Data science and AI have a growing role in drug discovery and development. Industry analyst, Michael Krigsman, speaks with a leader in data science for health care, Dr. Bülent Kiziltan, to learn more.

Dr. Bülent Kiziltan is an AI executive and an accomplished scientist who uses artificial intelligence to create value in many business verticals and tackles diverse problems in disciplines ranging from the financial industry, healthcare, astrophysics, operations research, marketing, biology, engineering, hardware design, digital platforms, to art. He has worked at Harvard, NASA and MIT in close collaboration with pioneers of their respective fields. In the past 15+ years he has led data-driven efforts in R&D and built multifaceted strategies for the industry. He has been a data science leader at Harvard and the Head of Deep Learning at Aetna leading and mentoring more than 200 scientists. In his current role, his data-driven strategies with machine learning, analytics, engineering, marketing, and behavioral psychology components had a disruptive impact on a multi-billion-dollar industry sector.

Transcript

This transcript was lightly edited.

Introduction

Michael Krigsman: Data science and pharma, that's our discussion for today on CXOTalk. Bulent Kiziltan, tell us about your work.

Bulent Kiziltan: I have been advising companies on how to leverage machine learning and build data science use cases for best outcomes, as you mentioned, multiple domains that include pharma and biotech.

Michael Krigsman: Set the stage for us. What are we actually talking about?

Bulent Kiziltan: Pharma is big business. Big companies that are trying to survive in that space. It's very competitive from a business perspective but also the public health impact of how pharma companies are conducting their R&D and how they're investing their efforts has a huge impact.

When we look at the drug discovery and production costs, it has been skyrocketing over the last decade or so. It's gone up from approximately $800 million per drug to develop from end-to-end. That's the number back from 2001. In 2016, the projected number was approximately $2.9 billion. When we look at the projections, it seems it's going to be unsustainable. Pharma companies are looking for ways to cut down the costs and coming up with new, innovative approaches to drug discovery to continue to be relevant and sustain their impact.

Michael Krigsman: AI and data science in drug discovery, basically it's like every other industry; they're trying to do it faster, do it better, and do it cheaper. Right, as a quick summary?

Bulent Kiziltan: Certainly, it's one dimension of that equation. I think it's even more important than that. Pharma companies, if they don't come up with means to cut down the costs, they will not be able to survive. They're going to maybe turn into another Nokia if they are not using the most innovative approaches. Whereas, some other companies, they can optimize certain procedures and increase revenue. There's certainly ROI in investing in data science, but I think pharma is at a critical spot and the challenge is monumental.

Data science and AI are of existential importance to big pharma in drug discovery and development

Michael Krigsman: I'm assuming then that using data science and AI techniques in pharma must be pretty far along since you're describing it as existentially important.

Bulent Kiziltan: I think it's an existential point of investment but, to the contrary of what you just said, I think the data science and machine learning investment in pharma is in its infancy. They're just learning to crawl in that space, mainly because pharma companies had to reinvent how they are going about R&D, how they're implementing the results from their research and development into their pipelines. You cannot have and remain the same with the same infrastructure and expect innovation to happen.

I think both culturally, and from an infrastructure perspective, pharma companies are very big. There's a lot of inertia internally. Adapting data science and going beyond a service provider internally to a critical stakeholder in the decision-making process takes time.

Michael Krigsman: Can you elaborate? When you say, "Make an evolution from being a service provider to a critical stakeholder and decision-maker," what does that mean?

Bulent Kiziltan: This is a generic problem in many businesses as well as pharma, where the risks are very high both financially and strategically. Traditionally, pharma has been very territorial. Back in the days when data science was not on the map and there were statisticians or people who do informatics, they were providing services to other stakeholders, to the main stakeholders that were driving the business and drug discovery pipelines.

They are currently being replaced by data science teams by just changing the label. That works to a certain extent for short-term gains, but I think pharma companies realize that keeping them siloed and not fully integrated into the decision making process at the very top will not work in the long-term. I think that transformation is taking place as we speak in big pharma.

Which areas of pharma can benefit from data science and machine learning?

Michael Krigsman: What are some of the applications, some of the domain areas within pharma that data science and AI seems particularly well suited for?

Bulent Kiziltan: Data science has the potential to make an impact in all operational pipelines, both in pharma and in other businesses. So far, in pharma, interestingly, data science has already made an impact in optimizing the clinical trials pipeline, which essentially cuts down the cost. The projections are, with just implementing rudimentary machine learning and data science into the clinical trials process can and has been cutting down the costs up to 20%. This is very significant. The clinical trial process has been using machine learning effectively and it's going to be even more effective as the years go by.

Also, pharma is trying to make financial predictions to understand the potential market impact of any drug they're trying to discover or work on. Machine learning has provided very powerful, predictive machines that give pharma companies powerful predictions about finances. It has also been showing promise in that space.

Then the main areas where pharma companies are trying to improve the drug discovery process, I think, data science implementation is still in its infancy. There are new technologies coming up from the data science domain that is currently being discussed and implemented the drug discovery process, but we are yet to see a drug discovery that is being done or empowered by deep learning or data science, in general.

Michael Krigsman: Why is that? What makes this so difficult?

Bulent Kiziltan: There are multiple dimensions to this question. One of them is, data science is an up-and-coming discipline. It's still not mature enough with its methodologies. The domain is changing and reinventing itself, transforming every six months.

For instance, this year, graph learning and graph theory is being used in data science and deep learning. One of the promises of graph learning and deep learning is drug discovery. On the pharma side of things, data plays an important role in any type of data science operations and use case. The data has not been clean and not well integrated, so pharma companies are working very hard to aggregate their data, clean them up as much as possible, and add public information to the data that they have to increase the predictive power of data science. The data, not aggregated data play an important role in moving forward.

Also, the integration internally in pharma companies. Data science is a fully collaborative effort. As I mentioned, traditionally, pharma has been very siloed and very territorial.

Unless there is an empowerment from the very top pushing the stakeholders to fully cooperate and approach this problem and frame it as, "How can we achieve what we want to achieve together?" rather than, "Provide me with those insights and I will do my end of things and I will do my job," I think are critical. The cultural transformation, the data transformation, and also the know-how that's being improved as years go on in data science, they all play a role in moving slowing in that space.

But also, there is a justified skepticism that I want to touch upon. Traditionally, again, pharma has been very siloed. The leadership that is leading data science efforts were not essentially data science domain experts. They were purely driven by their traditional ways of thinking of business and that sort of thinking has shown not to work with data science operations. That is also quickly transforming itself.

What cultural change is necessary for AI adoption in drug discovery?

Michael Krigsman: We have a few questions from Twitter. Sal Rasa asks, "What kind of culture change is required to connect the intention with the outcomes for the organizations that you are involved with? Another way of saying it is, what are the cultural disconnects that interfere with the use of data science and AI techniques for drug discovery in pharma?

Bulent Kiziltan: Data science is a collaborative effort. All the stakeholders have to be organically integrated and work for the same cause.

One of the challenges of big companies and especially in pharma is that the stakeholders have been siloed for a very long time, for many decades, and they have been very territorial. That mindset has to change.

Also, for data science to make an impact, they have to be critical stakeholders and they have to sit at the decision-making table rather than just being service providers. That is also changing.

