Optimizing microflow sequential coupling and cyclization of multiple linear substrates guided by Bayesian optimization

Abstract

Sequential coupling and cyclization of 2 linear substrates containing multiple reaction sites enables single-step synthesis of useful cyclic compounds; however, if the reaction sites are highly reactive, overreactions tend to occur. Although the use of microflow synthesis can prevent overreactions, careful optimization of multiple variables is required. Herein, we optimized the microflow synthesis of a cyclic sulfamide via sequential coupling and cyclization of 2 linear substrates with 2 highly reactive sites. The traditional one-variant-at-a-time-based approach revealed nonlinear correlations between the variables and the yield of cyclic sulfamide, and the optimal conditions afforded a 90% yield. Subsequent re-optimization using the Bayesian optimization (BO)-based approach identified significantly different optimal conditions, giving the product in 94% yield. Additional experiments and simulations were conducted to investigate the key factors influencing the optimal conditions for the BO-based approach.

Graphical Abstract

Optimization of the sequential coupling and cyclization of 2 linear substrates containing multiple reaction sites using BO-based approach afforded higher yield compared with that using the traditional OVAT-based approach. The reasons for the higher yield obtained using the BO-based approach were investigated.

Open in new tabDownload slide

1. Introduction

The sequential coupling and cyclization of multiple linear substrates A and B, both containing multiple reaction sites, represents a powerful approach for the step-economic synthesis of useful cyclic compounds C (Scheme 1a). However, when the sites are highly reactive, the risk of overreaction, generating AB₂, A₂B, and/or ABAB, increases. In pioneering work by Dye et al. in 1973, rapid mixing enabled by flow technology was used to synthesize macrolactams F from linear substrates D and E, containing multiple highly reactive sites (Scheme 1b).¹ Continuous-flow technology has revolutionized synthetic organic chemistry in both academic and industrial fields.^2–10 We previously reported the microflow synthesis of cyclic phosphotriesters and their analogs I via sequential coupling and cyclization of substrates G and H (Scheme 1c)¹¹; however, both desired and undesired reactions occurred rapidly during the synthesis. The concentration and ratio of flow rates of the solutions seemed to affect not only the mixing efficiency but also the stoichiometric ratio of the substrates in the microscopic area where the mixing of the solutions occurred. Therefore, optimization of the reaction conditions was not simple.

Bayesian optimization (BO) is an efficient approach to establishing optimal reaction conditions using a restricted number of experiments.^12–15 BO trains Bayesian machine learning surrogate models using an iterative approach, aiming to enhance the predictive power of the surrogates. The use of BO to optimize batch/flow reactions with or without automated synthesis technology has been reported in recent years.^16–27 Optimization based on a reaction model with rate equations is the most reliable and canonical approach for chemical processes; however, it is difficult in the initial stages of investigation of organic synthesis. Screening of reagents (e.g. solvents, reactants, and catalysts) can be incorporated into the BO. Thus, BO is becoming an increasingly popular tool for organic chemistry research and development.

We previously developed a microflow synthesis of linear sulfamides guided by BO²⁸ and became interested in developing a microflow synthesis of cyclic sulfamide 3²⁹ by sequential coupling and cyclization of multiple linear substrates 1 and 2 (Scheme 1d). Herein, we initially optimized the reaction conditions using a traditional one-variant-at-a-time (OVAT) approach. We then re-optimized the conditions using a BO-based approach and found that the optimal conditions identified in this approach (94% yield) significantly differed from those identified in the OVAT-based approach (90% yield). The reasons for the higher yield obtained using the BO-based approach were investigated and discussed.

2. Results and discussion

2.1 OVAT-based optimization of reaction conditions

Initially, we examined the bases used for the synthesis of 3 (Table 1). In all the examinations, we did not observe the generation of undesired products but detected unreacted 1 and desired product 3 (entries 1-9). Trace amounts of the desired 3 were obtained in the absence of a base (entry 1). The use of a tertiary aliphatic amine, DIEA, afforded desired product 3 in the highest yield (entry 3, 74%). Both stronger (entry 2) and weaker (entries 4-9) bases afforded desired compound 3 in lower yields (trace-60%). We presumed that the stronger base abstracted the proton of 1 (pK_a = ca. 12),³⁴ which facilitates the desired coupling of 1 with 2. The use of a stronger base, DBU, might cause undesired nucleophilic attack of DBU on oxalyl chloride³⁵ that lead to a decrease in yield (entry 2), whereas weaker bases did not sufficiently abstract the proton; thus, the desired reaction did not proceed smoothly.

