# Lecture 16.2: AI in Healthcare ⧸ Drug Discovery  (AI4Health & AI4Medicine)

[Recording](https://www.youtube.com/watch?v=KCx1EEzNs4M)

Automatic transcript; technical terms and equations may contain recognition errors.

00:00:00 And apparently, AI's impacts have reached healthcare.

00:00:06 Okay?

00:00:07 There's reasons like science.

00:00:10 There's other reasons like money.

00:00:12 But for all kinds of reasons, right,

00:00:14 AI is rippling its impact into healthcare.

00:00:18 So these are just some summaries of the kind of applications you can have in healthcare.

00:00:24 I talked about the virtual assistants before, right?

00:00:28 But now if you think about these virtual assistants in the healthcare, right,

00:00:33 it actually saves a lot of cost and time for both sides.

00:00:39 Okay?

00:00:40 Very often and more likely in future,

00:00:42 you will first check in with a virtual assistant before you go to hospital

00:00:46 and you will be directed somehow, right?

00:00:49 And not just the, you know, these appointments and records

00:00:57 of course and data collection are assisted by AI.

00:01:00 Some medical diagnosis will also be assisted by AI, especially,

00:01:05 so in other words, the decision making is also assisted by AI, right?

00:01:10 The diagnosis, right?

00:01:12 So including the medical image analysis, right?

00:01:16 So what's funny is that, you know, one of the fathers of AI actually said a few years ago that, you know,

00:01:33 one of the people who are interested in the field is radiologists, right?

00:01:36 Probably if you are entering the field, I would strongly recommend you not to study radiobiology

00:01:42 because these jobs will be replaced by AI products in a few years.

00:01:48 That was his comment a few years ago, okay?

00:01:51 But I think it's been a long time ever since and I don't think that is happening

00:01:57 because there's a lot of things that AI cannot replace.

00:02:01 Even for the relatively mature application of AI, that's image analysis, okay?

00:02:07 Drug discovery, that's a new wave, right?

00:02:10 As I said, you know, AI is making advances in what?

00:02:14 Natural language processing.

00:02:16 So you see that once the NLP products get matured, the technologies ripple into these medical assistants, right?

00:02:25 And electronic health record and so on, right?

00:02:29 And then the image, right? Computer vision, right?

00:02:32 It's another mature field for AI.

00:02:35 And then naturally, you know, the second wave of the AI into healthcare is in image analysis, right?

00:02:42 Right now, you know, a lot of the pharmaceutical companies are analyzing these images at all scales, right?

00:02:50 From the cell level, right? Molecule level, cell level, right? Tissue level and also in the history of these tissue images, right?

00:03:00 Or, you know, brain images.

00:03:02 So all these images are actually massively analyzed, right?

00:03:05 By AI products, which are assisting doctors to make decisions, right?

00:03:12 And now it's entering drug discovery, which is much more difficult.

00:03:17 And it's not clear exactly it's going to help.

00:03:22 But nevertheless, AI is entering drug discovery.

00:03:27 We'll talk a little bit more about it in the next slide.

00:03:30 It's a much more difficult problem.

00:03:32 And it's not necessarily a field that AI is already good at, okay?

00:03:37 Because a lot of the problems in drug discovery are not even well formulated.

00:03:41 And not to mention that some of them are formulated but not supported by a lot of data.

00:03:47 After all, AI is a lot driven by data for now, okay?

00:03:52 I know what the original definition is talking about.

00:03:56 But still, it's very often we are talking about the deep learning model that is as good as the data.

00:04:05 And that is not artificial intelligence, what you want to claim about.

00:04:11 But it's more than that, right?

00:04:14 There's a medical data security, automated workflow assistance, medical risk prediction, fraud detection, clinical trial design.

00:04:23 These are more in the healthcare part, okay?

00:04:26 So there are a lot of opportunities and career possibilities in all these different kinds, right?

00:04:33 There's also a lot of consulting companies needing experts who have both data science and domain science expertise.

