In the past five years, however, advancements in deep learning models, the increased availability of scientific data and high-quality biological knowledge graphs and the growth in computing power have made scientific AI much more powerful. Researchers are now using it to understand the intricate biology of disease—understanding the genes that contribute to illness, the structures of molecules that make effective drugs, and which diseases those drugs will be most effective in treating.
“Scientific AI integrates diverse and massive biological datasets, from chemical structures and patient journeys to genomic information, with cutting-edge AI models,” says McKinsey & Company partner Alex Devereson. “This powerful combination allows researchers to explore the scientific landscape in ways they previously could not.”
For instance, one leading pharmaceutical company is using AI to accelerate the preclinical design and validation of gene therapies. From millions of potential molecules, AI helps researchers identify more potent candidates in a quarter of the time it took previously. AI also reduces by a factor of three the time it takes to design the vector to carry the active ingredient into the patient. Finally, it accelerates and improves the process of translating the laboratory product into a therapy for clinical trials: It uses a database that includes the results of translational experiments to help researchers predict the clinical outcomes of different molecule and vector combinations.
In each of these three stages, researchers use scientific AI to simulate experiments before they are conducted in the lab so that they can optimize parameters and focus on the most promising options. They then feed the results back to the model. This closed-loop approach continuously refines the experiments, speeding the path to clinical trials.
Saving time, saving lives: AI in drug discovery and development
The journey of a new drug from concept to market is a marathon fraught with peril. For every approved medicine, 5,000 to 10,000 compounds are tested in discovery, of which only 10 to 20 move on to preclinical trials. Since the 1990s, researchers have used machine learning (ML) algorithms to analyze large datasets of molecular information to more quickly and reliably identify promising candidates for development, but its role was limited.
From identifying new disease targets to refining patient selection, AI is transforming drug discovery and development with its ability to analyze real-world data and simulate experiments.
i
i
At the heart of virtually every large organization lies a legacy tech system slowing it down. In fact, about 70% of the software that powers Fortune 500 companies was developed at least 20 years ago. Upgrading these legacy systems is essential, but the process can be daunting—a typical mainframe modernization project, depending on the amount of code, can require more than 200 engineers, take multiple years, cost millions of dollars, and even be outdated before it is completed.
Advances in generative AI are making it possible to modernize enterprise technology without all the traditional hurdles.
One of the most promising use cases for gen AI has been in updating software code. “With the unprecedented power of large language models, trained on software development, it has suddenly become feasible and economically viable for organizations to attempt these kinds of modernizations that were previously too difficult and too expensive,” said Dan Collins, associate partner at McKinsey. The emergence of agentic AI, a powerful new capability, has further revolutionized this process. Agentic AI allows developers to create entire teams of AI agents in different roles (e.g., data engineer, coder, tester) and instruct them to assess, update, and build new technology infrastructures that meet current and future business demands.
Scientific AI can also analyze real-world data (RWD) to identify new uses for existing mechanisms of action (MoAs). McKinsey worked with a client to test this approach for PD-1 inhibitors. It analyzed combined horizontal health record and claims data, cleaned of PD-1 data, for 3 million cancer patients, to identify 4,000 features in categories such as procedures and prescriptions. Then the company used ML to glean from a handful of indications susceptible to PD-1s, such as melanoma, the most relevant features. The ML then assessed other indications and created a prioritized list of candidates for treatment with PD-1s.
To validate the results, researchers defined those where PD-1 inhibitors were already approved as positive examples, and those where multiple Phase 3 trials had failed as negatives. Of McKinsey’s 25 top-ranked indications, more than 70% had positive validations, and of the top 50 indications, 90% were supported by literature or trial evidence, confirming that scientific AI can help in the hunt for further applications of known MoAs.
AI and RWD can improve the design of clinical trials. For instance, they can help refine the target patient population by discovering patient subgroups, refining eligibility criteria, and identifying patients who are unlikely to benefit from the treatment. One early-stage biotech was able to better identify patients whose disease was likely to progress faster than the time frame of a clinical trial. They were then able to design a trial that would bring those patients benefits in their time frame.
