Supporting Information for Manuscript titled

“Rapid and Accurate Prediction of pKa Values of C-H Acids Using Graph Convolutional Neu-
ral Networks”
by Rafał Roszak1,3, Wiktor Beker1,3, Karol Molga1, Bartosz A. Grzybowski1,2,3*

1
Institute of Organic Chemistry, Polish Academy of Sciences, ul. Kasprzaka 44/52, 01-224 Warsaw,
Poland
2
Institute for Basic Science, Center for Soft and Living Matter, Ulsan 44919, South Korea
3
Allchemy, Inc., 2145 45th Street #201, Highland, IN 46322, USA

CONTENTS:

Section S1. Overview of other methods for pKa prediction ............................................................... 1

Section S2. Quantum-chemical benchmark calculations ................................................................... 2
Section S3. Details on the composition of the training set ................................................................. 7
Section S4. Details of neural networks’ setup .................................................................................... 9
Section S4.1 Fingerprint based ......................................................................................................... 9
Section S4.2 Embedded fingerprint .................................................................................................. 9
Section S4.3 Chemistry-aware NN ................................................................................................. 10
Section S4.4 Graph convolutional neural network (GCNN) ...........................................................11
Section S5. Transfer learning experiment ........................................................................................ 12
Section S6. Other factors influencing reaction outcomes................................................................. 14
Section S7. Supplementary references ............................................................................................. 16

Section S1. Overview of other methods for pKa prediction

Methods of pKa prediction can be divided into two categories: i) ab initio calculations and ii) statistical
modeling. Ab initio calculations allow to estimate absolute pKa from Gibbs free energy of dissociation
reaction in a given solvent, (ΔGaq): HA = H+ + A-. This value can be obtained from a thermodynamic
cycle7a: ΔGaq = G(A-gas) – G(AHgas) + ΔGs(A-) + ΔGs(H+) - ΔGs(AH), where G denotes free energy of a
given species and ΔGs is the corresponding solvation energy. Since the proton solvation free energy
cannot be obtained directly from quantum-chemical calculations, it has to be estimated using
experimental data. Although this protocol allows for de novo estimation of pKa of any molecule, values
obtained in this manner “can be a few pKa units away from the experimentally measured values, even
for the simplest, most rigid molecules.”.6 To circumvent this problem and improve the prediction
quality, a set of empirical correction parameters tailored for specific molecules/motifs may be used. An
example of such an approach is the pKa prediction module of the Jaguar software developed by
Schrodinger.7b

S-1
 A common approximation taken in ab initio approaches is the continuum (implicit) solvent
model, which averages the internal states of the solvent and characterizes it with some global
parameters (usually the dielectric constant). Absolute pKa prediction with atomic resolution and
accounting for the conformational flexibility of both the compound of interest and its surroundings is
possible with ab initio molecular dynamics (MD) simulations (either Car-Parinello7c or Born-
Oppenheimer7d scheme of computation). However, these approaches involve significant computational
cost, which is probably the reason behind rather small number of such studies reported to date.7e
It is also worth to briefly narrate techniques useful for the correction of systematic errors in
theoretical pKa calculations, namely establishing the relationship between calculated parameters and
the dissociation free energy (strictly proportional to pKa). The most common type of this approach is
the so-called Linear Free Energy Relationship (LFER), whereby one assumes linear dependence
between pKa and calculated dissociation energy: pKa= a*ΔG+b. Empirical constants a and b are
obtained from linear regression to experimental data.7f This methodology, when properly tuned,7h yields
highly accurate pKa values, but its application is limited only to homologous series as regression
parameters are not transferable between different groups of compounds (e.g., phenols and carboxylic
acids would require different constants).
The last group of pKa prediction methods is focused on identifying statistical correlations
between molecular features and experimentally reported pKa values. Historically, the first such model
was based on Hammett and Taft parameters.8a Other descriptors were later developed leading to
modern QSAR methods.8b However, such approaches were usually intended for rather narrow groups
of compounds (e.g. phenols,8c amines,8d etc.). More recently, machine learning methods were applied to
pKa prediction as well,2a yet still restricted to functional groups ionizable in water solutions (i.e., not C-
H acids). Quality of these models depends strongly on the number of available experimental data.
Considering typical classes of pharmaceutical substances, tens of thousands of experimental pKa values
are available, allowing to achieve mean absolute error (MAE) in the range 0.5 – 0.9, depending on the
software used.5

