Finding furfural hydrogenation catalysts via predictive modelling

Article abstract of DOI:10.1002/adsc.201000308

We combine multicomponent reactions, catalytic performance studies and predictive modelling to find transfer hydrogenation catalysts. An initial set of 18 ruthenium-carbene complexes were synthesized and screened in the transfer hydrogenation of furfural to furfurol with isopropyl alcohol complexes gave varied yields, from 62% up to >99.9%, with no obvious structure/activity correlations. Control experiments proved that the carbene ligand remains coordinated to the ruthenium centre throughout the reaction. Deuterium-labelling studies showed a secondary isotope effect (k_H:k_D = 1.5). Further mechanistic studies showed that this transfer hydrogenation follows the so-called monohydride pathway. Using these data, we built a predictive model for 13 of the catalysts, based on 2D and 3D molecular descriptors. We tested and validated the model using the remaining five catalysts (cross-validation, R² = 0.913). Then, with this model, the conversion and selectivity were predicted for four completely new ruthenium-carbene complexes. These four catalysts were then synthesized and tested. The results were within 3% of the model's predictions, demonstrating the validity and value of predictive modelling in catalyst optimization.

Full text of DOI:10.1002/adsc.201000308

FULL PAPERS
DOI: 10.1002/adsc.201000308
Finding Furfural Hydrogenation Catalysts via Predictive
Modelling
Zea Strassberger,a Maurice Mooijman,b Eelco Ruijter,b Albert H. Alberts,a
Ana G. Maldonado,a Romano V. A. Orru,b,* and Gadi Rothenberga,*
a
Van ꢀt Hoff Institute of Molecular Sciences, University of Amsterdam, Science Park 904, 1098XH Amsterdam, The
Netherlands
Fax : (+31)-(0)20-525-5604; phone: (+31)-(0)20-525-6963; e-mail: g.rothenberg@uva.nl
Department of Chemistry & Pharmaceutical Sciences, Vrije Universiteit Amsterdam, De Boelelaan 1083, 1081 HV
b
Amsterdam, The Netherlands
Fax : (+31)-(0)-20-598-7488; phone: (+31)-(0)20-598-7447; e-mail: orru@few.vu.nl
Received: April 21, 2010; Revised: July 16, 2010; Published online: September 2, 2010
Supporting information for this article is available on the WWW under
http://dx.doi.org/10.1002/adsc.201000308.
Abstract: We combine multicomponent reactions, of the catalysts, based on 2D and 3D molecular de-
catalytic performance studies and predictive model- scriptors. We tested and validated the model using
ling to find transfer hydrogenation catalysts. An ini- the remaining five catalysts (cross-validation, R2 =
tial set of 18 ruthenium-carbene complexes were syn- 0.913). Then, with this model, the conversion and se-
thesized and screened in the transfer hydrogenation lectivity were predicted for four completely new
of furfural to furfurol with isopropyl alcohol com- ruthenium-carbene complexes. These four catalysts
plexes gave varied yields, from 62% up to >99.9%, were then synthesized and tested. The results were
with no obvious structure/activity correlations. Con- within 3% of the modelꢁs predictions, demonstrating
trol experiments proved that the carbene ligand re- the validity and value of predictive modelling in cat-
mains coordinated to the ruthenium centre through- alyst optimization.
out the reaction. Deuterium-labelling studies showed
a secondary isotope effect (kH:kD =1.5). Further
mechanistic studies showed that this transfer hydro- Keywords: carbenes; descriptor modelling; furfural;
genation follows the so-called monohydride pathway. multicomponent reactions; QSAR; quantitative
Using these data, we built a predictive model for 13 structure-activity relationship
Introduction
realistic. Yet, one way to move towards this goal is by
teaching “catalytic intuition” to a computer,[5] and
Although it is one of the most studied processes in harnessing this computational power to help solve
the history of chemistry, catalytic hydrogenation is problems in catalytic hydrogenation. As we recently
still full of surprises.[1] Changing the catalyst, by slight- showed,[6] predictive modelling can select active re-
ly altering a ligand or switching to a different metal gions in the catalyst space, provided that the following
precursor, often wreaks havoc in the results. Indeed, three conditions are met: first, you need sufficient ex-
it seems that the more knowledge that is gathered on perimental data for building a predictive model;
hydrogenation reactions, the more pathways and op- second, you should have acces to a large number of
portunities emerge for hydrogenation catalysts.[2,3] diverse ligand-metal complexes; and, finally, a robust
Polymers, arenes and heteroaromatics are just three model validation procedure must be available. The
recent examples.[4]
computer can then “synthesize” sets of virtual cata-
The problem is that whilst research on homogene- lysts (see flowchart in Figure 1), and predict their
ous catalytic hydrogenation has provided us with ef- characteristics (descriptor values) and performance
fective solutions to specific reactions, there is no (figures of merit). These predictions can then be vali-
“grand unified theory” for finding good hydrogena- dated experimentally, closing the cycle.
tion catalysts. Indeed, such a panacea is, perhaps, un-
Adv. Synth. Catal. 2010, 352, 2201 – 2210
ꢂ 2010 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
2201

Zea Strassberger et al.
