic and dopaminergic neuronal systems, and visualize where those interactions concentrate.
• Zero-shot human transfer: A checkpoint trained on mouse data was applied to previously unseen human brain samples, suggesting the learned morphology-from-expression mapping generalizes beyond the training species.

#Technical Details

Tera-MIND is a denoising diffusion model trained on whole mouse brain imaging data from the Brain Image Library, paired with spatially resolved transcriptomic expression as the conditioning signal; the released expression panel covers a large single-cell spatial atlas (distributed on Zenodo as patchwise gene expression for inference). The architecture combines patch-based, boundary-aware diffusion for the volumetric synthesis with a 3D gene-gene self-attention mechanism that exposes spatial molecular interactions. Training was performed on 2 × A100 40GB GPUs; region-of-interest generation at 8092 × 8092 uses 8 × A100 GPUs, and generating a full brain on a single A100 DGX node takes roughly one week. The authors release a pretrained checkpoint and MIT-licensed inference code via a download-checkpoint-then-infer workflow, with example outputs provided in whole-slide-image format viewable in QuPath.

#Applications

Tera-MIND is aimed at computational neuroscientists, spatial-omics researchers, and developers of brain atlases who need morphologically realistic 3D tissue that is explicitly linked to gene expression. Because generation is conditioned on spatial transcriptomics, the model can be used to study how molecular programs — such as glutamatergic and dopaminergic signaling — are spatially organized, to visualize gene-gene interaction maps across the volume, and to produce large synthetic datasets for method development where real whole-brain spatial atlases are scarce. Its demonstrated zero-shot application to human brain samples hints at translational use for relating molecular state to tissue structure across species.

#Impact

Tera-MIND is, to the authors' knowledge, the first framework to simulate brains at teravoxel scale in 3D while conditioning on spatial mRNA expression, pushing generative modeling well past the volumes typical of medical image synthesis. By coupling morphology to transcriptomics and surfacing interpretable gene-gene attention maps, it offers a new way to probe spatial molecular biology in silico rather than only observing it in fixed tissue. Practical adoption is constrained by substantial compute requirements — full-scale inference needs multiple A100 GPUs — and the weights are distributed alongside MIT-licensed code without a separately stated weights license. As a preprint with open code and a pretrained checkpoint, it provides a concrete, reproducible foundation for further work on expression-conditioned 3D tissue generation.