Once those two areas have been transformed culturally and internally, I think then the next step is to find the right leader and the right talent, assuming that the data is in place. I think, yes, we are expecting a huge impact in the pharma space, but I think that's going to take a few more years.

Michael Krigsman: What you're saying is, there's a kind of deep cultural divide between the way pharma has operated historically and the way data science must operate. Can you explain a little bit about the collaborative nature of data science and why it's so important in this context?

Bulent Kiziltan: There's a flip side of the coin where people are moving into data science have very strong academic backgrounds and they don't have, essentially, the optimized thinking of business folks. On the business side of things, things are very structured and the KPIs, the key performance metrics, don't necessarily always speak to data scientists.

It's essential to find a leader or translators to translate what the business side of things is expecting from the data science teams and also a person that can translate what can and cannot be done to prevent overpromising to the businesspeople to essentially prevent over-hyping what data science can deliver. I think speaking the same language is essential and there are very few leaders who can successfully do that.

What does the FDA say about data science in pharma?

Michael Krigsman: We have another question from Twitter. Arsalan Khan asks, "If pharma is using data science for clinical trials, then it would also be useful to regulators such as the FDA. Has there been a push by the government to get data science involved in this as well?"

Bulent Kiziltan: That's a very good point. We've seen that in the healthcare industry for regulators. We see the same thing in the financial world where the regulators were really behind the curve and the transformation, essentially, that AI is bringing into the domain.

In FDA, I know of efforts where the teams are being transformed to implement some of the new technologies that are up-and-coming and how they can be implemented in a healthy manner into the whole clinical trials pipeline. There was a lot of resistance to machine learning algorithms by the regulators. That has been changing and FDA and other government regulators are trying to transform themselves as well.

Misconceptions non-scientists have about data science and machine learning in pharma

Michael Krigsman: Zachary Jeans asks a really interesting question. He said, "What does the average person get wrong about what data science is and its use in drug discovery?" The misconceptions that non-scientists have about this, help us understand that.

Bulent Kiziltan: I think the perception is, you give data to a machine learning algorithm and it spits out the final product. That's not how machine learning works. Machine learning is incrementally implemented into the whole process, including clinical trials. But the drug discovery itself and the holy grail of drug discovery using machine learning is to make predictions on the therapeutics side of things with machine learning just by going into the compound libraries that pharmaceutical companies have and pipe in the clinical trials data, both the successful ones and the failures in the past, and ask the machine learning algorithm to spit out something that's predictive and useful in nature.

That has not happened yet. Maybe the layman understanding of machine learning is this and it might happen in the future, but we're not there yet.

Obstacles related to data in the drug development process

Michael Krigsman: You mentioned that one of the obstacles is the lack of data and the lack of normalized and prepared data. Can you give us some insight into what kind of data you're describing?

Bulent Kiziltan: There are clinical trials data that are all over the place. They are not standardized across even within a single pharma company because different stakeholders create their own data. They used to create their own data and they were not normalized.

Also, because data scientists or machine learning experts have not been a critical part of the decision-making process in the past, the type of data that is available, even if they were normalized and clean, might not be sufficient to extract the type of information that we're trying extract. Right now, with the proliferation of different techniques, sensors that we can put on the people who are a part of the clinical trials and continuously monitor them, not doing the trials only but after when they go home, I think that's the sort of data that we get from sensors that we can attach to individuals will be very powerful in the future.

That sort of information is currently not available. They are being implemented currently but, in order for machine learning algorithms to make predictions, we need data going back a few years. We don't have that yet.

Michael Krigsman: Is data the primary or lack of data the primary obstacle to broader adoption of AI and data science techniques in drug discovery, is it the set of cultural issues you were describing or is it a glomeration of both together?

Bulent Kiziltan: I will give you a third option. Both play an important role, but also our understanding of human physiology and disease is very primitive. Because I'm a physicist, I can tell you, in physics, we have a fundamental understanding of the laws that produce the universe that we see. But in biology, the interactions of proteins and how disease happens to be at the physiology, the fundamental understanding is not there yet.

This is both a pro and a con, how data science can make an impact. It's a limitation that we have because we don't have the fundamentals, but this is again one of the primary reasons why data science can make and will make an impact because we can do data-driven predictions even though we don't know how the fundamentals work. I think the disciplinary ignorance, if you will, the lack of deep information, I think, is one of the primary bottlenecks that prevent us from producing drugs that are going to be impactful.

What is the current state of machine learning and data science in drug discovery research?

Michael Krigsman: Can you give us an example of a situation where it's worked, where data science has supported machine learning, has supported drug discovery?

Bulent Kiziltan: For pharma companies, I think the primary objective right now is to build their infrastructure and be ready when the technology is ready and it can be scaled. Startups, because they are very focused, they are agile, they're dynamic, they don't have the cultural divide, everybody is working to achieve a single goal, there have been startups that have shown to produce predictions about molecular structure in three dimensions purely driven by data and purely produced in the computer and make those predictions about molecules. Once those molecules are produced and tested in the lab, they show efficacy that is being predicted.

I don't want to name names, but there are companies that have shown and published their work. Recently, Google DeepMind also published one of their tools that are called AlphaFold, which uses deep learning approaches to make predictions about the three-dimensional structure of proteins. It will be a very active domain within drug discovery and pharma. I am expecting that there will be a breakthrough in the coming year.

Michael Krigsman: What you're saying is—correct me if I'm wrong—it seems like a good idea but not yet.

Bulent Kiziltan: It depends on who is saying it's not a good idea. I think, for pharma companies, I think it's essential, critical to invest into that domain even though I predict, for the coming year or two, pharma companies will end up buying the solutions that startups have produced, mainly because it's much cheaper that way. Startups can move much quicker in their drug development efforts, currently.

I think it's essential for pharma companies to be ready and have a scalable infrastructure and talent pool when the time comes. There is a big opportunity cost if they don't do so.

Also for startup companies, I think moving quickly and attracting top talent, being culturally viable for data science is one of their advantages. But they have a big disadvantage. They don't have the data they require. So, I think it's also essential to build collaborative relationships between big pharma and startups to alleviate that problem.

Michael Krigsman: Are there any specific examples of that that you can point to that you're comfortable talking about?

Bulent Kiziltan: For instance, one of the startups that have proven that their predictions show promise, they have partnered with big pharma to use their compound libraries to inform their deep learning and machine learning algorithms to make further predictions. Obviously, the predictions that they're making are not public yet, but I would assume that they're making progress in that sense.

Michael Krigsman: Give us examples of where the use of data science and machine learning has not met expectations or failed. Give us some insight, at the same time, as to what went wrong.

Bulent Kiziltan: One reason data science has produced some justified skepticism in many domains, including pharma, is the leaders of that domain, as I said, they were essentially not domain experts. They had maybe unrealistic expectations of what data science can do today and were projecting for the future.

Data science has yet to produce therapeutics and medicines that can be used in cell therapy, in gene therapy. Those are areas where pharma is growing mainly because data science has to offer new insights in those domains.

Until now, we haven't seen data science-driven insights into drug discovery except a few startups that have shown some promise in that area. But we don't have a single drug yet that has been predicted and has gone through the pipeline to produce a drug that's viable and applicable.

You know we have to consider also that data science is not the only part to blame here. In regular drug discovery or drug development, 1 out of 20 drugs that go through the pipeline is successful. There is a very high attrition rate anyway in the standard approach, so data science is promising to increase those odds.