FULL PAPERS
The most fallible step in studies of this kind is, in
fact, the experimental validation. Generating large
virtual libraries in silico is relatively easy, but restrict-
ing these according to explicit synthetic rules often
robs them of their diversity. To avoid this pitfall, we
turned to making our catalysts via multicomponent
synthesis protocols. Multicomponent reactions
(MCRs) are convergent procedures that combine at
least three simple, easily accessible building blocks in
a one-pot process, thus powerful reactions for gener-
ating functionalized complex molecules.[7–15] MCRs
proceed with remarkably high atom and step econo-
my by reducing the number of functional group ma-
nipulations, thus avoiding the use of protective groups
and, as such, are ideally suited for the generation and
validation of catalyst libraries. Previously, we have
presented the scope and application of MCRs for syn-
thesizing 2H-2-imidazolines[14,15] which are key inter-
mediates en route to N-heterocyclic carbene (NHC)
complexes,[14] some of todayꢁs most promising hydro-
genation catalysts.[16] In this MCR (Scheme 1), an al-
dehyde or ketone I, a primary amine II and an a-
acidic isocyanide III are combined in a one-pot reac-
tion, giving the corresponding 2H-2-imidazoline. The
Figure 1. Flowchart of iterative modelling and experiments
cycle.
Scheme 1. Multicomponent synthesis-alkylation-complexation route to N-heterocyclic carbenes. Using the 29 building blocks
shown here, you can make 1,152 different ligands.
2202
asc.wiley-vch.de
ꢂ 2010 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Adv. Synth. Catal. 2010, 352, 2201 – 2210

Finding Furfural Hydrogenation Catalysts via Predictive Modelling
product is then easily alkylated with an alkyl halide was added, together with one equivalent of potassium
IV at position N-3, yielding the final NHC precursor. hydroxide, activating the catalyst. Then, twenty-five
In this paper, we demonstrate the utility of combin- equivalents of furfural were added and the mixture
ing multicomponent ligand synthesis and catalytic per- was heated to 608C and stirred under nitrogen for
formance studies with predictive modelling for cata- 24 h. Reaction progress was monitored by GC and
lyst discovery and optimization. Our case study focus- 1H NMR (detailed procedures are given in the Exper-
es on catalysts for hydrogenating furfural 1 to furfurol imental Section and the Supporting Information).
2 [Eq. (1)]. This is an industrially important reaction,
Table 1 shows the substrate conversion, product se-
lectivity and product yield. In the absence of any car-
bene ligand, the metal precursor [RuCl2ACTHUNTRGNEU(GN p-cymene)]2
already gives 94% conversion, with 84% yield of fur-
furol after 24 h. Nevertheless, adding a carbene ligand
can raise both numbers to >99.9%. Control experi-
ments showed that the hydrogen transfer reaction is
sensitive to the type and amount of base used. The
best conversions were obtained with a mix of potassi-
um tert-butoxide for deprotonating the imidazolinium
as furfural is readily obtained from cellulosic bio- precursor, and KOH as promoter.[22] No reaction oc-
mass,[17,18] and its hydrogenation is a key pathway in curred in the absence of KOH, and <80% yield was
bio-refining.[19] Moreover, the presence of both a het- observed when using 0.1 equivalents of KOH with re-
eroaromatic ring and a carbonyl group makes furfural spect to the Ru catalyst.
a versatile study subject, from which one can draw
One question that immediately springs to mind in
analogies to many similar compounds. Recently, we such systems pertains to the ligand dissociation equili-
showed the first proof of concept for using this MCR bria: what is the chance that, in the course of the cat-
in the synthesis of transfer hydrogenation catalysts.[20] alytic cycle, the carbene ligand(s) dissociate from the
Ru complex? However, our carbene ligands remain
coordinated throughout the cycle, since any free car-
Results and Discussion
bene in solution would lead to immediate acyloin con-
densation, analogous to the reactions reported by
The multicomponent synthesis gives us access to a Enders et al.[23] Control experiments showed that the
broad range of imidazolinium precursors. In fact, the acyloin condensation [Eq. (2), giving roughly 90%
range is already too broad for traditional synthesis
and screening studies. Using just the 29 building
blocks from Scheme 1 (all of which are commercially
available) already gives 1,152 different ligand precur-
sors. In theory, we could synthesize and test all of
them. In practice, having a model that can pinpoint
the active regions in this catalyst space is much more
efficient and convenient. We want to minimize the
number of experiments, without compromising the
search for good catalysts. This is where the iterative
cycle of experiments, descriptor modelling, and vali-
dation comes in.
