Subscriber access provided by UNIVERSITY OF TOLEDO LIBRARIES
Deoxyfluorination with Sulfonyl Fluorides: Navigating Reaction Space with Machine Learning Matthew K. Nielsen, Derek T. Ahneman, Orestes Riera, and Abigail G. Doyle J. Am. Chem. Soc., Just Accepted Manuscript • DOI: 10.1021/jacs.8b01523 • Publication Date (Web): 27 Mar 2018 Downloaded from http://pubs.acs.org on March 27, 2018
Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.
is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.
Page 1 of 5 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of the American Chemical Society
Deoxyfluorination with Sulfonyl Fluorides: Navigating Reaction Space with Machine Learning Matthew K. Nielsen, Derek T. Ahneman, Orestes Riera, and Abigail G. Doyle* Department of Chemistry, Princeton University, Princeton, New Jersey 08544, United States Supporting Information Placeholder ABSTRACT: Through fine-tuning of reagent and base structure, sulfonyl fluorides can efficiently fluorinate diverse classes of alcohols. We show that machine learning can map the intricate reaction landscape and enable accurate prediction of highyielding conditions for untested substrates.
Due to the multidimensionality of chemical reactivity, vast resources are currently expended on the translation of reported synthetic methods to new chemical structures.1 An ongoing conundrum in methods development is the trade-off between reporting a single protocol for many substrates vs. developing highly optimized conditions tailored to individual substrates. Chemists tend to seek a single set of “optimal” conditions with the objective of rendering the method easily adoptable. However, if these conditions are low-yielding or fail to deliver product for a new substrate, a user has little recourse for advancement. Alternatively, chemists can report distinct sets of conditions for different substrate classes. In this case, new users may be dissuaded from adopting the method due to the complexity of interpreting multidimensional chemical data. Several strategies have been advanced to address this multitude of options, including the commercialization of broad spectrum catalyst mixtures,2 the development of high-throughput informer libraries,3 and the publication of reviews featuring “how-to” manuals.4 As an alternative, we questioned whether it would be possible to employ machine learning (ML) to analyze existing reaction data in its full complexity and then predict optimal conditions on a per substrate basis.5 This approach would enable researchers to evaluate the feasibility of a specific transformation and identify high-yielding conditions for a new substrate a priori. Recently, our laboratory, in collaboration with scientists at Merck Research Laboratories, demonstrated that ML algorithms can successfully model and predict the outcomes of Buchwald–Hartwig aminations based on a training set comprising thousands of reactions performed via ultra-high throughput experimentation (HTE).6 One challenge was the necessity of isolating standards to calibrate ultra-HTE yields, a time-consuming process when hundreds of products are in play. To streamline our workflow, we employed a Glorius additive screen,7 testing numerous isoxazole additives against a small subset of amination reactions. Following this initial study, we hoped to expand our evaluation of ML utility to data sets containing more structural diversity and mechanistic non-uniformity, representative of the challenges faced during adoption of most synthetic methods. In
