AI is a Perfect Match for Complex Life Science Challenges

While artificial intelligence (AI) technologies have been around for several years, their rapid development lately has sparked questions about how to best use these tools for complex lab work, like new drug discovery. Chris Bouton, PhD, senior vice president and head of AI at Certara has over 20 years of experience in developing software and data analytics. He is driving ways to apply AI and deep learning to solve complex problems in life sciences. In this interview with Lab Manager, Chris offers insight into how lab managers can better understand how to apply AI to drug discovery.

Q: What is your vision of what AI can accomplish in drug discovery?

A: Generative pre-trained transformers (GPTs) [a type of large language model, the same technology that ChatGPT is built on] can accelerate drug discovery and innovation in the life sciences industry. Generative AI (genAI) can save time and money across the lifecycle from early discovery to clinical trials through regulatory submission. Current genAI use cases include screening literature, validating target identification, and data analysis. These examples serve to reduce manual labor and provide context to datasets. As AI continues to become integrated across labs and researchers become more comfortable with using the technology, I believe AI has the potential to expedite the identification of new drugs, ensure faster and more transparent regulatory approval and data-sharing, and ultimately bring life-saving medications and cures to patients faster.

However, to be effective and prevent AI hallucinations, these GPTs must be purpose-built for life science use cases and trained on biomedical datasets to enable specified output.

Q: What do lab scientists need to know about incorporating AI into their research?

A: With all the hype surrounding AI, lab scientists must first understand what AI truly is and its advantages relative to previous generations and types of technologies. Understanding AI’s general capabilities is the foundation for contextualizing its role in lab research. For example, pattern recognition is one of AI’s greatest strengths compared to the human brain or past technologies and has great potential in a laboratory setting. Comprehensive knowledge about [the] types of data a lab generates, and the privacy and security of that data, are also critical when determining how to best incorporate AI into the lab setting.

Q: What improvements are needed in AI tools to enable better drug discovery?

A: A common roadblock when using AI is not receiving the output one is looking for because the GPT isn’t being asked the right questions. That’s still a limitation of AI and always will be. Its output is only as good as its prompts. That makes it reliant on the ability of the researcher to ask the right prompt to coax the right answer out of the data.

Currently, in our industry, we cannot deliver data at scale to these algorithms in most lab environments. For example, many electronic lab notebooks still don’t have effective AI integration. There needs to be more awareness and training in the life sciences industry around how to surface data effectively to these technologies and ask the right questions.

In the context of a specific lab, that principal investigator is often highly averse to their data leaving the lab’s firewall. Therefore, we need to deploy specialized AI securely. All these aspects must come together to enable better drug discovery using AI.

Q: What limitations does current AI technology have that lab managers should be aware of?

A: AI has plenty of limitations. It is a useful, new, and exciting tool, but at the end of the day, it’s not magic. Users need to understand both its value and its drawbacks. A good example of this is tabular content. AI is still not adept at figuring out what was inputted into a spreadsheet without any context in the same way that a human can use context clues and make assumptions.

However, it’s important to note that focusing only on the limitations and not recognizing the value of this technology is equally hindersome.

Q: What is your suggestion of how to best incorporate AI in drug discovery workflows?

A: Understand what kind of data the lab is processing, and then understand what AI can do that is paradigm shifting. If you’re in an image-heavy lab, and you have a problem with object recognition in image classification, then AI has great potential to support your lab well beyond what was previously possible. If instead, all your lab’s data is in tables without any contextualization of those numbers, you must first figure out how to add metadata to that dataset.

Q: What are some leading indicators that AI is struggling to deliver in the lab?

A: Leading indicators that AI is not working include lack of clarity about what problems AI can most readily solve in the lab, the type of data needed to solve those problems effectively, and clearly defining what successful outcomes look like.

Q: What do lab researchers need to learn to effectively use modern AI tools?

A: Many misconceptions may hold lab researchers back from using AI to its full potential. First, you do not need to be a programmer to use AI. Don’t shy away from this technology because of apprehension around the workload associated with integrating a new tool and/or learning a new method. Plenty of fit-for-purpose platforms can be deployed as low-code/no-code capabilities within labs.

As discussed above, researchers must understand what AI can do for each type of data they are working on within their specific lab setting. They also need to learn and implement secure ways to deploy these technologies. Finally, recognize that GPTs are not specifically the product of a given company like OpenAI, and that there are plenty of ways to deploy GPTs.

Q: Are there any key laboratory needs that AI will finally unlock?

A: Science is inherently a "needle in the haystack" pursuit, as opposed to something like a web search which is inherently a "most common answer" pursuit. The goal of drug discovery is to find rare information – the unique insights that no other lab has identified. AI can access that rare information. AI uniquely handles data, making it excellent at unlocking these rarities.

GPTs are not just some form of a probabilistic character generator; they’re something more like a compressed worldview of what they’ve learned about. That compressed worldview has threads in it of rare information. When you’re specifically seeking out that kind of intelligence, you can find it more effectively than what’s been possible previously. Drug discovery and AI are a perfect match.