We prepared a set of 18 imidazolinium salts follow-
ing the synthetic pathway shown in Scheme 1 (see Ex-
perimental Section for details). These ligand precur- furoin 3a and 10% furyl 3b] is indeed very fast com-
sors (structures 4–21 in Scheme 2) were isolated and pared to the transfer hydrogenation reaction, with t1/2
purified prior to the catalysis screening tests. Then, in <15 min.[24] Thus, by using a 1:1 ligand:Ru ratio, we
a typical reaction [Eq. (1) above], an equivalent managed to avoid any acyloin condensation, confirm-
amount of the imidazolinium precursor was deproton- ing that the carbene ligand remains coordinated to
ated with potassium tert-butoxide for 30 min at 408C, the ruthenium complex throughout the catalytic cycle.
and the resulting carbene was coordinated in situ to
Figure 2 shows time-resolved reaction profiles for
the dimer Ru salt precursor, [RuCl2A(p-cymene)]2. For- the transfer hydrogenation reaction catalyzed by the
CTHUNGTRENNUNG
mation of this complex was shown clearly by 18 different ruthenium-carbene complexes made from
13C NMR.[21] Both reactions were complete within the imidazolinium salts in Scheme 2, where Ln de-
30 min. Subsequently, an excess of isopropyl alcohol, notes the ligand prepared using precursor n. For com-
used as both solvent and sacrificial hydrogen donor, pleteness, this figure also shows the profiles obtained
Adv. Synth. Catal. 2010, 352, 2201 – 2210
ꢂ 2010 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
asc.wiley-vch.de
2203

Zea Strassberger et al.
FULL PAPERS
Scheme 2. Chemical structures for the 18 ligands synthesized in the first iteration.
with the catalysts synthesized later, following the ferred only to the carbonyl carbon. Conversely, in the
ꢀ
ꢀ
model predictions (L22–L25) The catalysts are divid- dihydride pathway, the C H and O H from the hy-
ed in two groups (left and right graphs) for clarity. drogen donor are equivalent with respect to the hy-
The yield obtained without carbene ligand (i.e., just drogen transfer. Since hydroxide protons exchange
using the [RuCl2ACHTUNGTRENNUNG(p-cymene)]2 precursor) is shown as a rapidly with solvent protons, and since the transfer
continuous curve. We see that several of the Ru com- hydrogenation is reversible, the incorporation of la-
plexes, and especially those derived from imidazolini- belled hydrogen in the product OH group will not
um salts 4, 9 and 19, exhibit both fast and highly se- exceed 50% (Scheme 3, bottom). In our case, a series
lective catalysis.
of
experiments
using
isotopically
labelled
Delving deeper into the reaction mechanism, we (CH3)2CD(OH) gave a rate constant ratio of kH:kD =
tried to determine a pathway for the transfer hydro- 1.5. This value implies a secondary deuterium isotope
genation process.[25] Ruthenium-catalyzed hydrogen effect, corresponding to a route where the scission of
ꢀ
transfer typically involves either a monohydride or a the isopropyl alcohol C H bond is not the rate-deter-
dihydride species in the catalytic cycle. The major dif- mining step. Moreover, following the reactions with
ference lies in the location of the hydride in the hy- 1H NMR showed an 80:20 incorporation ratio of deu-
drogenated product, as shown by the elegant isotope terium on the carbon and the oxygen, respectively.
substitution experiments of Bꢃckvallꢁs group[26] and This implies at least partial participation of the mono-
Enthaler and co-workers.[22] If the catalytic cycle fol- hydride pathway.[27]
lows the monohydride pathway (Scheme 3, top), the
These experimental results are ideal for modelling
hydrogen bound to the carbinol carbon donor is trans- purposes, since the data show a wide range and
2204
asc.wiley-vch.de
ꢂ 2010 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Adv. Synth. Catal. 2010, 352, 2201 – 2210

Finding Furfural Hydrogenation Catalysts via Predictive Modelling
Table 1. Catalytic transfer hydrogenation of furfural 1 to fur-
furol 2.[a]
To solve this problem, we applied predicting model-
ling on our dataset. The idea behind building a pre-
dictive model is linking the ligand-complexes space
(the imidazolinium precursors) with a figure of merit
or measured response (e.g., the furfural hydrogena-
tion yield), through an intermediate space of molecu-
lar descriptors. A descriptor is a number that encodes
structural and/or chemical information.[30,31] The
ligand backbone, the cone angle and the reaction
pocket volume are well known examples. By choosing
the “correct” descriptors we can build a model that
takes them as input and gives the predicted values of
performance (furfurol yield) as output. However,
choosing the right descriptors for modelling a given
catalytic system is both crucial and difficult. Previous-
ly, we showed that one can select good descriptors by
simply calculating a large set of them and then sieving
out the less important ones by using, for example, var-
iable importance (VIP) analysis.[32,33] Here, we com-
bined two statistical methods, principal component
analysis (PCA) and partial least squares (PLS) regres-
sion, to build and validate our descriptor model (see
Experimental Section for details).