Michael Krigsman: Given the challenges, why do you say that data science is an existential necessity for pharma?

Bulent Kiziltan: As I just pointed to the cost of drug discovery. Right now, I think it's past $3 billion per drug to be produced and go to the market. The projections show it's going to hit $5 billion in the coming years. That projection, that cost is just not sustainable.

Either the price will go up to a level where it will not make sense for pharma companies to invest into them, which essentially will make their portfolio even smaller and the revenues will go down, or they have to come up with new ways. Data science and AI is potentially the only way that we know of right now to cut down the cost to optimize the whole process and even come up with new insights just given that pharma companies have huge compound libraries that scientists have not been able to effectively tap into to produce insights in the past.

I think it won't be farfetched to predict that AI and data science machine learning will produce new insights. We just don't know the level of impact it's going to have in the future, but I'm optimistic.

Michael Krigsman: What's the timeframe, do you think, that we will actually see some type of material result as opposed to the, shall we say, theory today that it seems like a good idea?

Bulent Kiziltan: Data science is already playing a role in cutting down the costs. Also, producing new insights, making the clinical trials process more effective, especially doing patient enrollment that produce data that will produce information that's more useful for the data science process. I would expect, since the data is already coming in over the past year or two, I would expect a drug or therapeutics to come out that has been largely influenced and empowered by data science processes, but that doesn't mean that data science and machine learning is yet powerful enough and implemented well enough to make predictions for a three-dimensional structure and to produce therapeutics from end-to-end.

I think that is the holy grail to be able to simulate things in the computer and produce information and empower the people who are developing the drugs from end-to-end, which will primarily drive down the costs significantly, which will make therapeutics accessible to people who have rare diseases. It was not financially viable for pharma companies to invest into what's called orphaned diseases, diseases that not more than 200,000 people in the U.S. are suffering, but they are significant. Now, data science processes will make those drug development efforts viable and possible for people with genetic disorders and mutations. It's one of the areas in which pharma companies are currently investing in cell therapeutics and gene therapeutics.

Michael Krigsman: We have another question from Twitter. Arsalan Khan asks a great question having to do with the data. "Given the fact that the data is so important, is this different from any other area, type of domain, where we have to gather and aggregate large amounts of data? In addition, specifically in drug discovery, are there perception and bias issues that are obstacles to progress and getting the results that we want?"

Bulent Kiziltan: Yes, aggregate data is very important in all domains. The problem in pharma is that the risks are high. You cannot just probabilistically make a prediction, see, and test it on the ground to get results and then basically produce that drug.

The stakes are very high. It's not like a marketing effort where you can produce a model that's based on probabilities that you're producing and then you can iterate in the market in real-time to inform your algorithms to perfect them. That's very difficult even in clinical trials because there are certain strict regulations that regulate what you can and cannot do.

You have to be very transparent. You have to have very finely sampled data rather than just granular data. You cannot aggregate data and average out certain aspects of that data, which you can really do in marketing data and some other domains. The risks are very high in pharma.

To the second part of the question about biases, yes, bias exists everywhere because we humans are biased. Our biases are reflected into the data.

There has been an ongoing discussion of whether machine learning algorithms are biased. I would argue that the algorithms reflect the bias that's in the data. Certainly, that bias also exists in the pharma space. We have certain methodologies to overcome those biases, but it's an ongoing effort by all parties, both in academia and in the industry.

Enterprise architecture in data science and pharma

Michael Krigsman: James McGovern asks, "Any thoughts on how enterprise architecture adds value, in general, to data science, but if there are specifics that you can talk about in pharma, that would be great."

Bulent Kiziltan: It's one of the most critical components of data science operations, I think. One of the reasons why data science has been slow in delivering and why the data science efforts have been hampered is mainly because IT and the enterprise architecture has not been incentivized to keep up with technology in the past. Going into an infrastructure that is not up to date, that doesn't use the latest technology, and does not speak organically with the data science efforts was one of the problems and has been one of the problems in larger companies.

In order for companies to move forward and produce the value that data science is promising, quickly, I think an effort that is company-wide where all stakeholders are on board is essential. The specifics of what type of enterprise architecture is to be used is very domain-specific. It's very specific to the business objectives of the company, the timelines, the resources, and the talent pool. I think that there are a lot of solutions in the market which can be customized to the company to make sure that data science can deliver today.

Michael Krigsman: We have another question from Twitter. This is from the @CXOTalk account. "What part of the drug discovery process do you see data science making the most significant contributions in drug discovery?"

Bulent Kiziltan: I think the frontier today is to make predictions on the three-dimensional structures of molecules and compounds with the computer using machine learning algorithms. Currently, those are all done manually with robots, but they are basically manually mixing certain components and trying to see the effect on target molecules. This is, again, done manually.

I think the frontier is where machine learning can go into the compound library and know what the target is and make predictions based on previous trials. I think this is where machine learning can and will make an impact, but it will take time.

Collaborative nature of data science

Michael Krigsman: Talk about the collaborative nature of data science and drill down and compare that to the siloed nature of historically the way pharma companies work.

Bulent Kiziltan: Coming from academia, the culture is, unfortunately, not something that we have always talked about. But as we move into the business and how things can operate and produce value, culture comes up at the top of our list all the time in many of the use cases that we are working on.

Culturally, the business mindset expects certain outcomes and they are very rigid in their thinking. This is the traditional way of approaching things. Also, in pharma, different silos and stakeholders, they want certain results and they are used to getting certain results.

Data science, by nature, has a significant exploratory component to it. I think this is not being appreciated by many business stakeholders. If you keep a rigid project management pipeline and expect certain outcomes without redefining how you operate, it will be very difficult for data scientists to use their creative capability to contribute to the operations and to the value that data science can produce.

There are two outcomes to this. One outcome is in cases where businesspeople who have been at the company are strong stakeholders. They will push the data science teams to keep producing the rigid outcomes that they're used to and it will limit the production and the creative contribution a data science team can do. Whereas, companies who are on the way of transforming and redefining their culture, sitting down with the data science leadership, having a dynamic interaction, producing the key performance metrics together, and to keep it dynamic, I think, is essential because data science, as I said, is very exploratory in nature.

This is why some of the project management styles that people have been trying to implement into data science simply does not work because the outcomes are not well defined. Whereas, you can borrow some of the ideas of project management such as Agile and some other project management styles and implement it but not as is.

If companies think of data science as a software engineering project, it will limit, significantly, their capability. For stakeholders that are a part of the discussion, I would advise them to not be so rigid and try to have a two-way dialog rather than asking data science teams to produce certain outcomes that they're used to having.

Large drug research companies working with startups to drive innovation

Michael Krigsman: You mentioned earlier the importance of large pharma companies working with startups. The larger companies have access to data and processes. The smaller companies have access to more advanced technical approaches. Once you start mixing those two, don't you layer onto it another whole set of very difficult cultural and economic challenges?

Bulent Kiziltan: That's correct. One of the solutions some of the companies have applied was to build an internal group that is somewhat isolated from the rest of the stakeholders internally, essentially acting like an internal startup. This is one way to solve some of the problems, potential problems, that you just mentioned.

I think the traditional approach nowadays is to collaborate and build relationships with startups. It really depends on how that relationship is being structured.

