This article originally appeared in Nutrition Business Journal’s Tech Issue.

AI is zipping past buzzword status and into operational reality for nearly every industry on the planet, but for supplement makers, the biggest opportunities for innovation are only beginning to be explored, including perhaps the biggest question of all: whether AI can formulate new products.

Tony Van Campen is confident the answer to that question is yes. Van Campen is a co-founder of Belgium-based ­Biomdrin, an AI/machine learning platform that is capable, the company says, of “developing cutting-edge nutraceuticals that prevent disease.” Campen says Biomdrin’s AI tool “looks at millions of possible interactions between human genes and protein expressions in a database of about 70,000 known ingredients, and it reads all the research and publications that have been done on these ingredients.” From that, Biomdrin can spit out a formulation that a human would then evaluate and decide whether it’s fit for trial and development.

“It looks at many more interactions from much more data than a human could do,” Van Campen says, “but it still takes the human to make that into something that’s practical.”

Akash Shah’s IngredientAI platform, now in the final weeks of pre-launch, does some of the same literature review, but the Care/of co-founder stops short of saying his new offering will formulate. It is instead a tool that formulators will use. “AI can sort of write those quick notes for you, extract all the things that you want to know and give you something to work with a lot faster than all of the tedious time that you might pour into trying to produce that for yourself,” he says. “But then the formulator is still going to do the formulator’s job.”

Related:I asked Google AI about dietary supplements. It did not go well.

Defining formulation

The question of whether AI can formulate supplements is actually easy to answer. Go to ChatGPT and ask it to design a dietary supplement for immunity, and you will get a formulation that includes vitamin C, quercetin, a probiotic and elderberry with dosages in milligrams and directions for use. The question of whether that formulation is new, or even particularly innovative, is also easy to answer: it’s not. More sophisticated tools like Biomdrin’s give more sophisticated answers, but Sebastian Balcombe, a medicinal chemist and the founder of ingredient innovation-focused Specnova believes the answer is still no. The AI applications put to use by ­companies like Biomdrin and IngredientAI are hardly new, says Balcombe. The machine learning technology behind much of what is presented as AI has been commonplace in pharmaceuticals for years, he says. To Balcombe, that’s not innovating, it’s prospecting. “That’s straight mathematical computation type stuff,” he says. “Over years and years, it has become smarter and better at figuring out these hits, but understand, those are just hits and not even leads.”

Balcombe isn’t against AI—Specnova puts it to work to search large amounts of data—but he is against overstating its abilities. Not all the hype comes from tech companies, he says. Much of it comes from the popular media. He recalls a “60 Minutes” segment that suggested AI had abilities it has yet to achieve. “They said AI has discovered all these pharmaceutical drugs, and it hasn’t. There’s one drug that it assisted in helping discover, and it’s not even on the market.” Reports that AI had successfully predicted protein folding also fall short of reality, he says. “It’s just so over-exaggerated, the extent of what it can do.”

Hype is a constant across many industries, but when that kind of shiny object falls into the hands of supplement industry, missteps might be more easily assumed, Balcombe contends. “It’s definitely, obviously, the big buzzword, right? Everybody wants to jump in and throw the term AI on whatever they’re talking about.” AI, he adds, is “a tool that can help the human accelerate their formulation process” but not a formulation engine on its own.

Not all AI innovators and advocates would disagree with that assessment. Shah puts further boundaries around what ­IngredientAI can do. “AI is very good at reading lots and lots and lots of text. It’s also very good at writing lots and lots and lots of text, but what it’s not good at is judgment. And it’s also not good at vision or mission.”

Guardrails

For Shah, getting IngredientAI to read “lots and lots of text” in nutrition research papers was step one in the “experiments” that proved the idea had merit, but it had to go further before the company felt comfortable describing it as tool for formulation. Among the most notorious qualities of today’s AI incarnations is the tendency to “hallucinate,” stating as fact evidence that was concocted from nothing in the depths of its computational black box. As Balcombe puts it, “It literally fabricates information.” Shah describes ­IngredientAI as AI with guardrails, and one of those guardrails is a built-in fact check feature that allows users to see the citation behind any claim the platform produces. “You want to make sure that there’s essentially quality assurance in addition to the human in the loop, so that people aren’t blindly making decisions.”

Whether those or any other guardrails are high enough or sturdy enough could become a question. Chris Lockwood, vice president of scientific affairs at Nutrabolt, has read enough scientific papers to know that simply linking to the citation could still leave a formulator wandering down the wrong path. Separating the good studies from the bad can take a practiced, and very human, eye. “How robust are they? What kind of risks do we have using these studies to support our claims?”

Balcombe also sees AI stumbling over the finer points. “There is so much nuance and interpretation in clinical research,” he says. “I would not be having AI summarize clinical research results and statistics at all.”

