Lean-Docking: Exploiting Ligands’ Predicted Docking Scores to Accelerate Molecular Docking

Abstract

In structure-based virtual screening (SBVS), a binding site on a protein structure is used to search for ligands with favorable nonbonded interactions. Because it is computationally difficult, docking is time-consuming and any docking user will eventually encounter a chemical library that is too big to dock. This problem might arise because there is not enough computing power or because preparing and storing so many three-dimensional (3D) ligands requires too much space. In this study, however, we show that quality regressors can be trained to predict docking scores from molecular fingerprints. Although typical docking has a screening rate of less than one ligand per second on one CPU core, our regressors can predict about 5800 docking scores per second. This approach allows us to focus docking on the portion of a database that is predicted to have docking scores below a user-chosen threshold. Herein, usage examples are shown, where only 25% of a ligand database is docked, without any significant virtual screening performance loss. We call this method "lean-docking". To validate lean-docking, a massive docking campaign using several state-of-the-art docking software packages was undertaken on an unbiased data set, with only wet-lab tested active and inactive molecules. Although regressors allow the screening of a larger chemical space, even at a constant docking power, it is also clear that significant progress in the virtual screening power of docking scores is desirable.

Full text

Lean-Docking: Exploiting Ligands’ Predicted
Docking Scores to Accelerate Molecular Docking
Francois Berengera,∗,†Ashutosh Kumara,‡Kam Y. J. Zhang,‡and Yoshihiro
Yamanishi†
†Department of Bioscience and Bioinformatics, Faculty of Computer Science and Systems
Engineering, Kyushu Institute of Technology, 680-4 Kawazu, Iizuka, Japan
‡Laboratory for Structural Bioinformatics, Center for Biosystems Dynamics Research,
RIKEN, 1-7-22 Suehiro-cho, Tsurumi-ku, Yokohama, Kanagawa 230-0045, Japan
E-mail: beren314@bio.kyutech.ac.jp
aShared first author.
Abstract
In Structure-Based Virtual Screening (SBVS), a binding site on a protein structure
is used to search for ligands with favorable non-bonded interactions. Because it is
computationally hard, docking is time-consuming and any docking user will eventually
encounter a chemical library which is too big to dock. This problem might arise be-
cause there is not enough computing power, or because preparing and storing so many
3D ligands requires too much space. In this study, however, we show that quality
regressors can be trained to predict docking scores from molecular fingerprints. While
typical docking has a screening rate of less than one ligand per second on one CPU core,
our regressors can predict about 5800 docking scores per second. This approach allows
us to focus docking on the portion of a database which is predicted to have docking
scores below a user-chosen threshold. Here, usage examples are shown, where only 25%
1

of a ligand database is docked, without any significant virtual screening performance
loss. We call this method “lean-docking”. To validate lean-docking, a massive docking
campaign using several state-of-the-art docking software was undertaken on an unbi-
ased dataset, with only wet-lab tested active and inactive molecules. While regressors
allow the screening of a larger chemical space, even at constant docking power, it is
also clear that significant progress in the virtual screening power of docking scores is
desirable.
Introduction
The conformation and pose of a small molecule in a protein binding pocket can be predicted
by several methods including molecular docking,1–4 de novo design5,6 and molecular dynamics
simulation.7–9 Among these methods, molecular docking is one of the most popular and has
been successfully applied not only to predict the binding mode but also to rank-order a
library of small molecules for cherry picking in inhibitor discovery campaigns.10–14 Owing to
the utility of molecular docking in drug discovery, several improvements have been reported in
the last decade.15–19 Despite significant developments, routine docking of screening libraries
with more than a billion compounds is still beyond reach of most people, due to the huge
computing power and storage required. Moreover, most docking programs lack the speed and
efficiency necessary to execute ultra-large virtual screening campaigns. This resource gap is
further widening as the size of commercially available chemical space is increasing rapidly,20
due to the rise of automated synthesis and increased availability of diverse building blocks.
In this context, docking protocols that can quickly screen ultra-large chemical libraries in the
absence of large computing and storage capacity would be of immense use in drug discovery.
To speed-up molecular docking of large chemical libraries, different filtering approaches
based on physicochemical properties, drug-likeness, similarity to known inhibitors etc. have
been proposed. Although, these approaches may lead to faster screening of large compound
libraries, many interesting scaffolds may be filtered out. In order to fully exploit the po-
2

Table 1: Overview of methods to accelerate docking (in chronological order of publication).
Machine-learning method
Name Reference Classifier Regressor Algorithm Molecular encoding Validation dataset Docking software Claimed acceleration Remarks
Progressive Docking Cherkasov et. al. (2006) N/A PLS Iterative 58 molecular descriptors Glide 1.2x to 2.6x Maintains 80% to 99% of hit recovery rates
Spresso Yanagisawa et. al. (2017) N/A Sequential 102 proteins from DUD-E Glide HTVS 200x
Conformal Prediction Svenson et. al. (2017) N/A Iterative 97 molecular descriptors 41 proteins from DUD-E Glide 10.6x
VirtualFlow Gorgulla et. al. (2020) N/A N/A Distributed Ready-to-dock ligands
Deep Docking Gentile et. al. (2020) N/A Iterative FRED 50x
Lean Docking This work N/A Linear SVR Sequential
3 proteins: SHBG, SARS
3CLPro, HIV1 reverse
transcriptase
Generalized
Sum of
fragments’
docking scores
formula
Rigid molecular
fragments
Fragmenting the whole database to screen
and docking all unique fragments is required
Random forests
based conformal
predictor
Wet-lab assay to label training set molecules
required
Kelch-like ECH-associated
protein 1 (KEAP1)
Smina Vinardo,
AutoDock Vina
Scaling behaviour
linear in the number
of cores
Access to a cloud computing platform /
supercomputer required
Feed-forward Deep
Neural Network
1024 bits Morgan
fingerprints (radius=2)
12 proteins: ADORA2A,
AR, AT1R, CAMKK2,
CDK6, ERα, GABAA,
GLIC, Nav1.7, PPARγ,
TBXA2R, VEGFR2
Training set of 250,000 to 1M docked
molecules recommended
Counted atom pairs
fingerprint
15 proteins: LIT-PCBA
AVE-unbiased
Gold, Autodock-
Vina, MOE-Dock
(partial data),
Glide (partial
data)
4x to 41x without loss
of top-scoring actives
(depending on
docking VS
performance on given
protein); raw scoring
acceleration
compared to CCDC
Gold > 58000x
Training set of 10,000 docked molecules
recommended
tential of expanded chemical space by maximizing the number of evaluated compounds,
several docking reduction approaches have been proposed (Table 1). In “Progressive Dock-
ing”,21 QSAR models predicting Glide docking scores are used iteratively to avoid docking
molecules predicted as non-binders. The authors report an acceleration between 1.2 and 2.6
times, while maintaining 80 to 100% hit recovery rates. Svensson et al. 22 and colleagues use
a conformal predictor classifier trained on a small set of assayed molecules. The authors ret-
rospectively validate their protocol on DUD-E,23 showing that 57% of the remaining actives
could be identified while screening only 9.4% of the remaining database (acceleration of 10.6
times). The inspirational “Deep Docking” approach24 (DD) uses deep-learning classifiers to
predict ligands with a low docking score and reduce the number of docking calculations. DD
was used to screen 1.36 billion molecules from the ZINC1525 database and its authors re-
ported an acceleration of 50 times compared to a classical docking campaign. Another work
at accelerating docking, due to Yanagisawa and colleagues,26 used docked ligand fragments
to pre-screen ligands that will require docking. They reported a 200 times acceleration com-
pared to standard docking. Another interesting approach to accelerate docking is distributed
computing. Gorgulla et al. 27 and colleagues created the VirtualFlow open-source program
for this purpose and screened more than one billion commercially-available molecules using
Autodock-Vina28 and Smina Vinardo.29 They eventually identified submicromolar protein-
protein interaction inhibitors for nuclear factor erythroid-derived 2-related factor 2 (NRF2)
3