If you have an agreement with a startup but have that startup act like an internal stakeholder, I think it won't solve some of the problems. Whereas, you can really build a relationship with a startup that isolates the startup as a separate entity but reports only to the top and then, organically, builds a very flat structure with some of the teams that can complement to the efforts, I think, is the way to move forward. I think companies are eager to build that sort of relationship. I've seen similar relationships being built in the industry.

Michael Krigsman: From this standpoint, there's really not much of a difference from looking at innovation approaches across different kinds of industries. The pharma space is not unique from that innovation standpoint of partnering with startups and trying various things to make it work out successfully for both sides.

Bulent Kiziltan: There are unique aspects to it because pharma is also very siloed, as I mentioned. That brings a certain level of secrecy. Unless that trust has been built, which takes some time and sometimes years to build, the lack of transparency hampers those efforts.

Once that relationship has been built, I think the companies and the stakeholders can move beyond that secrecy and share information transparency. It's one of the problems building those long-lasting relationships that are based on trust.

Michael Krigsman: Is the secrecy aspect in drug discovery really one of the most important obstacles to that innovation partnership?

Bulent Kiziltan: I wouldn't say it's one of the most important aspects, but it definitely contributes or hampers the effort of building those relationships. When people and companies sit down at the table and discuss how they can contribute to their efforts and how mutually they can benefit from that relationship, it typically takes several months to a year to move through bureaucracy and make sure that the data, the challenges, and the fiscal priorities are being shared transparently.

Disruption is an obstacle to innovation in the pharmaceutical industry

Michael Krigsman: To what extent do large pharma companies have an interest in deploying innovative processes that disrupt the established organization and disrupt established executive positions and, potentially, compensation? This is a generic problem. It's not just drug discovery, it seems like. How big an issue is it?

Bulent Kiziltan: I think it's a very big issue. There is a high risk that comes with cultural transformation. There comes a high risk when it comes to changing the status quo internally. What we see is, when you come as a stakeholder or leader in one of the branches within the company, to have that vision but you are not empowered or you don't share that vision with the board, the CEO, or the CTO, what we see is it never happens. It never matures to a level where it can be implemented and executed on.

I think companies that really have decided at the board level to drive that cultural transformation, which has to go in parallel with the digital transformation, change their leadership from the very top. They change their culture and they make sure that it precipitates down to the stakeholders, which will take time, yes. But I think the first and foremost transformational point is at the very top. Unless you have a board or leadership that drives that transformation, it will never happen from the bottom up.

VC investment startup drug companies

Michael Krigsman: Kanupriya Agarwal asks, "How much funding do you anticipate being set aside for AI and data science in drug discovery in the near future? What do you see the dynamics at play here?"

Bulent Kiziltan: It's a difficult one to answer because each company is making its own decisions. I think the trend moves in a direction where a significant amount of money will go into internal investment. But also, I think, at least in the short-term, because pharma companies won't be able to complete that digital transformation and cultural transformation in the short-term, they will continue to buy, partner, or invest in startups.

I think there will be a significant budget. Obviously, many billions of dollars that go into external partnerships and purchases to expand their portfolio in the target areas. But also, internally, I think it will be more incremental. But I think, in order to build capabilities that are scalable, it takes a long time and companies know that they cannot just wait, not invest internally, and expect to buy out startups.

Two years from now, when the technology is mature enough, they cannot just hire people and then do it internally, which will obviously be the choice to make. I think incrementally building that internal know-how and capability while they diversify their risk is the way to move forward. It's the trend that we see.

Michael Krigsman: In your view, there's an opportunity for startups who are able to gain success with this to be acquired by larger pharma companies going forward.

Bulent Kiziltan: It's the right time to invest in startups that build therapeutics in different domains in gene therapy, in proteins, in oncology, and in different kinds of therapeutics and drug discovery. I think it's a very good time to invest in those companies.

Michael Krigsman: Data, however, is going to have to come from the large companies, or is that not correct? Do startups have a chance of developing their own data?

Bulent Kiziltan: It'll depend on the company and what exactly they're doing. Some companies are just building algorithms. Some companies want to be a part of or drive the whole drug discovery process and, typically, those companies are already backed by big investors. There are companies that just tackle certain aspects of the clinical trials process and optimize that. It will really depend on the company, their funding, whether they're being backed by big companies and VCs.

Michael Krigsman: What about the role of VCs and investors in pushing this forward?

Bulent Kiziltan: There are several kinds of VCs. One type of VC wants a short-term gain. But I think, in data science and machine learning, particularly in biotech, the investors are more interested in midterm to long-term gains because it will be transformational. But VCs are interacting, obviously, across the industry. They know where things are going. If they're smart investors, they will diversify their risk.

There are some companies that potentially will have results in the short-term, but there are companies that are in for the long-term. I think data science machine learning in the context of drug discovery is somewhat of a midterm investment to a long-term investment whereas pharma companies internally will continue to invest in the short-term as well to be ready when the technology is mature enough.

Michael Krigsman: Bulent, as we finish up, what have I not asked you? Are there any fundamentally key issues that we have not covered that we need to talk about for the sake of completeness?

Bulent Kiziltan: One of the aspects that we have touched upon in our previous conversations is also very valid in this context, and that's leadership. A part of the cultural transformation, I think that the most important aspect of driving the transformation internally is the profile of the leader.

Pharma, as I mentioned, is traditionally very territorial. When you look at the leadership that is driving those efforts, they have been in the industry for more than 10, 15 years. The leaders that they're typically hiring have been in the pharma industry, out of academy, for many decades.

When you have data science efforts primarily led by business leadership, it certainly adds to the data science efficiency. But I think the business mindset does not work really well with long-term data science strategies. I think it is essential to find leadership that has domain expertise, has been in multiple domains, and can adapt to the pharma domain fairly quickly.

Repeating the same thing, keeping the same structure, having the same mindset and then expecting out-of-the-box transformational results just doesn't work. You have to have leaders that can think out-of-the-box, can bring the creative aspect of data science into the operations of your company, especially in pharma, so they have to be open to the hiring of leadership that is out-of-the-box.

Michael Krigsman: If I can push back slightly on what you just said, of course, it makes sense to hire leadership with domain expertise, both in data science and in drug discovery. However, if you don't bring a businessperson to that party, then you may not be able to construct an economically viable model and set of processes for enabling this to be durable for the longer-term.

Bulent Kiziltan: Absolutely. I didn't want to recommend hiring a person that does not consider business priorities. This is where collaboration is critically important and the approach to data science operations is to build a collaborative relationship with all stakeholders, especially with the business arm of the companies, to make sure that everybody has an open mind and can go through a process in which they educate each other rather than having a rigid mindset. That goes both to the business side of things and also to the data science of things.

Michael Krigsman: I notice that, on your bio, it says Stealth Startup. How would you like to totally spill the beans and no longer be stealth? [Laughter]

Bulent Kiziltan: For companies that move out of stealth mode, there are different models. If you have a product that you are ready to talk about and share with the public, I think it's a good time. But typically, nowadays, because the R&D loop is very fast, much quicker than before, companies prefer to come out of stealth way before the product is mature enough, but right after, when they have a proof of concept. With the companies that I'm involved in, both on the advising side of things and also building technology, some of them are not there yet.

