Processing math: 100%
Research article

DTiGNN: Learning drug-target embedding from a heterogeneous biological network based on a two-level attention-based graph neural network


  • Received: 30 December 2022 Revised: 05 February 2023 Accepted: 07 February 2023 Published: 21 March 2023
  • Motivation: In vitro experiment-based drug-target interaction (DTI) exploration demands more human, financial and data resources. In silico approaches have been recommended for predicting DTIs to reduce time and cost. During the drug development process, one can analyze the therapeutic effect of the drug for a particular disease by identifying how the drug binds to the target for treating that disease. Hence, DTI plays a major role in drug discovery. Many computational methods have been developed for DTI prediction. However, the existing methods have limitations in terms of capturing the interactions via multiple semantics between drug and target nodes in a heterogeneous biological network (HBN). Methods: In this paper, we propose a DTiGNN framework for identifying unknown drug-target pairs. The DTiGNN first calculates the similarity between the drug and target from multiple perspectives. Then, the features of drugs and targets from each perspective are learned separately by using a novel method termed an information entropy-based random walk. Next, all of the learned features from different perspectives are integrated into a single drug and target similarity network by using a multi-view convolutional neural network. Using the integrated similarity networks, drug interactions, drug-disease associations, protein interactions and protein-disease association, the HBN is constructed. Next, a novel embedding algorithm called a meta-graph guided graph neural network is used to learn the embedding of drugs and targets. Then, a convolutional neural network is employed to infer new DTIs after balancing the sample using oversampling techniques. Results: The DTiGNN is applied to various datasets, and the result shows better performance in terms of the area under receiver operating characteristic curve (AUC) and area under precision-recall curve (AUPR), with scores of 0.98 and 0.99, respectively. There are 23,739 newly predicted DTI pairs in total.

    Citation: Saranya Muniyappan, Arockia Xavier Annie Rayan, Geetha Thekkumpurath Varrieth. DTiGNN: Learning drug-target embedding from a heterogeneous biological network based on a two-level attention-based graph neural network[J]. Mathematical Biosciences and Engineering, 2023, 20(5): 9530-9571. doi: 10.3934/mbe.2023419

    Related Papers:

    [1] Zhizhou Ma, Yujun Qi, Weiqing Liu . A comparative study of pull-out performance of bolted joints in pultruded FRP with drilled holes or punched holes. Mathematical Biosciences and Engineering, 2019, 16(5): 4213-4228. doi: 10.3934/mbe.2019210
    [2] Peng Jing, Weichao Wang, Chengxi Jiang, Ye Zha, Baixu Ming . Determinants of switching behavior to wear helmets when riding e-bikes, a two-step SEM-ANFIS approach. Mathematical Biosciences and Engineering, 2023, 20(5): 9135-9158. doi: 10.3934/mbe.2023401
    [3] Linxuan Zhou, Jingwei Gao, Qiao Li, Cheng Hu . Simulation study on tractive performance of off-road tire based on discrete element method. Mathematical Biosciences and Engineering, 2020, 17(4): 3869-3893. doi: 10.3934/mbe.2020215
    [4] Yuriy Stoyan, Georgiy Yaskov, Tatiana Romanova, Igor Litvinchev, Sergey Yakovlev, José Manuel Velarde Cantú . Optimized packing multidimensional hyperspheres: a unified approach. Mathematical Biosciences and Engineering, 2020, 17(6): 6601-6630. doi: 10.3934/mbe.2020344
    [5] Guirong Jiang, Qishao Lu, Linping Peng . Impulsive Ecological Control Of A Stage-Structured Pest Management System. Mathematical Biosciences and Engineering, 2005, 2(2): 329-344. doi: 10.3934/mbe.2005.2.329
    [6] Jiaxin Nan, Wanbiao Ma . Stability and persistence analysis of a microorganism flocculation model with infinite delay. Mathematical Biosciences and Engineering, 2023, 20(6): 10815-10827. doi: 10.3934/mbe.2023480
    [7] Dongmei Zhang . Unveiling dynamics of urbanization, rural logistics, and carbon emissions: A study based on China's empirical data. Mathematical Biosciences and Engineering, 2024, 21(2): 2731-2752. doi: 10.3934/mbe.2024121
    [8] Hongying Shu, Wanxiao Xu, Zenghui Hao . Global dynamics of an immunosuppressive infection model with stage structure. Mathematical Biosciences and Engineering, 2020, 17(3): 2082-2102. doi: 10.3934/mbe.2020111
    [9] Hui Li, Mengyao Zhang, Chenbo Zeng . Circular Jaccard distance based multi-solution optimization for traveling salesman problems. Mathematical Biosciences and Engineering, 2022, 19(5): 4458-4480. doi: 10.3934/mbe.2022206
    [10] Guangxun Sun, Binxiang Dai . Stability and bifurcation of a delayed diffusive predator-prey system with food-limited and nonlinear harvesting. Mathematical Biosciences and Engineering, 2020, 17(4): 3520-3552. doi: 10.3934/mbe.2020199
  • Motivation: In vitro experiment-based drug-target interaction (DTI) exploration demands more human, financial and data resources. In silico approaches have been recommended for predicting DTIs to reduce time and cost. During the drug development process, one can analyze the therapeutic effect of the drug for a particular disease by identifying how the drug binds to the target for treating that disease. Hence, DTI plays a major role in drug discovery. Many computational methods have been developed for DTI prediction. However, the existing methods have limitations in terms of capturing the interactions via multiple semantics between drug and target nodes in a heterogeneous biological network (HBN). Methods: In this paper, we propose a DTiGNN framework for identifying unknown drug-target pairs. The DTiGNN first calculates the similarity between the drug and target from multiple perspectives. Then, the features of drugs and targets from each perspective are learned separately by using a novel method termed an information entropy-based random walk. Next, all of the learned features from different perspectives are integrated into a single drug and target similarity network by using a multi-view convolutional neural network. Using the integrated similarity networks, drug interactions, drug-disease associations, protein interactions and protein-disease association, the HBN is constructed. Next, a novel embedding algorithm called a meta-graph guided graph neural network is used to learn the embedding of drugs and targets. Then, a convolutional neural network is employed to infer new DTIs after balancing the sample using oversampling techniques. Results: The DTiGNN is applied to various datasets, and the result shows better performance in terms of the area under receiver operating characteristic curve (AUC) and area under precision-recall curve (AUPR), with scores of 0.98 and 0.99, respectively. There are 23,739 newly predicted DTI pairs in total.



