Exploring AI Challenges in Pharma with Fausto Artico

Sheikh Shuvo:
Hi, everyone. I'm Sheikh, and welcome to Humans of AI, where we meet the people who are building the tech that's changing our world.

Today's guest is Fausto Artico, Global Head of Data Science at GlaxoSmithKline, one of the world's largest pharmaceutical companies. He manages a large team around the world to execute AI initiatives across GSK. Super excited to share him with you and learn from what he's been up to. Thanks for joining us from London, Fausto.

Fausto Artico:
Thank you very much to you, Sheikh. It is a pleasure to be here with you and thanks for inviting me.

Sheikh Shuvo:
Yeah. The very first question I have for you, Fausto, you obviously have a very impressive title and I'm sure you're doing very complicated work. If you had to describe your job to a five-year-old, what would you say you do?

Fausto Artico:
I make things happen. That is what I really do. And you know, many times people are focusing only on the tech and instead the cultural challenges that we face or the emotional aspects of things are way more important, especially in big organizations that have a lot of regulatory constraints and so on. So to a five-year-old, let's do something cool. I'm the guy that makes those things happen with a team of people all together.

Sheikh Shuvo:
That sounds like a great T-shirt to have. Cool. Well, taking a step backward then, tell us just about your career and how you landed where you are.

Fausto Artico:
So I have multiple PhDs, multiple masters. I worked in Silicon Valley for a short period of time for a video corporation. And after that, I decided to change and to move from deep tech to help people in a different way, especially in a sector like biopharma and biotech. There was an opportunity with GSK in London. I moved here and very quickly, essentially, I went from being a manager to become Global Head. And, you know, I always work in innovation. This is a transformation. I help companies to transform. I was hired to develop new capabilities, services, products from zero to one. What I found is you can never scale quickly enough, doesn't matter how excellent you are. And therefore, for the last five years, even if I remain very hands-on, I moved and transitioned more on the business side of things. That implied that I had to take a lot of certifications and do them and get approved at MIT, University of California, Berkeley, Kellogg, Northwestern, and so on. And I really enjoy much more doing that than coding at this point in time, because really change can happen at scale only if you have enough people that are enthusiastic and are working together to deliver a common goal or objective.

Sheikh Shuvo:
Is there anything you miss about committing code?

Fausto Artico:
Not as much coding because I think that things could have been automated much better already a long time ago. But sometimes I miss that kind of investigation and research that I was doing to create a new algorithm or to try to solve a problem in a way that people could believe was not possible. That attraction in the real world and therefore a way in which the technological transfer can really happen and reach a final customer in the market. That is really what I love much more right now compared to coding before.

Sheikh Shuvo:
As you made the shift from working on deep tech to the pharmaceutical world, what are some of the things that have to change about your mental models about how you approach work?

Fausto Artico:
Work? Oh my God, the way of working in the military compared to academia compared to commercial is very different, in my opinion. And even in the commercial sector, when you work for a biopharma company, considering how essentially the pharma sector is structured, you don't really have direct control on a lot of moving parts. And you know, to develop a new drug and get it approved for commercialization can go anywhere from seven years to 12. So those kinds of projects are very similar to aerospace engineering projects for some things, but even more complicated because we do not really have a science well defined for a biologic body like the one that we have. And therefore, what I learned is that networking and creating an influence network of supporters is way more important in this sector than many others. Creating a legacy system that can stay in place and be fully integrated with a lot of other ones that were already there before is important so that if something happens, the system can continue to run in many different ways for many years. That is another thing that I learned. And more in general, I think that I'm not anymore believing as much as before in technology. I think that because people are running businesses, people are still, and will continue to remain the main component. And therefore, that is much better for a lot of people to learn aspects of how you relate to people, how you can create a strategy that is allowing to each one of the people that need to be part of the strategy to feel that they are growing and they get what they want, even if in a way that is allowing us for the greatest benefit of the organization. Those are the three things that I learned working.

Sheikh Shuvo:
Absolutely. Well, now shifting gears a bit to the teamwork side of things, it sounds like you've managed a very diverse team across the world. What's something different about managing a team that's working on an AI project versus any other type of tech?

