LigityScore: Convolutional Neural Network for Binding-afﬁnity

Predictions

Joseph Azzopardi

1 a

and Jean Paul Ebejer

1,2 b

Department of Artiﬁcial Intelligence, University of Malta, Msida, MSD 2080, Malta

Centre for Molecular Medicine and Biobanking, University of Malta, Msida, MSD 2080, Malta

Keywords:

Virtual Screening, Structure-based Virtual Screening, Scoring Function, Pharmacophoric Interaction Point,

Machine Learning, Deep-Learning, Convolutional Neural Networks, LigityScore.

Abstract:

Scoring functions are at the heart of structure-based drug design and are used to estimate the binding of ligands

to a target. Seeking a scoring function that can accurately predict the binding afﬁnity is key for successful

virtual screening methods. Deep learning approaches have recently seen a rise in popularity as a means to

improve the scoring function having as a key advantage the automatic extraction of features and the creation

of a complex representation without feature engineering and expert knowledge. In this study we present

LigityScore1D and LigityScore3D, which are rotationally invariant scoring functions based on convolutional

neural networks. LigityScore descriptors are extracted directly from the structural and interacting properties

of the protein-ligand complex which are input to a CNN for automatic feature extraction and binding afﬁnity

prediction. This representation uses the spatial distribution of Pharmacophoric Interaction Points, derived from

interaction features from the protein-ligand complex based on pharmacophoric features conformant to speciﬁc

family types and distance thresholds. The data representation component and the CNN architecture together,

constitute the LigityScore scoring function. The main contribution for this study is to present a novel protein-

ligand representation for use as a CNN based SF for binding afﬁnity prediction. LigityScore models are

evaluated for scoring power on the latest two CASF benchmarks. The Pearson Correlation Coefﬁcient, and

the standard deviation in linear regression were used to compare and rank LigityScore with the benchmark

model, and also to other models recently published in literature. LigityScore3D has achieved better overall

results and showed similar performance in both CASF benchmarks. LigityScore3D ranked 5

place for the

CASF-2013 benchmark , and 8

for CASF-2016, with an average R-score performance of 0.713 and 0.725

respectively. LigityScore1D ranked 8

place for the CASF-2013 and 7

place for CASF-2016 with an R-

score performance of 0.635 and 0.741 respectively. Our methods show relatively good performance when

compared to the Pafnucy model (one of the best performing CNN based scoring functions), on the CASF-

2013 benchmark using a less computationally complex model that can be trained 16 times faster.

1 INTRODUCTION

Structure-based virtual screening (SBVS) employs

the known 3D protein structure to apply computa-

tional methods that measure the ability of a small

molecule to bind to the target protein. Docking is

one of the most popular SBVS methods and is used

to validate the ability of small molecules to bind to

the target structure in a typical ’lock and key’ fash-

ion (Ching et al., 2018). During the docking process

many possible ligand conformers, or docking poses,

are iteratively tested at the binding site to ﬁnd a suit-

https://orcid.org/0000-0001-9058-5361

https://orcid.org/0000-0003-0888-2637

able ligand pose yielding the best binding afﬁnity.

The binding afﬁnity of a particular pose is determined

by the scoring function (SF) of the docking program.

The SF is therefore crucial for docking programs in

SBVS and can be deﬁned as ”Estimating how strongly

the docked pose of a ligand binds to the target” (Ain

et al., 2015). Scoring functions are typically used

for fast evaluation of protein-ligand interactions, so

building an efﬁcient and powerful SF is a means of

accelerating the virtual screening (VS) process. The

SF is considered the foundation in SBVS and used in

the following areas for hit discovery and optimization

(Ragoza et al., 2017):

1. Pose Prediction: Predict the shape of the ligand

Azzopardi, J. and Ebejer, J.

LigityScore: Convolutional Neural Network for Binding-afﬁnity Predictions.

DOI: 10.5220/0010228300380049

In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) - Volume 3: BIOINFORMATICS, pages 38-49

ISBN: 978-989-758-490-9

which gives the best binding afﬁnity.

2. Ranking: Ranking of molecules with known

binding pose in order of the binding afﬁnity for

a given protein target.

3. Classiﬁcation: Given the binding-pose, classify

whether a small molecule is active or inactive for

a given 3D structure of a target protein.

Intense research has been carried out over the

years on this problem, however improving the accu-

racy for binding afﬁnity prediction has proven to be

a non-trivial task (Ain et al., 2015). Despite sev-

eral advances in binding afﬁnity prediction, the cur-

rent binding afﬁnity estimates are still not accurate

enough leading to high false positive rates (Zheng

et al., 2019). The work described in this paper fo-

cuses directly on the SF layer of SBVS and tack-

les this problem by ﬁnding an alternative data repre-

sentation of the protein-ligand complex and then ap-

plies a Convolutional Neural Network (CNN) regres-

sion model to implement the SF. Deep Learning (DL)

models have achieved remarkable success in various

areas such as computer vision (Szegedy et al., 2015)

and natural language processing (Goldberg, 2016).

Inspired by this success, the use of DL and in particu-

lar CNNs have naturally become an obvious strategy

to apply for computer aided drug design. Conven-

tional ML techniques, such as Random Forests and

Support Vector Machines, are limited to process raw

data, requiring careful feature engineering and expert

domain knowledge (LeCun et al., 2015). Deep learn-

ing methods aims to reduce feature engineering and

automatically extract the salient feature information

from the input data using multiple hidden layers, pro-

vided it has a suitable representation of the molecu-

lar interactions between the protein and ligand (P

erez-

Sianes et al., 2019). Therefore, our strategy is to allow

deep learning models to learn the underlying molec-

ular interactions so that this learned information can

be reapplied to other protein targets for exploration

of novel ligands, without the need to incorporate ex-

pert chemical knowledge. Our work is evaluated us-

ing the CASF-2013 (Li et al., 2014) and CASF-2016

(Su et al., 2018) benchmarks and is also compared

to other recently-published SF methods that use the

same benchmarks.

The ﬁrst deep neural network (DNN) used for VS

was introduced by the winners of the 2012 Kaggle

Merck Molecular activity challenge (Kaggle, 2012)

where the team applied a multi-task deep feed for-

ward network for quantitative structure-activity. This

work was later published by Ma et al. (2015) and gen-

erated a lot of interest and excitement around the use

of DL in this ﬁeld. Ma et al. (2015) have achieved an

average Pearson correlation coefﬁcient of 0.496 us-

ing a multi-task DNN compared to the 0.423 obtained

when using a Random Forest model.

