Elucidating Structures of Complex Organic Compounds Using a Machine Learning Model Based on the ¹³C NMR Chemical Shifts

Abstract

We present a protocol that combines the support vector machine (SVM) model with accurate 13C chemical shift calculations at the xOPBE/6-311+G(2d,p) level of theory, denoted as SVM-M (i.e., SVM for magnetic property). We show here that this SVM-M protocol is a versatile tool for identifying the structural and stereochemical assignment of complex organic compounds with high confidence. Of particular significance is that, by utilizing the dual role of the decision values in SVM, the present SVM-M protocol provides an accurate yet efficient solution to simultaneously handle the classification issue (i.e., “is a given structure correct or incorrect?”) and the comparison-based problem (i.e., “which structure is more likely to be correct or wrong among several candidate structures?”). A significantly high success rate has been reached (i.e., ∼100% on a set of 760 sample molecules with 15928 13C chemical shifts), which makes the SVM-M protocol a powerful tool for routine applications in structural and stereochemical assignments, as well as in detecting mis-assignments, for complex organic compounds, including natural products.

Accurate Structural Elucidation of Samoquasine A and an Unknown Homologue Using a Computation-Based Machine Learning Protocol

“…Once the chemical shifts were computed, four statistical parameters (standard deviation (SD), mean absolute deviation (MAD), unsigned maximum deviation (MAX), and rootmean-squre-deviation (RMSD)) were evaluated and fed into the SVM-M model to obtain the decision value. As previously established, 18 the sign of decision value serves as a critical indicator of whether a given structure aligns potentially with the experimental NMR 13 C spectra, effectively classifying a given structure as correct with a positive decision value and incorrect with a negative decision value. Furthermore, the magnitude of the decision value reflects the goodness of the classification.…”

“…13−16 In the present work, the four-step hierarchical conformational analysis adopted in the SVM-M protocol was employed. 18 Initially, the conformers generated by RDKIT 19,20 were optimized at the PM7 21 For the calculations of the NMR shielding constants, the recently developed xOPBE 17 functional was employed in connection with the GIAO method. 23 The 6-311 + G(2d,p) 24 basis set was used for all atoms, and a PCM model 25 using chloroform as the solvent was employed to account for the solvent effect for all calculations of chemical shieldings in this work.…”

“…For accurate calculations of 13 C chemical shifts, it has been well recognized that conformational analysis is essential. − In the present work, the four-step hierarchical conformational analysis adopted in the SVM-M protocol was employed . Initially, the conformers generated by RDKIT , were optimized at the PM7 level of theory, and all conformers within 50.0 kJ/mol of the global minimum were subjected to further reoptimization at the ωB97X-D/3-21G level of theory.…”

“…17,18 This SVM-M protocol has been shown to be a powerful tool for identifying the structural and stereochemical assignment of natural product with high confidence. 18 Herein, we applied the SVM-M protocol to reevaluate the structural identification of samoquasine A (perlolidine, 2), as well as compounds 1, 3, and 4 by utilizing the dual roles of decision values in the SVM method. Moreover, we utilized the SVM-M protocol to unmask the structure of the heretofore unknown natural product X.…”

Study of heavy atom influence on poly-halogenated compounds using DP4/MM-DP4+/DP4+: insights and trends

“…For example, CP3 6 and MAE ΔΔ δ 7 are excellent strategies to assign groups of stereoisomers from which the NMR data are known. On the other hand, ANN-PRA, 8,9 SVM-M, 10 and DP5 11 are useful tools of structural validation. Nevertheless, the most popular and widely employed approaches are based on the correlation between one set of experimental data (a common situation in the natural products area where only one isomer out of many possible is isolated) with the calculated values of two or more plausible candidates selected in advance.…”