Bioptimus, a global AI company building models for biology, announced the launch of its Spatial Tissue Embedding Learning Atlas (STELA), a multinational spatial data generation initiative in partnership with 10x Genomics and Broad Clinical Labs. The company aims to solve issues in drug discovery and development, improving efficiency and decreasing costs.
Credit: Bioptimus
Building foundation models for biology requires data at a scale the field has never had. Unlike information scraped from the internet, patient data resides in hospitals, fragmented across institutions and governed by strict regulatory constraints. Bioptimus is now closing this gap by building the data infrastructure necessary to power M-Optimus, a multimodal and multiscale world model of biology.
“The main pain point to achieve this vision is to access data in large volumes, in large diversity, ethnicity, in terms of modalities that would feed the AI,” said Jean-Philippe Vert, CEO of Bioptimus. “We started by already making deals and partnerships with several companies to have access to proprietary data sets that already allowed us to build systems and commercialize AI systems that are already very good for some specific topics.”
The data problem biology hasn’t solved
STELA will serve as the backbone for M-Optimus, generating massive datasets designed to decrypt the complex organization of human tissues. The initiative aims to profile up to 100,000 patient tissue specimens, starting in oncology and immunology. This could represent a 20-fold increase over the largest spatial biology atlases that exist today. M-Optimus will leverage the dataset to map molecular and cellular interactions.
“We are creating a model that learns biology, not through a single entry point, a single modality, but by trying to process information that comes from the clinics, from images, from genomics, from DNA, from proteins and more,” said Vert. The company unveiled M-Optimus in December and is currently offering early access to a select group of pharma partners.
The model integrates histology images, RNA sequencing, spatial transcriptomics, and clinical data into a unified framework — the first of its kind to combine multiple biological modalities at this scale, the company said. Bioptimus’ models are already in use by 16 of the top 20 pharmaceutical companies.
The model can predict clinical outcomes and key gene expression markers and mutations, as well as infer cell composition. With STELA, the model will allow researchers to anticipate patient responses to novel therapies, accelerate drug development and design more effective immunotherapies.
From drug discovery to clinical trials
Beyond drug discovery, M-Optimus could also help pharma companies improve design and make decisions during clinical trials. The model will analyze early-phase patient response data to identify the right target population or flag when a trial should be stopped early.
“Today, the largest cohorts in the world have a maximum of a few thousand patients. What we want to do is just say, well, now that we demonstrated there is some signal that it starts working, we just need to scale the training set — to add volume, to add diversity, to maintain quality of data — and we will see progress in performance,” Vert said.
Now, the company is moving to expand the volume of data with the STELA initiative. Participating hospitals and research institutions will contribute samples under standardized protocols and, in return, receive access to rich spatial characterization and foundation model capabilities.
The company’s histopathology model, H-Optimus, offers a preliminary digital twin, which Vert referred to as a “digital barcode”. The model turns a patient’s biopsy image into a digital representation that can predict treatment response, gene expression and other outcomes.
“Today, most patients’ diagnostic data is used to inform decisions for only that individual. We envision a world where every patient can contribute insights to better inform the care and treatment outcomes of future patients,” Vert said.
