Article pubs.acs.org/accounts
Feedback in Flow for Accelerated Reaction Development Brandon J. Reizman† and Klavs F. Jensen* Department of Chemical Engineering, Novartis Center for Continuous Manufacturing, Massachusetts Institute of Technology, Room 66-542A, 77 Massachusetts Avenue, Cambridge, Massachusetts 02139, United States CONSPECTUS: The pharmaceutical industry is investing in continuous flow and highthroughput experimentation as tools for rapid process development accelerated scale-up. Coupled with automation, these technologies offer the potential for comprehensive reaction characterization and optimization, but with the cost of conducting exhaustive multifactor screens. Automated feedback in flow offers researchers an alternative strategy for efficient characterization of reactions based on the use of continuous technology to control chemical reaction conditions and optimize in lieu of screening. Optimization with feedback allows experiments to be conducted where the most information can be gained from the chemistry, enabling product yields to be maximized and kinetic models to be generated while the total number of experiments is minimized. This Account opens by reviewing select examples of feedback optimization in flow and applications to chemical research. Systems in the literature are classified into (i) deterministic “black box” optimization systems that do not model the reaction system and are therefore limited in the utility of results for scale-up, (ii) deterministic model-based optimization systems from which reaction kinetics and/or mechanisms can be automatically evaluated, and (iii) stochastic systems. Though diverse in application, flow feedback systems have predominantly focused upon the optimization of continuous variables, i.e., variables such as time, temperature, and concentration that can be ramped from one experiment to the next. Unfortunately, this implies that the screening of discrete variables such as catalyst, ligand, or solvent generally does not factor into automated flow optimization, resulting in incomplete process knowledge. Herein, we present a system and strategy developed for optimizing discrete and continuous variables of a chemical reaction simultaneously. The approach couples automated feedback with high-throughput reaction screening in droplet flow microfluidics. This Account details the system configuration for on-demand creation of sub-20 μL droplets with interchangeable reagents and catalysts. These droplets are reacted in a fully automated microfluidic system and analyzed online by LC/MS. Feeding back from the online analytical results, a design of experiments (DoE)-based adaptive response surface algorithm is employed that deductively removes candidate reagents from the optimization as optimal reaction conditions are refined, leading to rapid convergence. Using the automated optimization platform, case studies are presented for solvent selection in a competitive alkylation chemistry and for catalyst-ligand selection in heteroaromatic Suzuki−Miyaura cross-coupling chemistries. For the monoalkylation of trans1,2-diaminocyclohexane, polar aprotic solvents at moderate temperatures are shown to be favorable, with optimality accurately identified with dimethyl sulfoxide as the solvent in 67 experiments. For Suzuki−Miyaura cross-couplings, the optimality of precatalysts and continuous variable conditions are observed to change in accordance with the coupling reagents, providing insights into catalyst behavior in the context of the reaction mechanism. Future opportunities in automated reaction development include the incorporation of chemoinformatics for faster analysis and machine-learning algorithms to guide and optimize the synthesis. Adoption of this technology stands to reduce graduate student and postdoc time on routine tasks in the laboratory, while feeding back knowledge used to guide new research directions. Moreover, the application of this technology in industry promises to lessen the cost and time associated with advancing pharmaceutical molecules through development and scale-up.
■
INTRODUCTION
technologies that offer rapidly scalable results and accelerated reaction development beyond the throughput of a traditional batch workflow. A breadth of technologies and methodologies fall under the heading “accelerated reaction development.” Much of our focus has been in the development of microreaction technology,7,8 namely, the use of submillimeter dimension reactors to achieve
With the cost to discover and develop a drug now estimated to exceed $2 billion,1 the pharmaceutical industry is in search of innovative and cost-effective ways to reduce process footprint, minimize lead times, and accelerate scale-up. One path to achieving these goals is in the adoption of continuous flow technology,2,3 which lessens costs and risks associated with manufacturing by promoting safer,4 greener,5 and more aggressive6 flow synthetic routes. At the laboratory scale, the push for continuous has driven a similar demand for © XXXX American Chemical Society
Received: May 29, 2016
A
DOI: 10.1021/acs.accounts.6b00261 Acc. Chem. Res. XXXX, XXX, XXX−XXX
Article
Accounts of Chemical Research highly controlled chemical syntheses that can then be scaled to larger flow systems. This is made possible by the excellent rates of heat and mass transfer in microscale systems, which enable easier access to the intrinsic reaction kinetics and the opportunity for reaction characterization at conditions that would be hazardous to achieve in batch.9 Data collection from continuous flow systems is accelerated by the incorporation of online analytics,10 providing enhanced mechanistic understanding through real-time monitoring of intermediates and byproducts in response to process perturbations. In addition to easier scalability, the heading “accelerated reaction development” also applies to efforts to conduct more experiments in less time with automation. Automated highthroughput experimentation (HTE) tools imported from drug discovery allow for fast understanding of highly multivariate reaction systems, with full reaction maps assembled in time spans of days down to a few hours.11,12 Unfortunately, these techniques have disadvantages in chemical process optimization, ranging from the limited scalability of batch results to equipment limitations that restrict modulation of key factors such as reaction time or temperature. To alleviate these concerns, efforts have been made to employ rapid automated experimentation in flow, merging the scalability of microreaction technology with the speed of HTE.13 However, the explosion in experiment count with increasing numbers of variables (“the curse of dimensionality”) makes direct assimilation of HTE into complex chemical systems both expensive and tedious. This leads to either lost time, from running many experiments at conditions unlikely to be informative in guiding development, or lost information when adopting a simplification such as one-factor-at-a-time experimentation to limit complexity. The use of automated feedback in flow offers researchers an alternative strategy for efficient characterization of reactions.10,14,15 With feedback all experimental conditions are not screened upfront; rather one or multiple initial experiments are chosen as an initialization, then an optimization routine selects the next best experiment to conduct that moves the system toward finding an optimum (Figure 1). New experiments can be selected stochastically with an evolutionary algorithm or deterministically with the application of a model, which can be as simple as a piecewise response surface (a simplex) or as complex as to fully describe the kinetics within the system. With continuous flow and online analytics, smart automated systems have been constructed that use experimental data collected in real-time as feedback for rapid reaction development. Illustrated in Figure 2, these systems integrate automated control of reagent delivery, flow reaction conditions, and online sampling. Analytical results are returned to a computer executing an optimization algorithm, which determines the next set of automated experiments. Herein we review select examples of feedback optimization in flow before detailing our experiences in simultaneous continuous-discrete variable optimization, with an end goal of comprehensive optimization of reaction systems with the single push of a button.
