AI in Drug Development: Promise and Pitfalls

By Allison Proffitt 

April 22, 2025 | An expert panel of technology and pharmaceutical executives convened at this month's Bio-IT World Venture, Innovation, and Partnering event to discuss the transformative potential of artificial intelligence in drug development—and the significant challenges that still lie ahead. Participants included Rory Kelleher of NVIDIA, Bill Fitzgerald from Google Cloud, Bill Mayo of Bristol-Myers Squibb, Anthony Philippakis of GV (formerly Google Ventures), and Becky Stevenson of HSBC, moderated by Jeremy Goldberg of Arsenal Capital Partners. 

Computing Power Meets Biology 

Rory Kelleher highlighted NVIDIA's recently launched Blackwell computing platform, describing it as "the most advanced AI accelerator out there for training and also deploying AI models at scale." Kelleher explained that three key "scaling laws" are driving AI innovation: pre-training (acquiring broad knowledge), post-training (domain-specific learning), and test-time compute (intensive reasoning). 

These advances are particularly promising for biological applications. "What I am excited about… is these reasoning models, these test-time compute models, are going to be an incredible productivity gain in the world of science," Kelleher said. He noted AI's potential to review scientific literature, generate hypotheses, and even design in silico experiments. 

Understanding Causal Biology 

Bill Mayo of Bristol-Myers Squibb emphasized that AI's most crucial application in pharma is understanding causal human biology. "If we don't actually understand what's going on, then the greatest folding model, the greatest docking model, the greatest multi-parameter optimization...all of that isn't going to work if we actually have the wrong biological target," Mayo explained. 

He expressed concern that tech companies might "accidentally solve biology" and reduce pharmaceutical companies to contract research organizations. This drives BMS to maintain a competitive edge in biological understanding while leveraging AI tools. 

Data Generation and Partnerships 

Panelists agreed that generating quality biological data remains a significant challenge. Kelleher advocated for public-private partnerships to build the necessary datasets: "It could take a trillion cells before we can really start to understand the true causal relationships of biology." 

Bill Fitzgerald of Google Cloud highlighted the importance of collaboration and tooling: "We train on the same level of data that is done in the public available datasets... but we do that at a massive scale." Google has launched several open-source models for healthcare AI, including pathology models and repurposing models, designed to speed development for the biotech community. 

Investment Landscape 

Rebecca Stevenson of HSBC provided insight into the current investment environment, noting a "post-COVID recovery period" where valuations for later-stage companies have cratered while early-stage companies are seeing better valuations. "What we're seeing is a complete sort of stabilization back to hovering around one times of step-up valuations," Goldberg said. 

She highlighted the tension between tech investors and life science investors, who approach total addressable market (TAM) and valuation very differently: "Your life science people are focused on de-risking around your FDA, your regulatory constraints, and your tech guys are saying, ‘The sky's the limit.’” 

The Human Element 

Anthony Philippakis identified the shortage of interdisciplinary talent as perhaps the biggest challenge. "It doesn't work very well to talk biology and computer science across the table; you need to do it across the corpus callosum," he explained. “There really is a need for people who can put the pieces together and make very, very non-obvious connections.”   

Echoing a sentiment that came up in on the plenary stage of the main Bio-IT World conference, Philippakis said, “Figuring out how we can actually train the next generation of workforce that is actually fluent in both domains, I find that to be the biggest rate limiting step right now.” 

Philippakis later predicted that pure mathematics would be the first field dramatically transformed by AI, while "drug development is actually going to be the hardest one because the iteration cycles are longer, they’re multidisciplinary. It’s a lot harder to come up with benchmarks that show you a leaderboard. Protein folding is the exception, not the rule. Most problems don’t have such measurable outcomes.” 

When asked about what constitutes a defensible business advantage or "moat" in AI-driven biotech, Philippakis suggested that currently, "the single biggest moat is the number of people that can actually do the work of training the models and deploying them." Right now, these skill sets are not being developed in academia, he said, because universities can’t afford models with the needed number of parameters. “Human beings that are the scarce resource is my best guess.” Longer term, he believes the sales channel and customer connections will become crucial differentiators, especially given the high technology switching costs in healthcare. Kelleher noted that strong relationships and networks remain vital: "One thing that AI is not likely to replace anytime soon is relationships." 

Looking Forward 

Despite current market uncertainties, panelists expressed optimism about the future. “It’s hard to be optimistic at the moment, but actually I’m very optimistic, Stevenson observed. “There's never been more capital pointed at healthcare and life sciences, and there's never been a deeper desire by the general public to advance science forward." 

Goldberg challenged the panelists to share three-year predictions. Mayo foresaw that within three years, many current technical challenges will be overcome. Stevenson anticipated more M&A as companies struggle to get financing and move their platforms forward. Fitzgerald is bullish on increased collaboration between personalized medicine, diagnostics and pharma companies. Kelleher predicted that lab-in-the-loop and recursive, iterative learning will become the industry standard, resulting in faster target identification and lower costs in early drug discovery.  

As biotech continues its AI transformation, the panel suggested that companies must choose whether to lead with science or technology—and those that successfully bridge both domains may ultimately reshape drug discovery and development.