00:04:43 In this case, domain science being healthcare.

00:04:48 I want to talk about drug discovery a little bit because this is the latest wave.

00:04:53 And it's getting a little bit crazy.

00:04:56 And why people are, you know, seeing so much AI technologies rippling into drug discovery.

00:05:08 And so much that almost every major IT company has a branch dedicating to healthcare or maybe even specifically drug discovery already, right?

00:05:23 There are some bad news including Google Health recently, right?

00:05:28 But overall, they are just investing in healthcare a lot.

00:05:33 And a lot of the startups are focusing on drug discovery.

00:05:38 Because drug discovery is, in short, a very expensive, lengthy, and easy to fail profession.

00:05:51 Okay?

00:05:52 So if you look at how long a drug discovery and development takes, there is a pre-discovery, right?

00:06:01 So you have to, you know, find out the targets for a disease, right?

00:06:08 And then for the given targets, you have to identify hits, right?

00:06:13 You have to identify small molecules or biologics that can hit the target.

00:06:19 And then you will find some of the leads among them and then optimize their properties in all kinds of ways.

00:06:26 Pharmacodynamics, pharmacokinetics, all kinds of properties, right?

00:06:31 And then some of them will, you know, luckily make into the latter phases.

00:06:40 Okay?

00:06:41 So that first stage, the preclinical, drug discovery and preclinical, already takes three to six years in general.

00:06:49 And then it enters, so if you find something right here, so let's say here, right?

00:06:57 If you find some candidates, they will enter the clinical trials, right?

00:07:02 And some of them will make it, but even if we look at these candidates entering the clinical trial,

00:07:09 which is already a long way, right?

00:07:11 After six years, eventually less than 10% of those that enter clinical trials will eventually make it to the market.

00:07:21 Okay?

00:07:22 So in the clinical trials, there are three phases, right?

00:07:24 You essentially want to show that it's, you know, it's efficacy, it's safety and other properties, right?

00:07:32 That takes another, right, six to seven years.

00:07:36 Once the clinical trials, phase one, two, three passed, FDA review, and then, you know, you will get approval and all these considerations,

00:07:48 including the manufacturing scale up.

00:07:50 And afterwards, there's still post-marketing surveillance.

00:07:54 So if we add these years together, how many years are we talking about?

00:08:04 How many years?

00:08:05 Nine and a half, like 10 years, right?

00:08:10 So it's a 10 year.

00:08:12 So, and the success rate, well, I don't even say starting from here, you start with, you know, 10 to 10 molecule candidates, right?

00:08:23 Well, first you have to identify the targets already.

00:08:26 I mean, even if the targets are known, you may start with 10 to 10 molecules and you find maybe dozens of them.

00:08:33 But then, even if you enter the clinical trial, it's 10%.

00:08:37 Okay?

00:08:38 But going from here to here is, you know, before the clinical is one out of, I don't know, a million or less.

00:08:46 So success rate is really low.

00:08:49 And the average cost to bring a new drug to the market, this was an old number.

00:08:56 It definitely didn't consider the inflation.

00:08:59 I don't know whether the current number is roughly the same because on one hand, you know, inflation is increasing, right?

00:09:07 But on the other hand, does AI really help make it cheaper?

00:09:10 Well, I don't think so.

00:09:12 Not so much yet.

00:09:13 It's definitely the hope we will see later on.

00:09:16 But back then, the old number I got was $770 million, okay, on average, to make one, bring one drug to market.

00:09:26 Okay?

00:09:27 So, usually those big phamas have just handful or less patentized the drugs that basically, you know, let them flourish for two decades.

00:09:38 So, because of drug discovery is such a lengthy, low success rate, and expensive process, right?

00:09:53 And also, it's making a lot of money if you make it, right?

00:09:56 So, there's a lot of faith now.

00:09:59 I don't know, I don't think faith is the right word.

00:10:02 I think that, you know, you're looking for opportunities, right?

