00:03In 1987, you had American economist Robert Solow quote the productivity paradox.
00:09He was seeing IT investments, billions, tens of billions, hundreds of billions, pour into different companies,
00:15but not seeing a bottom line impact on productivity statistics in the American economy.
00:22And today, we're seeing something very similar when it comes to AI investments in R&D.
00:27There's tons of money being poured in, lots of advances, but we don't yet see an impact on bottom line
00:33productivity statistics across industries.
00:35In fact, it's stagnating, if not declining, in terms of productivity.
00:40So just like the advent of IT and the Internet back in the days,
00:45we're going to have here to rewire the innovation and the operational model of R&D organizations
00:50to fully capture the value that's available and see sustainable impact.
00:57We like to distinguish between two flavors of AI opportunity when it comes to R&D productivity.
01:04You have operational AI that looks at the end-to-end processes and identifies the bottleneck
01:11in terms of time and resource investment needed for each step,
01:17and then deploys a variety of AI and generative AI techniques to increase the speed
01:22and ultimately reduce the cost of these steps.
01:26In R&D organizations across industries, the kind of use cases in that space that are prioritized
01:30are typically around things like document generation for technical and regulatory purposes.
01:37Then you have scientific AI.
01:39And there, we're talking about generating insight to guide the discovery
01:43and the development process of new innovations.
01:48Scientific AI works directly on the scientific data,
01:51whether it's patients, diseases, molecules, materials, chemical reactions, physics, simulations.
01:59And the goal here is to increase the number of shots-on goals,
02:04to optimize the properties of the innovation candidates,
02:08to ultimately increase their value as you market them.
02:13Operational AI is great.
02:15The use cases are more mature.
02:17They're typically more generalizable and easier to realize value on the P&L.
02:21But this is not what's going to build you a sustainable strategic differentiator for your R&D organization.
02:29For that, you'll have to fundamentally increase the pace
02:32and the probability of success of your innovation process.
02:35You will need scientific AI.
02:40So, there's no doubt that the value that scientific AI can bring is enormous.
02:46I mean, if you take just the pockets of individual R&D transformation examples
02:52and the impact that we've seen,
02:54you can extrapolate easily to a doubling of R&D productivity.
02:58And that's the kind of aspiration that you often see life science
03:01or chemical agriculture companies set for their R&D department.
03:05We want to double R&D productivity.
03:10And one of the key recent innovations
03:13that's underpinning the advancements in scientific AI
03:16is the foundation model.
03:17So, let's talk a little bit about foundation models.
03:20Ultimately, just like ChatGPT is powered by a GPT model
03:26that's been trained on a huge corpus of text
03:29in order to learn something really profound about human language.
03:32It's almost mystical.
03:35It's then able to be used for a variety of cognitive tasks.
03:40The same applies to foundation models and other data modalities.
03:45You can have a foundation model on chemistry
03:47that's been trained on a large corpus of chemical molecules
03:50in order to learn something very profound about chemistries
03:53that can then be used in order to identify new molecules
03:58and optimize their properties.
03:59It goes way beyond the mere understanding of similarities
04:03between sequences of atoms.
04:05And the same exists across the modalities that you see here.
04:09You can discover new polymers, materials.
04:11You can use foundation models on images, chemicals, reactions,
04:15patients, diseases, etc.
04:18And what those foundation models allow you to do
04:20is create a representation on any one of the entities on the left
04:24that are just long vectors of digits.
04:28And you can enrich these representations
04:29by bringing data modalities from multiple of the areas on the left
04:37in order to create a computational space that's unified.
04:40That's very simple.
04:41You do algebra on these vectors.
04:42It allows you to do very powerful things like discover
04:45and optimize new innovation candidates.
04:51But ultimately, we're not saying you're not going to need to do wet lab
04:55and physical experiments anymore.
04:57You're still going to need to generate new data
04:59in order to explore the discovery space,
05:02the parts of the discovery space where the model is struggling,
05:05where the model does not understand what's happening.
05:07You need the new experimental data for that.
05:10You also need to generate new data in the wet lab
05:12in order to validate the hypotheses that you developed in silico.
05:17And this is going to happen in what we depict here
05:21as a closed-loop research system
05:22where you have the scientists and the data scientists working together
05:26such that the AI model is generating new hypotheses.
