• 6 months ago
The implementation of AI in healthcare spans from developing drugs to using AI in the real world. This panel from Imagination In Action's 'Forging the Future of Business with AI' Summit talks about how AI could manifest in the healthcare system.

Subscribe to FORBES: https://www.youtube.com/user/Forbes?sub_confirmation=1

Fuel your success with Forbes. Gain unlimited access to premium journalism, including breaking news, groundbreaking in-depth reported stories, daily digests and more. Plus, members get a front-row seat at members-only events with leading thinkers and doers, access to premium video that can help you get ahead, an ad-light experience, early access to select products including NFT drops and more:

https://account.forbes.com/membership/?utm_source=youtube&utm_medium=display&utm_campaign=growth_non-sub_paid_subscribe_ytdescript

Stay Connected
Forbes newsletters: https://newsletters.editorial.forbes.com
Forbes on Facebook: http://fb.com/forbes
Forbes Video on Twitter: http://www.twitter.com/forbes
Forbes Video on Instagram: http://instagram.com/forbes
More From Forbes: http://forbes.com

Forbes covers the intersection of entrepreneurship, wealth, technology, business and lifestyle with a focus on people and success.

Category

🤖
Tech
Transcript
00:00 (audience applauding)
00:03 - Thank you, John.
00:05 Good morning, everybody.
00:08 How's everyone feeling?
00:09 - This is amazing.
00:10 - So I'm Colin Hill.
00:13 I'm the CEO and co-founder of Aetia.
00:16 We are considered to be one of the first AI companies
00:19 in biomedicine and healthcare.
00:22 We were essentially doing AI before AI was cool,
00:25 is how people like to introduce me.
00:28 So really excited about the panel this morning
00:31 because now we're gonna talk about
00:32 where AI starts to matter, matter the most,
00:35 in discovering and developing new drugs to cure disease
00:39 and in really transforming healthcare.
00:42 So just to frame the discussion
00:45 on the discovery and development of new drugs,
00:49 what are the problems that we have to solve?
00:52 First off, we have to figure out
00:53 what's the right drug target
00:55 out of all of the gene products,
00:57 and we have hundreds of thousands of them.
01:00 What do we need to hit to stop Alzheimer's
01:02 or colorectal cancer?
01:04 Once we figure out what the right drug target is,
01:06 we have to then make that drug.
01:08 So what's the right drug against the right drug target?
01:11 Then once we have the right drug,
01:13 we have to figure out which patients will respond.
01:16 So on the side of creating new medicines,
01:18 these are the three problems that need to be solved.
01:21 And just a quick FYI,
01:24 it doesn't work out so well most of the time.
01:26 More than 80% of all drugs that enter clinical trials fail.
01:30 So how do we change this?
01:32 How does AI start to really change this paradigm?
01:36 Number one, we've been living through
01:38 the exponential increase in human multi-omic data.
01:42 The human genome was sequenced in 2001
01:44 at a cost of $11 billion and took more than a decade.
01:49 We can now get a genome in less than a day
01:55 at a cost of less than $1,000.
01:57 We of course have the exponential rise in compute power.
02:02 And now we're at the stage where the supercomputing power
02:07 can now reverse engineer systems at scale.
02:11 And we have also the emergence of new kinds of AI
02:15 that get all the way to mechanism and causality.
02:17 So these three have converged
02:20 to enable a wholly different way
02:23 of approaching drug discovery and development.
02:25 At our company, we use Gemini digital twins
02:28 that we create from the data, from the compute power,
02:31 and from causal AI to now start to reverse engineer
02:35 the hidden 95% of circuitry of human disease,
02:39 what really prevents us from making new drugs.
02:41 We're gonna hear from this panel
02:44 on other ways of approaching this revolution
02:48 from the discovery and development of new drugs
02:51 and then the delivery of that care in the real world.
02:54 So Jeff, I'm gonna turn the floor over to you
02:57 to introduce yourself and what you're doing at Flagship.
03:00 - Sure.
03:02 So it's wonderful to be here.
03:04 My name's Jeff, I'm a general partner of Flagship.
03:06 I'm an inventor and entrepreneur.
03:08 20 years ago, I was a undergraduate at MIT
03:10 majoring in chemical engineering.
