Ctrlk
For the complete documentation index, see llms.txt. This page is also available as Markdown.

M-Optimus

M-Optimus (and its initial iteration M-Optimus-1) is a multimodal, multi-scale biological foundation model developed by Bioptimus (Paris, France). Conceived as a "world model for biology," M-Optimus integrates phenotypic morphology with molecular genetics by cross-training on paired Hematoxylin and Eosin (H&E) whole-slide histology, bulk transcriptomics (RNA-seq), and high-plex spatial transcriptomics (10x Genomics Visium and Xenium).

A breakthrough capability of M-Optimus-1 is its ability to perform in silico spatial transcriptomics, reconstructing high-dimensional spatial gene expression profiles for 6,000+ genes directly from routine H&E whole-slide images—either in H&E-only mode or guided by paired bulk RNA-seq.

Architectural Principles & Modality Fusion

Most conventional digital pathology foundation models (e.g., UNI, Virchow, Phikon, Prov-GigaPath) are strictly unimodal: they operate as vision encoders mapping $256 \times 256$ px RGB image patches into latent representation vectors. While powerful for morphological feature extraction, they cannot directly infer unmeasured molecular transcripts or spatial multi-omics.

M-Optimus bridges the scale and modality divide across three layers of biological abstraction:

┌──────────────────────────────────────────────────────────────────────────────────────────────────┐
│                                   M-OPTIMUS MULTIMODAL PYRAMID                                   │
└──────────────────────────────────────────────────────────────────────────────────────────────────┘

  Level 3: Tissue Architecture & Spatial Niches ──►  In Silico Spatial Transcriptomics (6,000+ Genes)
                                                                 ▲
                                                                 │ (Cross-Modal Attention & Decoding)
  Level 2: Cellular Phenotypes & Morphology      ──►  Gigapixel Whole-Slide H&E (Tile Embeddings)
                                                                 ▲
                                                                 │ (Biological Co-Embedding Space)
  Level 1: Molecular Profiling & Transcriptomics ──►  Bulk RNA-seq Expression Vectors

Pre-Training Foundation & Scaling

  • Histological Breadth: Trained on tens of millions of H&E tiles spanning more than 50 human tissue and tumor types.

  • Multimodal Alignment: Incorporates thousands of multi-modal clinical donor datasets where physical histology slides are aligned with both bulk transcriptomics and spatial transcriptomics arrays.

  • Inference Modes:

    1. Multimodal Inference (H&E + Bulk RNA-seq): Ingests the whole-slide H&E image alongside patient-level bulk transcriptomics (or pseudo-bulked scRNA-seq), projecting bulk molecular abundance into high-resolution spatial coordinates across the slide.

    2. Unimodal Image-Only Inference (H&E-only): Infers 6,000+ spatial gene expression maps directly from morphological features alone, unlocking archival glass slides where tissue blocks or fresh RNA are unavailable.

Key Capabilities & Clinical Applications

1. In Silico Spatial Transcriptomics (SpT)

Wet-lab spatial sequencing requires specialized hardware, expensive consumables ($1,500–$5,000/slide), and destroys clinical tissue. M-Optimus-1 generates virtual spatial expression maps at compute cost alone (~$1–$10/slide).

  • Validated on unseen stage III-B ovarian papillary serous carcinoma against physical 10x Genomics Xenium Prime ground truth, achieving $r = 0.81$ mean Pearson correlation across tumor sub-populations and $r = 0.78$ for cancer-associated fibroblasts (CAFs) and VEGFA+ tumor cells.

2. Unsupervised Spatial Niche Discovery

M-Optimus embeds multi-gene spatial co-expression into an unsupervised clustering framework:

  • Decomposes whole slides into biologically coherent microenvironmental functional units (tumor core, hypoxic invasive margin, fibrotic stroma, tertiary lymphoid structures).

  • Predicted niche cell compositions matched Xenium ground-truth cellular densities at Pearson $r = 0.899$.

3. Spatial Biomarker Discovery in Archival Cohorts

Transforms historical clinical trial archives into high-dimensional discovery search spaces:

  • In a retrospective cohort of 288 ovarian cancer WSIs (TCIA), M-Optimus-1 identified that spatial colocalization between VEGFA+ tumor cells and CAFs was significantly enriched in patients refractory to bevacizumab anti-angiogenic therapy ($p = 0.003$).

  • Hierarchical clustering of spatial interactions revealed distinct TME sub-phenotypes among non-responders, providing a rational basis for combination regimens (e.g., anti-VEGF + PARP inhibitors + CAF depletion).

Bioptimus Foundation Model Ecosystem

Bioptimus maintains an integrated suite of open and proprietary AI platforms for life sciences:

  1. H-Optimus-0: Open-weights histology foundation model (ViT-Giant, 1B parameters) trained on 500,000 whole-slide images; widely recognized as an open benchmark for tissue representation.

  2. H0-mini: An 86M-parameter compact distillation of H-Optimus-0, engineered for low-latency cell segmentation and classification; powers the CellViT backbone of Owkin's [[HistoPLUS: Towards Comprehensive Cellular Characterisation of H&E Slides]].

  3. M-Optimus-1: Multimodal world model translating H&E slides into 6,000-gene spatial transcriptomic landscapes.

  4. STELA Initiative: Global multi-institutional data coalition launched to amass multimodal clinical cohorts across 100,000 patients to train next-generation biology world models.

Related Notes & Vault References

  • Literature Clipping: [[From Model to Patient: M-Optimus-1 for Spatial Biomarker Discovery in Ovarian Cancer]]

  • Cellular & Spatial Models:

    • [[HistoPLUS: Towards Comprehensive Cellular Characterisation of H&E Slides]]

    • [[CytoFormer: A Molecularly Supervised Cell Foundation Model for Histopathology Cell Classification]]

    • [[CellQuant-Net]] / [[CellPrior-Net: Prior-Guided Nuclei Detection and Classification for H&E Whole-Slide Images]]

    • [[NuClick]]

  • Spatial Visualization Infrastructure:

    • [[Celldega]] (Broad Institute high-performance WebGL spatial-omics visualization toolkit)

  • Catalogs & Syntheses:

    • [[Digital Pathology Software]]

    • [[Articles on computational, digital, and mathematical pathology]]

    • [[Digital Pathology]]

PreviousSTAMP
NextCelldega

Last updated