    The spread of infectious diseases threatens people's health and social stability. To control the spread of the epidemic, governments usually take some measures, such as developing vaccines and isolating patients carrying the virus. Before the successful promotion of the vaccine, the government's non-pharmaceutical control measures, which reduce people's contact rate through lockdown and isolation, play a leading role in controlling the epidemic. However, it cannot relieve people's fear of the virus and may deteriorate the national or regional economy [1,2].

    Mathematical models are useful tools to analyze the formation of social and economic phenomena under various backgrounds. In recent years, the Boltzmann-type equation, which uses the collision theory to study the statistical distribution of rarefied gas molecules [3], is adopted to investigate the socio-economic phenomena of multi-agent systems, such as opinion formation and wealth distribution (see [4,5,6,7,8]). To discover the socio-economic impact of epidemics, researchers extend the classical Susceptible-Infectious-Recovered (SIR) epidemic model [9,10,11,12], in which agents are separated into three categories: susceptible (S), infected (I) and recovered (R), and integrate the SIR model with the socio-economic phenomena of multi-agent systems, such as people's social contact and wealth exchange behavior [13,14].

    The description of socio-economic features involves knowledge of mathematics, sociology, economics, psychology and other fields [15,16,17,18]. In the physical-economic dynamic models, the wealth distribution is determined by the continuous trading features of a large number of agents according to a wealth exchange mechanism, which is characterized by the saving (exchange) propensity of agents and uncertain elements. The uncertainty includes market risks and speculative risks [5]. Saving means that agents are risk-conscious and only use part of their wealth for trading. The saving (exchange) propensity reflects agents' trading decision-making. Decision-making problems are common in life, such as commodity selection, product evaluation, company management, etc. (see [19,20,21]). Persons' decision-making behaviors are affected by psychological preferences, health status, wealth and other factors (see [5,13,22,23,24,25,26]).

    Dimarco et al. [13] introduce a kinetic system consisting of three Boltzmann-type equations to describe how a wealth distribution might evolve under the course of an epidemic. In the kinetic system, the transport parts reflect the movement of individuals in different infection statuses and the Boltzmann collision operators depict the change in wealth distributions caused by the wealth exchange behaviors. The wealth exchange mechanism is depicted by the Cordier-Pareschi-Toscani (CPT) model

    w=(1PL)w+PJw+ηLJw, (1.1a)
    w=(1PJ)w+PLw+˜ηLJw. (1.1b)

    In (1.1), the pre-trade wealth (w,w) of two agents with infection status L,J{S,I,R} becomes (w,w) after the transaction, and the constant PJ[0,1] is exchange propensity for the class of agents with the infectious state J. The random variables ηLJ and ˜ηLJ stand for the uncertain impact of market risks. Dimarco et al. [13] find that epidemics affect wealth distribution.

    In [13], agents in the same infection state have the same exchange propensity (see (1.1)) without considering the psychological differences of people when trading. In addition, the impact of government policies and measures on wealth distribution is not considered. Based on the work in [13], the motivation in our work includes two points. Firstly, we explore wealth distribution when the agent's trading decision is influenced by psychology. Secondly, we discuss the impact of contact control measures and vaccination on wealth distribution.

    In a speculative market with one stock and two populations of chartists and fundamentalists, Maldarella and Pareschi [27] consider the influence of agents' behavioral and psychological factors on their investment. The psychological impact is represented by an appropriate value function corresponding to the prospect theory in Kahneman and Tversky [28]. The prospect theory is applied to dynamic models of social and biological phenomena, such as social contacts, alcohol consumption, the age of first marriage, the size of cities, tumor growth, etc. (see [14,29,30,31]). In these studies, the micro modification of the social trait x is expressed as a single interaction model

    x=xΦϵδ(xˉx)x+ηx, (1.2)

    where x denotes the value of the trait after the interaction, and η is a random variable with mean zero and bounded variance. For the variable s=x/ˉx, the general function Φϵδ(s) is given by

    Φϵδ(s)=μeϵ(sδ1)/δ1eϵ(sδ1)/δ+1. (1.3)

    The constant parameters ˉx, μ(0,1), δ[1,1] and ϵ>0 are related to the phenomenon in society.

    The parameter δ plays a key role in determining the properties of the value function (1.3) and the trait's macroscopic features under the updating rule (1.2). Provided that δ0, the value function (1.3) is concave, and the stationary solution of the Boltzmann-type kinetic model is a lognormal density describing situations of call center service time and human behaviors (see [30,32,33]). When δ0, (1.3) possesses properties of the prospect theory in [28,34]. For example, the value function (1.3) is concave above ˉx and convex below it. In the case of δ>0, the stationary distribution of the trait x is a generalized gamma distribution depicting the alcohol consumption in [29]. δ<0 leads to an Amoroso-type distribution, which depicts the formation of social elites caused by social climbing activities [35].

    In this paper, to study the influence of epidemics and the psychology of agents on wealth distribution, we will generalize the CPT model (1.1) in [13]. Assuming that agents' trading behavior is affected by their characteristics (such as health state and wealth level) and the prevalence of infectious diseases, we extend the constant exchange propensity of (1.1) as a function of the infected fraction and wealth. Firstly, as a response to the severity of the epidemic, we multiply the constant exchange propensity PJ by a function of the infected fraction at time t to describe a possible way for agents to reduce the exchange propensity PJ. Secondly, considering that the rich have better anti-risk abilities than the poor, unlike the assumption that agents in the same infection state have the same trading propensity in [13], in this paper, we assume that the rich tend to use more wealth for trading to obtain better life enjoyment or investment return, while the poor tend to save wealth to ensure future life needs. We adopt a function Φ() (see (2.4)), which is a generalized form of (1.3) with δ=1, as a component of the exchange propensity to represent the disturbance of these psychological factors when making trading decisions. We introduce an additional parameter on the denominator of (1.3) to avoid the extreme situation that the increased investment proportion is too large (small) when the wealth level is high (low).