Next, the reaction temperature and solution concentration were examined (Table 2). Although the 0 and −20 °C conditions afforded the desired 3 in decreased yield probably due to the slow progress of the reaction (entries 1 and 2), the 20 and 40 °C conditions afforded the desired 3 in comparable yields (entries 3 and 4). Thus, we carried out the following examinations at 20 °C. The concentrations of the solutions were then examined; 0.20 and 0.15 M conditions afforded the desired 3 in the highest yields (entries 6 and 7), whereas both higher (0.40 M) and lower (0.05 M) concentrations afforded the lower yields (entries 5 and 8). When the higher concentration (entry 5) was employed, substrate 1 was not completely dissolved in CH₂Cl₂ and appeared as a suspension. We speculated that the use of a suspension of 1 decreased the efficiency of mixing with a solution of 2 in the T-shape mixer and reaction tube, decreasing the yield of 3. In contrast, the lower concentration (entry 8) slowed the desired coupling of 1 with 2, significantly decreasing the yield.

Finally, the stoichiometric ratio of 1 and 2 as well as the flow rates of syringe pumps I and II were examined (Table 3). Interestingly, although the use of equivalent amounts of 1 and 2 resulted in 74% yield of 3 (entry 4), and a slight excess of either 1 (entry 3: 1/2 = 1.2) or 2 (entry 5: 2/1 = 1.2) gave higher yields (84% and 88%, respectively). Excess use of 2 (entry 6: 2/1 = 1.5) afforded the desired 3 in the highest yield (90%). However, the further excess use of either 1 (entries 1 and 2: 1/2 = 2.5, 1.5) or 2 (entries 7 and 8: 2/1 = 2.0, 2.5) decreased the yields. Regarding these results, we speculated that when using equal amounts of substrates 1 and 2, the concentration of 1 and 2 decreases in the latter stage of the reaction, avoiding complete consumption of 1 and 2. On the other hand, using an excess of either 1 or 2 avoids the decrease in reaction rate in the latter stage, allowing further consumption of the substrate. The reason why the highest yield was obtained when 2 was used in excess will be discussed in detail in section 2.4. The decrease in yield when either 1 or 2 was used in large excess is believed to be due to the overreaction of either 1 or 2.

We then examined the influence of the flow rates of syringe pumps I and II on the yield of 3. Although the flow rates of Z and W = 1.6:1.6 gave comparable results (entry 6 vs. 10), the flow rates of Z and W = 1.2:2.0 gave a slightly lower yield (entries 6 vs. 9). The use of a 1.5 times excess amount of 2 against 1 and 1.67 times higher flow rate of the solution of 1 compared to that of the solution of 2 (entry 6) was identified as the optimal condition using the OVAT-based approach.

OVAT-based optimizations revealed that no variables had simple linear correlations with the yields (Fig. 1). In particular, the observed correlations between the relative amounts of 1 and 2 and the yield of 3 are complicated (Fig. 1d). We calculated the correlation coefficients between the employed variables in the OVAT-based reaction condition optimization and visualized them using a heatmap (for details, see Supplementary Information, section 5). The best conditions identified by the OVAT-based approach afforded the desired 3 in high yield (Table 3, entry 6, 90%); however, we were concerned that the real optimal conditions were overlooked because of the complicated correlations between variables and yields.

2.2 BO-based optimization of reaction conditions

We re-examined the reaction conditions using the BO approach. Four continuous variables (pK_aH of base P, temperature T, concentration V, and the ratio of flow rates Z/W, yielding 8 × 7 × 5 × 7 = 1,960 combinations) were selected (Table 4) based on the results of the OVAT-based optimizations. The Scikit-learn machine-learning Python library was used, based on our previous report.²⁸

We performed the optimization (Table 5 and Fig. 2) with a set of 5 initial experiments generated by LHS. The following 10 experiments based on BO identified the conditions {V = 0.10 M, P = 11.4 (DIEA), Z/W = 0.45, W = 30 °C)} affording slightly higher yield (Table 5, entry 15, 94%) than that obtained from the OVAT-based approach (Table 3, entry 6, 90%). We calculated the correlation coefficients between the employed variables in the BO-based reaction condition optimization and visualized them using a heatmap (for details, see Supplementary Information, section 6).

