Live on Bloomberg TV

00:00We've been speaking to the team at ISO Labs, Demis Hassabis and his colleagues, and they have forecast that they could reduce preclinical drug discovery through their A.I. models in a couple of years time, from years potentially to months, potentially even faster than that in a few years time. Is that realistic from what you are seeing? Well, let me just say that the technology that as at Isomorphic Labs are developing and also many other for that matter startups and companies are developing is very exciting. However, I do think that we need to be a little bit cautious because in order to come up with a molecule that is indeed useful and will make it all the way through the regulatory process is another thing. Hmm. So yes, discovering molecules, I think it will be tremendously accelerating. But the first phase of a clinical trial takes much more than just discovering the molecule. You need to discover who will benefit. You need to discover what does this medication should be given it to be safe and yet provide the efficacy it needs. So yes, we may discover a lot of potentially good molecules. We may be able to do much better preclinical studies and hopefully in the process make sure that fewer animals are used and maybe no animals are used to validate the drug and it works. And hopefully also we are going to make it all the way to a successful drug. And data is essential to this. And clinical data. Medical data can be siloed, it can be biased, it can be very messy. And the models are only as good as the data that goes into them. Good data and good data out bad in bad data out. How much of a hurdle is data for these accomplishments? Data is very much key, but that is where I'm very optimistic because there is quite a lot of data. There is data from the medical domain, from the clinical domain. There are quite a lot of datasets from prior clinical trials that may be pertinent to the current molecules. There is also a lot of biological molecular data. And yes, this data is messy, it contains missing data and all of that though, and they need to be harmonized across these different types of datasets. But this is where I really can excel. So there is a very big field that is called data centric AI, where the focus is no longer just to build an AI, to build models such as language models or the type of models like Alpha Fold, but also AI to improve the quality of the data, augment datasets, homogenized datasets, clean datasets. On the point of bias, though, we need to be very careful. We need to be very careful that these molecules are indeed created in a way in which in the end many people will benefit by the final drug. So this is why when I say we need to be excited about these technologies, that those are morphic labs and others are developing ways to be excited, but also cautious because at the very end we need to come up with drugs that are going to be better than the current drugs. And ideally a wide range of people can benefit from them. And that is an entire journey, not just the molecule. And part of that journey is the clinical phase. That under current structures can take what, 6 to 7 years, sometimes longer. How could I potentially speed up that part of the process? So this is exciting, but much more complicated. If you allow me maybe even more complicated in discovering a new molecule, because an entire clinical trial is very regulated and for good reason, and it needs to identify, for instance, in the first phase, who should be recruited in the trial, who would benefit from it. And here is where hopefully I can make the recruitment process more generous, more inclusive, because currently we tend to include in clinical trials the very wealthy and very healthy. And what we need is much better representation of we really are all also we need to potentially use the AI to personalise things like those currently in the first phase of a trial. We need to determine what's the safe dose and at the same time the dose with which we have the right efficacy. Being able to do personalized dosing is very, very important and this is where I can do maybe quite a lot. Then in the second phase we do really we conduct the trial and here is where maybe we decide that only a subgroup of patients will benefit or we may want. Will stop the court on trial and evolve it into a different trial. Or we may want to have what is called synthetic control arms. Right now, when we randomize, we randomized into an arm that receives the dragon one that a placebo arm. Increasingly eye is capable to understand what would happen to the patients that are not that are not receiving the drug for something that's called into the control arm so that we can predict what would happen to the patients that would have been all in the drug but are not in the in the trial but are not actually getting the drug. And I can himself cost and speed up the process. And at the very final stage when we have that, the the drug the trial has been really conducted. We passed through the analysis phase and here is where tremendous advances have been made in the last ten years. In I work on causality, causal effects in France, which enables us to identify not only the population that will benefit, but maybe whether subpopulations will not benefit, and trying to really understand how to target this medication to a particular subgroup. And then as the drug is hopefully approved, we go really to the deployment phase. And this is a situation where we have many drugs and doctors need to be able to identify the right drug for the right person. And here is where again, I can do a lot more in trying to identify which drugs are best for this individual. For how long? When can we stop taking the drug? When maybe the drug stops working and we need to move to another drug? What does that tell us about the time frame? How far away or close we are to personalized medicine then? Well, in terms of technology, a lot is already available both to determine and identify disease early, because we know that ideally you want to identify for most diseases the disease in its incipient phase because that's when it's more suitable. Then I can help with personalizing again, as I mentioned throughout medicine to give potentially to are those we should be giving it, but also how to monitor it in a personalized way. This particular patient, because many drugs unfortunately may have side effects, but also maybe it's a moment in time the drug may not need to be taken anymore. So how to monitor a patient that may have a particular disease, for instance, cancer over time, not in a one size fits all like apparently is happening, but rather in a personalized way where