Fausto Artico:
It's difficult to answer that question because, in my opinion, that assumes that a team is working only on the AI part of a project. But at least in pharma, because there are a lot of legacy systems, you need to involve a lot of different people with many different skills. For example, you cannot create an AI pure solution out of the blue without considering privacy and security aspects, the data engineering aspect, the infrastructure aspect to make it scale. So the reality is, yes, there are some teams that are more, let's say, AI pure than others, but it's difficult to have that kind of purity such that you can create a clear differentiation compared to a pure infrastructure team. However, to try to address your question, if we define AI as the algorithmic part and the model creation part, those people are sitting in a specific part of the pipeline of the whole ecosystem. And even if it works more like a flat graph, and there are many interactions and things going on in parallel, assuming that we can streamline the process, they are really sitting and talking with the data engineers to understand the contextualization of the data sources that they are receiving that are already cleaned. And they need to talk with the tracking, logging, and monitoring guys in the infrastructure part, because you want to make sure that if those models are going to drift, you can block the models and retrain the models and have humans take over when the models are running into production. So, an AI team is doing a lot of exploration in finding patterns in data. And usually, they come more from a mathematical, statistical, and computer science background. Other teams, like for example, the data engineering team that comes before them, are much more on the cleaning and contextualization of data. So they are computer scientists, but they do not need to know as much math and statistics as the pure AI team. And the infrastructure team that comes down the line for logging, tracking, deployment, and so on is much more on the side of things. So the guys definitely do not really need to be computer scientists. They need to come more from a kind of tech background where you can integrate and glue up tools and have the models essentially deployed in a way that is going to fit on the existing architectures and capabilities of the organization. So we are talking about a network and communication and perimeter and things like that. And in addition to that, yes, containers and other things that are so good in the cloud.

Sheikh Shuvo:
Yeah. Let's see now, let's say to develop the model itself, including all of the data cleanup, the training and fine-tuning takes X amount of time. How much time does it take to do everything else until things are in quote-unquote 'in production'?

Fausto Artico:
Much, much, much longer. So, for example, if you were to talk with a regulator and say I have an amazing new generative AI solution. The regulator is saying, 'you what?' Because they want to have some humans in the loop. Really, it's too far-fetched to think that there is an auto-wheeling generative AI complete autonomous solution that is discovering drugs, right? So, that process is running many times in parallel and it starts even earlier than when you want to try to develop a model because you need to educate those people. You need to involve them. Those are the people that essentially can say, 'yes, you can commercialize this solution or not'. So it starts way earlier. It involves all the levels of the organization and also outside partners like the regulators for different reasons because the financial guys want to understand how much this solution is going to make because the legal department wants to understand if we will get any kind of problem using this solution. And after, you know, you have the C-level stakeholders that want to understand if we can protect this solution and it's going to generate a really competitive advantage. And that is the reason why I'm saying to you that this takes much longer than just developing the solution. We start way earlier just to approve the fact that you can build a solution. And after that, there are a set of gatekeepers along the road where even if you can prove the technical feasibility of the solution, to bring that solution into production and start to have it adopted from a smaller number of users and scale up that kind of number of users requires additional procedures that can easily take between nine months and one year as you ramp up the work. Solution in production, that is the reason essentially the before and after that is going to generate why it's going to take so long.

Sheikh Shuvo:
Yeah, there, one of the words you used was explaining things to different regulators and obviously everyone has different frameworks that they use to assess things. And right now there are lots of companies that talk about different tools to make AI explainable. Are there any of those tools that you use or things that you've developed internally?

