Tera-MIND: Tera-scale mouse brain simulation via spatial mRNA-guided diffusion
Jiqing Wu, Ingrid Berg, Yawei Li, Ender Konukoglu, Viktor H. Koelzer · 2025
Read Tera-MIND: Tera-scale mouse brain simulation via spatial mRNA-guided diffusion by Jiqing Wu, Ingrid Berg, Yawei Li, Ender Konukoglu, Viktor H. Koelzer (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-01220.txt.
Introduction
Jiqing Wu and co-authors present Tera-MIND, a generative system that simulates a whole mouse brain in 3D, in a 2025 paper. New tissue profiling tools now map mammalian brains cell by cell, with gene activity pinned to place. These tera-scale atlases are too vast for easy modeling of brain structure in its native layout.
Tera-MIND uses a patch-based, boundary-aware diffusion model, guided by spatial gene expression data. It generates virtual mouse brains with detailed cell shapes at teravoxel scale. A gene-to-gene attention view also reveals spatial interactions in key pathways, such as the glutamate and dopamine systems. The method even transfers to human brain samples, pointing toward whole-organ simulation for biomedical research.
Catalog record
| CREATOR | Jiqing Wu, Ingrid Berg, Yawei Li, Ender Konukoglu, Viktor H. Koelzer |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 41 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-01220
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Tera-MIND: Tera-scale mouse brain simulation via spatial mRNA-guided diffusion Jiqing Wu, Ingrid Berg, Yawei Li, Ender Konukoglu, Viktor H. Koelzer Abstract Holistic 3D modeling of molecularly defined brain structures is crucial for understanding complex brain functions. Emerging tissue profiling technologies enable the construction of a comprehensive atlas of the mammalian brain with sub-cellular resolution and spatially resolved gene expression data. However, such tera-scale volumetric datasets present significant computational challenges in understanding complex brain functions within their native 3D spatial context. Here, we propose the novel generative approach Tera-MIND , which can simulate Tera -scale M ouse bra IN s in 3D using a patch-based and boundary-aware D iffusion model. Taking spatial transcriptomic data as the conditional input, we generate virtual mouse brains with comprehensive cellular morphological detail at teravoxel scale. Through the lens of 3D gene - gene self-attention, we identify spatial molecular interactions for key transcriptomic pathways in the murine brain, exemplified by glutamatergic and dopaminergic neuronal systems. Importantly, these in-silico biological findings are consistent and reproducible across three tera-scale virtual mouse brains. Therefore, Tera-MIND showcases a promising path toward efficient and generative simulations of whole organ systems for biomedical research. Project website: https://musikisomorphie.github.io/Tera-MIND.html 1 Introduction Recently, a network of researchers [ 47 , 49 , 21 , 35 ] has reached a groundbreaking milestone by charting the first complete cellular atlas of the adult mouse brain, the most extensively studied animal model in neuroscience. Using advanced spatial transcriptomics methods [ 3 , 25 , 5 ] capable of profiling molecularly defined brain anatomy and morphology at sub-cellular resolution, the collective studies established foundational resources and datasets, providing an unprecedented framework for systematically exploring the molecular-driven spatial organization of the mammalian brain. Consequently, this landmark achievement opens the door for next-generation investigations into intricate brain functions characterized by exceptional spatial complexity. However, these tera-scale volumetric datasets, including spatial molecular data ( e.g., mRNA readouts acting as proxies for gene expression) and paired morphological bioimages, pose substantial computational challenges [ 45 ] for in-silico modeling of the mouse brain. Integrating spatial gene expression with 3D brain morphology to uncover functional relationships at whole organ scale remains an open challenge. Meanwhile, generative artificial intelligence (GenAI) has emerged as a transformative tool for automating visual content creation and manipulation [ 4 , 48 ] . Given a text prompt that describes the desired visual content, GenAI models [ 14 , 15 ] are able to produce highly realistic images and videos [ 32 , 30 ] with remarkable efficiency. In biomedical research, the applications of GenAI have gained domain-specific recognition in various areas such as digital histopathology [ 2 ] , drug screening [ 43 , 20 ] , and molecule design [ 4 ] . However, the direct application of existing GenAI methods to whole organ simulation has currently remained infeasible due to technological limitations. Developed upon the framework of Generative Adversarial Nets (GAN) [ 14 ] , the GigaGAN [ 19 ] model was trained on 96-128 A100 GPUs to synthesize 