Michael Krigsman: We have been speaking with Bulent Kiziltan. He is a senior data scientist and C-level executive. Bulent, thank you for taking the time to talk with us today.

Bulent Kiziltan: Thanks for having me, Michael.

Michael Krigsman: It's been a very interesting and fascinating discussion. Before you go, please subscribe on YouTube and subscribe to our newsletter at the top of our website. You can hit the subscribe button. Thanks so much, everybody, and check out CXOTalk.com. We will see you again next time. We have great shows coming up. Bye-bye.

This transcript was lightly edited.

Introduction

Michael Krigsman: Data science and pharma, that's our discussion for today on CXOTalk. Bulent Kiziltan, tell us about your work.

Bulent Kiziltan: I have been advising companies on how to leverage machine learning and build data science use cases for best outcomes, as you mentioned, multiple domains that include pharma and biotech.

Michael Krigsman: Set the stage for us. What are we actually talking about?

Bulent Kiziltan: Pharma is big business. Big companies that are trying to survive in that space. It's very competitive from a business perspective but also the public health impact of how pharma companies are conducting their R&D and how they're investing their efforts has a huge impact.

When we look at the drug discovery and production costs, it has been skyrocketing over the last decade or so. It's gone up from approximately $800 million per drug to develop from end-to-end. That's the number back from 2001. In 2016, the projected number was approximately $2.9 billion. When we look at the projections, it seems it's going to be unsustainable. Pharma companies are looking for ways to cut down the costs and coming up with new, innovative approaches to drug discovery to continue to be relevant and sustain their impact.

Michael Krigsman: AI and data science in drug discovery, basically it's like every other industry; they're trying to do it faster, do it better, and do it cheaper. Right, as a quick summary?

Bulent Kiziltan: Certainly, it's one dimension of that equation. I think it's even more important than that. Pharma companies, if they don't come up with means to cut down the costs, they will not be able to survive. They're going to maybe turn into another Nokia if they are not using the most innovative approaches. Whereas, some other companies, they can optimize certain procedures and increase revenue. There's certainly ROI in investing in data science, but I think pharma is at a critical spot and the challenge is monumental.

Data science and AI are of existential importance to big pharma in drug discovery and development

Michael Krigsman: I'm assuming then that using data science and AI techniques in pharma must be pretty far along since you're describing it as existentially important.

Bulent Kiziltan: I think it's an existential point of investment but, to the contrary of what you just said, I think the data science and machine learning investment in pharma is in its infancy. They're just learning to crawl in that space, mainly because pharma companies had to reinvent how they are going about R&D, how they're implementing the results from their research and development into their pipelines. You cannot have and remain the same with the same infrastructure and expect innovation to happen.

I think both culturally, and from an infrastructure perspective, pharma companies are very big. There's a lot of inertia internally. Adapting data science and going beyond a service provider internally to a critical stakeholder in the decision-making process takes time.

Michael Krigsman: Can you elaborate? When you say, "Make an evolution from being a service provider to a critical stakeholder and decision-maker," what does that mean?

Bulent Kiziltan: This is a generic problem in many businesses as well as pharma, where the risks are very high both financially and strategically. Traditionally, pharma has been very territorial. Back in the days when data science was not on the map and there were statisticians or people who do informatics, they were providing services to other stakeholders, to the main stakeholders that were driving the business and drug discovery pipelines.

They are currently being replaced by data science teams by just changing the label. That works to a certain extent for short-term gains, but I think pharma companies realize that keeping them siloed and not fully integrated into the decision making process at the very top will not work in the long-term. I think that transformation is taking place as we speak in big pharma.

Which areas of pharma can benefit from data science and machine learning?

Michael Krigsman: What are some of the applications, some of the domain areas within pharma that data science and AI seems particularly well suited for?

Bulent Kiziltan: Data science has the potential to make an impact in all operational pipelines, both in pharma and in other businesses. So far, in pharma, interestingly, data science has already made an impact in optimizing the clinical trials pipeline, which essentially cuts down the cost. The projections are, with just implementing rudimentary machine learning and data science into the clinical trials process can and has been cutting down the costs up to 20%. This is very significant. The clinical trial process has been using machine learning effectively and it's going to be even more effective as the years go by.

Also, pharma is trying to make financial predictions to understand the potential market impact of any drug they're trying to discover or work on. Machine learning has provided very powerful, predictive machines that give pharma companies powerful predictions about finances. It has also been showing promise in that space.

Then the main areas where pharma companies are trying to improve the drug discovery process, I think, data science implementation is still in its infancy. There are new technologies coming up from the data science domain that is currently being discussed and implemented the drug discovery process, but we are yet to see a drug discovery that is being done or empowered by deep learning or data science, in general.

Michael Krigsman: Why is that? What makes this so difficult?

Bulent Kiziltan: There are multiple dimensions to this question. One of them is, data science is an up-and-coming discipline. It's still not mature enough with its methodologies. The domain is changing and reinventing itself, transforming every six months.

For instance, this year, graph learning and graph theory is being used in data science and deep learning. One of the promises of graph learning and deep learning is drug discovery. On the pharma side of things, data plays an important role in any type of data science operations and use case. The data has not been clean and not well integrated, so pharma companies are working very hard to aggregate their data, clean them up as much as possible, and add public information to the data that they have to increase the predictive power of data science. The data, not aggregated data play an important role in moving forward.

Also, the integration internally in pharma companies. Data science is a fully collaborative effort. As I mentioned, traditionally, pharma has been very siloed and very territorial.

Unless there is an empowerment from the very top pushing the stakeholders to fully cooperate and approach this problem and frame it as, "How can we achieve what we want to achieve together?" rather than, "Provide me with those insights and I will do my end of things and I will do my job," I think are critical. The cultural transformation, the data transformation, and also the know-how that's being improved as years go on in data science, they all play a role in moving slowing in that space.

But also, there is a justified skepticism that I want to touch upon. Traditionally, again, pharma has been very siloed. The leadership that is leading data science efforts were not essentially data science domain experts. They were purely driven by their traditional ways of thinking of business and that sort of thinking has shown not to work with data science operations. That is also quickly transforming itself.

What cultural change is necessary for AI adoption in drug discovery?

Michael Krigsman: We have a few questions from Twitter. Sal Rasa asks, "What kind of culture change is required to connect the intention with the outcomes for the organizations that you are involved with? Another way of saying it is, what are the cultural disconnects that interfere with the use of data science and AI techniques for drug discovery in pharma?

Bulent Kiziltan: Data science is a collaborative effort. All the stakeholders have to be organically integrated and work for the same cause.

One of the challenges of big companies and especially in pharma is that the stakeholders have been siloed for a very long time, for many decades, and they have been very territorial. That mindset has to change.

Also, for data science to make an impact, they have to be critical stakeholders and they have to sit at the decision-making table rather than just being service providers. That is also changing.

Once those two areas have been transformed culturally and internally, I think then the next step is to find the right leader and the right talent, assuming that the data is in place. I think, yes, we are expecting a huge impact in the pharma space, but I think that's going to take a few more years.

Michael Krigsman: What you're saying is, there's a kind of deep cultural divide between the way pharma has operated historically and the way data science must operate. Can you explain a little bit about the collaborative nature of data science and why it's so important in this context?