03:13 And I fell in love with biology
03:14 because of the kind of pictures that you're seeing here.
03:17 Like what we call a protein in daily life
03:19 is eggs, bacon, soy.
03:22 When you zoom in inside of us or inside those,
03:25 they are the most amazing machines, sensors,
03:28 and smallest technologies that we have access to.
03:32 And we don't understand how almost any of them work.
03:34 So based on the idea that during my life,
03:36 we might understand that
03:37 and we might be able to create new ones,
03:40 I was sort of head over heels.
03:44 That has been really difficult to do,
03:47 but for reasons that you'll hear on the panel today,
03:49 artificial intelligence is giving us access
03:52 to some of the rules that are present in DNA
03:54 and the ways that those encode the functions
03:56 of those devices that we call proteins.
04:00 And six years ago,
04:01 I co-founded a company called Generate Biomedicines,
04:06 where we started to imagine
04:07 that you might be able to perform the most valuable thing
04:10 in all of biotechnology today with generative AI,
04:14 which is to predict what the structure of a protein
04:17 would need to be to bind to a target
04:19 at a very specific location.
04:21 Often it's hard to be disciplined
04:23 about one's choices of what to do.
04:25 This is easy.
04:26 That's all of antibodies, all of peptides,
04:28 and almost all of protein therapeutics today.
04:31 And today we go fishing for them
04:33 in pretty laborious and inefficient ways.
04:35 And in some ways that I'll describe,
04:38 generative AI allows you to do extraordinary things,
04:42 of ask questions like,
04:43 what would the DNA sequence of an antibody be
04:46 that would hit this target on a site
04:49 that if you had access to a molecule for it,
04:52 you might be able to cure a disease
04:53 that today we have no medicines for.
04:56 So I'm excited to talk about what the future may hold.
04:58 Thanks, Tom. - Fantastic.
05:00 Ava.
05:01 - Yeah, so pleasure to be here.
05:03 My name is Ava Amini.
05:05 I'm a senior researcher at Microsoft Research
05:08 within a group that we call
05:09 the Biomedical Machine Learning Group.
05:11 And really our vision there is to develop
05:14 and pioneer new AI systems that can empower us
05:18 to make new biological discoveries,
05:21 and also to get at this vision
05:23 of actually optimizing biological design.
05:27 And when we think about what we mean there,
05:29 really the way we are thinking about it
05:31 is sort of at all scales of biology,
05:34 from the molecular level like Jeff described,
05:37 all the way up to the patient level.
05:39 And we see this as an opportunity
05:41 to now really integrate across those scales of resolution,
05:45 to now be able to deliver real solutions and insights
05:50 for everything from biotechnology companies and industries,
05:55 all the way up to the clinical implementation.
05:58 And sort of on background on myself,
06:01 I came to MIT initially as an undergraduate
06:04 and was enamored by this idea of seeing biology
06:08 as the most exquisite computing system that exists.
06:12 And I had this, fell in love with this idea
06:15 of how can we bring the lens of computer science
06:19 to better understand how biology computes, right?
06:24 And so that has kind of propelled my thinking
06:27 and the framework by which I approach research questions,
06:31 which is always driven first by
06:33 what does this mean for human health?
06:36 How can this actually impact human lives
06:40 and improve patient lives?
06:42 And even at a place like Microsoft
06:44 where we're very much technology driven,
06:47 this is the vision by which we're approaching
06:49 this intersection of AI and biology
06:53 in thinking about even when we come to define
06:56 the new AI systems that we create,
06:58 how can we guarantee that they actually deliver
07:01 those insights that will make an impact in the real world?
07:05 - Fantastic, Zach.
07:07 - Hi, it's great to be here.
07:09 My name's Zachary Ziegler.
07:11 I'm the co-founder and CTO of Open Evidence.
07:14 I'm someone, my background, I did my undergrad at Cornell
07:18 and I started a PhD at Harvard.
07:21 But like many other great folks, I dropped out of that
07:24 because this is too important a moment
07:27 to spend in academia.
07:29 This is a moment to be building real things
07:31 and building companies.
07:32 So that's why I'm here today.
07:34 At Open Evidence, we're building the future
07:36 of medical information and biomedical information.