and Kelch-like ECH-associated protein 1 (KEAP1). They report an acceleration linear in
the number of cores used during the distributed computation. There are good reviews about
ligand-based virtual screening30 or the combination of ligand- and structure-based meth-
ods.31,32 Jastrzębski et al. used an edge-attention graph convolutional network33,34 to train
a regressor of docking scores for the GLIDE docking software on the human protein 5-HT1B
(PDB:4IAQ; R2∼= 0.68). Some authors33,35 even manage to predict the binding mode of 2D
ligands. A docking score takes into account calculated non-covalent 3D interactions between
a ligand and a protein. Regression models can be trained to capture a mapping from 2D
ligands ato docking scores (for a given protein binding site and docking software). While this
might be counter-intuitive, there are publications indicating that predicting docking scores
from 2D ligands is possible.21,33,36 Related, but not exactly about docking, a random-forest
using only 2D ligand-based features could predict the mean binding affinity of a ligand for
its protein targets.36 Lyu and colleagues37 have docked 170 million make-on-demand com-
pounds against AmpC β-lactamase and the D4dopamine receptor. Their study provides key
insights about docking scores in a large library screen, with hundreds of molecules ordered
and tested after the docking screen, so that the total number of actives in a library could be
estimated by a statistical model. Cieplinski and colleagues38 have tried to predict docking
scores for the SMINA docking software39 on a few protein targets; another example showing
that this task is deemed useful.
In this article, three significant contributions are presented. i) A massive docking cam-
paign using several state-of-the-art software, on the recently published LIT-PCBA dataset.40
ii) How to train quality regressors of docking scores for 15 protein targets and four docking
software. Such regressors can score about 5800 ligands per second on a single CPU core. iii)
The lean-docking protocol to accelerate docking by several folds, with a low risk of loosing
top-scoring active molecules. However, the utility of this protocol ultimately depends on the
virtual screening power of docking scores on a given protein target.
aStrictly speaking and as noted by a reviewer, a molecular graph without spatial coordinates is 1D.
4

Methods
Lean-docking: docking only a fraction of a large chemical database
Figure 1: Lean-docking protocol overview.
In order to exploit fast predictors of docking scores, the lean-docking protocol is pro-
posed (Figure 1). Since training a regressor requires docking a partition of up to 10,000
ligands randomly selected from the database to screen, the first quartile docking score from
those ligands can be used as a rather conservative threshold to separate promising molecules
(predicted scores under the thresholdb) from the rest (molecules predicted to score above
the threshold). Such a protocol allows to divide several times the required docking power,
while it can have a negligible impact on the number of active molecules per N top-scoring
molecules.
bDocking scores are negative.
5

The LIT-PCBA dataset
Table 2: Number of PDB structures, molecules and random hit-rate for the 15 protein targets
of the AVE-unbiased LIT-PCBA dataset.
Training Validation
Protein PDBs Actives Inactives Random Actives Inactives Random
hit-rate hit-rate
ADRB2 8 13 234,363 0.00006 4 78,120 0.00005
ALDH1 8 4,032 77,606 0.04939 1,344 25,868 0.04939
ESR1+ 15 10 4,188 0.00238 3 1,395 0.00215
ESR1- 15 77 3,711 0.02033 25 1,237 0.01981
FEN1 1 277 266,552 0.00104 92 88,850 0.00103
GBA 6 125 222,039 0.00056 41 74,013 0.00055
IDH1 14 30 271,537 0.00011 9 90,512 0.00010
KAT2A 3 146 261,411 0.00056 48 87,137 0.00055
MAPK1 15 231 46,972 0.00489 77 15,657 0.00489
MTORC1 11 73 24,729 0.00294 24 8,243 0.00290
OPRK1 1 18 202,362 0.00009 6 67,454 0.00009
PKM2 9 410 184,143 0.00222 136 61,380 0.00221
PPARG 15 21 3,909 0.00534 6 1,302 0.00459
TP53 6 60 3,126 0.01883 19 1,042 0.01791
VDR 2 498 199,906 0.00248 165 66,635 0.00247
Docking experiments were conducted on the 15 protein targets of the LIT-PCBA dataset40
(Laboratoire d’Innovation Thérapeutique - PubChem Assays dataset), in its Asymmetric
Validation Embedding (AVE41) unbiased version (Table 2). An unbiasing procedure was
used to split protein target ligands into a training and a validation set. LIT-PCBA was
designed to benchmark ligand and structure-based virtual screening methods. It was care-
fully curated from 149 PubChem42 assays. It contains 7844 confirmed actives and 407,381
confirmed inactive molecules.
Rigid-protein flexible-ligand docking
Docking was performed using CCDC Gold,43 Autodock-Vina,28 FRED,44 Glide45 and MOE
(Chemical Computing Group). Since LIT-PCBA contains multiple receptor PDBs per tar-
get, Gold, Autodock-Vina and FRED docking was performed on each PDB and the lowest
6

scoring pose (among all PDBs of a given protein) was retained. Due to the availability
of limited license tokens, Glide and MOE docking was performed for 6 out of 15 targets
(ESR1+, ESR1-, MAPK1, MTORC1, PPARG and TP53) using only the highest resolution
PDB. First, the lowest energy 3D conformer of each ligand was generated with OpenEye
OMEGA.46 For each docking program, ligands and receptors for the dataset proteins were
prepared according to standard practice as mentioned below. Binding site for each dock-
ing method was defined using the co-crystallized ligand of receptor PDB in the dataset.
Docking with Gold. Ligands were further prepared with the OpenEye OEChem toolkit.
Hydrogens were added and charges were assigned using AM1BCC.47,48 Receptors were pre-
pared with UCSF Chimera49 v1.14. Hydrogens were added and partial charges were assigned
using AMBER ff14SB.50 Docking was performed using the default scoring function and evo-
lutionary parameters. A maximum of 10 poses per ligand were generated and the lowest
scoring one was retained.
Docking with Autodock-Vina. Ligands and receptors were prepared using the protein
and ligand preparation scripts from AutoDock Tools51 v1.5.4. Receptors were prepared by
adding hydrogens, assigning Gasteiger52 partial charges before saving coordinates in the
PDBQT format. Ligands were prepared using the same protocol. The search space was
restricted to a cubic box with 25Å side-length. Docking was performed using the default
search exhaustiveness (eight). Ten poses maximum per ligand were generated and the lowest
scoring one was retained.
Docking with FRED. Conformational ensemble for FRED docking was generated using
OMEGA. A maximum of 200 conformations per compound were generated. Receptors for
FRED docking were prepared using Spruce4Docking utility program (OpenEye Scientific
Software). Ligands were scored using Chemgauss4 scoring function and a single pose per
compound was saved.
Docking with Glide. The protein structures for Glide molecular docking were prepared
using Maestro (Schrodinger Inc.) by adding hydrogens and assigning protonation states of
7