Entry Imidazolinium
Ligand Precursor
Furfural
Furfurol
Conversion Selectivity Yield
[%][a]
[%]
[%][b]
1
2
3
4
5
6
7
8
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
100.0
95.9
97.9
62.4
89.1
100.0
99.9
63.3
98.9
98.9
99.8
99.4
98.8
99.2
99.3
99.9
100.0
91.6
94.1
99.9
99.1
99.4
98.9
98.8
98.3
99.5
99.2
98.1
98.1
98.3
98.7
99.2
98.5
99.6
99.6
99.3
98.2
90.0
99.9
95.0
97.4
61.7
88.0
98.3
99.4
62.8
97.1
97.0
98.2
98.2
98.1
97.8
99.0
99.5
99.3
90.0
84.6
9
10
11
12
13
14
15
16
17
18
19
20
21
First, we calculated a large number of various de-
scriptors for all of the catalysts derived from the pre-
cursors shown in Scheme 2 (168 descriptors for each
catalyst, see summary in Table 2 and complete list in
the Supporting Information). Then, we ranked these
descriptors according to their importance and exam-
ined the correlation between them using PCA.
no ligand
[a]
Standard reaction conditions: 5 mmol furfural 1,
0.1 mmol [RuCl2A(p-cymene)]2; 0.2 mmol 2-imidazolinium
CTHUNGTRENNUNG
precursor (structures shown in Scheme 1); 0.2 mmol KO-
t-Bu; 0.1 mmol KOH; 5 mL dried THF; 3 mL i-PrOH;
N2 atmosphere; 608C; 24 h and with a minimal TON of
20–25 (4% catalyst).
[b]
Yield determined by 1H NMR analysis as well as GC Figure 3 shows the relative position of the catalysts
analysis (n-octane as internal standard).
based on the two first PCs (the so-called ꢀscores
plotꢁ). These explain >72% of the variance in the
data. We see that the catalysts are well distributed.
spread of activity values.[28,29] Interestingly, pairing the This is important since experiments that are “bunched
reaction profiles in Figure 2 with the imidazolinium together” in a scores plot are redundant. Thus,
salt structures in Scheme 2, one sees two things: first, Figure 3 shows that this dataset is sufficiently diverse
the type of NHC ligand influences the catalytic activi- for building good predictive models. There is a single
ty. In most cases, the ligand enhances the catalysis, cluster, composed of ligands 5, 10, 11, 14, 19 and 20.
but not always. Second, we cannot point to any No outliers are observed, but low-yield data are
common structural factor that influences catalytic ac- scarce. This means that a model based on these data
tivity. For example, one would expect complexes de- will be biased towards high yields. For modelling pur-
rived from the precursor pairs {6, 7} and {11, 14} to poses, low-yield experiments are just as important as
give similar performance, due to their respectively high-yield ones.[6]
similar backbone structure,[6] but the results show oth-
The loadings plot, shown in Figure 4, indicates how
erwise. In each of these pairs, one ligand is an ex- much each descriptor contributes to a given PC. Un-
tremely good catalyst while the other is a poor cata- important descriptors have small loading values, and
lyst. In the case of 7, the performance is even lower will appear close to the centre. Conversely, important
than that of the carbene-free blank run! The elusive- descriptors will appear far from the centre. In this
ness of a simple explanation that tallies with our plot there are no symbols close to the centre. This
ꢀchemical intuitionꢁ does not mean that it does not means that all the variables are roughly equally im-
exist. It simply means that simple ChemDraw struc- portant. The 2D constitutional descriptors (ꢀ*ꢁ sym-
tures as depicted in Scheme 2 give us insufficient in- bols), appear in the top right and bottom left quarters,
formation. Moreover, the larger and more diverse the contributing equally to both PC1 and PC2. Similarly,
input data set is, the more difficult it becomes to pre- the 3D electrostatic descriptors (ꢀ+ꢁ symbols) con-
dict activity based on chemical intuition alone.
tribute equally to the first two PCs. Thus, these types
are the more important descriptors. Conversely, the
Adv. Synth. Catal. 2010, 352, 2201 – 2210
ꢂ 2010 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
asc.wiley-vch.de
2205

Zea Strassberger et al.
FULL PAPERS
Figure 2. Time-resolved reaction profiles for the Ru-carbene catalyzed hydrogenation of furfural 1 to furfurol 2, following
the standard reaction conditions as shown in Table 1. For clarity, the data are divided in two groups (left and right graphs),
since several profiles overlap. The blank experiment using [RuCl
continuous curve.
2ACHTUNGTREN(NUNG p-cymene)]2 precursor only (no ligand) is shown as a thick
Table 2. Definition and labelling of the descriptor types
shown in Figure 4.
Label Type
Examples
*
2D, constitu- Elemental composition, atom counts,
tional
atom types (no geometric, no elec-
tronic information)
ꢄ
2D, topologi- Connectivity, paths, shape and flexi-
cal bility of the molecule
3D, geomet- Inertia moment, 3D shadows, volume,
>
+
rical
surface, gravitation index
Charge distribution, electro negativi-
ty, polarity
3D, electro-
static
To predict the catalytic performance of new car-
bene-ruthenium complexes in this reaction, we used a
PLS regression model. We divided the original 18
ligand precursors (structures 4–21) in a training set of
13 and a validation set of five ligand precursors,
which resulted in a promising R2 =0.913.[34] Then, the
predictive power of the model for actual new struc-
Scheme 3. Two possible transfer hydrogenation pathways
from i-PrOH to furfural.
