In the following sections we describe how specific χDL protocols were communicated between our laboratories in Canada and Scotland. As part of the communication process all the protocols were validated after receiving them. By that we mean matching the characterisation data that is relevant to the synthetic task. This can be typical characterisation data for organic compounds (NMR/HPLC/Mass Spec.), the compound purity (measured by quantitative NMR), or the isolated yield within an error margin of 10%.
Transferring Protocols Between Platforms Of The Same Type
H2IMes•HBF4
To demonstrate that communication of χDL protocols enables reproducible chemical procedures regardless of location or hardware configuration, we chose a three-step synthesis of the carbene precursor 1,3-Bis(2,4,6-trimethylphenyl) imidazolinium tetrafluoroborate, 4 [28–30], see Scheme 1. As the chemistry is robust and well-studied, it provided a useful starting point to investigate the efficacy of our χDL communication procedure.
The original protocol was designed on the Chemputer in Canada where three executable χDL scripts were responsible for the three major synthetic transformations: bis-imination, reduction, and ring formation with salt exchange. Initially, these scripts contained manual interventions due to hardware limitations. These included moving connections, transferring solid materials, and crystallization of the final product in ethanol. The scripts with interventions provided compound 3 in 45% yield, which was then transformed to compound 4 with 77% yield. Manual isolation proved inconsistent, and the purity of the final product 4 varied between repeated attempts of the final step. These χDL scripts were then transferred to Scotland for validation on their Chemputer platform and subsequent optimisation.
The three scripts were consolidated into a single χDL, and the final crystallisation was automated thus achieving an entirely automated synthetic sequence with no human intervention. This optimised χDL returned compound 3 in 88% yield and compound 4 in 47% yield. Optimising a process between labs using χDL is much easier than via “traditional” approaches of communicating synthetic operations in prose or oral communication. Misinterpreting a procedure from another lab may result in the same synthetic steps being performed differently. These discrepancies may initially be insignificant, but over a large project can compound into major differences in results between labs. Recreating the synthesis via χDL entails confidence that along each step of the synthesis, all collaborating groups will be starting from and ending with identical material. With complete assurance that both groups were always starting the next step from the same position, optimal conditions for the following transformation can be found quickly, and with excellent results.
Btida, Esterification, Suzuki
Having established the reproducible transfer of χDL scripts between laboratories, we then developed the first fully digital cross-laboratory optimised, convergent synthesis utilising the Suzuki-Miyaura coupling, a reliable method for carbon-carbon bond formation in organic synthesis. We exploited TIDA-protected boronate esters, recently reported by the Burke lab [31], due to their greater stability towards hydrolysis over the MIDA-protected analogues. This permitted the presence of water during the coupling reaction which increased the solubility of the reagents and facilitated the development of an automated process, see Scheme 2.
The named reaction of 5 with 7 (Scheme 2, top left) was initially developed in Canada (59% yield) then sent to Scotland. Optimisation of the reaction duration was achieved by utilising a feedback-controlled system [32]. The digital procedure for this boronic acid esterification was divided into three subsections according to the best-practice guidelines for χDL [21]: preparation, reaction, and work-up/isolation. The feedback program begins by running the preparation χDL, loading the appropriate starting materials and solvents into the reactor. As the reaction begins, automated sampling every 10 minutes via on-line-HPLC tracks the boronic acid conversion. Meanwhile, the Dean-Stark trap that is connected to the refluxing reaction mixture is automatically emptied in 60 min intervals. These operations are repeated until the feedback-control system decides to stop the reaction and isolate the product. The algorithm makes this decision by comparing the change in the amount of remaining boronic acid in the sample against a user-dependent threshold value. For a more detailed explanation of the decision-making program, see supporting information. This decision-making program was able to determine sensible time points to stop the reaction for three boronic acids (one of them exemplarily shown in Figure SI 3).