The Convolutional Neural Network (CNN) is one

of the most common DL architectures used for SFs

(Rifaioglu et al., 2018; P

erez-Sianes et al., 2019). The

CNN uses a number of sequential layers of convolu-

tions and pooling modules to encode the hidden fea-

tures of the data, and then use a fully connected feed-

forward neural network for classiﬁcation or regres-

sion. One of the advantages of CNNs in the area of

structure-based drug design is its ability to capture lo-

cal spatial information interactions between protein-

ligand complexes. In recent years, many CNN models

have been applied to SF development (Ragoza et al.,

2017; Stepniewska-Dziubinska et al., 2017; Jim

enez

et al., 2018; Zheng et al., 2019; Liu et al., 2019).

Ragoza et al. (2017) were the ﬁrst to use CNNs

to implement a DL scoring function that predicts

the docking score for a drug target interaction which

was then used for SBVS and pose prediction. How-

ever, Sieg et al. (2019) later showed that their model

was effected by non-causal bias. One of the most

promising DL SF models, Pafnucy, was proposed by

Stepniewska-Dziubinska et al. (2017) where the au-

thors achieved a Pearson Coefﬁcient, R, for the pre-

dicted versus experimental binding afﬁnity of 0.70 on

the CASF-2013, and 0.78 on the CASF-2016 bench-

marks respectively. Pafnucy uses a 3D CNN model

with a 4D tensor to represent 19 protein-ligand fea-

tures in 3D. The 4D tensor includes discretised atom

location in the ﬁrst three dimensions, whilst the fea-

tures for the particular atom are encoded in the fourth

dimension.

Jim

enez et al. (2018) also utilise a 3D CNN model

termed K

deep

and achieved state of the art results on

the CASF-2016 test with an R value of 0.82. Their

input features use a 3D voxel representation where

each channel encodes a particular property of the

atom. Each protein-ligand complex is represented by

a 4D tensor, where each 3D hyperplane represents the

protein-ligand complex with respect to a particular

property only. The eight properties chosen for K

deep

include: hydrophobic, aromatic, hydrogen bond ac-

ceptor (HBA), hydrogen bond donor (HBD), cation,

anion, metallic, and excluded volume. A recent study

proposed by Zheng et al. (2019) compares their model

to Stepniewska-Dziubinska et al. (2017) and criticize

the Pafnucy model that the protein-ligand interactions

in a 3D grid box of 20

A are not sufﬁcient to capture

all the protein-ligand interactions, and suggest that

other long-range interactions outside the 20

A , termed

non-local electrostatic interactions are also important.

To capture all the interactions between protein-ligand

LigityScore: Convolutional Neural Network for Binding-afﬁnity Predictions

Figure 1: PIP pair interaction between two hydrogen bond donor protein PIPs (blue mesh), and a hydrogen bond acceptor

ligand PIP (red mesh). Other PIPs not shown for clarity. For each PIP interaction the distance between the geometric centres

of the PIPs is calculated.

complexes, Zheng et al. (2019) divide all the 3D space

of the binding site into a number of shells or zones and

count the number of different element-to-element in-

teractions within each shell. Their experiments show

that the shells closer to the ligand are more impor-

tant, as was intuitively expected, however they also

show that non-local interactions have signiﬁcant im-

portance. Zheng et al. (2019) compare their method

OnionNet to another recent model, AGL-Score by

Nguyen and Wei (2019a). Both methods use CASF

as an evaluation benchmark. Zheng et al. comment

that OnionNet provides a more complete local envi-

ronment and improves the afﬁnity prediction perfor-

mance with an R of 0.833. To date this represents the

best performing ML scoring function. The better re-

sults are achieved by adding novel features relating to

the physical and biological information of the com-

plex using graph theory.

One of the limitations of Pafnucy and K

deep

is the

dependency on the coordinate frame. The represen-

tation can be thought of as one snapshot of the struc-

ture. However, if the orientation from where the snap-

shot is changed, a different representation of the same

protein-ligand complex is obtained. The authors have

worked around this limitation by introducing different

systematic rotations of the same input during train-

ing. However, these might present additional chal-

lenges when testing novel complexes that can take

different orientations. This limitation has led us to

explore methods that are inherently rotationally in-

variant. One such model that is not dependent on

the coordinate frame is Ligity developed by Ebejer et

al. (2019). Ligity is a hybrid VS technique that col-

lects key interaction features within the protein-ligand

complexes. These key interaction points are known

as hot-spots and are deﬁned by considering speciﬁc

pharmacophoric features that lie within a predeter-

mined distance threshold between the protein and

ligand feature pairs. Each of these pharmacophoric

features that interact together are termed Pharma-

cophoric Interaction Points or PIPs. Once these pairs

are extracted, the Ligity descriptor for the ligand is

created by considering only the PIPs from the ligand

space. Three or four PIP combinations are considered

in the original Ligity method. The Ligity descriptor

is built using the spatial distribution of PIPs (i.e. the

distance between PIPs), and is, therefore, rotationally

invariant.

Pafnucy has motivated us to use the CNN mod-

els for automatic feature extraction, and to ﬁnd an

alternative representation for the protein-ligand com-

plex based on its structural and interacting properties.

On the other hand, Ligity was used as the basis of

our study and has inspired us to build a feature rep-

resentation using both the protein and ligand PIPs,

as opposed to Ligity that uses only ligand PIPs. In

this study we present LigityScore — LigityScore is a

novel rotationally invariant CNN based scoring func-

tion that utilises the interaction of pharmacophoric

features of the protein and ligand for its data repre-

sentation. In our approach we have therefore hypoth-

esised that these pharmacophoric interactions across

different feature types contain key information to suit-

ably represent the protein ligand structure and their

binding properties. We have further hypothesised that

this representation would be suitable to train a CNN

model for binding afﬁnity prediction. Our approach

introduces two techniques, LigityScore1D and Ligi-

tyScore3D, that make use of important structural fea-

tures of both the protein and ligand to create a suit-

able data representation of the protein-ligand com-

plex. LigityScore uses distance between PIPs, which

BIOINFORMATICS 2021 - 12th International Conference on Bioinformatics Models, Methods and Algorithms

remain the same irrespective of the structure’s orienta-

tion hence making the representation rotationally in-

variant. The PIP pair interactions from the protein

and ligand are illustrated in Figure 1. Other methods

such as OnionNet have inspired us to consider Phar-

macophoric Interaction Points (PIPs) that are further

apart, and to use InstanceNorm and ReLU to enhance

our CNN models.

