Merging AI and Quantum Computing To Boost Drug Discovery
Quantum computing is a growing trend due to recent major advances in hardware making this technology a practically feasible thing. Some experts compare today’s rapid progress in quantum computing with a similar period during the last century when personal computing was emerging through a cascade of technological advances both in hardware and software.
Will quantum change everything? When? Are we there yet?
To find our insights about the above questions and learn a thing or two about what the quantum computing industry looks like today, I sat down with Dr. Christopher Savoie, Co-founder, and CEO at Zapata Computing -- an American quantum software company on the cutting-edge of research in this area.
#advertisement
Christopher is a scientist and a serial tech entrepreneur with 20+ years of experience in the technology industry. He is the inventor of the Natural Language Understanding (NLU) technology behind Apple's Siri and he was recognized as a top business leader and innovator by Nikkei and MIT Technology Review.
Christopher co-founded Zapata Computing out of Harvard in 2017, along with his colleagues Alán Aspuru-Guzik, Jhonathan Romero, Jonathan Olson, Peter Johnson, and Yudong Cao.
The below text of the interview has been modified for size. Watch the original video for the complete interview.
Andrii: I would like to first start with a very general question -- there are headlines involving top companies like Google, for example, about “quantum supremacy” and a lot of theoretical work has been done in the field. Now it seems like the industry is finally getting to some practical valuable applications. Can you explain where we are right now with this technology?
Christopher: We're in the early days of this technology and the hardware itself is still being developed as we speak. There are many types of ways to make these qubits so the theory that a qubit can work has already been proven experimentally. We know that qubits work, we have them, we have these computers, they can do relatively simple calculations right now -- so the theory behind it is not even a question.
The question is, however, which of the different ways of making a qubit work are going to be the most effective and the most scalable solutions. So, Google and IBM have this superconducting type of qubit and that has its advantages including the speed of the gate operations that it does. Then there are other types of these computers that have nothing to do with superconducting: they are done with ion traps and there are neutral atom ways of creating qubits and photonic ways of creating qubits… So all of these competing platforms are trying to create a hardware device that will perform this theory of superposition and entanglement and all of these things to do the wonderful calculations and, at the end of the day, become fault-tolerant, meaning that they are robust enough to continue in their operations and do calculations perfectly. So they are all doing the same calculation, the same thing but they do them in very different ways. There's a lot of development in each one of these areas so it is very nuanced and very hard to know from a user's perspective or from a pharmaceutical company’s or, for that matter, from any other company's perspective which one of these platforms is going to be ideal for any given problem or solution. We are in the early days of discovery and I am of the opinion that it is not a zero-sum game -- it is not going to be just one of these solutions that wins everything. I think all of these various platforms will be good for certain things, certain applications.
RELATED: 12 Companies Using Quantum Theory To Accelerate Drug Discovery
For example, an ion trap is right now of a higher fidelity -- the qubits have less error built-in them than the superconducting qubits that IBM or Google might have. So companies like IQM and Honeywell have these machines but their gate speeds, the speed of operation, is much slower. So there's kind of a trade-off here. If you want to have lots and lots of samples really quickly -- then the superconducting ones are going to be preferred -- as long as you can tolerate a little more error. Contrary, if you have something that needs that fidelity very much -- maybe the ion traps are a better solution. So the choice very much depends on the needs, on the use case. In 2019 Google released a paper, they said they accomplished something that would be difficult to do on a classical computer, even a supercomputer, they compared it to the Oak Ridge computer. It wasn't though a very particularly useful example that they did -- it is not like the problem that they solved could be directly applicable to chemistry or biology or anything that we might care about or even drug delivery or anything close to that -- but it showed that at 53 qubits you could do something that a classical computer would have a difficult time doing. So we are getting to that regime where we can see that there will be some applications and use cases where a quantum computer, empirically and experimentally -- not just theoretically, is going to be superior to classical computers.
We have shown in recent results that we can get enhancements to, let’s say, machine learning using a quantum computer. Even a relatively small quantum computer can give us some advantage in the way that we do neural networks and things like generative adversarial networks (GANs). That points to perhaps a quite near-term possibility that things that we care about, like neural networks and machine learning, maybe be enhanced by quantum computers, so we are at that point where this is starting to get very real.
Andrii: What is the solution that Zapata is bringing to the table? What industries are you targeting and what are the use cases?
Christopher: My partner in crime, my co-founder ‪Alan Aspuru-Guzik -- it was in his lab that the variational quantum eigensolver (VQE) was invented, a very important first chemistry algorithm that can be run on these noisy near-term quantum computers. So, our group from the very beginning had a lot of experience in actually running these algorithms on various platforms like iron traps and the superconducting ones -- not just coming up with algorithms. VQE has been run on all of these platforms because it is maybe the first and most famous algorithm in this area. Actually, all of the algorithms that have run on noisy intermediate-scale quantum (NISQ) devices as we call them are in some ways the progeny of this VQE algorithm -- so because of this we have been able to become who we are today. As a group, we were able to attract 20 plus of the best PhD level scientists in this area to work at our company as researchers and so we have on the one side of the company some of the best people being able to look at a problem and figure out sometimes a novel algorithm to solve it. Sometimes there is not yet an algorithm that works on in this computer to do these jobs and we have the people who can look at the mathematics of something and come up with a novel way of actually solving that algorithmically on a near-term quantum computer or even for a fault-tolerant quantum computer that may be developed in the near future.