this publication, we apply ML to the analysis of deoxyfluorination with sulfonyl fluorides. Since fluorination yields may be accurately determined from reaction mixtures by 19F NMR, we were able to rapidly screen structurally and stereochemically diverse alcohols. This study identifies sulfonyl fluorides as a reagent class with tunable and broad-spectrum reactivity. However, distinct combinations of sulfonyl fluorides and bases are necessary to obtain high yields across different alcohol classes. Despite the complexity of the reaction space, we show that a random forest ML algorithm trained on HTE deoxyfluorination data can be used to predict yield and high-yielding conditions for entirely new substrates. Deoxyfluorination transforms alcohols to aliphatic fluorides, which are high-value targets in pharmaceutical development.8 Traditionally, most deoxyfluorinations have been carried out with DAST or other sulfur(IV) reagents that suffer from thermal instability.9 Recently, the Ritter group has developed Phenofluor and other fluoroimidazolium compounds capable of fluorinating complex substrates with exquisite selectivity.10 Our group and others have identified sulfonyl fluorides such as PyFluor and perfluorobutanesulfonyl fluoride (PBSF) as noteworthy alternatives.11 In addition to being inexpensive and thermally stable, the sulfonyl fluoride motif is highly modular. Diverse sulfonyl fluorides may be synthesized by both classical and modern methods,12 enabling one to fine-tune the reagent to specific alcohols. Notwithstanding, there has been no systematic investigation into the substrate or reagent scope of deoxyfluorination with sulfonyl fluorides. In our previous report, we found that the combination of PyFluor and DBU was optimal for minimizing elimination in the preparation of acyclic secondary fluoride 2a (see Figure 1).11c Although this protocol translates well to most unactivated alcohols, PyFluor delivers poor yields for activated substrates that are susceptible to nucleophilic attack by DBU, such as benzylic alcohol 1b. Conversely, strained cyclic alcohols (1c and 1d) are largely unreactive. To better understand sulfonyl fluoride reactivity, we screened alcohols 1a – 1d against a selection of sulfonyl fluorides and bases (Figure 1A). We evaluated the commercial reagents PyFluor and PBSF as well as the arylsulfonyl fluorides 3-Cl, 3-CF3, and 3-NO2, each of which is readily obtained from the commercial sulfonyl chloride. Based on experimentally determined relative rate constants, these sulfonyl fluorides span a broad range of reactivity that correlates with the leaving group ability of the corresponding sulfonate ester (Figures S1–S3). Reactions with PBSF are typically complete within seconds, whereas arylsulfonyl fluorides may require hours or even days to reach full conversion. Likewise, the
ACS Paragon Plus Environment
Journal of the American Chemical Society 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
sulfonyl fluoride (1.1 equiv) base (1.5 equiv)
OH
O O S R O
THF (0.4 M) room temperature, 48 h
1
2
alcohols
A
OH
Ph
Ph
N
BnO
1a
OH
OH
OH
Me
F
Boc
1b
CO2Me 1d
1c
sulfonyl fluorides SO2F N
Cl
SO2F
SO2F F3C
SO2F n-C4F9
O2N
SO2F
3-Cl
PyFluor
3-CF3
3-NO2
PBSF
1
1.7
4.9
15
44,000
krel: N
Me N
N N DBU
N MTBD
bases N Me
N Me
t-Bu Me N Me
BTMG
t-Bu N N P N N BTPP
Figure 1. Multi-dimensional screening of alcohols 1a – 1d. The 19F NMR yield for each combination of sulfonyl fluoride and base is displayed in the accompanying heat map. See SI for krel data.
bases feature steric diversity, ranging from the compact amidine DBU to the bulky phosphazene BTPP. As shown in Figure 1, alcohols 1a – 1d displayed very distinct reactivity profiles. Secondary fluoride 2a and benzylic fluoride 2b were obtained in highest yield with 3-CF3, whereas cyclic fluorides 2c and 2d were favored by the more reactive PBSF. The reaction surfaces of 2a and 2d show a complex interplay between base and sulfonyl fluoride, with relatively unencumbered MTBD performing best. In contrast, bulky BTPP proved optimal for benzylic alcohol 1b. Finally, cyclobutanol 1c displayed almost no reactivity except with PBSF, for which