The LigityScore method considers six pharma-

cophore features: hydrophobic, hydrogen bond ac-

ceptor, hydrogen bond donor, aromatic, cation, and

anion. Gund (1977) describes a pharmacophore as “a

set of structural features in a molecule that are recog-

nised at the binding site and is responsible for that

molecule’s biological activity”. Therefore a pharma-

cophore model represents a number of general struc-

tural features such as aromatic or hydrophobic re-

gions within the molecule, which are used to iden-

tify the features at the binding site that are respon-

sible for molecular binding and biological activity.

These features at the binding site may be used to ﬁnd

strong molecular binding interactions or hot-spots in

the protein-ligand complex which are used to extract

descriptors that represent the protein-ligand attributes.

These features can be used in a 3D pharmacophore

model and the spatial relationship between these phar-

macophoric features can also be used to represent the

protein-ligand complex (Leach et al., 2010).

The novel protein-ligand representation for use in

a CNN based scoring function for binding afﬁnity

prediction is our major contribution in the SBVS do-

main. The source code for LigityScore is available at

https://gitlab.com/josephazzopardi/ligityscore.

2 MATERIALS AND METHODS

LigityScore is a CNN based scoring function that

utilises a rotationally invariant data representation

extracted from interacting pharmacophoric features

in protein-ligand complexes. An overview of Ligi-

tyScore is illustrated in Figure 2, highlighting the pa-

rameters that can be changed for each module. The

major functional parts for both LigityScore1D and

LigityScore3D are detailed below:

1. Pre-processing. PDBbind (Liu et al., 2017a)

ﬁles are processed to build a dataset of the com-

plexes with their respective binding afﬁnity val-

ues. At this stage the molecular ﬁles are validated

to check that the protein-ligand complexes listed

have corresponding molecular ﬁles, and also to ﬁx

any errors that occur whilst loading the ﬁles us-

ing the RDKit cheminformatics library (Landrum,

2020).

2. PIP (Hot-spots) Generation. This module

loads the complexes and searches for the phar-

macophoric features using the RDKit BuildFea-

tureFactory class. All the possible pairs of phar-

macophoric features across the protein pocket and

the ligand are built, and are then run against a

number of constraints (see Table 1). The resul-

tant feature pairs represent the PIPs or interaction

hot-spots for a particular complex.

3. Generation of LigityScore Descriptors. The

LigityScore descriptors module utilises the hot-

spots dataset that includes both the ligand and pro-

tein PIPs to generate a feature descriptor for each

complex. LigityScore1D considers two hot-spots

at a time that correspond to a particular pharma-

cophric feature family pair (e.g. HBA-HBA). For

each possible family pair a feature vector is con-

structed, hence the name LigityScore1D. On the

other hand, LigityScore3D uses three PIPs at a

time, and the spatial information for the family

set (e.g. HBA-HBA-HBA) is encoded in a feature

cube. A family set is a combination of three PIP

types. The names of our models are derived from

the dimensionality of the spatial information used

to generate the features.

4. CNN Training. This module is built using the

Pytorch library (Paszke et al., 2019) and includes

a dynamic model to construct a CNN in order

to facilitate the testing and evaluation of differ-

ent CNN architectures. The CNN module is used

to train the network as a scoring function using

the LigityScore descriptors. The module tackles a

regression type of problem and therefore the out-

put is a continuous value predicting the binding

afﬁnity of the protein-ligand complex. This output

is compared with the experimental binding afﬁn-

ity so that the network parameters are updated

using stochastic gradient descent. Each epoch

is validated against the validation set, composed

of 1,000 randomly sampled complexes from the

PDBbind Reﬁned set, and the model with the low-

est root means squared error (RMSE) is stored to

disk for use for predicting unseen complexes.

5. CNN Predictions. The Predictions module is

used to load the best performing model and to

compute results for the Training, Validation, and

Test Sets.

6. Experiments Pipeline. This module combines

the Training, Validation and Testing in one

pipeline. This step uses a CSV ﬁle to describe a

series of experiments with different CNN param-

eters.

LigityScore: Convolutional Neural Network for Binding-afﬁnity Predictions

Figure 2: LigityScore schematic representation of the major functional components used in our approach to develop a CNN

based scoring function for virtual screening. The parameters used in each functional block are included as reference. For

example PIP Hot-spots extraction can take two parameters – Lipinski ﬁltering, and the hot-spots distance threshold factor.

2.1 Evaluation Dataset

The PDBbind dataset (Liu et al., 2017a) is regarded

as a golden dataset for the development of scor-

ing functions (Liu et al., 2017b), and was therefore

used for training and testing of LigityScore. The

PDBbind v2018 was also used as an additional ex-

periment for data augmentation since it has around

2,700 additional protein-ligand complexes. The PDB-

Bind dataset is manually curated and includes records

of experimentally measured binding afﬁnity data for

biomolecular complexes taken from the Protein Data

Bank (PDB) (Berman et al., 2003). Their binding

afﬁnity is expressed in terms of dissociation (K

), in-

hibition (K

) or half-concentration (IC

) constants.

No distinction is made between these constants and

they were converted into a negative log; pK

−log

, where K

can be K

, K

or IC

, and pK

is the binding afﬁnity (Stepniewska-Dziubinska et al.,

2017; Zheng et al., 2019). The PDBbind dataset is

split into the General and Reﬁned sets.

The Core set v2013 and v2016 were established as

part of the CASF-2013 and CASF-2016 benchmarks.

These benchmarks are meant to provide an objec-

tive platform to assess scoring functions, using high-

quality protein-ligand complexes selected from the

reﬁned set, through a systematic and non-redundant

sampling procedure (Su et al., 2018). The CASF

BIOINFORMATICS 2021 - 12th International Conference on Bioinformatics Models, Methods and Algorithms

benchmarks were used to assess the Scoring Power

of LigityScore1D and LigityScore3D. The scoring

power is quantitatively measured for evaluation using

the Pearson’s correlation coefﬁcient, R, and the stan-

dard deviation in linear regression (SD). The Scoring

Power measures the ability of the model to map a lin-

ear correlation of the predicted and known experimen-

tal afﬁnity values. This study is focused to predict the

binding afﬁnity and the scoring power CASF bench-

marks will be used for objective assessment and eval-

uation of the proposed scoring function.