    The stationary solution of the conservative socio-economic dynamic model, whose form is determined by the micro-interaction rules, plays a significant role in describing the large-time behavior of the trait under investigation. In this work, we utilize the Boltzmann system [13] to describe the evolutions of wealth distributions, and generalize the constant contact rate in [13] to a time-dependent function. Under the modified binary wealth exchange mechanism proposed in this paper, the Boltzmann system is transformed into the Fokker-Planck equations with quadratic drift and diffusion coefficients by using the quasi-invariant asymptotic limit method [7]. Specifically, the stationary wealth distribution of susceptible (recovered) agents fL(w), L{S,R} in this work satisfies an ordinary differential equation in the form

    σ2d2dw2[w2fL(w)]+ddw[(bw2+˜aLwc)fL(w)]=0, (1.4)

    where the non-negative coefficients σ, ˜aL, b, and c>0 rely on the parameters involved in the exchange rule and the epidemic environment. The solution of (1.4) is in the form of the product of an inverse gamma function and an exponential function. Since the microscopic interaction of the trait (alcohol consumption) in [29] is not a binary interaction, the stationary distribution in [29] satisfies Eq (1.4) with b0 and c=0. The stationary distribution in [13] satisfies Eq (1.4) with b=0 and c0. Compared with the work in [13], in our work, the new quadratic term bw2 in the drift coefficient of (1.4) is due to the nonlinear value function Φ() adopted in the exchange propensity.

    The contributions of this paper are summarized as follows. 1) We supplement the exchange propensity form in the wealth distribution literature. Compared with the constant exchange propensity (see [6,13,22,36,37]), the varying exchange propensity in our model reflects people's psychological features. Compared with the exchange propensity in [5], which is a monotone function of wealth, we take into account the asymmetry of the agent's trading psychology at different wealth levels. 2) We supplement the description of the steady-state wealth distribution function in the literature. Polk and Boghosian [38] point out that the expression form of the wealth distribution's tail is not universal. Based on the classical wealth exchange CPT model, if the dynamic model is involved in the feedback control, taxation and redistribution implemented by the government, or the heterogeneity of agents, the steady-state wealth distributions are inverse gamma function (see [5,6,13,22,39,40]). Compared with the inverse gamma wealth distribution obtained in [13], the presence of the exponential function part in the solution of (1.4) makes the wealth distribution curve decay faster as w, namely, the slimmer of the tail. This phenomenon corresponds to the decline in the number of super-rich. 3) We verify that the epidemic affects wealth distribution, contact control measures can curb the epidemic and aggravate wealth inequality. Ashraf [41] finds that the increase of COVID-19 confirmed cases has a negative impact on the stock market. Ashraf [2] discovers that social distancing measures are bad for the economy. Dosi et al. [42] point out that COVID-19 and contact controls may aggravate wealth inequality. 4) We illustrate that vaccination can curb the spread of epidemics and help to improve the economy. Hansen and Mano [43] find that vaccination can increase the mobility of the workplace, help to improve the economy and reduce inequality.

    The rest of the paper is constructed as follows. In Section 2, we set up the Boltzmann-type system describing the evolution of wealth distribution involving an epidemic environment, and introduce a wealth exchange mechanism containing the environment and agents' psychology. The macroscopic properties of the wealth exchange model, such as the proportion of each group population and their mean wealth, are discussed in Section 3. We obtain a Fokker-Planck system to describe the evolution of wealth distributions in Section 4, and acquire the steady-state wealth distribution under the background of the SIR model. In Section 5, we introduce the government's contact control measures and vaccination into the Boltzmann-type system of wealth distributions and discuss their impact on epidemics and wealth distributions. In addition, we numerically analyze the effect of agents' trading psychology on wealth distribution and verify the existence of bimodal stationary wealth distribution. Conclusions are given in Section 6.

    In this paper, under the framework of the SIR epidemic model [10,11], we investigate how wealth distribution might evolve under the course of an epidemic. In the SIR model [11], agents are separated into the susceptible (S), infected (I), and recovered (R) categories. The susceptible group refers to individuals who have no immunity and may be infected by the disease, the infected group includes those who are infectious, and the recovered group includes those who are cured or permanently immunized.

    As assumed in the wealth exchange dynamics in [5], we regard that agents are indistinguishable. In addition to the difference of infection status, each agent's state at any time t is characterized by wealth w0. We denote Ω={S,I,R}. Let fL(w,t) with LΩ represent the wealth distribution of agents in group L at time t. In the SIR epidemic model, the birth and mortality factors are not considered, which is a simplification of the actual epidemic dynamics. Then, the whole agent system's wealth distribution is f(w,t)=LΩfL(w,t), and satisfies the normalization condition R+f(w,t)dw=1.

    Following the idea of wealth exchange dynamics in [13], the wealth distribution fL(w,t), LΩ satisfies an integro-differential system of three Boltzmann-type equations

    fS(w,t)t=Π(w,t)fS(w,t)+QS(w,t), (2.1a)
    fI(w,t)t=Π(w,t)fS(w,t)rfI(w,t)+QI(w,t), (2.1b)
    fR(w,t)t=rfI(w,t)+QR(w,t), (2.1c)

    where

    QL(w,t)=JΩ1τLJQ(fL,fJ)(w,t),LΩ,

    is a combination of collision operators, and represents the interaction between heterogeneous agents [22,44]. The collision term Q(fL,fJ) describes the changes in wealth distribution resulting from the microscopic interactions between individuals from groups L and J. The relaxation parameters τLJ (L,JΩ) before the collision operators depend on the behavior patterns of agents and the share of transactions [45], reflecting the impact of the infectious disease on the frequency of wealth exchange.

    In (2.1), the time evolution of fL(w,t) (LΩ) is expressed as the additive coupling of collision process and transport terms. The transport terms Π(w,t)fS(w,t) and rfI(w,t) consider the shift of agents from one category to another, while the change in health status is due to contact with infected persons or medical treatment. In (2.1b) and (2.1c), constant r>0 is the recovery rate and 1/r represents the average infection period [11]. The spread of the epidemic is governed by the local incidence growth rate

    Π(w,t)=R+ρ(w,v,t)fI(v,t)dv,

    where ρ(w,v,t) is the contact rate of two agents possessing wealth w and v at time t.

    Let L(t)=R+fL(w,t)dw, LΩ represents the proportion of agents in group L. In this work, we take ρ(w,v,t)=ρ(t)>ρ0>0, where the constant ρ0 is a lower bound of ρ(t), then Π(w,t)=ρ(t)I(t).