2.3 Analysis of the optimal conditions obtained using the BO-based approach

Because the difference between the best yield of the OVAT-based approach (90%) and that of the BO-based approach (94%) was small, we performed 3 independent experiments under both conditions (Table 6, entries 1 and 2). Thus, the reproducibility and accuracy of the obtained yields were confirmed. All the examined variables, except for the concentration of the solution of 1, differed between the 2 best conditions (entries 1 vs. 2). We then attempted to identify the key factors under the optimal conditions obtained from the BO-based approach (entry 2). Initially, we increased the relative amount of 2 against 1 Y/X (from 1.5 to 2.2) in the best conditions identified by the OVAT-based approach, and a decrease in yield was observed (entries 1 vs. 3). Next, we raised the reaction temperature (from 20 to 30 °C) with an increased relative amount of 2 against 1 Y/X (2.2). The observed yield (entry 4) was comparable to that obtained under the best conditions identified by the OVAT-based approach (entry 1), but lower than that obtained under the best conditions identified by the BO-based approach (entry 2). Then we altered relative flow rates Z/W (from 1.67 to 1.00 or 0.60) with the increased relative amount of 2 against 1 Y/X (2.2) and raised temperature conditions (30 °C). The observed yields (entries 5 and 6) were again comparable to those obtained under the best conditions identified by the OVAT-based approach (entry 1), but lower than those obtained under the best conditions identified by the BO-based approach (entry 2). None of the examined variables seemed to have a dominant effect on yield improvement, but combinations of multiple variables seemed to influence the yield. In the OVAT-based optimization, 20 and 40 °C conditions resulted in similar yields (74% and 75%, Table 2, entries 3 and 4). Therefore, the temperature difference between 2 optimal conditions (20 vs. 30 °C, Table 6, entry 1 vs. 2) seemed not a crucial factor. We presumed that the combination of concentration and equivalence ratio was important for this reaction.

2.4 Reaction modeling and simulation

Numerical simulation based on reaction rate analysis was performed to evaluate the effect of the combination of concentration and equivalent ratio on the yields under the identified optimal conditions (Table 6, entries 1 and 2). We hypothesized that the yield of the desired 3 is mainly affected by the following 3 reactions: the desired reaction affording 3 (Scheme 2a), the undesired overreaction of 2 to 1 affording 4 (Scheme 2b), and the base-mediated undesired decomposition of 2 (Scheme 2c). Omitting unimportant reactions before fitting is important to reduce the number of parameters and avoid overfitting. Therefore, we did not consider the undesired overreaction of 1 to 2 because both optimal conditions used an excess amount of 2 against 1. The undesired reaction (Scheme 2c) induced by DIEA was experimentally examined using in-line IR analysis and ¹³C NMR analysis. Both examinations indicated that the decomposition of COCl₂ occurred in the presence of DIEA (for details, see Supplementary Information, section 10).

We assumed 3 rate Equations (1)-(3) for the 3 reactions shown in Scheme 2a–c, respectively:

The rate constants k_des, k_und1, and k_und2 were determined to reproduce the yield of 3 under the 2 optimal conditions (Table 6, entries 1 and 2), as well as the conditions examined in the OVAT-based optimization (Table 2, entry 3) and the modified optimal conditions identified by the OVAT-based approach (Table 6, entry 4). Eliminating the activation energies is important to avoid overfitting the results and to elucidate how the combination of the concentration and equivalence ratio influences the reaction. Therefore, we assumed kinetic constants are identical although these conditions have different temperatures of 20 and 30 °C. The rate constants were determined to be k_des = 42.1 L³·mol⁻³·s⁻¹, k_und1 = 48.1 L²·mol⁻²·s⁻¹, and k_und2 = 0.34 L⁴·mol⁻⁴·s⁻¹ by minimizing the sum of the square of the relative errors. The yields of 3 under each condition were reproduced with an error of 1% (Scheme 2d and e). In addition, the yields of 3 under the other 2 conditions with different concentrations and/or equivalent ratios (Table 2, entry 3 and Table 6, entry 4) were also reproduced with an error of 1% (for details, see Supplementary Information, section 12).