Fausto Artico:
So look, if you, for example, the answer is no. And the reason why I cannot find them, they are not flexible enough for a big pharma company that is very complex, from early-stage discovery to clinical development to supply chain manufacturing, commercialization, and so on, to be a good fit. That is the first aspect, why I'm not using anything that I can find in the market, even if it can be used. The other aspect is interpretability, explainability, simplicity, and so on. And the problem with that is if you give to one of those tools a large language model, there is not much that the tool can do. What you can do is a problem of the foundational principles that were used to create that kind of model and architecture that are purely based on statistics. And this is also the problem that you see many times when you want a solution adopted. You can do something incredibly amazing in deep learning. But people cannot simply relate to a statistical explanation. A biologist would like to know, from the scientific point of view, why that solution is proposing something. And today, the answer is, because statistically, it works. And I say, 'I don't care about statistics, I want to know, from the biological point of view, why it works.' A much better way is to develop solutions that are hybrids. So with some principles from science, as we know today, and some statistical principles that can come from deep learning, so that at least the final results can be explained in some ways to the main knowledge experts, so that they can understand why some choices were made by the model. And if the science is still not good enough to explain why, at least we have statistically and how, with some kind of principle and feedback to the scientific part of the company so that the science can grow. And some new explanation can be developed starting from first principles that today we don't have, but the machine is pointing from the correlative point of view saying, 'if you do this, this is what happens.' And you could not see it before simply because there was too much data. But now I tell you that effectively there is a correlation over that. That is what I find. It works. Other problems are related to the responsible use of AI. Those are better in the sense that you can architect and design ways in such a complicated and constrained set of mechanisms that the model can really run on some guardrails. And if they try to go out, you can stop everything. That is easier. But yeah, a tool that is explaining why some things are happening, no, I still did not find it. And therefore, we are working on that.

Sheikh Shuvo:
Okay. We'll just chalk it up to science fiction for now.

Fausto Artico:
Yeah. That is also the reason why it's easier to start to create models that are simple from the start or hybrids so that some people can understand why they're making some choices. Nice.

Sheikh Shuvo:
Well, in some ways that you, as an individual interested in this space, stay up to date on the latest and greatest. Are there any resources you can recommend? Any events you'd recommend, things like that?

Fausto Artico:
There is a lot of hype. What I would, yeah. It depends on the kind of audience and people, but there are many opportunities at conferences that are not technical and not academic, where people can get a feeling of what the trends are and how companies are doing things. And most importantly, can network. Because really, the interesting discussions, in my opinion, are happening during the networking part or during the evening when we have a dinner together after meeting at a conference. So those kinds of events and all opportunities in your city, if there are meetups, can provide to everybody some kind of insight that is useful. Books are still too technical. I don't think many people need that from the regulatory point of view too early. So just network a lot and hear things from many different angles and many different sectors so that even if you are not a technical person, you can start to get a feeling of what is coming down the pipe and if it's realistic or not. That is my advice, and it's a lot of fun. I'm Italian. I like to eat, I like to drink, I like to talk to people, so you will have a lot of fun too, I'm sure. Absolutely.

Sheikh Shuvo:
Now, at the very last, the question I have is, let's say if you were finishing up your PhD now and thinking about how to start your career in the world of AI, what would you be doing right now?

Fausto Artico:
So look, both PhDs were with NVIDIA GPUs and NVIDIA GPUs are not really the best, in my opinion, for those kinds of very fast generative AI problems. There are other non-mainstream competitors that from the hardware point of view are providing a lot of interesting opportunities. Guys, if you are young, you do not have the opportunity to develop large-scale models. And that is what many companies look for. However, you have a knowledge that is fresh and recent. So try to be a little entrepreneurial. In the world of tomorrow, you will not have a safe job. You will need to be very proactive. Network a lot. I invested a lot of time in creating super complex technical things when I saw many other people in life get a lot of fun just because they were meeting people and that kind of network was bringing them opportunities. And many times you cannot figure out the opportunity. It's not about working for a big company. If you work for a big company, it's fine. There are some advantages, but your career path is completely different than, 'Oh, let's create a startup.' Okay. And so that kind of balance is important. Don't do too much of only technical work or don't do too much of only networking. Try to find the right middle path. And I think amazing things will happen because you will find a job as a scientist at the moment, for sure. But don't stop there. The sector is moving fast. When you get 40 years old, you start to get tired and you need to have a network in place by then so that there are opportunities to help people in a way that you cannot figure out today, but that are very well perfectly fine for you because during the 20 years, from 20 to 40, you built a lot of holistic ways of seeing things and seeing so many different parts of how things click together that maybe have nothing to do with technology or for which the technology is only one part, but that are important essentially in the world of informal networks to make things happen in the real world.

Sheikh Shuvo:
Okay. Awesome. Well, Fausto, thank you for sharing your story and your advice. This has been a ton of fun. That was great.

Fausto Artico:
Thank you to you.