Bulent Kiziltan: There's a flip side of the coin where people are moving into data science have very strong academic backgrounds and they don't have, essentially, the optimized thinking of business folks. On the business side of things, things are very structured and the KPIs, the key performance metrics, don't necessarily always speak to data scientists.

It's essential to find a leader or translators to translate what the business side of things is expecting from the data science teams and also a person that can translate what can and cannot be done to prevent overpromising to the businesspeople to essentially prevent over-hyping what data science can deliver. I think speaking the same language is essential and there are very few leaders who can successfully do that.

What does the FDA say about data science in pharma?

Michael Krigsman: We have another question from Twitter. Arsalan Khan asks, "If pharma is using data science for clinical trials, then it would also be useful to regulators such as the FDA. Has there been a push by the government to get data science involved in this as well?"

Bulent Kiziltan: That's a very good point. We've seen that in the healthcare industry for regulators. We see the same thing in the financial world where the regulators were really behind the curve and the transformation, essentially, that AI is bringing into the domain.

In FDA, I know of efforts where the teams are being transformed to implement some of the new technologies that are up-and-coming and how they can be implemented in a healthy manner into the whole clinical trials pipeline. There was a lot of resistance to machine learning algorithms by the regulators. That has been changing and FDA and other government regulators are trying to transform themselves as well.

Misconceptions non-scientists have about data science and machine learning in pharma

Michael Krigsman: Zachary Jeans asks a really interesting question. He said, "What does the average person get wrong about what data science is and its use in drug discovery?" The misconceptions that non-scientists have about this, help us understand that.

Bulent Kiziltan: I think the perception is, you give data to a machine learning algorithm and it spits out the final product. That's not how machine learning works. Machine learning is incrementally implemented into the whole process, including clinical trials. But the drug discovery itself and the holy grail of drug discovery using machine learning is to make predictions on the therapeutics side of things with machine learning just by going into the compound libraries that pharmaceutical companies have and pipe in the clinical trials data, both the successful ones and the failures in the past, and ask the machine learning algorithm to spit out something that's predictive and useful in nature.

That has not happened yet. Maybe the layman understanding of machine learning is this and it might happen in the future, but we're not there yet.

Obstacles related to data in the drug development process

Michael Krigsman: You mentioned that one of the obstacles is the lack of data and the lack of normalized and prepared data. Can you give us some insight into what kind of data you're describing?

Bulent Kiziltan: There are clinical trials data that are all over the place. They are not standardized across even within a single pharma company because different stakeholders create their own data. They used to create their own data and they were not normalized.

Also, because data scientists or machine learning experts have not been a critical part of the decision-making process in the past, the type of data that is available, even if they were normalized and clean, might not be sufficient to extract the type of information that we're trying extract. Right now, with the proliferation of different techniques, sensors that we can put on the people who are a part of the clinical trials and continuously monitor them, not doing the trials only but after when they go home, I think that's the sort of data that we get from sensors that we can attach to individuals will be very powerful in the future.

That sort of information is currently not available. They are being implemented currently but, in order for machine learning algorithms to make predictions, we need data going back a few years. We don't have that yet.

Michael Krigsman: Is data the primary or lack of data the primary obstacle to broader adoption of AI and data science techniques in drug discovery, is it the set of cultural issues you were describing or is it a glomeration of both together?

Bulent Kiziltan: I will give you a third option. Both play an important role, but also our understanding of human physiology and disease is very primitive. Because I'm a physicist, I can tell you, in physics, we have a fundamental understanding of the laws that produce the universe that we see. But in biology, the interactions of proteins and how disease happens to be at the physiology, the fundamental understanding is not there yet.

This is both a pro and a con, how data science can make an impact. It's a limitation that we have because we don't have the fundamentals, but this is again one of the primary reasons why data science can make and will make an impact because we can do data-driven predictions even though we don't know how the fundamentals work. I think the disciplinary ignorance, if you will, the lack of deep information, I think, is one of the primary bottlenecks that prevent us from producing drugs that are going to be impactful.

What is the current state of machine learning and data science in drug discovery research?

Michael Krigsman: Can you give us an example of a situation where it's worked, where data science has supported machine learning, has supported drug discovery?

Bulent Kiziltan: For pharma companies, I think the primary objective right now is to build their infrastructure and be ready when the technology is ready and it can be scaled. Startups, because they are very focused, they are agile, they're dynamic, they don't have the cultural divide, everybody is working to achieve a single goal, there have been startups that have shown to produce predictions about molecular structure in three dimensions purely driven by data and purely produced in the computer and make those predictions about molecules. Once those molecules are produced and tested in the lab, they show efficacy that is being predicted.

I don't want to name names, but there are companies that have shown and published their work. Recently, Google DeepMind also published one of their tools that are called AlphaFold, which uses deep learning approaches to make predictions about the three-dimensional structure of proteins. It will be a very active domain within drug discovery and pharma. I am expecting that there will be a breakthrough in the coming year.

Michael Krigsman: What you're saying is—correct me if I'm wrong—it seems like a good idea but not yet.

Bulent Kiziltan: It depends on who is saying it's not a good idea. I think, for pharma companies, I think it's essential, critical to invest into that domain even though I predict, for the coming year or two, pharma companies will end up buying the solutions that startups have produced, mainly because it's much cheaper that way. Startups can move much quicker in their drug development efforts, currently.

I think it's essential for pharma companies to be ready and have a scalable infrastructure and talent pool when the time comes. There is a big opportunity cost if they don't do so.

Also for startup companies, I think moving quickly and attracting top talent, being culturally viable for data science is one of their advantages. But they have a big disadvantage. They don't have the data they require. So, I think it's also essential to build collaborative relationships between big pharma and startups to alleviate that problem.

Michael Krigsman: Are there any specific examples of that that you can point to that you're comfortable talking about?

Bulent Kiziltan: For instance, one of the startups that have proven that their predictions show promise, they have partnered with big pharma to use their compound libraries to inform their deep learning and machine learning algorithms to make further predictions. Obviously, the predictions that they're making are not public yet, but I would assume that they're making progress in that sense.

Michael Krigsman: Give us examples of where the use of data science and machine learning has not met expectations or failed. Give us some insight, at the same time, as to what went wrong.

Bulent Kiziltan: One reason data science has produced some justified skepticism in many domains, including pharma, is the leaders of that domain, as I said, they were essentially not domain experts. They had maybe unrealistic expectations of what data science can do today and were projecting for the future.

Data science has yet to produce therapeutics and medicines that can be used in cell therapy, in gene therapy. Those are areas where pharma is growing mainly because data science has to offer new insights in those domains.

Until now, we haven't seen data science-driven insights into drug discovery except a few startups that have shown some promise in that area. But we don't have a single drug yet that has been predicted and has gone through the pipeline to produce a drug that's viable and applicable.

You know we have to consider also that data science is not the only part to blame here. In regular drug discovery or drug development, 1 out of 20 drugs that go through the pipeline is successful. There is a very high attrition rate anyway in the standard approach, so data science is promising to increase those odds.

Michael Krigsman: Given the challenges, why do you say that data science is an existential necessity for pharma?

Bulent Kiziltan: As I just pointed to the cost of drug discovery. Right now, I think it's past $3 billion per drug to be produced and go to the market. The projections show it's going to hit $5 billion in the coming years. That projection, that cost is just not sustainable.