The simulated time courses of the concentrations of 1, 2, and 3 under the 2 optimal conditions identified by the OVAT-based (Fig. 3a) and BO-based approaches (Fig. 3b) are shown. The time course in Fig. 3a shows that the reaction proceeded rapidly in the early stages, and even after 10 s, ca. 86% of 3 was produced. However, the reaction slowed, and after 30 s, 91% of the target product was obtained, 3% of 1 remained unreacted, and the remaining 6% of 1 was consumed in an overreaction with 2 (Scheme 3b). Similarly, the time course in Fig. 3b shows that the reaction proceeds rapidly in the early stages; however, the reaction of 1 is almost complete owing to the larger excess amount of 2. This reaction yielded ca. 93% of the desired product after 30 s, with the remaining 5% of 1 being consumed in an overreaction with 2 (Scheme 2b). These simulation results suggest that a larger excess use of 2 under the optimal conditions identified by the BO-based approach contributed to the complete consumption of 1, which led to a higher yield. This is because 2 was continuously consumed by 2 undesired reactions (Scheme 2b and c).

Not only the equivalence ratio but also the concentration of the solutions is important in this reaction. As already shown, increasing the amount of 2 by simply increasing the concentration of a solution of 2 from optimal conditions identified by the OVAT-based approach resulted in a decrease in yield (Table 6, entries 1 vs. 3). This is presumably due to the undesired overreaction of 2 (Scheme 2b). The rate equation for this overreaction would be second order to a concentration of 2. When the concentration of the solution of 2 was doubled, the rate of the desired reaction also doubled, but that of the undesired overreaction quadrupled. The simulated reaction profile for increased concentrations of 2 (Supplementary Fig. S4b) agrees well with the experimental results and the above explanation. The simulated yield was 90% (Supplementary Table S1), which is identical to that of the experiment. The desired reaction was completed in 5 s; however, 10% of 1 was consumed during the overreaction (Supplementary Fig. S4b). It is conceivable that increasing the amount of 2 by increasing the flow rate of the solution of 2 while keeping its concentration low avoids the severe overreaction of 2 (Scheme 2b) under the optimal conditions identified by the BO-based approach. This is a plausible reason for the higher yield.

3. Conclusion

We optimized the conditions for the microflow synthesis of cyclic sulfamides 3 from linear substrates 1 and 2, both bearing multiple highly reactive sites, via sequential coupling/cyclization. Although the optimal conditions from the OVAT-based approach afforded the desired product 3 in 90% yield, no variable had simple linear correlations with the yields; thus, we were concerned that the real optimal conditions may not have been identified. Therefore, we re-optimized the reaction conditions using a BO-based approach and found that the desired product 3 could be obtained in 94% yield under the new conditions. Although the difference between the 2 optimal yields was only 4%, the accuracy and reproducibility of the results were confirmed by multiple independent experiments. Interestingly, almost all the optimized variables established in the 2 optimization approaches differed. We performed additional experiments and simulations based on the reaction rate analysis to understand why the yield under the optimal conditions established using the BO-based approach was higher than that using the OVAT-optimization approach. We conclude that the following 2 points are key for the higher yield: (i) the use of a larger excess of 2, which is continuously consumed in 2 undesired reactions (Scheme 2b and c), enables the complete consumption of 1; (ii) increasing the amount of 2 by increasing the flow rate of the solution of 2 while maintaining a low concentration, avoids super-stoichiometric attack of 2 on 1.

A fair comparison of the performance of OVAT- and BO-based approaches is not simple and is beyond the scope of this study; however, we believe that the use of a BO-based approach to optimize sequential microflow coupling and cyclization of multiple linear substrates containing multiple highly reactive sites is useful and can accelerate the optimization process. In addition, consideration of the optimal conditions identified by the BO-based approach can provide interesting insights into the reaction process and mechanism.

Supplementary data

Supplementary material is available at Bulletin of the Chemical Society of Japan online.

Funding

This work was supported by a Grant-in-Aid for Transformative Research Areas (A) from JSPS under Grant Number JP21A204 (no. 24H01072) and a Moonshot R&D Program from JST under Grant Number JPMJMS2236-10.

Data availability

Experimental procedures and spectroscopic data are available in Supplementary material.

References

Shinichiro Fuse

Shinichiro Fuse received his PhD degree from Tokyo Institute of Technology. He was a researcher at ChemGenesis Incorporated from 2005 to 2006, and a postdoctoral fellow at Harvard University from 2006 to 2008. In 2008, he joined the faculty at Tokyo Institute of Technology as an assistant professor. He moved to the Chemical Resources Laboratory at the same university as an associate professor in 2015. He was promoted to a full professor in the Graduate School of Pharmaceutical Sciences, Nagoya University in 2019. His research interests include microflow synthesis, automated synthesis, and machine-learning.

Author notes