    In the background of an epidemic, we consider the binary wealth exchange model

    w=w[PLH(I(t))+Φ(w¯W)]w+[PJH(I(t))+Φ(v¯W)]v+ηw, (2.2a)
    v=v[PJH(I(t))+Φ(v¯W)]v+[PLH(I(t))+Φ(w¯W)]w+ηv. (2.2b)

    In the wealth exchange rule (2.2), an agent with infection status LΩ and wealth w interacts with another agent with infection status JΩ and wealth v, and their post-trade wealth is w and v, respectively.

    According to (2.2), the change of each agent's wealth in a single interaction comes from two mechanisms, the deterministic exchange part and the random fluctuation caused by market risks. The certain exchange part of each agent's wealth depends on the effect of infectious diseases and the impact of the individual's psychology on the transaction. In terms of a mathematical expression, the exchange propensity of the agent with wealth w and infection state L is represented by PLH(I(t))+Φ(w¯W). Specifically, PL is the exchange propensity of agents with infection state L under normal circumstances, which is the same as that in the CPT model (1.1). H(I(t))1 is a non-increasing function of the infected fraction I(t) and is used to measure the impact of the epidemic on people's participation in economic activities. We take H(I(t))=11+I(t). The value function Φ(w¯W), which measures the psychological impact of an agent's wealth level on his trading decision, is an application of the prospect theory [28]. The constant ¯W represents a reference wealth level, such as average wealth. The uncertainty is represented by the random variable η with mean zero and variance σ. To ensure that the post-trade wealth is nonnegative, we suppose that ηPLH(I(t))+Φ(w¯W)1. In addition, to guarantee that the kinetic equations (2.1) is well posed [13], we assume that

    maxL{S,I,R}{[PLH(I(t))+Φ(w¯W)]2+[1PLH(I(t))Φ(w¯W)]2}+σ<1. (2.3)

    For the variable s0, we consider the value function

    Φ(s)=μeϵ(s1)1νeϵ(s1)+1,ν1, (2.4)

    where the constant μ[0,1) measures the influence of individual psychology on trading decision-making, and the constant 0<ϵ<1 denotes the intensity of the frequency of wealth interaction. For fixed μ, ϵ and s, function |Φ(s)| decreases with respect to parameter ν. (2.4) is a generalization of the value function Φϵ1(s) in [29],

    Φϵ1(s)=μeϵ(s1)1eϵ(s1)+1.

    Compared with Φϵ1(s), parameter ν>1 in (2.4) prevents the possibility that when the wealth is large, the proportion of increased investment is too high, and when the wealth is small, the conservative psychology leads to little investment. The images of function Φ(s) with μ=0.1 and different values of parameters ν and ϵ are presented in Figure 1.

    Figure 1.  (a): The images of function Φ(s) with different values of parameters ν and ϵ. (b): The value of function Φ(s) around the reference point s=1.

    The Eq (2.4) satisfies the properties of the value function in [28], which are described as follows:

    1) Φ(s) monotonically increases with s and has finite upper and lower bounds

    lims+Φ(s)=μνΦ(s)lims0Φ(s)=μν[11+ννeϵ+1]<0. (2.5)

    2) s=1 is the reference point, i.e., Φ(s=1)=0.

    3) In the case of ϵ<logν, we have [Φ(s)]<0 for every s0. Namely, Φ(s) is a concave function on R+. In the case of logν<ϵ<1, s=11ϵlogν(0,1) is the inflection point, and

    [Φ(s)]{>0,s<11ϵlogν,<0,s>11ϵlogν.

    4) Since ν1, we get

    Φ(1+ζ)<Φ(1ζ),Φ(1+ζ)<Φ(1ζ),ζ(0,1).

    Property 4) states that the value function (2.4) is asymmetric to the point s=1, which can also be observed in Figure 1(b). It corresponds to a phenomenon that the decision-making behavior of agents possessing wealth higher than the reference level ¯W is more robust, namely, the fluctuation of decision-making affected by psychology is more gentle. Moreover, this property also indicates that agents with wealth under the reference level ¯W tend to save more wealth.

    According to the boundedness (2.5), the parameters in the binary exchange rule (2.2) satisfy PL+μ/ν<1, ηPL+μ/ν1 and PL+μν(11+ννeϵ+1)>0 for every ϵ(0,1), LΩ. Since μν(1+ννeϵ+11) increases with ϵ, we have PLμe1ν+e, LΩ.

    In (2.1), the Boltzmann-type collision operator Q(fL,fL) (LΩ) denotes the binary interaction of agents with the same infection state L, and Q(fL,fJ) with LJ (L,JΩ) represents the binary interaction of agents with different infection states L and J. For any test function ψ(w)C0(R+), we have [13]

    R+Q(fL,fJ)(w,t)ψ(w)dw=R2+ψ(w)ψ(w)fL(w,t)fJ(v,t)dvdw, (3.1)

    where is the expectation operator with respect to the random variable η.

    We take the relaxation time τLJ=1 (L,JΩ) in (2.1). Then, the weak form of the system (2.1) is

    ddtR+fS(w,t)ψ(w)dw=ρ(t)I(t)R+fS(w,t)ψ(w)dw+˜QS, (3.2a)
    ddtR+fI(w,t)ψ(w)dw=ρ(t)I(t)R+fS(w,t)ψ(w)dwrR+fI(w,t)ψ(w)dw+˜QI, (3.2b)
    ddtR+fR(w,t)ψ(w)dw=rR+fI(w,t)ψ(w)dw+˜QR, (3.2c)

    where

    ˜QL=JΩR+Q(fL,fJ)(w,t)ψ(w)dw,LΩ.

    Substituting ψ(w)=1 into (3.2), we obtain that the fractions of the three agent groups satisfy

    dS(t)dt=ρ(t)S(t)I(t), (3.3a)
    dI(t)dt=ρ(t)S(t)I(t)rI(t), (3.3b)
    dR(t)dt=rI(t) (3.3c)

    and S(t)+I(t)+R(t)1, which is due to the normalization condition of f(w,t). If ρ(t)=ρ is a constant, (3.3) is the classical SIR model, which is mathematically and epidemiologically well-posed [11]. In the SIR model, the contact number γ, which refers to the average number of contacts of an infected person during the period of infection, is equal to the ratio of contact rate ρ(t) to recovery rate r (i.e. γ=ρ(t)/r) [11]. Let γ0=ρ/r. For the classical SIR model (3.3), Hethcote [10] gives the properties of its solution, which is repeated as the following Theorem. Denote L=limt+L(t), fL(w)=limt+fL(w,t) with LΩ.