Section S2. Quantum-chemical benchmark calculations

Because the quality of the LFER approximation depends strongly on the configuration of QM
calculations, we performed a comprehensive validation of the corresponding settings. Due to
computational cost, we decided to limit ourselves to Density Functional Theory (DFT) methods and to
not try any post-Hartree-Fock approaches, like Coupled Cluster or Møller–Plesset perturbation theory.
Looking for a configuration providing the best correlation between calculated G and the experimental
pKa values of 26 sulfone compounds, we independently evaluated the effects of the choice of density
functional (Table S1), atomic radii (defining molecular cavity in implicit solvent model, Table S2) and
the basis set (Table S3). All calculations were performed using Gaussian09 software.S1
First, a series of 50 density functionals (Table S1) was examined using 6-31+G(d) basis set with
integral-equation-formalism variant of polarizable continuum model (IEF-PCM) of DMSO solvent
(with UFF atomic radii scaled by 1.1). Each one of these DFT methods gives a very good correlation
with experimental data, with coefficient of determination R2 > 0.95 and mean absolute error (MAE) in
the range from 0.65 to 1.16. The worst correlation was observed for functionals with empirical
dispersion correction (e.g. wB97XD, APFD, B97D), whereas the best result was obtained for those
including a PBE correlation term. Among all considered methods, the best result – that is, the lowest
MAE and the highest R2– was obtained for HISSbPBE functional.
In the continuum model of solvent, a molecule is placed inside a cavity within a continuous
dielectric medium. The influence of the environment on the molecule is then modeled by point charges

S-2
induced on the surface of the cavity by molecular electric field. The cavity itself is constructed from a
set of overlapping spheres centered on molecule’s atoms, with the radius of each one defined as an
atom-type-dependent parameter multiplied by a scaling factor. In order to find an optimal cavity
definition, three types of atomic radii were tested (Bondi, UFF, Pauling) with different scaling factors
(Table S2). This test was performed with B3LYP functional with 6-31+G(d) basis set. The best
correlation was obtained using UFF atomic radii: the lowest MAE was observed with scaling factor
1.4, whereas the highest R2 was obtained with scaling factor 1.6. We decided to take the midpoint value
of the scaling factor (1.5) for the final model.
Finally, we evaluated how the choice of the basis set affected model’s accuracy. The
calculations were performed with B3LYP functional and continuum model of DMSO using IEF-PCM
formalism with UFF atomic radii (scaling factor=1.1). Four families of basis sets were selected for tests
(Table S3), namely: i) Pople (entries 1-9), ii) Duning (entries 9-15), iii) Jensen (16-22) and iv) Weigend
(entries 23-27). We decided to not include computationally demanding quadruple zeta (or higher) basis
sets. In the Pople’s basis set series, mean absolute error of pKa obtained with a widely used double-zeta
6-31+G(d) basis set was 0.82, yet even higher MAE was observed with corresponding triple-zeta 6-
311+G(d) basis set. Any improvement in this family requires inclusion of significant number of
polarization functions, like 6-31+G(2df,p). However, even such augmented basis sets gave slightly
higher MAE than Jensen’s double-zeta basis sets (pc1, pcseg-1 or pcSseg-1) comprised of significantly
fewer functions (ca. half). In the Dunning family, MAE of double-zeta cc-pVDZ basis set is rather
high, namely 0.85; incorporation of diffusion functions (entries 10-14) does not improve the correlation
in the tested series of compounds. In this series, MAE was lower only in the case of rather large triple
zeta cc-pVTZ, with the value of 0.61. Within Jansen’s basis set family, the result does not depend
strongly on the inclusion of diffusion functions, whereas extension to triple-zeta reduces MAE by only
0.06. Double-zeta basis sets from Weigend, def2svp and def2svpp, provide high MAE of 3.26 and 2.53,
respectively, which is significantly higher than any other tested double-zeta basis set. On the other
hand, triple-zeta def2tzvp gives low MAE of 0.68. To summarize, among all tested basis sets, the best
results were obtained with the Jensen family, namely 0.57 in the case of pc-s-seg-2 (triple-zeta class)
and 0.63 for pc-seg-1 (double-zeta class).
The choice of the basis set involves a trade-off between the computational cost (increasing with
the size of the basis set) and accuracy (increasing with the basis set size). Hence, we decided to
compare selected basis sets using production setup (HISSbPBE functional and UFF radii with scaling
factor 1.5 for solute cavity definition). For the final comparison, we decided to choose two basis sets:
pcseq-1, which was the best among tested double-zeta basis sets, and triple-zeta def2tzvp (Table S4).
The def2tzvp basis set was ultimately chosen as a compromise between accuracy and size – this basis
set gives slightly higher MAE that triple zeta basis sets from Jansen and Duning but contains ~15% less
functions.

