BrainBeacon: A Cross-Species Foundation Model for Single-cell Resolved Brain Spatial Transcriptomics

Chengming Zhang, Yiwen Yang, Yifeng Jiao, Qianqian Yang, Xin Guo, Jing Xu, Jiyang Li, Yuanyuan Zhou, Zhen Liu, Yushuai Wu, Yu Zhang, Jiang Chen, Jianyun Ma, Yuxuan Liu, Yao Zhang, Kai Xiao, Yidi Sun, Cirong Liu, Lei Han, Qi Yuan, Kazuyuki Aihara, Chengyu Li, Yuan Cheng, Luonan Chen, Wei WubioRxiv 2025 · DOI

TL;DR

Researchers have developed BrainBeacon, a novel foundation model for spatial transcriptomics that integrates data from multiple species and platforms to create a unified understanding of brain cellular architecture. This model enables accurate cross-species cell type annotation, spatial mapping, and even allows for in silico experiments to study gene function and aging.

What Problem Does This Paper Address?

Understanding the brain's complexity requires mapping its intricate molecular, cellular, and circuit organization. Spatial transcriptomics is a powerful tool for this, providing cell-type-specific expression patterns within their spatial context. However, existing methods struggle to integrate data from different species and technologies, leading to a lack of a unified, cross-species model for whole-brain analysis. This gap hinders our ability to compare brain structures across evolution, study aging, and understand disease mechanisms comprehensively.

Methods

BrainBeacon is built upon a Transformer-based architecture with a two-stage training strategy:

  1. Intra-cell modeling (Gene-level Transformer): This stage encodes each cell's transcriptomic profile into a sequence of tokens, including gene expression ranks, spatial density, species, and platform information. A Transformer encoder learns molecular representations by modeling gene-gene dependencies within a cell, using a masked gene prediction objective.
  2. Inter-cell modeling (Cell-level Transformer): This stage takes the cell embeddings from the first stage and models spatial relationships among cells within a tissue slice. It uses a variational autoencoding framework to reconstruct gene expression profiles of masked cells, enabling context-aware inference and uncertainty estimation.

The model also incorporates unified gene representation and alignment by standardizing gene nomenclature and using homology graphs and protein embeddings to learn cross-species gene relationships.

Datasets

BrainBeacon was trained on a massive BrainST-133M dataset, comprising 133 million spatially resolved cells from:

  • Four species: Human, macaque, marmoset, and mouse.
  • Five spatial transcriptomics platforms: MERFISH, Xenium, STARmap, Slide-seqV2, and Stereo-seq.

The model was evaluated on several publicly available datasets, including:

  • Heffel2024 (human hippocampus and cortex)
  • Fang2022 (human cortex and mouse cortex)
  • Allen2022-Aging (mouse cortex and striatum)
  • Allen2022-LPS (mouse cortex and striatum)
  • Ma2024-Aging (mouse brain)

Key Insights and Ideas

  • Unified Cross-Species Representation: BrainBeacon successfully integrates diverse spatial transcriptomics data, creating a single framework that can generalize across species and platforms. This is a significant step towards a universal brain atlas.
  • Spatial Context is Crucial: The model's ability to capture both intra-cellular gene dependencies and inter-cellular spatial interactions is key to its success, allowing it to learn biologically meaningful representations of brain structure.
  • In Silico Perturbation for Causal Inference: The novel "cell2niche" perturbation framework allows researchers to simulate gene knockouts and study their effects on cellular states and their microenvironments, opening doors for mechanistic studies of gene function and disease.
  • Aging Mechanisms Uncovered: BrainBeacon's analysis of aging-associated genes revealed their involvement in neuronal plasticity, cell survival, and mitochondrial metabolism, providing insights into the molecular shifts during aging.

Key Results

  • State-of-the-Art Performance: BrainBeacon consistently outperforms existing methods in unsupervised clustering tasks, achieving high Adjusted Rand Index (ARI) scores (up to 0.79) and Silhouette Width (ASW).
  • Accurate Cross-Species Annotation: The model achieves remarkable accuracy (>98% at the class level, ~87% at the subclass level) in transferring cell type labels across species and platforms in a zero-shot setting.
  • Biologically Meaningful Embeddings: BrainBeacon's learned representations accurately identify hippocampal regions and neuronal subtypes, reflecting known anatomical organization and evolutionary relationships.
  • Effective Aging Analysis: The model successfully identified key genes involved in aging and demonstrated that perturbations of these genes can shift aged cells towards a younger state, highlighting potential targets for intervention.

Limitations and Future Work

  • Transcriptomic Input Only: Currently, BrainBeacon primarily operates on transcriptomic data. Integrating other modalities like chromatin accessibility, protein expression, or histological imaging could further enhance its resolution and interpretability.
  • Coupling with Histology: The authors suggest that coupling spatial transcriptomics with co-registered histological sections could facilitate joint modeling of molecular states and tissue architecture for high-fidelity reconstruction of cell-tissue interactions.
  • Extensibility: The framework is designed to be generalizable and extensible, paving the way for future research in comparative neuroscience, brain atlas construction, and translational studies in aging and neurodegenerative diseases.

Summary updated

Abstract

Abstract The brain’s functional complexity emerges from spatiotemporal heterogeneity in molecular, cellular, and circuit organization, shaped by evolution, aging, and disease. Single-cell spatial transcriptomics provides a transformative framework to decode this complexity, resolving cell-type-specific expression patterns, region-specific regulation, evolutionary divergence, and disease-associated disruptions. Despite recent advances in spatial transcriptomics, a unified foundation model capturing cross-species, whole-brain cellular architecture remains lacking. Here, we introduce BrainBeacon, the first cross-species foundation model for whole-brain spatial transcriptomics. Trained on 133 million spatially resolved cells covering a total area of 210194 mm 2 from human, macaque, marmoset and mouse brains profiled across five major platforms, BrainBeacon integrates gene expression rankings, spatial cellular organization, and evolutionarily conserved genetic relationships into a unified transformer-based architecture. A two-stage training strategy captures both intra-cellular transcriptional dependencies and inter-cellular spatial interactions, yielding biologically grounded and interpretable representations of brain structure. Subsequently, BrainBeacon enables cross-species alignment of brain cell types and anatomical regions, facilitating integrative analyses of conserved and divergent cellular architectures. Moreover, it uncovers spatial regulatory mechanisms of aging by predicting bidirectional perturbation responses between cells and their local niche. Together, BrainBeacon serves as a foundational cross-species model for whole-brain spatial transcriptomics, enabling virtual brain atlas construction, systematic in silico perturbation studies, and mechanistic insights into brain disease pathogenesis.
BibTeX
@misc{zhang2025brainbeacon,
  author = {Chengming Zhang and Yiwen Yang and Yifeng Jiao and Qianqian Yang and Xin Guo and Jing Xu and Jiyang Li and Yuanyuan Zhou and Zhen Liu and Yushuai Wu and Yu Zhang and Jiang Chen and Jianyun Ma and Yuxuan Liu and Yao Zhang and Kai Xiao and Yidi Sun and Cirong Liu and Lei Han and Qi Yuan and Kazuyuki Aihara and Chengyu Li and Yuan Cheng and Luonan Chen and Wei Wu},
  title = {{BrainBeacon: A Cross-Species Foundation Model for Single-cell Resolved Brain Spatial Transcriptomics}},
  year = {2025},
  doi = {10.1101/2025.07.08.663729},
  howpublished = {bioRxiv},
}
Figure