Drug discovery has always been a game of improbable odds. A new medicine can take many years to develop, consume substantial investment, and still fail before reaching a single patient. For biologic medicines, therapies built from engineered proteins rather than synthetic chemistry, the challenge is even steeper. The number of possible molecular combinations is so vast that no human team could systematically explore it. This is the problem AI is now being asked to solve, and the approach is more sophisticated than most coverage suggests.
The Build-Measure-Learn Loop That Is Shrinking Drug Timelines
The core shift happening in pharmaceutical R&D is not simply that AI is being added to existing workflows. It is that AI is restructuring the logic of the entire discovery process.
At AstraZeneca, the approach follows what Puja Sapra, senior vice president and head of R&D biologics engineering and oncology targeted discovery, describes as a build-measure-learn loop. AI generates or ranks candidate molecules computationally, predicting which designs are most likely to succeed before any physical experiment takes place. Scientists then direct laboratory resources exclusively toward the top-ranked candidates. The result is a tighter feedback cycle: fewer dead ends, faster iteration, and the ability to pursue disease targets that were previously considered unreachable.
McKinsey estimates that generative AI, combined with other computational tools, could cut drug discovery timelines by as much as 50%. That figure deserves context. It does not mean AI is replacing scientific judgment. It means the ratio of productive experiments to wasted ones is improving, and the time between hypothesis and result is compressing.
What makes this possible is data. AstraZeneca’s datasets are proprietary and multimodal, incorporating molecular structures, binding measurements, safety profiles, and manufacturing outcomes. The company has also invested in deep screening technologies to generate additional datasets at the volume required to continuously refine and validate its models. As Sapra puts it: “Data is our differentiator.”
Designing Drugs That Did Not Previously Exist
Beyond accelerating timelines, AI is opening a category of medicines that traditional methods could not reach at all.
Conventional biologic drugs typically target a single disease pathway. The next generation, sometimes called multi-specific biologics, can hit multiple targets simultaneously or deliver therapeutic payloads to specific cells with precision. Designing these molecules requires optimizing across many variables at once: potency, stability, manufacturability, and safety. That kind of multi-dimensional optimization is exactly where AI models have a structural advantage over manual approaches.
Sapra describes a scenario where AI models could identify which two or three biological targets to prioritize, then optimize across all relevant parameters in parallel. “Drugging the undruggable is becoming a reality,” she says. Targets once considered impossible to reach are now within scope.
The longer-term vision is what the field calls “de novo” design: AI generating entirely new protein sequences from scratch, designed to fit a precise set of drug properties, with safety and manufacturability built into the prediction from the start. AstraZeneca is building toward this with a “lab of the future” facility in Kendall Square, Cambridge, Massachusetts. The facility is designed as a closed-loop discovery system where AI makes predictions, robotic systems execute experiments, and instruments generate data that feeds directly back into the models. Automated high-throughput systems will eventually evaluate thousands of molecular interactions on a weekly basis, producing AI-ready data at a scale that traditional workflows cannot match.
Safety prediction is the hardest unsolved piece. Predicting whether a computationally generated molecule will be safe in the human body remains one of the most consequential challenges in de novo design. AstraZeneca is addressing this through advanced cell systems and micro-scale organ models that function as physical testbeds, paired with AI that learns from their outputs. These are, in effect, virtual clinical trials.
Why This Matters Beyond the Laboratory
The transformation underway in biologics is not primarily a story about technology. It is a story about what becomes possible when the bottleneck shifts.
For decades, the constraint in drug discovery was not scientific imagination. It was the sheer cost and time required to test ideas. Most promising candidates were never explored because the pipeline could not accommodate them. AI does not eliminate that constraint entirely, but it changes the ratio dramatically. More candidates can be evaluated, more disease targets become viable, and the feedback loop between experiment and insight tightens.
Three prerequisites still stand between the current state and fully realized de novo drug design: richer and more standardized training data across the industry, robust evaluation benchmarks for AI-generated candidates, and teams capable of working at the intersection of machine learning and biology. None of these are trivial. But the direction is clear.
Critically, the human role in this process is not diminishing. Scientists remain central to oversight, judgment, and strategic direction. The goal is not autonomous drug design without human input. It is a system where AI handles the combinatorial volume that humans cannot, while scientists focus on the decisions that require biological intuition, ethical reasoning, and accountability to patients.
In Short
AI is not replacing drug discovery. It is restructuring its logic. By narrowing the search space computationally, enabling multi-target drug design, and building closed-loop systems where every experiment improves the next prediction, pharmaceutical R&D is moving toward a model where the limiting factor is no longer the number of experiments a team can run. The goal, still in progress, is a completely AI-generated biologic designed from scratch to a clinical candidate. The prerequisites are significant, but the trajectory is no longer speculative.
Based on reporting from MIT Technology Review.