3D geometrical descriptors (ꢀ>ꢁ symbols) appear tures was evaluated. As noted above, the multicompo-
mainly on the right half of the plot, and the 2D topo- nent reaction provides facile access to a large number
logical descriptors (ꢀꢄꢁ symbols) appear mainly on of possible ligands. We predicted the activity and se-
the top left quadrant, so these are less relevant for lectivity of various structures using the model, and
PC1 and PC2, respectively.
chose to synthesize four new ligand precursors (struc-
2206
asc.wiley-vch.de
ꢂ 2010 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Adv. Synth. Catal. 2010, 352, 2201 – 2210

Finding Furfural Hydrogenation Catalysts via Predictive Modelling
Figure 3. Scores plot for the Ru complexes of the 2-imidazolinium salts used in this study.
Figure 4. Loadings plot for the 168 descriptors. In order to
study the effect and the relevance of the descriptor families
in the PC space, we have added different labels to different
descriptors. See Table 2.
tures 22–25 in Figure 5). Figure 5 and Table 3 show
the results. The experimental tests show that our aver-
age prediction error is ~3%. This is quite small con-
sidering the near-quantitative yield. With the large
number of data points in the high-yield area, howev-
er, such a high variation is somewhat surprising. It
may reflect a similarity in the test set structures. Im-
portantly, the model was not only able to find high-
yield compounds, but also a low-yielding one (ligand
Figure 5. The four new ligands 22–25 (top) and their predict-
ed vs. experimental yield (bottom).
25). These results demonstrate the value of using pre- ligands and calculating their descriptors is much less
dictive descriptor modelling, since drawing these four work than their actual synthesis.
Adv. Synth. Catal. 2010, 352, 2201 – 2210
ꢂ 2010 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
asc.wiley-vch.de
2207

Zea Strassberger et al.
FULL PAPERS
Table 3. Predicted vs. experimental yield for ligands 22–25.
tion times and/or NMR spectra to those of authentic sam-
ples. Ligands (4, 9, and 13) are known compounds, and were
synthesized following a published procedure[14,15] Ligand 5
was purchased from Fluka. The other eighteen ligands are
new compounds, synthesized using a modified procedure as
outlined below. Detailed synthesis procedures and character-
ization data for these compounds, as well as details of the
NMR characterization control experiments, are included in
the Supporting Information.
Ligand Predicted yield
[%]
Experimental yield
[%]
Error
[%]
22
23
24
25
100.0
96.9
100.0
79.4
97.2
98.7
94.8
80.8
2.9
1.8
5.4
1.7
General Procedure for the Synthesis of 2H-2-
Imidazoline Archetypes
Conclusions
Seven archetypes of 2H-2-imidazolines (A–G), the struc-
tures of which can be found in the Supporting Information,
were synthesized via a three-component Mannich-type reac-
tion according to literature procedures ([A–D]: ref.[16], [E
and F]: ref.[9], G: ref.[12]).
Combining multicomponent synthesis and predictive
modelling facilitates the quest for new catalysts.
MCRs give easy access to a variety of potential li-
gands, each of which can be synthesized in high yields
and selectivities. Predictive modelling guides the syn-
thesis efforts by highlighting “good regions” in the
catalyst space. In this way, one can avoid “dead ends”
and focus the experimental effort on promising struc-
tures. Importantly, we have shown that descriptor
models can give high correlations even in situations
where structure/activity relationships are elusive.
Moreover, this combined approach allows facile si-
multaneous variation of the parameters (e.g., chang-
ing substituents on several sites of the heterocyclic
ring) since in the descriptor model the dataset is
taken as a whole. We believe that such an approach
will expand rapidly, enhancing the performance of
conventional catalysis research. That said, computers
will not replace chemists, and data mining methods
will not replace mechanistic studies. These methods
will simply be part of the chemistꢁs toolbox in the 21st
century.
General Procedure for the Synthesis of 2-
Imidazolinium Salts
Reactions were carried out at a concentration of 1M of imi-
dazoline in acetone. The halide (1 equiv.) was added to a
stirred solution of the 2H-2-imidazoline and NaI (1 equiv.).
The reaction mixture was stirred at 258C for 18 h. Then, the
reaction mixture was filtered over Celite and concentrated
under vacuum.
Example: Imidazolinium iodide 23
According to general procedure II for the synthesis of 2-imi-
dazolinium salts, the reaction between 2-imidazoline D
(297.2 mg, 1.07 mmol), NaI (160.5 mg, 1.07 mmol) and
Experimental Section
Materials and Instrumentation
chloroACHTNUTRGNEmNUG ethyl methyl ether (86.1 mg, 81.3 mL, 1.07 mmol) af-
forded 2-imidazolinium iodide 23 as a yellow foam; yield:
481.7 mg (1.07 mmol, 100%). 1H NMR (500 MHz, CDCl3):
d=9.87 (s, 1H), 7.72–7.69 (m, 4H), 7.51–7.49 (m, 2H), 7.45–
7.43 (m, 2H), 4.757 (s, 2H), 4.30 (s, 2H), 3.06 (s, 3H), 1.70
(s, 9H); 13C NMR (125 MHz, CDCl3): d=157.1 (CH), 142.4
(C), 140.1 (C), 131.1 (CH), 129.4 (CH), 124.5 (CH), 120.8
(CH), 73.6 (C), 58.7 (CH2), 57.8 (CH2), 56.9 (CH3), 28.1
(3CH3); IR (neat): n=2988, 2947, 2877, 1616, 1449, 1374,
1250, 1172, 1107, 1105, 918, 766, 733 cmꢀ1; HR-MS: m/z=
321.1947, calcd. for C21H25N2O (MꢀIꢀ): 321.1961.