07:40 And we started here talking about
07:43 how do we build new drugs, how do we put that into care?
07:49 And one thing that is at the core of Open Evidence
07:52 is that there's a huge amount of excitement
07:55 around new drugs and new things,
07:56 but at the same time, there's already an enormous amount
08:00 that we already know about biomedical science.
08:04 There's truly just a ridiculously huge haystack
08:07 that is known, that we know as a species,
08:09 and it's a haystack that is growing every single day
08:14 and the rate at which that's growing is increasing.
08:18 And so at Open Evidence, we're closing the barrier
08:21 between just this enormous amount of information
08:24 and our ability to access that information
08:26 and draw really valuable insights,
08:27 both for pharma and drug discovery,
08:29 as well as for physicians and healthcare as well.
08:32 - Awesome, thank you.
08:34 Tanishq.
08:36 - Yeah, hi, my name is Tanishq.
08:40 I am currently a research director at Stability AI
08:43 and also the CEO of this medical research organization
08:47 called MedArc, which I founded.
08:49 Yeah, I've been interested in the medical AI space
08:52 for quite some time,
08:52 and especially applying generative AI to medicine.
08:55 In fact, during my PhD, I was also working
08:58 on applying image generation to microscopy and pathology.
09:02 Back then, we were using generative adversarial networks,
09:05 which most people are using now, diffusion models,
09:08 but back then, that was the state of the art.
09:11 So this is something I've been quite interested in
09:13 since about 2019 or so.
09:15 And now more recently at Stability AI and MedArc,
09:19 I'm most interested in building foundation models
09:22 for medicine, so that's the kind of research
09:23 we've been doing.
09:24 And more importantly, trying to build these models openly
09:28 and collaboratively.
09:30 So we've been collaborating
09:31 with a bunch of academic institutions
09:33 such as Stanford, Princeton, Columbia,
09:37 and we've been working on building
09:39 these open foundation models.
09:40 So we've built some models in the radiology space
09:43 where we built image generation models for chest X-rays
09:48 and a multimodal model for chest X-rays.
09:50 We've also been working in neuroscience
09:52 where we've been doing fMRI to image reconstruction,
09:55 where basically you take the brain activity
09:57 and reconstruct what a person is seeing
09:58 just from the brain activity.
10:00 And we've also started to explore
10:02 medical language models now as well.
10:04 So yeah, overall, I think that open foundation models
10:07 is the future of medicine and healthcare,
10:09 and I'm really excited to be working in this space.
10:12 - Great.
10:14 So one really deep question is, well, why now?
10:17 Why is biomedicine at a tipping point
10:20 in 2024 when it comes to AI?
10:24 What has changed that has enabled this moment
10:28 in what is the most important industry
10:30 and application of technology?
10:33 So Av, I wanna start with you.
10:36 - Yeah, it's a really fascinating question.
10:38 I think before we dive into why now,
10:40 I wanna kind of take a step back and say,
10:43 in fact, we've been building up to this moment
10:46 for quite some years.
10:48 So even a tool like AlphaFold, right?
10:51 AlphaFold, you may have heard of.
10:52 It's this AI model that allows us to predict the structure,
10:56 the three-dimensional shape of a protein
10:59 from basically a single-letter code,
11:03 a sequence of that protein.
11:05 AlphaFold is based on many, many years
11:08 of biochemical research that actually shows
11:13 these principles of how we can use the relatedness
11:17 of proteins to each other to now compare their similarity
11:21 and make predictions about them.
11:23 So we're kind of standing on the shoulders of giants
11:26 when we think about the history of medical
11:29 and biological research that has gotten us to this point.
11:32 But really, I think the inflection point
11:35 comes down to two reasons.
11:37 One, because of the tremendous inflection we see
11:40 in the technology, the AI models themselves,
11:43 in their ability to reason over large amounts of data.
11:48 And secondly, I think we're seeing this shift in biology
11:51 that is taking us from what we think of
11:54 as a reductionist point of view,
11:56 of having a very specific question
11:58 and a very specific hypothesis that we want to go test
12:01 biologically in the lab,
12:03 versus how can we take an unbiased point of view
12:07 and look for patterns and reason about patterns in data
12:11 without necessarily having a hypothesis to begin with.