A validation set, composed of 1,000 randomly

chosen complexes from the reﬁned set, was selected

to evaluate the training progress after each epoch and

select the CNN model with the smallest error. The

validation set was also used for Early Stopping func-

tionality in the training module so that training is

stopped after a number of epochs with no loss im-

provements. The complexes in the validation set are

chosen entirely from the reﬁned set, as these provide

higher quality protein-ligand complexes and are more

reliable for the development of scoring functions (Liu

et al., 2017b). Each of the core sets (2013 and 2016)

were used entirely as the two test sets to simulate

new and unseen protein-ligand complexes during the

prediction stage. None of these test structures were

used during training and validation. The remainder of

the protein-ligand complexes were used as the train-

ing set. In our study, training was not performed

on individual protein families but a generic CNN

model was developed for all protein families which is

a common approach in ML-based scoring functions

(Stepniewska-Dziubinska et al., 2017; Jim

enez et al.,

2018; Ragoza et al., 2017; Zheng et al., 2019;

Ozt

urk

et al., 2018).

2.2 LigityScore Implementation

The LigityScore scoring function consists of the fea-

ture generation process to extract a representation of

the protein ligand complex in LigityScore space, and

the CNN model for automatic feature extraction and

representation for binding afﬁnity predictions. The

two components are described next.

2.2.1 Feature Descriptors

The protein-ligand feature representation is split into

two phases. In the ﬁrst phase the PIPs of each in-

dividual protein-ligand complex are extracted to cre-

ate the PIP dataset using all the PDBBind com-

plexes. The algorithm used for PIP generation is

based on the Ligity methodology described in Ebe-

jer et al. (2019). The PIPs are extracted from the

query protein-ligand complex using the open-source

cheminformatics package RDKit BuildFeatureFac-

tory class (Landrum, 2020). The BuildFeatureFac-

tory uses SMARTS patterns to identify these pharma-

cophoric features within the molecule. These PIPs,

from both the protein and the ligand, are then ﬁltered

by a set of rules constraining feature family-pairs at a

speciﬁc distance threshold so as to capture only the

stronger interactions between the protein’s and lig-

and’s pharmacophoric features. The allowed feature

family pairs and their corresponding distance thresh-

old are listed in Table 1. The euclidean distances be-

tween the features are calculated using the centre of

the atoms making up the feature. For example in a

six-membered aromatic ring PIP, only the centre of

the atomic structure is considered. In order to extract

all conformant PIPs from the protein-ligand complex

a cartesian product of all PIPs from the protein and

ligand is performed, followed by the ﬁltering of the

allowed family pairs, and further ﬁltering by the max-

imum distances allowed. PIP interactions are illus-

trated in Figure 1 showing the calculated distances

between centres of PIPs.

In our approach we have also considered using

longer distance thresholds than those stated in Ta-

ble 1 which varies from the approach of Ebejer et al.

(2019), and was inspired by the non-local electrostatic

interactions used in Zheng et al. (2019). The PIP gen-

eration module provides a distance threshold-factor

argument that can be used to multiply this baseline

distance. A number of experiments were carried out

using a varying distance threshold-factor between 1.0

and 1.6 in order to capture additional PIP interactions

in our feature representation. This additional infor-

mation includes also other weaker interactions, since

the protein and ligand features are further apart, which

can lead to a more information-rich representation of

the protein-ligand complex.

Table 1: Pharmacophoric features and distance thresholds

used to extract PIPs from the protein-ligand complex, re-

produced from Ebejer et al. (2019).

Interacting Protein-Ligand

PIP Family Pairs

Distance Threshold (

hydrophobic, hydrophobic 4.5

hydrogen bond acceptor, donor

3.9

cation, anion 4.0

aromatic, aromatic 4.5

cation, aromatic 4.0

The second phase of the protein-ligand complex

representation uses the PIP dataset described in the

previous section to create a feature matrix, or a feature

cube collection for every complex for LigityScore1D

or LigityScore3D respectively. The feature cube col-

lection generation process for LigityScore3D is high-

lighted in Figure 3. Each feature cube collection is

LigityScore: Convolutional Neural Network for Binding-afﬁnity Predictions

Figure 3: LigityScore3D Feature Cube Collection Genera-

tion. A protein-PIP set and a ligand-PIP set are extracted

from a protein-ligand complex. From each 3-PIP combina-

tion taken from the PIP pool, the family-set, and their dis-

crete distances are used to update the binning count in the

Feature Cube. A Feature Cube is built for every available

family-set.

calculated using the PIPs from the PIP-dataset related

to the particular protein-ligand complex. The PIPs for

the ligand side and those of the protein side are ex-

tracted to obtain two separate sets – the ligand PIP

set, and the protein PIP set.

The feature cube collection is constructed by con-

sidering all the possible combinations when choosing

one PIP from the ligand-PIP set and two PIPs from the

protein-PIP set, and vice-versa. This 3-PIP combina-

tion creates a triangular structure amongst the PIPs as

shown in Figure 3 and generates a set of three dis-

tances. The three distances are discretised using a

A resolution to extract a binning coordinate in 3D

space. Additionally, each 3-PIP family combination

represents a unique feature cube within the feature

cube collection. Taking three, out of six pharma-

cophoric families considered with replacement cre-

ates a total of 56 possible 3-family set combinations.

The unique family set combination is used to index

the particular feature cube to update the binning count

using the coordinates from the three discretised dis-

tances. The voxel, or bin at this location is then in-

cremented by one. Considering the example in Fig-

ure 3, one ligand PIP and two protein PIPs are con-

sidered. These generate a PIP-family combination of

HBA-HBA-HBD, so that its cube will be updated at

the (10, 8, 3) voxel location. The family combina-

tions were sorted using both names and distance, to

ensure the correct feature cube is updated.

Considering a maximum distance of 20

A in each

dimension, each feature cube has a dimension of

21 × 21 × 21. Since each 3-PIP family set has its own

feature cube, 56 features cubes are stacked together

to create a protein-ligand LigityScore3D representa-

tion of size 1176×21×21. As indicated by Figure 3,

the PIP distances are calculated by using combina-

tions across both the ligand PIPs and the protein PIPs.

In LigityScore3D, a combination of 3-PIPs is consid-

ered at a time. All the possible combinations using

two PIPs for the protein, and one PIP from the ligand,

plus the combinations where two ligand PIPs and one

protein PIP are considered. This contrasts with the

approach used in Ebejer et al. (2019) where only the

ligand PIP pool was considered to take 3-PIP and 4-

PIP combinations. Our hypothesis is that since the

protein structure is essential for SBVS, considering

also the protein PIPs in the feature generation process

strengthens our model.