GC analysis was performed using an Interscience GC-8000
gas chromatograph with a 100% dimethylpolysiloxane capil-
lary column (VB-1, 30 mꢄ0.325 mm). 1H and 13C nuclear
magnetic resonance (NMR) spectra were recorded on
1
Varian Mercury 300 (300 MHz for H), Bruker Avance 250
(250.13 MHz for 1H and 62.90 MHz for 13C), or Bruker
1
Avance 500 (500.23 MHz for H and 125.70 MHz for 13C).
Chemical shifts (d) reported in reported in ppm, internally
referenced to residual solvent resonances for CDCl3 (1H d=
7.26 ppm; 13C{1H} d=77.00 ppm), and coupling constants (J)
are reported in Hz. Samples for GC analysis were diluted in
1 mL EtOH. GC conditions: isotherm at 608C (2 min);
ramp at 508C minꢀ1 to 808C; isotherm at 808C (3 min);
ramp at 18C minꢀ1 to 908C; ramp at 508C minꢀ1 to 2508C;
isotherm at 2508C (3 min). All reactions were performed
under N2 using standard Schlenk techniques. Unless other-
wise noted, all chemicals were purchased from commercial
sources and used as received. All products are known com-
pounds and were identified by comparing of their GC reten-
Procedure for the Preparation of Ru-Carbene
Complexes
A solution of [RuCl2ACHTUNTGRENUNG(p-cymene)]2 (61 mg, 0.1 mmol) was
stirred under N2 in a 25-mL Schlenk tube together with the
2-imidazolinium salt (0.2 mmol), KO-t-Bu (22 mg, 0.2 mmol)
and KOH (6 mg, 0.1 mmol) in 5.0 mL dried THF and
3.0 mL i-PrOH for 30 min at 408C.
2208
asc.wiley-vch.de
ꢂ 2010 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Adv. Synth. Catal. 2010, 352, 2201 – 2210

Finding Furfural Hydrogenation Catalysts via Predictive Modelling
tion. This gives a better initial structure for the geometry
optimization. The descriptors were computed with the Co-
dessa software package[36] and analyzed using Matlab
scripts.[37] A total of 168 descriptors (2D and 3D) were cal-
culated for four different classes: constitutional, topological,
geometrical and electrostatic. A full list of these descriptors
is included in the Supporting Information.
Example: [RuCl2ACHTUNGTRENNUNG(p-cymene)]2/Ligand 4
ACHTUNGTRENNUNG[RuCl2ACHTUNGTRENNUNG(p-cymene)]2 (61 mg, 0.1 mmol), 2-imidazolinium salt
4 (83.4 mg, 0.2 mmol), KOtBu (22 mg, 0.2 mmol) and KOH
(6 mg, 0.1 mmol) were stirred in 5.0 mL dried THF and
3.0 mL i-PrOH for 30 min at 408C.
Procedure for Catalytic Transfer Hydrogenation
Data Analysis and Model Validation
In a 25-mL Schlenk tube under N2, the catalyst was first pre-
pared according to the previous procedure. After adding fur-
fural 1 (480 mg, 5 mmol) and the internal standard octane,
in a furfural-octane mass ratio of 2:1, the reaction mixture
was heated to 608C and stirred for 24 h. Periodically, sam-
ples were taken out and analyzed by GC.
The descriptors matrix was analyzed using principle compo-
nent analysis (PCA). A detailed description of this tech-
nique is available elsewhere.[34,38] Putting it simply, PCA re-
duces a large data matrix into two smaller matrices, that are
easier to examine and interpret. Using PCA, you can extract
the key factors. These are the principal components, or PCs
(also called the latent variables). Mathematically speaking, if
X is an (IꢄJ) matrix that contains J variables for I reactions,
PCA divides this matrix into a systematic part TPT (the
PCA model), and a residuals part E. T is the scores matrix.
It represents the spread of the reactions within the model
space. P is the loadings matrix. It describes the relationships
between the variables.
After calculating the ligand descriptors and ranking them,
we used partial least squares (PLS) regression to relate
them to the yield of furfurol (the figure of merit, FOM). For
this model, we computed 12 latent variables, retaining the
five most important ones that explained 89.2% of the var-
iance in the data.