From this data set we extracted the distillation time for 4-bromophenyl boronic acid and created one hard-coded χDL script containing setup, distillation of an appropriate duration, and work-up. This script was sent to Scotland for validation, where the target boronic ester was synthesized on the first attempt (59% yield) without any modification of the reaction parameters. The combination of χDL and graph file explicitly details not only the operation, but the exact method of its execution. This clarity allowed the Scotland team to spot a suboptimal detail in the drying of the organic phase. The original procedure used an additional round bottom flask containing the drying agent in combination with an in-line filter. This drying flask was replaced with an in-line drying cartridge, reducing the amount of lost material during the work up. Such a detail would have normally been hidden by the interpretation of a common phrase like “the organic layer was dried using sodium sulphate” in the experimental protocol. Upon this optimisation of the work-up, the yield of compound 8 increased to 71%, which was then counter-validated in Canada (72% yield). Comparing the spectral data in NMR 1 indicates that compound 8 was produced in comparable purity. An elaborated purity analysis can be found the supplementary information.
Whilst the synthesis of the first building block was being developed in Canada, the synthesis of the second building block, (4-(methoxycarbonyl)phenyl)boronic acid, 9, was developed in Scotland (78% yield) and counter-validated in Canada (72% yield).
The Scotland team developed the χDL script for the Suzuki-Miyaura-reaction in three sections: The first consisted of the automated coupling procedure under an inert gas atmosphere. The second handled the workup of the crude mixture using the separator and rotary evaporator modules, the third and final section performed a catch & release purification protocol as described by the Burke lab [31]. All sections were performed using one Chemputer capable of performing each sub-task (Fig. 3).
The automated coupling procedure begins with ensuring the inert atmosphere in all relevant modules. This is completed by evacuating the reactor module and refilling it with Argon as well as purging the solvent and reagent vessels with Argon. All this is automatically completed by the pneumatic manifold (shown in green in Fig. 3) that has a connection to an Argon line and to the fume hood vacuum as well as various outlets to supply all the different flasks. After this, the Chemputer backbone gets equally flushed with Argon to ensure that the lines are clean before adding the reagents to the reactor flask and starting the reaction by heating to 60°C for 12 hours. After cooling down to room temperature the crude reaction mixture is transferred through a Celite cartridge to remove Palladium impurities and ends up in the separation module where it is washed with brine. To ensure quantitative transfers, additional ethyl acetate and water is added to the reactor flask and transferred to the separation module after 15 min of stirring. The separation itself is achieved as follows. Aliquots of 1 mL are taken from the bottom port of the separator funnel and their electric conductivity is determined with a flow sensor that is connected in-line. Depending on the conductivity reading the software can decide if the current phase is aqueous (high conductivity) or organic (low conductivity) and sends them to the respective destinations. The organic layer gets transferred to the rotary evaporator through a sodium sulphate cartridge to remove residual water first and afterwards the organic solvent in the evaporator.
After this workup procedure, the purification protocol starts by redissolving the crude product in the evaporator flask and moving it to a collector flask intermediately. The catch & release purification works by firstly capturing the product via precipitation on a Silica/Celite cartridge that with flushed with hexane beforehand. This “catch” process is completed by repeatedly flushing the cartridge with hexane and precipitating small quantities of product into it until the crude product mixture is used up completely. Subsequently, the cartridge is flush with diethyl ether and methanol in diethyl ether (1.5 volume per cent) various times to remove impurities. In the end, the product is “released” from the cartridge by flushing it with tetrahydrofuran. After removing the tetrahydrofuran in the rotary evaporator, the desired product 7 was obtained in 40% yield.
Transferring Protocols Between Different Robotic Platforms
Thus far, the execution of chemical synthesis protocols has been focused on Chemputer-type platforms, which are primarily used for prototyping hardware, χDL steps, and chemistry. [27] Next, we demonstrate the platform independence of the χDL standard by executing the same code on platforms of different types. Four robots were used in this demonstration – a Kinova multi-axis cobotic robot arm (Fig. 4a), an Opentrons pipetting robot (Fig. 4b), and two Chemputer platforms (Fig. 4c/d) – one for prototyping and experience and the other one for final validation. Each platform has strengths suited for different synthetic tasks. The Opentrons robot operates on small volumes (down to 1 µL) with a capacity of 120 reaction vials on a two-dimensional grid. The vials can be heated up to 99°C and agitated by shaking the vial racks. Depending on the transfer volume, the χDL software on the Opentrons robot decides which of the attached pipette arms is best suited for completing the desired transfer in the quickest way. In case of volatile solvents, the software automatically adds prewetting cycles of the pipette tip to ensure that the transfer is accurate. If a library of reagent combinations is to be realised on this robot, the end user does not have to explicitly hardcode the addition of every reagent individually because the χDL software can automatically infer all possible reagent combinations if a list of reagents is given. Moreover, χDL allows the definition of blueprints as mentioned in the next section. Blueprints encode a generic reaction procedure and can be called various times within a χDL protocol with different reaction conditions (e.g., with a new temperature or substrate concentration every time). These features allow to frame chemical procedures for high throughput screenings very concisely and with only a minimal amount of high-level code. All the tedious low-level unit operations that are necessary to complete the procedure will be automatically inferred by χDL and communicated to the hardware API.