The method used for LigityScore1D is similar to

LigityScore3D but considers a combinations of 2-

PIPs at a time, and hence one inter-PIP distance. Each

PIP family pair (example HBA-HBD) represents a

different row in the feature matrix for the complex.

Therefore the PIP-family pair is used to index the row

of the feature matrix, whilst the discretised distance

is used to index the column of the PIP. These coordi-

nates in the feature matrix are then used to increment

the bin count of that location. A total of 21 family

combinations are possible, where each vector has 21

discrete locations corresponding to a feature matrix of

21 × 21 per protein-ligand complex.

2.2.2 CNN Architecture

The architecture used for LigityScore is a deep con-

volutional neural network with a single regression

output neuron used for prediction of binding afﬁn-

ity. The patterns extracted should differentiate the

spatial information between different complexes cap-

tured from the PIP interactions. The CNN architec-

ture used for LigityScore3D is illustrated in Figure 4.

The input is normalised and size for LigtyScore3D is

transformed to 98 × 98 × 54, whilst LigityScore1D is

BIOINFORMATICS 2021 - 12th International Conference on Bioinformatics Models, Methods and Algorithms

Figure 4: CNN Architecture for LigityScore3D. The input is reshaped to (98 × 98 × 54). Four convolutional layers are used

with instance normalisation, RELU activation, and spatial dropout. The output of the last convolution layers is ﬂattened to

feed a fully connected network with 4 hidden layers. The output is a single neuron predicting the binding afﬁnity.

kept at 21 × 21. These inputs are treated as 2D and

3D tensors respectively, and our approach treats them

similar to a greyscale and colour image respectively.

This analogy allowed us to explore and use image pro-

cessing techniques to optimise the scoring function

model such as InstanceNorm (Ulyanov et al., 2017).

LigityScore3D inputs are processed using four

convolutional layers with ﬁlter dimensions of 64, 128,

256, and 512 respectively, initialised using the Kaim-

ing method (He et al., 2015). This initialising method

is well suited for use with the RELU activation func-

tion as it keeps the standard deviation of the layer’s

activations close to one. Correctly initialising the

weights of the network is important for the training

of deep neural networks as this prevents the output

of the activation layers from exploding or vanishing.

The PyTorch Conv2D module is used for each of the

convolutional blocks. A convolutional kernel size of

5 × 5 is used, with a padding and stride of one. Each

convolution layer includes InstanceNorm, RELU ac-

tivation, and spatial dropout (Tompson et al., 2015)

components, which is then followed by a maxpooling

layer with a patch of two to reduce the dimensions by

half. LigityScore1D uses a similar CNN architecture

but uses three convolution layers (64, 128, 256), and

a padding of two.

The output of the last convolution layers is ﬂat-

tened to be used to feed four fully connected lay-

ers. LigityScore1D had a dimension of 3 × 3 × 256

at the input of the fully connected layers whilst Lig-

ityScore3D had an 5 × 5 × 2 × 512 input. To cater

for the difference in dimensionality the fully con-

nected layers were assigned dimensions of (2000,

1000, 500, 200) and (6000, 2000, 1000, 200) respec-

tively. Stochastic gradient descent with the Adam op-

timisation (Kingma and Ba, 2014) is used with default

parameters for momentum scheduling (β

= 0.99,

= 0.999) to train the network with a learning rate

of 10

−5

and L2 weight decay of λ = 0.001, using a

mini-batch size of 20. Various experiments were car-

ried out to tune hyperparameters.

3 RESULTS AND DISCUSSION

Several experiments were performed to ﬁnd the best

performing CNN architecture and LigityScore data

representation. A considerable improvement in pre-

diction performance was achieved when PIP distance

threshold factors greater than one were applied to the

values listed in Table 1. This implies that pharma-

cophoric hot-spots that are further apart are also con-

sidered during the PIP generation. The PIP threshold

factor of 1.4 showed an improvement in the R-score

of 19% over the baseline model. This may indicate

that long-range interactions also play a role in protein-

ligand binding.

LigityScore1D achieved best results when using

InstanceNorm at the convolution layers, a PIP thresh-

old factor of 1.4, and spatial dropout of 0.1 and ob-

tained an R-score of 0.725 for CASF-2016 and 0.695

for CASF-2013 test sets. Spatial dropout was applied

after the second convolution layers (middle layer with

128 channels) similar to the usage described in Tomp-

son et al. (2015). The best results for LigityScore3D

were achieved with spatial dropout probability of 0.2

on all the convolution layers and obtained a prediction

performance on the Core-2016 R-score of 0.739, and

a Core-2013 R-Score of 0.745. Spatial dropout im-

proved the CNN as it made it more resilient to over-

ﬁtting allowing the network to achieve higher predic-

tions scores.

The mean and standard deviation of 10 tests of

the best performing models were taken, to remove

any bias that might be caused from testing using a

LigityScore: Convolutional Neural Network for Binding-afﬁnity Predictions

Table 2: Performance of LigityScore1D when trained with PDBbind v2016 and v2018, and LigityScore3D trained with

PDBbind v2016, showing average and standard deviation for 10 tests using different validations sets taken from the reﬁned

set. LigityScore3D has the better overall performance for Core2013 and Core2016 test sets.

Set RMSE (±std) MAE (±std) SD (±std) R(±std)

LigityScore1D (v2016)

Training 0.406 (0.151) 0.323 (0.118) 0.393 (0.157) 0.974 (0.027)

Validation 1.438 (0.038) 1.144 (0.031) 1.432 (0.032) 0.698 (0.020)

Core2016 1.556 (0.039) 1.234 (0.031) 1.555 (0.038) 0.699 (0.018)

Core2013 1.861 (0.076) 1.485 (0.051) 1.701 (0.042) 0.657 (0.021)

LigityScore1D (v2018)

Training 0.964 (0.295) 0.764 (0.237) 0.947 (0.287) 0.845 (0.076)

Validation 1.447 (0.037) 1.158 (0.033) 1.436 (0.029) 0.684 (0.017)

Core2016 1.516 (0.066) 1.223 (0.058) 1.461 (0.038) 0.741 (0.016)

Core2013 1.831 (0.072) 1.472 (0.064) 1.743 (0.054) 0.635 (0.028)

LigityScore3D (v2016)

Training 0.621 (0.077) 0.490 (0.059) 0.531 (0.116) 0.957 (0.021)

Validation 1.479 (0.020) 1.182 (0.013) 1.435 (0.021) 0.692 (0.009)

Core2016 1.509 (0.034) 1.224 (0.031) 1.497 (0.034) 0.725 (0.015)

Core2013 1.676 (0.050) 1.335 (0.040) 1.583 (0.044) 0.713 (0.019)

Table 3: LigityScore evaluation on the CASF-2013 Scor-

ing Power benchmark ranked using the Pearson Correlation

Coefﬁcient, R. Our results are in bold achieving 5

and 8

placings from the scoring functions listed in the benchmark,

as well as other literature marked with (*) where authors

also used the CASF-2013 benchmark for evaluation. En-

tries without an (*) are taken directly from Li et al. (2014a)

– only the top 10 are included.

