DSAGLSTM-DTA: PREDICTION OF DRUG-TARGET AFFINITY USING DUAL SELF-ATTENTION AND LSTM

Machine Learning and Applications: An International Journal (MLAIJ) Vol.9, No.2, June 2022
DOI:10.5121/mlaij.2022.9201 1
DSAGLSTM-DTA: PREDICTION OF
DRUG-TARGET AFFINITY USING DUAL
SELF-ATTENTION AND LSTM
Lyu Zhijian, Jiang Shaohua and Tan Yonghao
College of Information Science and Engineering,
Hunan Normal University, Changsha, China
ABSTRACT
The research on affinity between drugs and targets (DTA) aims to effectively narrow the target search
space for drug repurposing. Therefore, reasonable prediction of drug and target affinities can minimize the
waste of resources such as human and material resources. In this work, a novel graph-based model called
DSAGLSTM-DTA was proposed for DTA prediction. The proposed model is unlike previous graph-based
drug-target affinity model, which incorporated self-attention mechanisms in the feature extraction process
of drug molecular graphs to fully extract its effective feature representations. The features of each atom in
the 2D molecular graph were weighted based on attention score before being aggregated as molecule
representation and two distinct pooling architectures, namely centralized and distributed architectures
were implemented and compared on benchmark datasets. In addition, in the course of processing protein
sequences, inspired by the approach of protein feature extraction in GDGRU-DTA, we continue to
interpret protein sequences as time series and extract their features using Bidirectional Long Short-Term
Memory (BiLSTM) networks, since the context-dependence of long amino acid sequences. Similarly,
DSAGLSTM-DTA also utilized a self-attention mechanism in the process of protein feature extraction to
obtain comprehensive representations of proteins, in which the final hidden states for element in the batch
were weighted with the each unit output of LSTM, and the results were represented as the final feature of
proteins. Eventually, representations of drug and protein were concatenated and fed into prediction block
for final prediction. The proposed model was evaluated on different regression datasets and binary
classification datasets, and the results demonstrated that DSAGLSTM-DTA was superior to some state-of-
the-art DTA models and exhibited good generalization ability.
KEYWORDS
Drug-Target Affinity, BiLSTM, pooling, Graph Neural Network, Self-Attention.
1. INTRODUCTION
According to incomplete statistics, the development of a new drug that can obtain marketing
authorization is expected to cost hundreds of millions of dollars, and the rate of drug approval for
clinical trials is only about 10%. Furthermore, due to the bottleneck of technological
development, the development of new drugs is more difficult. Driven by these factors,
researchers have to explore novel and more efficient approaches in drug discovery. Under this
circumstances, the exploration of new uses of developed drugs has become a new hot spot.
Discovering new associations between drugs and targets is critical for drug development and
repurposing, However, the traditional study of drug-protein relationships in the wet laboratory
[1][2] is time-consuming and expensive due to the huge range of chemical spaces to be searched,
to solve this problem, some virtual screening(VS) has been proposed to accelerate the
experimental drug discovery and reposition studies in silico [3], some of the more commonly

2
used VS methods, like structure-based VS, ligand-based VS and sequence-based VS have
contributed to drug development to a large extent [4][5]. However, these VS methods have their
own defects in application. For example, if the structural information of the protein is unknown,
the structure-based approach cannot play its role. There is still a long way to go before accurately
constructing the structure of proteins, to this end, some structure-free methods have sprung up.
In recent years, with the development and maturity of deep learning technology and its great
breakthrough in the field of computer vision(CV) and natural language processing(NLP) [6][7],
many people in the field of drug research have begun to turn their attention to deep learning.
Methods based on deep learning to study drug-target relationships are usually computer-aided
methods, and these methods can effectively speed up the virtual screening process of potential
drug molecules because they can minimize unnecessary biological and chemical experiments by
adjusting the search space. Moreover, with the advent of more and more biological activity data,
a great deal of computer-aided work based on these data has been carried out to investigate the
relationship between drugs and targets. These work is usually divided into two categories, one is
binary classification-based approach, that is, to determine whether drugs and targets interact
through data labels(i.e., active or inactive), and the other is a regression-based approach, which
uses binding tightness(specific values) to describe the relationship between the drug and the
target.
In binary classification-based drug-target (DT) prediction tasks, deep learning technologies seem
to be used by more researches to deal with drug-target interaction (DTI) problems. When doing
DTI prediction tasks in the past, compounds and proteins were represented using manually
crafted descriptors and the final interaction prediction was made through several fully connected
networks [8][9]. The problem with this approach is that the descriptors are designed from a
specific perspective, that is, the design angle is too single, in addition, it remains fixed during the
training process, so it cannot learn and adjust according to the results, and thus cannot extract
task-related features. Therefore, some end-to-end models were proposed. Du et al. proposed a
model called wide-and-deep to predict DTIs [10] where a generalized linear model and a deep
feed-forward neural network were integrated to enhance the precise of DTI prediction. Molecular
structural information is also of great significance for feature extraction. To learn the mutual
interaction features of atoms in a sequence, Shin et al. proposed a Transformer-based DTI model
[11], which used multilayer bidirectional Transformer encoders [12] to learn the high-
dimensional structure of molecules from the Simplified Molecular Input Line Entry System
(SMILES) strings. Some researchers obtained structural information of compounds or proteins
from another perspective, they represented the corresponding compounds or proteins as graphs
and utilized graph neural networks to extract their spatial features, related work such as
GraphCPI [13], Graph-CNN [14], etc.
Compared with the binary classification model, it seems more convincing to describe the
relationship between drug and target through a regression task, the use of regression model can
provide us with more information about the relationship between compounds and proteins, since
continuous values can tell us how strongly the two are bound. In the early stages of this field,
drugs and proteins were represented by researchers using human experience or skillfully designed
mathematical descriptors. Related studies include KronRLS [15] and SimBoost [16], both of
which based on regression and utilized the similarity information of drugs and targets to predict
DTAs. Although these approaches achieve good results in DTA tasks, they rely on chemical
insights or expert experience, which in turn limits further optimization of these models. What’s
more, the rapid advancement in deep learning has also largely facilitated the affinity prediction of
DT pairs and various data-driven models are applied to the description of drugs and targets.
DeepDTA [17] is the first framework for predicting drug and target affinity based on deep
learning, which utilized two CNN blocks to process SMILES strings of drugs and amino acid