The Kinova robot arm also operates on small volumes, but it has the added benefit of unrestricted movement in three-dimensional space. The high versatility of the system, which is primarily run on a python interface, allows it to adapt to a wide variety of hardware modules and made it easy to integrate it into the χDL software standard.
The Chemputer, conversely, is fixed in position but operates as a universal synthesiser on batch scale, using larger volumes to transfer reagents and products along a liquid-handling backbone. As outlined in previous sections, the Chemputer is characterised by its high adaptability stemming from the modular architecture that allows to connect any module that is able to execute χDL commands to the liquid-handling backbone. It was developed as a robotic equivalent to a human chemist operating on gram scale and covers the majority of common bench chemistry unit operations like liquid reagent addition, filtration, separation, evaporation, heating, cooling, or drying. Those, in turn, can be combined to accomplish typical laboratory tasks such as recrystallisation or extraction of aqueous mixtures with subsequent drying over a drying agent and removal of the extraction solvent under reduced pressure. As being the “gold-standard” for running χDL protocols, the Chemputer platform is ideal to cross-validate any χDL protocol that was developed on a different platform or even hardware independently. When combined with a robot that is capable of running high throughput screenings as the ones mentioned before, the Chemputer can facilitate a workflow where successful reaction conditions arising from the small-scale high throughput screening are scaled up seamlessly by just running the same χDL blueprint with adjusted scaling factors on the Chemputer platform. For further specifications and abilities of the Chemputer platform we refer the reader to related publications. [25]
CDI
To demonstrate the χDL execution on these platforms, we chose to perform a CDI (Carbonyl diimidazole) assisted amide coupling of 2-methoxybenzoic acid with different amines (Scheme 3).
Optimisation of the amide coupling was first developed with the Kinova robotic arm in Canada, however, the procedure was hardcoded in Python. [32] Unfortunately, this approach requires the operator to have a strong understanding of the Python language before making any reaction modifications. We were able to translate the programmed workflow of the platform and convert the defined actions into χDL executable steps. The system operator no longer requires an understanding of Python and can instead use χDL to design any synthesis. Using the χDL driven Kinova platform and following the reaction progress via HPLC analysis allowed the synthesis of the target amide 13 with 83% conversion but without final purification. Having demonstrated that a χDL protocol was able to replicate a previously optimised synthetic procedure, we then scaled up the reaction and ran the same χDL on the Chemputer in Canada with comparable results. Endpoint HPLC analysis shows complete conversion of starting acid to product amide on the Chemputer, outcompeting the Kinova robot’s 83% conversion. The χDL protocol was then validated on the Chemputer in Scotland producing amide 13 in 93% yield. Next, the team in Scotland framed the χDL in a “Blueprint” template. This χDL feature leaves the chemical information in the code unchanged but wraps it into a coding construct that can be called numerous times as a unit following a functional programming paradigm. The Blueprint feature is usually used to indicate that a χDL procedure has been validated adequately and is considered to work for different substrates or reaction conditions. This framing of the code was crucial to allow for meta- χDL-features like iterating over different values of an input variable.
By this means, the Opentrons robot was programmed to iterate over five different amines, including the already validated 2-(2-chlorophenyl)ethan-1-amine, applying the χDL Blueprint for the CDI assisted amide coupling to each of the reagent combinations. All the amines were successfully converted into the corresponding amides using a common χDL Blueprint (see SI). The χDL Blueprint was then used with upscaled parameters for Benzylamine on a Chemputer platform in Scotland again to counter-validate that the χDL protocol works equally fine on different platforms. The desired amide was obtained in 74% yield on the Chemputer.