Scoring Function Rank SD R

AGL* (Nguyen and Wei, 2019a) 1 1.45 0.792

LearningLigand* NNScore+RDkit

(Boyles et al., 2020)

2 - 0.786

OnionNet* (Zheng et al., 2019) 3 1.45 0.782

EIC-Score* (Nguyen and Wei, 2019b) 4 - 0.774

PLEC-nn* (W

ojcikowski et al., 2019) 4 1.43 0.774

LigityScore3D 5 1.58 0.713

Pafnucy*

(Stepniewska-Dziubinska et al., 2017)

6 1.61 0.700

DeepBindRG* (Zhang et al., 2019) 7 - 0.639

LigityScore1D 8 1.74 0.635

X-Score 9 1.77 0.622

X-ScoreHS 10 1.77 0.620

X-ScoreHM 11 1.78 0.614

X-ScoreHP 12 1.79 0.607

dSAS 13 1.79 0.606

ChemScore@SYBYL 14 1.82 0.592

ChemPLP@GOLD 15 1.84 0.579

PLP1@DS 16 1.86 0.568

PLP2@DS 17 1.87 0.558

GScore@SYBYL 18 1.87 0.558

* other literature using CASF-2013 benchmark

single holdout validation set. Table 2 summarises

these average results for LigityScore1D trained using

PBDbind v2016 and PDBbind v2018, and for Lig-

ityScore3D using PDBbind v2016. LigityScore3D

shows a signiﬁcant performance improvement for the

Core-2013 model with an average of 0.713 R-score

that is well above the 0.657 and 0.635 achieved for

LigityScore1D trained on PDBBind v2016 and PDB-

Bind v2018 respectively. On the other hand the results

for Core-2016 for LigityScore3D shows comparable

Table 4: LigityScore evaluation on the CASF-2016 Scor-

ing Power benchmark ranked using the Pearson Correlation

Coefﬁcient, R. Our results are in bold achieving 7

and 8

placings from the scoring functions listed in the benchmark,

as well as other literature marked with (*) where authors

also use the CASF-2016 benchmark. Entries without an (*)

are taken directly from Su et al. (2018) – only the top 10 are

included.

Scoring Function Rank SD R

AGL* (Nguyen and Wei, 2019a) 1 - 0.830

EIC-Score* (Nguyen and Wei, 2019b) 2 - 0.826

LearningLigand NNScore+RDkit

(Boyles et al., 2020)

2 - 0.826

deep

* (Jim

enez et al., 2018) 3 - 0.820

PLEC-nn* (W

ojcikowski et al., 2019) 4 1.26 0.817

OnionNet* (Zheng et al., 2019) 5 1.26 0.816

∆VinaRF20 5 1.26 0.816

Pafnucy*

(Stepniewska-Dziubinska et al., 2017)

6 1.37 0.780

LigityScore1D 7 1.46 0.741

LigityScore3D 8 1.50 0.725

X-Score 9 1.69 0.631

X-ScoreHS 10 1.69 0.629

∆SAS 11 1.70 0.625

X-ScoreHP 12 1.70 0.621

ASP@GOLD 13 1.71 0.617

ChemPLP@GOLD 14 1.72 0.614

X-ScoreHM 15 1.73 0.609

Autodock Vina 16 1.73 0.604

DrugScore2018 17 1.74 0.602

* other literature using CASF-2016 benchmark

performance to the LigityScore1D (PDBBind v2018)

models with only 0.01 difference in R-score. Due to

the similarity in results obtained for both CASF-2013

and CASF-2016, LigityScore3D is chosen as the best

performing model with R-score of 0.725 and 0.713.

The additional scoring power of approximately 10%

for CASF-2013 comes at the expense of a more com-

plex network. The LigityScore3D model has 94M

learnable parameters, whilst the best model for Lig-

ityScore1D has only 3.9M parameters.

BIOINFORMATICS 2021 - 12th International Conference on Bioinformatics Models, Methods and Algorithms

Figure 5: Experiment vs Predicted Binding Afﬁnity in pK

for best LigityScore3D model. The left scatter plot represents the

core 2016 set, whilst the right scatter plot represents the core 2013 set.

The scatter plots for the predicted afﬁnity versus

the experimental afﬁnity for best performing Ligi-

tyScore3D model is shown in Figure 5. The scat-

ter plots represent the Core-2016 (left) and the Core-

2013 (right) sets showing good correlation between

predicted and experimental afﬁnities. The ideal model

would produce a plot of function y = x, as the pre-

dicted afﬁnity should be equal to the experimental

value. The ranking of LigityScore for CASF-2013

and CASF-2016 are presented in Tables 3 and 4.

Apart from the scoring function evaluated directly in

CASF, Tables 3 and 4 include other scoring functions,

marked with an asterix (*), that represent results re-

ported in literature (in individual publications) that

also utilise the CASF benchmarks. Tables 3 and 4

thus provide, to the best of our knowledge, a compre-

hensive list of the scoring functions developed in re-

cent years to date, that compare and rank the different

scoring functions available. LigityScore3D achieved

place in the CASF-2013 benchmark with an av-

erage R-score of 0.713, and exceeds the reported

CASF-2013 score for Pafnucy. On the CASF-2016

benchmark, LigityScore models achieve the 7

and

places.

4 CONCLUSIONS

In this study we explored the use of CNNs to develop

a scoring function, called LigityScore, for binding

afﬁnity prediction. Machine learning scoring func-

tions have been developed to address the limitations

of classical models, such as the use of linear models,

imposed functional form, and their inability to learn

from new data. However, conventional ML based

scoring functions still rely on a degree of feature engi-

neering that requires expert knowledge to preprocess

the data. This, in turn, led to the introduction of deep

learning methods. To this effect we have developed

two different protein-ligand representations that are

extracted directly from the 3D structure of both the

protein and ligand using pharmocophoric features.