■
Figure 1. Generalized feedback loop for automated reaction optimization.
Figure 2. Workflow for automated feedback optimization.
variables that lead to improvement in an objective. Such a model-free strategy is commonly referred to as “black box” optimization. A notable disadvantage of the black box approach is that with no modeling structure supplied to the optimization, little to no fundamental information can be gained from the final output; hence optimal results have no guarantee to transfer across scales. However, if the feedback system is engineered to determine intrinsic reaction rates, such as under the conditions of fast heat and mass transfer and low dispersion found in microscale flow systems, the results of black box optimization can still have great utility. This is especially true when little is known about the chemistry in advance of experimentation. Some of the earliest methods for incorporating black box optimization algorithms into automated microreactor experiments were developed by Krishnadasan et al.16 and by Jensen and co-workers.17,18 Krishnadasan et al. employed a Stable Noisy Optimization by Branch and Fit (SNOBFIT19) algorithm to optimize the automated synthesis of CdSe quantum dots. McMullen and Jensen compared the SNOBFIT algorithm with two local-search black box optimization algorithms, Nelder−Mead Simplex and steepest descent, in studying the Knoevenagel condensation reaction of panisaldehyde and malononitrile. The gradient-based steepest descent method gave the fastest convergence to the optimum.
FEEDBACK OPTIMIZATION IN FLOW
Reaction Optimization “from Scratch”
With no prior information or models, a plausible optimization strategy must accept input factors known to influence the desired response and interpret relationships among these B
DOI: 10.1021/acs.accounts.6b00261 Acc. Chem. Res. XXXX, XXX, XXX−XXX
Article
Accounts of Chemical Research Though only a local search strategy, the Simplex method20 has been used extensively in black box flow optimization
the H-cube, and demonstrated Simplex optimization of Scheme 3 with respect to hydrogen pressure, temperature, and flow rate.
Scheme 1. Heck Reaction Optimization Studied by McMullen et al.18
Scheme 3. Hydrogenation Optimization Studied by Fabry et al.23
Sans et al.24 used inline nuclear magnetic resonance (NMR) to optimize imine formation in the reaction of 4-fluorobenzaldehyde and aniline (Scheme 4). To minimize analysis time, Scheme 4. Optimization of Imine Formation by Inline NMR from Sans et al.24
Holmes et al.25 developed an online mass spectrometry (MS) optimization strategy for the synthesis of N′-methylnicotinamide (Scheme 5) and compared the optimum estimated with a Scheme 5. Optimization of N′-Methylnicotinamide Synthesis by Online MS from Holmes et al.25
Figure 3. Automated feedback loop used by McMullen et al. for the optimization in Scheme 1. Reproduced from ref 18. Copyright 2010 John Wiley and Sons.
systems. This strategy is both simple to implement and does not require approximation of a gradient, which caters well to expensive experimentation. In one application, McMullen and Jensen demonstrated the Simplex method in a four-dimensional optimization of benzaldehyde production, increasing the yield from 21% to 80%.17 The Simplex method was further applied in optimizing the number of alkene equivalents and reaction time in a selective Heck reaction (Scheme 1).18 Both of these approaches used a microreactor and high-performance liquid chromatography (HPLC) feedback loop, which is illustrated in Figure 3 in application to the Heck reaction. Using the SuperModified Simplex algorithm,21 Parrott et al.22 optimized the methylation of primary alcohols by dimethyl carbonate in supercritical CO2 (Scheme 2). Here the advantage of the flow system was both in rapid optimization and in the accessibility of supercritical reaction conditions. Recent applications of flow optimization systems have demonstrated incorporation of more advanced automated synthesis and analytical systems into the Simplex and SNOBFIT algorithm frameworks. Fabry et al.23 coupled LabVIEW software to a commercial flow hydrogenation system,
traditional central composite experimental design to that identified with SNOBFIT feedback optimization. Pushing the possibilities of robotic synthesis, both Fitzpatrick et al.26 and Skilton et al.27 demonstrated remote-controlled automated platforms, allowing experimenters without flow hardware access to automated optimization equipment on different sides of the globe. Despite the popularity of the method, Simplex optimization is known to scale poorly to larger multivariable systems, and for all systems the algorithm’s heuristic convergence and limitation as a local optimization method make efficiency problemdependent and accuracy initialization-dependent.28 Though advantageous as a global search strategy, SNOBFIT has also been shown to be slow to converge and inefficient for wellbehaved chemical systems.17,25 Among black box strategies, gradient-based optimization strategies generally offer much faster convergence rates and an estimate of response surface curvature amenable to kinetic investigations, at the expense of the extra experiments involved in gradient calculation. To demonstrate, Moore and Jensen29 optimized production rate for the Paul−Knorr reaction in Scheme 6 using Fourier transfer infrared spectroscopy with three different gradient-based searches: steepest descent, conjugate gradient, and conjugate gradient with an Armijo30 step size. The rate of convergence effectively doubled with the conjugate gradient method compared to the steepest descent method (Figure 4), because
Scheme 2. Methylation of Primary Alcohols in Supercritical CO2 Studied by Parrott et al.22
C
DOI: 10.1021/acs.accounts.6b00261 Acc. Chem. Res. XXXX, XXX, XXX−XXX
Article
Accounts of Chemical Research
estimated reaction kinetics and heat of reaction upon the residence time distribution and temperature models in the Corning Advanced-Flow reactor. In an example of applying the D-optimal design of experiments (DoE) strategy to a more complex reaction pathway, we considered the estimation of reaction kinetics in a series-parallel nucleophilic aromatic substitution (SNAr) reaction (Scheme 8).34 Unlike for the bimolecular Scheme 7, we
Scheme 6. Paul−Knorr Reaction Optimization Studied by Moore and Jensen29
of the additional gradient history incorporated into the selection of the conjugate gradient search direction, and tripled with a smarter selection of step size following the Armijo rule.
