As language models like ChatGPT and Gemini have ushered in a new age of AI in Silicon Valley, the world’s most powerful tech companies are looking ahead to drug discovery and digital biology. By Richard Nieva and Alex Knapp By Richard Nieva and Alex Knapp
As Nvidia CEO Jensen Huang scanned the audience at the JPMorgan Healthcare conference this January in San Francisco, the biggest health tech event of the year, he acknowledged that he was on unusual ground. “You’re not my normal crowd,” he said to the room of health and biology technologists, during a fireside chat with Recursion, a drug discovery firm that Nvidia pumped $50 million into last year. The audience may not have been part of his core demographic, but he’s hoping that will change. Over and over again, Huang has touted digital biology as the “next amazing revolution” in technology. As the AI boom has swept Silicon Valley, Nvidia has built a more than $60 billion a year business and last summer became one of the few companies with a market cap in the trillions. In health and biotech, it sees more opportunities to fuel its growth. “It’s been declared we are the next many-billion dollar business for Nvidia,” Kimberly Powell, Nvidia’s vice president of healthcare, told Forbes . She said the company aims to provide chips, cloud infrastructure and other tools to more biotech firms. Now that large language models like OpenAI’s ChatGPT and Google DeepMind’s Gemini have mainstreamed generative AI, several of the world’s most powerful tech companies are looking to biotech as the next frontier in artificial intelligence — a frontier where AI isn’t generating funny poems from a prompt, but rather the next life-saving drug. “It’s been declared we are the next many-billion dollar business for Nvidia.” At Nvidia, arguably a backbone of the AI revolution because of its powerful GPU chips, the bulk of investments at the company’s Nventures VC division over the past two years have been in drug discovery. At DeepMind, the Google AI lab’s AlphaFold model — a groundbreaking tool for predicting protein structures — has been used by academic researchers over the past year to develop a “molecular” syringe to inject medicine directly into cells, and to research crops that are less dependent on pesticides. The interest in biotech is industry-wide: Microsoft, Amazon and even Salesforce have protein design projects as well. While using AI in drug discovery is not exactly a new trend — DeepMind first unveiled AlphaFold in 2018 — executives at both DeepMind and Nvidia told Forbes that this is a breakthrough moment, thanks to a confluence of three things: the mass of training data now available, the explosion of computing resources and advancements in AI algorithms. “The three ingredients are here for the very first time,” Powell said. “This was not possible five years ago.” AI has great potential in the biotech space because of its sheer complexity — just take the problem that AlphaFold targets. Proteins are the basic machinery of your body, managing a wide variety of functions. All of these functions are reliant on the three dimensional shape of a protein. Every protein is made up of a sequence of amino acids, and interactions between those amino acids and the external environment determine how the protein “folds” — which dictates its ultimate shape. Being able to predict the shape of a protein based on its amino acid sequences is of intense interest to biotech companies, which can use those insights to design everything from new drugs to improved crops to biodegradable plastics. “It was always this kind of lunatic, fringe thing. Very much out of the mainstream.” This is where deep learning comes in: training AI models on hundreds of millions of different protein sequences and their underlying structures help those models uncover patterns in biology without necessarily needing to do the expensive computations required by a true molecular dynamics simulation. Fully simulating proteins requires such intense computational resources that institutions have designed and built supercomputers specifically to handle this type of problem, such as the Anton 2 at the Pittsburgh Supercomputing Center. The boom in drug discovery tech isn’t only coming from the AI tech giants. Since 2021, there have been 281 venture capital deals worldwide in AI drug discovery startups, accounting for $7.7 billion in investment, according to Pitchbook. The biggest spike occurred in 2021 as the pandemic took hold, when 105 deals were made, up from 65 the year before, tapering to 67 deals in 2023. In a report published earlier this month, the analyst firm noted that there’s still a strong level of enthusiasm “for early-stage firms integrating AI into drug discovery and development.” The rise of generative AI has also sparked increased interest, said David Baker, director of the Institute for Protein Design at the University of Washington. “It was always this kind of lunatic, fringe thing. Very much out of the mainstream,” said Baker. Now, he said, “everyone is talking about it.” Since