Some diseases have confounded researchers for decades. But the application of artificial intelligence (AI) to medicine promises to drastically accelerate the discovery of new drugs to treat them. Tackling hard-to-crack challenges and reaching more patients excites Molly Gibson, cofounder and chief strategy and innovation officer at Generate Biomedicines. Based in the Boston area, Generate is a therapeutics company launched in 2018.
Gibson recently spoke with McKinsey’s Christian Fougner and Lydia The about how AI and machine learning can speed up drug discovery—and what it could mean for patients everywhere. An edited version of their conversation follows.
Lydia The: You were part of several biotech firms before you cofounded Generate Biomedicines. Tell us about Generate’s potential for revolutionizing medicine.
Molly Gibson: Generate is focused on transforming drug discovery in the protein therapeutics space. Instead of thinking about how to discover new molecules from natural processes and evolution, we use computers to learn the rules of proteins and how proteins function. We can then generate completely de novo molecules that do the things that we want so that we can create better, cheaper, safer drugs.
When we first started thinking about the idea of Generate, we drew upon examples of different fields that have been revolutionized by the idea that a computer can learn principles of systems, then take those principles and invent new and novel things that never have existed before. An example is in the field of computer vision. A computer can look at hundreds of millions of pictures of human faces and videos of people talking and ultimately learn the principles by which a human face comes together. It can then create images or videos of people that look realistic, like you or me, but don’t exist and statistically never will exist.
We’re excited about applying that to the world of proteins and biologics and therapeutics. We’ve tried to predict biology for decades. But what if you could actually invent new biology? The algorithms can get better over time and are constantly learning. You could then start to program therapeutics versus discover them. Nature has given us this huge set of beautiful proteins to learn from, and it’s given us the rules that have been discovered through evolution. We can learn those rules, extract from them, and generate proteins that have never existed before.
‘More shots on goal’
Lydia The: We know that drug development has a 90 percent failure rate. Why do you believe that AI could improve on this?
Molly Gibson: That’s an incredibly daunting number, and one that I feel we, as an industry, should be obsessed about solving. It’s a complex problem with lots of steps—from selecting the target, to creating the molecule, to identifying the patient populations. We’re seeing AI have an impact across all spectrums of that complex equation. AI will give us an improved understanding of biology, so we’ll be able to identify better targets and then create molecules that better intervene in that biology and are safer for patients, as well as identify the patients who are going to be most affected by those medicines in disease populations.
We’re going to see an inflection point in our ability to learn and optimize for molecules and diseases that we know. We know how to affect the disease, but we just don’t have the tools to create the medicines that we want at the speed that we want. My hope is that we’re going to see an acceleration of molecules that are invented by computers, and that, for those diseases where we understand the biology, we’re going to find ways to treat them more quickly.
One recent example of how this could look in the future is what we did for the development of a vaccine for SARS-CoV-2. We were able to program mRNA molecules to express the specific antigen that our immune system responded to, in order to create protection against future infection. We understood enough about the connection between an mRNA molecule and the protein structure of that antigen. But what if, in the future, we actually know the connection between a DNA sequence and a protein that not only elicits an immune response but treats a disease? We could completely change the lives of patients. Our ability to program biology is going to change the pace, speed, and cost at which patients get new medicines.
What if, in the future, we actually know the connection between a DNA sequence and a protein that not only elicits an immune response but treats a disease? We could completely change the lives of patients.
The combination of all of that is going to start to chip away at the failure rate [for new drugs] that we see today. We’re going to go from a world where we see one or two therapeutics a year come out to a world where you could imagine a tenfold jump—where we could see ten to 20 therapeutics, if not more, that are ready to go into clinical development every year.
Lydia The: The biopharmaceutical industry is focused on a handful of diseases or on small patient populations. People always say it’s hard to move beyond that in ways that are cost-effective or that apply to infectious diseases that mostly affect the developing world. If we think that AI’s going to make drug development faster and cheaper, could that also mean it will address a broader reach of patients?
Molly Gibson: Yes. I’m really excited about the opportunity for AI to change the economic equation for drug development. We should see a complete change in priorities. We’ll be able to pursue different types of patient populations: those in developing countries or places where clinical trials are higher risk or cost more. If you can bring the drug development costs and time down significantly, you have more shots on goal, and you can do new and novel things that just haven’t been possible with models in the past.
Not just faster, cheaper, and better
Lydia The: What do you think the lab of the future will look like? What types of work will labs be focused on?
Molly Gibson: This is an exciting topic. When we think about what type of data we need to generate for AI, it’s not going to look like the types of experiments that you’ve done before. We’re going to constantly be thinking about how to miniaturize experiments and how you make them high throughput in a completely different way so that you constantly have data feeding the algorithms, and you’re learning from every protein you generate.
One of the things that we’ve invested heavily in at Generate is fabricating our own microfluidics devices to be able to assay proteins with high throughput rates that traditional techniques just can’t do. A lot of times, people think about layering AI on top of how we already do things. What gets me more excited is the idea that AI gets to transform what we do. It’s not just about doing the things that we already do faster, cheaper, and better. It’s about being able to do things that we weren’t able to do before.
Lydia The: This seems like it’s going to be very expensive and require a lot of data. How will it affect larger, well-funded biotech companies or institutions, versus smaller upstarts or researchers at a university that may not have the same resources?
Molly Gibson: There’s going to be a play across all these different sizes of companies, from the large companies that can afford to generate large amounts of data to start-ups. As much as I believe that there’s going to be a huge advantage to having data, I still think human ingenuity will have an important role in innovation. How you approach the problem and how you execute upon that problem will be important.
We’ll also likely see more and more public data sets being generated for start-ups, and people with new ideas and benchmarks to test new algorithms against. And with all those new ideas, it’s execution—your ability to turn those ideas into reality—that will make all the difference.
Of course, the data that we have access to is going to dictate the types of problems that we solve and the areas in which AI and machine learning will make the biggest impact the earliest. A good example of this is DeepMind.1 They have been able to make enormous strides and solve the protein-folding problem—which had been a holy grail in the field for decades—because there is just such a well-curated data set in protein structure. That’s been a key part of their ability to innovate.
Lydia The: What do you think are the main barriers to realizing the vision of AI in changing the cost structure and ability to develop drugs?
Molly Gibson: One of the challenges will be to create data sets without bias. This is a huge area of focus across biology and in patient data. I saw recently that most of the medical data that have been used to train AI algorithms came from three states: California, Massachusetts, and New York. So we’re only seeing a very small portion of the population. Different types of bias creep into all different types of data sets, so it’ll be very important to make sure that the data sets that everybody is using to develop new models are fair and representative.
But the biggest barrier, by far, is that biology is hard. Human biology is incredibly complex. Integrating our understanding of biology with engineering, translating that into an ability to create novel molecules—bringing it all together will be very challenging. And moving into more complex diseases, expanding our reach to the most in-need patient populations, understanding disease at an earlier state so we can treat or reverse disease before it even starts—mastering all of that complexity and being able to do these impactful things is going to be really hard.
But it’s such an exciting opportunity. The future is not going to look like the past. The way that we discover drugs will not look like it did ten years ago. What we have done in drug discovery historically is going to look incredibly analog to the way that we discover drugs in the future.