Scheme 8. Series-Parallel SNAr Reaction from the Kinetic Study by Reizman and Jensen34
Figure 4. Convergence of the Paul−Knorr reaction from the automated system of Moore and Jensen.29 (Solid red square) Steepest descent algorithm. (Solid green up triangle) Conjugate gradient algorithm with fixed step size. (Open blue circle) Conjugate gradient algorithm with Armijo step size. (Open black square) Optimal conditions. Normalized production rate is defined as conversion per minute residence time.
Kinetics in Flow: A Route to Faster Scale-up
found several kinetic parameters from an initial screen to have high uncertainty because of the low sensitivity of the kinetic model to changes in experimental conditions. This problem was overcome by isolating samples of the two monosubstituted products and decoupling the reaction pathway to identify the kinetics of isolated reaction steps. Though successful in this application, translation of this divide-and-conquer approach to more challenging chemical syntheses is likely to be expensive because of the complex physical system needed to handle so many combinations of reagents in flow.
Although black box strategies are valuable tools for identifying improved reaction conditions in less time than combinatorial or one-factor-at-a-time screening, predictable scalability of results can only come from a complete understanding of the reaction mechanism and kinetics. Identification of reaction kinetics in flow offers researchers the advantages of fast heat and mass transfer and precise control of reaction conditions. Additionally, with feedback an automated system can determine which kinetic experiments are most valuable to run in order to select optimal rate parameters and to discriminate among candidate rate laws. As an example of using feedback in the determination of reaction kinetics, McMullen and Jensen31 used a silicon microreactor flow system with online HPLC sampling to study the Diels−Alder reaction of isoprene and maleic anhydride (Scheme 7). Using Shannon’s entropy maximization,32 a feedback algorithm selected experiments designed to maximize the discrepancy between the predictions of four kinetic rate laws. After identification of the correct kinetic model, a D-optimal experimental design strategy33 was employed to minimize the uncertainty in the best-fit kinetic parameters. A 500-fold scale-up was realized by overlaying the
Optimization with Discrete Variables
Unlike the single-stream flow platforms presented above, droplet flow systems35,36 allow experimenters the opportunity to manipulate discrete variables such as catalysts or solvents in addition to the standard continuous variables of temperature, reaction time, and concentration, greatly expanding the design space for the synthesis. By performing reactions in the confines of isolated droplets, reagent compositions can be controlled accurately, with recirculation patterns within the droplet producing mixing and heat transfer profiles mimicking those of a microscale batch reactor.37 Being low-dispersion, these systems are not required to hold until steady-state for an accurate measurement to be taken, and the amount of reaction material consumed can be nearly as small as the size of the online sample itself. In an example application of droplets to discrete variable optimization, Kreutz et al.38 incorporated a global optimization algorithm into high-throughput catalyst screening for methane oxidation (Figure 5). After a generation of 48 flow experiments was completed, the fitness of each reaction was assessed by a
Scheme 7. Diels−Alder Reaction Used in the Kinetic Study by McMullen and Jensen31
D
DOI: 10.1021/acs.accounts.6b00261 Acc. Chem. Res. XXXX, XXX, XXX−XXX
Article
Accounts of Chemical Research
Figure 5. Feedback loop for the optimization of methane oxidation by Kreutz et al.38 Reproduced in part from ref 38. Copyright 2010 American Chemical Society.
color indicator, and a genetic algorithm39 was applied to identify a new generation of catalyst, cocatalyst, and ligand combinations to be screened. After eight generations, the optimal active catalyst system was identified in agreement with catalyst systems already reported in literature. Substantiating the tremendous time and cost savings achievable with the application of feedback to discrete variable screening, the optimization required fewer than 400 experiments be conducted out of greater than 8 million possible catalyst/ cocatalyst/ligand combinations.
