Machine learning and market fit in health care

Molly Gibson, PhD, an expert in computational biology and computer science and a principal at Flagship Pioneering, discusses infusing innovative AI into health care startups.

Twelve colored AI icons

Overnight success is applauded in many industries, though not typically in health care, which rests on a sturdy foundation of painstaking science. Scientists are often viewed as measured and analytical to a fault. Of course, that overlooks the deep creativity and upending curiosity behind breakthroughs like the mRNA vaccines from Moderna and Pfizer/BioNTech that pushed forward at unprecedented speed a worldwide effort to stop the worst ravages of SARSCoV2 — all while being built on a very sturdy foundation of more than a decade of mRNA research.

Flagship Pioneering, the biotech innovation company behind Moderna, describes its visionary style in directing scientific innovation and entrepreneurship as a journey from unreasonable propositions — deliberately cultivated, then thoroughly tested — to transformational outcomes. More simply, Flagship cultivates outrageous “what ifs” and nurtures the most promising, eyebrow-raising “it turns out” responses.

Molly Gibson, PhD, a principal at Flagship Pioneering and co-founder at Generate Biomedicines, is an entrepreneur and scientist with expertise in computational biology and computer science. She answers our questions on creating successful AI health care start-ups and the approach Flagship Pioneering takes toward venture projects.

Edited and condensed for clarity

What are the keys for an AI company to differentiate itself and find product-market fit in health care?

Typically, AI and machine learning, or ML, are interchangeable in the way that we’re using them here. The implications of machine learning and biology in health care are remarkable. We’ve only begun to see what’s going to be possible over the next 10 years as biology transitions from its current state of exploration and discovery to one where we can actually engineer and predictably influence the biology of medicine. There’s a ton of white space, but there’s also a ton of noise. So, one major challenge is that investors, partners, talent — all of the people you need when you’re building an AI health care company — need to understand and be inspired by your story. Platform-product market fit is essential. You need to be really crisp and clear on where your platform creates value. What’s the value proposition you’re building in the product you’re creating? Whether it’s a diagnostic, a therapeutic, a piece of software, or whatever you plan to use to impact the market and ultimately people and patients, you have to know how machine learning affects that.

We think about a platform as a core technology that multiple applications can be created from. This is incredibly common in technology fields — think Apple and the App Store — and it’s becoming more common in medicine and therapeutics. The concept is simple: what technology will give you multiple verticals of opportunity and multiple applications? Be crisp and quantitative on how your technology, your machine learning capability, truly transforms the way you do something versus improving it incrementally.

How does company creation and creating value for AI startups differ in health care compared to other sectors?

I think the biggest difference is the data and the depth of domain knowledge required. AI drove really significant transformations that people are familiar with in technology, commerce and advertising — you know, Google’s ability to pinpoint what you ate for breakfast and then place that bowl of Cheerios in front of you. In those types of industries, more data always seems to be better. Even if there’s a reasonable amount of noise in the data, you can often collect enough that you wade through that noise and are able to identify a signal. Also, in these industries, data has temporal context: what you like today might not be what you liked six months ago, or what you will like in two years. So, the updated information is really important.

Biological data is incredibly complex and hard to collect systematically. It’s really hard to get a passive flow of data like Google can with the millions of searches going on a daily basis. With health care data, there are significant regulations. Even if you had access to these types of data sources, often there are limitations on how it’s aggregated and how you can search or learn on it.

When you move into biology, it’s really important that you have the domain knowledge to understand what you’re modeling. There are a lot of things you don’t actually know how to learn because you don’t know if the right level of prior knowledge exists. For example, we know proteins start as a 1D sequence of amino acids, then fold into a 3D structure. How the 1D sequence of a protein encodes for the underlying three-dimensional structure is a longstanding challenge in biology, which DeepMind tackled through machine learning with AlphaFold. The team needed to infuse the right amount of biologic priors into the machine learning models in order to be able to solve that problem. That breakthrough didn’t come because they applied standard machine learning; it came because they understood the biological system in a really deep way and included that in their model. It’s our ability to understand, connect the biology, connect the machine learning, and connect the data together that matters.

Can you tell us about Flagship Pioneering and its approach to venture projects?

A lot of what we spend our time doing is exploring the extreme white spaces, trying to look for novel insights, essentially by connecting fields that hadn’t been previously connected or identifying the most unbelievable component. We explore ideas, we test them out, we’ll refine them and learn something new. And in this exploration process, some ideas become the seed of a prototype company, or ProtoCo, that we fund.

ProtoCos get a small amount of money to do key killer experiments. Depending on strengths and weaknesses uncovered, this may give us the confidence to fund the company — now called a NewCo — which will continue to grow and be funded by Flagship. All of these companies are wholly owned by Flagship, which allows them to hire almost entirely scientists during their first few years. Traditionally, a Flagship partner will be a CEO of a company, which permits them to maintain an innovation mindset for quite a while until they’ve really grown up. We raise external funds and continue to build them. So, Flagship is really an ecosystem of companies more than any individual company, which allows for lots of collaboration across a range of industries and technologies.

— Francesca Coltrera

Continue the conversation by joining Molly Gibson at the upcoming Designing and Implementing AI Solutions for Health Care program, or connect with us on Twitter @HMS_ExecEd or with Molly Gibson @gibsmk.