Table S1: Mean absolute errors (MAE) and coefficients of determination (R2) for different density
functionals. Calculations were performed for 26 sulfones, PhSO2-R, using 6-31+G* basis set and
integral equation formalism variant of polarizable continuum model of DMSO solvent (ε = 46.826,
UFF radii with scaling factor 1.1)

Entry Functional R2 MAE

1 B3LYP 0.96795 0.81711
2 HCTC/406 0.96623 0.76564
3 HCTH/147 0.96630 0.73145
4 HCTH/93 0.96526 0.73766
5 BLYP 0.95730 0.90410
6 tHTCH 0.96801 0.73107
S-3
Entry Functional R2 MAE
7 BMK 0.95635 1.05251
8 tHTCHhyb 0.96973 0.76594
9 BPBE 0.96716 0.71087
10 MPWPBE 0.96695 0.74513
11 PW91PBE 0.96736 0.74998
12 wB97 0.95842 1.01437
13 wB97X 0.96168 0.98052
14 wB97XD 0.94982 1.15367
15 M06 0.96056 0.98845
16 M06HF 0.95208 1.15923
17 M06L 0.96297 0.84693
18 M11 0.95889 1.03419
19 M11L 0.95836 0.83824
20 MN12L 0.95744 0.98311
21 MN12SX 0.96652 0.87897
22 N12 0.96469 0.77204
23 N12SX 0.97529 0.65423
24 SOGGA11 0.96170 0.84898
25 SOGGA11X 0.97219 0.79256
26 M062X 0.96712 0.88886
27 X3LYP 0.96784 0.82808
28 B1B95 0.97602 0.69945
29 B3P86 0.97520 0.65900
30 B971 0.96944 0.79930
31 B972 0.97479 0.66320
32 B97D 0.95413 1.01965
33 Bb98 0.96853 0.81366
34 BhandH 0.97470 0.78554
35 BhandHLYP 0.96676 0.90800
36 HISSbPBE 0.97830 0.64696
37 HSEH1PBE 0.97570 0.66561
38 OHSE1PBE 0.97542 0.66832
39 OHSE2PBE 0.97622 0.65782
40 mPW1LYP 0.96611 0.87323
41 mPW1PW91 0.97564 0.65565
42 mPW3PBE 0.97494 0.65749
43 PBE1PBE 0.97622 0.65496
44 PBEh1PBE 0.97603 0.66767
45 PBEPBE 0.96702 0.73907
46 PW91PW91 0.96729 0.75778
47 TPSSh 0.97224 0.70149
48 TPSSTPSS 0.96813 0.74869
49 APF 0.97583 0.64786
50 APFD 0.95902 0.97976

S-4
Table S2: Mean absolute errors (MAE) and coefficients of determination (R2) for different density
functionals and for different definition of atom radii defining molecular cavity. Calculations were
performed for 26 sulfones, PhSO2-R, using B3LYP functional with 6-31+G* basis set and integral
equation formalism variant of polarizable continuum model of DMSO solvent (ε = 46.826).

Entry Radius type Scaling factor R2 MAE

1 1.0 0.9603 0.9681
2 1.1 0.9679 0.8175
3 1.2 0.9703 0.7422
4 1.3 0.9725 0.6769
5 UFF 1.4 0.9736 0.6711
6 1.5 0.9739 0.6930
7 1.6 0.9742 0.7055
8 1.7 0.9739 0.7333
9 1.8 0.9735 0.7497
10 1.0 0.6156 2.3946
11 1.1 0.9518 1.0697
12 1.2 0.9637 0.9002
13 Bondi 1.3 0.9691 0.7947
14 1.4 0.9709 0.7377
15 1.5 0.9728 0.6853
16 1.6 0.9739 0.6714
17 1.1 0.9436 1.3589
18 1.2 0.9527 1.1716
19 Pauling 1.3 0.9615 0.9445
20 1.4 0.9661 0.8244
21 1.5 0.9675 0.7687

Table S3: Mean absolute errors (MAE) and coefficients of determination (R2) for different basis sets.
Calculations were performed for 26 sulfones, PhSO2-R, using B3LYP functional and integral equation
formalism variant of polarizable continuum model of DMSO solvent (ε = 46.826, UFF radii with
scaling factor 1.1)

entry Family Basis set R2 MAE

1 6-31+G(d) 0.9679 0.8171
2 6-31++G(d) 0.9677 0.8234
3 6-31+G(2d) 0.9737 0.7153
4 6-31+G(2d,p) 0.9739 0.6980
Pople
5 6-31+G(2df,p) 0.9760 0.6536
6 6-31+G(2df,2p) 0.9756 0.6629
7 6-311+G(d) 0.9678 0.8367
8 6-311+G(d,p) 0.9688 0.8135
9 cc-pVDZ 0.9648 0.8554
10 apr-cc-pVDZ 0.9623 0.9324
Dunning
11 may-cc-pVDZ 0.9622 0.9332
12 jun-cc-pVDZ 0.9623 0.9324

