Bioptimus Launches Multimodal Tissue Atlas Across Three Continents

Bioptimus launched the Spatial Tissue Embedding Learning Atlas (STELA), a large-scale, clinically linked multimodal dataset spanning oncology and immunology. The global initiative, spanning data from patients from three continents, was developed with 10x Genomics and Broad Clinical Laboratories. 

Bioptimus is a Paris-based company building large AI models trained on biological data, combining different types of data such as tissue images, genomics, and clinical records into a single system. Its models, including H-Optimus for histology and the upcoming M-Optimus for broader biology, are trained on datasets spanning billions of images and contributions from thousands of clinical practices, with the goal of helping researchers and clinicians better understand disease and guide treatment decisions.

Dataset Design and Scale

STELA is designed to profile up to 100,000 patient tissue specimens, combining spatial transcriptomics with histopathology imaging, multi-omics layers such as genomics and proteomics, and longitudinal clinical records. The dataset spans sites in the United States, Europe, and Asia, and represents an estimated order-of-magnitude scale increase compared to existing spatial biology atlases.

Participating hospitals and research centers contribute samples under standardized protocols and receive access to resulting datasets and modeling tools. This aligns data generation, processing, and model development within a unified framework, reducing variability across institutions.

In oncology and immunology, spatial profiling is increasingly valuable because it captures immune niches, immune-cell organization, and local cell-cell interactions that bulk and dissociated assays miss. Reviews in translational immunology and oncology argue that these spatially organized immune states are directly relevant to tumor evolution, immune evasion, biomarker discovery, clinical outcomes, and therapy response.

Technology Stack and Infrastructure Partners

The program is standardized on 10x Genomics’ Xenium spatial transcriptomics platform, and is a coordinated effort to produce reproducible, AI-compatible spatial biology data across multiple institutions, to support foundation models that connect molecular biology to patient outcomes. 

10x Genomics is a biotechnology company that develops tools for analyzing biology at high resolution, including single-cell and spatial technologies that show how genes are expressed within tissues. Its instruments, reagents, and software are widely used in research across areas like cancer, immunology, and neuroscience.

10x Genomics’ Xenium is designed specifically for spatial transcriptomics, allowing researchers to measure where genes are active within cells and tissues while preserving their exact location. It can detect hundreds to thousands of RNA molecules at subcellular resolution, making it possible to map gene expression patterns directly onto tissue structure. 

Broad Clinical Laboratories and the Broad Institute provide the lab capacity to process large numbers of samples. Their multiyear Bioptimus collaboration also includes joint development of AI-based quality control metrics and workflow optimization tools intended to improve data consistency and assay performance at scale.

Broad Clinical Laboratories is a subsidiary of the Broad Institute that provides large-scale genomic and multi-omics data generation and analysis for research and clinical use. Operating under CLIA and CAP standards, it supports projects ranging from basic research to clinical diagnostics, with experience reportedly spanning hundreds of thousands of sequenced genomes.

From Atlas to Foundation Model

STELA dataset is intended to serve as the foundation for Bioptimus’ M-Optimus system, a multimodal foundation model designed to link molecular and cellular features of tissues to disease progression and treatment response, allowing to simulate and analyze experiments in silico. 

See also: What are Foundation Models in Biology and Healthcare?

Currently, M-Optimus is available via early access, with the formal launch expected sooner this month.

STELA is entering a fast-growing group of large tissue-atlas programs built around standardized spatial profiling, multimodal integration, and downstream AI use. A January 2026 review in Cancer Cell described the field as moving toward multimodal integration and spatial foundation models, and the eventual conversion of high-plex discoveries into scalable AI-enabled clinical assays, while also pointing to persistent problems in standardization and scalability.

In a similar effort, in November 2025, the Garvan Institute, University of Tokyo, and 10x Genomics also launched ASTRA, a Xenium-based pan-cancer atlas across 2,000 tumors in Asia-Pacific. Meanwhile, the Broad-Illumina Spatial Flagship Project is generating large-scale spatial data from hundreds of samples, and public programs such as NCI’s HTAN and NIH’s HuBMAP continue to define the reference standard for interoperable atlas infrastructure. 

To expand the diversity of training data for biological AI, Basecamp Research recently launched the Trillion Gene Atlas, a global effort to collect genomic data from over 100 million species to expand training datasets for biological AI beyond the limits of existing public repositories.

Topic: AI in Bio