Biopharma companies won’t fully capture the benefits of AI unless they reorganize their R&D units, according to McKinsey.
Interest in applying AI to drug development has surged in recent months. Eli Lilly, Bristol Myers Squibb and Incyte unveiled deals in May, with Alnylam and Merck following in June. The field features startups with deep pockets, including an Alphabet business, Isomorphic Labs, that raised $2.1 billion in May. One question for established companies is how to integrate AI into existing workflows.
Researchers at management consultancy firm McKinsey argue in a new report that applying AI to the current R&D operating model may make individual steps more efficient, but will not “compound learning.” While AI can accelerate decisions and reshape inflection points, programs still move linearly through stage gates without creating “systematic feedback across decisions.”
McKinsey has proposed reorganizing R&D around five connected decision points to maximize AI’s value. The first point is understanding patients and disease biology, while the last point is improving the impact of approved therapies for patients.
The three middle points cover activities such as validating targets and running clinical trials—similar to decisions made using the traditional R&D model. The difference, in McKinsey’s view, is that a closed-loop R&D model ensures that “every pivotal decision generates data that informs the next and refines the one that preceded it.”
Outputs become inputs to drive a “continuous cycle of learning,” McKinsey said. That idea is already fundamental to many companies, particularly techbio startups. Recursion Pharmaceuticals, for example, characterizes its platform as combining “large-scale phenomics, emerging omics layers, AI-driven chemistry design and clinical development intelligence into a single, closed-loop system.”
McKinsey named Robin, a multiagent system developed by the nonprofit FutureHouse, and Google DeepMind’s Co-Scientist as examples of the real-world potential of AI-powered loops. A recent Nature paper described the cycle through which Robin generates hypotheses, proposes experiments, interprets experimental results and generates updated hypotheses.
Robin proposed enhancing retinal pigment epithelium phagocytosis to treat dry age-related macular degeneration. Having generated the hypothesis, the system identified and confirmed in vitro efficacy for the combination of approved eye drops ripasudil and an experimental small molecule called KL001 that interacts with cryptochrome. Robin also proposed and analyzed a follow-up RNA sequencing experiment.
The high cost and long duration of clinical development make human testing the biggest opportunity to make R&D cheaper and faster. McKinsey pitched the closed-loop model as a way to “compress both trial duration and decision cycles.”
In that context, the loop includes predictive models to optimize patient selection and trial design, data integration and synthesis to combine clinical trial and real-world data, and agentic systems to orchestrate operations. AI models and agents can identify how patients in the control arm are likely to perform and screen draft submissions against prior agency feedback, McKinsey said.
Companies considering implementing AI-powered loops should create a blueprint for the system they ultimately want across all five decision points, the consultancy said. McKinsey is encouraging companies to use the blueprint to identify early wins and guide investment toward elements that can be added over time.