S-5
 entry Family Basis set R2 MAE
13 jul-cc-pVDZ 0.9622 0.9149
14 aug-cc-pVDZ 0.9619 0.9162
15 cc-pVTZ 0.9750 0.6143
16 pc-1 0.9742 0.6564
17 pc-seg-1 0.9753 0.6344
18 pc-s-seg-1 0.9747 0.6384
19 Jensen aug-pc-seg-1 0.9699 0.8411
20 pc-2 0.9776 0.5746
21 pc-seg-2 0.9773 0.6013
22 pc-s-seg-2 0.9779 0.5702
23 def2svp 0.9636 3.2601
24 def2svpp 0.9641 2.5265
Weigend
25 def2tzvp 0.9769 0.6840
26 def2tzvpp 0.9765 1.2438

Table S4: Mean absolute errors (MAE) and coefficients of determination (R2) for different basis sets.
Calculations were performed for 15 nitriles using HISSbPBE functional and integral equation
formalism variant of polarizable continuum model of DMSO solvent (ε = 46.826, UFF radii with
scaling factor 1.5)

entry Basis set R2 MAE

1 pc-seq-1 0.9726 0.9002
2 def2tzvp 0.9823 0.7358

Table S5: Mean absolute errors (MAE) and coefficients of determination (R2) for different series of
compounds. Calculations were performed at the HISSbPBE level with def2tzvp basis set and integral
equation formalism variant of polarizable continuum model of DMSO solvent (ε = 46.826, UFF radii
with scaling factor 1.5)
entry Number of compounds Compounds type R2 MAE
1 27 R-SO2Ph 0.982 0.6232
2 15 R-CN 0.9823 0.7358
3 14 R-NO2 0.9555 0.7866
4 11 R-COOEt 0.9207 1.0680

S-6
Section S3. Details on the composition of the training set
Table S6.Molecular make-up of the dataset. Each compound was assigned to class/group according
to the patterns provided in the middle column. In case of multiple matches, the one higher in the Table
was used for classification. Group ‘Other’ includes compounds like DMSO, methane, etc.

No. of No. of
Class Group Name Group Pattern exp. ex- theoretical
amples examples
I Nitroalkanes [C!H0][NX3+](=[O])[O-] 14 56
II [SX3+] [C!H0][SX3+,SX4+] 4 0
Total 261 156
1,3-diketones [#6][CX3](=[OX1])[CX4!H0][CX3](=[OX1])[#6] 14 0
Cyclic ketones [CX4][CX3R]([CX4])=[OX1] 13 0
Acetophenones c1cc([*])ccc1[CX3](=[OX1])[CX4H3] 19 0
Phenylketones c1ccccc1[CX3](=[OX1])[CX4H2][*] 14 0
Alkyl ketones [CX4H3][CX3](=[OX1])[CX4H2][*] 8 0
Other ketones [C!H0][CX3]=[OX1] 16 0
III
Esters [C!H0][CX3](=[OX1])[OX2] 31 72
Amides [C!H0][CX3](=[OX1])[NX3H0] 8 0
Thioamides [C!H0][CX3](=[SX1])[NX3] 3 0
Nitriles [C!H0]C#N 55 42
Sulfones [C!H0][SX4](=[OX1])(=[OX1]) 63 42
Sulfoximides [C!H0][SX4](=[N])(=[OX1]) 7 0
[PX4+] [C!H0][PX4+] 10 0
Total 119 0
Thioethers [C!H0][SX2H0] 22 0
[PX3] [C!H0][PX3] 1 0
[OX1]=[PX4] [C!H0][PX4]=[OX1] 4 0
IV
[NX4+] [C!H0][NX4+] 2 0
Ethers [C!H0][OX2H0] 7 0
Benzylic [C!H0][c] 78 0
Allylic [CX4!H0][CX3]=[CX3] 5 0
V Arenes [cH] 6 194

Other Compounds not belonging to either of above groups 10 2

S-7
Table S7. Similarity between groups in the dataset. In all comparisons, carbon atom with the lowest
pKa value was marked as C-14 isotope before calculation of ECFP4 fingerprint. In this way, we
intended to enforce comparisons of active sites between molecules. Average Tanimoto similarities
within each group, as well as between the group and the rest of the dataset, are provided.

