Bioptimus's first multimodal, multi-scale World Model for biology, integrating histology, transcriptomics, and clinical data from millions of patients into a unified embedding space.
M-Optimus is Bioptimus's first "World Model for biology" — a multimodal, multi-scale foundation model that integrates disparate biological measurement types into a single unified embedding space. Announced in December 2025, it extends the company's earlier work on histopathology foundation models (the H-Optimus series) toward a broader goal: learning a shared representation that connects what a tissue looks like, what genes it expresses, and how a patient ultimately responds to treatment. Rather than treating histology, sequencing, and clinical records as separate analyses, M-Optimus is designed to reason across them jointly.
The model addresses a structural problem in translational biology and drug development: clinical and molecular data are routinely collected in different formats, at different scales, and in different institutions, which makes it hard to learn from them together. Bioptimus reports assembling training data from multimodal patient cohorts at large scale — spanning millions of patients, more than 50 organ types, and hundreds of medical centers — using its proprietary "STELA" data engine. The modalities span Hematoxylin and Eosin (H&E) histology whole-slide images, routine and bulk RNA sequencing, spatial transcriptomics, and structured clinical data.
M-Optimus is at the announcement and early-access stage. As of the December 2025 launch there is no published paper, preprint, code repository, or public model weights; access is via a limited early-access program for select pharmaceutical and research partners. The descriptions below reflect Bioptimus's stated capabilities and should be read as company claims pending independent peer-reviewed validation.
Bioptimus describes M-Optimus as a multimodal, multi-scale foundation model built on a proprietary neural architecture, trained on multimodal patient cohorts assembled via the STELA data engine. The training corpus integrates H&E whole-slide images, routine/bulk RNA sequencing, spatial transcriptomics, and clinical data drawn from hundreds of medical centers and more than 50 organ types, at a reported scale of millions of patients. Beyond this framing, key technical specifics have not been disclosed: no parameter count, training compute, context length, modality-encoder design, or benchmark scores are public, and there is no accompanying paper or code release. Claims of accuracy or state-of-the-art performance therefore cannot yet be independently verified.
The stated applications center on oncology and translational research. M-Optimus is positioned to predict gene expression from H&E slides, stratify patients and forecast clinical outcomes such as Objective Response Rate and Progression-Free Survival, identify drug targets and mechanisms of treatment response, and support clinical trial design through inclusion strategies, biomarker identification, and indication prioritization. Bioptimus also describes generating "digital twins" of cells, tissues, and patients to run in silico trials. The primary beneficiaries are pharmaceutical companies and research groups in the early-access program, who would fine-tune the model into bespoke downstream tools for specific therapeutic programs.
If its capabilities hold up under independent evaluation, M-Optimus would represent a notable step from single-modality pathology foundation models toward integrated, patient-scale models of biology — a direction Bioptimus signaled with the H-Optimus series. CEO and co-founder Jean-Philippe Vert framed the launch as assembling "the first critical components of our journey to crack the code of biology by combining multiple modalities at scale." Its real-world influence remains to be demonstrated: there is no peer-reviewed paper, no public weights or API, and no independent benchmark, so adoption is currently limited to early-access partners. The most important caveat is that the model is announcement-stage, and its clinical and research claims await external validation before they can be relied upon for decision-making.