12:14 And I think ultimately,
12:16 it's probably going to be a combination of those two
12:18 that get us the most insights,
12:20 but the coupling of the powerful AI models
12:24 and this greater appreciation in biology
12:27 for that unbiased perspective has, in my opinion,
12:30 brought us really to this point.
12:32 - And so really the marriage of hypothesis-free,
12:35 data-driven approaches with more now expert-driven approaches
12:40 is what you're saying we're at a tipping point for.
12:42 I completely agree with that.
12:45 In our case, after several years
12:48 of being a platform company partnering with Pharma,
12:51 we made a big change across the Rubicon
12:54 to making drugs ourselves
12:55 in the area of neurodegenerative diseases
12:57 such as Alzheimer's and Huntington's and the rest.
13:01 Let's also talk about the business models in this industry,
13:05 because while AI is really starting to infuse
13:08 what's happening here,
13:10 there's also the question of how to ultimately create value
13:15 for stakeholders and ultimately deliver value for patients.
13:19 So Jeff, since you're a VC,
13:23 what would you say about the business models
13:25 that are now being enabled by this transition
13:29 where we have the data, we have the AI,
13:30 and we have just new fundamental ways of doing things?
13:33 - Yeah, first of all, let me clarify.
13:35 Everything I do is based on technologies
13:37 that I'm a co-inventor of,
13:38 so I'm often not very good at kind of knowing
13:41 what the external world is doing.
13:42 I'm kind of like an ostrich with my head underground.
13:45 But let me bridge it to Ava's really nice answer
13:48 in that the reason to believe there's any new business model
13:51 would be based on the rules having changed,
13:55 and the rules have really changed.
13:57 Like the reason why not before is pretty simple.
14:01 There's a deep irony for everyone
14:03 who's been a life scientist
14:04 that although our intelligence is biological,
14:07 our intelligence sucks at understanding biology.
14:11 So like we've only really had a glimpse,
14:14 sort of like our ancestors long ago seeing a few stars
14:18 and not being able to imagine the expanses of the universe,
14:22 either in terms of how biology really works
14:24 or the full remit of what could be biological technology.
14:32 And biology is this wonderful information technology.
14:35 Like DNA is the blueprint of the living world around us.
14:40 But while we've been seeing that code,
14:43 we haven't been able to read and comprehend that code.
14:46 Like there isn't a life scientist on the planet
14:48 that can glance at a page full of DNA
14:51 and describe to you quantitatively what that DNA produces
14:55 and what its function is.
14:56 So I think from a business model perspective,
15:00 rather than imagine that AI
15:02 might just kind of tweak biology a little bit,
15:04 I think the better starting point
15:07 is to imagine everything being flipped upside down.
15:09 Like some of the dirty laundry of drug discovery
15:13 is that once you've entered the clinic
15:16 with a therapeutic that you've spent
15:18 hundreds of millions of dollars on and a decade plus,
15:21 it still has a five to 10% chance
15:24 of actually becoming a medicine.
15:26 So 90 to 95% of the time,
15:29 we have failed to predict a molecule
15:32 that could benefit patients
15:33 without adding risk to their vulnerable condition.
15:37 And if that becomes 50%, all of medicine changes.
15:42 If that becomes 90%, all of medicine changes.
15:44 And I think you're going to hear variations
15:47 on how the decision making in medicine
15:50 can be vastly improved
15:51 with an overlay of machine intelligence.
15:53 So I don't know what the business models
15:57 of the future are going to be,
15:57 but I think the best starting point
15:59 is a not function on the way that we do things right now.
16:02 - Fantastic.
16:04 Tanish, what do you think is going to have
16:07 the greatest impact over the next five to 10 years
16:10 in this space,
16:11 whether it's what your company is working on
16:13 or what you see more broadly?
16:15 - Yeah, I think I speak more from,
16:18 I guess, healthcare medicine background.
16:20 I think there'll be a huge impact
16:26 with the development of multimodal foundation models
16:29 that are able to process a wide diversity of clinical data
16:35 and that will be assistance for doctors
16:40 for making clinical decisions.
16:42 And I think, yeah, that's kind of where
16:44 the biggest impact, in my opinion, kind of stands
16:47 in terms of changing patient lives.
