Biology remains, by researchers’ own admission, the most complex system that humanity has attempted to understand. Despite decades of accumulating data and building knowledge bases, medicine still does not have a robust predictive framework to rationally anticipate the effect of a treatment or the evolution of a disease. This “incomplete map” of life translates into a simple figure: almost 90% of drugs that reach clinical trials fail, after years of development and hundreds of millions invested.
Artificial intelligence has transformed biology in a piecemeal manner. AlphaFold, developed by Google DeepMind and crowned with the 2024 Nobel Prize in Chemistry, was a game-changer by solving a problem that had remained open for decades, by predicting the three-dimensional structure of proteins from their amino acid sequence. But if AlphaFold has revolutionized structural biology, it only covers a fragment of the overall problem. Understanding life involves linking genes, proteins, cells, tissues, organs, the environment and the medical history of each patient.
This is precisely the ambition of Bioptimus, a Parisian startup which positions itself in a territory that is both scientific and industrial and wants to build a multi-scale and multi-modal foundation model for biology, a “GPT of life” which generates biological simulations.
From diagnostic AI to the foundation model for all biology
In one year, the company has already brought to market H-Optimus-0, an open source foundation model for pathology, presented as the largest AI model dedicated to this discipline. Independently, teams from Harvard and the University of Leeds have shown that it outperforms existing models on several tasks: prediction of gene expression from tissue morphology, subtyping of ovarian cancers, detection of biomarkers making it possible to better estimate the response to certain treatments.
H-Optimus-0 has already been downloaded more than 100,000 times, which, in a field as specialized as histopathology, is more than satisfactory. For Bioptimus, this model is only the first step, the roadmap provides for the release of a first multimodal model capable of linking genetic data and imaging, with direct applications in oncology, in particular for the detection of mutations causing cancer and the improvement of patient stratification.
In the longer term, the objective is to build digital twins, digital twins of patients, integrating their genome, their cellular and tissue data, their clinical history and, ultimately, environmental and lifestyle factors. These digital twins would be used to simulate the effect of treatments even before administration, with the hope of drastically reducing failures in the clinic and personalizing therapeutic strategies.
Three critical resources: data, compute and team
To claim to model biology at all its scales, Bioptimus had to secure three assets that have become strategic.
First, data, because unlike text on the Internet, the amount of relevant, good-quality biological data deep enough to train foundation models is limited, often proprietary, and rarely public. Some of the data sets are purchased at a high price, sometimes several million for a few thousand patients, in a market that is still poorly stabilized where a base paid for today can be a year later, an open standard or greatly depreciated. The startup combines public data, hospital partnerships and access to patient data from Owkin, which incubated it, to power its models on concrete clinical cases.
Another necessary asset is computing power. In just under a year, Bioptimus has raised the equivalent of 65 million euros, a significant portion of which is invested in GPU clusters and dedicated cloud infrastructures. A large part of the capital raised is absorbed by the training of these massive models, capable of integrating several modalities and operating on heterogeneous volumes of data.
Finally, the team. Bioptimus has brought together profiles from Google DeepMind, Owkin and leading academic laboratories. In the field, this translates into teams where machine learning specialists, pathologists, biologists, HPC engineers and data scientists coexist, each with extreme depth on a specific segment of biology or AI.
Data, ethics and diversity: a complexity more political than technical
Beyond the volume of data, Bioptimus must deal with strong qualitative and ethical constraints. Not all data is equal, a pathology model can over-interpret artifacts linked to a type of scanner, a slide preparation protocol or report writing habits of certain pathologists. The diversity sought is therefore not only demographic (gender, origin, age), but also technological (types of scanners), organizational (types of hospitals and practices) and geographical.
Managing these potential biases is central, firstly for reasons of scientific performance, then for regulatory reasons and acceptability of models by authorities and clinicians. It is also in this context that the establishment of a Scientific Advisory Board is made up of leading researchers in biology, medical sciences and machine learning, responsible for guiding scientific choices, opening doors on the pharma and data provider side, and challenging model architectures under construction, including Sarah Teichmann, PhD, Chair of the SAB (University of Cambridge), Andrea Califano, Dr. (Columbia University, CZ Biohub NYC) and Caroline Uhler, PhD (MIT, Broad Institute).
Capital, competition and product trajectory
Competitively, Bioptimus operates in a landscape where several players pursue a similar ambition, sometimes with much greater financial resources, like Isomorphic Labs (Alphabet) or XtalPi in China.
In January 2025, the startup completed a fundraising 41 million euros in Series A led by Cathay Innovation, with the participation of Bpifrance (via its Large Venture fund), Sofinnova Partners, Andera Partners, Hitachi Ventures, Sunrise, Boom Capital Ventures and Pomifer Capital, bringing its cumulative financing to the equivalent of approximately $76 million. Bioptimus claims a disciplined positioning in the allocation of capital: choice of targeted pathologies, arbitration between early commercialization and pursuit of a long-term roadmap, selection of data sets that are truly differentiating rather than “spectacular” but quickly trivialized. This capital should make it possible to accelerate the development of the first commercial multimodal model, to expand access to rare data and to pursue the objective, within five to ten years, of building digital twins capable of simulating the biology of each patient, from microscopic to macroscopic.
Behind Bioptimus is a collective of elite scientists, engineers and entrepreneurs, headed by Jean-Philippe Vert, co-founder and CEO, former research director at Owkin and former research manager at Google Brain. He has a very high academic level (École Polytechnique, PhD in mathematics, more than 190 publications) and solid experience in AI for life sciences.
Alongside him are several co-founders from the ranks of Google DeepMind and Owkin, including Rodolphe Jenatton (Chief Technology Officer), Zelda Mariet (VP Research), Felipe Llinares-López (VP AI), as well as David Cahané and Eric Durand, who bring cutting-edge scientific skills in machine learning, AI applied to biology, bioinformatics and biomedical data modeling. This founding foundation is complemented by operational and strategic leadership with Mathilda Ström, appointed Chief Operating Officer in November 2024.
