by Alex Devereson, Chris Anagnostopoulos, David Champagne, Hugues Lavandier, Lieven van der Veken, Thomas Devenyns, and Ulrich Weihe
with Alex Peluffo, Benji Lin, Jennifer Hou, and Maren Eckhoff
It has been two years since the emergence of generative AI (gen AI) reset the expectations of what’s possible in business, and organizations are beginning to recognize its real value, using gen AI to boost the productivity of software engineers, for example, or increase the efficiency of marketing campaigns and customer-service operations. These early results are consistent with McKinsey research that found that 75 percent of the value of gen AI use cases would come from four areas: marketing and sales, software engineering, customer operations, and R&D.
Of those four, R&D remains the least appreciated and, perhaps, the most compelling. That’s because beyond boosting productivity and efficiency of laboratory researchers, recent developments in AI have the potential to transform the entire R&D process, fundamentally accelerating the metabolic rate at which ideas are explored and even creating altogether new hypotheses to investigate. GPT-4—which powers ChatGPT—and other large language models (LLMs) are already leveraging advancements in human-language processing to deliver progress across scientific disciplines. A recent Microsoft preprint, for example, demonstrates how GPT-4 shows a strong potential to analyze and synthesize complex scientific information in fields including biology, drug discovery, computational chemistry, and materials design.1 Because so much scientific information is text-based, LLMs have been used with success in various contexts, such as detecting patterns in DNA expression in cancer cells and helping solve engineering problems in aerospace design.2
But realizing the full potential of AI for scientific discovery and development requires a different approach to AI—an approach we’ve defined as “Scientific AI.”
What is Scientific AI?
As its name implies, Scientific AI uses AI to generate and test scientific hypotheses. It comprises AI tools and practices developed specifically for scientific applications, such as methods that use large amounts of scientific data, especially raw objective measures, to power cross-disciplinary scientific insights. In contrast with AI tools used to drive efficiency and productivity in operations—tools more likely to become commoditized—Scientific AI leverages proprietary data and expertise, with the potential to supercharge innovation and become a true competitive differentiator.
We believe that Scientific AI has the potential to solve some of the thorniest, long-standing challenges faced by researchers across broad branches of science—such as chemistry, biology, materials, and physics—helping propel innovation across all industries where science matters. It can do so in several important ways.
Working across silos
Scientific discovery has long relied on siloed approaches, with data and methods developed to address highly specific problems. Eight years ago, in an article urging the adoption of more interdisciplinary approaches, Nature Biotechnology lamented, “Research silos continue to hamper biological research.”3 More recently, a 2023 article in Nature Synthesis argued that the “siloed nature of the conventional research efforts on the discovery of new materials and molecules” remains an issue in chemicals and materials discovery.4
Scientific domains such as genome analysis and physics modeling have similarly hit a plateau of innovation, in part because of the difficulty of synthesizing quantitative insights produced by modeling techniques that are highly tailored to individual data types. Scientific AI provides a pathway out of that impasse, enabling approaches with the potential to break down silos in scientific discovery and product development. Scientific AI can leverage advances in foundation models that are multimodal in nature and generalize across data domains or even industries, such as using chemistry foundation models to enhance drug and polymer discovery, and leverage advances in the broader AI ecosystem, such as synthetic data generation and causal machine learning.
By integrating a diverse corpus of data, foundation models can extrapolate beyond their strict training perimeters to derive net new insights. For example, models originally developed for drug discovery that combine molecular information and imagery can be repurposed for specialty chemicals synthesis; models developed for computer vision in plants can be tailored to cancer detection in human cells; or models developed to predict antibody protein-folding can be used to engineer food applications enzymes.
Going beyond text and language
Many of today’s most popular and powerful AI and gen AI applications are text-based, but the core of critical data in R&D is not based on text and is instead highly heterogenous, encompassing images, molecular structures, dynamical systems, activity measures, and statistical responses. This core of critical data is drawn from diverse sources such as petri dishes, formulations, clinical trials, microscopes, radars, and other scientific instruments. Because most of industrial R&D relies on these data modalities to discover new products, the impact potential of new foundation models based on such data sources is very high.
For example, protein engineering is a core driver of R&D in several pivotal industries, such as pharmaceuticals (such as therapeutic antibodies), medical diagnostics (such as antibody-based affinity matrices), industrial chemistry (such as technical enzymes for detergents), and renewables (such as plastic-digesting enzymes). All these industries are beginning to benefit from breakthrough protein foundation models, such as RoseTTAFold and AlphaFold 3. (The lead researchers behind these technologies were awarded the 2024 Nobel Prize in chemistry and have raised more than $1 billion in Series A funding to continue to translate these technologies to industry.5 ) Similarly, foundation models such as Uni-Mol, FM4M, and SPMM6—which are trained on the properties of chemical structures—allow researchers to predict the nature of small-chemical molecules and even generate previously unknown ones.