the Institute of Protein Design’s founding in 2012, more than 20 startups have been spun out of the program, said Baker. Ten of them — including Archon Biosciences, which develops nanomaterials for regenerative medicine and cancer, and Lila, creating treatments for fibrotic diseases — have come in recent years, since 2021. At DeepMind, it wasn’t until the Covid-19 pandemic hit that researchers truly understood the stakes of their research. They’d worked for almost 5 years to develop AlphaFold, and as they were retraining the model for its second generation, the entire world began sheltering in place because of a mysterious virus. “That really brought home the importance of the problem,” Pushmeet Kohli, vice president of science for DeepMind, told Forbes . The result of DeepMind’s retraining was AlphaFold 2, a groundbreaking model that could so accurately predict protein structures that organizers at CASP, a worldwide research contest for protein folding, emailed DeepMind to ask if the company somehow cheated, Kohli recalled, laughing. The effort has been so promising that cofounder Demis Hassabis spun out a separate company at Alphabet based on AlphaFold’s breakthroughs in 2021. Called Isomorphic Labs, the startup focuses on drug discovery and is helmed by Hassabis himself. Just this year, for example, Isomorphic Labs inked research deals with Lilly and Novartis collectively worth up to nearly $3 billion if all the milestones are met — and that doesn’t count lucrative royalties from potential drug sales that result from these partnerships. In 2022, Nvidia unveiled BioNeMo, a generative AI platform that helps developers accelerate the training, deployment and scaling of large language models for drug discovery. At Nventures, the chip maker’s venture capital arm, seven of the unit’s 19 overall deals have been in AI drug discovery startups, including Genesis Therapeutics, Terray and Generate Biosciences — the largest of any investment category. “The computer-aided design industry created the first $2 trillion chip company,” Powell said, referring to Nvidia and its stratospheric rise over the past year. “Why wouldn’t the same computer-aided drug discovery industry build the next trillion dollar drug company?” She added, “Which is why we are investing in the way that we are.” Several other tech giants have their own protein folding efforts. Last year, Salesforce debuted ProGen, a protein-generating AI model, and Microsoft released EvoDiff, a similar, but open source model. Amazon also released protein folding tools for SageMaker, its AWS machine learning platform. Even ByteDance, TikTok’s parent company, appears to be recruiting for science and drug design teams, Forbes reported in January. Still, as promising and hyped as AI drug discovery is, there are setbacks. It still takes years to get drugs through clinical trials, and while the FDA so far has okayed clinical trials for over 100 new drug candidates that use AI or machine learning for development, it it will likely be years before any hit the market. In some cases, the difficulties associated with drug discovery has caused big tech companies to abandon that research. Last August, Facebook parent Meta shuttered its protein folding team. The researchers at the unit later struck out on their own, founding a company called EvolutionaryScale, Forbes reported last year. Meta declined to comment on the reason for shutting down the project. One important bottleneck that tech companies will need to focus on is having enough training data. Newer foundational models like GPT are reliant on reinforcement learning, a method where algorithms can process unlabeled information through trial and error. That makes them even more reliant on high-quality data, Anna Marie Wagner, head of AI for synthetic biology company Ginkgo Bioworks told Forbes . Last summer, her company entered into a five-year strategic partnership with Google Cloud to pair its AI expertise with Ginkgo’s ability to quickly generate biological data in its automated labs, which can then be immediately put back into the AI model as new training data. This combination, she said, helps to better optimize the discovery process. Additionally, she said, Gingko has the ability to validate model predictions quickly. Counterintuitively, that makes the quirk of AI models sometimes hallucinating — producing wrong or misleading results to a prompt — “a feature, not a bug” because it can lead to interesting discoveries that might have been unimaginable to scientists. “We want the model to come up with the crazy stuff, because that’s where we start seeing order of magnitude improvements.” Kohli puts the data problem more bluntly: “Garbage in, garbage out.” Still, as the industry works to fix those issues, he has already seen the impact that AI has had on biological research. “When I go to conferences and see the change in how biologists were doing their work earlier, and how they’re doing their work today, that’s an amazing transformation,” he said.
This content was originally published here.