4,096\times 4,096 color images. Using up to 6,144 H100 GPUs, Movie Gen [ 31 ] , a diffusion-based model [ 15 ] , achieved video generation results with 1,092\times 1,080\times 256 spatial-temporal resolution. In comparison, even after image registration and quality control, the mouse brain atlas remains at tera-scale volume, consisting of 100 whole slide images (WSIs; 0.108\mu m/\mathrm{pixel} ) paired with size-matched gene expression arrays, each with a spatial resolution of 73,216\times 105,984 . This immense data scale of WSIs ( e.g., 73,216\times 105,984\times 100=0.77\times 10^{12} voxels) as well as gene expression arrays significantly amplifies the computational demands associated with GenAI modeling of a virtual mouse brain. Addressing memory bottlenecks in high-resolution image generation has previously motivated the development of patch-based approaches [ 22 , 23 , 46 , 28 , 12 ] , which enforce boundary consistency [ 45 ] between generated neighboring patches. Similar to many standard GenAI models, these approaches can also leverage the interplay between textual and visual modalities and create desired visual content conditioned on text prompts. While this design choice has shown remarkable success in natural image synthesis and artistic content creation, it is suboptimal for biomedical applications – particularly for the accurate 3D molecular-driven reconstruction of mammalian brains. The inherent complexity of brain morphology controlled by spatial molecular interactions demands a generative framework that can capture molecular-to-morphology spatial associations with precision. In response to these challenges, we propose Tera-MIND , a novel GenAI approach designed for the paired data structure of spatial mRNA arrays (gene expression, prompts) and WSIs (brain morphology, bioimages). Specifically, Tera-MIND employs a newly introduced patch-based and boundary-aware diffusion model, which allows the reconstruction of tera-scale mouse brains with high fidelity. Owing to the patch-wise training and inference paradigm, Tera-MIND models in-silico mammalian brain(s) with computational efficiency. As a result, the whole simulation process can be efficiently executed on a single DGX A100 machine, significantly lowering hardware demands while maintaining scalability for tera-scale data processing. This efficiency underscores the potential wide applicability of Tera-MIND for biomedical applications. Our contributions can be summarized as follows: • By conditioning on 3D spatial gene expression as the input prompt, Tera-MIND enables the seamless generation of in-silico mouse brains at the scale of 0.77 teravoxels. Moreover, our approach accurately preserves fine-grained morphological structures across cellular, tissue, and slice-wise scales, supporting detailed exploration and comparison of brain architecture. • Leveraging 3D gene - gene self-attention mechanisms, we quantify and visualize spatial molecular interactions of key pathways that contribute to fundamental brain functions, including those involved in glutamatergic ( Slc17a6 and Slc17a7 ) and dopaminergic ( Nr4a2 and Th ) neuronal systems. • Lastly, the in-silico biological findings obtained by Tera-MIND can be efficiently reproduced across three tera-scale virtual mouse brains. This confirms the robustness and reliability of Tera-MIND for biomedical applications. Figure 1: Overview of the Tera-MIND architecture and workflows for virtual organ simulation (Main result) . a . The 3D reference map of major brain regions (left) alongside their corresponding labels (right) for guiding the navigation of brain structures. b . The 3D visualization of the ground-truth (GT, left) and generated (right) mouse brain, as captured by a sequential stack of DAPI (4’,6-diamidino-2-phenylindole) and PolyT-stained (total mRNA signals) WSIs at the scale of 0.77\times 10^{12} voxels. The bottom part of generated and GT results overlay brain region maps, aiding in the navigation of the complex spatial brain architecture. c . The conceptual illustration of patch-based diffusion model training. Here, we train Tera-MIND with noisy DAPI and PolyT bioimage patches so that the model learns to output clean patches guided by paired spatial mRNA array. Apart from the standard reconstruction path (black arrows), we supply the boundary-aware path (gray arrows) to impose boundary consistency between generated neighboring patches. d . The conceptual illustration of patch-wise generating virtual mouse brain using Tera-MIND. At this stage, we only run through the boundary-aware path to seamlessly generate the tera-scale mouse brain. 2 Results 2.1 Model overview of Tera-MIND To resolve the high-volume computational demand, we develop a novel patch-based and boundary-aware diffusion model Tera-MIND , designed for the scalable and seamless generation of teravoxel mouse brains (See Fig. 1 (a, b)). Considering the spatial linkage of