[Digital Today reporter Seung-ah Yoo] As artificial intelligence (AI) becomes core infrastructure for biopharmaceutical development, methods for finding and validating new drug candidates are changing rapidly.

MIT Technology Review reported on July 23 that Puja Sapra (푸자 사프라), a senior vice president at UK drugmaker and biotech company AstraZeneca and head of R&D biopharmaceutical engineering and oncology target discovery, explained that computing has strengthened every stage from design to production, validation and analysis. She said this is shortening development cycles while raising productivity and innovation together.

Biopharmaceuticals are far more complex to design than synthetic chemistry-based medicines. Molecules must bind to desired targets, remain stable in the human body and be capable of mass production. Potential molecular combinations are vast, but only a very small number of candidates reach actual patients.

AstraZeneca uses a build-measure-learn cycle. When AI predicts molecules with a high chance of success or sets priorities, researchers focus experimental resources on top candidates. Sapra said the process reduces trial and error and increases iteration speed, making it possible to approach targets previously considered difficult to drug.

AI is changing not only development timelines but also drug design itself. Next-generation biopharmaceuticals are evolving toward simultaneously attacking multiple targets rather than aiming at a single disease pathway, or delivering therapeutic substances precisely to specific cells. Sapra said AI plays an important role in prioritising multiple targets and balancing efficacy, safety and manufacturability, and assessed that an era is opening in which targets once considered untreatable can be tackled.

Data is the core of AI competitiveness. McKinsey estimated that using generative AI along with other computational tools can cut drug discovery timelines by up to 50 percent. AI performance, however, depends on the quantity and quality of training data. Sapra explained that AstraZeneca's competitiveness lies in its self-built multimodal dataset. The data include molecular structures and binding measurements, safety profiles and manufacturing results, and are built to cover a range of diseases and drug types. She said the company is also investing in precision screening technologies to fine-tune and continuously validate the latest AI models based on this foundation.

AstraZeneca is building a future lab that combines AI and robotic automation in Kendall Square in Cambridge, Massachusetts. In a closed-loop cycle, AI predicts experiments, robots carry them out, equipment generates data and the data are then used again for AI training. Sapra projected that automated high-throughput systems will eventually be able to create and evaluate thousands of molecular interactions each week.

She stressed that scientists must still play a central role even as autonomy expands. The role of providing oversight, judgement and strategy so that AI results are explainable, reliable and truly helpful to patients ultimately belongs to people, she said.

The ultimate goal is de novo design. In this approach, AI designs entirely new protein sequences to match desired drug properties, while optimising structure, safety, how they function in the body and manufacturability. Sapra said the industry is moving toward developing fully AI-generated biopharmaceuticals in which AI handles the entire process from design through selection of clinical candidates.

She noted that achieving this will require richer and standardised training data, robust benchmarks to evaluate AI-designed candidates, and talent that understands both machine learning and biology. She stressed that the most difficult task is predicting whether computer-generated molecules will be safe in the human body.

AstraZeneca is introducing what is effectively a virtual clinical trial approach, using advanced cell systems and miniature organ models to address this. The strategy is to use the resulting data for AI training and narrow the gap between AI-designed candidates and the actual clinical stage.

Sapra projected that agentic AI systems that generate molecular candidates while also predicting efficacy and safety will spread. But as systems become more autonomous, human oversight will become even more important, she said, adding that scientists will work with AI and take on roles that verify and synthesise results.

She said AstraZeneca's engineering team is developing AI systems that serve as a "thinking partner" rather than a black box. Engineers are designing systems to solve complex problems such as multimodal data fusion, closed-loop optimisation, uncertainty quantification and interpretability in clinical decision-making stages, and she stressed that the combination of AI technology and scientific expertise will be key to next-generation biopharmaceutical development.