Average similarity Average similarity No. of mem-

Class Group Name
within group to the rest of the dataset bers

I Nitroalkanes 0.285 0.060 70

II [SX3+] 0.244 0.104 4
Whole class 0.136 0.088 417
Ketones 0.201 0.106 84
Esters 0.326 0.077 103
Amides 0.240 0.103 8
III Thioamides 0.183 0.100 3
Nitriles 0.216 0.071 97
Sulfones 0.318 0.129 105
Sulfoximide 0.430 0.128 7
[PX4+] 0.353 0.125 10
Whole class 0.161 0.098 119
Thioethers 0.186 0.103 22
[PX3] Not applicable 0.136 1
[OX1]=[PX4] 0.387 0.083 4
IV
[NX4+] 0.045 0.074 2
Ethers 0.319 0.127 7
Benzylic 0.192 0.122 78
Allylic 0.050 0.048 5

V Arenes 0.145 0.056 200

Other Whole class 0.098 0.086 12

S-8
Section S4. Details of neural networks’ setup
All neural networks were built in Python using Keras library with TensorFlow backend. Molecular
descriptors and fingerprints were obtained with RDKit library (versions 2018.09 and 2017.09). The
details of the training of NNs shown in the main-text Figure 4 (panels a-c) are described in the
following subsections.

Section S4.1 Fingerprint based

Fingerprint-based input vector has 604 dimensions and consists of three parts:
1. Description of proton-donating atom, four features: Gasteiger charge of the neutral atom, Gasteiger
charge of atom in anionic form (after proton dissociation), number of lone electron pairs and
number of pi electrons on the atom.
2. Fingerprint representation of the core atom: ECFP2 fingerprint of radius 1i centered at the proton-
donating atom. Such fingerprint was then cast into a bit vector of length 300.
3. Fingerprint representation of proton-donating atom neighbors: for each one, a bit vector was
created in the same fashion as above; then, all those vectors were summed up.
Note that such a representation effectively describes environment of a proton-donating atom within
topological distance of 2, since i) the nearest neighbors of center atom are explicitly described in the
feature vector; ii) each of these neighbors is characterized by their environment of radius 1; and iii) the
atoms within the topological distance of 2 from the center atom are simultaneously within distance of 1
of central atom’s nearest neighbors. After the network optimization with 5-fold cross-validation,
following hyper-parameters gave the lowest MAE: dropout = 0.05, L2 regularization = 0.005, learning
rate = 0.002, batch size = 5 and ELU as a activation function.
We have also tested other input vectors (all having 604 dimensions).
1. The same as described in the paragraph above but radius of 2 was used for all fingerprints
(thus the effective radius of proton-donating atom’s environment covered by the model was
3). Such representation leads to higher MAE of 3.8 which is not surprising because of very
significant overlap between the environments of central and neighboring atoms. Optimal
network architecture consists of two hidden layers with 100 and 20 neurons respectively,
dropout = 0.005, L2 regularization = 0.002, learning rate = 0.002, batch size = 5 and ELU as
a activation function).
2. A representation sharing the same four features for proton-donating atom (Gasteiger charges,
numbers of pi electrons and the lone pairs), but having only a ECFP4 fingerprint of proton
donating atom (cast into a 600-bit vector). Best model of dense neural network with such an
input gives MAE of 4.5 pKa units. We also tested radii of value 3 and 4, which led to MAE
4.7 and 4.9, respectively.

Section S4.2 Embedded fingerprint

Here, we tested a fingerprint embedding based on the concept of word embedding (WE), which is
currently a standard technique in natural language processing (NLP). The basic idea of WE is to assign
a high-dimensional vector to each word in such a way that the distance between two words corresponds

i
In the Extended Connectivity Fingerprints (ECFP; ref 17 in the main text), the last digit corresponds to the maximum
diameter of neighborhood considered during computation. Diameters of value 4 and 6 (corresponding to the neighborhood
radii 2 and 3, respectively) are most often used. Here, we implemented the radius 1, corresponding to the diameter 2 – thus,
the proper name is ECFP2.
S-9
to the probability of their simultaneous occurrence in the same context. In the mol2vec approach,17b a
molecule is converted into a list of ECFP4 identifiers (unique numerical representations of molecular
fragments). Then, each identifier is treated as a ‘word’, whereas the whole molecule corresponds to a
‘sentence’. Identifiers are ordered according to the occurrence of atoms in canonical SMILES. Such
representation is then used for embedding with word2vec, which is an unsupervised WE method
requiring to be trained on a large text corpus. Here, we use 19 million compounds taken from Zinc 15
database as such corpus. The embedding was trained with the skipgram method with 10 words window
and the dimensionality of the resulting vector space set to 300. Each word that occurred in the corpus
less than three times was replaced by string ‘UNSEEN’. Embedded fingerprints for atoms were built in
a following way: all identifiers of radius 1 from ECFP4 fingerprints (which describe just the atom and
its immediate neighborhood) were embedded into the aforementioned 300-D space and the resulting
vectors were summed up. Input vector for the NN was constructed in the same way as in previous
section, but instead of ECFP4 fingerprints, the embedded fingerprints were used.
After the network optimization with 5-fold cross-validation, the following hyper-parameters gave the
lowest MAE: dropout = 0.01, L2 regularization = 0.0005, learning rate = 0.002, batch size = 5 and
ELU as an activation function.