Operating in iterative cycles
AI models propose designs, laboratory researchers and engineers test these proposals, and the resulting data are incorporated into the AI to derive new insights. This process of generation, testing, and refining drives innovation through data enhancement and continuous learning. In a world in which more gen AI models are becoming open-source and talent is free to move from one player to the next, the path to differentiation relies on data and the training of these models through active-learning loops. Access to data is a core competitive advantage that can translate into value only if data are properly integrated and can flow back and forth from AI to laboratories. We are already seeing investors in Scientific AI favor companies with active-learning loops that create proprietary insights via fine-tuning in specific data sets.
And these iterative cycles will grow only more robust with the advent of agentic AI, which will allows researchers to conversationally interact and share expertise with AI-driven knowledge agents trained on a broad base of scientific knowledge and historical data spanning multiple industries. More colloquially, this is a world in which an AI companion can tell researchers, “Don’t run this experiment; it has been done before and failed” or “The last person who ran this analysis made the best progress through this next step.”
It is important to understand that the abundance of new evidence likely to be generated by Scientific AI will come with different grades of confidence, depending on the amount of data, the number of training cycles, and the extent of external validation. In other words, all evidence is not created equal. To get the most out of Scientific AI, organizations will need to build business processes that are able to leverage and, where needed, further validate insights of varying degrees of certainty.
Getting started with Scientific AI
Many industries—ranging from pharmaceuticals and agriculture to automotive, aeronautics, and energy—stand to reap considerable value from the deployment of Scientific AI. Our analysis finds that this value will come from two sources: accelerating productivity through speed and a higher probability of success (making the wheels spin faster) while simultaneously enabling new solutions and domains (creating entirely new wheels). That said, because Scientific AI deeply affects the entire R&D process, it requires a set of transformation building blocks for successful adoption at scale. A McKinsey framework that can enable successful analytics transformations comprises six key dimensions that must be addressed: a blueprint linked to scientific and business value, digital and analytics capabilities, data architecture, technical architecture, talent and an agile operating model, and an adoption and scaling plan that details the road map from the first superusers to broad adoption. By addressing all six dimensions throughout the process of scientific discovery, organizations can ensure that Scientific AI is deployed at scale, with a direct impact on strategic priorities.
Even with those factors in place, organizations must be careful about attempting too much at once. Rather than launching an array of proofs of concept, companies are better served by deploying a smaller number of initiatives (even just one or two) that can be tied to strategic and business objectives and serve as a foundation for future initiatives. A strategy to drive adoption over the long term also is essential. Scientific AI may be powerful, but it’s also a new capability that could take some time to deliver. Overpromising on early efforts could result in disappointment and skepticism among users that undermine the technology’s long-term potential. But with the proper framework and a sound organization-wide strategy, Scientific AI heralds a new age of creativity, innovation, and transformation.
Alex Devereson is a partner in McKinsey’s London office, where David Champagne is a senior partner and Maren Eckhoff is a distinguished data scientist; Chris Anagnostopoulos is a partner in the Athens office; Hugues Lavandier is a senior partner in the Paris office, where Alex Peluffo is a consultant; Lieven van der Veken is a senior partner in the Lyon office; Thomas Devenyns is a partner in the Geneva office; Ulrich Weihe is a senior partner in the Frankfurt office; and Benji Lin is an associate partner in the Boston office, where Jennifer Hou is a senior asset leader.
1 Microsoft Research AI4Science and Microsoft Azure Quantum, “The impact of large language models on scientific discovery: A preliminary study using GPT-4,” arXiv, November 2023.
2 Wenpin Hou and Zhicheng Ji, “Assessing GPT-4 for cell type annotation in single-cell RNA-seq analysis,” Nature Methods, March 25, 2024, Volume 21; Matthew J. Ha and Kristopher C. Pierson, Usage of ChatGPT for engineering design and analysis tool development, AIAA SCITECH 2024 Forum meeting paper, AIAA 2024-0914, January 2024.
3 “So long to the silos,” Nature Biotechnology, 2016, Volume 34.
4 Milad Abolhasani and Eugenia Kumacheva, “The rise of self-driving labs in chemical and materials sciences,” Nature Synthesis, 2023, Volume 2.
5 “David Baker: Facts,” Nobel Prize Outreach, accessed January 8, 2025; Annalee Armstrong, “New AI drug discovery powerhouse Xaira rises with $1B in funding,” Fierce Biotech, April 24, 2024.
6 Qiankun Ding et al., “Uni-Mol: A universal 3D molecular representation learning framework,” ChemRxiv, March 6, 2023; “Introduction to IBM foundation models for materials (FM4M),” GitHub, accessed January 8, 2025; Jinho Chang and Jong Chul Ye, “Bidirectional generation of structure and properties through a single molecular foundation model,” Nature Communications, 2024, Volume 15.