Pfizer Sees NLP for Scientific Research and Deep Learning to Analyze Scans As Core AI Use Cases

Peter Henstock, Machine Learning & AI Technical Lead, Pfizer

Peter sees AI as a core toolset helping pharmaceutical companies to get new medicines to patients faster. He sees lots of general tools out there but what companies like Pfizer need are industry-specific capabilities. NLP focused on understanding scientific texts is one place he’d like to see more emphasis from startups. Another area is deep learning for imaging to characterize and flag regions in images that are difficult for a human to catch.

Amanda: To start off, I’d love to hear a bit about your role and what your team structure looks like.

Peter: I’m somewhat of a lone wolf in that I don’t currently have an AI team. I’m on a mission to try and get AI across the company. I mostly work in drug discovery but have projects that go into the legal, clinical, safety and all sorts of other areas.

Amanda: For AI initiatives at Pfizer, do they all run through you or are there team-specific initiatives where there will be an AI lead for, say, a certain clinical team?

Peter: We have an AI Center of Excellence that’s working on putting solutions in place for areas around RPA, customer engagement and a few other strategic areas. Then there are individual research efforts that are focused on specific areas of scientific problems. All of these operate somewhat independently. I am often asked to advise on different projects as they come in or join teams.

Amanda: It sounds like there’s AI being applied in a lot of different areas. How do you learn about new applications and the different vendors out there?

Peter: I try to attend conferences and talk to all the vendors to learn about their latest offerings because the landscape is changing very quickly. Quite a few vendors contact me daily. I usually prefer conferences as a way of finding out the latest technologies. Journals and proceedings provide the background and details on the ever-changing areas.

Amanda: Which conferences are most relevant for you?

Peter: It’s a range that are specific to our space. Bio-IT World is a big one for our industry. They’ll have several hundred vendors attend that who are focused on the life sciences and pharma spaces. Beyond that, there’s the KDD Conference which is more technical and not just based on pharma but around general AI approaches across industries.

Amanda: Got it. So when did you first start seeing AI being applied and what were the earliest applications?

Peter: In pharma, AI has been around for 20+ years. A lot of this AI work has been focused on the bioinformatics and cheminformatics spaces. The predictions of what compound might bind to which particular protein is all being done through AI models. We’ve used these for many years. They’ve evolved and continue to get better. Those were some of the early uses. The more recent applications are the use of natural language processing to handle our unstructured text that is coming in all forms from safety reports, FDA letters, and clinical trials. There are also a number of imaging projects to analyze cell responses better and faster while extracting different types of information. It’s an exciting time for these areas that are advancing quickly.

Amanda: Are most of these initiatives done in-house or are you bringing in third-party vendors?

Peter: We use a combination. We constantly compare how well our internal capabilities match against the state-of-the-art externally. Google and Facebook develop all sorts of great applications. Everyone is leveraging their work and applying these tools to their specific problems. Certainly, there are some spaces where we don’t have the expertise or bandwidth to take on the project. We are trying to outsource these to strategic partners.

Amanda: Can you give some examples of the projects you’ve taken on recently?

Peter: We’re trying to understand research literature. There are more articles published than any single person can read and the ability to leverage the latest papers is critical for research. Getting the right papers to the right people can change the course of projects. For example, finding a new association between a gene and a particular disease can help us understand the target space, the associated risks of side effects, and impact the direction of the project.

Pfizer has also had a large collaboration with IBM Watson around identifying immune-oncology drug targets. It aimed to figure out the set of genes associated with boosting one’s immune response as a way to treat cancer.

Amanda: How about on the image recognition front — what are the specific applications that you’re using AI models for there?

Peter: A core area is drug safety. If you have tissue samples, you might want to look for certain effects potentially caused by a drug candidate. Being able to quickly locate problem areas is a big challenge because the images are massive. A single standard image would fill 100s to 1000s of laptop monitor screens and the issues occur at different resolutions. You’re constantly zooming in and out with a microscope looking at small patches and it’s a challenging task for anyone. It requires a pathologist to really understand the images. Deep learning can help the pathologist by flagging regions with potential issues.

Amanda: When you are using third parties to help build these kinds of models, are you mostly partnering with companies like Google and IBM or are you spending a lot of time with startup vendors?

Peter: We’re doing both. Our internal teams are not usually conducting research into the newest Deep Learning methods, for instance, but are applying the best available tools to solve our problems. Model tuning actually is one interesting place where some startups are focusing. Larger companies often provide more general and encompassing tools.

Amanda: Have you hired a lot of AI experts to learn how to use these tools?

Peter: We have some, but do not have enough expertise in-house yet to solve the 1000s of problems where AI could be applied. It’s difficult to hire experts because a lot of the good ones are taken by tech companies and even those don’t understand the business that we’re in. We’re looking to fill a really challenging skillset. We need people who understand the math, statistics, and computer science as well as the science, biology, and pharma spaces. It’s a lot to ask for.

Amanda: Within Pfizer, what’s the general sentiment around AI? Is there concern that part of the job of data science is being automated?

Peter: I haven’t heard a lot of concerns about being displaced. There are just so many problems that we have to tackle as an industry. The success rates in clinical trials are still below 14% which is abysmal and shows the complexity of developing new medicines. We have the best experts. We have people with postdocs and 10+ years of experience trying to discover drugs and trying to take them through clinical trials successfully. It’s still such a huge challenge.

The patients need and deserve all the expertise we can get — the best of science, best of approaches, best of algorithms. AI is helping to fill that gap and provide an extra benefit of toolsets and capabilities that we’re all looking for to get the medicines to the patients faster.

Amanda: So if you were to advise AI startups wanting to solve some of your problems, what areas would you direct them to? Where are there gaps/holes that solutions today aren’t filling?

Peter: We need more advanced tools that can be applied to our pharmaceutical problems. There are lots of great algorithms that exist out there off the shelf. We can take those and apply them directly ourselves. The question for the small startups is whether they can demonstrate an advantage over these tools. In many cases, it’s not clear. We are looking for more innovation to improve upon those solutions that are already readily available.

There’s an ever-growing number of companies and we find there’s a lot of noise. It’s very hard to figure out who actually has novel technology versus who is recycling what everyone else is doing and selling us back the open-source approaches.

Note: post represents personal opinions of the interviewee and does not necessarily reflect the opinions or positions of Pfizer