Section S4.3 Chemistry-aware NN

Input vector for this network consists of 4 parts: the first one describes proton-donating atom whilst the
rest characterize the neighbors. As previously, the proton-donating atom was described by four
features: Gasteiger charge of the neutral atom, Gasteiger charge of the atom in anionic form (after
proton dissociation), number of lone electron pairs, and the number of pi electrons on the atom. Each of
substituents was assigned with a nine-element vector:
 Gasteiger charge of the atom
 Difference between Pauling electronegativity of the atom and electronegativity of carbon
 number of lone pairs on the atom
 number of pi electrons
 distance to the nearest functional group with known Hammett constant as defined in S2
 F parameter describing field/induction effect of substituent as defined in S2
 R parameter describing resonance effect of substituent as defined in S2
 Hammett substituent constant σ for meta position
 Hammett substituent constant σ for para position
When proton-donating atom had less than three substituents (hydrogen was not counted as a
substituent) corresponding vector was filled with zeros.
After the network optimization with 5-fold cross-validation following hyper-parameters gave the
lowest MAE: droupout = 0.1, L2 regularization = 0.001, learning rate = 0.002, batch size = 5 and ELU
as an activation function.

S-10
Section S4.4 Graph convolutional neural network (GCNN)

Below, we include the details of the best GCNN architecture examined in the study:
 Input dense layer
256 hidden features, activation = ReLU, L2 regularization = 0.005, dropout = 0.2
 Graph convolution part
4 x 256 hidden features, local addition layers, activation = ReLU, L2 regularization=0.005,
dropout = 0.2
 Output dense layer
1 hidden feature, activation = ReLU, L2 regularization = 0.005

Table S8 conatins results of tests with architectures obtained by variation of the following parameters:
-number and type of Graph convolutional layers,
-number of hidden features,
-L2 regularization,
-dropout probability.

Table S8. Cross-validation results involving random splits.

Type of con- No. of convo- Hidden L2 regularization Drop CV MAE Test MAE
volutional lutional layers units out MAE difference
filter
4 5*10-3 0.2 2.07 ± 0.08 2.18 0.11
256
6 5*10-3 0.2 2.11 ± 0.06 2.23 0.12
6 10-3 0.1 2.3 ± 0.1 2.7 0.4
4 10-3 0.2 2.4 ± 0.1 2.9 0.5
Local addi- 6 10-5 0.1 2.4 ± 0.1 3.2 0.8
tion layers 4 10-5 0.1 2.4 ± 0.1 2.6 0.2
6 10-5 0.2 2.4 ± 0.1 2.4 0.0
6 10-3 0.2 2.5 ± 0.1 2.6 0.1
320
4 10-5 0.2 2.5 ± 0.1 2.7 0.2
4 10-3 0.1 2.5 ± 0.1 2.4 -0.1
Second or- 2 10-3 0.2 2.6 ± 0.1 2.8 0.2
der Cheby- 3 10-3 0.2 2.6 ± 0.1 2.7 0.1
shev poly-
nomial

In order to estimate the importance of features used to describe atoms in molecules, we performed an
‘elimination test’ for each descriptor. This procedure involved removal of the corresponding column
from atomic feature matrix and subsequent 5-fold cross-validation of the model with the same
configuration (that is, number of layers, hidden features, and regularization parameters). We decided to
keep the atomic number, hybridization and the total number of attached hydrogen atoms as the ‘base’
model, and to focus on the evaluation of remaining features (Figure S1).

S-11
Figure S1. Performance of the best GCNN architecture upon removal of input descriptors. Labels
on the y-axis describe models trained with the same architecture as the final model, but with a different
set of features used to describe the input data. For the sake of comparison, three other models are
presented as well: Base model = model including atomic number, hybridization and number of attached
hydrogen atoms; random descriptors = model with random vector assigned to each distinct chemical
element (see in the main text); full model = the best model discussed in the study. Blue bars represent
the results from five-fold cross-validation on 90% of data (with error bars indicating the standard error
of the resulting mean), whereas the orange series depict results obtained on the 10% held-out test set.

Section S5. Transfer learning experiment

We performed additional experiments to test whether a more robust model could be obtained with a
transfer learning techniqueS3. The idea is to initially train a model on a large dataset different from the
one considered here, and then to use its weights as a starting point in training our pKa model. We test
this approach with the QM9 datasetS4, which is one of popular benchmark sets for GCNNsS5. The set
contains ~134,000 small molecules (up to 9 heavy atoms) for which several quantum-mechanically
calculated quantities (HOMO/LUMO energy, dipole moment, polarizability, heat of formation, zero-
point energy, etc.) were reported at B3LYP/6-31G(2df,p) level of quantum chemistry. It seemed
possible that features useful in predicting these quantities would be as effective in prediction of pKa.
We found the presented architecture to perform well on this dataset (Table S9), comparably to models
referred in the MoleculeNet benchmarkS5.

