Following MedImmune’s Second Annual Translational Science Forum, a Q&A with keynote speaker Philip Nelson, director of engineering at Google Accelerated Sciences
What is the true capacity that digital technologies have for improving today’s health outcomes? MedImmune, the global biologics research and development arm of AstraZeneca, recently convened the second annual Translational Science Forum in Silicon Valley, CA to take on this question. Together with our partners at the University of California, San Francisco (UCSF)—as well as industry leaders and experts in the fields of drug development—the day-long forum undertook the timely and critical topic of Digital Health: Disrupting Drug Development.
Philip Nelson, a pioneer in the fields of science and machine intelligence and director of engineering at Google Accelerated Sciences, keynoted the event, discussing the advancement of diagnostics and drug discovery with machine learning. In this Q&A, Nelson shares his thoughts on how digital technologies are revolutionizing the way we approach the discovery of new therapeutics and health care.
Q: Can you share what your Accelerated Science team at Google is prioritizing now and in the future to contribute to digital health?
A: We’re working on two big initiatives – the first is interpreting empirical drug screening data. When people run a drug screen, they might be screening compounds of varying doses and trying to measure the effects. Traditionally, you pick a few measures to pull out of those microscopy images using software. There’s technology flowing out of deep learning called embedding, which essentially lets you look at the complete image and ask questions like “did something interesting happen relative to the controls?” And, “where something happened, which are similar?”.
With machine learning, you must be very careful to not project what you’re seeing and trick yourself – trying to decipher what is a normal variation versus identifying when something interesting happens is what is key. Differentiating the two will open up an amazing world of research.
In terms of mechanism of action studies, we’re seeing very early phenotypes of particular effects of drugs. You can knock down genes and see the morphological effects and potentially connect the effect of small molecules to the effect of genes and bridge that mapping. You can potentially stratify patients who might clinically manifest in a similar way with a disease, but from a cellular point of view they might have very different morphologies, and that might tell you something about the origin of the disease.
The second class is around computational chemistry using deep learning – we’ve published a variety of papers on this. An example is taking a compound that has an effect and finding a comparable compound that has a similar effect, and using techniques to predict other medicinal properties like solubility.
Q: The Translational Science Forum discussed innovating in personalized biologics, vs. “one size fits all” medicines. How can machine learning and intelligence help develop personalized health care solutions for patients?
A: Personalized medicine is very interesting and complex. In some of the areas we work in there are very few therapies, so you’re not picking off a deep menu in terms of how to treat the patient. In order to help discover new personalized therapies, you need really good characterizations of a patient.
Let’s use a patient’s voice as an example. ALS (amyotrophic lateral sclerosis) is a degenerative disease that starts from the limbs inward and eventually impacts your voice. Can we learn an objective biomarker from a patient’s voice and the change in a patient’s voice? Can we take signals information and build classifiers that can stratify the patients and then use those signals in order to drive their care? It is important to do both. You need to identify what the signals are and then you have to identify what the best practice for this type of patient is. We believe in the aspiration of personalized medicine, but at this stage we are still trying to understand the characteristics of individuals, diseases and therapies, and then do a much more intelligent mapping of the three.
…great advances in one field don’t necessarily carry over into another field without a lot of work, and that’s really where the opportunity lies… There’s so much expertise in the world, and the revolutionary translational science happens by bringing them together through collaboration.
Q: Can you recall one technology your team helped develop that improved health outcomes that you were particularly proud of?
A: One technology I’m very proud of hasn’t improved health outcomes yet, but we’ve done groundbreaking working in reading retinal fundus images. Diabetic retinopathy is the fastest growing cause of blindness in the world and it’s completely preventable. We are working very closely with hospitals in the developing world – particularly in India – to dramatically improve the screening rate.
In India, there are hospitals doing incredible work, but they need the tools. Our hope is by partnering with them, we will be able to drive the screening rate while also treating patients with something as simple as metformin, which could dramatically impact quality of life. We’re hoping these technologies, whether it’s retinal screening or some of the others, can really advance preventive care.
Q: In your opinion, how has big data and digital health transformed health care over the past 10 years?
A: I had the opportunity to shadow some amazing doctors doing rounds at Boston Children’s Hospital, and it was tough to watch the head of the NICU (neonatal intensive care unit) negotiating with her computer terminal to find data. We discussed concerns about the long hours the doctors work trying to retrieve and analyze data, but how mistakes can still get made in handoffs and transfers. All of this data is coming online, but it’s not quite as accessible or as useful as it can be.
Right now, we’re building the foundations of what could be amazing technology in this area, but it hasn’t been fully realized yet. I’m optimistic this new round will prove promising, especially with machine learning technologies. In the ICU (intensive care unit) it’s easy to be overwhelmed with data, but at the end of the day you need to make sense of that data and make decisions from it. This is the promise of the technology.
Q: Google collaborates closely with world class research partners to help solve important problems with large scientific or humanitarian benefit, which is also a priority for MedImmune. Why are these collaborations important, and what qualities of a research partnership make it successful?
A: My team is primarily mathematicians, computer scientists and computer researchers, and we’re humble in approaching amazing experts from other fields. By working with people who are top of their field – in biology, sensing, genomics – the hope is to bridge the gap between them. By bringing our expertise, which comes from the computer science and consumer world, we learn about biology and teach the biologists about what we do. There are amazing people who can actually bridge both worlds, and working together is truly where the magic is going to happen.
Things are still a little bit too isolated – great advances in one field don’t necessarily carry over into another field without a lot of work, and that’s really where the opportunity lies. Can we harness the innovations you’re seeing all around? Can we harness those for health care purposes? It’s really an incredible calling and an incredible mission, but the collaborations are absolutely necessary to execute them. There’s so much expertise in the world and the revolutionary translational science happens by bringing them together through collaboration.
In order to help discover new personalized therapies, you need really good characterizations of a patient…. At this stage we are still trying to understand the characteristics of individuals, diseases, and therapies — and then do a much more intelligent mapping of the three.