The recent post here on automation in chemistry (especially medicinal chemistry) is a good intro for this paper in ACS Med. Chem. Letters. It’s from David Parry, who led Cyclofluidic, and I’ve blogged about them a few times over the years. That was a company formed in 2008 in the UK to try to develop a “closed loop” of automated medicinal chemistry, where compounds would be synthesized, then tested in a primary assay on the same platforms, with the results fed back into the software to inform the next round of automated synthesis, etc. This has been a longstanding idea in the field, and there are no theoretical barriers to realizing it.
There are, however, plenty of practical ones. Fully automating the synthetic chemistry alone is not exactly trivial, given the variety of reactions and their different conditions. That problem by itself has kept several very capable research teams in both industry and academia busy for many years now, and it’s just the first step. Automating the assays, of course, has been the work of high-throughput screening teams for at least the last twenty-five years, so there’s a lot of experience to build on there – but connecting the output of the automated synthesis to such screening (without having some human operators in the middle of the process for purification and compound handling) presents a whole new set of challenges. And then you have the “evaluate and send back around” step, which calls for software that can suggest reasonable new analogs after evaluating screening data, which is yet another longtime goal that has kept many other folks fully employed for a long time now as well.
So Cyclofluidic had their work cut out for them. And Parry himself even more so:
Starting out at the helm of a new company on day one is both incredibly exciting and daunting in equal measure. The challenges were numerous, while both a business and technical plan were in place as part of the funding process that still left many open questions. I was very comfortable with the scientific and technical aspects to be undertaken having been increasingly involved with the evolution of the proposal but may have taken a slightly different stance on some of the more challenging technical milestones had I known I was going to be responsible for their delivery.
That last part will, I’m sure, bring on some shivers of sympathy among many readers, many of whom may have experienced some form of the “OK, whose problem is this going to be? Ah. I see. Mine” phenomenon. A big early decision was whether to build the machine from the ground up or adapt commercial equipment, naturally, and the company took the reasonable route of using available technology whenever possible, and doing its own development at the points where it was most needed. And it was certainly needed:
The assay platform initially utilized a custom glass chip for the assay with channel dimensions of approximately ca. 80 μm; early observations of laminar flow with colored aqueous dyes rapidly opened my eyes to the challenges of working in this environment. For practical reasons (lead time, cost, and ease of replacement) the glass chip was replaced with 75 μm ID capillary tubing, a simple switch that proved remarkably effective. Pumping at the very low flow rates with the required relatively high pressures determined by the assay chip or capillaries was not so straightforward. . .
. . .Much was learned along the way including the challenges of minimizing the adherence of reagents to glass surfaces, accurate pumping to achieve a gradient of reagents into the assay chip with time, and ensuring solubility of the reagents at all times.
I can well imagine, having done a fair amount of flow chemistry myself (and without the added twist of doing flow biology!) The paper goes into much interesting detail on both of these, and on the software needed to keep the whole system running. But there were even larger factors at work, which hadn’t been anticipated:
When Cyclofluidic was conceived, the business plan was to build a prototype platform and then sell the platform or its component parts to pharma and biotech companies. During the early years of the company, the pharma sector tended to be moving away from large internal technology platforms and not making significant investments in internal capabilities. With input from the shareholders and more widely, it was agreed that the business should look at alternative options for revenues based upon the rapidly evolving capabilities. This came down to a choice of two options, either become a technology integrator providing the expertise to tackle complex automation and integration projects in the life sciences or to provide services based upon the envisaged capabilities of the Cyclofluidic platform. . .
They opted for the latter, and were able to engage in a number of collaborations (only some of which have been published). The market for companies that want to assist in early-stage drug discovery is a large one, though, with pretty ferocious competition. And by this time, Cyclofluidic wasn’t just competing with other small service providers; they were, in many cases, competing with internal efforts at a number of large pharma companies to automate their own medicinal chemistry efforts. In general, these (at AbbVie, Lilly, Roche, and others) were devoted to increasing chemistry productivity or perhaps assay throughput: few (or perhaps none) were trying to close the compound-optimization loop completely with automation.
So that was the Cyclofluidic selling point, and the big question was whether that was enough. The all-in-one aspect was unique, but it also meant (as Parry notes) that at any given time, a good amount of the system was idle as things moved around the loop. The chemistry efforts at other companies were generally higher-throughput for a given fixed transformation, and could be combined in some cases in modular fashion. They weren’t integrated with assay technology for the most part, but how much of an disadvantage was that, practically? From what I can see, it came down to a choice between the Big Pharma model, where the individual parts of the process were highly optimized but not connected together as smoothly, versus the Cyclofluidic platform, which had far better connections but had limitations in each of the component parts. And those limitations were imposed by the connections themselves.
In the end, the company never reached takeoff speed: they had good feedback from their collaborators, but not enough of them seemed to feel that this was technology that was compelling enough to invest in further. Cyclofluidic ended up positioning itself as a service provider, but selling a service that people weren’t sure if they needed or not. As Parry mentions, it would have been very interesting to compare the progress of their platform versus a traditional med-chem team, and ideally to do this over several different projects.
An ideal test would be to hand the same starting set of potential lead compounds to several traditional med-chem teams (not in communication with each other!) and to a Cyclofluidic-style effort as well, and to compare the results. What were the compounds they arrived at after (say) a year of work, and how good were they? You’d want to do this several times (as mentioned, given the natural variability of drug discovery) and honestly I’m not even sure how many trials you’d want to run before drawing firm conclusions. I am quite sure, though, that it would be more than anyone would be willing to fund. The human-team versus human-team aspect would be fascinating enough even without the man-versus-machine aspect; I’m sure that nothing like this has ever been run under controlled conditions. Bored billionaires apply here.