yields were largely independent of base identity. These results hint at a complex reaction space that defies simple axiomatic interpretation. As such, we expected that this method would be an ideal candidate for reaction modeling via ML. To develop a suitable training set, we subjected a diverse range of alcohols 1a – 1af to the same twenty-reaction screen as shown in Figure 1, resulting in a total of 640 distinct reactions. The highest yielding conditions for each substrate are summarized in Figure 2.13 Unactivated primary alcohols (Figure 2A) were fluorinated in as high as 96% yield with arylsulfonyl fluorides and non-nucleophilic bases such as BTMG and BTPP (2e, 2g, 2i, 2j). For acyclic secondary alcohols, we were unable to identify superior conditions to those previously disclosed with PyFluor (2a, 2h).11c Timepoint studies indicated that second-order kinetics
Page 2 of 5
are operative in the fluorination of both primary and secondary alcohols, consistent with a rate-limiting SN2 substitution between fluoride and the intermediate sulfonate ester. In contrast, tertiary alcohols required the exceptionally reactive perfluorobutanesulfonate leaving group and displayed distinct first-order kinetics indicative of a unimolecular SN1 pathway (2f). In the case of α-amino alcohol 1j, we suspected that the aniline might engage in neighboring group participation. In support of our hypothesis, piperidine 1k reacted to form both 2k and the rearranged azepane 2k′, presumably via the intermediacy of an aziridinium ion.14 Cyclic substrates (Figure 2B) typically have higher energy transition states arising from deformation of the ideal SN2 trigonal bipyramidal geometry. Indeed, we found that cyclic alcohols 1c, 1d, and 1l – 1q were fluorinated in up to 94% yield with highly activated PBSF. Full stereochemical inversion was observed including for hydroxyproline diastereomers 2c and 2l as well as cyclohexanol diastereomers 2o and 2p. In contrast to 4and 5-membered cyclic alcohols, cyclohexanols performed poorly due to the absence of ring strain in the ground state, which raises the ΔG‡ for substitution.15 Cyclohexanol 1o, for which the alcohol is locked in an equatorial conformation, delivered 2o in only modest yield due to the highly congested backside trajectory. However, axial diastereomer 1p afforded almost exclusively E2 elimination owing to the constrained anti-periplanar geometry. Benzylic fluorides (Figure 2C) were obtained in as high as 90% yield by employing bulky non-nucleophilic bases (2b, 2r – 2v). Fluoride 2v exhibited retention of configuration, consistent with selective ring-opening of an aziridinium intermediate.16 Primary allylic fluoride 2w was formed with high selectivity (12:1) against the branched SN2′ product; however, the secondary fluoride 2x exhibited only 4:1 rr, and the tertiary fluoride 2y was dominated in 1.5:1 rr by the linear fluoride. Homobenzylic alcohols 1z and 1aa (Figure 2D) afforded high yields, but may form via an arenium ion. The analogous competing homoallylic rearrangement resulted in the low yield of 2ab.17 The α-hydroxycarbonyl compounds 1ac and 1ad performed reasonably well, but β-hydroxycarbonyl substrates afforded exclusively α,β-unsaturated products. A rare exception, 2ae was formed in 24% yield with BTPP due to shielding of the α carbon by the bulky trityl group. Fluorination of hemiacetal 1af likely involves an oxocarbenium intermediate as evidenced by the dependency of diastereoselectivity on the sterics of the base, which serves as the fluoride counterion. In general, activated substrates perform optimally with comparatively electron-rich sulfonyl fluorides that form stable sulfonate esters, thus minimizing elimination and undesired side reactions, whereas deactivated cyclic and tertiary alcohols react best with the highly reactive PBSF. Unencumbered