00:10:05 You're trying to make this, you know, drug discovery process a little bit shorter or more likely to succeed or cheaper.

00:10:13 You're trying to find all help you can, right?

00:10:15 And they're now looking into AI.

00:10:17 But I want to remind you, I don't have it here, but there is a cover I, you know, I retrieved from the cover of Time magazine.

00:10:28 Okay?

00:10:29 The cover says the breakthrough of drug discovery, and then it says computer-aided drug discovery.

00:10:36 Do you want to guess when, which year was that cover of Time?

00:10:48 It was early 2000s, more than, you know, almost 20 years ago.

00:10:53 I'm just going to tell you that, to some people, right, they feel like computer-assisted drug discovery, aided drug discovery, is not a new concept.

00:11:03 Now, you know, they feel like it's now just disguising a new buzzword of AI.

00:11:08 So there's a lot of disbelief, okay?

00:11:11 But nevertheless, there are a lot of investment into these companies.

00:11:18 There are a lot of startups, okay?

00:11:20 You know, I strongly recommend you, some of you are looking for internships in the summer.

00:11:25 I strongly recommend you to look into these startups.

00:11:28 They are partnering with a lot of big pharma.

00:11:32 They are looking at a lot of important problems, okay?

00:11:37 I don't know whether I have this example here.

00:11:40 No.

00:11:41 So if you studied probabilistic graphical models and use a textbook by Daphne Kohler, if you know about Daphne Kohler, she started the company four, five years ago.

00:11:56 We'll see the company's name again in five slides-ish, called In-Citro.

00:12:01 And they basically, you know, got investment from Gilead at the very beginning.

00:12:07 Gilead is a major pharmaceutical company.

00:12:09 So Gilead basically says that, well, let's spend two phases.

00:12:13 In the first phase, if you can find up to how many targets for these diseases, right, that's quite a success.

00:12:19 This is the amount of money I'll give.

00:12:22 And in the second phase, for the given protein target, if you find some amount of the, you know, lead small molecules, right, as promising drug candidates, that's another amount of investment I will give, right?

00:12:35 And that was just the beginning.

00:12:38 So there are a lot of such startups now.

00:12:40 This was 2019, right?

00:12:42 This estimate was 2018, actually.

00:12:49 Back then, the market of AI in medicine was valued at 700 million and estimated to reach 18,000 million with a cumulative aggregated gross rate of close to 50%.

00:13:02 And of course, between 2018 and 2022 this year, a lot of things happened, noticeably COVID.

00:13:11 So on one hand, COVID definitely stalled a lot of things.

00:13:15 But on the other hand, COVID also accelerated some of the things.

00:13:19 For instance, messenger RNA-based vaccines, right?

00:13:23 So I don't have the right number right now, but I'll look into it.

00:13:28 I suspect that it wouldn't be very far from their estimation.

00:13:31 And I have other numbers to back up.

00:13:33 So you have seen AlphaFold.

00:13:37 This was the first version of AlphaFold, okay?

00:13:40 But I think the real breakthrough is AlphaFold 2.

00:13:44 That was, you know, the year before last year and published last year.

00:13:48 And so I think I should update the left.

00:13:52 And then there's generative models, right?

00:13:54 So some people are interested in faking images.

00:14:00 Some are interested in faking videos, right?

00:14:04 So, but some other scientists, let me not use the word fake, are interested in creating or synthesizing small molecules or drugs, right?

00:14:15 So this is, I think this is in Silicon Medicine, right?

00:14:21 So they published a paper that drastically accelerated this drug discovery stage, right?

00:14:29 Within 40, 50 days, they found a promising candidate, actually.

00:14:34 And the roadblocks, as I said, you know, some people have been in the pharmaceutical companies for long, right?

00:14:44 They remember the last wave of computer-aided drug discovery, right?

00:14:50 So they know that there are a lot of, you know, things that are more complicated than what AI can address.

00:14:57 But I think I really like this code, right?

