Paris-HQ’d AI startup Bioptimus, which develops generative AI models for biotechnology, has raised a $41m Series A less than one year after securing a $35m seed round and before commercialising its product.
The round was led by global VC Cathay Innovation and included public bank Bpifrance, French investors Sofinnova Partners and Andera Partners, as well as Hitachi Ventures, the venture branch of global tech company Hitachi. UK VC Sunrise and US VCs Boom Capital Ventures and Pomifer Capital also participated.
Bioptimus didn’t disclose whether the fundraise consisted of debt or equity, and didn’t share its valuation.
The company was co-founded by Jean-Philippe Vert, a Google alumni and the chief R&D officer of French AI unicorn Owkin, who teamed up with fellow Owkin employees David Cahané and Eric Durand. Vert and Durand still work at Owkin.
Owkin, which develops AI tools for drug discovery and precision medicine, incubated the startup and backed its seed round.
The co-founding team also includes ex-Google researchers Rodolphe Jenatton, Zelda Mariet and Felipe Llinares.
Bioptimus builds foundational AI models — similar to the large-language models developed by Paris-based Mistral AI and US tech giant OpenAI. But the startup says that instead of generating text, its models will be able to understand and simulate the complex behaviour of organisms, thanks to training datasets that range from genetic code and DNA profiles to biopsy imaging and clinical data.
ChatGPT for biology
Bioptimus says that it is developing multi-modal and multi-scale AI models. This means that the technology will include different modalities of biological data (such as imaging, genetics and DNA) as well as different scales (molecules, cells, organic tissues).
Vert says that historically research has looked at these different modalities and scales in silo — yet connecting them is key to understanding how organisms work as a whole.
“We’ve understood that we can train AI models so that, when we show them lots of data, they can understand the logic,” says Vert. “We’re using this approach to understand the logic of how genes link to a cell, or how cells function together to create organic tissue.”
The startup’s ambition, says Vert, is to create a model that would eventually enable doctors to simulate how a tiny change at the microscopic level could impact a patient’s everyday life.
Bioptimus’s roadmap
In July, Bioptimus released its first model in open source — an early version of the technology that was only trained on images of organic tissue, and which the company says was able to detect cancerous cells and abnormalities. The model hasn’t generated any revenue yet.
In 2025, says Vert, the startup will release its first multi-modal and multi-scale model, which will build on the first version and also include genetic code. “The model will learn to not only read images, but also to link between these images and the genetic activity going on in the cells that it sees on the images,” says Vert.
This model could be used to better understand how a microscopic genetic mutation could lead to a tumour visible at the macroscopic scale, says Vert.
The company then plans to keep building models that will gradually include data that is more macroscopic, such as patient reactions to a treatment, as well as microscopic, like DNA.
Bioptimus’s data acquisition strategy is crucial to the success of the company, says Vert. In addition to publicly-available data, the startup has signed partnerships with research labs to access proprietary datasets. In the future, Bioptimus will also be able to access Owkin’s patient data.
“It’s a big difference compared to most LLMs that all use the same data and have to differentiate themselves with compute power or talent,” says Vert. “We complement public data with lots of non-public data, and that creates a lot of value.”
Commercialisation
Bioptimus’s future models will not be “completely” open source, says Vert, except for academic research purposes.
The startup is planning to monetise the technology by signing paid research partnerships with pharmaceutical companies, in which Bioptimus will develop tailor-made use cases for customers — like developing a new drug or increasing the success of clinical trials.
In parallel, it will make the model available on the cloud through APIs, similar to tools like ChatGPT. This means that developers will pay to access the model and to deploy it for their own use cases.
“Our models will be useful for any actor that wants to develop AI with biomedical data,” says Vert. “So, it’s much larger than pharmaceutical research. It’s relevant to biotech, cosmetics, the food industry, anyone working in genetics or medical imagery, and so on.”
With Bioptimus planning to commercialising its 2025 model, Vert says that the company will add to its 17 staff by hiring across commercial and sales.
The startup is growing in a competitive space, with deep-pocketed Big Tech companies like Google and Microsoft increasingly taking an interest in applying AI to health and the life sciences. Through its subsidiary DeepMind, Google is behind leading AI model AlphaFold, which can predict the structure of proteins based on sequences of molecules called amino acids.
Vert says that he sees this as a positive signal. “It’s an opportunity,” he says. “Big Tech companies need to provide cloud services, and the service we provide has huge added value. It could be very complementary.”
Vert declined to specify if Bioptimus has received any offers for acquisition yet.
Read the orginal article: https://sifted.eu/articles/bioptimus-41m-series-a/
