In January last year, German biotech company BioNTech acquired African AI startup Instadeep for more than $550 million, a deal closed in July of the same year. Instadeep, whose outlet is currently the largest in Africa, has been operating under the German pharmaceutical umbrella for just over a year. Now is a good time to see how it has fared since the acquisition.
Instadeep uses advanced machine learning techniques to bring AI into business applications. Its products range from GPU-accelerated insights to self-learning decision-making systems. Prior to last year's acquisition, the Tunisian-born, Paris- and London-based enterprise AI startup raised more than $108 million from several global investors, including Google, Deutsche Bahn and BioNTech. These three strategies were also among the startup's largest partners and clients.
Specifically, the decade-old startup collaborated with BioNTech to develop an early warning system that could detect high-risk COVID-19 variants months in advance during the pandemic. Instadeep worked with Google DeepMind to create an early detection system for desert locust outbreaks in Africa. He also collaborated on a lunar project to automate rail scheduling for Deutsche Bahn, Europe's largest rail operator.
While these partnerships show several applications for Instadeep's solutions, its acquirer had a clear use case: using AI to develop therapies and vaccines for various cancers and infectious diseases, something it is now doubling down under its new owner.
Fifteen months after the completion of the BioNTech acquisition, co-founder and CEO Karim Beguir told TechCrunch in an interview that Instadeep has made significant progress on that front, even as the AI company, which continues to operate independently, still offers solutions to external clients. biotechnology.
“We are strategically aligned with BioNTech on the objectives to be pursued in biology and bioAI capabilities,” said the Instadeep boss. “But we also have room to maneuver and remain a force in AI in Africa and in general, while we continue to develop technologies that expand the frontier of innovation in other verticals such as industrial optimization.”
Increasing capabilities within biotechnology
Beguir notes that Instadeep's goal in the last year since its acquisition has been to implement AI at every step of BioNTech's process to improve existing processes.
Share an example of histology, which involves tissue analysis and the visible task of labeling different tissues, such as identifying tumor cells or healthy cells. According to him, BioNTech experts traditionally performed this work manually. However, Instadeep's technology has helped speed up the process by implementing visible AI and segmentation systems, speeding up this tissue labeling workflow by 5x.
Another is the completion of its RiboMab project, which involves mRNA-encoded antibodies that have now become part of BioNTech's toolkit as an immunotherapy company to combat cancer and other diseases. InstaDeep introduced this project on its DeepChain platform, which designs proteins and analyzes biological data, during its first collaboration in 2020.
Biotechnology involves a lot of sensitive health data. Collecting and analyzing them is one thing. Keeping them safe is another. Ask 23andMe, once heralded as a disruptor in the biotech space before becoming the victim of a massive breach that exposed the data of nearly 7 million people, half of its customer base.
Curiously, BioNTech is no stranger to these types of events. In 2020, hackers illegally accessed documents related to their COVID-19 vaccine, developed with Pfizer, by attacking the European Medicines Agency (EMA), the European medicines regulator, which evaluates medicines and vaccines. While Pfizer and BioNTech confirmed that their test systems and data remained secure, the incident highlights how vulnerable organizations, including regulatory ones, can be to cyberattacks.
As any CEO would tell you, Beguir tells me that Instadeep and BioNTech are very cautious with healthcare data, especially since the partnership is currently using AI to augment data assets, allowing them to identify precise protein sequences and potentially unlock new ones. targets for cancer and other immunotherapies. use cases.
But there is a segmentation in the data that both companies use. BioNTech handles personal data from real-life patients, and Instadeep typically develops models and trains them with publicly available data. This is how it trained, for example, its Nucleotide Transformer, a series of AI genomics models, which today is the most downloaded and fashionable AI genomics model in the world. [Thanks in part to this open-source deal.]
“Instadeep developed and trained the nucleotide model using public data,” says Beguir. “However, when we wanted to implement the model on specific use cases and real-life patient data, we did so at the BioNTech level, with all the privacy guarantees that come from its position as one of the leading biopharmaceutical players that operates under strict conditions. regulations and following rigorous quality protocols.
Develop new technologies within BioNTech and outside biotechnology
When asked what Instadeep's next milestones are within BioNTech, Beguir mentions the startup's “latest breakthrough”: Bayesian Flow Networks (BFN), a new generative AI model for proteins that significantly outperforms autoregressive models. and diffusion, according to the company. BioNTech CEO Ugur Sahin, in a statement, describes it as a “next-generation technology.”
