From Code to Clinic: How AI Is (and Isn’t) Rewriting the Life of a Drug

00:21, Chris Thonis

Hey everybody. Welcome back to William Blair Presents. Today is Wednesday, April 29th, 2026. I'm joined by Matt Larew. He's a partner and a research analyst covering life science tools and diagnostics at William Blair. Matt, thanks for joining us.

00:35, Matt Larew

Hey, Chris. Happy to do it.

00:36, Chris

So, I've invited Matt to chat through his recent industry thematic. It's called, From Code to Clinic: How AI Is (and Isn't) Rewriting the Life of a Drug, which automatically concludes that while AI may not replace the physical lab work of drug discovery, it is reshaping it. So, to kick this off, Matt, hoping you can start with a high-level overview of what the report is about. And, you know, why the topic is relevant right now. And then I figure from there we can dig a little deeper.

01:03, Matt

Yeah. Good deal. So, look, Chris, drug discovery is hard, expensive, and time consuming. And just to put some numbers on that, okay, so hard: about 10% are drug candidates that enter human clinical trials ultimately become approved drugs. Expensive: typically, it takes more than $2 billion per approved drug. Time consuming: on average, it's about ten years or more from idea to approval.

So, you know, like many industries right now where AI has been adopted in pursuit of efficiency, there's optimism that AI can help bring more drugs to market on quicker timelines and for less cost. That sounds great. But the life science tools and services industry that I cover, it really exists to feed the scientific research discovery engine. And so, if you walk into, you know, any lab in the world, you'll see instruments with my company's names on the front of them, test kits that have their names and so on and so forth.

And so, as investors have observed over the last few months, you know, the rising realization of the consumer facing LLMs, and improved performance, and announcement of some life science specific capabilities from some of the large frontier AI labs. The stocks in my space were under tremendous pressure, in part on this wet lab obsolescence thesis, essentially, that future scientific insights will be increasingly generated via AI and not empirically in the lab, thus reducing the need for some of the products and instruments my companies offer.

So that really was the focus - to examine how AI is and isn't influencing drug discovery.

02:44, Chris

Got it. You mentioned wet lab work, it’s a core framework in the report, you know, this is the distinction between dry lab and wet lab work. How should listeners think about those concepts and why is that distinction so important?

02:58, Matt

Yeah, I mean think of dry lab as compute, right? So, researchers use data models that generate ideas. You know, what target to go after, what molecule to try. And then think of wet lab as the physical or real-world side of this. So, you know, proving whether those ideas are actually true in physical biology. And, you know, drug discovery operates through a design, build, test, learn cycle.

So, design a hypothesis, build and test it in a lab, and then learn from those results and cycle back through. And AI is and really has been, you know, increasingly leveraged as a dry lab tool in those design and learn stages, especially generating and making hypotheses. But, ultimately, those hypotheses still have to be validated in the wet lab because decision grade and, this is the key here, regulatory grade evidence, it requires physical proof.

And so, when you map out the workflow, and we do this in the report, most of the work that's required to bring a drug candidate into human testing, it's actually in those build and test phases and the validation phases. Not as much the design and learn phases. So, you know, the view is if AI is successful in improving hypothesis generation by creating more candidates faster, hopefully better candidates too, that should create even more demand for that wet lab validation work.

And the final point here, really, the quality of outputs from a biological model, so on the dry lab side, it's directly predicated on the quality and utility of the data its trained on. And the biology today really lacks the requisite inputs, so standardized, fit for purpose, experimental data to generate the desired outputs, so novel drug ideas. So, for AI to reach its potential in biology, you know, our view is there's a significant amount of new experimental data that needs to be generated in the wet lab, which we think will create a durable tailwind for companies we cover sitting at those data bottlenecks.

So, you know, in sum, our view is that the wet lab obsolescence narrative I mentioned above is really overstated and frankly, mis framed.

05:16, Chris

That makes sense. So, the report emphasizes that drug discovery continues to be expensive. It's lengthy. It's uncertain. What specific challenges is AI best positioned to address within that reality?

05:28, Matt

Yeah. So, I mentioned the data points earlier, ten years, 10% success rate, more than $2 billion to get a therapy to market. And importantly, you know, unlike many technologies and technology areas that have benefited from Moore's Law over the years, the cost of developing a new drug has roughly doubled every nine years since the 1950’s, so it's, in effect, termed Eroom’s law in drug development.

So, with that context, we think the biggest near-term opportunity is compressing timelines, and particularly search and triage timelines, or improving the funnel, you know, ranking and filtering candidates earlier so you're wasting less time and lab capacity on low probability ideas. And AI adoption is growing in those capability areas. And we've seen data points out there, at times, compressing the timeline from idea to pre-clinical candidate from five plus years down to 12 to 18 months. So that's really one big area is compressing the timeline.