printed manifold (Solid Concepts, Inc.) continuously sparged with inert gas (Figure 6e). Droplet Transport and Reaction
Droplets were carried through the microsystem by inert gas, to avoid complications arising from the partial miscibility of a fluorinated carrier phase with organic reagents41 and “budding” that occurred with flow rate acceleration.42 Being especially challenging to degas, fluorinated compounds were also unsuitable for oxygen-sensitive chemistries. The use of inert gases such as nitrogen or argon avoided the limitations of using fluorinated solvents, though wetting, gas compressibility, and gas permeability all became new problems to address. Carryover was minimized with the use of 750 μm inner diameter Teflon fluorinated ethylene propylene (FEP) as the tubing material and superflangeless fittings and ferrules (Upchurch Scientific) to minimize the number of constrictions along the flow path. To limit pulsations in gas delivery yet maintain control of flow rate, gas was pushed through the Teflon FEP tubing from a stainless steel syringe (Harvard 8 mL) with a syringe pump (Harvard PhD 2000) operating at 6.9 bar. Pressure in the system was controlled with a 40 mL inert gas-regulated Parr bomb, which was drained automatically during refill of the gas delivery syringe. For controlled activation of the reaction entering the reactor, we adopted the approach of Hatakeyama et al.43 of introducing the activating reagent online through a T-junction. The online injection reagent was stored under an inert atmosphere and sampled by a 100 μL or a 250 μL glass syringe (Gastight, Hamilton Company) driven by a syringe pump (Harvard PhD 2000). As a droplet passed through the 500 μm T-junction (Upchurch Scientific), the syringe pump infused a volume of 2−10 μL reagent into the droplet. Refractive index sensors (EESPX613, Omron Corporation) were attached to the tubing before and after the T-junction to correctly time the online reagent injection and to verify that the droplet volume was within an acceptable tolerance following the online injection. Glass syringes and Luer connections (Upchurch Scientific) were regularly inspected for leaks. The microreactor comprised 240 μL Teflon FEP tubing inserted into a 1.6 mm groove cut into in an aluminum block (Figure 6f). A sheet of polycarbonate, compressed against a raised lip on the aluminum block, allowed for pressurization of the reactor to 6.9 bar. With this device, we were able to rapidly heat and cool the reactor tubing between 30 and 120 °C and to neutralize gas permeation out of the reactor. Residence times in the reactor were maintained between 1 and 10 min at gas flow rates of 15−250 standard μL/min. Blank droplets were not
■
A SYSTEM FOR AUTOMATED DISCRETE AND CONTINUOUS VARIABLE OPTIMIZATION Building upon the droplet flow approach, we constructed an automated system that allowed for on-demand manipulation of both continuous variables and discrete variables using microfluidic reaction droplets. Coupled with online feedback, such a system we envisioned would be first of its kind in solving problems deductively with the full arsenal of experimental variables available to the process chemist. The system operated as a continuous feedback loop (Figure 6a), generating from a supply of reagents sub-20 μL droplets surrounded by inert gas that were activated, reacted, and quenched before being sampled online by liquid chromatography/mass spectrometry (LC/MS) (Figure 6b−d). Real-time data were then fed to an algorithm, which selected new combinations of reagents and reaction conditions to test en route to optimization. Droplet Preparation
For ease of reagent handling, we screened liquid phase reactions and homogeneous catalysts. On-demand reagent sampling was accomplished using an automated liquid handling robot (Gilson GX-271, Gilson, Inc.) controlled with LabVIEW software (National Instruments, ver. 8.6). Reagents in solution for a given experiment were aspirated sequentially into the liquid handler probe, with probe rinses included to minimize carryover between reagent containers. Following sample aspiration, the sample was “stirred” three times in the probe under inert conditions by oscillating the flow of the aspirated droplet with a syringe pump (Harvard PhD 2000). The reagents were then transferred into a 6 port-2 way injection valve (Cheminert 10S-0503H, Valco Instruments Co. Inc.) containing a 14-μL sample loop. Switching of the sample loop to the inject position created a droplet. To clean the autosampler probe, valve, and system between experimental droplets, three blank droplets were prepared before every ondemand droplet experiment. For inertness, samples were stored under polytetrafluoroethylene-lined rubber septa in a 3DE
DOI: 10.1021/acs.accounts.6b00261 Acc. Chem. Res. XXXX, XXX, XXX−XXX
Article
Accounts of Chemical Research
Figure 6. (a) Concept diagram for on-demand preparation, reaction, analysis, and feedback in automated reaction flow screening. (b) Diagram of the droplet flow optimization system developed by Reizman and Jensen.40 (c) System components: syringe pumps, pump refill valves, automated liquid handler and probe, and pressure control. (d) System components: inert gas manifold, droplet injection valve, pressurized tubular reactor, temperature controller, and LC/MS injection valve. (e) Septum-sealed inert gas manifold for reagent storage. (f) Pressure-sealed reactor comprising a Teflon fluorinated ethylene propylene (FEP) tube in an aluminum housing, polycarbonate cover, thermocouple, and tetrahydrofuran droplets.
prepared and introduced into the flow system until a reaction droplet had traversed a full reactor volume, to protect against droplet acceleration from thermocapillary flow.44
(Rheodyne MXP7960−000), the sample was again mixed with quench solution and transported to a second 6-port, 2-way valve (Agilent G1158A) with a 1 μL polyether ether ketone sample loop. This smaller sample was injected into an LC/MS (Agilent G1312B binary pump, G1329B ALS, G1316A column compartment, G1365C multiwavelength detector, 6120 quadrupole MS). The sample was filtered (0.5 μm filter, Upchurch Scientific) and then split by pressure difference between a 1.8 μm particle diameter column (Agilent Zorbax SB-C18 2.1 × 50 mm) and a 4.6 μm particle diameter column (Agilent Zorbax
Analysis and Automation
Downstream of the reactor, droplets were mixed at room temperature with a continuously flowing quench solution. A third refractive index sensor was placed downstream of the quench to reproducibly time sampling by HPLC. Following sampling with a 30 μL sample loop in a 6-port, 2-way valve F
DOI: 10.1021/acs.accounts.6b00261 Acc. Chem. Res. XXXX, XXX, XXX−XXX
Article
Accounts of Chemical Research SB-C18 2.1 × 50 mm). The sample from the 1.8 μm particle diameter column was detected by ultraviolet (UV) absorbance and MS. The completed UV chromatogram was integrated using ChemStation software and passed to MATLAB (Mathworks, ver. R2011a) for automated calculation of product yields. LabVIEW software controlled valve manipulation, refractive index monitoring, and serial communication to all pumps, the liquid handler, and the reaction temperature controller. Data for individual droplet experiments were recorded in a single MATLAB matrix which was regularly accessed and updated by LabVIEW. To optimize both continuous and discrete variables simultaneously, we developed an optimization program in MATLAB that used an adaptive response surface methodology coupled with an overarching branch-and-bound framework to eliminate candidate discrete variables as experiments progressed. The algorithm, illustrated in Figure 7, operated
to create even less uncertainty in the model prediction following the next experimental iteration.