primary alcohols require the bulky bases BTMG or BTPP, while congested secondary alcohols are frequently higher-yielding with the more compact bases. Notwithstanding, these observations fail to encompass the full complexity of the enclosed reaction space. In this regard, ML is the ideal tool for capturing such intricacies. Previously, we demonstrated that the random forest ML algorithm, which is specifically designed to avoid overfitting by averaging the output of many randomly generated decision trees,18 affords superior predictive ability over a range of ML algorithms. In order to train a random forest model for sulfonyl fluoride deoxyfluorination, we assembled a table of descriptors for the alcohols, sulfonyl fluorides, and bases
ACS Paragon Plus Environment
13
Page 3 of 5 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of the American Chemical Society A
C
Unactivated Alcohols F
F
Ph
F
Ph
2e 89% yield
(±)-2a a 79% yield
2f 18% yield
3-CF3 BTPP 24 h
PyFluor DBU 48 h
PBSF BTPP 12 h
O
Ph
F N
N
N
Ph
N Me
2g 96% yield
O
3-CF3 BTPP 1h
3-CF3 BTPP 3h
N
Me
N
N
N
3-CF3 BTMG 24 h
(±)-1k
B
Me
N
N
(±)-2x b,e 39% yield
3-CF3 BTPP 1h
3-CF3 BTPP 24 h
BnO 2c c,d 84% yield PBSF BTMG 30 min
O
AcO
H
H
F
O
2n c,d 94% yield PBSF BTPP 30 min
CO2Me
Boc
2d c,d 66% yield
O
O
N
F
O
PBSF BTMG 30 min
N
Me
O
N
O
F
Boc N
O
PBSF BTPP 30 min
F H (±)-2q c,d 53% yield PBSF BTMG 30 min
Me H
3-NO2 BTMG 24 h
O
PBSF MTBD 30 min
PBSF BTMG 30 min
Hemiacetal:
O Me
F
HO Me
O
Me F
H H
PBSF MTBD 30 min
F
F TrtHN
CO2Me
O (±)-2ac 78% yield
H
2ab b,f,g 13% yield
α/ β-Hydroxy carbonyl:
O
H
F 2aa b 88% yield
H
2p b,c,d 15% yield
F
N
Cl
PBSF MTBD 12 h
O
Me
NO2
2z 86% yield
Me Me
N
2o c,d 46% yield
PBSF BTPP 48 h
Me
MeO
2m c,d 88% yield
PBSF MTBD 30 min
F
O
CO2Me
2l c,d 75% yield
PBSF MTBD 3h
(±)-2y b,e 9% yield
F
O
N
Boc
Me
Additional Substrate Classes
N
O N
Me
2w b,e 46% yield
Ph
(±)-2k' 54% yield
Me Me
F
F
Me
Homobenzylic/homoallylic:
Ph
(±)-2k 32% yield
F
Me
F F
Cyclic Alcohols F
PBSF BTPP 3h
Me
D
F F PBSF BTPP 24 h
Me 2v d,f 90% yield
(±)-2u 60% yield 3-CF3 BTMG 24 h
Me
3-NO2 BTPP 48 h
N
Ph
NO2
O O
F
Me
Allylic: Me
2j 82% yield
Ph
3-NO2 BTMG 1h
MeO
3-Cl BTPP 1h
F
N
2i 74% yield
OH
2s 41% yield
MeO
F
2t e 57% yield
Cl
NO2
OBz OBz
N
N Me
Me
N
N
O
2r 85% yield
F N
F
NBz2 F
2b 90% yield
PyFluor DBU 48 h
N
S O O
Me
(±)-2h a,b 81% yield
3-CF3 BTMG 24 h
N
Ph
Me
N N N Trt N
F
F
F
O
Me N
Cl N
Me
Me
Me Me
Ph
Me
Activated Alcohols
Benzylic:
Me
H
O O
O
O H
Me 2ad b 47% yield
2ae b 24% yield
3-CF3 BTPP 3h
PBSF BTPP 3h
2af b,c,e,g 54% yield 3-NO2 BTMG 30 min
Figure 2. Scope of deoxyfluorination. The reagents used for each alcohol are listed below the yield. Reactions were performed on 0.5 – 1.0 b mmol scale with 1.1 equiv sulfonyl fluoride and 1.1 – 2 equiv base in THF (0.1 – 0.5 M) at room temperature. a Solvent: toluene (1 M). c d e 19 f g Co-isolated with side product (see SI). Inversion observed. >20:1 diastereoselectivity. F NMR yield. Retention observed. 4:1 dr.
employed in each of the 640 screening reactions. These descriptors included computed atomic and molecular properties (ie. electrostatic charge, NMR shift) as well as binary categorical identifiers (ie. primary, secondary, tertiary, cyclic). Using R statistical software,13 we applied a random forest algorithm to a training set containing 70% of the screening entries, and evaluated the model using a test set comprising the remaining 192 reactions. Remarkably, the trained model predicted the yields of the test set reactions with a root-mean-square error (RMSE) of only 7.4% yield (Figure 3A). In comparison to our previous work, this training set was smaller (448 vs. 2898 training reactions), encompassed much broader substrate diversity, and incorporated multiple mechanisms (SN1, SN2, anchimeric assistance), highlighting the robust predictive ability of the random forest algorithm.