On the other side of the company, we have a bunch of people who have enterprise software experience like looking at how we actually deliver an algorithm in a machine learning context at scale on real data. For example, there are 10 to the seventh small molecules we wanna screen. Okay, how would I actually do that? Where's that data sitting? How do I arrange that data? How do I get everything to work in a workflow in an actual enterprise system which is not an easy thing? You can come up with the algorithm and have the most beautiful algorithm in the world -- and we have seen this in Artificial Intelligence space, for example. But if you actually want to apply that to drug discovery -- that is a harder thing to do to put that in production.
So we have these kinds of two sides to the company -- we have the capability of looking at a problem and coming up with the algorithm, and the other side of the company is our ability to actually engineer the software that would deliver that. I guess the important thing to understand here is that you never do just quantum computing in isolation -- that will never work out. A lot of what we have to do is still “classical”: getting the data, preparing the data, understanding what we have to run, understanding the problem, pre-processing things, sending it to the quantum computer that maybe lives on the cloud, bringing that result back, post-processing that result and so on. Most of the work is classical and it is “boring” software engineering -- but you have to do it to make this actually practical. So that's really the combination of what Zapata is -- we have some of the brightest minds in algorithmic theory for quantum computing combined with the ability to actually software engineer that into a system that will work in a real company to deliver a real solution. Because, at the end of the day, if you talk to the C-suite at biopharma companies, they just want the answer. They really don't care if it is a quantum computer or something else that's giving them the answer -- it may be a pocket calculator, whatever.
RELATED: Combining AI and Quantum Mechanics to Improve Drug and Vaccine Discovery
They want to know which hit is going to become a lead that's going to become the right drug. Then, which patient should be in the clinical trial, what exclusion criteria -- all of these kinds of things. Then how to distribute the drug -- that is also a big computational problem.
So, giving those answers to people is what really matters -- this is what we do. We look at a problem and we can tell a customer: “Nope, I don't think quantum is your solution, you should just use a classical neural network for this and here's an example of how we can do it”. Our job is to compare and contrast the various solutions and then deliver the right solution to customers -- a scalable enterprise-grade software solution, not just an algorithm.
Andrii: You mentioned drug discovery -- what Zapata is doing in that sense, and what are the relevant use cases with quantum computing and theory? Are there any “low-hanging fruits” there to start with?
Christopher: I used to run a pharmaceutical discovery company, and I can say at every stage of drug discovery there are mathematical problems. Quantum will get us to a point where it's possible to find optimal solutions for many of them -- from chemical synthesis all the way through clinical development and even supply chain and demand chain modeling. For instance -- distributing vaccines for tackling COVID-19 rapidly and efficiently -- is a big problem and it's not optimal right now. So there are a lot of places along the entire value chain in biopharma that will eventually benefit from quantum technologies. The key is “when”. Some of these will have more near-term uses because we can use these noisy intermediate-scale devices to solve those problems and other problems will require fault tolerance -- basically a perfect quantum computer which may be a lot farther away along the road. It is nuanced. For instance, our dream is that we will do some very complex things with quantum computers that will be able to do the ab initio Chemistry and look at the intermediate states -- these highly correlated states. So when you take a small molecule and you bind it to a protein, a GPCR receptor, or something like that -- which one is going to be the best binder? We would love to be able to predict that and actually even predict a biochemical cascade that happens in a systems biology -- before we go to the lab and do a lot of expensive experiments before we do combinatorial chemistry at a large scale, and so on -- that would be the ideal situation. Unfortunately, to do that accurately we need a very powerful quantum computer and we've had some research papers out recently that show that we're going to need thousands of millions of qubits to do that, with very high fidelity, meaning they don't make errors very often. That's not where we are today. It does not mean we shouldn't start doing the research and pushing that field forward because the payoff is huge.
Having said that there are some problems that will be nearer-term and highly valuable. I believe that generative adversarial networks (GANs), for instance, can be used to do wonders in terms of predicting completely new molecules with desired pharmacological properties. We have shown if you use a combination of classical neural networks, like GANs, and some bits of quantum -- we can start to use these nearer-term devices a lot quicker to do something that's quite useful in that space. That is an area that I am excited about -- the chemoinformatics side of discovery, I think that'll be nearer term.
In the other places of the value chain machine learning is being used, for example, in clinical trial work to try and do patient selection, identification of biomarkers, and other things like that. Quantum technologies can provide nearer-term enhancements for all such use cases, improving and complementing the machine learning side of things.
Doing some clinical trial design work may seem not as exciting for quantum scientists as doing the drug discovery and the ab initio chemistry calculations, but it can really move the needle for companies because at the end of the day you need to get these drugs to patients.
Andrii: Artificial intelligence, or more specifically, machine learning and deep learning, is a growing trend in the pharmaceutical industry right now and you just mentioned complementarity between quantum technologies and artificial intelligence technologies. As I understand, quantum technology is going to be one of the enabling technologies for accelerating progress in the area of artificial intelligence, correct?