■
DISCRETE AND CONTINUOUS FEEDBACK OPTIMIZATION IN CHEMICAL SYSTEMS
Solvent Optimization in Alkylation
As a demonstration of system versatility, we considered optimization of the alkylation reaction in Scheme 9 for ten Scheme 9. Alkylation Reaction Studied for Simultaneous Solvent Screening and Reaction Optimization by Reizman and Jensen40
organic solvents and three continuous variables, including elevated temperature.40 Using the optimal DoE approach, our algorithm refined response surface predictions for yield of the desired monosubstituted product as a function of temperature, reaction time, and electrophile concentration and progressively removed candidate solvents from the screen as they became less likely to produce an overall optimum (Figure 8a). Collectively 67 experiments were required to identify the best solvent as dimethyl sulfoxide (DMSO) and the best conditions at which to run the reaction to achieve the highest yield of monoalkylated product. A gradient-based quasi-Newton search algorithm was then employed to validate the continuous variable optimum, converging upon an optimal yield of 62%. Batch scale-up of the same optimal conditions generated the desired product in 61% yield by HPLC and 59% yield upon isolation. From Figure 8b, we observed that our application of a response-surface-based feedback strategy had the desired effect of concentrating continuous variable experiments and consolidating discrete variable experiments near the eventual optimum, producing fast convergence and, with the application of optimal DoE, relatively little uncertainty in the best reaction conditions. Although the solvents in our study were treated as uncorrelated integer variables, our software also rapidly deduced that polar aprotic solvents gave faster conversion to both the monosubstituted and oversubstituted adducts. The algorithm recognized that experiments with these solvents were best conducted at more moderate temperatures whereas solvents such as tetrahydrofuran (THF), toluene, and 1,2dichloroethane were much more likely to produce better yield of the desired product at higher temperatures and longer reaction times (Figure 8c). From the final optimization results, we observed preferentiality in the optimization for polar aprotic solvents such as DMSO, N,N-dimethylformamide (DMF), and pyridine, and found that discrete-variable optimal yields correlated strongly with the solvent hydrogen bond basicity.40
Figure 7. Real-time discrete and continuous variable optimization decision diagram.
similarly to feedback DoE,45 initializing with two fractional factorial experimental designs employed to populate response surfaces with respect to each discrete variable. These response surfaces were then optimized, with the optima of candidate discrete variables assessed on the likelihood of being the overall discrete variable optimum. Discrete variables with poor performance were removed from subsequent experimentation. For the remaining candidate variables, new experiments were chosen at continuous conditions that best challenged the predicted optima (using a G-optimality criterion46), in an effort
Precatalyst Selection in Suzuki−Miyaura Cross-Couplings
In very few instances are the integration of discrete and continuous variables as pronounced as in organometallic catalysis, with one of the most industrially significant examples being the Suzuki−Miyaura cross-coupling reaction.47,48 As a test of the adaptability of our approach to optimization in catalytic systems, we examined four case studies of optimizing palladacycle-ligand precatalyst selection and conditions in G
DOI: 10.1021/acs.accounts.6b00261 Acc. Chem. Res. XXXX, XXX, XXX−XXX
Article
Accounts of Chemical Research
Figure 8. Simultaneous solvent screening and reaction optimization by Reizman and Jensen.40 (a) Predicted maximum desired product yield as a function of solvent after n experiments. (b) Map of experiments conducted. (c) Model-predicted optimal temperature and yield as a function of solvent (blue, optimal temperature 70−90 °C; yellow, optimal temperature 90−110 °C; red, optimal temperature ≥ 110 °C). Reproduced in part from ref 40 with permission from the Royal Society of Chemistry.
Scheme 10. Suzuki−Miyaura Cross-Coupling Reactions Studied for Simultaneous Precatalyst Screening and Reaction Optimization and Optimal Conditions Identified49
Suzuki−Miyaura cross-couplings (Scheme 10).49 To minimize catalyst loading while generating high yield, systems were optimized for the turnover number (TON) of product relative
to catalyst loading, with the constraint that the yield be no less than 90% of the overall maximum. Illustrated in Figure 9a and b, the preference for different ligands at optimized conditions H
DOI: 10.1021/acs.accounts.6b00261 Acc. Chem. Res. XXXX, XXX, XXX−XXX
Article
Accounts of Chemical Research
of being good electron donors, continued to promote oxidative addition even in the presence of the chloride substrate. The shift from boronic acid pinacol ester substrate in cases I and II to the boronic acids in cases III and IV effected preferentiality for both dialkylbiarylphosphine ligand precatalysts and shorter reaction times of below 10 min. We attributed the favorability for biarylphosphine ligands to faster oxidative addition and rapid transmetalation to L1Pd(Ar)Cl intermediates in these cases, allowing the coupling process to outcompete decomposition of the boronic acid and product. For case IV, this gave rise to the response surface for P1-L1 in Figure 10, with an optimum at 97 °C and 5 min later validated
Figure 10. Response surface generated by automated optimization system for Scheme 10 case IV at optimal catalyst loading of 1.0% P1L1 (P1-XPhos). Figure 9. (a) Precatalyst selection frequency and (b) optimal precatalyst and TON by case. Model-predicted optimal temperature and yield as a function of precatalyst (blue, optimal temperature < 80 °C; yellow, optimal temperature 80−100 °C; red, optimal temperature ≥ 100 °C) in Scheme 10 (c) case II and (d) case IV.