As further validation, we evaluated substrates 1ag – 1ak which do not appear in the training set (Figure 3B). Again, our model successfully predicted the yields of 2ag – 2ai with reasonable accuracy (8 – 13% RMSE). Substantially higher variation (~20% RMSE) was observed in the case of 2aj and 2ak which contain distinct functionality not found in the training set; however, the error was largely systematic, indicating that the algorithm successfully identified underlying trends. Notably, for each substrate, the experimental yield corresponding to the predicted optimal conditions lay within 4% of the highestyielding conditions. This level of precision is more than sufficient for enabling new users to evaluate reaction feasibility and select initial reaction conditions. For example, although experimental yields for 2ai ranged from 16 – 97%, the predicted optimal conditions afforded 93% yield, only 4%
ACS Paragon Plus Environment
Journal of the American Chemical Society 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Page 4 of 5
ASSOCIATED CONTENT
Supporting Information Experimental procedures, additional reaction optimization, and spectroscopic data for all new compounds. This material is available free of charge via the Internet at http://pubs.acs.org.
AUTHOR INFORMATION
Corresponding Author
[email protected] Notes The authors declare no competing financial interests.
ACKNOWLEDGMENT Financial support was generously provided by NSF (CHE1565983), Janssen, and Celgene Corporation.
REFERENCES
B
N N N
F
Cl
F
N
2ag Predicted: 87% yield Observed: 85% yield
2ah Predicted: 75% yield Observed: 62% yield
3-CF3 BTMG
PBSF MTBD
RMSE: 8.4%
RMSE: 11%
Cl
Cl
Me F
O N 2ai Predicted: 82% yield Observed: 93% yield
N
F
F PhO
F Me
Me
(±)-2aj Predicted: 62% yield Observed: 86% yield
(±)-2ak Predicted: 65% yield Observed: 82% yield
3-CF3 BTPP
PBSF BTPP
PBSF BTPP
RMSE: 13%
RMSE: 20%
RMSE: 23%
Figure 3. Performance of random forest model. (A) Plot of observed vs. predicted yields for the test set and validation substrates 1ag – 1ak (see legend). The highest predicted and highest observed yields are circled. (B) Comparison of predicted and observed yields for the predicted optimal conditions of 1ag – 1ak.
below the best experimental conditions. On the other hand, the best conditions for 2ai, which delivered 97% yield, were predicted to provide only 65% yield, suggesting that expanded training sets along with advances in the description and modeling of the chemical space are necessary to use ML for predicting the best conditions. In conclusion, we have demonstrated that the flexibility of the sulfonyl fluoride scaffold allows deoxyfluorination of a broad range of alcohols, significantly expanding upon the scope initially reported with PyFluor. Moreover, we show that a random forest model can accurately predict reaction outcomes and aid in identifying optimal conditions for new alcohols. We envision that this tool will enable users to navigate the multitude of distinct deoxyfluorination conditions, and that similar models could prove beneficial in the development of other synthetic methods. Continual expansion of our training set with new substrates and additional variables (ie. stoichiometry, concentration, solvent, and temperature) will lead to more accurate and comprehensive coverage of the sulfonyl fluoride deoxyfluorination reaction space.