3
sequences of proteins, respectively. Works related to DeepDTA include WideDTA [18]. The
improvement of WideDTA over DeepDTA is that it combined several characters as words and
proposed a word-based sequence representation method. In order to better capture the topological
structure features of compounds, Nguyen et al. proposed GraphDTA [19] to predict drug and
target affinity which utilized RDKit technology to represent drug strings as graphs that could
reflect its structural characteristics, and used graph convolutional neural network to extract its
spatial features. Furthermore, Lin proposed a similar approach called DeepGS [20], which used
advanced techniques to encode amino acid sequences and SMILES strings. DeepGS also
combined a GAT model to capture the topological information of molecular graph and a BiGRU
model to obtain the local chemical context of drug.
In the above two categories of research tasks, a new mechanism called attention is increasingly
gaining the favor of researchers, since the performance of either the binary classification-basedor
regression-based approach can be enhanced by introducing attention mechanisms. Examples of
two typical applications are AttentionDTA [21] and HyperAttentionDTI [22], the novelty of
HyperAttentionDTI and AttentionDTA lies in that they utilized an attention mechanism for
learning important parts of each other's sequences. Zhang et al. proposed a different attention
mechanism in SAG-DTA [23], which is based on graph structure rather than sequence. In SAG-
DTA, a self-attention pooling network was used to learn the structural features of drug molecular
graphs, in which the features of each atom node in the molecular graph were weighted using an
attention score before being aggregated as molecule representation. In this work, inspired by the
above attention mechanism, we proposed a novel framework based on self-attention to predict
DTAs. In the process of extracting molecular graph features of drugs, the same as SAG-DTA, we
utilized a novel attention structure that introduces self-attention mechanisms for node pooling
named self-attention graph pooling (SAGPool) [24], in which the self-attention graph pooling
approach is adopted to molecular graph representation. Moreover, two different self-attention
network architectures called centralized pooling and distributed pooling were constructed and
compared. For protein sequence feature extraction, due to the limitation of the convolution kernel
size, the field of vision of the convolutional neural network (CNN) is limited, so it is difficult to
effectively capture the long amino acid sequence features that are context-dependent, so we
usedbidirectional Long Short-Term Memory(BiLSTM) networks to extract protein features. On
this basis, we applied another self-attention mechanism to protein sequences, we combined the
final hidden states of each element in the batch with each unit output of the LSTM as the final
features of proteins to obtain comprehensive representations. Experimental results demonstrated
that our model greatly improves the performance compared to previous models.
2. MATERIALS AND METHODS
DSAGLSTM-DTA is a computation-based end-to-end deep learning algorithm that takes the
features of drugs and targets as inputs and affinity/interaction values between them as output.
DSAGLSTM-DTA contains two prediction tasks, DTA and DTI, in which the context data of
drug-protein pairs are input into the network for feature extraction, which is then used by the
algorithm of the model to evaluate the inner association, and the predicted value is obtained. In
this work, a more complicated graph neural network was implemented by using the self-attention
pooling mechanism. Specifically, the features of the nodes were learned by attention scores to
weight the atom nodes. Furthermore, attention scores were also utilized to sort and filter atom
nodes. For the protein feature extraction, we hypothesized that each hidden layer cell of LSTM
actually contains some extra features of protein, so it should not be ignored, therefore, we
performed weighted attention calculation between the output of LSTM and its hidden layer cell to
effectively capture its comprehensive representation. We believed that the above strategy will
enable the network to pay more attention to the most important parts and thus obtain a more pure

4
feature representation for the prediction task. The overall flow of DSAGLSTM-DTA is expressed
in Figure 1. As can be seen in Figure 1, the SMILES strings of drugs are preprocessed into
molecular graphs, which are then fed into graph neural network with self-attention pooling for
feature extraction. For the protein, the amino acid sequences of proteins are fed into the
bidirectional LSTM network for contextual feature extraction. Subsequently, the output of the
LSTM is weighted with the state of final hidden layer, and the result of the calculation is
functioned as the final feature representation.
CCN(CCNCl)CCNC(=O)
Drug SMILES
MQGGNMSWSPSDPY
Protein sequences
MOLECULAR GRAPH
BiLSTM
Attention
Final hidden state
Attention-
GCN
Output
Prediction
Protein
Representation
Drug
Representation
Fig. 1. Overall flow of DSAGLSTM-DTA.
2.1. Datasets
In our experimental evaluation, we used the two datasets most commonly used in DTAs
prediction, namely Davis [25] and KIBA [26]. The Davis stores selectivity as say data for the
kinase protein families and the relevant inhibitors, along with their respective disassociation
constant (Kd) values, moreover, it contains 72 compounds and 442 proteins, and their
corresponding affinity values, where the affinity values are measured by Kd values (kinase
dissociation constant). There are a total of 30056 affinity values in Davis, and they range from
5.0 to 10.8. We converted Kd into the value of the corresponding logarithmic space, pKd, as
follows:
P𝐾d = − log10 (
𝐾d
109
) (1)
The KIBA dataset contains 2116 compounds and 229 proteins, as well as 118,254 drug and target
affinity values, where the affinity values range from 0.0 to 17.2. And the affinity value in KIBA
is represented by KIBA score which is computed by combining heterogeneous information
sources, i.e., IC50, Ki and Kd. The data in this dataset have high quality, since the integrated
heterogeneous metrics alleviate the coupling associated with using a single source of information.
In addition to using a regression task for drug-target relationship prediction, the proposed model
was also applied to a binary classification task for evaluation, and two binary benchmark datasets
called human [27] and BindingDB [28] for CPI prediction were used for the experiments. The
positive CPI pairs in the Human dataset are derived from DrugBank [29] and Matador [30], and
the negative CPI samples in this dataset is highly credible. BindingDB is another commonly used
CPI prediction dataset, which is characterized by containing pre-processed training, validation,
and test sets. The overview information of the four benchmark datasets is summarized in Table 1.

5
Table 1. Summary of the benchmark datasets
Datasets Compound Protein Binding Entities Task Type Ref
Davis 72 442 30056 DTA(regression) [25]
KIBA 2116 229 118254 DTA(regression) [26]
Human 1052 852 3369(+)/3359(-) CPI(binary-class) [27]
BindingDB 53253 1696 39747(+)/31218(-) CPI(binary-class) [28]
2.2. Overview of the Network Architecture
In this section, we will introduce overview network architectures of our model. As mentioned
earlier, DSAGLSTM-DTA contains two different network architecture, namely centralized and
distributed architectures, the difference between these two architectures lies in the different
positions of self-attention pooling in drug feature extraction. The centralized pooling architecture
is presented in the left panel of Figure 2, consists of three graph neural network layers, and the
outputs of each layer are concatenated and fed into the self-attention layer. These reserved nodes
are then fed into a global_max_pool layer and several fully connected layers for final
representations. The distributed pooling architecture is illustrated in the right panel of Figure 2,
the difference between this architecture and the previous is that a single self-attention pooling is
used after each graph convolutional network layer. Subsequently, the outputs of each layer after
self-attention pooling are concatenated, that is, pooling in a distributed way. The above two
architectures are consistent in the structure of the protein network. The features of proteins are
differentiated into two parts (i.e. output and the hidden state) after passing through the
bidirectional LSTM network block, and these two branches are then input into the attention block
for weighting calculation.