    Theorem 3.1. ([10]) Let (S(t),I(t)) be a solution of the classical SIR model in

    {(S(t),I(t))|S(t)0,I(t)0,S(t)+I(t)1}.

    If γ0S01, then I(t) decreases to zero as t. If γ0S0>1, then I(t) increases to a maximum Imax=I0+S01/γ0[ln(γ0S0)]/γ0 and then decreases to 0 as t. The susceptible proportion S(t) is a decreasing function and S(0,1/γ0) is the unique root of the equation

    S0+I0S+1γ0lnSS0=0. (3.4)

    Theorem 3.1 implies that the infected fraction I(t) tends to zero in the steady-state (i.e., I=0). Combined with the non-negativity of the distribution function, the wealth distribution of the infected population in the steady-state is zero, i.e., fI(w)=0. Then, in the steady-state, the order moments of wealth distribution of the infected population are zero. When I=0, we have limt+Π(w,t)=0.

    Let mL(t)=R+wfL(w,t)dw (LΩ) represent the mean wealth of group L. According to (2.2), the whole agent system's wealth is conserved in the mean, i.e., w+v=w+v, then LΩmL(t):=m, t0.

    Let EL(t)=R+w2fL(w,t)dw (LΩ) represent the second moment of fL(w,t). Denote mL=limt+mL(t) and EL=limt+EL(t).

    Taking ψ(w)=w and substituting (2.2) into (3.1) arise

    R+Q(fL,fJ)(w,t)wdw=[PLH(I(t))]mL(t)J(t)+[PJH(I(t))]mJ(t)L(t)J(t)R+Φ(w¯W)wfL(w,t)dw+L(t)R+Φ(v¯W)vfJ(v,t)dv.

    Denote

    FL(t):=R+Φ(v¯W)vfL(v,t)dv.

    We get

    dmS(t)dt=ρ(t)I(t)mS(t)+MS(t), (3.5a)
    dmI(t)dt=ρ(t)I(t)mS(t)rmI(t)+MI(t), (3.5b)
    dmR(t)dt=rmI(t)+MR(t), (3.5c)

    where

    ML(t)=H(I(t))L(t)JΩPJmJ(t)PLH(I(t))mL(t)+L(t)JΩFJ(t)FL(t),LΩ.

    If I=0, we have mI=0. Denote FJ=limt+FJ(t). From (3.5a) and (3.5c), the mean wealth mS and mR satisfy

    S(PSmS+PRmR)PSmS+S(FS+FR)FS=0, (3.6a)
    R(PSmS+PRmR)PRmR+R(FS+FR)FR=0, (3.6b)

    where R=1S and mS+mR=m. Since Φ(v¯W) is a nonlinear function of variable v, the mean wealth mS and mR cannot be solved explicitly from (3.6), but their upper and lower bounds can be obtained. The value function (2.4) has an upper and lower bound (2.5), so we have FS[μeϵ1νeϵ+1mS,μνmS] and FR[μeϵ1νeϵ+1mR,μνmR]. Then we obtain

    mS[PR+μeϵ1νeϵ+1DRSSm,PR+μνDSRSm],mR[PS+μeϵ1νeϵ+1DSRRm,PS+μνDRSRm],

    where DSR=S(PR+μν)+R(PS+μeϵ1νeϵ+1) and DRS=S(PR++μeϵ1νeϵ+1)+R(PS+μν).

    The Boltzmann-like system (2.1) describes how wealth distribution evolves under the course of an epidemic, but its explicit solutions are difficult to obtain. Utilizing the quasi-invariant limit method [40], the Boltzmann-type equations are turned into the Fokker-Planck equations. We can use the stationary solution of the Fokker-Planck system to depict the large-time behavior of wealth distribution.

    Letting 0<ϵ1 and employing the scaled quantities

    PJϵPJ,ηϵη,JΩ. (4.1)

    The binary wealth exchange model (2.2) becomes

    ww=[ϵPLH(I(t))+Φ(w¯W)]w+[ϵPJH(I(t))+Φ(v¯W)]v+ϵηw, (4.2a)
    vv=[ϵPJH(I(t))+Φ(v¯W)]v+[ϵPLH(I(t))+Φ(w¯W)]w+ϵηv, (4.2b)

    where L,JΩ.

    The scaled micro wealth exchange model (4.2) indicates that a single interaction arises only small changes in wealth. To preserve the effects of all parameters in the exchange rule (2.2) when taking the limit ϵ0, the agents need to do much more interactions within a time period, that is, to increase the interaction frequency. Therefore, we take τLJϵτLJ, which also makes the evolution of the distributions independent of the scaling parameter ϵ. Then, the weak form of the scaled system (2.1) reads

    ddtR+fS(w,t)ψ(w)dw=R+Π(w,t)fS(w,t)ψ(w)dw+˜QϵS, (4.3a)
    ddtR+fI(w,t)ψ(w)dw=R+Π(w,t)fS(w,t)ψ(w)dwrR+fI(w,t)ψ(w)dw+˜QϵI, (4.3b)
    ddtR+fR(w,t)ψ(w)dw=rR+fI(w,t)ψ(w)dw+˜QϵR, (4.3c)

    where

    ˜QϵL=JΩ1ϵτLJR+Qϵ(fL,fJ)(w,t)ψ(w)dw,LΩ.

    Taking a Taylor expansion of ψ(w) around w, and substituting the expansion into the collision terms on the right of (4.3), we obtain

    1ϵτLJR+Qϵ(fL,fJ)(w,t)ψ(w)dw=1ϵτLJR2+ψ(w)(ww)+ψ(w)2(ww)2fL(w,t)fJ(v,t)dvdw+RLJ(ϵ). (4.4)

    The residual RLJ(ϵ) describes the higher-order terms of ϵ. In fact, for any test function ψ(w) belonging to the second-order H¨older space C2+α0(R+) with α(0,1), we have

    RLJ(ϵ)=1ϵτLJR2+ψ(˜w)ψ(w)2(ww)2fL(w,t)fJ(v,t)dvdw,=O(ϵ1+α),

    where ˜w=θw+(1θ)w for some θ(0,1).