charged residues at neutral pH. Ligands were prepared using LigPrep (Schrodinger Inc.).
Tautomeric and ionization states of all ligands were determined using Epik.53 Molecular
docking was performed using the standard precision mode of Glide. Ligands were scored
using Glidescore with Epik penalties and a single pose per compound was saved.
Docking with MOE-Dock. The protein structures for MOE-Dock docking were prepared
using QuickPrep utility in MOE. Molecular docking was performed using rigid receptor
docking protocol. Ligand conformations were generated on-the-fly and were placed into the
binding pocket using Triangle Matcher placement method which were then scored using Lon-
don dG scoring function. Thirty poses were retained for each compound and were subjected
to rigid receptor refinement. The poses were energy minimized in the binding pocket and
were scored using GBVI/WSA dG scoring function. Finally, a single pose per compound
was saved.
A regressor to predict docking scores
Ligands were docked using the previously described protocol. To train a regressor for a given
protein target, a random partition of at most 10,000 ligands was drawn from all the training
set ligands docked on this target. If a target’s training set had less than 10k molecules,
all of them were used to train the regressor. Prior to regression modeling, 2D ligands were
standardized using Francis Atkinson’s standardizer.54 After standardization, ligands were
encoded using unfolded counted atom pairs (molenc software55) using the ChEMBL2456
feature dictionary (17,368 features). Only heavy atoms of a molecule are considered. An
atom pair consists of two atom types and their shortest distance (in bonds) on the molec-
ular graph. Distances go from zero and up to the graph diameter of a molecule. An atom
type is the tuple made of its number of pi electrons, atomic symbol, number of adjacent
(bonded) heavy atoms and formal charge. Stereochemistry is ignored. Because such a fin-
gerprint (a sparse vector of positive integers) is high dimensional and a simple but fast
regressor was wanted, the L2-regularized linear Support Vector Regressor (SVR) from lib-
8

linear57 was used (linwrap58 software). Only the Cparameter was optimized using 10 folds
cross-validation; leaving the parameter to its default value of zero. Cwas chosen in the set
{0.001,0.002,0.005,0.01,0.02,0.05,0.1,0.2,0.5,1,2,5,10,20,50}. It is not claimed that this
molecular fingerprint and linear-SVR combination is the best possible regressor. However,
it was good enough in practice. The test computer runs Ubuntu Linux on a 2.10GHz Intel
Xeon CPU. It takes about 80 seconds to train a regressor on a single core (40s on 16 cores;
scaling is not linear). A trained model can predict about 5800 docking scores per second on a
single core. AUC values and ROC curves were computed by the croc-curve Python script.59
R2and BEDROC values were computed by the Classification and Regression Performance
Metrics library.60
Results
Training regressors to predict docking scores
Our results suggest that it is possible to train quality regressors of docking scores for all
protein targets of the dataset (Figure 2). High accuracy was achieved for almost all targets
with a validation R2>= 0.78. The only exception is OPRK1, with TR2 = VR2 = 0.65. The
average validation R2was found to be 0.85, which suggests that good regressors were trained
irrespective of the protein target. A vast majority of the predictions are within +/-10 of
their actual Gold scores. For example, it is the case for 97% of the predictions on the largest
validation set (IDH1; Figure 3 and Supporting Information Figure S1 for all targets). After
inspecting Figures 3 and S1, we don’t see an easy and natural way to define an applicability
domain for the regressors. Interested readers might look into the recent literature for some
pointers.61 Moreover, the validation set R2being very close to the training set one indicates
that models are not overfitting the training set. The biggest difference between TR2 and
VR2 is 0.02 only, on the TP53 target. Also, note that the size of the regressor’s training
set used was chosen to be 10,000 molecules maximum (Figure 4), which means that most
9

Figure 2: Regressors trained using 10 folds cross validation on a random partition of at
most 10,000 molecules from the training set of each protein target (T:10xCV, green circles),
then used to predict the validation set (V, purple circles). The protein target name, best C
parameter, training and validation set R2(TR2 and VR2) are shown at the bottom right of
each plot. Lines of best fit are also shown.
10

IDH1
0 0.2 0.4 0.6 0.8 1
Tanimoto distance to nearest in training set
-30
-20
-10
0
10
20
30
Gold score prediction error
1
10
100
Figure 3: Docking score prediction error on the validation set of IDH1 (largest dataset).
Density scatter plot for the actual minus predicted Gold docking score for all molecules in
the validation set, as a function of the Tanimoto distance to the nearest molecule in the
training set. Plots for all targets can be seen in Supporting Information Figure S1.
0
0.2
0.4
0.6
0.8
1
100 200 500 1k 2k 5k 10k 20k
10 Folds Cross Validation R2 +/- σ
Training Set Size
Figure 4: Performance of the regressor as a function of the training set size in 10 folds
cross-validation experiments over all protein targets. The regressor training set of size N (N
∈[100,200,500,1k,2k,5k,10k,20k]) was drawn randomly in each experiment. If a target has
only M < N docked molecules, then only M molecules are used to train the regressor.
11

