In a prior videos, I explained how AI can affect innovation, what type of innovation, how organizational structures from practices, employment change, managerial change are all needed for firms to really effectively leverage AI. That may seem a little abstract even though I gave you many examples throughout the way. For example, IBM Watson, Holocene, the car chassis, the arts, but let me provide a deep drive into the drug industry, the pharmaceutical industry, to give you an example of what's really going on that drug space. Drug innovations are important, especially we have just gone through the pandemic. In total, there are about 30,000 diseases by various bacteria viruses and only a third of them can be treated and drugs are extremely expensive to produce. It takes about more than 10 years and a typical drug costs $2 billion to produce. As in part of that complexity of this drug discovery is very large. The human biological system is extremely complex and that involves a combinatorial space of more than 10^60 molecules. Because of that, we've seen the COVID-19 pandemic, we've seen superbugs, anti resistant of superbugs. We've seen Ebola outbreak a few years earlier. I'm going to show you how AI could potentially address these issues. This is an issue that's been at the mind of many firms big and small. For them, AI has ability to correctly interpret external data and learn from such data and use these learning to achieve a specific goal and tasks is critical. What is the effect of AI on drugs? How do we use AI to exploit data about known drug compound and predict whether that compound actually can be developed into a drug that human can consume safely? Atomwise and many other startups have developed complex, deep learning neural networks and deployed them on the rich data we have about drugs, about biological interactions, about medicine. As a result of these large corpus of data, two new drug compound can be discovered for Ebola disease within one week and this corresponding patterns has been filed and that's actually quite typical for many other conditions as well. This is AI investment Pharma industry. [inaudible] are larger trend but especially in the more recent years, there's a huge AI investment BioPharma firms. What we are really interested in is, drug as innovation. To what extent do AI can really help solving this a really long, complex innovation process? Remember, in our general studies we shown the AI is really great at finding recombination innovation or intermediate novelty in time citations scale it like it is best at the sweet spot between 50, 60 percent coming from your own knowledge and 40s-50 percent from other people's knowledge. We replicate this for drug novelty. Here drug novelty is measured by chemical novelty in a [inaudible]. How is this structure novel compared to all existing chemical compound out there? Interestingly we find very similar things, even those not exactly the same measurement recombination that we find that in generally, AI is great at the middle range. AI is now great at finding really brand new chemical compounds, I just haven't seen before, is effective for incremental but not as great. But the biggest [inaudible] is with an intermediate level. But AI really can help you, can help pharmaceutical firms finding drugs that intermediate level novelty. If you think about intermediate level novelty is probably likely to be some kind of recombination innovation. Remember, AI and data analytic is a really a pattern matching machine. It finds all kinds of correlations and use those correlation to make predictions. But we all know that correlation is not causation and if you find the wrong correlation that means a $2 billion mistake in the drug industry. Even though we can come up with many different possibility of potential common compound in addressing a disease condition and many of that correlation could just be spurious. It may be wrong. How do we know which one is which? Because the fact can be dramatic. AI could have a negative federal drug innovation if the wrong target or wrong compound was ping-picked for further development. We looked at this issue by classifying these drugs. It's compound into, do we know about a mechanism impact or how they target or treat a disease, or if we don't know the mechanism? We look at the AI's effect on these drugs or these new chemical compound that we discovered through AI. The first bar is, the case is when no mechanism has been found. We don't know the mechanism impact on how this particular drug target disease. There are many drugs we know it works but we don't exactly know how. If we do know exactly how maybe it's easier to create drugs. That's why we sometimes a lot of diseases caused by viruses are very difficult to treat because the mechanism is shifting we don't know exact mechanism. Where bacteria we do know because we know how antibiotics kills bacteria. In that case we do know what a mechanism is. Looking at the first bar when no mechanism is known you see the AI's effect is minimal. It's pretty much zero. AI does not help you find compound when you don't know the mechanism impact. It makes sense because you're blindly looking for correlation, you don't know which one correlation is right, and whatever correlation you find may be spurious. But however, if you do know the mechanism in the case of bacteria halicin. When you find one you know it will work, you know why it will work. We have the ability to verify because you already know the mechanism and AI can have a great effect. But even where there are no mechanisms you see that in terms of chemical novelty in those three graphs, it's really the middle or medium novelty that's driving the effect. Even if you know the mechanism, if it's really incremental effect, basically the chemical compound has incremental improvement of existing one, and it's effect is relatively small. It's positive is relatively small. When is very novel, very new, again the effect is almost zero. Most of the effects that we're seeing are AIs effect on a known mechanism. It come from drug's candidate of intermediate novelty. The main takeaway is that is really hard to know which drug candidates are real, which ones are false. Again AI is a great data mining machine finding all kinds of correlation and pattern hidden links you have never thought about before, you did not seen before. For example, a new way of combining things in a new way, or a way to link two things you haven't seen before. But in terms of finding really novel drug compounds which then the affect is relatively minimal. But again that it's not saying that this new combining a new thing, a new way, is not beneficial. They're extremely beneficial. This is how we find halicin. Potentially helping us address our next superbug problem. This is how we find the COVID-19 vaccine using AI. Just because they're not necessarily radical, just because they're a new way of combining things is still means is extremely useful. But the key here to think about whether innovation is really a recombination and innovation, or whether you've already known the mechanism of how it affects the disease. The mechanism impact really help you to discern between spurious correlation and true causation. The way we think about this novel innovation, really the first of this new type. If you think about it as I mentioned before in our earlier study we don't see in patterns study, we don't see it's effects on this first of its kind innovation. Remember a D in a prior video, the first technology of this class. If you think about how AI can affect this class of innovation in drug case too is hard because they're so new, there's very limited data about that particular drug, about that particular condition. There's lots of tacit knowledge in the clinical experience of people, and that's not exactly codifiable. It's very limited in terms of what AI, machine learning, and data analyst can do to aggregate and detect patterns when this tacit knowledge cannot be used because it cannot be codified. If you think about the breakthrough drugs, the novel therapy for malaria. This is discovered by Dr. Youyou Tu, she's a 2015 Nobel Laureate. This drug come from a single line of ancient Chinese text, combined with a little clinical expertise, and deep understanding of malaria conditions. This is how Dr. Youyou Tu link the single line of ancient text to the creation of the first malaria drug therapy. Again, in this case, AI really can't do much. Limited data, even there's more than one line of data AI can't use that, can't use any of a lot more data than that. Lots of this clinical experience, tacit knowledge is not codifiable for AI to use. Maybe one day they are when they're ready to be used, maybe we can make breakthrough. But right now, tacit knowledge are not codifiable yet, and they're very limited things we can do so we use AI for this type of discovery. The main takeaway is that AI can help drug discovery at a very early stage in the compound discovery stage. The mechanism for doing that is that it can go through millions of compounded in a very short amount of time, and finding a lots of correlation, lots of pattern that you haven't seen before and this pattern are critical to generate hypothesis about specific drugs. The effect is especially strong when you already know the mechanism impact for that drug is because that's a great way to verify this drugs works or not before you go down a two billion dollar path in clinical trials and figure out whether this thing works or not in the clinical trial, and if it's failing that is investment down the drain. We've seen that in general innovation using patent data, the AI really helps drug discovery by finding the intermediately novel compound, and that are pretty much similar to recombination innovation. But for really novel drug therapy, by novel, I mean chemical novelty. It is still limited to use because there's isn't data out there about that particular drug. Much of that we have almost still relies on human ingenuity.