    Using

    limϵ01ϵΦ(s)=limϵ01ϵμeϵ(s1)1νeϵ(s1)+1=μν+1(s1)μν+1, (4.5)

    the limit of (4.4) is expressed as

    limϵ01ϵτLJR+Qϵ(fL,fJ)(w,t)ψ(w)dw=1τLJR+{w[HLJ(w,t)fL(w,t)]+σ22w2[J(t)w2fL(w,t)]}ψ(w)dw, (4.6)

    where

    HLJ(w,t)=[PLH(I(t))+μν+1(w¯W1)]wJ(t)+[PJH(I(t))μν+1]mJ(t)+μ(ν+1)¯WEJ(t).

    Substituting (4.4) into (4.3), setting τLJ=1 (L,JΩ), taking ϵ0 and combining with (4.6), we obtain

    fS(w,t)t=Π(w,t)fS(w,t)w[BS(w,t)fS(w,t)]+122w2[σw2fS(w,t)], (4.7a)
    fI(w,t)t=Π(w,t)fS(w,t)rfI(w,t)w[BI(w,t)fI(w,t)]+122w2[σw2fI(w,t)], (4.7b)
    fR(w,t)t=rfI(w,t)w[BR(w,t)fR(w,t)]+122w2[σw2fR(w,t)], (4.7c)

    where

    BL(w,t)=[PLH(I(t))+μν+1(w¯W1)]w+H(I(t))JΩPJmJ(t)μν+1m+μ(ν+1)¯WJΩEJ(t),LΩ.

    The boundary conditions are

    BL(w,t)fL(w,t)|w=+w=0=0,w[w2fL(w,t)]|w=+w=0=0,t0,LΩ.

    The system (4.7) maintains the memory of the micro-dynamic (2.2) through relevant parameters μ, ν, ¯W, σ and PJ (JΩ).

    As a special case of explicitly finding the steady-state solution of (4.7), we suppose that the contact rate ρ(t)=ρ>0 is a constant. From Theorem 3.1, there are I=0, fI(w)=0 and S is the unique solution of (3.4). The mean wealth mS and mR are bounded and belong to [0,m].

    Proposition 4.1. When the steady-state infected fraction I=0, the moments of the stationary solutions of the Boltzmann-like system (4.3) satisfy the recursive relationship

    μ(ν+1)¯WMn+1,L=((n1)σ2+μν+1PL)Mn,L+(J{S,R}PJmJμmν+1+μ(ν+1)¯WJ{S,R}EJ)Mn1,L, (4.8)

    where

    Mn,L=R+wnfL(w)dw,L{S,R},n=1,2,3,...,

    denotes the n-order moment of the stationary wealth distribution of agents with infection state L.

    Proof. When I=0, we have fI(w)=0 and Mn,I=0.

    For the case of n=1. Substituting ψ(w)=w into (4.3), and using the scaled exchange rule (4.2), then the steady-state mean wealth mL and energy EL satisfy

    (μν+1PL)mLμ(ν+1)Lm+LJ{S,R}PJmJ=μ(ν+1)¯W(ELLJ{S,R}EJ), (4.9)

    which is consistent with (4.8) when n=1.

    When n2, substituting ψ(w)=wn into (4.3) and (4.4), letting ϵ0 and using (4.5), we have

    J{S,R}R2+{[PL+μν+1(w¯W1)]w+[PJ+μν+1(v¯W1)]v}wn1fL(w)fJ(v)dvdw+(n1)σ2J{S,R}R2+wnfL(w)fJ(v)dvdw=0. (4.10)

    Then the recursive relation (4.8) is deduced through a simple calculation of (4.10).

    In the scaling (4.1), condition (2.3) for ϵ0 becomes

    minL{S,I,R}[PL+limϵ01ϵΦ(s)]>σ2. (4.11)

    Using the inequality in (4.5), a sufficient condition for (4.11) is PLμν+1σ2>0 with L{S,I,R}. Employing the H¨older's inequality, we have

    (mL)2LEL,(EL)2mLM3,L,L{S,R}. (4.12)

    Let n=2 in (4.8). Using (4.9) and (4.12), EL has an upper bound

    ELPLμν+1PLσ2μν+1(mL)2L,L{S,R}. (4.13)

    In the case of μ0, according to the recursive formula (4.8) and the boundedness of the first two moments of wealth distributions (see (4.12) and (4.13)), we can obtain the boundedness of higher-order moments of wealth distributions in the stationary situation.

    From (4.7), the stationary distribution fL(w) (L{S,R}) satisfies the ordinary differential equation

    (bw2aLwcL)fL(w)+σ2ddw[w2fL(w)]=0,

    where

    aL=μν+1PL0,b=μ(ν+1)¯W0,cL=μ(ν+1)¯WELL+(PLμν+1)mLL>0.

    From (4.9), we have

    cL=μ(ν+1)¯WELL+(PLμν+1)mLL=J{S,I,R}PJmJμν+1m+μ(ν+1)¯WJ{S,I,R}EJ:=c,

    which is independent of the subscript L. Then the stationary wealth distribution of the susceptible (recovered) group is expressed as

    fL(w)=LCLw2(aLσ)σe2cσ1w2bσw,L{S,R}, (4.14)

    where CL is a regularization parameter such that R+fL(w)dw=L. Note that

    dfL(w)dw=2σ1w2fL(w)[bw2+(σaL)wc].

    Then, fL(w) is a concave function in (0,+) and reaches the maximum value at ˜wL=c/(σaL) in the case of μ=0. When μ0, fL(w) gets the maximum value at ˜wL=[(aLσ)+(aLσ)2+4bc]/(2b).

    The whole stationary wealth distribution is given by

    f(w)=SCSw2(aSσ)σe2cσ1w2bσw+RCRw2(aRσ)σe2cσ1w2bσw. (4.15)

    Proposition 4.2. The combination (4.15) of two unimodal distributions (4.14) is a unimodal or bimodal distribution.

    Proof. Let

    Υ(w)=L{S,R}fL(w)[bw2+(σaL)wc].

    Then

    df(w)dw=2σ1w2Υ(w).

