CAS approaches artificial intelligence with a level head, ‘not a magic wand’

A full room of conference attendees and members of the ACS Board of Directors gathered for Sunday’s open board meeting in Denver to hear CAS’s assistant director of strategy, Andrea Jacobs, talk about using artificial intelligence (AI) to enable science.

“AI is a very important topic for us.” Manny Guzman, president of CAS, said in his opening remarks. “It is an active dialogue in our organization today.”

Jacobs opened her talk by gauging the feelings of the audience toward AI. She used words like “potentially exciting,” “scary,” “challenging,” and “not sure what to think.” She also acknowledged that it would be impossible to resolve all the tension around AI in the length of her talk but hoped that specific examples of how CAS is applying AI would help.

CAS is currently applying AI to three areas: internal operations, customer-facing solutions, and customer AI initiatives.

Internal operations: Bioactivity

Last year, CAS announced that it was expanding its work to support the life sciences, which would include exhaustive biological sequences and bioactivity information—an effort to accelerate early stage drug discovery research. With that expansion comes the need to curate new and different information. Specifically in Jacobs’ example: bioactivity. When CAS begins to curate new information, they often look back at what work has been done previously: what scientific concepts, chemical substances, or reactions were defined when a paper was originally curated. CAS then uses its sophisticated search technology to find other papers based on the identified information that might have the terms related to bioactivity in them.

This approach resulted in 40–50% accuracy in finding a paper that included bioactivity. When CAS put the same information into an AI model, they were able to improve the results to 80–90% accuracy.

Customer-facing solutions: CAS BioFinder

In May, CAS launched the CAS BioFinder in an effort to support medicinal chemists in analyzing drug-target interactions and optimizing planning for drug discovery projects. As Jacobs describes, BioFinder offers an AI tool that is a predictive analytics panel.

CAS acquired the BioFinder AI tools through a partnership, and the tools were originally trained on information mostly from other public sources. CAS saw this as an opportunity to marry the highly structured data that CAS has compiled to these AI models to see just how much better they could be, Jacobs said.

“We saw really phenomenal results,” Jacobs said. “Fifty-six percent improvement in prediction accuracy and a dramatic reduction in noise.”

“The lesson here is that AI isn’t really just about the technology, AI is really about the data,” Jacobs said. CAS curates that data in a harmonized, structured way, making this a great example of the power of using data in this manner.

Jacobs acknowledges that, as scientists, there’s a high threshold for a tool like this to meet one’s willingness to use it: “If it’s spitting out garbage, that’s not going to get you anywhere.”

Customer AI initiatives: Retrosynthesis tool

Retrosynthesis is an important part of a chemist’s job.

“All of us at some point need to make something or have someone who works with us make something, and we need to figure out how to do that,” Jacobs said.

The retrosynthesis tool in SciFinder allows a chemist to use CAS’s reaction data and predictive technology to obtain a suggested route or a set of synthetic routes to a compound. “We know that retrosynthesis can find candidate paths, but more importantly, we need to be able to find the best candidate path,” Jacobs said.

The retrosynthesis engine uses a hybrid approach: search and pattern recognition from CAS literature and AI-based methods for certain aspects of what it’s trying to accomplish—like helping to predict something that is not empirically known.

AI in the scientific industry

AI is everywhere.

“We’re seeing everyone in the scientific industry play with AI in some way,” Jacobs said. “If there’s a company you can name that is not experimenting with AI in some way, I’d really like to know because I’m not sure I can come up with one.”

Jacobs divides the companies that are experimenting with AI into three big categories: content experts like CAS, algorithm experts like those in big tech, and domain experts like those in Big Pharma. The categories represent different backgrounds and skills sets, all of which are relevant to AI as a challenge in the sciences.

CAS thinks of these categories in the context of what is known as the “triangle of success,” Jacobs says.

“If you’re doing AI—and not to dismiss the sophistication of ChatGPT, but it doesn’t require science—there is no specialization needed to power that type of an algorithm; it’s really mostly a mathematical problem and a generic data problem,” Jacobs says.

When it comes to scientific challenges, Jacobs says that success requires the three big categories of expertise. CAS has a paper forthcoming in collaboration with Selvita that showcases this notion.

“The thing that we’re thinking about is valuing innovation but also keeping level heads,” Jacobs said.

She revealed a diagram of the Gartner hype cycle to explain her point. She explains that the AI we’re seeing all over the news with all the hype around it is at the beginning of that diagram, leading to “inflated expectations” on the hype cycle. “This tends to be a specific type of AI, like large language models or ChatGPT,” Jacobs said. “However, there are other types of AI technologies that aren’t as much in the news that are on the productivity part [of the diagram] . . .”

CAS is thinking about AI cautiously—distinguishing between the buzz and the reality. “Yes it’s important to listen, yes it’s important to stay aware, no we can’t get caught up in the hype,” Jacobs said.

A lot of the hype is that AI is going to be magic and solve everything, Jacobs said. What she sees is a need for us to understand the problem we’re trying to solve or the need we’re trying to address—and to let that come first, and let the technology come second.

“There are lots of choices for technology, but we can’t choose the technology because we like it, because it’s buzzy. We have to choose the thing that is going to solve the problem,” Jacobs said. “It’s not a magic wand—it’s not a throw everything at AI, and AI will solve everything.”

Making AI successful in the sciences requires a clear understanding of a problem, an understanding of the scientific information that will support solving that problem, and the ability to bring that together.

“AI can do a lot of things, but it can’t do a lot of things without us,” said Wayne Jones Jr., chair of the ACS Board of Directors, summing up Jacobs’ talk and the Q&A.