regressor training sets are several times smaller than the whole available training set (e.g.
ADRB2, ALDH1, FEN1, GBA, IDH1, KAT2A, MAPK1, MTORC1, OPRK1, PKM2, VDR).
The power of such regressors to retrieve top-scoring molecules in docking can be seen in
Supporting Information Figure S2.
The virtual screening power of predicted docking scores
Table 3: Performance of Gold docking and corresponding regressors on the LIT-PCBA
validation set. Interesting cells have a green background (AUC >= 0.6,BEDROC >= 0.1,
EF >= 2 or R2>= 0.8).
V set Gold Docking QSAR
Target AUC BEDROC AUC BEDROC
ADRB2 0.246 0.000 0.00 0.00 0.00 0.244 0.000 0.00 0.00 0.00 0.88 0.89
ALDH1 0.584 0.126 1.49 1.83 1.76 0.582 0.125 2.02 1.75 1.78 0.93 0.93
ESR1+ 0.523 0.046 0.00 0.00 0.00 0.525 0.033 0.00 0.00 0.00 0.93 0.92
ESR1- 0.580 0.032 0.00 0.00 0.00 0.566 0.035 0.00 0.00 0.00 0.96 0.96
FEN1 0.370 0.041 0.00 0.00 1.52 0.360 0.037 0.00 1.09 0.65 0.78 0.79
GBA 0.631 0.164 9.76 7.32 2.93 0.633 0.166 7.32 6.10 3.90 0.85 0.86
IDH1 0.628 0.233 11.11 11.11 4.44 0.653 0.216 11.11 11.11 4.44 0.82 0.83
KAT2A 0.426 0.060 2.08 1.04 1.25 0.451 0.059 2.08 1.04 1.25 0.80 0.80
MAPK1 0.625 0.084 1.31 0.66 1.58 0.634 0.099 2.63 1.98 1.84 0.88 0.88
MTORC1 0.511 0.008 0.00 0.00 0.00 0.556 0.035 0.00 0.00 0.00 0.84 0.85
OPRK1 0.715 0.177 0.00 0.00 3.33 0.724 0.201 0.00 8.33 3.33 0.65 0.65
PKM2 0.548 0.064 2.21 1.47 1.18 0.519 0.048 0.00 1.10 0.88 0.81 0.80
PPARG 0.709 0.173 0.00 8.18 3.32 0.721 0.172 0.00 8.18 3.32 0.95 0.95
TP53 0.722 0.225 5.79 2.76 3.34 0.707 0.180 0.00 2.76 3.34 0.89 0.87
VDR 0.365 0.011 0.00 0.00 0.12 0.360 0.015 0.62 0.62 0.37 0.80 0.80
avg 0.546 0.096 2.25 2.29 1.65 0.549 0.095 1.72 2.94 1.67 0.85 0.85
stddev 0.136 0.077 3.55 3.46 1.45 0.138 0.072 3.15 3.52 1.54 0.08 0.08
median 0.580 0.064 0.00 0.66 1.52 0.566 0.059 0.00 1.10 1.25 0.85 0.86
EF1% EF2% EF5% EF1% EF2% EF5% R2
train R2
valid
Predicted docking scores possess some of the virtual screening power of actual dock-
ing scores. While this is not the intended use of the models (promising molecules should
be docked in the end), it shows that when docking works, the corresponding predictor of
docking scores also works (Figures 2 and 5). Figure 5 clearly shows ROC curves with sim-
ilar trend when comparing the docking curve (in green) to the corresponding regressor (in
blue). Our proposed protocol is to first screen with ligand-based models, because they are
so fast. However, at the end of a structure-based virtual screening pipeline, ligands should
12

Figure 5: ROC curves for CCDC Gold (green line) or the corresponding regressor (blue line)
on the LIT-PCBA validation set; example legend at bottom right.
13

be available as protein-ligand complexes, so that other structure-based filtering is possible
(rescoring,62–66 pharmacophore filter,67 molecular dynamics simulation after docking,68 etc.).
Also, experiments show that docking usually has a better very early enrichment (EF1%) than
the corresponding regressor (Table 3).
Docking only 25% of a large chemical library
Table 4: Number of active molecules found among the N top-scoring molecules (LIT-PCBA
validation set; N in {100, 200, 300, ...}). #ligands: number of ligands to dock depending on
the protocol used; #err: number of ligands which failed in docking or LBVS preparation;
lean-docking 25% (lean) or classical docking (docked). Threshold: docking score threshold
used by lean-docking. N/A appears when there are not enough molecules to compare sets of
equal sizes. Tani: Tanimoto score between the sets of actives retrieved by the two methods
(NaN means “Not a Number” / impossible to compute).
target method #ligands #err threshold 100 200 300 500 1000 2000 3000 5000
ADRB2 lean 19683 161 -80.83 0 0 0 0 0 0 0 0
docked 77963 0 NaN 0 NaN 0 NaN 0 NaN 0 NaN 0 NaN 0 NaN 0 NaN
ALDH1 lean 6951 96 -80.29 8 18 21 42 86 167 223 352
docked 27116 8 1.00 18 1.00 21 1.00 42 1.00 86 1.00 167 1.00 222 0.96 345 0.91
ESR1+ lean 337 33 -64.77 0 1 1 N/A N/A N/A N/A N/A
docked 1365 0 NaN 1 1.00 1 1.00 N/A NaN N/A NaN N/A NaN N/A NaN N/A NaN
ESR1- lean 345 32 -65.03 1 3 7 N/A N/A N/A N/A N/A
docked 1230 1 1.00 3 1.00 7 1.00 N/A NaN N/A NaN N/A NaN N/A NaN N/A NaN
FEN1 lean 20879 2547 -53.55 0 0 0 0 0 0 4 7
docked 86395 0 NaN 0 NaN 0 NaN 0 NaN 0 NaN 0 NaN 3 0.75 7 1.00
GBA lean 18498 153 -63.72 2 2 2 3 5 6 6 8
docked 73901 2 1.00 2 1.00 2 1.00 3 1.00 5 1.00 6 1.00 6 1.00 8 1.00
IDH1 lean 23039 358 -79.71 0 0 0 0 2 2 2 3
docked 90163 0 NaN 0 NaN 0 NaN 0 NaN 2 1.00 2 1.00 2 1.00 2 0.67
KAT2A lean 20341 215 -71.14 0 1 1 1 1 2 3 4
docked 86970 0 NaN 1 1.00 1 1.00 1 1.00 1 1.00 2 1.00 3 1.00 4 1.00
MAPK1 lean 3760 72 -72.66 0 1 1 4 8 20 24 N/A
docked 15662 0 NaN 1 1.00 1 1.00 4 1.00 9 0.89 19 0.86 23 0.88 N/A NaN
MTORC1 lean 1979 0-91.96 0 0 0 0 1 N/A N/A N/A
docked 8267 0 NaN 0 NaN 0 NaN 0 NaN 1 1.00 N/A NaN N/A NaN N/A NaN
OPRK1 lean 14704 81 -71.37 0 0 0 0 0 0 2 3
docked 67379 0 NaN 0 NaN 0 NaN 0 NaN 0 NaN 0 NaN 0 0.00 2 0.67
PKM2 lean 15544 77 -93.44 1 1 1 3 4 4 8 13
docked 61439 1 1.00 1 1.00 1 1.00 3 1.00 4 1.00 4 1.00 8 1.00 14 0.93
PPARG lean 314 32 -65.53 2 2 3 N/A N/A N/A N/A N/A
docked 1276 2 1.00 2 1.00 3 1.00 N/A NaN N/A NaN N/A NaN N/A NaN N/A NaN
TP53 lean 299 18 -64.27 8 9 N/A N/A N/A N/A N/A N/A
docked 1043 8 1.00 9 1.00 N/A NaN N/A NaN N/A NaN N/A NaN N/A NaN N/A NaN
VDR lean 15463 154 -81.55 0 0 0 0 0 1 1 2
docked 66646 0 NaN 0 NaN 0 NaN 0 NaN 0 NaN 1 1.00 1 1.00 3 0.67
average
lean
10809.1 268.6 -73.3 1.5 1.0 2.5 1.0 2.6 1.0 4.8 1.0 9.7 1.0 20.2 1.0 27.3 0.8 43.6 0.9
stddev 8645.9 615.5 10.8 2.7 0.0 4.7 0.0 5.4 0.0 11.8 0.0 24.2 0.0 49.3 0.0 65.6 0.3 109.1 0.1
median 14704.0 81.0 -71.4 0.0 1.0 1.0 1.0 1.0 1.0 0.0 1.0 1.0 1.0 2.0 1.0 3.5 1.0 4.0 0.9
average
docked
44454.3 268.6 N/A 1.5 1.0 2.5 1.0 2.6 1.0 4.8 1.0 9.8 1.0 20.1 1.0 26.8 0.8 42.8 0.9
stddev 35464.0 615.5 N/A 2.7 0.0 4.7 0.0 5.4 0.0 11.8 0.0 24.2 0.0 49.3 0.0 65.4 0.3 106.9 0.1
median 61439.0 81.0 N/A 0.0 1.0 1.0 1.0 1.0 1.0 0.0 1.0 1.0 1.0 2.0 1.0 3.0 1.0 4.0 0.9
Tani100 Tani200 Tani300 Tani500 Tani1000 Tani2000 Tani3000 Tani5000
In Table 4, the effect of using the proposed protocol to dock only 25% of a chemical library
is shown. A docking score threshold was extracted from the first quartile of the regressor
training set (a maximum of 10,000 randomly-chosen molecules with docking scores). In
14