00:15:01 A lot of the AI in healthcare research papers or technologies are actually focusing on the low-hanging fruit,

00:15:10 which is not a bad thing, okay?

00:15:12 Because, you know, you don't solve the ultimate problem right away.

00:15:17 You have to, as I said, artificial intelligence, great question, but you need to plan for the milestones

00:15:23 and gradually approach the ultimate question, right?

00:15:26 You cannot ask, can machines think like humans?

00:15:29 Then how do you approach that problem?

00:15:31 And how could the answer be useful, right?

00:15:34 Same here.

00:15:35 It's impossible, right?

00:15:37 Almost impossible in the current setting to think about end-to-end AI product

00:15:44 that will just solve the whole drug discovery and development process.

00:15:50 That's almost impossible and unnecessary either.

00:15:53 So it's very natural that a lot of the AI research nowadays, including the journal called papers I was referring you to,

00:16:01 are focusing on the earlier stages of drug discovery that has the problems well formulated, right?

00:16:10 You are looking for proteins, I'm sorry, you are looking for drugs that can bind to the targets well, right?

00:16:17 And then it has good ADME properties and so on, right?

00:16:22 So those problems are well formulated and those problems are focused on the chemistry more than biology.

00:16:29 And the chemical rules are more or less better understood than biology.

00:16:33 And more importantly, there are way more data, right?

00:16:36 You can quickly essay these small molecule compounds and get a lot of labeled data, right?

00:16:42 So that's why I say that I feel like a lot of problems are focusing, AI technologies are focusing on the low-hanging fruit.

00:16:51 But if you want to make this process drastically faster or cheaper, you have to look at the higher high-income fruit or the tougher bottlenecks, right?

00:17:04 I think that's fine. I mean, I think that will gradually, you know, right, get there, right?

00:17:10 When you cultivate a new technology, you don't want to just, you know, throw it right away into the most challenging environment, right?

00:17:19 You see what I mean? It's all just like, you know, when you learn to swim, right?

00:17:23 Right? It's impossible that your first swimming lesson is in the Atlantic Ocean or the Pacific Ocean, right?

00:17:30 Right? Can you imagine that being taken to the middle of the Pacific Ocean by a ship and then your coach tell you,

00:17:38 now jump and swim. It's impossible, right? Where do you start?

00:17:42 You start with a swimming pool, five feet, right? That's how you start, right?

00:17:48 You gradually get to seven feet, four feet. You start with four feet.

00:17:52 You gradually get into seven feet and then maybe you will get to open water and deep water, right?

00:17:58 The same for AI. So I think that's good.

00:18:01 As long as we are not, we, the AI medicine or AI health researchers, are not content with the comfort zone slash low-hanging fruits, then that's good.

00:18:14 So let's keep humble in front of the demands in the healthcare industry.

00:18:25 This is recent news just published in February and actually a lot of developments.

00:18:37 The state of AI in drug discovery in 2020. So there's, you know, as I said, starting from the biology part, you have the protein-protein interaction network.

00:18:50 And AI is used a lot to model the disease and discover the target, right?

00:18:56 And then with the target known, AI is used a lot for virtual screening, finding the candidates that may work for the desired targets.

00:19:07 And then retro synthesis and even small molecule generation. The screening is against a known library.

00:19:15 Generation is starting from nowhere, like starting from scratch, right?

00:19:20 And then they may also enter the clinical development, right?

00:19:26 You probably noticed that there is a whole, you know, process between chemistry and the clinical development in the drug discovery that's not reflected in this image yet, right?

00:19:37 Because that's a higher hanging fruit for now.

00:19:39 But also toward the end, the clinical development is also achievable by, you know, AI.

00:19:45 You can stratify the patients, you can design the clinical trials, and then you can predict the clinical trial outcomes.

00:19:54 And so that is still a well-formulated, conceived and data-supported field, okay?

00:20:01 Not as abundant data as chemistry, but still.

00:20:06 A bunch of news already.