by traditional screening, and the overall preferentiality for palladacyle-dialkylbiarylphosphine precatalysts at temperatures between 80 and 100 °C illustrated in Figure 9d. Palladacycletrialkylphosphine catalysts, being less active at these more moderate temperatures, were most suitable above 100 °C and at consequently shorter reaction times, where product degradation was less detrimental to yield. Given our system’s ability to discriminate among catalysts with no prior information, we anticipate that this approach together with chemoinformatics tools will ultimately allow for automated elucidation of shared attributes of successful discrete variables, serving in the future as a guide to more intelligent screening and mechanism development.
changed depending upon the coupling partners, as expected. Unique to simultaneous discrete and continuous variable optimization was that the optimal catalyst loading, time, and temperature also differed based on the chemistry. This led the algorithm to identify precatalytic systems that performed well at higher temperatures that would have been removed from consideration in a screen at more moderate conditions, evidenced by comparing cases II (where P1-L5 was optimal) and IV (where P1-L1 was optimal) in Figure 9(c-d). In the context of the Suzuki−Miyaura reaction mechanism,50 we reasoned that the change in precatalyst preference implied a transition in rate-limiting step when converting substrates from the aryl bromide in case I to the aryl chlorides in cases II, III, and IV. Figure 9b shows the significant decline in optimal turnover number observed with the application of P1-L4 (P1XantPhos) with chloride substrates, whereas the optimal performance for the other precatalysts in the study only declined slightly or improved, as in the case of P1-L5 (P1PCy3). This suggested that the effectiveness of the bidentate ligand P1-L4 in accelerating the rate of transmetalation with the bromide was offset by a slow rate of oxidative addition with the chloride. By contrast, precatalysts incorporating trialkylphosphine ligands L5 and L7 (PtBu3) and dialkylbiarylphosphine ligands L1 (XPhos), L2 (SPhos), and L3 (RuPhos), by nature
■
CONCLUSIONS In short time, the field of continuous feedback optimization has grown from novelty to necessity. As syntheses increase in complexity, scientists are looking toward smarter approaches to solving more integrated problems and to technologies designed to save time by producing more rapidly scalable results. The approach presented herein makes strides in these directions, all while allowing users a “set-it-and-forget-it” approach to reaction characterization. While challenges of organic solvent miscibility and reagent carryover have been overcome, opportunities still exist to expand automated feedback technology to multiphase reaction systems, more dynamic reaction times, and less stable feed solutions. Faster analytical methods will enable higher throughput and the technology will be enhanced by real-time I
DOI: 10.1021/acs.accounts.6b00261 Acc. Chem. Res. XXXX, XXX, XXX−XXX
Article
Accounts of Chemical Research
(8) Jensen, K. F.; Reizman, B. J.; Newman, S. G. Tools for Chemical Synthesis in Microsystems. Lab Chip 2014, 14, 3206−3212. (9) Hartman, R. L.; McMullen, J. P.; Jensen, K. F. Deciding Whether To Go with the Flow: Evaluating the Merits of Flow Reactors for Synthesis. Angew. Chem., Int. Ed. 2011, 50, 7502−7519. (10) Fabry, D. C.; Sugiono, E.; Rueping, M. Online Monitoring and Analysis for Autonomous Continuous Flow Self-Optimizing Reactor Systems. React. Chem. Eng. 2016, 1, 129−133. (11) Schmink, J. R.; Bellomo, A.; Berritt, S. Scientist-Led HighThroughput Experimentation (HTE) and Its Utility in Academia and Industry. Aldrichimica Acta 2013, 46, 71−80. (12) Buitrago Santanilla, A.; Regalado, E. L.; Pereira, T.; Shevlin, M.; Bateman, K.; Campeau, L. C.; Schneeweis, J.; Berritt, S.; Shi, Z. C.; Nantermet, P.; Liu, Y.; Helmy, R.; Welch, C. J.; Vachal, P.; Davies, I. W.; Cernak, T.; Dreher, S. D. Nanomole-Scale High-Throughput Chemistry for the Synthesis of Complex Molecules. Science 2015, 347, 49−53. (13) Elvira, K. S.; i Solvas, X. C.; Wootton, R. C. R.; deMello, A. J. The Past, Present and Potential for Microfluidic Reactor Technology in Chemical Synthesis. Nat. Chem. 2013, 5, 905−915. (14) McMullen, J. P.; Jensen, K. F. Integrated