In drug development, where timelines are long, costs are high, and many bets fail, any tool that helps reduce risk is worth keeping. But only if we’re honest about what it can’t do yet.
For all the advancements in machine learning over the last decade, from transformer architectures, diffusion models, and protein structure prediction breakthroughs, the real test of AI in biotech remains unchanged: does it improve the likelihood of developing a viable drug?
At this year’s PEGS Boston summit, a panel of experts from industry and academia confronted that question well beyond abstract terms.
The discussion, led by Peter Tessier of the University of Michigan, cut through the common framing of AI as disruption, instead examining where it actually works, where it falls short, and where, frankly, it may never belong.
AI is already making itself useful in a few well-defined corners of work. Panelists pointed to three in particular: predicting developability, guiding sequence humanization, and modeling structure. At companies like Merck and Adimab, these tools are built into everyday workflows, spotting sequences that might clump up, reducing parts that could trigger immune responses, and making it possible to model proteins without needing crystal structures.
The panelists went a bit into the weeds on application-specific requirements but the point is that they’ve made some common steps faster, easier, and less error-prone.
These are not revolutionary shifts, they argue, but they’re operationally significant. What matters isn’t novelty, it’s that AI is actually improving early-stage confidence and importantly, it’s also reducing costly late-stage failures.
As Kadina Johnston, senior specialist of Discovery Biologics at Merck noted, the real gain isn’t necessarily in speed, but in the compound advantage of avoiding bad paths before they become expensive.
But that clarity doesn’t extend to all areas of AI application. De novo protein design, which has long been the centerpiece of AI-centric drug discovery, was acknowledged as promising but still constrained.
As Sarel Fleishman, professor of Biomolecular Sciences at the Weizmann Institute of Science explained, most AI design models tend to make simple, tube-like protein shapes that are easy to build and simulate, but don’t work well for the deep, complex binding sites found on hard targets like GPCRs and ion channels. For reference, these are types of proteins that sit on the surface of cells and play key roles in how cells send and receive signals. They’re important in many diseases but their binding sites are often narrow, tricky to reach, and chemically complex, which makes them especially difficult to design around using today’s AI tools.
In practice, this means current AI design tools often aren’t built to handle the toughest drug targets. They tend to create neat, ideal shapes that look good on paper but don’t actually bind well to the complex surfaces found on many real-world proteins. Until these tools can be trained to produce more varied and useful shapes (like those found in natural antibodies) their role will stay limited. Not useless, but only helpful in certain cases.
This insight reflects a broader tension in the field. As AI/ML becomes more accessible and scalable, the temptation to overuse it grows.
The panel made clear that usefulness does not equal applicability. AI is only effective when it's solving a problem that's well-defined, measurable, and for that matter, embedded in a much larger system where all the old validation mechanisms work.
That’s where benchmarking becomes essential (and became a topic of conversation among panelists).
One of the more uncomfortable truths the panel laid bare is that many of the claims being made about AI-designed biologics aren’t actually based on how those molecules perform in the real world. They’re based on internal scores ( guesses about how strong a binder is, or how stable a sequence looks in theory). These are helpful starting points, but they don’t always translate into what happens in a lab dish, let alone in a human body.
Without direct, apples-to-apples comparisons between AI-generated molecules and those designed using traditional methods, panelists agreed it’s hard to know what the models are really getting right.
This was one of the few places where every panelist, whether in academia and industry, landed on the same note. If AI is going to be part of the design process, it has to be held to the same experimental standards as everything else.
Where the conversation grew more optimistic was around optimization, or using AI to improve molecules that already show promise. This is where AI/ML is proving not just reliable but deeply useful.