16:51 Of course, I think there's a lot of value
16:54 in drug discovery as well,
16:56 but I guess I don't work too much in that space.
16:59 But I think there's just a lot of value
17:03 in especially with the amount of data
17:06 there is available in health records
17:11 and kind of also discovering new things
17:15 from these health records as well.
17:18 I think the foundation models will also be very beneficial
17:24 in kind of gleaning new insights
17:25 from these health records that,
17:28 again, you have this notion of big data
17:33 and learning from big data that
17:34 traditional approaches struggle
17:39 to kind of work with these sorts of large data sets.
17:44 And I think these sorts of foundation models
17:46 will help us to find new insights
17:49 that will change patient lives.
17:51 And I think that's a very exciting area to be working in.
17:56 Great. Now, same question to you, Zach.
17:59 Yeah, well, in thinking about what all you guys have been saying,
18:05 there's one thing that's really critical to me
18:07 to, I think, all this discussion,
18:09 which is what you asked before,
18:10 which is what is underlying
18:13 why we're sitting here today, specifically?
18:15 What's changed?
18:16 What's the flip that's switched?
18:18 And the way I understand the moment we're in
18:21 is that we've learned something very specific.
18:24 And that very specific thing is not that we've learned
18:27 across the board what the right incantation is to say
18:32 before we change a generative AI model
18:34 that makes Gen AI work.
18:35 What we've learned is that it's essentially
18:37 the exact alignment of three things
18:40 that makes Gen AI take a step function
18:44 in terms of an improvement.
18:45 And that's scale of model, scale of compute,
18:48 and importantly, scale of data.
18:50 And so when I think what's important,
18:52 what's the world in five years going to look like?
18:56 What I see is that we've just taken
18:59 this actually truly enormous step function forward
19:01 in terms of our capabilities,
19:02 and specifically our capabilities
19:05 when these three stars align and we have these things.
19:08 And I think that as a community,
19:10 we haven't even begun to scrape 1% of what we can do
19:14 with this new glorious world that we live in.
19:17 And so I think what we're going to see in five years
19:20 is that the companies that are successful
19:21 are the companies that identify that,
19:24 are laser focused on using these new capabilities
19:27 to meet real actual needs.
19:29 And what that looks like is anyone's guess,
19:33 but that's where I see the future going.
19:36 And for my fellow folks in this panel,
19:39 I'm intrigued as someone more fully
19:41 in the language model space,
19:42 for folks like Jeff Neva
19:45 in a more kind of traditional biology space.
19:47 How do you get those three things,
19:50 and especially data?
19:51 Great question.
19:55 Should we pile in?
19:57 Sure.
19:57 So Ava mentioned some of the ways
20:01 in which a couple rare factors align
20:06 in historical data sets,
20:07 which are high quantity, high quality data sets.
20:10 We've used two of those in Generate,
20:13 all DNA sequences that code for proteins.
20:15 The DNA sequencing community has been obsessing
20:18 over the fidelity rate of seeing the code of life.
20:21 The second is the three-dimensional structure of proteins.
20:24 So we started to imagine that you might be able
20:25 to perform that most valuable thing in biotechnology
20:28 of predicting what a protein would need to look like
20:32 to bind to the outside of a protein.
20:33 If you could learn the way that amino acids
20:36 do like to interact with one another inside of proteins
20:40 from all of the crystal structures
20:42 that scientists over the past five decades
20:44 have put on the cover of Nature
20:45 and the cover of Science nearly every week.
20:47 And that wasn't enough.
20:51 So that allowed us to kind of crawl,
20:53 but we had to build a laboratory around machine intelligence
20:59 so that millions of generated examples
21:01 could be constructed, quantitatively tested,
21:04 and every one of those successes and failures
21:06 run back through the model
21:09 so as to update their ability
21:11 to be able to make this kind of prediction.
21:12 And now we've been able to do some amazing things.
21:15 So I'll give you one quick example,
21:16 which is three years ago,
21:18 we challenged the team to take the top $50 billion
21:22 of antibody therapeutic sales
21:23 and generate new antibodies
21:25 that would hit the same target in the same epitope
21:28 in the same binding pose
21:29 with the same conformational structure at the interface
21:33 with comparable or better affinity
21:36 without going anywhere near the intellectual property
21:38 of those parent molecules.