Microreactors for Reaction Automation: New Approaches to Reaction Development. Annu. Rev. Anal. Chem. 2010, 3, 19−42. (15) Moore, J. S.; Jensen, K. F. Automation in Microreactor Systems. In Microreactors in Organic Chemistry and Catalysis, Second ed.; Wirth, T., Ed.; Wiley-VCH Verlag GmbH & Co. KGaA: Weinheim, Germany, 2013. (16) Krishnadasan, S.; Brown, R. J. C.; Demello, A. J.; Demello, J. C. Intelligent Routes to the Controlled Synthesis of Nanoparticles. Lab Chip 2007, 7, 1434−1441. (17) McMullen, J. P.; Jensen, K. F. An Automated Microfluidic System for Online Optimization in Chemical Synthesis. Org. Process Res. Dev. 2010, 14, 1169−1176. (18) McMullen, J. P.; Stone, M. T.; Buchwald, S. L.; Jensen, K. F. An Integrated Microreactor System for Self-Optimization of a Heck Reaction: From Micro- to Mesoscale Flow Systems. Angew. Chem., Int. Ed. 2010, 49, 7076−80. (19) Huyer, W.; Neumaier, A. SNOBFIT–Stable Noisy Optimization by Branch and Fit. ACM T. Math. Software 2008, 35, 9. (20) Nelder, J. A.; Mead, R. A Simplex Method for Function Minimization. Comput. J. 1965, 7, 308−313. (21) Routh, M. W.; Swartz, P. A.; Denton, M. B. Performance of the Super Modified Simplex. Anal. Chem. 1977, 49, 1422−1428. (22) Parrott, A. J.; Bourne, R. A.; Akien, G. R.; Irvine, D. J.; Poliakoff, M. Self-Optimizing Continuous Reactions in Supercritical Carbon Dioxide. Angew. Chem., Int. Ed. 2011, 50, 3788−3792. (23) Fabry, D. C.; Sugiono, E.; Rueping, M. Self-Optimizing Reactor Systems: Algorithms, On-Line Analytics, Setups, and Strategies for Accelerating Continuous Flow Process Optimization. Isr. J. Chem. 2014, 54, 341−350. (24) Sans, V.; Porwol, L.; Dragone, V.; Cronin, L. A Self Optimizing Synthetic Organic Reactor System Using Real-Time In-Line NMR Spectroscopy. Chem. Sci. 2015, 6, 1258−1264. (25) Holmes, N.; Akien, G. R.; Savage, R. J. D.; Stanetty, C.; Baxendale, I. R.; Blacker, A. J.; Taylor, B. A.; Woodward, R. L.; Meadows, R. E.; Bourne, R. A. Online Quantitative Mass Spectrometry for the Rapid Adaptive Optimisation of Automated Flow Reactors. React. Chem. Eng. 2016, 1, 96−100. (26) Fitzpatrick, D. E.; Battilocchio, C.; Ley, S. V. A Novel InternetBased Reaction Monitoring, Control and Autonomous Self-Optimization Platform for Chemical Synthesis. Org. Process Res. Dev. 2016, 20, 386−394. (27) Skilton, R. A.; Bourne, R. A.; Amara, Z.; Horvath, R.; Jin, J.; Scully, M. J.; Streng, E.; Tang, S. L. Y.; Summers, P. A.; Wang, J.; Perez, E.; Asfaw, N.; Aydos, G. L. P.; Dupont, J.; Comak, G.; George, M. W.; Poliakoff, M. Remote-Controlled Experiments with Cloud Chemistry. Nat. Chem. 2015, 7, 1−5. (28) Wright, M. H. Direct Search Methods: Once Scorned, Now Respectable. Pitman Res. 1996, 191−208.
reaction modeling and chemoinformatics coupled with machine-learning algorithms. Overall the outlook for this and other automated feedback optimization systems is positive, given the needs in academia and industry for tools that can rapidly and accurately characterize reactions using minimal material and time. We anticipate that with continued advancement, automated reaction characterization systems will provide researchers with greater insight into reaction mechanisms, guiding the strategy toward process optimization and scale-up, and ultimately fulfilling the vision of “accelerated reaction development.”
■
AUTHOR INFORMATION
Corresponding Author
*E-mail:
[email protected]. Fax: 1-617-258-8224. Tel: 1-617253-4589. Present Address †
B.J.R.: Eli Lilly and Company, 1555 Harding Street, Indianapolis, IN 46221, USA.
Funding
We thank the Novartis Center for Continuous Manufacturing for funding this research. Notes
The authors declare no competing financial interest. Biographies Brandon J. Reizman received a B.S. in chemical engineering from the University of IllinoisUrbana−Champaign and a M.S.C.E.P. and Ph.D. in chemical engineering from the Massachusetts Institute of Technology. He is an engineer at Eli Lilly and Company working on the design and development of continuous processes. Klavs F. Jensen received a M.Sc. in chemical engineering from the Technical University of Denmark and a Ph.D. in chemical engineering from the University of WisconsinMadison. He is the Warren K. Lewis Professor of Chemical Engineering at the Massachusetts Institute of Technology where he works on flow chemistry and engineering.