Christopher: Absolutely! I think that's really the nearest term where we're going to see tangible use cases with quantum tech. A relatively small noisy quantum computer can also still have an enhancement in that area and so we look at this as a software problem with quantum components. Zapata is really not in the quantum software industry because there's no quantum software industry alone, without classical software. We are in the data science and machine learning business -- and quantum will be a part of every, or almost every, data science and machine learning workflow in biopharma in the future. I do believe it will be an integral part of it. If you can get a more accurate model by using quantum tech -- why wouldn't you do that, after all?
Andrii: To me, it sounds like with quantum technologies it will be data science and machine learning “on steroids” -- exciting! I feel quantum computers will bring a lot of disruptive use cases for the pharmaceutical industry. But speaking of today, what would be a smart strategy for a pharmaceutical company to start embracing quantum technologies right now? What could they practically do to start implementing this with a future competitive advantage in mind?
Christopher: It depends on the culture of each company. Some companies are more willing to partner with outside vendors, than trying to create an internal process. Also, it depends on the existing workforce. Are you going to really double down on having internal experts who are quantum algorithms experts, for example? If you don’t have enough relevant talent, you have to start hiring those teams and developing that kind of expertise. Alternatively -- start partnering with folks who are in the quantum industry, the hardware makers, or companies like Zapata.
Probably, collaboration is the best bet here. You can have one or two, or a handful of relevant experts, but having a hundred cross-disciplinary experts who really understand the algorithms, understand the math behind it, and have domain expertise in biopharma -- seems challenging even for the biggest pharma companies. But you need to also develop internal expertise to make successful partnerships. You need to have, probably, one or two people who at least understand enough about the hardware, enough about the software systems, and quantum tech to take advantage of it. A lot of these companies have a quantum chemist who can at least understand the language of quantum mechanics enough to get involved -- become an internal “champion” for pilot projects and collaborations. Start small with some proof of concept that's going to lead to a result that will help you benchmark, you know.
At Zapata, we work with companies to come together and really work through the use cases first. The next step is to look at these use cases at a deeper level and really look at the math and be very truthful to ourselves -- where quantum can truly provide value, and where it is an overkill.
If you can find something that's going to create a real business impact, a tangible ROI, and that's technically feasible with the nearer term quantum devices -- that's the sweet spot to try. When you get the first ROI, the investment will start to turn around and we'll be able to invest more into this area.
My feeling is that quantum technologies will be picked up primarily in the first place by companies who already invest in things like deep learning, they are already well versed in digital and data science innovation.
Andrii: How does your company work with clients? What does a typical partnership workflow look like?
Christopher: Our main offering is our Orquestra platform. At the end of the day, we are a software company. We can develop applications -- complex data science applications -- over this system that we've developed, Orquestra. We can do that with a pharmaceutical client together, helping them deploy applications, or work with a vendor systems integration consultant or vendor of the large companies that we know that already are a vendor to these companies.
Sometimes there's a question is there even an algorithm, is there even a way to do this on a quantum computer -- we also do those engagements and that's where our scientists get involved and look deeply at the problem and try to find a novel solution so we can partner in that way as well and build the application with the partner. At the end of the day, the deliverable is a workflow that will work in your R&D environment. We make our money by offering the solution that will actually run these systems in production -- that's where we want to get our partners to, it is the ability, maybe, to create for themselves the intellectual property. We are not interested in taking the intellectual property and owning it, or having any percentage of drug revenues because we offer a computing solution -- that's not our model at all. We see our role as a way of enabling biopharma to make use of this novel means of computing in their value chain in any way that we can bring it to them, to help them create the intellectual property and create the drugs, and bring those to patients -- that is our goal.
Andrii: It sounds like a very customizable kind of solution and a very iterative process of creating value for the customer.
Christopher: It has to be because we're in the early days! You know, anyone who says they know the perfect solution for this, or they know how this is going to actually work -- they are just lying. Besides, every company is different, like I said, some companies have five or ten people already working on quantum projects -- that is one case scenario. The other case is you only have one person who is like the “quantum champion” -- and there is not yet budget or awareness among executive teams -- those are different situations and they have to be handled differently. We are flexible to work with a partner, meet them where they are on this journey. We see our goal in enabling them to get further along -- and that will be different from company to company, group to group, and project to project.
Andrii: One last question, what's your major ambition for this year for the company?
Christopher: I would say in practical terms I am really excited about version 2.0 of Orquestra -- that's on our roadmap to be released later this year and we're really excited about the features. I can't talk about it openly today so much, but under an NDA we can discuss this with anyone who wants to hear about our roadmaps. We are extremely excited about the new level of software engineering capabilities it will give us and the ability to do some of these really cool things with data science and machine learning.
Particularly, I think, people will be pleasantly surprised with what we have there and I think we'll be making some announcements with some of the results that we're having in biopharma and other industries. We are excited about the pace of the research and some of the algorithmic developments. It is a great year to be in quantum information science and in quantum computing!
Topic: AI in Bio