6
Graph
Convolution
Graph
Convolution
Graph
Convolution
Concatenate
SAGPool
Global_max_pool
Fully Connected
Dropout
Molecule
Representation
BiLSTM
Attention
Fully Connected
Dropout
Protein
Representation
Combined Representation
Fully Connected
Dropout
Fully Connected
Dropout
Output
a
Final
hidden
state
Output
Graph
Convolution
Graph
Convolution
Graph
Convolution
SAGPool
Global_max_pool
Fully Connected
Dropout
Molecule
Representation
Fully Connected
Dropout
Protein
Representation
Combined Representation
Fully Connected
Dropout
Fully Connected
Dropout
Output
SAGPool
SAGPool
Global_max_pool
Global_max_pool
BiLSTM
Attention
Concatenate
b
Final
hidden
state
Output
Figure 2. Network architectures of DSAGLSTM-DTA. Substructures surrounded by dashed lines indicate
molecular graph representation, which is the major difference of the two architectures. (a) Centralized
pooling architecture. (b) Distributed pooling architecture.
2.2.1. Data Preprocessing
The feature extraction of drugs and targets are two independent input channels. Before drugs and
targets are input into their respective feature extraction blocks, data preprocessing is required for
drugs and targets, respectively. The implementation details are as follows.
2.2.1.1. Drug Representation
For data preprocessing of drugs, we used the same method as GraphDTA, we utilized the open
source technology RDKit to convert the SMILES strings of drugs into corresponding 2D
molecule graphs. The molecule graph was denoted as 𝐺 = (𝑉, 𝐸), and the vertexes 𝑉 were
represented as atoms and the edges 𝐸 were represented as bonds, where |𝑉| = 𝑁 is the number of
nodes in the graph and |𝐸| = 𝑁𝑒
is the number of edges. Each atom was embedded with 78-
dimensional features such as the atom's type, degree, implied valence, aromaticity, and the
number of hydrogen atoms attached to the atom. The feature of the node was encoded as a one-
hot vector of shape (𝑁, 78). The chemical bonds index was encoded as (2, 𝐸) vector, which is
used to store the edges of the undirected graph. The schematic diagram of the SMILES string of a
drug converted into a two-dimensional molecule graph by rdkit technology is as follows:

7
CN1CCN(C(=O)c2cc3cc(Cl)ccc3[nH]
2)CC1
Drug SMILES
Rdkit
Fig. 3. Convert SMILES string to graph.
2.2.1.2. Target Representation
The sequence length of each protein is different and varies greatly. For uniform feature
representation, we fixed the length of all protein sequences as 1000 according to the average
length of protein sequences, if the sequence length of the protein exceeds 1000, the part more
than 1000 will be cut off, and otherwise, the part less than 1000 will be padded with 0. In
addition, protein sequences were represented by different combinations of 25 amino acids, and
each represented by the one-letter code. We then mapped each amino acid to an integer, and each
integer was embedded as a 128-dimensional feature.
2.2.2. Graph Convolution Layer
2.2.2.1. GatedGraph
In the graph neural network block, GatedGraph [31] convolution algorithm was utilized by us to
extract the 2D molecular graph features of drugs and its details are as follows. GatedGraph is a
feature learning technique that studies graph-structured inputs, it modifies previous graph neural
network work using gated recurrent units (GRU) and modern optimization techniques, and then
extends to output sequences, so this method can make full use of long-distance information and
fit well with our model of extracting protein features. In addition, GatedGraph has favorable
inductive biases relative to purely sequence-based models when dealing with graph structure
problems, and thus is a flexible and widely useful class of neural network models. The features of
the node are updated as follows:
hi
(0)
= xi || 0 (2)
m𝑖
(𝑙+1)
= ∑ e𝑗,𝑖 ∙ Θ ∙ h𝑗
(𝑙)
𝑗∈𝑁(𝑖)
(3)
ℎ𝑖
(𝑙+1)
= GRU (𝑚𝑖
(𝑙+1)
, ℎ𝑖
(𝑙)
) (4)
Where in formula (2), h(0)
i is the input state, xi ∈ R𝐹
is the feature of node i, xi || 0 represents
padding 0 after feature xi to the specified dimension. In formula (3), Θ is the parameter matrix to
be learned, that is, the aggregation information of surrounding nodes. Formula (4) is to use a
GRU unit to take the above two formulas as input and get an output, which can be functioned as a
new feature of node i.
2.2.3. LSTM Block
2.2.3.1. LSTM
When CNN is used to extract the context dependences of long sequences, the field of view is
limited due to the influence of the size of convolution kernel, and multiple CNN layers need to be

8
used, which makes the model bloated and complex. In order to overcome the inability of CNN
and RNN to deal with long-distance dependence, LSTM (Long-Short Term Memory) [32] is
proposed. LSTM captures long-term dependencies by controlling the circulation and loss of
features using a "gating" mechanism. As Figure 4 illustrated, LSTM unit is composed of a cell, a
forget gate, an input gate, and an output gate. These gates are responsible for controlling the
interactions among different memory cell.
Fig. 4. Overview of a LSTM unit.
Some important elements contacted to LSTM can be expressed as follows:
𝑓𝑡 = 𝜎(𝑊𝑓𝑥𝑡 + 𝑈𝑓ℎ𝑡−1 + 𝑏𝑓) (5)
The forget gate 𝑓𝑡 is used to evaluate which features of previous cell state should be retained for
calculation. Where 𝜎 is the sigmoid function, through which the data can be transformed into a
value in the range of 0~1 to act as a gating signal. 𝑥𝑡 is the input of the current node, ℎ𝑡−1 is the
hidden state passed down by the previous node, and this hidden state contains the relevant
information of the previous node. 𝑈𝑓 and 𝑊𝑓 are the corresponding weight matrices, respectively.
𝑖𝑡 = 𝜎(𝑊𝑖𝑥𝑡 + 𝑈𝑖ℎ𝑡−1 + 𝑏𝑖) (6)
𝐶
̃𝑡 = tanh(𝑊
𝑐𝑥𝑡 + 𝑈𝑐ℎ𝑡−1 + 𝑏𝑐) (7)
𝐶𝑡 = 𝑓𝑡 ⊗ 𝐶𝑡−1 + 𝑖𝑡 ⊗ 𝐶
̃𝑡 (8)
The input gate 𝑖𝑡 is responsible for updating current information in real time. 𝐶
̃𝑡 represents the
update value of the cell state at the current moment, which is obtained from the input data and the
hidden state through a neural network layer, and the activation function usually uses tanh. ⊗ is
the most important gate mechanism of LSTM, representing the unit multiplication relationship
between two data.
𝑜𝑡 = 𝜎(𝑊
𝑜𝑥𝑡 + 𝑈𝑜ℎ𝑡−1 + 𝑏𝑜) (9)
ℎ𝑡 = 𝑜𝑡 ⊗ tanh (𝐶𝑡) (10)
The output gate 𝑜𝑡 controls the output of the cell state into the rest of the network. The hidden
state ℎ𝑡 is obtained by the output gate 𝑜𝑡 and the cell state 𝐶𝑡. The calculation method of 𝑜𝑡 is the
same as that of 𝑓𝑡 and 𝑖𝑡. It is worth noting that in Equation (9), the LSTM can be made to
approximate its variant GRU by initializing the mean of 𝑏𝑜 to 1 [33].
The feature extraction process of LSTM can be shown in Fig. 5.