lean-docking, while all molecules have their docking scores predicted by the ligand-based
regressor, only molecules with a predicted docking score less than the threshold will be
docked. The number of active molecules among the N top-scoring molecules was counted (N
in {100, 200, 300, ...}). While lean-docking allowed to dock about four times less molecules
compared to classical docking, there was no significant impact on the number of true actives
among top-scoring docked molecules.
Discussion
The AVE-unbiased LIT-PCBA dataset. The docking campaign on this dataset unveiled
that it is a very hard dataset for docking. However, the authors of LIT-PCBA also noticed40
that their docking campaign only had a median EF1% >2on five targets (ADRB2, FEN1,
GBA, OPRK1, PPARG), out of the final 15 protein targets. Another drawback of this
dataset is that several protein targets have less than 10 actives in the validation set (ADRB2:
4, ESR1+: 3, IDH1: 9, OPRK1: 6, PPARG: 6); which creates awkward ROC curves (cf.
those five targets in Figure 5). Also, most targets have several PDBs (Table 2), which
make an exhaustive docking campaign (as was undertaken here) very costly computationally.
For CCDC Gold, more than 15 ∗106ligands have been docked in total. This required an
approximate 40,000 hours of supercomputer time, and approximately three times more for
AutoDock-Vina.
Advantages. Predicting docking scores using a trained regressor is significantly faster
than a real docking screen. We estimate the raw scoring speed to be about 58,000 times
faster than CCDC Gold. Lean-docking allows to screen virtual chemical libraries whose size
is far beyond the computational power or storage capacity of a standard docking user. In our
experiments, one only needs to dock a rather small set of 10,000 ligands on a given target
binding-site in order to train a regressor. We did not observe a significant improvement of the
regressor’s R2performance if 20,000 docked ligands were used (Figure 4). Using a docking
15

score threshold, this regressor can be used to select a smaller portion of the database for
actual docking. Compared to recent methods using deep-learning to accelerate docking,24,33
lean-docking requires a rather small training set and a regressor can be optimized and trained
in less than two minutes on a single core of our test computer. Hence, lean-docking should
be particularly applicable in situations where users only have access to limited computing
power.
Drawbacks. The method we propose does not predict the binding mode of a ligand;
only a docking score. See Chupakhin et al.35 or Jastrzębski et al. 33 for methods that predict
a binding mode. In our experience, not all docking programs and protein binding sites are
amenable to regression modeling. For example, we tried but failed for OpenEye FRED
3.5.0.444 (Supporting Information Figure S3). On all but one protein target (MTORC1, 10
folds cross validation training set R2=0.17 only), the best regressor has a negative R2. The
Chemgauss4 scoring function of FRED depends heavily on 3D information.69,70 This score
might be impossible to predict by a regressor and from a molecular graph alone. FRED’s
scores might be predictable for a given binding site from a ligand 3D conformer (associated
with a 3D-sensitive molecular encoding). However, having to prepare conformers of the huge
library one wants to screen would lessen the practicality of lean-docking. For MOE-Dock
(Chemical Computing Group), although we have only partial data (Supporting Information
Figure S4), very good regressors were obtained for five out of six targets (ESR1-, MAPK1,
MTORC1, PPARG and TP3) and an acceptable regressor for ESR1+. For Schrodinger Glide
Release 2018.3,45 we also have partial data (Supporting Information Figure S5). On four
(ESR1+, ESR1-, PPARG, TP53) of the six targets analyzed, it seems that it is possible to get
an acceptable regressor. Although we were able to see some trend for MAPK1 and MTORC1
trend, adequate regressors were not obtained probably due to our partial docking protocol
(using the highest resolution PDB only). MAPK1 is a challenging target for docking due to
the activation loop dynamics and the selection of a suitable receptor conformation is required
owing to the conformation selective nature of its inhibitors. MTORC1 might be difficult
16

under the partial docking protocol due to its large binding pocket. In an unrelated project,
it was found that CCDC Gold has a tendency to give very low docking scores to molecules
which are rather big, highly flexible and hydrophobic. Unfortunately, a regressor can capture
this tendency. Since finding such molecules is not the goal of a docking campaign, it was
found that filtering out PAINS71 compounds from the database that was screened removed
most of those unwanted molecules.
0
1000
2000
3000
4000
5000
6000
7000
-100 -90 -80 -70 -60 -50 -40 -30 -20 -10
Target protein: GBA
1 2 3
Number of molecules
Gold docking score
Classical docking
Lean-docking
Figure 6: Distribution of docking scores for docking versus lean-docking 25% on the GBA
target (validation set). Green arrows: actives docked upon lean-docking. Blue arrows:
actives whose docking was skipped because of lean-docking.
Is it possible to accelerate more than four times? The previously mentioned
predicted docking score threshold must not be set too low, unless the user knows the docking
scores of several known inhibitors of the protein target, under a given docking protocol. In
the experiments on the whole LIT-PCBA dataset, the first quartile docking score was used to
divide by four the required docking power for a given chemical library; a rather conservative
choice. Even in a case where docking works, 25% can skip the docking of some known actives
(Figure 6). However, for a protein target where docking works well, and where there are
several known actives, it is possible to accelerate much more. For example, on the ALDH1
target, docking can be accelerated up to 41 times, using a threshold docking score of -94.74
(computed on the regressor’s left-out training set). In this specific case, also with no loss in
17

terms of active molecules found among the top-scoring molecules upon docking the validation
set.
Conclusion
In this study, a massive docking campaign was undertaken on the LIT-PCBA dataset, using
several docking software. Regressors could be trained to predict docking scores from 2D
ligands, on most protein targets and with good accuracy (average training set 10 folds cross
validation R2= 0.851; validation set R2= 0.852). We also show a use case of such predicted
docking scores, to significantly reduce the docking power required to screen a large chemical
library, without any significant impact on the virtual screening power (in terms of number
of actives among top-scoring docked molecules).
The authors are interested to know if users of the protocol encounter docking programs
for which it seems a regressor cannot be trained for a given protein target, or confirmation of
the opposite event. The authors are especially interested in open-source docking programs
or with a free license for academia. In our hands, CCDC Gold, Autodock-Vina (Supporting
Information Figure S6) and MOE-Dock (Supporting Information Figure S4) were amenable
to regression modeling, on several protein targets.
There might be other creative use-cases for accurate but fast prediction of docking scores
from 2D ligands. For example, to help in the computational design of ligands with intended
polypharmacology, or to obtain a cheap docking oracle.72 However, the biggest step forward
would be to improve the virtual screening power of docking scores.
Acronyms
AM1BCC: AM1 semi-empirical quantum mechanical wave-function with Bond-Charge Cor-
rections; AMBER: molecular dynamics package; AUC: Area Under the receiver operating
characteristic Curve; BEDROC: Boltzmann-enhanced receiver operating characteristic curve
18