Van Campen says Biomdrin saw an immediate need for a human dose of “common sense” in formulation. Offering the hypothetical of an AI formulation calling for 5 grams of garlic extract, he says, humans will see “practical” realities more clearly than AI. But Biomdrin also built the ultimate guardrails into its value proposition. The company doesn’t just conceive formulations, it tests them, moving from theoretical to in vitro on the way to clinical. AI absolutely needs to meet the real world in the lab, he says. “Let’s iron out those mistakes and come to a point that we are 99% sure that what comes out makes sense and will run successfully into clinical trial. And that’s the only way we want to take products to market,” Van Campen says. “When we formulate products, we put them through clinical trials.”

The result, Van Campen insists, will be a product that “actually works for you because it’s based on deep insight and not just, no disrespect, a random product formulator learning that ashwagandha is popular.”

Reality check

What Van Campen describes—products moving from theory to practice with clinical trials as a mandatory step—is clearly not the rule in the U.S. supplement industry. Clinical trials for finished product formulations are far from routine. Veteran ingredient innovator Lauren Clardy, a consultant at Nutrimarket Business, would welcome AI-driven formulations that are both novel and clinically validated, but what she expects to see from AI formulation isn’t too different from what she sees so often: formulations casually thrown together and then copied over and over again. “It will be the same, just faster,” says Clardy.

She points to “so many brands popping up on Amazon that we’ve never heard of before” and wonders how many of them will skip the step of professional formulation in favor of an AI-generated formulation that could have been cut and pasted from some other nondescript brand. It’s not going to drive anything new. “I think it leads to more sameness on the formulation side,” she says.

Balcombe sees the same prospects. From his vantage, inspiration plays an integral role in formulation and ingredient discovery. “It [AI] takes out the creativity, the human creativity piece, which is the gold when it comes to formulation,” Balcombe says. “That’s how new categories have been created. That’s how things that are really impactful have happened in our industry, not through some huge mathematical screening process.”

Consumer confusion

If supplement executives don’t understand AI’s capabilities, as Balcombe fears, it’s a far leap to expect consumers to see through the inevitable hype. It’s nearly a certainty that “Optimized by AI” will show up on a bottle, perhaps from one of the Amazon-only brands Clardy decries, and it’s a good bet that such a bottle is already rolling off the line or will be by year’s end. Another feature to overhype, in an industry that critics charge is built on hype, is hardly doing consumers and responsible brands any favors. Put simply, Clardy calls AI “an opportunity for shenanigans.”

Shah is not so sure consumers will be hoodwinked. IngredientAI isn’t designed for formulation, but if he did see a brand claiming its products were formulated or “optimized” by AI, he wouldn’t buy it, and he’s not sure consumers would be attracted to it either. “If anything, I think consumers are going to have the reverse response to something like that,” Shah says. “If I saw an AI-optimized formula on a package, I don’t think I would want to pick that up and buy it. I think I would be very skeptical. And I’d wonder what the hell was in it.”

Lockwood is confident he will see “AI” on a label. While less than enthusiastic about the concept, he’s more curious than alarmed. “I can definitely see where it will be marketed that way by certain brands, and I’ll be interested to see what does that brand look like? And who is that consumer that thinks that that’s a better alternative?”

For Balcombe, the hype factor is just more proof that few people understand how AI can work in supplements. “I won’t be surprised when I see it. And you’re right, it probably will be as soon as, like, next SupplySide.”

Real world applications

Nobody, especially not Balcombe, is dismissing AI as having no place in supplements. Balcombe says it’s a tool for analyzing large amounts of data in a mathematical and mechanical way. Lockwood believes it will be good for the “block and tackle” basics of matching claims to evidence. For Shaw, it provides an escape from the time-consuming aspects of product development, something that surprised him in co-founding Care/of. “One of the biggest observations I’ve made, and it’s kind of been shocking to me as a technologist, is how manual the innovation process is for companies like Care/of that are actually digitally native and super technology forward, but then also for companies like Bayer, which I’ve now had a chance to see how they do things, too.” Bayer acquired a majority share of Care/of in 2020. “Everyone innovates very manually,” he explains. “It’s kind of crazy to me.”

Building IngredientAI meant teaching the program to read scientific studies and match ingredients to claims, building dossiers that matched the efforts of human formulators in seconds rather than months. Still, he describes the output more like “bullet points” and not a recipe for a new product. “This is just kind of simple human bullet points where I can very quickly understand not just the ingredient that is available for this benefit, but also the claims that I can make off of this ingredient within just a few seconds.”

Van Campen says Biomdrin also does not purport to create ready-for-the-shelf formulations. The human review of the practicalities and the commercial realities is built into the process. But what he also promises is that AI will keep formulators and brands from chasing the mirage of false promise. It’s not darts on the wall and hunches. The formulations that AI can suggests and the clues and pointers it offers will lead to better and more efficacious products, he insists. With the confidence delivered by AI, companies can make the right bets and spend the money on development and clinical trials to create only the best products. It’s not just discovery, he says, it’s “weeding out the bad from the good.” Fewer SKUs, but SKUs that work.

“I can have an idea today, testing tomorrow and know if I’m on the right path,” Van Campen says. “That’s an acceleration of discovery that I think will be very beneficial for the industry.”

Stay up to date on the rapidly changing industry by subscribing to Nutrition Business Journal.