According to Beguir, the model produces the purest looking and best performing protein proteins on the market by allowing the systems to look for specific properties in the heavy chain of an antibody, including chemical characteristics, hydrophobicity or sequence length. These models are crucial for understanding complex protein functions and designing new therapeutic proteins.
“We are excited about the potential for AI innovations like ours to identify real-world use cases, collaborate closely with BioNTech, and create products that will be tested in labs and clinics and ultimately save patients' lives,” he said. Beguir. “If we consider where we are today in biology and artificial intelligence, it is comparable to where we were with pure language processing in 2020 with GPT-3. The systems were starting to work and their capabilities were impressive, but there was still room for improvement.”
Instadeep launched the new AI model last week alongside a new near-exascale supercomputer that the companies say puts the partnership in the top 100 in computing and infrastructure and in the top 20 in H100 GPU clusters globally.
Both developments highlight where Instadeep, under BioNTech, deploys AI in various life sciences use cases. On the other hand, it independently manages its other line of business, which involves AI and deep reinforcement learning for industrial optimization.
One example is its 12-year project to automate rail planning and dispatch for Deutsche Bahn, one of its long-standing partners and Europe's largest rail operator. Comparably, the Tunisia- and London-based AI company has stepped up its efforts to develop other industrial optimization use cases, such as collaborating with Fraport in Germany to optimize complex airport operations with AI.
“Overall, we also see the potential of AI agents as very attractive for the future. We believe that industrial optimization and agent-based systems, working hand-in-hand with human colleagues, will revolutionize industrial efficiency. So this is also another area in which we have been working for many years and in which we will continue to invest,” said Beguir.
Meanwhile, Instadeep, earlier this month, launched in San Francisco the professional version of its DeepPCB (Deep Printed Circuit Board) product, a fully assisted hardware or printed circuit board design with autonomous AI powered by reinforcement learning. Beguir says the company's competitors are small AI startups in specific areas it operates in, such as Riyadh-based Intelmatix.
The head of Instadeep is proud of his company's work to solve more complex AI use cases (e.g. Gen AI for DNA or proteomics or agent workflows for combinatorial optimization) and move away from simple cases like Gen AI for NLP. He says that, in addition to the BioNTech acquisition, this technology plays an important role in driving customer interest in the US, where the AI company now has two offices, and also throughout Europe: Berlin, Paris and the United Kingdom, in particular.
Although BioNTech spent $500 million on Instadeep to boost its biotech capabilities, it keeps the AI company operationally independent for reasons like this, while funding its activities to serve customers beyond the biotech industry.”
“Because we bring value by being leaders in AI, and AI skills can be improved across multiple sectors,” Beguir responded when asked why BioNTech still allows the AI company to work on non-biotech projects. “It's the same technology stack, so time working on AI outside of biotech is not wasted time at all. BioNTech also implements InstaDeep in tasks outside of biotechnology R&D, such as operations optimization.”
Beguir explains that while InstaDeep was not forced to sell, it was the shared vision and successful projects with BioNTech since 2019, long before the acquisition, that convinced the AI company to move forward with the deal. He believes the trust built over years of collaboration is the reason InstaDeep will remain independent under BioNTech. The key for InstaDeep now is to maintain its momentum, maintain high-quality results and continue to innovate for as long as possible.”
Since the acquisition, InstaDeep has grown to more than 400 employees worldwide. This includes its team in Africa, based in a new office in Kigali, which leads the company's geospatial intelligence work.
Initially an on-the-ground effort in partnership with Google to detect locust breeding grounds in Africa, Instadeep now uses past tag data and satellite imagery to infer with high quality and 80-85% accuracy where locust passageways will be. lobsters in the next 30 days. Beguir says InstaGeo, the company's framework that uses multispectral satellite imagery from NASA or the European Space Agency (ESA), is open source and available for other companies to develop scalable solutions across the continent.
“This is a current example of how technology and AI capabilities are having an impact. Instead of collecting samples in the field or relying on ground infrastructure, we can deliver that knowledge via satellites at scale and notify multiple governments and actors to address a growing challenge to food security, especially given global climate issues. continent”.