Avak Kahvejian, a general partner at Flagship Pioneering, and Molly Gibson, a principal at Flagship, originate companies that employ machine learning and artificial intelligence to generate new insights into the complexities of biology. Jason Pontin, a senior advisor at the enterprise, spoke to them in March. The following are highlights from that conversation.
Jason Pontin: Can deep learning make sense of disease in ways that human intelligence cannot?
Avak Kahvejian: That's for sure. I think the number of data points and the diversity of data that one has to take into account when trying to understand a disease or trying to understand even basic biology is so immense that only a computational approach can address it and elucidate potential explanations and solutions.
JP: What's the current state of the art in machine learning and its application to biology?
AK: Flagship uses a variety of machine learning algorithms with the intent to solve a particular kind of problem or to answer a particular type of question. And given the kind of data that we're dealing with, there is no one-size-fits-all approach—but there are general algorithms that are applicable to many types of biological questions and data. We’re using combinations of them, including generative algorithms, in all of the companies.
For instance, one of our companies, Cellarity, is using machine learning to understand how cells behave in the context of health and disease, without necessarily trying to tease out a reductionist explanation for the mechanism by which these things are happening. It's predicting those behaviors and predicting the types of drugs that could work to influence those behaviors.
Without understanding the mechanism, we can use these approaches, for example, to choose a number of drugs that could tilt a population of cells from a state of disease to a state of health. That’s just one example where we're applying this agnostic and non-reductionist thinking that's only possible today with machine learning to make predictions about drug action.
We can even extend this capability to design new molecules and infer the chemical structures that can impart these behaviors and no longer just rely on predicting things but begin to generate things.
JP: Can machine learning be used for precision medicine to predict better outcomes in patients, where we can match molecules to specific populations?
Molly Gibson: Oftentimes when we think about a disease, we think about it at a high level. And there's much more complexity in subpopulations than we can see through the lens that we're using today. We aren't able to de-convolute, so we're not treating patients and targeting the specific disease that we could.
One way to get more precise would be to use tools like machine learning to get closer to the right population groups that we want in order to treat a specific disease. A lot of neurological disorders fall into these categories where we have heterogeneous populations of people but we classify them as having one disease and use one approach to treat it. But if we could actually elucidate those underlying biological processes that are leading to the disease and better pinpoint patients, we could precisely treat them.
Machine learning suggests a completely new approach to drug development. Instead of selecting from an array of molecules that have already been created, either through nature or some kind of synthetic method, you'd identify the specific molecule that would do what you want it to do and just directly generate that molecule.
This requires an understanding of biology in a way that current drug discovery does not. By using machine learning, we can turn the entire system on its head. Instead of just pulling out existing molecules, we can actually generate entirely new matter, entirely new modalities and concepts, and new molecules that can transform medicine.
Editor's note: This conversation has been edited for length and clarity.