9
2.2.3.2. BiLSTM
For some specific tasks, the information at a certain moment is not only related to the previous
state, but also has some connection with the later state. When dealing with such problems, the
traditional unidirectional LSTM is obviously not competent, therefore, the bidirectional LSTM is
introduced. For protein sequences, we considered that the features of a certain part of the protein
sequence are not only related to the previous part, but also related to the later part. Therefore, in
our work, we used bidirectional LSTM networks to extract the amino acid sequence features of
the proteins.
LSTM Unit
c0
h0
y1
x1
LSTM Unit
y2
x2
LSTM Unit
yN
xN
LSTM Unit
yN-1
XN-1
cN
hN
c1
h1
c2
h2
cN-2
hN-2
cN-1
hN-1
Fig. 5, the output of each stage is jointly determined by the hidden and cell state of its
previous stage and the input of the current stage.
BiLSTM is composed of two unidirectional LSTMs with opposite directions. At each moment,
the input will fuse the outputs of the two opposite LSTMs at the same time, and the output is
jointly determined by these two unidirectional LSTMs. The feature extraction process of
BiLSTM is as follows:
ht
⃗⃗⃗ = LSTM(xt,ht−1
⃗⃗⃗⃗⃗⃗⃗⃗ , Ct−1
⃗⃗⃗⃗⃗⃗⃗⃗ ) (9)
ht
⃖⃗⃗⃗ = LSTM(xt,ht−1,
⃖⃗⃗⃗⃗⃗⃗⃗⃗⃗ Ct−1
⃖⃗⃗⃗⃗⃗⃗⃗⃗) (10)
ht = wtht
⃗⃗⃗ + vtht
⃖⃗⃗⃗ + bt (11)
Where the function LSTM () represents a series of LSTM operations of the network layer. Ct−1
⃗⃗⃗⃗⃗⃗⃗⃗
and Ct−1
⃖⃗⃗⃗⃗⃗⃗⃗⃗represent the cell state of the previous stage. ht
⃗⃗⃗ and ht
⃖⃗⃗⃗ represent the hidden layer
state in the corresponding direction, respectively. wt and vt represent the weights corresponding
to the forward hidden layer state ht
⃗⃗⃗ and reverse hidden layer state ht
⃖⃗⃗⃗ of the bidirectional LSTM at
time t, respectively. bt represents the bias corresponding to the hidden layer state at time t. The
feature extraction process of BiLSTM networks can be shown in Fig. 6.

10
LSTM Unit
cN
hN
y1
x1
LSTM Unit
y2
x2
LSTM Unit
yN
xN
LSTM Unit
yN-1
xN-1
c0
h0
cN-1
hN-1
c2
h2
cN-2
hN-2
c1
h1
LSTM Unit
c0
h0
LSTM Unit LSTM Unit
LSTM Unit
cN
hN
c1
h1
cN-2
hN-2
c2
h2
cN-1
hN-1
Fig. 6. the output of each stage is jointly determined by the hidden state and cell state of its previous and
subsequent stages and the input of the current stage.
2.2.4. Self-Attention Graph Pooling Block
The self-attention graph pooling (SAGPool) was first proposed by Lee et al[24]. It was
subsequently applied by SAG-DTA to molecular graph feature extraction for drug and target
prediction. The innovations of SAGPool are: 1.The SAGPool method can learn a hierarchical
representation in an end-to-end manner with relatively few parameters; 2.Self-attention
mechanisms are utilized to distinguish the nodes that should be deleted from those that should be
kept; 3.Self-attention mechanisms based on graph convolution to calculate attention scores, and
node features and graph topological structure are taken into account. In general, in SAGPool, a
total of four graph convolution methods are used to obtain the self-attention score of each node in
the molecular graph, then these nodes are ranked according to the corresponding attention score
and a certain ratio is used to determine the preserved atoms. Finally, the pooled molecular graph
is obtained through the mask operation. The process of self-attention pooling is illustrated in
Figure 7.
Graph
Convolution
Activation
Function
Top-rank
Selection
Output
Graph
Masking
Attention
Mask
Input
Graph
Fig. 7. the self-attention scores of the input graph are obtained using the graph convolution method, the
nodes corresponding to the self-attention scores are ranked, and finally the mask operation is applied to
obtain the output graph.
Since there are various graph convolution methods for obtaining node self-attention scores in
SAGPool, according to SAG-DTA, we used GraphConv (Graph convoluton) as our final scoring
method. The GraphConv scoring method is presented as Equation (12).
Z = 𝜎(ℎ𝑣Θ1 + ∑ ℎ𝑢Θ2
𝑢∈𝑁(𝑣)
) (12)