Procedure for Acyloin Condensation of Furfural 1 to
Furoin 3a[23]
In a 100-mL Schlenk tube equipment under N2, 2-imidazoli-
um salt (5) (28 mg, 1 mmol) and KO-t-Bu (122 mg, 1 mmol)
were dissolved in 50 mL of dry THF and stirred for 15 min
at 248C. After addition of distilled furfural 1 (3200 mg,
33 mmol) the reaction mixture was stirred overnight, and
quenched with 2 mL formic acid (98%). The solvents were
evaporated and the crude product was recrystallized from
EtOH (85 mL), to give 3a as a yellow solid (yield: 3030 mg,
31.56 mmol, 94%) containing 2–3% of furyl 3b as side prod-
uct.
The model was validated using both internal and external
prediction. For the internal validation, we used a training set
of 13 ligands and a validation set of five ligands. Cross-vali-
dation was used to improve the modelꢁs stability. For the ex-
ternal validation, we used four newly synthesized ligands,
that were “unknown” to the model. The error percentage
and the DFOM for the external validation were calculated
using %error=jDFOM/expFOMjꢄ100, and DFOM=
jexpFOMꢀpredFOMj, respectively. A more detailed note
on the prediction error and a residuals plot are included in
the Supporting Information.
Procedure for Transfer Hydrogenation of Furfural
with i-PrOH-2d1 as Hydride Source
In a 10-mL Schlenk tube under nitrogen atmosphere,
0.05 mmol [RuCl2ACHTUNGTRENNUNG(p-cymene)]2, 0.1 mmol 2-imidazolinium
salt 5, 0.1 mmol KO-t-Bu, 0.05 mmol KOH were dissolved in
3 mL dried THF and 1.5 mL i-PrOH-2d1 and stirred for
30 min at 248C. The reaction mixture was heated to 608C
for 5 h after the addition of 2.5 mmol furfural 1 and the in-
ternal standard, octane. The solvent was removed in vacuum
and the residue solved in CDCl3. The conversion was deter-
1
1
mined by H NMR and by GC analysis. H NMR (300 MHz,
CDCl3): d=7.41 (s, 1H), 6.35 (d, 1H), 6.30 (d, 1H), 4.70
(s,1H), 4.62 (s, 2H), 3.73 (s, 1H).
Supporting Information
Detailed synthesis and analysis procedures including calibra-
tion studies and characterization details for ligands, a full
list of the descriptors used, and an explanation and discus-
sion of the modelling and validation approach are available
as Supporting Information.
Computational Methods
All computations were performed on a Dell Latitude D630
laptop with a double core 4.0 GHz processor.
Ligand Geometry Optimization and Descriptor
Calculation
Geometry optimization for calculating the 3D descriptors References
was performed using Hyperchem.[35] We used the MM+
force field in combination with a conjugate gradient optimi-
zation method (Polak–Ribiere). The 3D optimization of the
ligand-metal complexes turned out to be difficult. The
common algorithms for geometric optimization yielded only
high energies (~120 kcal) instead of the minima of 20–30
kcal. Having a 908 angle between the imidazole aromatic
ring and the substituent aromatic 3-ring spiro moiety is cru-
cial. The solution is to force a change into the 3D configura-
[1] M. Albrecht, K. J. Cavell, in: Organometallic Chemis-
try, (Eds.: I. J. S. Fairlamb, J. M. Lynan), Vol. 25,
R.S.C., Cambridge, pp 47–61, 2009.
[2] Q. Shuang, G. C. Fu, J. Org. Chem. 1998, 63, 4168.
[3] J. Collot, J. Gradinaru, N. Humbert, M. Skander, A.
Zocchi, T. R. Ward, J. Am. Chem. Soc. 2003, 125, 9030.
[4] J. G. De Vries, C. J. Elsevier, Handbook of Homogene-
ous Hydrogenation, Wiley-VCH, Weinheim, 2007.
Adv. Synth. Catal. 2010, 352, 2201 – 2210
ꢂ 2010 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
asc.wiley-vch.de
2209

Zea Strassberger et al.
FULL PAPERS
[5] C. Klanner, D. Farrusseng, L. Baumes, M. Lengliz, C.
Mirodatos, F. Schꢅth, Angew. Chem. 2004, 116, 5461;
Angew. Chem. Int. Ed. 2004, 43, 5347.
[6] A. G. Maldonado, J. A. Hageman, S. Mastroianni, G.
Rothenberg, Adv. Synth. Catal. 2009, 351, 387.
[7] R. V. A. Orru, M. de Greef, Synthesis 2003, 1471.
[8] A. Dçmling, Chem. Rev. 2006, 106, 17.
tions (see Experimental Section and Supporting Infor-
mation for details).
[22] S. Enthaler, R. Jackstell, B. Hagemann, K. Junge, G.
Erre, M. Beller, J. Org. Chem. 2006, 71, 4652.
[23] D. Enders, U. Kallfass, Angew. Chem. 2002, 114, 1822;
Angew. Chem. Int. Ed. 2002, 41, 1743.