Table S9. Mean Absolute Error comparison of our GCNN (GC) model and benchmark models
reported in MoleculeNet paperS5 on the QM9 datasetS4. Only quantities exhibiting correlation with
pKa of relevant compounds are shown. Minimum values in each row are indicated by bold font.

S-12
Quantity our GC benchmark GC benchmark DTNN benchmark MPNN
Dipole moment 0.475 0.583 0.244 0.358
HOMO 0.00384 0.00716 0.00388 0.00541
LUMO 0.00442 0.00921 0.00513 0.00623
Gap 0.0055 0.0112 0.0066 0.0082
R2 39.9 35.9 17 28.5
U0 2.32 3.41 2.43 2.05
ZPVE 0.00169 0.00299 0.00172 0.00216

Next, we transferred the weights form this model to the one modified for node regression (the
modification involves simple replacement of the last, node-gathering layer with ReLU unit) and trained
the whole neural net on the pKa values. Because QM9 set contains only second-row elements, we
restricted the pKa dataset to entries containing only H, C, N, O and F for consistency. Furthermore, we
trained two baseline models: (i) one with random initialization of weights and (ii) one with transferred
weights frozen during training (only parameters of the last ReLU unit were optimized). Results
presented in Figure S2 contradict the initial assumption about features learned by the model on the
QM9 dataset. First, the behavior of the model with frozen weights suggests that hidden features useful
in prediction of chosen ‘quantum’ properties can hardly be adapted to pKa prediction. Second, the
difference between model initialized randomly and the one with pre-trained weights seems negligible at
the end of training. Although we could not establish whether the information from QM9 is erased or
rebuilt during the first 10 epochs of training, this final convergence suggests that the transfer learning
attempt did not provide any improvement.

Figure S2. Results of transfer learning experiment. Curves present evolution of pKa mean
absolute error (calculated by 5-fold cross-validation with random splits) with training epochs.
‘Random initialization’ – GCNN model with randomly initialized weights, ‘Transfer learning’ –
GCNN model initialized with weights pre-trained on QM9, ‘Frozen weights’ – GCNN model
with fixed weights from QM9 model (optimization involves only last ReLU unit). Please note
that the X-axis is in logarithmic scale.
S-13
Section S6. Other factors influencing reaction outcomes.

Figure S3. Additional examples for the prediction of the second deprotonation site made available by
the use of two base equivalents and of a dianion. a) Experimentally observed outcomes of allylation of
cyclohexanone when one or two equivalents of base are usedS6-S7. b) Screenshots of the
https://pka.allchemy.net/ web-app illustrating correct prediction of the first (left) and the second (right)
deprotonation sites. c,d) Similar comparison for the alkylation of pyrrolidinone via dianion.S8 e,f)
Another example of alkylation of a cyclic ketoester from Corey’s synthesis of Desogestrel.S9

S-14
Figure S4. Deprotonation controlled by pre-coordination. a) The CH abstraction step (C D) occurs in
an intramolecular fashion after pre-coordination of the Lewis-acidic metal (here, lithium) with the
Lewis-basic directing group, DG. Additionally, the directing group stabilizes the obtained
organolithium compound before the reaction with an electrophile. b) Strong and c) moderate directing
groups controlling deprotonation. d) Initial steps in the total synthesis of Hydrangenol and
PhyllodulcinS10 are controlled by dimethylamide which is a stronger directing group than an aryl
methyl ether.

S-15
Figure S5. Selective functionalization of ketones via kinetic and thermodynamic enolates. a)
Hydroxymethylation of less acidic position performed during total synthesis of Jiadifenin S11 occurs via
more stable tetrasubstituted enolate. b) Depending of the conditions used, the same substrate may
deliver selectively one of the two possible products. For example, methylation of 2-
methycyclohexanone leads to α,α-dimethyl cyclohexanone when treated with NaH in THF and
equilibrated with 0.15 eq. of HMDS (in Baran’s synthesis of Maoecrystal V reported in S12) or α,α’–
dimethyl isomerS13 when deprotonated with LiHMDS and converted to Mn enolate prior to addition of
electrophile.