11
Where h represents the feature of the corresponding node, N(v) is the set of all nodes adjacent to
node v, Θ1 and Θ2 correspond to trainable convolutional weights with specific feature dimension.
𝜎 represents the activation function ReLU.
The mask operation is based on the self-attention scores of the atoms computed by the scoring
method to determine which parts in the input graph should be reserved. The mask operation is
illustrated in Equation (13).
idx = top − rank(Z, ⌈𝑘𝑁⌉), 𝑍𝑚𝑎𝑠𝑘 = 𝑍𝑖𝑑𝑥 (13)
Where the k ∈ (0, 1] is the pooling ratio to decide which portion of nodes should be retained. Z
represents the self-attention scores of nodes. ⌈𝑘𝑁⌉ is the number of reserved nodes, and the order
of nodes is ranked according to self-attention score.
2.2.5. Attention Block
The BiLSTM layer generates new protein representation P ∈ 𝑅𝑀×𝑁×𝐹
and final hidden state H ∈
𝑅2×𝑀×𝑓
, where N represents data batch, M is the length of the protein sequence, f and F are the
output sequence feature dimensions. Subsequently, the two matrices are multiplied after being
transformed, and we obtain an attention matrix with weight called A ∈ 𝑅𝑁×𝑀
. After that, the
elements of the weight matrix are mapped to the interval (0, 1) by utilizing the softmax activation
function. Finally, the feature matrix of the protein 𝑃′
∈ 𝑅𝑁×𝑀×𝐹
is weighted by the activated
weight matrix A′
to obtain the final representation of the protein 𝑃𝑓𝑖𝑛𝑎𝑙 ∈ 𝑅𝑁×𝐹
. The related
operations can be formulated as Equation (14-17).
A = 𝑃′
× 𝐻′
(14)
A′
= 𝑆𝑜𝑓𝑡𝑚𝑎𝑥(𝐴) (15)
𝑃𝑓𝑖𝑛𝑎𝑙 = 𝑃′
× 𝐴′
(16)
𝑃𝑟𝑜𝑡𝑒𝑖𝑛 = 𝐹(𝑊
𝑝𝑃𝑓𝑖𝑛𝑎𝑙 + 𝑏𝑝) (17)
Where 𝑃′
and 𝐻′
represent the matrix of protein features and hidden states after a series of matrix
transformations. The series of operations include dimension increase and reduction, matrix
transposition, etc. Softmax is a nonlinear activation function, which is usually used for
classification tasks. × stands for matrix multiplication. F(∙) is a non-linear activation function
(e.g., ReLU), 𝑊
𝑝 ∈ 𝑅𝐹′×𝐹
is the weight matrices, and 𝑏𝑝 is the bias vector.
2.2.6. MLP Block
The features of the drug and protein are concatenated after being extracted and then fed into the
Multilayer Perceptron block for final prediction. The Multilayer Perceptron block consists of two
fully connected layers, each of which is followed by a Dropout of rate 0.5 to prevent over fitting.
The activation function of fully connected layer is the Rectified Linear Unit (ReLU). The output
of the last layer identifies the final predicted affinity value for the drug and protein.
2.3. Implementation
DSAGLSTM-DTA was implemented in Pytorch[34] and its extension library PyTorch Geometric
(PyG)[35]. We used the Adam optimizer with the default learning rate of 2e-4 for regression

12
tasks and 1e-3 for binary classification tasks. In addition, the number of training epochs for
regression task and binary classification task was set to 2000 and 1000, respectively. The batch
size was set to 512 and dropout rate was set to 0.5. According to the experimental conclusion of
SAG-DTA, the pooling ratio of SAGPool was set as 1.
The SMILES string for each drug was converted into 2-dimensional molecular graph where each
node of the molecular graph was embedded with 78-dimensional features. For the centralized
architecture, GNN block consists of three stacked GNN layers with 78, 156 and 312 output
features, respectively. Features extracted by each layer of graph convolution were concatenated
and filtered through a self-attention pooling layer, which then followed by a global max pooling
layer to get the most striking features. The distributed architecture contains three same graph
neural networks with the 78 output feature, which is same with the hierarchical architecture used
by SAG-DTA and Lee et al. However, the output of each layer of graph convolution is directly
concatenated after being filtered by its own self-attention pooling layer instead of continuing to
the next layer, which is different from the hierarchical structure. The purpose of our design is to
expect the input graph will not lose important information due to excessive screening. The
features extracted by the above two structures get the final drug representation after passing
through two fully connected layers with 1024 and 128 neurons respectively.
The protein input embedding is of size 128, which means that we represent each character in
amino acid sequence with a 128-dimensional dense vector. The number of layers of BiLSTM was
set to 1, and the hidden feature dimension of LSTM was 32. The output and hidden state of the
LSTM were weighted and passed through a fully connected network with 128 neurons as the
final representation of the protein.
The prediction block is made up of three fully connected layers, in which the numbers of neurons
are 1024, 512 and 1, respectively. Each drug and protein are converted into a 128-dimensional
vector after their respective feature extraction, and are concatenated into a 256-dimensional
vector for the final prediction. The number of training epochs is set to 2000 for regression tasks
and 1000 for binary classification tasks. Our experiments are run on Windows 10 professional
with Intel(R) Core(TM) i5-10400F CPU @ 2.90GHz and GeForce GTX 1660Ti(6GB).
3. EXPERIMENTS AND RESULTS
3.1. Evaluation Metrics
MSE (Mean Squared Error), CI (Concordance Index) and r2
m (Regression toward the mean) are
the most commonly used evaluation metrics in regression tasks to study drug-target interactions
[15-21]. Therefore, we continue to use these evaluation metrics, the details of each metric are as
follows:
MSE is the mean square error, which is used to measure the gap between the predicted value of
the model and the actual label value. The smaller the gap is, the better the performance of the
model is; otherwise, the worse the performance of the model is.
MSE =
1
𝑛
∑(𝑃𝑖 − 𝑌𝑖)2
𝑛
𝑖=1
(18)
Where Pi is the prediction value, Yi corresponds to the label value and n is the total number of
samples.

13
CI is the Concordance Index, which is a measure of whether the order of predicted binding
affinity values for two random drug-target pairs is consistent with their true values, which value
exceeds 0.8 indicates a strong model.
CI =
1
𝑍
∑ ℎ(𝑝𝑖 − 𝑝𝑗) (19)
𝑦𝑖>𝑦𝑗
ℎ(𝑥) = {
1, 𝑥 > 0
0.5, 𝑥 = 0
0, 𝑥 < 0
(20)
In (19), sample i has a bigger label value than sample j.
R2
m index is the regression toward the mean, which is used to evaluate the external predictive
performance. The metric can be described as follows:
𝑟𝑚
2
= 𝑟2
∗ (1 − √𝑟2 − 𝑟0
2
) (21)
Where r2
and r2
0 are the squared correlation coefficients with and without intercept, respectively.
A value of r2
m above 0.5 is considered an ideal model.
To demonstrate the generalization ability of our model, we also applied the proposed model to a
binary classification problem, namely drug and target interaction prediction (DTI). Precision and
recall are the most frequently used metrics to evaluate binary classification tasks, therefore, in
this work, we continue to use them to evaluate our model. They can be formulated as follows:
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =
𝑇𝑃
𝑇𝑃 + 𝐹𝑃
(22)
𝑅𝑒𝑐𝑎𝑙𝑙 =
𝑇𝑃
𝑇𝑃 + 𝐹𝑃
(23)
Where TP, FP, and FN represent the sample numbers of true positive, false positive, and false
negative, respectively. In addition to the above two evaluation metrics, other commonly used
strategies like the area under the receiver operating characteristic curve (AUROC) and the area
under the precision recall curve (AUPRC) are also used by us to measure the performance of the
model.
3.2. Results
3.2.1. Comparison with other Regression Models
The DSAGLSTM-DTA model combines GNN and BiLSTM, and we conducted experiments on
two different datasets, Davis and KIBA. We compared the model with traditional learning
methods and some current deep learning methods, where for deep learning methods, we divided
it into two types according to the dimensional representation of the drug. In above models, drugs
and proteins are represented by different descriptors, and these descriptors are subsequently
extracted by various feature extraction approaches. In this experiment, we divided the dataset into
five equal parts, four of which were used as training set and one was used as test set, in order to