21:40 In three months, the team was able to do it for all of them.
21:43 And by contrast, like you wouldn't be able to do that
21:47 in a decade with a large pharmaceutical's
21:50 traditional discovery approaches.
21:52 So it's just like the very start of the rules changing
21:57 when those stars align.
21:58 But unfortunately, a lot of life science data is really messy.
22:02 And so those stars don't align very often.
22:03 So we're going to see a lot of labs built to facilitate this.
22:06 - So let's talk about that a little bit.
22:08 When the stars don't align,
22:10 because those of us who've been in biomedicine
22:12 and healthcare for some time know that it's always harder
22:16 and more difficult to make impact than we'd like.
22:20 What holds us back?
22:22 What will be, what are the critical factors for us
22:27 to overcome for this revolution to really take hold
22:31 and not be like the world post the genomics revolution,
22:35 where it really took a long time to see any impact?
22:38 - Yeah, I really resonate and agree with a couple of points
22:43 that Jeff raised and what you just raised, Colin,
22:46 in terms of what are those bottlenecks
22:48 and what happens when the data is not that good.
22:52 Because it is true that for many of these public data sets,
22:55 some of them are high quality,
22:56 but some of them it is a pain realistically to pre-process
23:01 and get that all in a format that's going to power
23:04 these next generation AI models.
23:06 I think one of the biggest needs is really that intimacy
23:11 of knowledge on both the biological side
23:14 and the computational side.
23:16 To really realize success, we need to be asking
23:19 the right questions biologically and clinically,
23:21 and we need to reciprocate that with the ability
23:25 to deliver the technology to answer those questions.
23:28 A concrete way of doing this at Microsoft Research
23:33 is we have a very integrated partnership
23:36 with the Broad Institute of MIT and Harvard,
23:39 where we are specifically trying to realize
23:41 this idea of closing the loop
23:43 on that experimental computational cycle
23:47 in the context of cancer specifically.
23:49 So we have the ability to generate our own data
23:52 that can train and improve the AI models that we develop
23:57 and actually take the generations or the predictions
24:00 from those models and test them experimentally
24:03 in the real world.
24:04 And so I think that having that close intimate partnership
24:09 between both sides, whether in the individuals
24:12 and the researchers that we train in the next generation
24:16 or through working through collaborations
24:18 collaboratively together,
24:20 realizing these closed loop platforms
24:22 will be really critical to unlock those bottlenecks
24:26 and release them.
24:28 So scientific bottlenecks,
24:29 and of course there's regulatory bottlenecks
24:31 and eventually reimbursement bottlenecks.
24:34 In our last minute here,
24:36 I want each of you to just state a word or a phrase
24:40 that's gonna describe where we're gonna be in five years,
24:43 starting with you, Tanish.
24:44 - Um, I guess the two words, I'll take two words,
24:50 which is open and multimodal.
24:52 I think that there's gonna be a huge focus on,
24:57 yeah, trying to incorporate different types of modalities
25:00 of data from images, clinical notes, time series,
25:05 all kinds of data.
25:05 I think that will be very crucial.
25:07 - Okay, fantastic.
25:07 Zach.
25:08 - Empowered.
25:09 We're not talking about replacing biologists
25:11 or replacing physicians, we're talking about empowering them.
25:13 - Ava.
25:14 - I would say imagination, right?
25:18 The ability to dream and think about new hypotheses
25:21 and actually test them with the aid
25:24 of an empowering tool like AI.
25:25 - Fantastic.
25:26 Jeff, the last word.
25:27 - Cures.
25:28 We've been really timid to use the word cure in medicine
25:32 for good reasons.
25:32 It's been almost impossible.
25:34 I think we're gonna be rightfully more and more courageous
25:38 in aspiring to and achieving that goal.
25:40 - I love it.
25:41 We got more ambitious as we went down the line.
25:44 Let's end there.
25:45 (audience applauding)
25:46 Thank our speakers.
25:48 (upbeat music)
25:51 (upbeat music)
25:53 (upbeat music)
25:56 (upbeat music)
25:58 (upbeat music)
26:01 (upbeat music)
26:04 (upbeat music)
26:06 [BLANK_AUDIO]

Recommended