Someone recently sent me The Virtues of Virtual — And Why We’re Devirtualizing and wanted to get my take on it, given that I’m clearly in the virtualization-is-the-future camp. The upshot is that Padlock Therapeutics, until recently an entirely virtual organization built on CROs, is now bringing some of the work in-house on the argument that CROs just shuffle money and people around without actually changing much in the fundamentals of drug discovery while introducing frictions of their own.
Specifically, there’s a debate about the extent to which medicinal chemistry has become a commodity and whether or not that’s a good thing for the industry.
My perspective on this starts from the observation that there is a difference between an activity (chemistry, machining, baseball) and the guidance of that activity. In general, different people design the part in CAD and make it on a CNC mill; these are different skill sets. The two may be cross-trained to some extent, but they’re really separate roles. Decoupling these things is an important step in the path of increasing sophistication on the design side, supported by reliable abstractions.
For various reasons, the life sciences still couple action and guidance to a much greater degree than many other fields. This is reflected in Michael Gilman’s (Padlock’s CEO) concern that a decline in hands-on bench chemistry will lead to a decline in effective medicinal chemistry leadership in the long run:
I would add that, unfortunately, medicinal chemistry is increasingly regarded as a commodity in the life sciences field. And, worse, it’s subject to substantial price competition from CROs in Asia. That — and the ongoing hemorrhaging of jobs from large pharma companies — is making jobs for bench-level chemists more scarce. I worry, though, because it’s the bench-level chemists who grow up and gather the experience to become effective managers of out-sourced chemistry, and I’m concerned that we may be losing that next generation of great drug discovery chemists.
There’s a bit of confusion here on what the phrase medicinal chemistry covers. The process of organic chemistry appears to be a commodity, which for the reason mentioned above and others, is a good thing. This commodification of chemistry is true to the extent that there are many places to hire bench chemists who will produce a certain number of molecules per month — “if you employ ten chemists for six months, you know roughly how many compounds they’re going to get,” as Gilman writes. But medicinal chemistry is also not commoditized in the sense that it extends far beyond rote benchwork.
(Relatedly, synthetic organic chemistry has also been proven to be end-end automatable. Such automation would be a differentiating advantage over “commodity” armies of human chemists due to lower cost, higher yield, and greater speed, but isn’t widely used yet.)
At any rate, getting a certain number of molecules per week isn’t quite the same thing as progress in medicinal chemistry. Guidance of the process — using organic chemistry to optimize a molecule to be a drug — is not yet a commodity. Because the two are so conflated today Gilman worries that a decline in one threatens the other, but that’s kind of a strange worry, in the same way that better compilers don’t put programmers out of jobs. It’s not at all obvious to me why one needs extensive hands-on bench experience to become an “effective manager of out-sourced chemistry,” as that’s a fairly separate challenge. Yes, you need detailed chemistry knowledge to do it, but in a world with abundant cheap synthesis technology the molecular engineering becomes a data problem. As Gilman recognizes:
[M]edicinal chemistry has become nearly a commodity, so we are not hiring bench chemists. We have, however, hired a second seasoned drug-hunter to help with the management of all of our outsourced chemistry.
So there are two halves: doing the synthesis and figuring out what to synthesize. The story of technological progress is in large part one of converting manual labor into knowledge work, and as tools have improved over the last several decades, the daily practice of biology has risen in level of abstraction, leading to a virtuous cycle of more powerful tools leading to more valuable results creating more investment in even better tools.
One similar problem that has become highly decoupled is DNA synthesis, now a widespread commodity. Like synthetic chemistry, DNA synthesis benefits from a very clear interface. After signing a deal demonstrating that a 50% decline in cost of synthesis would correspond to a 4x increase in demand, Barry Canton of Ginkgo Bioworks wrote “it seems plausible that all in vitro DNA synthesis, assembly, and editing will happen at specialized firms and the role of the customer will be focused on the equally interesting challenge of introducing that DNA into their organisms.” It’s easy to imagine writing something similar about a well-characterized synthetic chemistry service, and indeed that’s how Gilman uses Evotec for Padlock.
Other than capital efficiency, a big reason to aggregate workflows across customers is that it allows you to learn from the data and get better over time. Similar to Canton’s observation for DNA synthesis, it makes a lot of sense to have one central group that has all the knowledge for how to synthesize — or optimize — a compound rather than mixed levels of experience widely dispersed. Ultimately the customers don’t care about number of molecules generated, cells screened, or assays run, but hit rates and quality scores.
Now, interestingly, the management role that Gilman talks about keeping in house is endangered by centralization of synthesis to a greater extent than I think many people realize. Gilman expresses a belief that while benchwork may be automated away, the design work remain high-value knowledge work for the indefinite future, to which I give him [1] [2] [3] [4] [5]. This is driven by several trends, but an important one is the centralization of data, something which is substantially accelerated by CROs. Just as every DNA sequence synthesized by Twist feeds into a machine learning model to improve the synthesis process, we can expect that every molecule synthesized in the future will help make the system better.
Of the challenges with CROs Gilman lists, it feels like they largely reduce to “the interface wasn’t clear enough”:
But although sometimes you want them for their minds, other times you just want their bodies. You want them to shut up, swallow their objections to your stupid ideas, and just do the damn experiment that you’re paying them to do.
and:
It’s not unusual to discuss a proposed experiment only to get on the phone two weeks later to find out it wasn’t done or was done differently than we’d planned — not because of malfeasance but rather because of simple miscommunication or misunderstanding.
and, especially:
Scientists at a CRO will never feel the sense of urgency and the sheer existential angst that focuses our minds in a startup.
These quotes highlight the difference between outsourcing and cloud. Cloud is about developing clean abstractions that allow execution and design to be decoupled, opening a powerful toolbox for higher level thinking. Outsourcing is about HR arbitrage and turning capex into opex. Outsourcing is a subset of cloud, but far from the whole story.
At the end of his article, Gilman refers to chemistry as a reasonable business for CROs but comments that biology may be somehow fundamentally different. Biology, just like chemistry, has design and execution components that can be decoupled. Biology has been more resistant to this so far due to its intrinsic complexity and sensitivity, and our lack of comparatively powerful tools. Innovations like acoustic liquid handling, optofluidics, advanced robotics and statistical design of experiment methods have begun to give us greater control over execution in search of higher reproducibility.
Design for biological assays and systems is very complex, but it’s a data problem, whether solved by humans as it is today or machines possibly in the future. The issue isn’t the CROs, it’s the interface, and technology behind the scenes that makes such an interface possible. (You can only allow the customer to specify things as precisely as you can actually do them.)
I suspect it says something fundamental about the structure of the problem that CROs fulfilled by humans have been so resistant to clear abstractions. It’s not like CROs are a new idea, or that they have mediocre people working for them. They aren’t, and some clearly have very smart teams. But businesses are responsive to what their customers ask for and humans are so flexible that it undermines the integrity of the product in a very Christensen way. We’ve certainly seen elements of this at Transcriptic, too, with our human-supported custom implementation services. Robots require a clarity of specification that force the matter. I really hope that the Padlocks of a few years from now won’t think of themselves as “virtual” biotechs, but as legitimate information technology companies. There’s a whole frontier there opened up by better interfaces and powered by data that look similar but stand on very different foundations.