14
prevent data overfitting, we also divided the validation set from the training set and used cross-
validation to train the data set.
The experimental results in Table 2 demonstrate that compared with other DTA methods,
DSAGLSTM-DTA has a huge improvement in performance on Davis dataset. For each DTA
model, we used its optimal data for comparison. In the analysis based on below table, the two
different architecture approaches of the proposed model outperform the baseline model to
varying degrees in three indicators, where the optimal data of our model is 0.01 lower than the
optimal data of the baseline model in MSE, an increase of about 4.6%. In addition, CI and r2
m
increased by 0.013 and 0.041 respectively. Although distributed architecture of DSAGLSTM-
DTA is slightly inferior to centralized architecture in comprehensive performance, it still far
exceeds other baseline models. It is also worth noting that the CI values of two methods of the
proposed model all exceed 0.9, which proves that they have strong consistency, moreover, the r2
m
values are all over 0.7, indicating that they have strong external prediction performance and is an
acceptable model.
Table 2. Results of various DTA prediction models on the Davis dataset
Method Protein Compound CI MSE r2
m
Traditional methods
KronRLS[15] S-W Pubchem 0.871 0.379 0.407
SimBoost[16] S-W Pubchem 0.872 0.282 0.644
1D Representation-based Approaches
WideDTA[18] CNN CNN 0.886 0.262 —
DeepDTA[17] CNN(PS+PDM) CNN(LS+LMCS) 0.878 0.261 0.630
AttentionDTA[21] CNN CNN 0.893 0.216 0.677
DeepGS[20] GAT+Smi2Vec CNN(Prot2Vec) 0.882 0.252 0.686
GraphDTA[19] CNN GNN 0.893 0.229 —
DSAGLSTM-
DTA(Distributed)
BiLSTM GNN 0.904 0.208 0.711
DSAGLSTM-
DTA(Centralized)
BiLSTM GNN 0.906 0.206 0.718
Table 2. Italics represent the best data of the baseline model, and bolds represent the data that is
better than the baseline model. The models in the above table are arranged in descending order of
MSE (The following table is the same).
The proposed model was also evaluated on the KIBA dataset, which has more data and these data
sources are more extensive, so the effects on the KIBA dataset are more convincing. The
performance of the model on the KIBA dataset is presented in Table 3, and its experimental
parameter settings are consistent with the Davis dataset. Similarly, the centralized architecture
achieves the best performance in the evaluation of all models with an MSE of 0.131, a CI of
0.898 and a r2
m of 0.805. Although the improvement effect of the distributed architecture is not as
obvious as that of the centralized architecture, it still has a considerable improvement. In
addition, it should be emphasized that the r2
m of the centralized architecture exceeds 0.8 for the
first time, which is a huge improvement over previous models.

15
Table 3. Results of various DTA prediction models on the KIBA dataset
Method Protein Compound CI MSE r2
m
Traditional methods
KronRLS[15] S-W Pubchem 0.782 0.411 0.342
SimBoost[16] S-W Pubchem 0.836 0.222 0.629
DeepDTA[17] CNN(PS+PDM) CNN(LS+LMCS) 0.863 0.194 0.673
WideDTA[18] CNN CNN 0.875 0.179 —
AttentionDTA[21] CNN CNN 0.882 0.155 0.755
DeepGS[20] GAT+Smi2Vec CNN(Prot2Vec) 0.860 0.193 0.684
GraphDTA[19] CNN GNN 0.891 0.139 —
DSAGLSTM-
DTA(Distributed)
BiLSTM GNN 0.893 0.134 0.786
DSAGLSTM-
DTA(Centralized)
BiLSTM GNN 0.898 0.131 0.805
The baseline data in table 2 and table 3 above is obtained from [15-21]. From table 2 and table 3,
it is not difficult to analyze that the 2D representation of drug is more advantageous than its 1D
representation. In addition, the introduction of the self-attention block also greatly improves the
comprehensive performance of the model. Finally, the approach of extracting features for drugs
and proteins also has a big impact on the final result. In summary, the results of proposed model
illustrate that our model has better performance than some other DTA models, which is of great
significance to the research of DTA and will greatly promote the development of DTA.
3.2.2. Evaluation of Performance on Binary Classification Tasks
In order to demonstrate that our model has good generalization ability, we applied our model to
two binary datasets, namely human and bindingDB for drug and target interaction experiments.
The difference between the binary classification task and the regression task lies in that the final
result is a label value (i.e., 0, 1) instead of a continuous value. Since our model is based on a
regression task, we used the activation function sigmoid to map the predicted values to the
interval (0, 1) when processing the output of the model, after that, the round function were
utilized by us to convert the mapped values to interaction values. The binary classification model
and its data are shown in Table 4 and Table 5.

16
Table 4. Performances of various DTI prediction approaches on the Human dataset.
Models AUROC AUPRC Precision Recall
K-NN[36] 0.860 0.927 0.798
RF[36] 0.940 0.897 0.861
L2[36] 0.911 0.913 0.867
SVM[36] 0.910 0.966 0.969
GraphDTA[19] 0.960±0.005 0.882±0.040 0.912±0.040
GCN[37] 0.956±0.004 0.862±0.006 0.912±0.010
CPI-GNN[38] 0.970 0.918 0.923
DrugVQA[39] 0.964±0.005 0.897±0.004 0.948±0.003
TransformerCPI[
36]
0.973±0.002 0.916±0.006 0.925±0.006
DSAGLSTM-
DTA(Distributed)
0.982±0.002 0.983±0.002 0.917±0.017 0.953±0.015
DSAGLSTM-
DTA(Centralized)
0.983±0.003 0.984±0.002 0.917±0.014 0.950±0.007
In order to fully validate the performance of the model and prevent overfitting, we used five-fold
cross-validation on the human dataset. In this work, we compared the model with some classic
machine learning algorithms, including k-nearest neighbors (k-NN), random forest (RF), etc. In
addition, some deep learning methods based on molecular graphs, such as CPI-GNN and
DrugVQA, were also functioned as evaluation references. From the performance of the above
models, it can be concluded that the two architectures of the proposed model far exceed the
performance of other models on AUROC and AUPRC. Even though they are slightly lower than
support vector machines (SVMs) on Precision and Recall, they still far superior to other models.
In other words, after accounting for all the metrics, our model still performed best.
In addition to evaluating on the human dataset, another widely used binary classification dataset,
BindingDB, was also considered for our measurement. The BindingDB dataset is characterized in
that the dataset has been preprocessed in advance. Therefore, we directly conducted experiments
on the preprocessed dataset. The evaluation results on the BindingDB dataset are summarized in
Table 5. We selected some recent 2D molecular graph-based models for comparison, and the
results illustrate that our two architectures outperform the baseline model on all metrics, which
demonstrates the excellent adaptability of our model.
Table 5. Performances of various DTI prediction approaches on the BindingDB dataset.
Models AUROC AUPRC Precision Recall
CPI-GNN[38] 0.603 0.543
GCN[37] 0.927 0.913
GraphDTA[19] 0.929 0.917
TransformerCPI
[36]
0.951 0.949
DSAGLSTM-
DTA(Distributed)
0.952 0.950 0.913 0.844
DSAGLSTM-
DTA(Centralized)
0.959 0.960 0.898 0.899