The choice of representation of the protein-ligand

structure determines the ﬂexibility and expressiveness

that the model is able to learn and ultimately its scor-

ing power. Although deep learning methods extract

features automatically during training, correct repre-

sentation of the complex is critical for the feature ex-

traction ability of the DL model. As a point for im-

provement for LigityScore performance, future work

would focus on the data representation component of

the protein-ligand complex to build on the existing

representation and possibly seek ways to incorporate

alternate types of features within LigityScore. In this

regard one of the research tasks would be to look into

additional pharmacophoric feature families (or types)

that could help create an enriched descriptor. Other

features such as the spatial distribution count for dis-

tances between key atom combinations could be con-

sidered as another dimension to the LigityScore rep-

resentation. Additionally, since CNNs are difﬁcult to

interpret, in future work we would apply techniques

such as SHAP (Lundberg and Lee, 2017) to determine

critical features used for predictions.

The major contribution for this study is in the pre-

sentation of a novel protein-ligand representation for

use as a CNN scoring function for binding afﬁnity

prediction adapted from Ebejer et al. (2019). Rep-

resentation engineering is required when using CNN

for SBVS as the data needs to represent the protein-

ligand structure. Representation engineering is nec-

essary since the protein-ligand complex cannot be in-

put directly into the CNN as in the case of an im-

age. In our approach we use spatial distances be-

tween key pharmacophoric features which is simpler

than creating a mathematical model to describe the

protein-ligand interactions. LigityScore still relies on

the automatic feature extraction of CNNs for feature

LigityScore: Convolutional Neural Network for Binding-afﬁnity Predictions

extraction. Since LigityScore is based on distances

between pharmacophoric features, it also presents a

rotationally invariant representation. Additionally,

the method shows relatively good performance that

marginally exceed the Pafnucy R-score performance

on the CASF-2013 benchmark by 0.01 on average, us-

ing a less computationally complex model that can be

trained 16 times faster. The LigityScore models can

potentially be used for afﬁnity predictions for novel

molecules, and as a scoring function for docking in

virtual screening.

A recent paper by Shen et al. (2020) highlights

the importance of assessing the scoring function in all

four powers (scoring, ranking, docking, and screen-

ing) of the CASF benchmark for a 360 degree per-

formance evaluation. Due to the recent release of

Shen et al. (2020) it was not possible to extend eval-

uation of LigityScore on the rest of the powers. This

is a limitation in the sense that these results are not

known, and future work would consider testing Ligi-

tyScore for the other powers in the CASF benchmark.

Recent literature for deep learning scoring functions

also focused on only the scoring power aspect such as

Stepniewska-Dziubinska et al. (2017), Jim

enez et al.

(2018), and Zheng et al. (2019), and therefore a simi-

lar approach was taken. Shen et al. (2020) has shown

that Pafnucy and OnionNet do not perform well on the

rest of the benchmark powers, and even report perfor-

mance lower than the classical functions.

Although the ideal scoring function should per-

form well in all CASF benchmark powers, we argue

that this is not necessarily the case and a particular

ML scoring function may not be suited for every sce-

nario. Therefore a different version, trained for a par-

ticular power, may be better suited. As an example,

the screening power would require the scoring func-

tion model to differentiate between actives and inac-

tives. However, the models trained with the PDB-

bind dataset do not include any inactive information.

Due to the lack of experimentally-validated inactives

there are no evaluation datasets that include inactive

molecules highlighting the need for better and more

complete datasets (Sieg et al., 2019). ML models,

including DL, use learning by representation to ex-

tract the underlying function in the data. If the dataset

does not include the inactive class it is intuitive that

the model may not respond well when presented with

inactive molecules. A ML scoring function can be

developed to cater for the particular power, leverag-

ing on the ﬂexibility they provide to adjust and derive

their parameters from the given training data automat-

ically.

Finding a suitable representation of the protein-

ligand complex is a major challenge when building a

scoring function, and is key for accurate predictions

using deep learning techniques. In our work we have

successfully found a suitable representation that to the

best of our knowledge was never used for binding

afﬁnity prediction, which provides good results and

ranked 5

in the CASF-2013 benchmark. Therefore,

although our work did not outperform the top scoring

function we deem it is still a valid contribution to the

area and may be further enhanced in future work, or

may also serve as motivation and inspiration for other

researchers to seek out alternative methods that in-

crease the effectiveness of scoring functions and vir-

tual screening in general.

We believe a deeper understanding of CNN in the

domain of SBVS is still required, and a breakthrough

like the work of Krizhevsky et al. (2012) in the com-

puter vision domain is still being sought after in this

challenging domain. Nonetheless, we also believe

that ML and DL techniques will lead the future of the

development of scoring functions.

ACKNOWLEDGEMENTS

We would like to thank the AWS Research Credits

Team for supporting our research with AWS credits

to develop our models.

REFERENCES

Ain, Q. U., Aleksandrova, A., Roessler, F. D., and Ballester,

P. J. (2015). Machine-learning scoring functions to

improve structure-based binding afﬁnity prediction

and virtual screening. Wiley Interdisciplinary Re-

views: Computational Molecular Science, 5(6):405–

424.

Berman, H., Henrick, K., and Nakamura, H. (2003). An-

nouncing the worldwide protein data bank. Nature

Structural & Molecular Biology, 10(12):980–980.

Boyles, F., Deane, C. M., and Morris, G. M. (2020). Learn-

ing from the ligand: using ligand-based features to

improve binding afﬁnity prediction. Bioinformatics,

36(3):758–764.

Ching, T., Himmelstein, D. S., Beaulieu-Jones, B. K.,

Kalinin, A. A., Do, B. T., Way, G. P., Ferrero, E.,

Agapow, P.-M., Zietz, M., Hoffman, M. M., et al.

(2018). Opportunities and obstacles for deep learn-

ing in biology and medicine. Journal of The Royal

Society Interface, 15(141):20170387.

Goldberg, Y. (2016). A primer on neural network models

for natural language processing. Journal of Artiﬁcial

Intelligence Research, 57:345–420.

He, K., Zhang, X., Ren, S., and Sun, J. (2015). Delving

deep into rectiﬁers: Surpassing human-level perfor-

mance on imagenet classiﬁcation. In 2015 IEEE In-

BIOINFORMATICS 2021 - 12th International Conference on Bioinformatics Models, Methods and Algorithms

ternational Conference on Computer Vision (ICCV),

pages 1026–1034.