[24] In theory, reductive elimination of an intermediate (di)-
hydride to an imidazole species that does not catalyze
acyloin condensation may also take place, but since our
reactions are under basic conditions, the base would re-
store the carbene in situ. Hence, any carbene coming
off as an H-imidazole, would be restored and then cata-
lyze the above acyloin condensation. Given that this
condensation was not observed, and since a series of
control experiments confirmed that neither OHꢀ nor t-
BuOꢀ in itself promotes the acyloin condensation of
furfural, we maintain that the carbene ligands remain
coordinated throughout the reaction.
[25] M. Albrecht, J. R. Miecznikowski, A. Samuel, J. W.
Faller, R. H. Crabtree, Organometallics 2002, 21, 3596.
[26] J. S. M. Samec, J.-E. Bꢃckvall, P. G. Andersson, P.
Brandt, Chem. Soc. Rev. 2006, 35, 237.
[27] S. Gladiali, G. Mestroni, Transition Metals for Organic
Synthesis 2nd edn., Wiley-VCH, Weinheim, 2004.
[28] E. J. Ras, S. Maisuls, P. Haesakkers, G. J. Gruter, G.
Rothenberg, Adv. Synth. Catal. 2009, 351, 3175.
[29] E. Burello, D. Farrusseng, G. Rothenberg, Adv. Synth.
Catal. 2004, 346, 1844.
[9] B. Groenendaal, E. Ruijter, R. V. A. Orru, Chem.
Commun. 2008, 5474.
[10] N. Elders, F. J. J. de Kanter, E. Ruijter, M. B. Groen,
R. V. A. Orru, Chem. Eur. J. 2008, 14, 4961.
[11] R. S. Bon, C. Hong, M. J. Bouma, R. F. Schmitz, F. J. J.
de Kanter, M. Lutz, A. L. Spek, R. V. A. Orru, Org.
Lett. 2003, 5, 3759.
[12] R. S. Bon, B. van Vliet, N. E. Sprenkels, R. F. Schmitz,
F. J. J. de Kanter, C. V. Stevens, M. Swart, F. M. Bickel-
haupt, M. B. Groen, R. V. A. Orru, J. Org. Chem. 2005,
70, 3542.
[13] N. Elders, R. F. Schmitz, F. J. J. de Kanter, E. Ruijter,
M. B. Groen, R. V. A. Orru, J. Org. Chem. 2007, 72,
6135.
[14] R. S. Bon, F. J. J. de Kanter, M. Lutz, A. L. Spek, M. C.
Jahnke, F. E. Hahn, M. B. Groen, R. V. A. Orru, Orga-
nometallics 2007, 26, 3639.
[15] R. S. Bon, N. E. Sprenkels, M. M. Koningstein, R. F.
Schmitz, F. J. J. de Kanter, A. Domling, M. B. Groen,
R. V. A. Orru, Org. Biomol. Chem. 2008, 6, 130.
[16] V. Dragutan, D. Ileana, D. Lionel, D. Albert, Coord.
Chem. Rev. 2007, 251, 765.
[17] K. J. Zeitsch, The Chemistry and Technology of Furfu-
ral and its Many By-products, Vol. 13, 1st edn., Elsevier,
Amsterdam, 2000.
[18] C. Sievers, I. Musin, T. Marzialetti, M. Olarte,
B. V. P. K. Agrawal, C. W. Jones, ChemSusChem 2009,
2, 665.
[19] R. H. Kottice, in: Kirk-Othmer Encyclopedia of Chemi-
cal Technology, Volume supplement, 4th edn. (Eds.: J.
Kroschwitz, M. Home-Grant), John Wiley and Sons,
New-York, 1997, p 155.
[30] G. Rothenberg, Catalysis: concepts and green applica-
tions, Wiley-VCH, Weinheim, 2008.
[31] J. A. Hageman, J. A. Westerhuis, H. W. Frꢅhauf, G.
Rothenberg, Adv. Synth. Catal. 2006, 348, 361.
[32] E. Burello, G. Rothenberg, Int. J. Mol. Sci. 2006, 7, 375.
[33] E. Burello, G. Rothenberg, Adv. Synth. Catal. 2005,
347, 1969.
[34] A. Golbraikh, A. Tropsha, J. Molec. Graph. Model.
2002, 20, 269.
[35] Hyperchem for Windows (Molecular Modeling
System), Version 7.5 ed., Hyperchem, Hypercube, Inc.,
2002.
[20] Z. Strassberger, M. Mooijman, E. Ruijter, A. H. Al-
berts, C. de Graaff, R. V. A. Orru, G. Rothenberg,
Appl. Organomet. Chem. 2010, 24, 142.
[36] Codessa, Codessa, Version 2.642 edn., 1994.
[37] The Mathworks, Inc. Matlab, Version R2007a ed.,
Matlab, 1984–2007.
[21] Characterising this complex by NMR requires rigorous
[38] S. Wold, K. Esbensen, P. Geladi, Chemom. Intell. Lab.
exclusion of water and oxygen under glove-box condi-
Syst. 1987, 2, 37.
2210
asc.wiley-vch.de
ꢂ 2010 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Adv. Synth. Catal. 2010, 352, 2201 – 2210