05:29It's guiding the design of experiments
05:31that are executed by the scientists in the wet lab
05:34to generate data that then feeds back into the model
05:37to improve the model
05:38and create a virtuous circle
05:40that's built on proprietary know-how
05:43and proprietary data
05:44and ultimately will give you
05:46a strategically differentiating capability
05:48that increases your innovation productivity.
05:53And this applies across scientific domains.
05:57So here we have an example in the physics space,
06:00in aerospace,
06:01where deep learning surrogates,
06:03a special type of model
06:05that learns to predict the physical behaviors of components,
06:09can be used to optimize the design of aircraft wings
06:12instead of traditional methods
06:14that would go all the way down
06:15to fundamental physics equations
06:17and on the micro elements of the design,
06:20try to compute item by item
06:23the properties,
06:25the aerodynamic properties of the wings.
06:27That allows you here,
06:28you talk about four, five orders of magnitude
06:32acceleration in the simulation time,
06:34which allows you for a much more aggressive exploration
06:37of the discovery space
06:38and in many cases a much better design in the end.
06:42It also applies in chemistry,
06:45where here we have an example
06:47from a biotechnology company
06:49and their drug discovery process,
06:51where we brought together
06:52a range of data modalities
06:54from microscopy images,
06:57the chemical structure of molecules,
06:59the downstream proteomic signature
07:02of these molecules in the cells.
07:05Sorry, I've got the wrong slide.
07:09In order to understand
07:13the biological signatures
07:21of a complex disease.
07:22And once you have that,
07:24you're able to screen through drug candidates
07:26and chemical compounds
07:27for the ones that are able to move
07:29yourself from a diseased state
07:32back into a healthy state.
07:35You're then able to also
07:37optimize the downstream properties
07:40of that chemical
07:41so that the safety profile
07:43of the drug is optimized.
07:45And that's where you're able to transform
07:47not only the end-to-end ways
07:49of doing drug discovery,
07:51but the scientific strategy itself.
07:55Going back to the previous example,
07:57which is then in the chemistry space,
08:00the previous was more focused on biology.
08:02In the chemistry space here,
08:04we've got a mining company
08:05that was sitting on piles of ore
08:08with lots of very valuable metals
08:11stuck in the rock.
08:12And they needed to discover new reagents
08:15that would help flush the metal
08:18out of these piles of rocks.
08:19And so we use foundation models
08:21in chemistry in order to discover
08:23new molecules and actually reuse
08:26existing molecules in some cases
08:28that had better properties
08:29in terms of the extraction yield
08:31and in terms of the downstream
08:34environmental footprint.
08:36So in this case,
08:38you ended up being able
08:40to create new patentable molecules
08:43and implement a new operational process
08:47for mining extractions.
08:56But this is not just about building AI kits.
09:01It never is.
09:03In order to truly capture the value
09:07and build sustainable momentum,
09:10you have to rewire the R&D organization
09:13from its data substrate,
09:15the landscape of tech systems
09:17that are supporting the organization,
09:20the processes, the talent,
09:22the ways of working.
09:24And what we see often
09:25is companies setting up
09:27a large transformation program
09:28that has these six work streams,
09:30but they're just loosely coupled
09:32with some of the work streams
09:34on the IT side saying
09:35your IT system transformation
09:37will be ready in 18 months,
09:39the data organization saying
09:41we're building an exquisite data lake
09:44that'll be ready in 12 to 15 months,
09:46and then you have a bunch of pilots
09:47on the AI side.
09:49What you need to do
09:49is work on these six streams
09:51but in a much more synchronized manner,
09:53and that's where we often talk
09:55about quarterly value releases.
09:57You establish an execution discipline
10:00that allows you to deliver
10:01concrete business value every quarter.
10:04Every three months,
10:05you can go back to the XCOM,
10:06you can go back to the board
10:07and say this is how we now work using AI.
10:10We're able to demonstrate
10:12a new capability,
10:13but it's not just vaporware.
10:15It's built on an incrementally
10:17more sophisticated
10:18and more modernized foundation
10:20because you've synchronized
10:21the backlogs of activities
10:24of the different teams
10:26that are working on the six aspects
10:27of the transformation.
10:29This builds and sustains the momentum
10:31and allows us to achieve
10:33what is an incredible amount
10:35of transformation work
10:37that needs to be done
10:37in order to capture
10:38the full potential of scientific AI
10:40in this brave new world.
10:43That's it.
10:44Thank you.
10:44Thank you.
10:46Thank you.
10:47Applaudissements
Commentaires