(1) Santanilla, A. B.; 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. Science 2015, 347, 49. (2) Fors, B. P.; Buchwald, S. L. J. Am. Chem. Soc. 2010, 132, 15914. (3) Kutchukian, P. S.; Dropinski, J. F.; Dykstra, K. D.; Li, B.; DiRocco, D. A.; Streckfuss, E. C.; Campeau, L.-C.; Cernak, T.; Vachal, P.; Davies, I. W.; Krska, S. W.; Dreher, S. D. Chem. Sci. 2016, 7, 2604. (4) Chatterjee, A. K.; Choi, T.-L.; Sanders, D. P.; Grubbs, R. H. J. Am. Chem. Soc. 2003, 125, 11360. (5) The Sigman lab has introduced a data-intensive strategy to predict selectivity and gain mechanistic insight in catalytic reactions using linear regression analysis. For a leading reference, see: Milo, A.; Neel, A. J.; Toste, F. D.; Sigman, M. S. Science 2015, 347, 737. (6) Ahneman, D. T.; Estrada, J. G. ; Lin, S.; Dreher, S. D.; Doyle, A. G. Science 2018, DOI: 10.1126/science.aar5169. (7) Collins, K. D.; Glorius, F. Acc. Chem. Res. 2015, 48, 619. (8) Yerien, D. E.; Bonesi, S.; Postigo, A. Org. Biomol. Chem. 2016, 14, 8398. (9) (a) Middleton, W. J. J. Org. Chem. 1975, 40, 574. (b) Lal, G. S.; Pez, G. P., Pesaresi, R. J.; Prozonic, F. M.; Cheng, H. J. Org. Chem. 1999, 64, 7048. (c) Beaulieu, F.; Beauregard, L.-P.; Courchesne, G.; Couturier, M.; LaFlamme, F.; L’Heureux, A. Org. Lett. 2009, 11, 5050. (d) Umemoto, T.; Singh, R. P.; Xu, Y.; Saito, N. J. Am. Chem. Soc. 2010, 132, 18199. (e) Messina, P. A.; Mange, K. C.; Middleton, W. J. J. Fluorine Chem. 1989, 42, 137. (10) (a) Sladojevich, F.; Arlow, S. I.; Pingping, T.; Ritter, T. J. Am. Chem. Soc. 2013, 135, 2470. (b) Goldberg, N. W.; Shen, X.; Li, J.; Ritter, T. Org. Lett. 2016, 18, 6102. (11) (a) Shimizu, M.; Nakahara, Y.; Yoshioka, H. Tetrahedron Lett. 1985, 26, 4207. (b) Bennua-Shalmowski, B.; Vorbrüggen, H. Tetrahedron Lett. 1995, 36, 2611. (c) Nielsen, M. K., Ugaz, C. R.; Li, W.; Doyle, A. G. J. Am. Chem. Soc. 2015, 137, 9571. (12) (a) Dong, J.; Krasnova, L.; Finn, M. G.; Sharpless, B. K. Angew. Chem., Int. Ed. 2014, 53, 9430. (b) Davies, A. T.; Curto, J. M.; Bagley, S. W.; Willis, M. C. Chem. Sci. 2017, 8, 1233. (c) Tribby, A. L.; Rodríguez, I.; Shariffudin, S.; Ball, N. D. J. Org. Chem. 2017, 82, 2294. (13) See Supporting Information for full data set and documentation. (14) Déchamps, I.; Pardo, D. G.; Cossy, J. Eur. J. Org. Chem.. 2007, 4224. (15) Mange, K. C.; Middleton, W. J. J. Fluorine Chem. 1989, 43, 405. (16) Hamman, S.; Beguin, C. G.; Charlon, C.; Luu-Duc, C. J. Fluorine Chem. 1987, 37, 343. (17) Néder, Á.; Uskert, A.; Nagy, É.; Méhesfalvi, Z.; Kuszmann, J. Acta Chim. Acad. Sci. Hung. 1980, 103, 231. (18) (a) Mitchell, J. B. Wiley Interdiscip. Rev.: Comput. Mol. Sci. 2014, 4, 468. (b) Kleinberg, E. M. Ann. Statist. 1996, 24, 2319. (c) Breiman,
4
ACS Paragon Plus Environment
Page 5 of 5 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of the American Chemical Society L. Mach. Learn. 2001, 45, 5.
TOC graphic:
5
ACS Paragon Plus Environment
