Consultant sheds light on the AI hype machine

Artificial intelligence (AI) applications in the natural products industry run the risk of being overhyped and misused, a long-tenured consultant says.

A recent United Natural Products Alliance webinar focused on intellectual property issues raised by the use of AI tools. One of the key questions the event delved into was how much additional work a human must contribute before an invention belongs to the inventor and not to the developer of the AI tool the inventor used.

During the event Tim Avila, a long-tenured product development consultant with Systems Bioscience, voiced some doubts about how AI tools were being employed in some corners of the industry.

AI is current hot topic but idea itself is not new

Artificial intelligence recently has become a buzzword within the dietary supplement industry (and many others). It can connote an aura of being in step with the times, of a company that is using the latest development tools available to create better new products more rapidly and reliably.

But the idea is hardly new. The notion that a computer could use available data and apply reasoning to derive new connections and ideas was first postulated by English mathematician and polymath Alan Turing in a famous 1950 paper.

The term ‘artificial intelligence’ was coined in 1956 as the title of an academic conference on the subject at Dartmouth College. At that conference, the first AI program, called Logical Theorist, was presented. The program, designed to mimic the problem-solving skills of a human, was funded by RAND Corporation. 

From that start, AI and machine learning (an often-used synonym) programs have developed to the point that companies in many industries are starting to view them as indispensable.

AI has led to some notable blunders

There have been, however, been some notable debacles using the technology. Real estate listings site Zillow had to write off more than $300 million and lay off more than 25% of its workforce in 2021 after its "Zestimate’" appraisal tool, based on an AI algorithm, overestimated home values. That led the company to purchase a bevy of "fix-and-flip" properties at prices too high to make the transactions profitable.

Other issues have focused on the HR realm, with bias against minorities, women and older people creeping into the hiring process when AI tools were involved.

How supplements use AI

Within the natural products field, several companies have gained prominence for their expertise in using AI tools. Earlier this year, startup microbiome products company Verb Biotics partnered with Evogene to use its AI platform to speed new product discovery. [Natural Products Insider toured Verb’s facilities in the research triangle in North Carolina. See our coverage here.] Brightseed and Nuritas are two additional examples of companies putting the application of AI front and center in their R&D process.

Avila says AI is a way to leverage already powerful tools to discover new compounds and new metabolic pathway targets for those compounds.

“There has been an evolution in computation biology,” Avila told Natural Products Insider in a telephone interview.

“Computational biology has been used in drug development in the biotech world and in our space,” he said.

For an example of how AI can supercharge that process, Avila noted he is working with a company in the protein space. With more than 37 million possible protein structures to sort through, he said it’s obvious that an AI tool can sort the proverbial wheat from that chaff more quickly than a human could.

However, that output cannot be taken for granted, Avila said. What he’s concerned about is some marketers who might either be less well versed in the inner workings of AI or perhaps overly enthralled with how it can cut costs. Some errors might creep into the process unnoticed. AI is powerful, novel and fast, but it’s not infallible, Avila noted.

AI conclusions need verification

A German academic group focused on AI recently noted that when AI is used in drug discovery, it can generate results that might be correct, but at the same time are also inexplicable.

Drug companies have used AI tools called graph neural networks (GNNs) to predict how strongly a given molecule might bind to a certain target protein. But when the German researchers tested six different GNNs, it was not clear how those programs arrived at their conclusions.

“How GNNs arrive at their predictions is like a black box we can't glimpse into," said Prof. Dr. Jürgen Bajorath of the Universität Bonn in a post on the university’s website.

The university has developed an analytical tool dubbed EdgeSHAPer that helps shed light on how GNNs analyze and employ chemical data to make the decisions they do.

“The development of methods for explaining predictions of complex models is an important area of AI research,” Bajorath said. There are also approaches for other network architectures such as language models that help to better understand how machine learning arrives at its results.”

Bottom line: AI can’t supplant humans – yet

Avila said he, too, has been wary of the results that sometimes come out of AI. A highly qualified and experienced human still needs to be in the loop to make sure the output makes sense.

“I’ve noticed significant flaws in their insight and in how they crunch the data,” he said. “I’m in the camp that AI is really good at raising money, but it has been overhyped.”