AstraZeneca's AI loop can cut drug discovery timelines by 50%, exec says

The brutal economics of drug development are well-known: a decade or more of work, billions in sunk costs, and a failure rate that would embarrass most industries. But a quiet computational revolution is changing how the sausage gets made, especially for complex biologic medicines.

AstraZeneca is now running a “build-measure-learn” loop where machine learning models, not human intuition, generate and prioritize the most promising molecular candidates. Puja Sapra, who leads biologics engineering at the pharma giant, describes a world where everything is “computationally enhanced.” The payoff is stark. McKinsey estimates that this combo of AI and automation can slash discovery timelines by up to 50%. Instead of a scientist squinting at a list of thousands of theoretical molecules, the AI pre-screens them, allowing the wet lab to focus only on top-ranked designs. That’s a tighter feedback cycle with far fewer expensive dead ends.

What’s genuinely new here isn’t just speed. Sapra points to “drugging the undruggable”—designing multi-specific biologics that can hit several disease pathways at once or deliver a payload to a specific cell type. That is a multi-variable optimization nightmare for a human brain, but it’s standard pattern-matching for a model trained on enough data. And that’s precisely the moat AstraZeneca is digging. Sapra calls data their “differentiator,” citing proprietary, multimodal datasets that include molecular structures, safety profiles, and manufacturing outcomes. Every failed experiment is a data point that refines the next model iteration.

The company is building a physical “lab of the future” in Cambridge, Massachusetts, to completely close the loop. Sapra compares it to a self-driving car: AI makes a prediction, robotic automation runs the experiment, instruments generate data, and that data immediately updates the model. She’s adamant that scientists aren’t being sidelined, but it’s clear the boring work of pipetting and hypothesis generation is being handed off to a system that never sleeps. The open question isn’t whether this works—it’s how quickly this becomes a competitive advantage rather than an industry standard.