Reinvestigation of a robotically revealed reaction

Abstract

In a recent paper in this journal, Cronin et al. created an organic synthesis robot controlled by machine learning1. The robot was used to discover four new reactions. One such product (compound 1 in Fig. 1; compound 22 in ref. 1) was claimed on the basis of spectral data that does not support the proposed structure. We used the original spectroscopic data to identify an alternative connectivity for the product molecule, and then repeated the reported reaction. We found that two diastereomers were produced (rather than the single diastereomer that had originally been described), and confirmed through several lines of enquiry that our revised structures were the correct products. We were intrigued by the Cronin group’s report1 of robot-guided discovery, and were particularly attracted to the possibility of using molecules such as compound 1 in biological screening experiments. However, on noting inconsistencies in the product characterization, we calculated the expected nuclear magnetic resonance (NMR) shifts for the minimum-energy diastereomer of 1 (1a; Extended Data Fig. 1) using both a parameterized method (MestReNova) and density functional theory (DFT) calculation (at the B3LYP/6-31G* level of theory). The results (Extended Data Table 1) did not support the assigned product. A more plausible interpretation of the NMR data would assign the signal at 169 p.p.m. to an ester carbonyl, and the signal at 124 p.p.m. to the β-position of a vinyl ether or vinyl ester. This agrees almost precisely with the known spectral assignments for comparable enol acetates5 (Extended Data Fig. 1). Taking into account mechanistic possibilities as well as 2D-NMR (heteronuclear multiple bond correlation) data published by Cronin et al.1, the most reasonable product is therefore the constitutional isomer 2 (Fig. 1). As illustrated by our proposed mechanism (Extended Data Fig. 2), C-acylation of the enolate intermediate D would give rise to compound 1, whereas O-acylation would give 2. In addition to providing less steric congestion, O-acylation would afford a greater degree of electronic conjugation in the product. Repetition of the reaction in our own laboratory led to an approximately 1:1 mixture of diastereomers (Extended Data Fig. 3). Separation of these two products and full characterization established that they were products 2a and 2b. Spectroscopic data for 2a matched that provided by Cronin et al.1 for their isolated product. Possessing both an acid-sensitive vinyl ether linkage and a basesensitive activated ester, 2a and 2b would be expected to have poor stability. We found that 2b was particularly unstable, and a low yield of the product was obtained during chromatographic separation from the diastereomeric mixture. This presumably explains why Cronin et al. observed a single compound from their reaction1; 2b probably decomposed before isolation. To further confirm the presence of the ester group in the products, diastereomer 2a was subjected to saponification. As expected, the hydrolysis product was clearly observed by NMR spectroscopy, and was characterized by accurate mass analysis. This product is even less stable than 2b, however, and could not be isolated in pure form. We also attempted to isomerize 2a to produce genuine compound 1, but this was unsuccessful.