How AI helps scientists design the next generation of medicines

Drug discovery has always been a numbers game nobody wins easily: years of work, huge costs, and most candidates falling apart before they ever reach a patient. That problem gets worse with biologics, drugs built from engineered proteins rather than simple chemistry, where scientists have to sift through enormous pools of possible molecules hunting for the rare one that binds correctly, holds up inside the body, and can actually be manufactured. AstraZeneca is betting that AI can make that search less brutal.

Puja Sapra, who heads R&D biologics engineering and oncology targeted discovery at the company, describes a build-measure-learn loop where AI proposes and ranks candidate molecules computationally before any lab work begins. Only the top-ranked designs get real bench time. That single shift, she says, tightens the feedback cycle, cuts dead ends, and opens the door to targets that used to be considered untouchable. The sheer scale of possible molecular combinations is simply too large for a human team to explore by hand, which is exactly why narrowing the field computationally has become central to how biologics get designed.

The ambitions go beyond speed, though. Traditional biologics are built to hit one disease pathway. Sapra says AI-driven models could help design multi-target drugs instead, ones that balance potency, stability, manufacturability, and safety all at once, and even help decide which two or three biological targets are worth prioritizing in the first place. McKinsey has estimated that generative AI paired with other computational tools could shave up to half off drug discovery timelines, though as Sapra points out, any model is only as useful as the data feeding it. AstraZeneca frames its proprietary, multimodal datasets, covering molecular structures, binding data, safety profiles, and manufacturing outcomes, as the real differentiator, not the algorithms themselves.

To pull all of that together, the company is building what it calls a lab of the future in Kendall Square, Cambridge, Massachusetts, meant to link AI predictions, robotic experimentation, and instrument-generated data into one closed loop. Sapra compares it to a self-driving car reading its environment and adjusting in real time. The long-term goal, what the field calls de novo design, is an AI system that generates an entirely new protein sequence from scratch, one built to fit specific drug properties from the very start, all the way through to a clinical candidate.

Getting there depends on more than clever models. Sapra flags safety prediction, figuring out whether a computationally dreamed-up molecule will behave safely inside a human body, as maybe the hardest and least talked-about piece of the puzzle. AstraZeneca is approaching this with advanced cell systems and micro-scale organ models acting as physical testbeds, paired with AI trained on their output, alongside a broader move toward agentic systems that can generate and evaluate candidates simultaneously. Sapra is careful to insist scientists stay central throughout, providing the judgment and oversight that keeps the process explainable rather than a black box.

What that adds up to, in her telling, is less a story about algorithms replacing scientists and more one about engineers, data scientists, and biologists building tools meant to act as thinking partners. The problems involved, she notes, are genuinely hard: fusing multimodal data, closing feedback loops, quantifying uncertainty, making systems interpretable enough to trust at the point of clinical decision-making. Whether a fully AI-generated biologic candidate actually reaches patients is still, by her own admission, a matter of time rather than a done deal.