(early recovery biased ROC curve); AVE: Asymmetric Validation Embedding;41 CCDC:
Cambridge Crystallographic Data Centre; ChEMBL: Chemistry database of the European
Molecular Biology Laboratory; CPU: Central Processing Unit; DUD-E: Directory of Useful
Decoys, Enhanced;23 DD: Deep Docking;24 EF: Enrichment Factor; ff14SB: a force field;50
LIT-PCBA: Laboratoire d’Innovation Thérapeutique - PubChem Assays dataset;40 MOE:
Molecular Operating Environment; PAINS: Pan Assay Interference Compounds; PDB: Pro-
tein Data Bank; PDBQT: file format for docking software (atomic coordinates, partial
charges and atom types); ROC: Receiver Operating Characteristic; R2: Coefficient of De-
termination; SMINA: docking software39 fork of Autodock-Vina; SVR: Support Vector Re-
gression; TR2: Test R2; UCSF: University of California San-Francisco; VR2: Validation R2;
ZINC: ZINC Is Not Commercial (free online database of commercially-available compounds
for virtual screening).
Data and Software Availability
Dataset (SMILES, molecular fingerprints and docking scores): https://zenodo.org/record/4588239
(accessed 2021-03-08). Molecular standardizer: https://github.com/flatkinson/standardiser (ac-
cessed 2019-06-01; pip package chemo-standardizer). Molecular encoder: https://github.com/
UnixJunkie/molenc (accessed 2021-01-25). Software to train regressors to predict docking
scores: https://github.com/UnixJunkie/linwrap (accessed 2021-01-25).
Supporting Information
Supporting Information Available: Density scatter plots for Gold regressors (Figure S1),
recovery plots for Gold (Figure S2), regression plot for FRED (Figure S3), regression plots for
MOE-Dock (Figure S4), regression plots for Glide (Figure S5), regression plots for Autodock-
Vina (Figure S6).
19

Acknowledgement
This work was supported by JST AIP-PRISM [grant number JPMJCR18Y5] and JSPS
KAKENHI [grant numbers 18H03334 and 18H02395]. We acknowledge RIKEN ACCC for
computing resources on the Hokusai BigWaterfall supercomputer. FB acknowledges the use
of ChemAxon JChem 20.13 http://www.chemaxon.com (accessed 2020-11-25). We thank Dr.
Andrew Robertson from Kyushu university for improving the abstract.
Competing interests
The authors declare no competing financial interest.
References
(1) Yuriev, E.; Agostino, M.; Ramsland, P. A. Challenges and Advances in Computational Docking: 2009
in Review. J. Mol. Recognit. 2011,24, 149–164.
(2) Yuriev, E.; Ramsland, P. A. Latest Developments in Molecular Docking: 2010-2011 in Review. J. Mol.
Recognit. 2013,26, 215–239.
(3) Yuriev, E.; Holien, J.; Ramsland, P. A. Improvements, Trends, and New Ideas in Molecular Docking:
2012-2013 in review. J. Mol. Recognit. 2015,28, 581–604.
(4) Irwin, J. J.; Shoichet, B. K. Docking Screens for Novel Ligands Conferring New Biology. J. Med. Chem.
2016,59, 4103–4120.
(5) Kutchukian, P. S.; Shakhnovich, E. I. De Novo Design: Balancing Novelty and Confined Chemical
Space. Expert Opin. Drug. Discov. 2010,5, 789–812.
(6) Loving, K.; Alberts, I.; Sherman, W. Computational Approaches for Fragment-based and de Novo
Design. Curr. Top. Med. Chem. 2010,10, 14–32.
(7) Dror, R. O.; Pan, A. C.; Arlow, D. H.; Borhani, D. W.; Maragakis, P.; Shan, Y.; Xu, H.; Shaw, D. E.
Pathway and Mechanism of Drug Binding to G-protein-coupled Receptors. Proc. Natl. Acad. Sci. U S
A2011,108, 13118–13123.
20

(8) Buch, I.; Giorgino, T.; De Fabritiis, G. Complete Reconstruction of an Enzyme-inhibitor Binding
Process by Molecular Dynamics Simulations. Proc. Natl. Acad. Sci. U S A 2011,108, 10184–10189.
(9) Dror, R. O.; Green, H. F.; Valant, C.; Borhani, D. W.; Valcourt, J. R.; Pan, A. C.; Arlow, D. H.;
Canals, M.; Lane, J. R.; Rahmani, R.; Baell, J. B.; Sexton, P. M.; Christopoulos, A.; Shaw, D. E.
Structural Basis for Modulation of a G-protein-coupled Receptor by Allosteric Drugs. Nature 2013,
503, 295–299.
(10) Kellenberger, E.; Springael, J.-Y.; Parmentier, M.; Hachet-Haas, M.; Galzi, J.-L.; Rognan, D. Identifi-
cation of Nonpeptide CCR5 Receptor Agonists by Structure-based Virtual Screening. J. Med. Chem.
2007,50, 1294–1303.
(11) Kumar, A.; Ito, A.; Takemoto, M.; Yoshida, M.; Zhang, K. Y. J. Identification of 1,2,5-Oxadiazoles
as a New Class of SENP2 Inhibitors Using Structure Based Virtual Screening. J. Chem. Inf. Model.
2014,54, 870–880.
(12) Jiang, X.; Kumar, A.; Liu, T.; Zhang, K. Y. J.; Yang, Q. A Novel Scaffold for Developing Specific or
Broad-Spectrum Chitinase Inhibitors. J. Chem. Inf. Model. 2016,56, 2413–2420.
(13) Matsuoka, M.; Kumar, A.; Muddassar, M.; Matsuyama, A.; Yoshida, M.; Zhang, K. Y. J. Discovery of
Fungal Denitrification Inhibitors by Targeting Copper Nitrite Reductase from Fusarium oxysporum. J.
Chem. Inf. Model. 2017,57, 203–213.
(14) Jiang, X.; Kumar, A.; Motomura, Y.; Liu, T.; Zhou, Y.; Moro, K.; Zhang, K. Y. J.; Yang, Q. A Series of
Compounds Bearing a Dipyrido-Pyrimidine Scaffold Acting as Novel Human and Insect Pest Chitinase
Inhibitors. J. Med. Chem. 2020,63, 987–1001.
(15) Grinter, S. Z.; Zou, X. Challenges, Applications, and Recent Advances of Protein-ligand Docking in
Structure-based Drug Design. Molecules 2014,19, 10150–76.
(16) Ferreira, L. G.; Dos Santos, R. N.; Oliva, G.; Andricopulo, A. D. Molecular Docking and Structure-
based Drug Design Strategies. Molecules 2015,20, 13384–421.
(17) Wong, C. F. Flexible Receptor Docking for Drug Discovery. Expert Opin. Drug. Discov. 2015,10,
1189–200.
(18) Pagadala, N. S.; Syed, K.; Tuszynski, J. Software for Molecular Docking: a Review. Biophys. Rev.
2017,9, 91–102.
21