I point out the biggest long-term opportunity is probably in reducing clinical failure, especially at phase two, though I think the AI toolset is less developed in this area at the moment. Relative to, again, kind of, timeline compression, search, and triage.

06:48, Chris

You note that in biology, you know, data is often the limiting factor for AI progress. What makes biological data uniquely challenging compared with data used in other AI applications?

07:02, Matt

Sure. So, I think, maybe it's worth from a start to point out why biology is an attractive target for AI. Because it's programable. And that might not be obvious at first instinct, but you know, it relies on code in the same way computers do. It's just that the code of DNA and maybe you might remember from, kind of, high school biology, A T C G are the four letters of the code, rather than the ones and zeros that are used in computers. But it's still a code.

And then within DNA, so the code of DNA, A T C G, there are regions called genes that the human body turns into proteins. And proteins are what carry out every functional task in the body. And DNA proteins, and then, sort of, the intermediate molecule which is mRNA, they all have distinct code that can be read, and they have physical and spatial properties which can be observed, which can be imaged.

And so, forming a complete picture of DNA, that intermediate molecule mRNA, and the proteins that are involved in diseases, that complete picture really helps researchers discover and design drugs to target those diseases. And so, in theory, if you can read and model that code, you should be able to make much better guesses about diseases and drugs.

But here's the catch versus maybe typical LLMs. Biological data is not like internet text, right? So, many AI applications you can train on these massive, cheap, already digitized, you know, data sets and the training signal is almost, you know, free. But, in biology, the most important questions that you're trying to answer like does this compound actually change a cell state? Does this target matter? Does the protein, as we sequence it, model it? Does it actually function in the human body in the way we think that it will? Those all require a wet lab experimental outcome. And there's a plethora of open-source biological data that is out there, in the public ecosystem, and, frankly, most of the open-source biological models that exist. Listeners may be familiar with things like AlphaFold, which won a Nobel Prize. Those are then built largely on that data.

But there's a problem with that, which is those data sets were assembled over decades from individual experiments conducted, you know, for a diverse set of biological purposes. They were not conducted, designed as machine learning transits with standardized protocols, materials, controls.

And the reality is, if you don't have consistency, clean labels, and enough biological diversity, the models get biased and noisy. And this has been pointed out by a number of AI companies in the field, that there's been a data wall in the field, where progress, at this point, is increasingly constrained, not by algorithms, but by access to new, well curated, and experimentally grounded data.

And that's really, you know, the key we think, you know, going forward is the generation of new data.

10:22, Chris

This is actually not a question I put in here, but just based on what you just said, because it's super interesting. All this data that goes back decades, I mean, is there a possibility of AI finding a way to take that data and actually turn it into usable data? Or is just the way that it has been….

10:38, Matt

Yeah, I'd say it, you know, to date it has been, you know, very useful. We've learned things, again, I mentioned AlphaFold, there's a variety of other models out there, Bolt and others that have, I mean, the reason they won the Nobel Prize is because it is able to predict the way proteins fold from a sequence alone.

And so, there's protein folding models, there's protein language models. There's been tremendous progress made. And I referenced, you know, pharma companies have, sort of, broadly adopted some of these models at some of those key phases around identifying new targets and designing new drugs, right? So, the premise being, if I have a target, I can use a model to design a thousand potential drugs, potential antibodies, for example, and then test those individually.

And that's something that wasn't possible before a lot of these of these models. But what we found and, of course, I mean, not me, but for the field, is that, you know, they're increasingly constrained by the lack of new, well curated data and data that's focused in specific disease areas or in specific tissue or organ areas, you know, really to find new drugs.

So, they're a great starting point. Many of these models, what we've seen is, sort of, the next leg of growth is companies taking those models and layering on their own proprietary wet lab data in specific focus areas. So, they’re a wonderful starting point. And they've done, you know, a big credit to the field.

But, one of the points in our report is that's the starting point for how AI and how AI models are going to be impactful.

12:16, Chris

Got it. So, there's a perception that increased AI adoption could reduce the need for laboratory work. Based on your research, how does AI actually affect experimental demand?

12:28, Matt

Yeah. So, I think, you know, two things can be true of the same time, right? So, AI can and does reduce waste in certain areas of drug development, right? You know cutting out duplicative, low value experiments by doing more screening and filtering in silico or in the dry lab. And we referenced, in the report, that at target discovery and drug discovery, AI tools are widely adopted by big pharma, in particular, to reduce some of that unnecessary lab work.

But again, the hypotheses generated, whether they come from traditional wet lab experiments or from dry lab, they'll always need to be empirically validated in the real world. And I think this is a good, you know, question to maybe interject a point on where actually cost and drug development occurs, right? Because there's a perception, I guess, that, you know, the desirable outcome of AI is to do less lab work.