Either the price will go up to a level where it will not make sense for pharma companies to invest into them, which essentially will make their portfolio even smaller and the revenues will go down, or they have to come up with new ways. Data science and AI is potentially the only way that we know of right now to cut down the cost to optimize the whole process and even come up with new insights just given that pharma companies have huge compound libraries that scientists have not been able to effectively tap into to produce insights in the past.

I think it won't be farfetched to predict that AI and data science machine learning will produce new insights. We just don't know the level of impact it's going to have in the future, but I'm optimistic.

Michael Krigsman: What's the timeframe, do you think, that we will actually see some type of material result as opposed to the, shall we say, theory today that it seems like a good idea?

Bulent Kiziltan: Data science is already playing a role in cutting down the costs. Also, producing new insights, making the clinical trials process more effective, especially doing patient enrollment that produce data that will produce information that's more useful for the data science process. I would expect, since the data is already coming in over the past year or two, I would expect a drug or therapeutics to come out that has been largely influenced and empowered by data science processes, but that doesn't mean that data science and machine learning is yet powerful enough and implemented well enough to make predictions for a three-dimensional structure and to produce therapeutics from end-to-end.

I think that is the holy grail to be able to simulate things in the computer and produce information and empower the people who are developing the drugs from end-to-end, which will primarily drive down the costs significantly, which will make therapeutics accessible to people who have rare diseases. It was not financially viable for pharma companies to invest into what's called orphaned diseases, diseases that not more than 200,000 people in the U.S. are suffering, but they are significant. Now, data science processes will make those drug development efforts viable and possible for people with genetic disorders and mutations. It's one of the areas in which pharma companies are currently investing in cell therapeutics and gene therapeutics.

Michael Krigsman: We have another question from Twitter. Arsalan Khan asks a great question having to do with the data. "Given the fact that the data is so important, is this different from any other area, type of domain, where we have to gather and aggregate large amounts of data? In addition, specifically in drug discovery, are there perception and bias issues that are obstacles to progress and getting the results that we want?"

Bulent Kiziltan: Yes, aggregate data is very important in all domains. The problem in pharma is that the risks are high. You cannot just probabilistically make a prediction, see, and test it on the ground to get results and then basically produce that drug.

The stakes are very high. It's not like a marketing effort where you can produce a model that's based on probabilities that you're producing and then you can iterate in the market in real-time to inform your algorithms to perfect them. That's very difficult even in clinical trials because there are certain strict regulations that regulate what you can and cannot do.

You have to be very transparent. You have to have very finely sampled data rather than just granular data. You cannot aggregate data and average out certain aspects of that data, which you can really do in marketing data and some other domains. The risks are very high in pharma.

To the second part of the question about biases, yes, bias exists everywhere because we humans are biased. Our biases are reflected into the data.

There has been an ongoing discussion of whether machine learning algorithms are biased. I would argue that the algorithms reflect the bias that's in the data. Certainly, that bias also exists in the pharma space. We have certain methodologies to overcome those biases, but it's an ongoing effort by all parties, both in academia and in the industry.

Enterprise architecture in data science and pharma

Michael Krigsman: James McGovern asks, "Any thoughts on how enterprise architecture adds value, in general, to data science, but if there are specifics that you can talk about in pharma, that would be great."

Bulent Kiziltan: It's one of the most critical components of data science operations, I think. One of the reasons why data science has been slow in delivering and why the data science efforts have been hampered is mainly because IT and the enterprise architecture has not been incentivized to keep up with technology in the past. Going into an infrastructure that is not up to date, that doesn't use the latest technology, and does not speak organically with the data science efforts was one of the problems and has been one of the problems in larger companies.

In order for companies to move forward and produce the value that data science is promising, quickly, I think an effort that is company-wide where all stakeholders are on board is essential. The specifics of what type of enterprise architecture is to be used is very domain-specific. It's very specific to the business objectives of the company, the timelines, the resources, and the talent pool. I think that there are a lot of solutions in the market which can be customized to the company to make sure that data science can deliver today.

Michael Krigsman: We have another question from Twitter. This is from the @CXOTalk account. "What part of the drug discovery process do you see data science making the most significant contributions in drug discovery?"

Bulent Kiziltan: I think the frontier today is to make predictions on the three-dimensional structures of molecules and compounds with the computer using machine learning algorithms. Currently, those are all done manually with robots, but they are basically manually mixing certain components and trying to see the effect on target molecules. This is, again, done manually.

I think the frontier is where machine learning can go into the compound library and know what the target is and make predictions based on previous trials. I think this is where machine learning can and will make an impact, but it will take time.

Collaborative nature of data science

Michael Krigsman: Talk about the collaborative nature of data science and drill down and compare that to the siloed nature of historically the way pharma companies work.

Bulent Kiziltan: Coming from academia, the culture is, unfortunately, not something that we have always talked about. But as we move into the business and how things can operate and produce value, culture comes up at the top of our list all the time in many of the use cases that we are working on.

Culturally, the business mindset expects certain outcomes and they are very rigid in their thinking. This is the traditional way of approaching things. Also, in pharma, different silos and stakeholders, they want certain results and they are used to getting certain results.

Data science, by nature, has a significant exploratory component to it. I think this is not being appreciated by many business stakeholders. If you keep a rigid project management pipeline and expect certain outcomes without redefining how you operate, it will be very difficult for data scientists to use their creative capability to contribute to the operations and to the value that data science can produce.

There are two outcomes to this. One outcome is in cases where businesspeople who have been at the company are strong stakeholders. They will push the data science teams to keep producing the rigid outcomes that they're used to and it will limit the production and the creative contribution a data science team can do. Whereas, companies who are on the way of transforming and redefining their culture, sitting down with the data science leadership, having a dynamic interaction, producing the key performance metrics together, and to keep it dynamic, I think, is essential because data science, as I said, is very exploratory in nature.

This is why some of the project management styles that people have been trying to implement into data science simply does not work because the outcomes are not well defined. Whereas, you can borrow some of the ideas of project management such as Agile and some other project management styles and implement it but not as is.

If companies think of data science as a software engineering project, it will limit, significantly, their capability. For stakeholders that are a part of the discussion, I would advise them to not be so rigid and try to have a two-way dialog rather than asking data science teams to produce certain outcomes that they're used to having.

Large drug research companies working with startups to drive innovation

Michael Krigsman: You mentioned earlier the importance of large pharma companies working with startups. The larger companies have access to data and processes. The smaller companies have access to more advanced technical approaches. Once you start mixing those two, don't you layer onto it another whole set of very difficult cultural and economic challenges?

Bulent Kiziltan: That's correct. One of the solutions some of the companies have applied was to build an internal group that is somewhat isolated from the rest of the stakeholders internally, essentially acting like an internal startup. This is one way to solve some of the problems, potential problems, that you just mentioned.

I think the traditional approach nowadays is to collaborate and build relationships with startups. It really depends on how that relationship is being structured.

If you have an agreement with a startup but have that startup act like an internal stakeholder, I think it won't solve some of the problems. Whereas, you can really build a relationship with a startup that isolates the startup as a separate entity but reports only to the top and then, organically, builds a very flat structure with some of the teams that can complement to the efforts, I think, is the way to move forward. I think companies are eager to build that sort of relationship. I've seen similar relationships being built in the industry.

