AI and the dietary supplement industry: A conversation with Brightseed's Jim Flatt

Nutritional Outlook (NO): Can you talk about the processing and energy demands for using AI with regard to R&D/product development? What efforts are being made to ensure that stability and sustainability of using AI as the energy requirements increase?

Jim Flatt (JF): So if we look at how we're using AI at Brightseed, the energy intensiveness really comes in two areas. The 1st is actually in a building and training the models where again we are taking hundreds of thousands or more data points and then going through an iterative process to build the best AI trained models.

We also generate computationally intensive tasks and actually making [with] them the predictions. So we build all these models in order to be able to identify bioactive small molecules that are present in all of our dietary plants and microbes for proven health benefits.

And so, a couple of things, just to keep in perspective, at least at Brightseed, we're not using these large language models that literally train on hundreds of millions of data points. So, the relative amount of energy that we use is much smaller than you may see.

Having said that though, again, we do use a component of efficiency to figure out the shortest way and least amount of computation required to get to an answer. So, over time, we iterate our models and we do measure their performance, including how much time it takes, which corresponds to really the energy usage. This is a way that we continue to improve the productivity and the energy efficiency with our use of artificial intelligence methods.

NO: When it comes to generative AI, people talk about hallucinations. Is there anything comparable that when it comes to using machine learning for finding new compounds and related health benefits for developing new products? How do you weed out bad data if there is any?

JF: So, the source of hallucinations, which are again things that are spurious or not real really are derived from poor quality training data. It is that old axiom of garbage in, garbage out, andof course this is what leads to those hallucinations. In our case, a hallucination would be, you know, equivalent to a series of poor predictions that just aren't true, compounds that in fact won't have those effects. And so the way we guard against that is we have a three-step process of curating, validating, and human oversight at each step of the process.

That enables us to avoid that problem and so to be a little more granular in the curation process, we actually use statistical methods that are overseen by trained scientists. So, decades worth of experience and deep knowledge of the field to identify outliers or spurious data that could be that garbage that we don't want to filter into our model.

Once we do build the models and then make our predictions, we do have human oversight there to look at the predictions and say, “Hey, do these or do these make sense based on my knowledge as a scientist ?” And of course you have to be open too, sometimes you do get something unexpected that's good. But oftentimes, some of those very unexpected things can be again the result of inaccurate data. So in our early days we went through, you know, several cycles, iterations to test, to build our models, test them, refine and iterate and go through that builds-test cycle again.

The next step we do and I think what’s important is we actually do have a built-in validation step where we actually test the health effects of these compounds through what are called in vitro cell assay models. So, we will make a prediction of a compound that might promote deep restorative sleep through a specific biological mechanism and we can then actually take the compounds that are predicted to have that activity and actually test them out in the lab and this is our way to establish which of those predictions are in fact true and have the desired activity. So again, a second step where we weed out anything that may not be accurate and then of course, as I mentioned at each step of the process, we have human oversight by individuals trained in those steps and in that underlying science, which again just provides that assurance that we're building AI models that will lead to the answers that we want. In fact, will ultimately be proven to improve health and wellness.

NO: How can AI benefit the dietary supplement industry as a whole?

JF: I think [AI can benefit] three distinct areas. A we know, a lot of the history of the industry really originated with many complex plant, herb botanical sources that form the basis for traditional medicine practices around the world.And that has also given rise to very specific botanical herb and plant extract products you see on the market. But one of the challenges with those products is we have a very incomplete understanding of those products just as we talked about the cranberry and as such, because these are natural products, if we don't really know the components, we don't really know all of the compounds that can affect health and how growing conditions even can change the compositions of those products.

We often see conflicting clinical data about the benefits and do they work or do they not. And we believe that often times it's not that these products don't work, but rather that because of our lack of knowledge, it’s very difficult to quality assure them and ensure efficacious levels of the beneficial component. So, one opportunity here is to create superior, you know, botanical, herb and plant extract products for health benefits that will be not only superior in efficacy, but also give a much more consistent response and experience to consumers.

The second use case: In the industry we have a number of these, what we'll call hero ingredients, that we build products around and these are specific components that are oftentimes present in some of these plant, herb, and botanical sources.

For specific health benefits, one example that you know well is DHA omega-3 for clinically proven benefits for brain and eye development. And the question is what's that next sort of hero ingredient that's really going to get consumers excited about products and the potential of this industry. And so using Forager AI, we're able to start with an end health benefit, say deep restorative sleep as that has been announced with our partner Pharmavite, and we're able in this process to identify compounds that work through pharmacologically proven modes of action or processes in the body that these natural compounds can interact with and ultimately, in modulating that process, can yield the benefit and in the consumer case this is promoting deep restorative sleep.

That is the biggest unmet need of existing products such as melatonin for example, and so here we're able to generate essentially new to the world hero products that are in fact sourced from nature’s intelligence, ones that work through proven modes of action that actually will be better accepted by the medical and healthcare practitioner communities because we actually can show: Here's the product. Here's how it works. Here's the clinical proof of benefit, which gives them reasons to believe and reasons to recommend these products, perhaps instead of much more expensive therapeutics that may potentially even have some side effects or be just something people don't want to take.

And then the final aspect, and this is the new frontier, is personalized nutrition and personalized health. With it being so inexpensive now to understand your own genome, and understand what makes each person truly genetically unique, we're able to then through that knowledge, make predictions about products that will actually work particularly well for a given individual or groups of individuals that share some common genetic features. And in this way, it really offers the potential to deliver science based personalized nutrition products that hopefully will give individuals a better experience and better meet their health and wellness benefits.

So, these are the three use cases in the at the end of the day the proposition here is: these methods including, Forager AI, really are the ticket to really accelerating innovation, new product innovation in the industry, being able to do it at a fraction of the cost of traditional methods and ultimately give greater surety and probability of success in ending up with a product that actually makes it to market and delivers on its promise.