These advances can shorten the timeline of trials, bringing treatments more quickly to patients who need them. AI and RWD can also help refine clinical trial endpoints and how they are measured. For a rare disease trial, one biopharma company replaced an existing endpoint, which was an infrequent event, with endpoints that occurred more frequently and could be measured with blood tests, reducing the length of their trials by 15% to 30%.
AI is making solid progress in biopharma R&D, but there are significant challenges for organizations that want to take advantage of it. While data is much more available than a few years ago, it requires significant effort to validate it and integrate different datasets. It is also challenging to bridge the gap between technical AI capabilities and tangible business outcomes, and the benefits can’t be realized unless the AI is integrated into existing workflows.
Successful AI implementations depend on rewiring the entire R&D process, adjusting how change management is done, and working together with scientists, data scientists, and expert translators. These translators can help bridge the gap between technology, science, and business. After all, it’s never just the technology on its own.
“Scientific AI is helping researchers develop new treatments faster, more efficiently, and with a higher probability of success,” says McKinsey senior partner David Champagne. “And it is helping them bring life-saving treatments to patients sooner.”
Note: This article was created by McKinsey & Company.
For instance, one leading pharmaceutical company is using AI to accelerate the preclinical design and validation of gene therapies. From millions of potential molecules, AI helps researchers identify more potent candidates in a quarter of the time it took previously. AI also reduces by a factor of three the time it takes to design the vector to carry the active ingredient into the patient. Finally, it accelerates and improves the process of translating the laboratory product into a therapy for clinical trials: It uses a database that includes the results of translational experiments to help researchers predict the clinical outcomes of different molecule and vector combinations.
In each of these three stages, researchers use scientific AI to simulate experiments before they are conducted in the lab so that they can optimize parameters and focus on the most promising options. They then feed the results back to the model. This closed-loop approach continuously refines the experiments, speeding the path to clinical trials.
Scientific AI can also analyze real-world data (RWD) to identify new uses for existing mechanisms of action (MoAs). McKinsey worked with a client to test this approach for PD-1 inhibitors. It analyzed combined horizontal health record and claims data, cleaned of PD-1 data, for 3 million cancer patients, to identify 4,000 features in categories such as procedures and prescriptions. Then the company used ML to glean from a handful of indications susceptible to PD-1s, such as melanoma, the most relevant features. The ML then assessed other indications and created a prioritized list of candidates for treatment with PD-1s.
To validate the results, researchers defined those where PD-1 inhibitors were already approved as positive examples, and those where multiple Phase 3 trials had failed as negatives. Of McKinsey’s 25 top-ranked indications, more than 70% had positive validations, and of the top 50 indications, 90% were supported by literature or trial evidence, confirming that scientific AI can help in the hunt for further applications of known MoAs.
AI and RWD can improve the design of clinical trials. For instance, they can help refine the target patient population by discovering patient subgroups, refining eligibility criteria, and identifying patients who are unlikely to benefit from the treatment. One early-stage biotech was able to better identify patients whose disease was likely to progress faster than the time frame of a clinical trial. They were then able to design a trial that would bring those patients benefits in their time frame.
These advances can shorten the timeline of trials, bringing treatments more quickly to patients who need them. AI and RWD can also help refine clinical trial endpoints and how they are measured. For a rare disease trial, one biopharma company replaced an existing endpoint, which was an infrequent event, with endpoints that occurred more frequently and could be measured with blood tests, reducing the length of their trials by 15% to 30%.
AI is making solid progress in biopharma R&D, but there are significant challenges for organizations that want to take advantage of it. While data is much more available than a few years ago, it requires significant effort to validate it and integrate different datasets. It is also challenging to bridge the gap between technical AI capabilities and tangible business outcomes, and the benefits can’t be realized unless the AI is integrated into existing workflows.