Michael Krigsman: From this standpoint, there's really not much of a difference from looking at innovation approaches across different kinds of industries. The pharma space is not unique from that innovation standpoint of partnering with startups and trying various things to make it work out successfully for both sides.

Bulent Kiziltan: There are unique aspects to it because pharma is also very siloed, as I mentioned. That brings a certain level of secrecy. Unless that trust has been built, which takes some time and sometimes years to build, the lack of transparency hampers those efforts.

Once that relationship has been built, I think the companies and the stakeholders can move beyond that secrecy and share information transparency. It's one of the problems building those long-lasting relationships that are based on trust.

Michael Krigsman: Is the secrecy aspect in drug discovery really one of the most important obstacles to that innovation partnership?

Bulent Kiziltan: I wouldn't say it's one of the most important aspects, but it definitely contributes or hampers the effort of building those relationships. When people and companies sit down at the table and discuss how they can contribute to their efforts and how mutually they can benefit from that relationship, it typically takes several months to a year to move through bureaucracy and make sure that the data, the challenges, and the fiscal priorities are being shared transparently.

Disruption is an obstacle to innovation in the pharmaceutical industry

Michael Krigsman: To what extent do large pharma companies have an interest in deploying innovative processes that disrupt the established organization and disrupt established executive positions and, potentially, compensation? This is a generic problem. It's not just drug discovery, it seems like. How big an issue is it?

Bulent Kiziltan: I think it's a very big issue. There is a high risk that comes with cultural transformation. There comes a high risk when it comes to changing the status quo internally. What we see is, when you come as a stakeholder or leader in one of the branches within the company, to have that vision but you are not empowered or you don't share that vision with the board, the CEO, or the CTO, what we see is it never happens. It never matures to a level where it can be implemented and executed on.

I think companies that really have decided at the board level to drive that cultural transformation, which has to go in parallel with the digital transformation, change their leadership from the very top. They change their culture and they make sure that it precipitates down to the stakeholders, which will take time, yes. But I think the first and foremost transformational point is at the very top. Unless you have a board or leadership that drives that transformation, it will never happen from the bottom up.

VC investment startup drug companies

Michael Krigsman: Kanupriya Agarwal asks, "How much funding do you anticipate being set aside for AI and data science in drug discovery in the near future? What do you see the dynamics at play here?"

Bulent Kiziltan: It's a difficult one to answer because each company is making its own decisions. I think the trend moves in a direction where a significant amount of money will go into internal investment. But also, I think, at least in the short-term, because pharma companies won't be able to complete that digital transformation and cultural transformation in the short-term, they will continue to buy, partner, or invest in startups.

I think there will be a significant budget. Obviously, many billions of dollars that go into external partnerships and purchases to expand their portfolio in the target areas. But also, internally, I think it will be more incremental. But I think, in order to build capabilities that are scalable, it takes a long time and companies know that they cannot just wait, not invest internally, and expect to buy out startups.

Two years from now, when the technology is mature enough, they cannot just hire people and then do it internally, which will obviously be the choice to make. I think incrementally building that internal know-how and capability while they diversify their risk is the way to move forward. It's the trend that we see.

Michael Krigsman: In your view, there's an opportunity for startups who are able to gain success with this to be acquired by larger pharma companies going forward.

Bulent Kiziltan: It's the right time to invest in startups that build therapeutics in different domains in gene therapy, in proteins, in oncology, and in different kinds of therapeutics and drug discovery. I think it's a very good time to invest in those companies.

Michael Krigsman: Data, however, is going to have to come from the large companies, or is that not correct? Do startups have a chance of developing their own data?

Bulent Kiziltan: It'll depend on the company and what exactly they're doing. Some companies are just building algorithms. Some companies want to be a part of or drive the whole drug discovery process and, typically, those companies are already backed by big investors. There are companies that just tackle certain aspects of the clinical trials process and optimize that. It will really depend on the company, their funding, whether they're being backed by big companies and VCs.

Michael Krigsman: What about the role of VCs and investors in pushing this forward?

Bulent Kiziltan: There are several kinds of VCs. One type of VC wants a short-term gain. But I think, in data science and machine learning, particularly in biotech, the investors are more interested in midterm to long-term gains because it will be transformational. But VCs are interacting, obviously, across the industry. They know where things are going. If they're smart investors, they will diversify their risk.

There are some companies that potentially will have results in the short-term, but there are companies that are in for the long-term. I think data science machine learning in the context of drug discovery is somewhat of a midterm investment to a long-term investment whereas pharma companies internally will continue to invest in the short-term as well to be ready when the technology is mature enough.

Michael Krigsman: Bulent, as we finish up, what have I not asked you? Are there any fundamentally key issues that we have not covered that we need to talk about for the sake of completeness?

Bulent Kiziltan: One of the aspects that we have touched upon in our previous conversations is also very valid in this context, and that's leadership. A part of the cultural transformation, I think that the most important aspect of driving the transformation internally is the profile of the leader.

Pharma, as I mentioned, is traditionally very territorial. When you look at the leadership that is driving those efforts, they have been in the industry for more than 10, 15 years. The leaders that they're typically hiring have been in the pharma industry, out of academy, for many decades.

When you have data science efforts primarily led by business leadership, it certainly adds to the data science efficiency. But I think the business mindset does not work really well with long-term data science strategies. I think it is essential to find leadership that has domain expertise, has been in multiple domains, and can adapt to the pharma domain fairly quickly.

Repeating the same thing, keeping the same structure, having the same mindset and then expecting out-of-the-box transformational results just doesn't work. You have to have leaders that can think out-of-the-box, can bring the creative aspect of data science into the operations of your company, especially in pharma, so they have to be open to the hiring of leadership that is out-of-the-box.

Michael Krigsman: If I can push back slightly on what you just said, of course, it makes sense to hire leadership with domain expertise, both in data science and in drug discovery. However, if you don't bring a businessperson to that party, then you may not be able to construct an economically viable model and set of processes for enabling this to be durable for the longer-term.

Bulent Kiziltan: Absolutely. I didn't want to recommend hiring a person that does not consider business priorities. This is where collaboration is critically important and the approach to data science operations is to build a collaborative relationship with all stakeholders, especially with the business arm of the companies, to make sure that everybody has an open mind and can go through a process in which they educate each other rather than having a rigid mindset. That goes both to the business side of things and also to the data science of things.

Michael Krigsman: I notice that, on your bio, it says Stealth Startup. How would you like to totally spill the beans and no longer be stealth? [Laughter]

Bulent Kiziltan: For companies that move out of stealth mode, there are different models. If you have a product that you are ready to talk about and share with the public, I think it's a good time. But typically, nowadays, because the R&D loop is very fast, much quicker than before, companies prefer to come out of stealth way before the product is mature enough, but right after, when they have a proof of concept. With the companies that I'm involved in, both on the advising side of things and also building technology, some of them are not there yet.

Michael Krigsman: We have been speaking with Bulent Kiziltan. He is a senior data scientist and C-level executive. Bulent, thank you for taking the time to talk with us today.

Bulent Kiziltan: Thanks for having me, Michael.

Michael Krigsman: It's been a very interesting and fascinating discussion. Before you go, please subscribe on YouTube and subscribe to our newsletter at the top of our website. You can hit the subscribe button. Thanks so much, everybody, and check out CXOTalk.com. We will see you again next time. We have great shows coming up. Bye-bye.