    Using limw0+Υ(w)<0 and limw+Υ(w)>0, thus the function Υ(w) has at least one positive root, which corresponds to the maximum point of f(w). Taking the derivative of Υ(w) with respect to w, we get

    dΥ(w)dw=2σ1w2L{S,R}fL(w)ϕL(w),

    where

    ϕL(w)=b2w4+b(σ2aL)w3+[(σ2aL)(σaL)2bc]w22c(σaL)w+c2. (4.16)

    Since fS(w)0, fR(w)0 and the number of variations in the sign of the polynomial ϕL(w) is 2. Using the Descartes' rule of sign, the number of positive roots of dΥ(w)/dw is 0 or 2. When dΥ(w)/dw has no positive root, the function Υ(w) has only on positive root. When dΥ(w)/dw has two positive roots, the function Υ(w) has three positive roots, corresponding to two maximum points and one minimum point of f(w), respectively.

    The distribution (4.14) is the product of an inverse-gamma function and an exponential function, which is equivalent to

    fL(w)=~CLgL(w;aL,c,σ)h(w;b,σ), (4.17)

    where

    ~CL=LCLσ2bΓ(12aLσ)(2cσ)2aLσ1,gL(w;aL,c,σ)=1Γ(12aLσ)(2cσ)12aLσw2aLσ2e2cσ1w,h(w;b,σ)=2bσe2bσw

    and Γ() is the Gamma function. Note that gL(w;aL,c,σ) is the density of the inverse-gamma distribution IGL(12aLσ,2cσ), and h(w;b,σ) is the density of the exponential distribution E(2bσ).

    It is fascinating to note that the appearance of the exponential function part (i.e., h(w;b,σ)) is mainly due to the function Φ() in (2.2). When the psychological factors are not considered in the wealth exchange (i.e., μ=0), (4.14) becomes an inverse-gamma function, which is the same as the stationary wealth distribution in [13], showing polynomial decay when w. The decomposition form (4.17) suggests that, due to the exponential function h(w;b,σ), the tail of the wealth distribution (4.15) is slimmer than that of the inverse gamma distribution curve under appropriate parameters, and corresponds to a society with fewer super-rich people.

    Under the framework of the SIR epidemic dynamic, by coupling the binary wealth exchange rule (2.2) with the Boltzmann-type dynamic system (2.1), we obtain the stationary wealth distribution (4.15), which is impacted by the epidemic-related parameters and the value function (2.4). In this part, we introduce the government contact control measures and vaccination into the kinetic system, and analyze the impact of different measures on epidemic dynamics and wealth distribution. Finally, in a fixed epidemic background, we explore the influence of the value function (2.4) on the stationary wealth distribution, and verify the possible situation that the whole stationary wealth distribution exists double peaks.

    The Gini coefficient takes value in [0,1] and measures wealth inequality. A larger Gini coefficient indicates a more unequal wealth distribution. The Gini coefficient is equal to the ratio of the area between the equality line and the Lorentz curve to the area below the equality line. Define the Lorenz curve as [40]

    L(F(w))=w0vf(v)dv,F(w)=w0f(v)dv,

    then the Gini coefficient is

    Gini=1210L(F(w))dF(w).

    In the following simulations, if there is no special declaration, we always take ρ=0.3, I0=0.01, S0=0.99, PS=0.075, PR=0.15, m=1, σ=0.1, ¯W=1 and ν=3.

    Considering that the government takes non-drug containment measures to control the spread of infectious diseases. We take ρ(t)=ρκI(t)>0, where the constant κ[0,min{ρ/I(t)}) measures the strength of contact control. Then, (3.3) is rewritten as

    dS(t)dt=[ρκI(t)]S(t)I(t), (5.1a)
    dI(t)dt=[ρκI(t)]S(t)I(t)rI(t), (5.1b)
    dR(t)dt=rI(t). (5.1c)

    System (5.1) is an example of the implicit behavioral SIR model [46], and the contact rate ρ(t) decreases with the infected fraction I(t). When κ=0, (5.1) becomes the classical SIR model in [10]. When κ0, compared with the classical SIR model, a nonlinear term κS(t)I2(t) appears in (5.1). The contact number γ1=[ρκI(t)]/rρ/r varies with the infected fraction I(t). Because of 0<ρκI(t)ρ, S(t)0 is monotonically decreasing. Similar to Theorem 3.1, the infected fraction in the steady-state is zero, i.e., I=0, as verified in Figure 2(a). Then, the wealth distribution of the infected group is zero in the steady-state, and the whole wealth distribution is still in the form of (4.15), in which S and R are the stationary solutions of (5.1).

    Figure 2.  Epidemic dynamics and wealth distribution with different values of κ and r. (a): The solid line denotes susceptible proportion and the dotted line indicates infected proportion. (b): The wealth distributions in epidemic backgrounds of (a). (c): The Lorentz curves of the wealth distributions in (b).

    Figure 2 depicts the epidemic dynamics and the wealth distribution with different values of parameters κ and r. We take μ=0. The parameter value ρ=0.3 and r=0.04 are close to the epidemiological parameters estimated in [14], which are fitted from the COVID-19 data before the lockdown in three European countries (France, Italy and Spain). Figure 2(a) shows that with the increase of the recover rate r or the contact control strength κ, the peak value of the infected fraction becomes smaller, and more agents keep the uninfected state. In particular, when r=0.25, more than half of the susceptible people have the opportunity to remain uninfected during the epidemic.

    The Gini coefficients of the wealth distributions in Figure 2(b) are Gini(r=0.04,κ=0)=0.3127, Gini(r=0.15,κ=0)=0.3739, Gini(r=0.25,κ=0)=0.4257, Gini(r=0.15,κ=0.5)=0.3901 and Gini(r=0.15,κ=1)=0.3989. In the epidemic backgrounds in Figure 2(a), the wealth distribution curves and Lorentz curves in Figure 2(b) and their corresponding Gini coefficients show that the increase in r or κ leads the wealth distribution curve shifts to the left, and aggravates the wealth inequality. Combined with Figure 2(a), a reasonable explanation for this phenomenon is that when the contact number γ1 is small, the susceptible agents in the steady-state take a large proportion. Affected by the conservative exchange psychology of susceptible agents, that is, the exchange tendency of susceptible individuals is less than the tendency of recovered agents, the inequality of wealth distribution becomes worse.