17
In conclusion, the superior performance of the proposed model on the two types of tasks exhibits
the excellent generalization ability of the model. In addition, by comparing the performance of
centralized and distributed architecture of proposed model on two types of tasks, it is not difficult
to conclude that the centralized architecture is more competitive than the distributed architecture,
which also verifies the point of Lee et al. Finally, the comparison with the data of GraphDTA
also fully confirms the role of feature extraction and self-attention mechanism in improving the
performance of the model, which thus demonstrates the effectiveness of our innovation.
4. CONCLUSION
In this work, we proposed a novel DTA prediction method named DSAGLSTM-DTA, which
introduced self-attention mechanisms in different ways in the feature extraction process of drug
molecular graphs, and formed two architectures called centralized and distributed. To capture the
context dependencies in long amino acid sequences of proteins, a bidirectional LSTM network
was applied. In addition, in order not to ignore the hidden features in the LSTM unit, an attention
block was used as the weight calculation. We applied the model to regression tasks and binary
classification tasks for drug and target prediction, respectively. Evaluation of the model on
benchmark datasets demonstrated that the proposed model achieves superior performance to that
of various existing DTA and DTI prediction methods, suggesting the effectiveness of the
proposed approach in predicting the interaction of drug and protein pairs. Furthermore, it also
demonstrated the good generalization ability of the extraction method as well as the effectiveness
of the self-attention mechanisms.
REFERENCES
[1] A. Ezzat, M. Wu, X. L. Li, and C. K. Kwoh, “Computational prediction of drug-target interactions
using chemogenomic approaches: An empirical survey,” Brief. Bioinform., vol. 20, no. 4, pp. 1337–
1357, 2018, doi: 10.1093/bib/bby002.
[2] X. Chen et al., “Drug-target interaction prediction: Databases, web servers and computational
models,” Brief. Bioinform., vol. 17, no. 4, pp. 696–712, 2016, doi: 10.1093/bib/bbv066.
[3] A. S. Rifaioglu, H. Atas, M. J. Martin, R. Cetin-Atalay, V. Atalay, and T. Doǧan, “Recent
applications of deep learning and machine intelligence on in silico drug discovery: Methods, tools
and databases,” Brief. Bioinform., vol. 20, no. 5, pp. 1878–1912, 2019, doi: 10.1093/bib/bby061.
[4] E. H. B. Maia, L. C. Assis, T. A. de Oliveira, A. M. da Silva, and A. G. Taranto, “Structure-Based
Virtual Screening: From Classical to Artificial Intelligence,” Front. Chem., vol. 8, no. April, 2020,
doi: 10.3389/fchem.2020.00343.
[5] M. Himmat, N. Salim, M. M. Al-Dabbagh, F. Saeed, and A. Ahmed, “Adapting document similarity
measures for ligand-based virtual screening,” Molecules, vol. 21, no. 4, pp. 1–13, 2016, doi:
10.3390/molecules21040476.
[6] Y. Lecun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, pp. 436–444, 2015,
doi: 10.1038/nature14539.
[7] A. Hazra, P. Choudhary, and M. Sheetal Singh, Recent advances in deep learning techniques and its
applications: An overview. Springer Singapore, 2021.
[8] M. Wen et al., “Deep-Learning-Based Drug-Target Interaction Prediction,” J. Proteome Res., vol. 16,
no. 4, pp. 1401–1409, 2017, doi: 10.1021/acs.jproteome.6b00618.
[9] K. Tian, M. Shao, S. Zhou, and J. Guan, “Boosting compound-protein interaction prediction by deep
learning,” Proc. - 2015 IEEE Int. Conf. Bioinforma. Biomed. BIBM 2015, pp. 29–34, 2015, doi:
10.1109/BIBM.2015.7359651.
[10] Y. Du, J. Wang, X. Wang, J. Chen, and H. Chang, “Predicting drug-target interaction via wide and
deep learning,” ACM Int. Conf. Proceeding Ser., pp. 128–132, 2018, doi: 10.1145/3194480.3194491.
[11] B. Shin, S. Park, K. Kang, and J. C. Ho, “Self-Attention Based Molecule Representation for
Predicting Drug-Target Interaction,” pp. 1–18, 2019, [Online]. Available:
http://arxiv.org/abs/1908.06760.