Jim

enez, J., Skalic, M., Martinez-Rosell, G., and De Fabri-

tiis, G. (2018). K deep: protein–ligand absolute bind-

ing afﬁnity prediction via 3d-convolutional neural net-

works. Journal of chemical information and model-

ing, 58(2):287–296.

Kaggle, M. (2012). Kaggle: Merck molecular activity chal-

lenge. https://www.kaggle.com/c/MerckActivity, Ac-

cessed Feb 8, 2019.

Kingma, D. P. and Ba, J. (2014). Adam: A

method for stochastic optimization. cite

arxiv:1412.6980Comment: Published as a con-

ference paper at the 3rd International Conference for

Learning Representations, San Diego, 2015.

Landrum, G. (2020). Rdkit: Open-source cheminformatics.

Accessed April, 2020.

Leach, A. R., Gillet, V. J., Lewis, R. A., and Taylor, R.

(2010). Three-dimensional pharmacophore methods

in drug discovery. Journal of medicinal chemistry,

53(2):539–558.

LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learn-

ing. nature, 521(7553):436–444.

Li, Y., Liu, Z., Li, J., Han, L., Liu, J., Zhao, Z., and Wang,

R. (2014). Comparative assessment of scoring func-

tions on an updated benchmark: 1. compilation of the

test set. Journal of chemical information and model-

ing, 54(6):1700–1716.

Liu, Z., Cui, Y., Xiong, Z., Nasiri, A., Zhang, A., and

Hu, J. (2019). Deepseqpan, a novel deep convolu-

tional neural network model for pan-speciﬁc class i

hla-peptide binding afﬁnity prediction. Scientiﬁc re-

ports, 9(1):794.

Liu, Z., Su, M., Han, L., Liu, J., Yang, Q., Li, Y., and

Wang, R. (2017a). Forging the basis for developing

protein–ligand interaction scoring functions. Accounts

of chemical research, 50(2):302–309.

Liu, Z., Su, M., Han, L., Liu, J., Yang, Q., Li, Y., and

Wang, R. (2017b). Forging the basis for developing

protein–ligand interaction scoring functions. Accounts

of chemical research, 50(2):302–309.

Lundberg, S. M. and Lee, S.-I. (2017). A uniﬁed approach

to interpreting model predictions. In Advances in Neu-

ral Information Processing Systems 30, pages 4765–

4774. Curran Associates, Inc.

Nguyen, D. D. and Wei, G.-W. (2019a). Agl-score: Al-

gebraic graph learning score for protein–ligand bind-

ing scoring, ranking, docking, and screening. Journal

of chemical information and modeling, 59(7):3291–

3304.

Nguyen, D. D. and Wei, G.-W. (2019b). Dg-gl: Differen-

tial geometry-based geometric learning of molecular

datasets. International journal for numerical methods

in biomedical engineering, 35(3):e3179.

Ozt

urk, H.,

Ozg

ur, A., and Ozkirimli, E. (2018). Deepdta:

deep drug–target binding afﬁnity prediction. Bioinfor-

matics, 34(17):i821–i829.

Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J.,

Chanan, G., Killeen, T., Lin, Z., Gimelshein, N.,

Antiga, L., Desmaison, A., Kopf, A., Yang, E., De-

Vito, Z., Raison, M., Tejani, A., Chilamkurthy, S.,

Steiner, B., Fang, L., Bai, J., and Chintala, S. (2019).

Pytorch: An imperative style, high-performance deep

learning library. In Advances in Neural Information

Processing Systems 32, pages 8024–8035. Curran As-

sociates, Inc.

erez-Sianes, J., P

erez-S

anchez, H., and D

ıaz, F. (2019).

Virtual screening meets deep learning. Current

computer-aided drug design, 15(1):6–28.

Ragoza, M., Hochuli, J., Idrobo, E., Sunseri, J., and Koes,

D. R. (2017). Protein–ligand scoring with convolu-

tional neural networks. Journal of chemical informa-

tion and modeling, 57(4):942–957.

Rifaioglu, A. S., Atas, H., Martin, M. J., Cetin-Atalay, R.,

Atalay, V., and Dogan, T. (2018). Recent applications

of deep learning and machine intelligence on in silico

drug discovery: methods, tools and databases. Brief.

Bioinform, 10.

Sieg, J., Flachsenberg, F., and Rarey, M. (2019). In need

of bias control: Evaluating chemical data for machine

learning in structure-based virtual screening. Jour-

nal of chemical information and modeling, 59(3):947–

961.

Stepniewska-Dziubinska, M. M., Zielenkiewicz, P., and

Siedlecki, P. (2017). Pafnucy–a deep neural network

for structure-based drug discovery. stat, 1050:19.

Su, M., Yang, Q., Du, Y., Feng, G., Liu, Z., Li, Y., and

Wang, R. (2018). Comparative assessment of scoring

functions: the casf-2016 update. Journal of chemical

information and modeling, 59(2):895–913.

Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S.,

Anguelov, D., Erhan, D., Vanhoucke, V., and Rabi-

novich, A. (2015). Going deeper with convolutions.

In Proceedings of the IEEE conference on computer

vision and pattern recognition, pages 1–9.

Tompson, J., Goroshin, R., Jain, A., LeCun, Y., and Bregler,

C. (2015). Efﬁcient object localization using convo-

lutional networks. In Proceedings of the IEEE Con-

ference on Computer Vision and Pattern Recognition,

pages 648–656.

Ulyanov, D., Vedaldi, A., and Lempitsky, V. (2017). Im-

proved texture networks: Maximizing quality and di-

versity in feed-forward stylization and texture synthe-

sis. In The IEEE Conference on Computer Vision and

Pattern Recognition (CVPR). IEEE.

ojcikowski, M., Kukiełka, M., Stepniewska-Dziubinska,

M. M., and Siedlecki, P. (2019). Development of

a protein–ligand extended connectivity (plec) ﬁnger-

print and its application for binding afﬁnity predic-

tions. Bioinformatics, 35(8):1334–1341.

Zhang, H., Liao, L., Saravanan, K. M., Yin, P., and Wei,

Y. (2019). Deepbindrg: a deep learning based method

for estimating effective protein–ligand afﬁnity. PeerJ,

7:e7362.

Zheng, L., Fan, J., and Mu, Y. (2019). Onionnet: a multiple-

layer intermolecular-contact-based convolutional neu-

ral network for protein–ligand binding afﬁnity predic-

tion. ACS omega, 4(14):15956–15965.

LigityScore: Convolutional Neural Network for Binding-afﬁnity Predictions