[Abstract neural network image from Firefly]
At the beginning of the year, IQVIA announced it had forged a deal with the AI giant NVIDIA to develop agentic automation of “complex and time-consuming workflows across the therapeutic life cycle,” as an announcement noted.
When asked about the backstory of the collaboration with NVIDIA, Shankar said the tech giant reached out to IQVIA because “we are the largest provider of not just clinical trials, but almost all kinds of services to life sciences companies,” he said.
At present, IQVIA is building out a variety of agentic approaches with its partners. “One of those is going to go into production soon, and then at some point towards the end of this year, we’ll deploy some of these agents into workflow,” Shankar said. Areas where agentic AI holds promise in clinical trials include spotting potential compliance issues early on, analyzing enrollment rates, identifying enrollment delays tied to specific sites and adapting as necessary.
The economics of better and faster with AI/ML tools
Raja Shankar
The applications of artificial intelligence, whether agentic AI or more traditional ML, are so multifaceted in life sciences that little, in theory, is out of its reach, Shankar said. “Everything that we do, you can use AI to make it better, faster, better quality, and so on,” he explained.
With so many potential uses, it is helpful to focus on the economics of the AI project, including, of course, the ROI. “What is most important to focus on is not going to come from the fact of ‘is this something which we can improve with AI?’ It’s going to be more where is the most bang for the buck that we would get.”
Shankar sees content generation as one of the first areas where genAI can help. “Content generation of all kinds of things — informed consent form generation, CSR [clinical study report] generation, proposal generation, you can do literature reviews — all of these things are content generation where there is a huge value from AI,” he said.
With content creation in a highly regulated industry, however, it’s not a matter of simply giving a bot a complex assignment to finish in one go. “Especially with agentic AI, you’re not going to write the whole content in one go.” Shankar said. “For any content, you need source material, you need to extract information, author it for different sections, and have another agent check the previous agent’s work.”
While AI can accelerate many individual tasks in clinical trials, Shankar points to a fundamental constraint that technology alone cannot solve: patient recruitment. “If there are 10 companies doing the trial in the same indication, they’re all competing for the same patients,” Shankar said. “It’s a finite pool of patients… you cannot create new patients.”
Keeping genAI systems on track
A core technical challenge for generative AI has been the risk of “hallucination,” or generating false information. For high-stakes environments like clinical trials, accuracy is non-negotiable. Shankar explained how this can be managed. “The hallucination piece, partly is to a large extent solved if you just say, ‘Look at the source of information and only provide me information or do the analysis based on this source of information,'” he noted. By grounding the AI in specific data and adding verification steps and adding scaffolding with techniques like retrieval augmented generation, reliability increases dramatically. “With that, you can actually get these to work very well.”
Shankar notes that IQVIA rigorously evaluates the quality of the output of such systems before deploying them. “If you write 100 CSRs, we have to have benchmarks that 99% of the time, or whatever the benchmark is, we are getting it done the right way,“ he said. “Plus then we also need the human in the loop, because it might give us a very good draft — say 70% of the way there, even with agentic. And then you have the human in the loop that does the final test to make sure it’s appropriate. You don’t have it 100% automated, but you really have it significantly automated. Then the human can do more than what they’re able to do today.“
Toward human/machine ‘dream teams’
While keeping a human in the loop is important for AI workflows, Shankar also touched on the promise of including humans with increasingly sophisticated networks of AI and ML models. The collaboration between human and machine thus becomes more multidimensional. For instance, agentic frameworks can invoke other AI/ML models to create a multi-agent framework, a sort of “dream team” combining diverse AI models with human expertise to address longstanding R&D challenges more comprehensively.
Such “dream teams” can be complex to orchestrate. “With generative AI or foundation models, instead of having one model for each task, we are able to leverage these models for multiple tasks,” Shankar explained. “We are exploring the use of generative AI models to even do some of those predictions with few-shot learning that may be as good as using a traditional single task machine learning model.”
The ultimate goal: Multimodal foundation models
Beyond accelerating individual tasks, the convergence of diverse data types through multimodal approaches promises to transform how we understand disease biology and predict treatment outcomes. Multimodal approaches are fundamentally suited to life sciences as biology is inherently multimodal. For instance, genes affect proteins, which affect cell function, which affects tissues and organs. And at a macro-level, a host of factors such as demographics, environmental conditions, treatment history, and real-world clinical practices influence clinical trial outcomes. On top of that, there are a variety of imaging modalities, electronic health records, real-world evidence, and patient-reported outcomes. “If you’re able to combine all of this together and create a multimodal foundation model across multiple types of biological data at scale, that will also be something that could potentially change the game,” Shankar said.
IQVIA has been building machine learning models for over 10 years, using approaches like XGBoost and random forest for a host of applications. “We use those models to build models for country and site selection, enrollment prediction,” Shankar said. IQVIA also uses traditional ML to inform clinical trial design. “If you want to find likely responders to different mechanisms of action, we can build machine learning models to find responders,” he added. Such machine learning models can also help to find people who are likely to progress from one treatment to another treatment and forecast clinical trial success, regulatory approval likelihood and commercial potential.
Now, IQVIA is integrating generative AI approaches with individual machine learning models ranging from random forest tree-based approaches to neural networks. “What’s changed with generative AI or foundation models is, instead of having one model for each task, we are able to leverage these models for multiple tasks,” he said.
IQVIA is also exploring how generative AI might not just complement but potentially replace traditional machine learning approaches for some regression or classification tasks. “It’s early days yet,” Shankar added. “But we are seeing some promise where you can even use these generative AI models with few-shot learning to make predictions that may be as good as using a traditional single task machine learning model.”
Even with buy-in, the journey from a working model to a scalable, enterprise-grade tool is a major undertaking. A successful demo is one thing; a robust, production-ready system is another. “What happens in a lot of companies is you will have some really good data scientists who would mock up something and build something that’s very good and works very well. But then if you need to deploy it and give it to 1,000 users in the company or externally, then you need the entire software engineering around it.”
Filed Under: Data science, machine learning and AI