■
REFERENCES
(1) DiMasi, J. A.; Grabowski, H. G.; Hansen, R. W. Innovation in the Pharmaceutical Industry: New Estimates of R&D Costs. J. Health Econ. 2016, 47, 20−33. (2) Baxendale, I. R.; Braatz, R. D.; Hodnett, B. K.; Jensen, K. F.; Johnson, M. D.; Sharratt, P.; Sherlock, J.-P.; Florence, A. J. Achieving Continuous Manufacturing: Technologies and Approaches for Synthesis, Workup, and Isolation of Drug Substance. May 20−21, 2014 Continuous Manufacturing Symposium. J. Pharm. Sci. 2015, 104, 781− 91. (3) Anderson, N. G. Using Continuous Processes to Increase Production. Org. Process Res. Dev. 2012, 16, 852−869. (4) Kockmann, N.; Roberge, D. M. Harsh Reaction Conditions in Continuous-Flow Microreactors for Pharmaceutical Production. Chem. Eng. Technol. 2009, 32, 1682−1694. (5) Poechlauer, P.; Colberg, J.; Fisher, E.; Jansen, M.; Johnson, M. D.; Koenig, S. G.; Lawler, M.; Laporte, T.; Manley, J.; Martin, B.; O’Kearney-McMullan, A. Pharmaceutical Roundtable Study Demonstrates the Value of Continuous Manufacturing in the Design of Greener Processes. Org. Process Res. Dev. 2013, 17, 1472−1478. (6) Hessel, V.; Kralisch, D.; Kockmann, N.; Noel, T.; Wang, Q. Novel Process Windows for Enabling, Accelerating, and Uplifting Flow Chemistry. ChemSusChem 2013, 6, 746−789. (7) Hartman, R. L.; Jensen, K. F. Microchemical Systems for Continuous-Flow Synthesis. Lab Chip 2009, 9, 2495−2507. J
DOI: 10.1021/acs.accounts.6b00261 Acc. Chem. Res. XXXX, XXX, XXX−XXX
Article
Accounts of Chemical Research (29) Moore, J. S.; Jensen, K. F. Automated Multitrajectory Method for Reaction Optimization in a Microfluidic System using Online IR Analysis. Org. Process Res. Dev. 2012, 16, 1409−1415. (30) Armijo, L. Minimization of Functions Having Lipschitz Continuous First Partial Derivatives. Pac. J. Math. 1966, 16, 1−3. (31) McMullen, J. P.; Jensen, K. F. Rapid Determination of Reaction Kinetics with an Automated Microfluidic System. Org. Process Res. Dev. 2011, 15, 398−407. (32) Box, G. E. P.; Hill, W. J. Discrimination among Mechanistic Models. Technometrics 1967, 9, 57−71. (33) Box, G. E. P.; Lucas, H. L. Design of Experiments in Non-Linear Situations. Biometrika 1959, 46, 77−90. (34) Reizman, B. J.; Jensen, K. F. An Automated Continuous-Flow Platform for the Estimation of Multistep Reaction Kinetics. Org. Process Res. Dev. 2012, 16, 1770−1782. (35) Song, H.; Chen, D. L.; Ismagilov, R. F. Reactions in Droplets in Microflulidic Channels. Angew. Chem., Int. Ed. 2006, 45, 7336−7356. (36) Dressler, O. J.; Maceiczyk, R. M.; Chang, S. I.; deMello, A. J. Droplet-Based Microfluidics Enabling Impact on Drug Discovery. J. Biomol. Screening 2014, 19, 483−496. (37) Gunther, A.; Jhunjhunwala, M.; Thalmann, M.; Schmidt, M. A.; Jensen, K. F. Micromixing of Miscible Liquids in Segmented GasLiquid Flow. Langmuir 2005, 21, 1547−1555. (38) Kreutz, J. E.; Shukhaev, A.; Du, W. B.; Druskin, S.; Daugulis, O.; Ismagilov, R. F. Evolution of Catalysts Directed by Genetic Algorithms in a Plug-Based Microfluidic Device Tested with Oxidation of Methane by Oxygen. J. Am. Chem. Soc. 2010, 132, 3128−3132. (39) Goldberg, D. E. Genetic Algorithms in Search, Optimization, and Machine Learning; Addison-Wesley Pub. Co.: Reading, MA, 1989. (40) Reizman, B. J.; Jensen, K. F. Simultaneous Solvent Screening and Reaction Optimization in Microliter Slugs. Chem. Commun. 2015, 51, 13290−13293. (41) Gladysz, J. A.; Curran, D. P.; Horváth, I. T. Handbook of Fluorous Chemistry; Wiley-VCH: Weinheim, 2004. (42) Hawbaker, N.; Wittgrove, E.; Christensen, B.; Sach, N.; Blackmond, D. G. Dispersion in Compartmentalized Flow Systems: Influence of Flow Patterns on Reactivity. Org. Process Res. Dev. 2016, 20, 465−473. (43) Hatakeyama, T.; Chen, D. L.; Ismagilov, R. F. Microgram-Scale Testing of Reaction Conditions in Solution Using Nanoliter Plugs in Microfluidics with Detection by MALDI-MS. J. Am. Chem. Soc. 2006, 128, 2518−2519. (44) Burns, M. A.; Mastrangelo, C. H.; Sammarco, T. S.; Man, F. P.; Webster, J. R.; Johnsons, B. N.; Foerster, B.; Jones, D.; Fields, Y.; Kaiser, A. R.; Burke, D. T. Microfabricated Structures for Integrated DNA Analysis. Proc. Natl. Acad. Sci. U. S. A. 1996, 93, 5556−5561. (45) Atkinson, A. C.; Donev, A. N. Optimum Experimental Design; Clarendon Press; Oxford University Press: Oxford, New York, 1992. (46) Steinberg, D. M.; Hunter, W. G. Experimental-Design - Review and Comment. Technometrics 1984, 26, 71−97. (47) Miyaura, N.; Suzuki, A. Palladium-Catalyzed Cross-Coupling Reactions of Organoboron Compounds. Chem. Rev. 1995, 95, 2457− 2483. (48) Hartwig, J. F. Carbon-Heteroatom Bond Formation Catalysed by Organometallic Complexes. Nature 2008, 455, 314−322. (49) Reizman, B. J. Automated Feedback in Flow for Accelerated Reaction Screening, Optimization, and Kinetic Parameter Estimation. Ph.D. Dissertation, Massachusetts Institute of Technology, Cambridge, MA, 2015. (50) Miyaura, N. Metal-Catalyzed Cross-Coupling Reactions of Organoboron Compounds with Organic Halides. In Metal-Catalyzed Cross-Coupling Reactions; Wiley-VCH Verlag GmbH, 2008; pp 41− 123.
K
DOI: 10.1021/acs.accounts.6b00261 Acc. Chem. Res. XXXX, XXX, XXX−XXX