some patients may not need to be monitored too frequently while others much more. And finally, side effects. Increasingly, I can also determine whether a particular drug may have side effects for a particular individual, but not for another. And that is all part of, let's say, a more holistic and a more personalized medicine when it comes to the regulators, what would need to happen? What would they need to see to have confidence in an discovered drug? I think that they even have them. The way in which this drug was developed is not transparent. And let's be honest, many of the machine learning models that is in today, we don't understand always why something has happened. And I hear that's fine when you discover a molecule, but the process of really passing for a rigorous clinical trial that really is trying to identify whether this drug is indeed working, if it's working and for whom. What is the largest inclusion criteria? Are most people benefiting from it? And then really identifying true benefit for a particular subpopulation is very important. So the clinical trial process can benefit by, but needs to be a very remain a very rigorous process to be able to identify whether this newly discovered molecule by is indeed useful. And for whom is it realistic to think that that clinical phase stage of the process, 6 to 7 years is going to be shrunk significantly so and within a short time frame? Or is that still a far off proposition? I hope it will. I hope it will. But I think that what people need to realize is that as we are building these molecules that are more sophisticated, potentially are targeting subpopulations or very specific types of disease, we need also to identify correct biomarkers as to who would benefit by the drug. I can help there as well, but we need to be cautious and identify these types of of drugs of of biomarkers such that indeed we can identify who will benefit because increasingly we are building drugs that are very specialized and they could be very potent. But we need to identify who benefits and hopefully keep the. costs low, As well. Such that indeed these drugs are making a difference in the general population. Not very well. Not only very wealthy. So the only. So, yes, I share the excitement, but I also want to say we need to be cautious and identify ways in which we can speed up the entire process of validating the drug and then delivering the drug in a way in which is equitable and it is affordable for the majority of people. Otherwise, we didn't solve any problem. And you're an academic with deep expertise on AI and machine learning, but you also worked in industry as well. You've got dozens of patents to your name. How does this technology, if it evolves as you expect, change the balance of power for the farmer industry? Well, I think that we may see very different new blends of future companies that will build drugs. I think that increasingly the pharmaceutical industry is joined by this very new startups that are building molecules, that are building labs. So I do think that the entire process will change. For now. I think pharma, often one sees an interesting molecule, is buying that particular company. But if companies of the type office or morphic labs are starting to bring molecules to market. I don't think that's going to be that easy. So I on the positive side, hope that this competition that is going to happen is going to be actually very healthy. And I do think the pharmaceutical industry with which at times I work together with, sees the need to to revamp the way of thinking. And I hope that we are going to see a lot of competition, a lot of innovation and hopefully the find finally the patient is going to benefit at large. So I'm excited because there are many diseases, cancer, including where we still don't have the miracle drug. And I hope that this competition is going to energize everybody and provide us with exciting new drugs that will be sold in the lot of current diseases that for which we can't have a solution to the cure to cancer as a result of discovered drugs. Does that happen in in our lifetimes? I hope so. And again, it's not only about drugs, it's about using AI also to identify diseases much more thoroughly and then finding the right drug for the right person. And that is where, again, I can contribute a lot, but that's a different blend of that. So it's not only about molecules, it's also about building the right decision support systems for the clinicians to be able to identify what patient, what drug. And do these innovations ultimately democratize medicine or lead to further gaps between wealthy nations and wealthy patients in terms of who gets access to these medicines? This is why I said I'm very excited, but I think we need to be very cautious because the technology has the potential to bring a lot of drugs to market much faster and to a much wider population. And I hope, though, that the people who are building these drugs are planning also to to think through as to who would benefit from it, because if it if it's only a very small population will pay a very large amount of money. We didn't solve a real world problem. On the other hand, it will be drugs whose cost can go down very fast, partially because the pipeline, but currently the farmers are asking at times a large amount of money because it took them a large amount of money to develop these drugs due to failures potentially, or because the entire pipeline is very long and expensive. Hopefully by reducing the pipeline as Isomorphic Labs is saying, we are capable also to reduce the cost of bringing a drug to market. And also due to the competition, the cost may be reduced even further. There is a potential that the drugs will become much more accessible and at a lower cost, and I'm excited about that. But we need to make sure that indeed companies work towards this goal rather than just increasing the revenue, forgetting the fact that actually they have a duty to and a commitment to the finalization.

No schedule data available.

BTV Channel FinderWatch BTV in your area

Mihaela van der Schaar on AI, Machine Learning for Drug Discovery & Medicine

Mihaela van der Schaar, Director of Cambridge Center For AI In Medicine, says caution is required when artificial intelligence-powered tools are used in the medicine and pharmaceutical industries. "We need to come up with drugs that are going to be better than the current drugs. And ideally a wide range of people can benefit from them. And that is an entire journey, not just the molecule," she said. Van der Schaar spoke to Bloomberg's Tom Mackenzie on September 5. (Source: Bloomberg)

More From Bloomberg Technology
All episodes and clips