AstraZeneca says artificial intelligence is becoming a working layer of biologic drug discovery, not a side demo bolted onto the lab. The point is practical: biologic medicines, which are engineered proteins rather than small-molecule chemicals, are hard to design, hard to manufacture and prone to failure before they ever reach patients.
Puja Sapra, AstraZeneca’s senior vice president and head of R&D biologics engineering and oncology targeted discovery, said the company now uses computational tools across design, production, testing and analysis. In AstraZeneca’s telling, AI helps shorten the cycle by ranking candidate molecules before scientists spend lab time on them.
The mechanism is not magic. Researchers define the disease target and the properties a molecule needs, such as binding behavior, stability, safety signals and manufacturability. Models generate or prioritize candidates. Scientists test the strongest options in the lab. Those results go back into the system, improving the next round. That build, measure and learn loop is the pitch, with fewer doomed molecules eating time and budget.
Why biologics are a hard target
Biologics can be used across major acute and chronic diseases, but their design space is enormous. A team cannot manually inspect every plausible protein sequence or molecular configuration. AstraZeneca argues that AI is useful because it can narrow that search to candidates more likely to meet several constraints at once.
Sapra said the same approach could support more complex medicines, including multi-specific biologics that act on more than one target, or drugs that deliver a therapeutic payload to particular cells. That requires balancing potency, stability, manufacturability and safety at the same time. The company says AI models may help choose which targets to combine and how to tune the molecule around them.
McKinsey has estimated that generative AI and other computational methods could reduce drug discovery timelines by as much as 50%. That is an estimate, not a clinical outcome, and it depends on the quality of the models and the data used to train them.
The data problem is the product
AstraZeneca says its advantage is proprietary, multimodal data, including molecular structures, binding measurements, safety profiles and manufacturing results. Failed experiments matter too, because they show the model what does not work. Sapra said the company has also invested in screening technologies to generate more data for training and validation.
The company is building a facility in Kendall Square in Cambridge, Massachusetts, which it describes as a “lab of the future.” The plan is to connect AI prediction, robotic experimentation and instrument-generated data into a closed loop. AstraZeneca says automated systems could eventually create and evaluate thousands of molecular interactions each week, with robotic sample handling, quality checks and data pipelines feeding results back into models.
Designing from scratch remains unfinished
Sapra said the long-term goal is de novo biologic design: models that create new protein sequences from scratch for a desired therapeutic profile. That would require predicting structure, behavior in the body, manufacturability and safety before a candidate moves toward the clinic.
Several pieces are still missing, according to Sapra: richer standardized training data, credible benchmarks for AI-generated molecules and teams that understand both machine learning and biology. She identified safety prediction as one of the hardest problems. AstraZeneca says it is pairing AI with advanced cell systems and micro-scale organ models that act as physical testbeds, a kind of preclinical proving ground for model-generated designs.
The company also says scientists remain in charge of oversight and judgment as autonomous systems take on more of the generation and testing workflow. That human-in-the-loop caveat is doing real work here. Models can propose molecules quickly. They do not make a medicine safe, effective or approved by saying so.
AstraZeneca disclosed that the material describing this work was initiated and funded by the company and produced by MIT Technology Review’s custom content unit, not its editorial newsroom.
This story draws on original reporting from MIT Technology Review.