00:20:08 At the end of year 2021 and early 2022, already a lot of things are happening.

00:20:14 Even though we feel like we are still recovering from COVID.

00:20:17 I want you to pay attention to those companies in the bold.

00:20:21 These are major farmers.

00:20:24 And those companies that are in bold italics, those are the AI startups focusing on medicine.

00:20:33 So I talked about in-sitro, right?

00:20:37 And you see that I think the biggest, one of the biggest is, I think should be Roche.

00:20:46 So Roche has a deal with recursion and aiming for up to 40 new medications for cancer and neurological conditions, starting with $150 million investment.

00:21:00 Sanofi has a deal with another AI company.

00:21:04 They are aiming at 15 novel small molecule drug candidates in oncology and immunology.

00:21:10 $100 million upfront.

00:21:13 And if all future milestones are hit, that can reach $5.2 billion.

00:21:17 So a lot of interest, a lot of passion and a lot of money are flowing into the field.

00:21:25 So I strongly recommend that if you don't know about some of this company's name, I should italicize this one.

00:21:32 As the name suggests that this is apparently focusing on the de novo protein design, right? Generated models.

00:21:41 Some are small molecules. Some are proteins like EPCI or generated biomedicines.

00:21:49 I strongly recommend you be familiar with these opportunities because now the opportunities are everywhere.

00:21:55 There are three kinds of opportunities in industry if you're looking at AI health.

00:21:59 There are traditional IT companies that are venturing into health.

00:22:04 So their working environment is more IT, but they do a lot of relevant research in health and collaborating with domain scientists.

00:22:14 So almost every major IT companies have their health branch.

00:22:20 Another kind of employers are major farmers.

00:22:26 So almost every major farmer has its data science and AI branch now.

00:22:32 Some of you must have known some alumni of our own department who went to major farmers.

00:22:41 And they are working on AI.

00:22:44 Not necessarily directly related to medicine, but there's no news at all.

00:22:50 And the third kind of employers are these italicized startups.

00:22:56 So they are directly working on the development and application of AI technologies for some kind of drug discovery.

00:23:07 Small molecule or proteins and so on.

00:23:11 So I strongly recommend you to pay attention to these opportunities because it's very exciting and it's happening.

00:23:21 Okay.

00:23:22 And somehow every time I show this image there are some issues.

00:23:30 Now let's come back and ask ourselves, right?

00:23:33 So now that we have so much data, we can actually think about our problems differently.

00:23:38 Previously for the simple problem, well not so simple, for the problem of protein structure prediction,

00:23:44 you start with a sequence and end up with a structure.

00:23:48 Back then I asked you how do you solve this problem computationally?

00:23:51 I got two kind of answers, right?

00:23:53 One is, you know, what is the causal relationship, right?

00:23:58 I think that's what's, hey, mean, right?

00:24:01 So in other words, what's the principle that can guide me solve this problem?

00:24:06 We know that that's what we did.

00:24:08 In the midnight zone, we used the principle that's the energy minimum principle, right?

00:24:13 And another kind of question is, you know, what kind of data can support me make such prediction?

00:24:20 Just like the machine learning, if the system is too sophisticated, but I have abundant IO data,

00:24:26 then I can learn machine learning model to recreate the observed IO relationships, right?

00:24:33 In both cases, right, we have seen, right?

00:24:37 But our learning of the IO relationship back then was very naive, isn't it?

00:24:45 Back then we look at small amount of data and we try to see what's the sequence similarity,

00:24:52 how the sequence similarity translates to structure similarity,

00:24:56 how similar two sequences need to be so that their structures are similar,

00:25:00 so that we can directly borrow the template's structure, right?

00:25:05 The sequence template structure.

00:25:07 So back then, the identification of the signal was naive

00:25:12 and the utilization of the signal was also naive, right?

00:25:16 So now actually there are apparently abundant data as we've already seen from the success, right?

00:25:25 So you can think of it now from the data perspective, right?