(19) pinzi, l.; rastelli, g. Molecular Docking: Shifting Paradigms in Drug Discovery. Int. J. Mol. Sci. 2019,
20 .
(20) Walters, W. P. Virtual Chemical Libraries. J. Med. Chem. 2019,62, 1116–1124.
(21) Cherkasov, A.; Ban, F.; Li, Y.; Fallahi, M.; Hammond, G. L. Progressive Docking: A Hybrid
QSAR/Docking Approach for Accelerating In Silico High Throughput Screening. J. Med. Chem. 2006,
49, 7466–7478.
(22) Svensson, F.; Norinder, U.; Bender, A. Improving Screening Efficiency through Iterative Screening
Using Docking and Conformal Prediction. J. Chem. Inf. Model. 2017,57, 439–444.
(23) Mysinger, M. M.; Carchia, M.; Irwin, J. J.; Shoichet, B. K. Directory of Useful Decoys, Enhanced
(DUD-E): Better Ligands and Decoys for Better Benchmarking. J. Med. Chem. 2012,55, 6582–6594.
(24) Gentile, F.; Agrawal, V.; Hsing, M.; Ton, A.-T.; Ban, F.; Norinder, U.; Gleave, M. E.; Cherkasov, A.
Deep Docking: A Deep Learning Platform for Augmentation of Structure Based Drug Discovery. ACS
Cent. Sci. 2020,6, 939–949.
(25) Sterling, T.; Irwin, J. J. ZINC 15 - Ligand Discovery for Everyone. J. Chem. Inf. Model. 2015,55,
2324–2337.
(26) Yanagisawa, K.; Komine, S.; Suzuki, S. D.; Ohue, M.; Ishida, T.; Akiyama, Y. Spresso: an Ultrafast
Compound Pre-screening Method Based on Compound Decomposition. Bioinformatics 2017,33, 3836–
3843.
(27) Gorgulla, C.; Boeszoermenyi, A.; Wang, Z.-F.; Fischer, P. D.; Coote, P. W.; Padmanabha Das, K. M.;
Malets, Y. S.; Radchenko, D. S.; Moroz, Y. S.; Scott, D. A.; Fackeldey, K.; Hoffmann, M.; Iavniuk, I.;
Wagner, G.; Arthanari, H. An Open-Source Drug Discovery Platform Enables Ultra-Large Virtual
Screens. Nature 2020,580, 663–668.
(28) Trott, O.; Olson, A. J. Autodock Vina: Improving the Speed and Accuracy of Docking with a New
Scoring Function, Efficient Optimization, and Multithreading. J. Comput. Chem. 2010,31, 455–461.
(29) Quiroga, R.; Villarreal, M. A. Vinardo: A Scoring Function Based on Autodock Vina Improves Scoring,
Docking, and Virtual Screening. PLOS ONE 2016,11, 1–18.
(30) Ripphausen, P.; Nisius, B.; Bajorath, J. State-of-the-art in Ligand-based Virtual Screening. Drug Dis-
covery Today 2011,16, 372 – 376.
22

(31) Drwal, M. N.; Griffith, R. Combination of Ligand- and Structure-based Methods in Virtual Screening.
Drug Discovery Today: Technol. 2013,10, e395 – e401.
(32) Kumar, A.; Zhang, K. Y. Hierarchical Virtual Screening Approaches in Small Molecule Drug Discovery.
Methods 2015,71, 26–37.
(33) Jastrzębski, S.; Szymczak, M.; Pocha, A.; Mordalski, S.; Tabor, J.; Bojarski, A. J.; Podlewska, S.
Emulating Docking Results Using a Deep Neural Network: A New Perspective for Virtual Screening.
J. Chem. Inf. Model. 2020,60, 4246–4262.
(34) Shang, C.; Liu, Q.; Chen, K.-S.; Sun, J.; Lu, J.; Yi, J.; Bi, J. Edge Attention-based Multi-Relational
Graph Convolutional Networks. 2018.
(35) Chupakhin, V.; Marcou, G.; Baskin, I.; Varnek, A.; Rognan, D. Predicting Ligand Binding Modes from
Neural Networks Trained on Protein-Ligand Interaction Fingerprints. J. Chem. Inf. Model. 2013,53,
763–772.
(36) Boyles, F.; Deane, C. M.; Morris, G. M. Learning From the Ligand: Using Ligand-based Features to
Improve Binding Affinity Prediction. Bioinformatics 2019,36, 758–764.
(37) Lyu, J.; Wang, S.; Balius, T. E.; Singh, I.; Levit, A.; Moroz, Y. S.; O’Meara, M. J.; Che, T.; Algaa, E.;
Tolmachova, K.; Tolmachev, A. A.; Shoichet, B. K.; Roth, B. L.; Irwin, J. J. Ultra-large Library
Docking for Discovering New Chemotypes. Nature 2019,566, 224–229.
(38) Cieplinski, T.; Danel, T.; Podlewska, S.; Jastrzebski, S. We Should at Least Be Able to Design Molecules
That Dock Well. 2020.
(39) Koes, D. R.; Baumgartner, M. P.; Camacho, C. J. Lessons Learned in Empirical Scoring with Smina
from the CSAR 2011 Benchmarking Exercise. J. Chem. Inf. Model. 2013,53, 1893–1904.
(40) Tran-Nguyen, V.-K.; Jacquemard, C.; Rognan, D. LIT-PCBA: An Unbiased Data Set for Machine
Learning and Virtual Screening. J. Chem. Inf. Model. 2020,60, 4263–4273.
(41) Wallach, I.; Heifets, A. Most Ligand-Based Classification Benchmarks Reward Memorization Rather
than Generalization. J. Chem. Inf. Model. 2018,58, 916–932.
(42) Kim, S.; Thiessen, P. A.; Bolton, E. E.; Chen, J.; Fu, G.; Gindulyte, A.; Han, L.; He, J.; He, S.;
Shoemaker, B. A.; Wang, J.; Yu, B.; Zhang, J.; Bryant, S. H. PubChem Substance and Compound
Databases. Nucleic Acids Res. 2016,44, D1202–D1213.
23