Section S7. Supplementary references

(S1) Frisch, M. J.; Trucks, G. W.; Schlegel, H. B.; Scuseria, G. E.; Robb, M. A.; Cheeseman, J.
R.; Scalmani, G.; Barone, V.; Mennucci, B.; Petersson, G. A.; Nakatsuji, H.; Caricato, M.; Li, X.;
Hratchian, H. P.; Izmaylov, A. F.; Bloino, J.; Zheng, G.; Sonnenberg, J. L.; Hada, M.; Ehara, M.;
Toyota, K.; Fukuda, R.; Hasegawa, J.; Ishida, M.; Nakajima, T.; Honda, Y.; Kitao, O.; Nakai, H.;
Vreven, T.; Montgomery, J. A., Jr.; Peralta, J. E.; Ogliaro, F.; Bearpark, M.; Heyd, J. J.; Brothers, E.;
Kudin, K. N.; Staroverov, V. N.; Kobayashi, R.; Normand, J.; Raghavachari, K.; Rendell, A.; Burant, J.
C.; Iyengar, S. S.; Tomasi, J.; Cossi, M.; Rega, N.; Millam, J. M.; Klene, M.; Knox, J. E.; Cross, J. B.;
Bakken, V.; Adamo, C.; Jaramillo, J.; Gomperts, R.; Stratmann, R. E.; Yazyev, O.; Austin, A. J.;
Cammi, R.; Pomelli, C.; Ochterski, J. W.; Martin, R. L.; Morokuma, K.; Zakrzewski, V. G.; Voth, G.
A.; Salvador, P.; Dannenberg, J. J.; Dapprich, S.; Daniels, A. D.; Farkas; ; Foresman, J. B.; Ortiz, J. V.;
Cioslowski, J.; Fox, D. J. Gaussian 09; Gaussian Inc.: Wallingford, CT, 2009.
(S2) Hansch, C.; Leo, A.; Taft, R. W. “A Survey of Hammett Substituent Constants and
Resonance and Filed Parameters. Chem. Rev. 1991, 91, 165-195.
(S3) Pan, S. J.; Yand, Q. A. Survey on Transfer Learning. IEE Trans. Knowl. Data End. 2010,
22, 1345-1359.
(S4) Ramakrishnan, R.; Dral, P. O.; Rupp, M.; Von Lilienfeld, O. A. Quantum chemistry
structures and properties of 134 kilo molecules. Scientific data 2014, 1,140022.
(S5) Wu, Z.; Ramsundar, B.; Feinberg, E. N.; Gomes, J.; Geniesse, C.; Pappu, A. S.; Leswingd
K; Pande V. MoleculeNet: a benchmark for molecular machine learning. Chem. Sci., 2018, 9, 513-530.

S-16
 (S6) Aburel, P. S.; Rømming, C.; Ma, K.; Undheim, K. Synthesis of α-Hydroxy and α-
Oxospiranes Through Ruthenium(II)-Catalyzed Ring-Closing Metathesis. J. Chem. Soc. Perkin Trans.
1 2001, 12, 1458–1472.
(S7) Xue, F.; Seto, C. T. A Comparison of Cyclohexanone and Tetrahydro-4H-Thiopyran-4-One
1,1-Dioxide as Pharmacophores for the Design of Peptide-Based Inhibitors of the Serine Protease
Plasmin. J. Org. Chem. 2005, 70, 8309–8321.
(S8) Davis, F. A.; Bowen, K. A.; Xu, H.; Velvadapu, V. Synthesis of Polysubstituted Pyrroles
from Sulfinimines (N-Sulfinyl Imines). Tetrahedron 2008, 64, 4174–4182.
(S9) Corey, E. J.; Huang, A. X. A Short Enantioselective Total Synthesis of the Third-
Generation Oral Contraceptive Desogestrel. J. Am. Chem. Soc. 1999, 121, 710–714.
(S10) Watanabe, M.; Sahara, M.; Kubo, M.; Furukawa, S.; Billedeau, R. J.; Snieckus, V. Ortho-
Lithiated Tertiary Benzamides. Chain Extension via o-Toluamide Anion and General Synthesis of
Isocoumarins Including Hydrangenol and Phyllodulcin. J. Org. Chem. 1984, 49, 742–747.
(S11) Carcache, D. A.; Cho, Y. S.; Hua, Z.; Tian, Y.; Li, Y.-M.; Danishefsky, S. J. Total
Synthesis of (±)-Jiadifenin and Studies Directed to Understanding Its SAR: Probing Mechanistic and
Stereochemical Issues in Palladium-Mediated Allylation of Enolate-Like Structures. J. Am. Chem. Soc.
2006, 128, 1016–1022.
(S12) Cernijenko, A.; Risgaard, R.; Baran, P. S. 11-Step Total Synthesis of (−)-Maoecrystal V. J.
Am. Chem. Soc. 2016, 138, 9425–9428.
(S13) Reetz, M. T.; Haning, H. α-Methylation of Ketones via Manganese-Enolates: Absence of
Undesired Polyalkylation. Tetrahedron Lett. 1993, 34, 7395–7398.

S-17