18
[12] A. Vaswani et al., “Attention Is All You Need,” CoRR, vol. abs/1706.0, 2017, [Online]. Available:
http://arxiv.org/abs/1706.03762.
[13] Z. Quan, Y. Guo, X. Lin, Z. J. Wang, and X. Zeng, “GraphCPI: Graph Neural Representation
Learning for Compound-Protein Interaction,” Proc. - 2019 IEEE Int. Conf. Bioinforma. Biomed.
BIBM 2019, pp. 717–722, 2019, doi: 10.1109/BIBM47256.2019.8983267.
[14] W. Torng and R. B. Altman, “Graph Convolutional Neural Networks for Predicting Drug-Target
Interactions,” J. Chem. Inf. Model., 2019, doi: 10.1021/acs.jcim.9b00628.
[15] T. Pahikkala et al., “Toward more realistic drug-target interaction predictions,” Brief. Bioinform., vol.
16, no. 2, pp. 325–337, 2015, doi: 10.1093/bib/bbu010.
[16] T. He, M. Heidemeyer, F. Ban, A. Cherkasov, and M. Ester, “SimBoost: a read-across approach for
predicting drug-target binding affinities using gradient boosting machines,” J. Cheminform., vol. 9,
no. 1, pp. 1–14, 2017, doi: 10.1186/s13321-017-0209-z.
[17] H. Öztürk, A. Özgür, and E. Ozkirimli, “DeepDTA: Deep drug-target binding affinity prediction,”
Bioinformatics, vol. 34, no. 17, pp. i821–i829, 2018, doi: 10.1093/bioinformatics/bty593.
[18] H. Öztürk, E. Ozkirimli, and A. Özgür, “WideDTA: prediction of drug-target binding affinity,” 2019,
[Online]. Available: http://arxiv.org/abs/1902.04166.
[19] T. Nguyen, H. Le, and S. Venkatesh, “GraphDTA: prediction of drug–target binding affinity using
graph convolutional networks,” BioRxiv, p. 684662, 2019, doi: 10.1101/684662.
[20] X. Lin, K. Zhao, T. Xiao, Z. Quan, Z. J. Wang, and P. S. Yu, “Deepgs: Deep representation learning
of graphs and sequences for drug-target binding affinity prediction,” Front. Artif. Intell. Appl., vol.
325, no. i, pp. 1301–1308, 2020, doi: 10.3233/FAIA200232.
[21] Q. Zhao, F. Xiao, M. Yang, Y. Li, and J. Wang, “AttentionDTA: Prediction of drug-target binding
affinity using attention model,” in Proceedings - 2019 IEEE International Conference on
Bioinformatics and Biomedicine, BIBM 2019, Nov. 2019, pp. 64–69, doi:
10.1109/BIBM47256.2019.8983125.
[22] Q. Zhao, H. Zhao, K. Zheng, and J. Wang, “HyperAttentionDTI: improving drug–protein interaction
prediction by sequence-based deep learning with attention mechanism,” Bioinformatics, vol. 38, no.
3, pp. 655–662, 2022, doi: 10.1093/bioinformatics/btab715.
[23] S. Zhang, M. Jiang, S. Wang, X. Wang, Z. Wei, and Z. Li, “Sag-dta: Prediction of drug–target
affinity using self-attention graph network,” Int. J. Mol. Sci., vol. 22, no. 16, 2021, doi:
10.3390/ijms22168993.
[24] J. Lee, I. Lee, and J. Kang, “Self-attention graph pooling,” 36th Int. Conf. Mach. Learn. ICML 2019,
vol. 2019-June, pp. 6661–6670, 2019.
[25] M. I. Davis et al., “Comprehensive analysis of kinase inhibitor selectivity,” Nat. Biotechnol., vol. 29,
no. 11, pp. 1046–1051, 2011, doi: 10.1038/nbt.1990.
[26] J. Tang et al., “Making sense of large-scale kinase inhibitor bioactivity data sets: A comparative and
integrative analysis,” J. Chem. Inf. Model., vol. 54, no. 3, pp. 735–743, 2014, doi:
10.1021/ci400709d.
[27] H. Liu, J. Sun, J. Guan, J. Zheng, and S. Zhou, “Improving compound-protein interaction prediction
by building up highly credible negative samples,” Bioinformatics, vol. 31, no. 12, pp. i221–i229,
2015, doi: 10.1093/bioinformatics/btv256.
[28] K. Yingkai Gao, A. Fokoue, H. Luo, A. Iyengar, S. Dey, and P. Zhang, “Interpretable drug target
prediction using deep neural representation,” IJCAI Int. Jt. Conf. Artif. Intell., vol. 2018-July, pp.
3371–3377, 2018, doi: 10.24963/ijcai.2018/468.
[29] D. S. Wishart et al., “DrugBank: A knowledgebase for drugs, drug actions and drug targets,” Nucleic
Acids Res., vol. 36, no. SUPPL. 1, pp. 901–906, 2008, doi: 10.1093/nar/gkm958.
[30] S. Günther et al., “SuperTarget and Matador: Resources for exploring drug-target relationships,”
Nucleic Acids Res., vol. 36, no. SUPPL. 1, pp. 919–922, 2008, doi: 10.1093/nar/gkm862.
[31] Y. Li, R. Zemel, M. Brockschmidt, and D. Tarlow, “Gated graph sequence neural networks,” 4th Int.
Conf. Learn. Represent. ICLR 2016 - Conf. Track Proc., no. 1, pp. 1–20, 2016.
[32] S. Hochreiter, “Long Short-Term Memory,” vol. 1780, pp. 1735–1780, 1997.
[33] H. V. Adikane, R. K. Singh, D. M. Thakar, and S. N. Nene, “Single-step purification and
immobilization of penicillin acylase using hydrophobic ligands,” Appl. Biochem. Biotechnol. - Part A
Enzym. Eng. Biotechnol., vol. 94, no. 2, pp. 127–134, 2001, doi: 10.1385/ABAB:94:2:127.
[34] A. Paszke et al., “PyTorch: An imperative style, high-performance deep learning library,” Adv.
Neural Inf. Process. Syst., vol. 32, no. NeurIPS, 2019.

19
[35] M. Fey and J. E. Lenssen, “Fast Graph Representation Learning with PyTorch Geometric,” no. 1, pp.
1–9, 2019, [Online]. Available: http://arxiv.org/abs/1903.02428.
[36] L. Chen et al., “TransformerCPI: Improving compound-protein interaction prediction by sequence-
based deep learning with self-attention mechanism and label reversal experiments,” Bioinformatics,
vol. 36, no. 16, pp. 4406–4414, 2020, doi: 10.1093/bioinformatics/btaa524.
[37] T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” 5th
Int. Conf. Learn. Represent. ICLR 2017 - Conf. Track Proc., pp. 1–14, 2017.
[38] M. Tsubaki, K. Tomii, and J. Sese, “Compound-protein interaction prediction with end-to-end
learning of neural networks for graphs and sequences,” Bioinformatics, vol. 35, no. 2, pp. 309–318,
2019, doi: 10.1093/bioinformatics/bty535.
[39] S. Zheng, Y. Li, S. Chen, J. Xu, and Y. Yang, “Predicting drug–protein interaction using quasi-visual
question answering system,” Nat. Mach. Intell., vol. 2, no. 2, pp. 134–140, 2020, doi:
10.1038/s42256-020-0152-y.

DSAGLSTM-DTA: PREDICTION OF DRUG-TARGET AFFINITY USING DUAL SELF-ATTENTION AND LSTM

More Related Content

Similar to DSAGLSTM-DTA: PREDICTION OF DRUG-TARGET AFFINITY USING DUAL SELF-ATTENTION AND LSTM (20)

Recently uploaded (20)

DSAGLSTM-DTA: PREDICTION OF DRUG-TARGET AFFINITY USING DUAL SELF-ATTENTION AND LSTM