(43) Jones, G.; Willett, P.; Glen, R. C.; Leach, A. R.; Taylor, R. Development and Validation of a Genetic
Algorithm for Flexible Docking. J. Mol. Biol. 1997,267, 727 – 748.
(44) McGann, M. FRED and HYBRID Docking Performance on Standardized Datasets. J. Comp. Aid. Mol.
Des. 2012,26, 897–906.
(45) Friesner, R. A.; Murphy, R. B.; Repasky, M. P.; Frye, L. L.; Greenwood, J. R.; Halgren, T. A.;
Sanschagrin, P. C.; Mainz, D. T. Extra Precision Glide: Docking and Scoring Incorporating a Model
of Hydrophobic Enclosure for Protein-Ligand Complexes. J. Med. Chem. 2006,49, 6177–6196.
(46) Hawkins, P. C. D.; Skillman, A. G.; Warren, G. L.; Ellingson, B. A.; Stahl, M. T. Conformer Generation
with OMEGA: Algorithm and Validation Using High Quality Structures from the Protein Databank
and Cambridge Structural Database. J. Chem. Inf. Model. 2010,50, 572–584.
(47) Jakalian, A.; Bush, B. L.; Jack, D. B.; Bayly, C. I. Fast, Efficient Generation of High-quality Atomic
Charges. AM1-BCC Model: I. Method. J. Comput. Chem. 2000,21, 132–146.
(48) Jakalian, A.; Jack, D. B.; Bayly, C. I. Fast, Efficient Generation of High-quality Atomic Charges.
AM1-BCC Model: II. Parameterization and Validation. J. Comput. Chem. 2002,23, 1623–1641.
(49) Pettersen, E. F.; Goddard, T. D.; Huang, C. C.; Couch, G. S.; Greenblatt, D. M.; Meng, E. C.; Fer-
rin, T. E. UCSF Chimera—A Visualization System for Exploratory Research and Analysis. J. Comput.
Chem. 2004,25, 1605–1612.
(50) Maier, J. A.; Martinez, C.; Kasavajhala, K.; Wickstrom, L.; Hauser, K. E.; Simmerling, C. ff14SB:
Improving the Accuracy of Protein Side Chain and Backbone Parameters from ff99SB. J. Chem. Theory
Comput. 2015,11, 3696–3713.
(51) Morris, G. M.; Huey, R.; Lindstrom, W.; Sanner, M. F.; Belew, R. K.; Goodsell, D. S.; Olson, A. J.
AutoDock4 and AutoDockTools4: Automated Docking with Selective Receptor Flexibility. J. Comput.
Chem. 2009,30, 2785–2791.
(52) Gasteiger, J.; Marsili, M. Iterative Partial Equalization of Orbital Electronegativity a Rapid Access to
Atomic Charges. Tetrahedron 1980,36, 3219 – 3228.
(53) Shelley, J. C.; Cholleti, A.; Frye, L. L.; Greenwood, J. R.; Timlin, M. R.; M., U. Epik: a software
program for pK( a ) prediction and protonation state generation for drug-like molecules. J. Comput.
Aided Mol. Des. 2007,21, 681–691.
24

(54) Atkinson, F. Molecular Standardisation Tool. https://github.com/flatkinson/standardiser, accessed
2019-06-01.
(55) Berenger, F. Molenc: a Molecular Encoder. https://github.com/UnixJunkie/molenc, accessed 2019-11-
19.
(56) Gaulton, A.; Hersey, A.; Nowotka, M.; Bento, A. P.; Chambers, J.; Mendez, D.; Mutowo, P.; Atkin-
son, F.; Bellis, L. J.; Cibrián-Uhalte, E.; Davies, M.; Dedman, N.; Karlsson, A.; Magariños, M. P.;
Overington, J. P.; Papadatos, G.; Smit, I.; Leach, A. R. The ChEMBL database in 2017. Nucleic Acids
Res. 2016,45, D945–D954.
(57) Fan, R.-E.; Chang, K.-W.; Hsieh, C.-J.; Wang, X.-R.; Lin, C.-J. LIBLINEAR: A Library for Large
Linear Classification. J. Mach. Learn. Res. 2008,9, 1871–1874.
(58) Berenger, F. Wrapper on Top of Liblinear-tools. https://github.com/UnixJunkie/linwrap, accessed
2020-11-26.
(59) Swamidass, S. J.; Azencott, C.-A.; Daily, K.; Baldi, P. A CROC Stronger than ROC: Measuring,
Visualizing and Optimizing Early Retrieval. Bioinformatics 2010,26, 1348–1356.
(60) Berenger, F. Classification and Regression Performance Metrics Library. https://github.com/
UnixJunkie/cpmlib, accessed 2020-11-26.
(61) Sheridan, R. P.; Karnachi, P.; Tudor, M.; Xu, Y.; Liaw, A.; Shah, F.; Cheng, A. C.; Joshi, E.; Glick, M.;
Alvarez, J. Experimental Error, Kurtosis, Activity Cliffs, and Methodology: What Limits the Predictiv-
ity of Quantitative Structure–Activity Relationship Models? J. Chem. Inf. Model. 2020,60, 1969–1982.
(62) Zhenin, M.; Bahia, M. S.; Marcou, G.; Varnek, A.; Senderowitz, H.; Horvath, D. Rescoring of Docking
Poses under Occam’s Razor: Are There Simpler Solutions? J. Comp. Aid. Mol. Des. 2018,32, 877–888.
(63) da Silva Figueiredo Celestino Gomes, P.; Da Silva, F.; Bret, G.; Rognan, D. Ranking Docking Poses
by Graph Matching of Protein-Ligand Interactions: Lessons Learned from the D3R Grand Challenge
2. J. Comp. Aid. Mol. Des. 2018,32, 75–87.
(64) Da Silva, F.; Desaphy, J.; Rognan, D. IChem: A Versatile Toolkit for Detecting, Comparing and
Predicting Protein-Ligand Interactions. ChemMedChem 2018,13, 507–510.
(65) Jacquemard, C.; Tran-Nguyen, V.-K.; Drwal, M. N.; Rognan, D.; Kellenberger, E. Local Interaction
Density (LID), a Fast and Efficient Tool to Prioritize Docking Poses. Molecules 2019,24, 2610.
25

(66) Horvath,; Marcou,; Varnek, Generative Topographic Mapping of the Docking Conformational Space.
Molecules 2019,24, 2269.
(67) X, Q.; XY, L.; De Raeymaecker, J.; J, T.; K, Z.; M, D. M.; A, V. Pharmacophore Modeling: Advances,
Limitations, and Current Utility in Drug Discovery. J. Recept., Ligand Channel Res. 2014,7, 81–92.
(68) Guterres, H.; Im, W. Improving Protein-Ligand Docking Results with High-Throughput Molecular
Dynamics Simulations. J. Chem. Inf. Model. 2020,60, 2189–2198.
(69) McGann, M. R.; Almond, H. R.; Nicholls, A.; Grant, J. A.; Brown, F. K. Gaussian Docking Functions.
Biopolymers 2003,68, 76–90.
(70) McGann, M. FRED Pose Prediction and Virtual Screening Accuracy. J. Chem. Inf. Model. 2011,51,
578–596.
(71) Baell, J. B.; Holloway, G. A. New Substructure Filters for Removal of Pan Assay Interference Com-
pounds (PAINS) from Screening Libraries and for Their Exclusion in Bioassays. J. Med. Chem. 2006,
49, 6177–6196.
(72) Boitreaud, J.; Mallet, V.; Oliver, C.; Waldispühl, J. OptiMol: Optimization of Binding Affinities in
Chemical Space for Drug Discovery. J. Chem. Inf. Model. 2020,60, 5658–5666.
26

Graphical TOC Entry
0
1000
2000
3000
4000
5000
6000
7000
-100 -90 -80 -70 -60 -50 -40 -30 -20 -10
Target protein: GBA
1 2 3
Number of molecules
Gold docking score
Classical docking
Lean-docking
Distribution of docking scores for classical dock-
ing versus lean-docking 25%. The top three low
scoring active molecules are shown on the left.
Green arrows: actives docked upon lean-docking.
Blue arrows: remaining actives.
27