And, as I referenced, yeah, it'd be great to do a little bit less duplicative work, but more than half of R&D spend at pharma is people. And a good chunk of the balance of it is overhead and non-lab costs. And we estimate that only around 15% of R&D spend is lab spend. So, with that context it becomes clear, you know, the biggest opportunity for pharma to reduce cost per program is to compress timelines, not necessarily cut lab spend.

So, you know, that's the first point, is the biggest bang for the buck pharma is going to get is by doing, you know, more high quality, entering more high-quality candidates into the clinic per year, you know, leveraging in that huge, fixed cost that they have rather than, you know, nickel and diming a specific program.

But it goes back to the point I mentioned on models, Chris, which is that if you're going to be relying on the dry lab models to take on more idea generation, they have to be high quality models, right?

And they have to have, you know, good data driving them. And so, there's been a new, you know, we think what's going to be the start of a, sort of, an age of industrial biological data generation to help feed those models. So, you actually have more demand for some components of lab work to build out these billion cell projects and virtual cell atlases that have been announced in recent months.

14:54, Chris

So, the report walks through drug discovery step by step, you mentioned that a little bit, using antibody development as an example. Where is AI having its greatest impact today? And, where do limitations clearly remain, would you say?

15:06, Matt

Yeah. So, it's, you know, in the report we use, as reference, antibody development as an example. And that's, in part, because at this point in biologics, they’re more than half of all drug sales, they’re biggest part of the pipelines, and monoclonal antibodies are the biggest category within biologics. So, we think it's a good example to use.

And again, the biggest impact for AI is these, kind of, design and learn phases. So, the first being target identification, again finding a target in human biology that if you modulate, you mean, upregulate, downregulate, turn off, turn on has in a downstream effect that is desirable. You know, stopping a cell from replicating, killing a virus, whatever it might be.

So, that's the first piece where AI is impactful. And that dates back, frankly, Chris, to even, sort of, machine learning replacing, you know, manual review of new literature. Where maybe, 20 years ago you might have an army of PhDs combing through the literature just to find new targets, you know, applying ML to that. Imagine, now, expanding on that to find target, so that that's the first area, and, again, widely adopted today within pharma.

And then the second piece where it's having a big impact is on the drug discovery or drug design phase, right? You know, this is where you found a target. And you know that if you affect that target, it will have a desirable, downstream, effect. Now, what's the drug that will affect that target in a way I want, again, upregulate, downregulate, turn off, turn on, and only do that.

So, very specifically, you know, hit my target. And so, it's also having a big impact there. And again, I reference things like AlphaFold and some of these protein folding models. So, using AI to spit out, you know, hit lists, basically, of potential drug candidates.

Now, where AI hits a wall is in some of the downstream pieces, both downstream of target ID, meaning target validation, and downstream of drug ID or design. So, validating that the drug works.

So, this is, you know, things like binding, which is antibodies, kind of, a core thing they do is they bind to their target. The functional biology of what actually happens in the human body after that binding occurs. Developability, this is, kind of, an underrated thing, but when I actually turn this drug candidate into a drug that needs to be manufactured, stored, ultimately, delivered to a patient, you know, what occurs and how physically do I need to affect or edit this drug so that it can do all of those things?

And then, ultimately, the final piece before you're in clinical trials, it's called CMC and IND enabling studies. Again, these are inherently physical tasks where you're essentially doing test runs of how you're going to manufacture the drug, scale the drug, store the drug, etc.. And, as we referenced at the outset, the majority of the wet labs spend occurs in the categories I'm characterizing as really physical tasks, where AI, we believe, will have a limited impact.

So, you know, huge opportunity for AI to, kind of, compress the timelines via, you know, via the, sort of, search and triage and filtering by quality at the target and drug ID phases, but we think less of an impact, less of a potential impact in some of these inherently physical steps of drug discovery.

18:40, Chris

So, you make an important distinction throughout the conversation between hypothesis creation and then hypothesis validation. Why is this distinction so central in understanding AI's role in drug discovery?

18:56, Matt

And I think it's both important when assessing the risk to tools, right? You know, viewed writ large, at, sort of, a higher level. I think we would just say AI, you know, can have a big impact on compressing timelines and drug discovery. When you take a more nuanced view as how that's going to impact the wet lab and the actual tool spend, that's where it's important.

Because if you actually break down, step by step, what an antibody discovery program looks like, the majority of the spend, and really where public companies, public tools companies’ revenue exposure sits, is in those hypothesis validation phases. And you're right, the distinction we make in the report is that hypothesis creation, I mean, it's inherently a learning application, right?

