BioNTech became a big name in biopharma thanks to its Pfizer-partnered Covid-19 vaccine. It’s betting that artificial intelligence will be key to its future.
This week, BioNTech and its subsidiary InstaDeep hosted their first artificial intelligence summit, where executives spent two hours building the case that their own technologies and infrastructure will give them an edge in an increasingly competitive field.
The companies unveiled a new AI model for designing antibodies from scratch, the latest in a growing list of tools from academics and companies alike. It also showed how AI can be used to rapidly analyze slides of tumor tissues, make sense of mysterious genomes and identify tidbits of cancer proteins that the immune system can target.
BioNTech acquired InstaDeep for about $440 million in early 2023. But the event left many observers wondering if they missed the most obvious question: Which drug programs in BioNTech’s expansive pipeline of cancer therapies bear the fingerprints of InstaDeep’s technology?
Ryan Richardson“We do track it internally, but we don’t disclose it,” BioNTech chief strategy officer Ryan Richardson told Endpoints News in an interview Wednesday. “This is a long-term thesis.”
InstaDeep enjoys a long leash, allowing many of its 370 employees to work on fundamental problems in AI. And so far, BioNTech seems pleased with the results. Richardson said that after just a year or two of work on some projects, the tech startup’s AI models were beating BioNTech’s models that it had been developing for 10 years — although he wouldn’t discuss the details.
The company hopes that by giving the group freedom, it can attract talent that might otherwise go to the tech industry.
“The idea is to continue to be a leader in AI and continue to attract top AI talent that otherwise would perhaps not go work in biotech, but go work for Google, DeepMind, OpenAI,” InstaDeep CEO Karim Beguir told Endpoints.
Cancer vaccines and AI-generated antibodies
BioNTech has at least one program powered by AI already in clinical testing: its personalized, mRNA-based cancer vaccines.
It started designing the experimental vaccines on computers in 2011, encoding bits of cancer genes found in an individual’s tumor in hopes that it would train a patient’s immune system to attack the cancer. BioNTech began testing the concept in patients a decade ago, and has been designing the vaccine sequences with algorithms, with no human intervention required, since 2017, Richardson said.
The results look promising in clinical studies so far, and although Richardson told Endpoints there’s room for improvement, he wouldn’t say exactly how InstaDeep’s technology will factor into further refinements.
Karim BeguirInstaDeep has 45 researchers focused on inventing new algorithms, Beguir said. While text generators are largely based on autoregression models and image generators are based on diffusion models — both where vast datasets of text and photos can be scraped from the internet — the InstaDeep team has created what it calls a Bayesian Flow Network, or BFN, which Beguir believes will be more useful for scientific data that are continuous, discrete and heterogenous.
Beguir told Endpoints that BFNs are “a fundamental breakthrough in AI,” and that the technology will have applications beyond protein generation. That BFN model underpins InstaDeep’s new technology for creating proteins, dubbed ProtBFN. The company released a preprint describing the AI model last week, and on Tuesday it debuted a twist on the model for designing antibodies called AbBFN.
A researcher showed how the model could be used to generate new structures to optimize specific parts of antibodies, including the so-called CDR-H3 loop, the sticky molecular fingertip that helps an antibody grab its target. But Richardson said that BioNTech doesn’t necessarily view itself as a competitor to the wave of AI-for-antibody startups that have recently emerged, including Xaira Therapeutics, which launched with more than $1 billion earlier this year.
“We’re less excited by just applying the new, latest AI model to classic antibody technology. But when you combine cutting edge AI with novel therapeutic platforms like we have, I think that’s where you’re going to see the greatest scope for disruption,” Richardson said. That includes mRNA medicines and personalized medicines, he said.
The company emphasized how it was publishing its AI models and releasing the code for others to use. The open-source ethos is common in tech, and other AI-focused biotechs like Generate:Biomedicines have also released their protein models.
Beguir told Endpoints that getting feedback from the scientific community helps InstaDeep improve its technology and helps attract and retain talented AI researchers. But there are limits. The company’s protein modeling software is only available through non-commercial licenses. “So only BioNTech has the ability to build products and commercialize them with this,” he added.