And so the, you know, moving that from traditional wet lab techniques to the dry lab, you know, it doesn't necessarily impact the companies that we're covering because most of the spend for them occurs in more the validation step. So, hypothesis creation, yes, it's transitioning from the wet lab to the dry lab.

Hypothesis validation, we believe will always be a physical task that requires, you know, wet lab spend. And, again, the key insight here, I think, and maybe a good way to distinguish this world versus others is you ultimately need regulatory grade evidence. Because the ultimate outcome of all of this work is creating a drug that, you know, thousands, or millions, or billions of people are going to be putting into their body, and you need to know exactly what's going to happen after that occurs.

20:37, Chris

All right. So, looking ahead, what signal should listeners watch for as AI adoption moves from experimentation towards more sustained real-world integration?

20:48, Matt

Yeah. So, you know, I should say from the outset, I think if AI is successful and impacting drug discovery, there's no question that a quantifiable sign of that, and probably the first sign that we see, is more drugs entering clinical trials in the coming years.

So, I think, sort of, full stop, that is an objective thing that we should see if AI is impactful. Now, that will be the result of some other signal, which will be companies characterizing timeline compression.

And so, there's a number of public and private companies that are, sort of, AI native drug discovery companies who have described timeline compression to get drugs into clinic or to create pre-clinical candidates, again, reducing those timelines by 50%, 60%, 70% or more. And we would expect those metrics or those narratives to begin to be shared broadly by Big Pharma companies and established biotech companies, as well.

You know, the second piece is, I referenced, in recent months in particular, there have been announcements of these really ambitious, kind of, industrial scale data generation projects, spatial atlases, virtual cell projects, and the like.

And so, I think another tangible, example will be that more and more of those are being created, being announced, being funded, because that will connect to either the actual or presumed success of the models, you know, these new models being built, in, you know, delivering new biological insights, you know, the target identification or drug identification.

You know, another will be, I would say more automated lab setups. So, we've seen announcements recently from a large, one of the biggest AI players in the world and two of the largest pharma companies in the world, who are creating automated lab in a loop.

They're investing billions of dollars in creating these hybrid dry lab, wet lab setups, where experiments will be run 24/7 and fed into onsite GPUs to generate new insights and design new experiments. Versions of that are being created at other pharma companies. And there is also, have been a number of partnerships announced between software or other nontraditional, kind of, pharmaceutical ecosystem players and traditional researchers and tools companies, in pursuit of, kind of, a distributed lab in the loop model. So, I think that'll be another tangible piece of evidence.

One thing we don't expect to see, and this is a key point, because it faces often the pushback we get, we do not expect to see a reduction of R&D budgets by pharma, either on a gross dollar or percentage of revenue basis. And we point out in the report that if you track back, you know, two decades, before sequencing was really democratized, before the introduction of interesting technologies like cryo-EM, before the introduction of AI into Big Pharma or AlphaFold.

So, prior to the, you know, litany of technology advancements that have reduced the cost per biological insight, pharma R&D spend has grown not just on a gross dollar basis, but has grown dramatically as a percentage of sales for pharma. And what that should tell you, we think, is that if, indeed, AI is successful in reducing the cost per insight for pharma, they're going to do more drug discovery, not less.

That's certainly what history tells us. 24:35, Chris All right, so before we wrap, what is one misconception about AI in drug discovery that this report is intended to clarify?

24:44, Matt

You know, I think the biggest thing we wanted to get across is that AI and drug discovery is not a headwind for life science tools. And, I would say, based on our conversations with investors over the first couple of months of the year, there's no question that was the consensus. But, it was the consensus without conviction, frankly.

And that was the goal of this was report, was to try to dig into that and provide some conviction. And we were open to it being a headwind if that was where the evidence led us. But, you know, after talking to, you know, 20 companies in the space spanning from Big Pharma, to some of these AI labs, to antibody discovery CROs, to public companies, to private companies, we just did not hear that message at all.

Again, the message was much more: AI is going to enable us to do more work faster. So, I think that's the biggest misconception.

Again, then, the second piece I would mention is, if AI is going to be successful, it is not going to be based on the data sets and models that exist today. And, I think, again, that's another misconception is that we, sort of, have what we need, and it can jump off from here, sort of, push button, get drug.

We'll probably never get to that level of simplicity, but if we approach it, it will be because of significant investment in new data generation to actually transform these foundation models into truly, you know, new drug discovery engines.

26:09, Chris

All right. So, that's all the time we have today. But, Matt, thanks again for joining us. For those interested in reading the report, it's called From Code to Clinic How AI Is (and Isn’t) Rewriting the Life of a Drug. You can request a copy by visiting Williamblair.com/contact-us. Thanks for listening. We'll see you next time on William Blair Thinking Presents.