    One possible way to deal with the negative effect in Section 5.1.1 is vaccination. Through vaccination, the susceptible agent transforms directly into the recovered person. Compared with (2.1), we introduce a term Ξ(w,t)fS(w,t) to represent the transfer of agents from the susceptible group to the recovered group due to vaccination. The Boltzmann-type system (2.1) is rewritten as

    fS(w,t)t=Π(w,t)fS(w,t)Ξ(w,t)fS(w,t)+QS(w,t), (5.2a)
    fI(w,t)t=Π(w,t)fS(w,t)rfI(w,t)+QI(w,t), (5.2b)
    fR(w,t)t=rfI(w,t)+Ξ(w,t)fS(w,t)+QR(w,t), (5.2c)

    in which we take the transfer rate Ξ(w,t)=R+θ(t)fI(v,t)dv. To explore the effect of vaccination on the wealth distribution, we take the contact rate ρ(t)=ρ as a constant and θ(t)=βΨ{tT0}. The constant β>0 represents the vaccination rate, T0 is the time of starting vaccination and Ψ{tT0} denotes the indicator function on {t|tT0}. Integrating the Eqs (5.2) over R+ yields

    dS(t)dt=[ρ+θ(t)]S(t)I(t), (5.3a)
    dI(t)dt=ρS(t)I(t)rI(t), (5.3b)
    dR(t)dt=[r+θ(t)S(t)]I(t). (5.3c)

    Since θ(t)0, the consideration of vaccination does not affect the steady-state infected fraction I=0, which is verified in Figure 3(a). Therefore, the stationary solution obtained from (5.2) has the same form as (4.15), while the susceptible and recovered fractions S and R are calculated from (5.3).

    Figure 3.  The impact of vaccination on epidemic and wealth distribution. (a): The solid line denotes the susceptible proportion, the dashed line indicates the infected proportion and the dotted line represents the recovered proportion. (b): The wealth distributions in the epidemic backgrounds of (a). (c): The Lorentz curves of the wealth distributions in (b).

    The impact of vaccination on epidemic and wealth distribution is shown in Figure 3, in which we take T0=30, r=0.15 and μ=0. Figure 3 indicates that as the vaccination rate β increases, more agents become recovered. The total number of infected patients declines, and the distribution of wealth shifts to the right, reflecting an improvement in wealth inequality. The Gini coefficients of the wealth distributions in Figure 3(b) are Gini(β=0)=0.3739, Gini(β=0.3)=0.3551 and Gini(β=0.7)=0.3395. Therefore, according to the simulations in Figure 3, we infer that vaccination plays a role in preventing and controlling the epidemic and improving the economy. This change can not only effectively control the spread of the epidemic and reduce the infected fraction, but also have a positive impact on reducing wealth inequality.

    In the exchange mechanism (2.2), the exchange propensity of agents is affected by their psychology, which is measured by the value function (2.4). The stationary wealth distribution of susceptible (infected) populations is obtained in the form of (4.14), which is the product of an inverse-gamma function and an exponential function (see (4.17)). It is different from the inverse-gamma steady-state distribution obtained from the usual socio-economic dynamic models [6,13].

    In Figure 4, we compare the wealth distribution f(w) with different μ. Moreover, to verify the existence of bimodal distribution, we simulate the distribution f(w) with different values of the risk parameter σ. We take PR=0.2 and choose the epidemic background corresponding to a recovery rate r=0.25. The Gini coefficients of the wealth distributions in Figure 4(a) are Gini(μ=0)=0.4476, Gini(μ=0.001)=0.4444, Gini(μ=0.01)=0.4258, Gini(μ=0.05)=0.3856 and Gini(μ=0.08)=0.3654. The Gini coefficients of the wealth distributions in Figure 4(b) are Gini(μ=0)=0.2989, Gini(μ=0.001)=0.2982, Gini(μ=0.01)=0.2923, Gini(μ=0.05)=0.2712 and Gini(μ=0.08)=0.2591.

    Figure 4.  The impact of value function (2.4) on the wealth distribution.

    Figure 4 indicates that with the increase of psychological influence (i.e., the increase of μ), more agents are at the middle wealth level, the number of agents of both low and high wealth levels has declined. The Gini coefficient of different risk parameter values decreases with the increase of parameter μ. It also verifies the improvement of wealth inequality. Additionally, Figure 4(b) illustrates that the wealth distribution's tail becomes slimmer with the increase of μ, which corresponds to a society with fewer super-rich people. From the perspective of mathematical analysis, the larger the value of μ, the smaller the shape parameter 12aL/σ of the inverse-gamma function gL(w;aL,c,σ). This corresponds to a fatter tail. Therefore, the existence of exponential function h(w;b,σ) plays a key role in thinning the tail.

    Real economic behavior is complex and is affected by various factors, such as the market environment and the psychology of traders. Under the background of an epidemic, we utilizes the method of mixed rarefied gas dynamics to investigate the distribution of wealth. In the wealth exchange mechanism, the impact of agents' psychology on trading decisions is expressed by a value function, which satisfies the properties of the prospect theory [28]. Under the SIR epidemic, we obtain that the steady-state wealth distribution of the susceptible (recovered) group is the product of an inverse-gamma function and an exponential function, showing a single peak. The emergence of the exponential function part is caused by the nonlinear utility function (2.4). We prove and numerically verify that the wealth distribution of the whole agent system presents a bimodal distribution with appropriate parameters.

    An increase in the recovery rate or the strength of contact control helps to curb the epidemic, keeping more people uninfected and reducing the peak value of the infected fraction. However, these measures are imperfect and may lead to economic recession, because the conservative trading psychology of susceptible agents caused by the fear of the virus is not changed. Vaccination, which transfers the susceptible person directly into a recovered individual, is numerically verified to have a positive impact on curbing epidemics and reducing wealth inequality. In addition, numerical simulations state that, with the increase of the psychological effect of agents, the population in the middle wealth level is increased, the Gini coefficient is decreased, and the tail of the multi-agent system's wealth distribution becomes thinner. These results indicate a reduction in the super-rich and wealth inequality.

    This research is supported by the National Natural Science Foundation of China (No. 11471263). The authors are very grateful to the reviewers for their valuable and meaningful comments of the paper.

    The authors declare there is no conflict of interest.



    [1] J. A. DiMasi, R. W. Hansen, H. G. Grabowski, The price of innovation: new estimates of drug development costs, J Health Econ., 22 (2003), 151–185. https://doi.org/10.1016/s0167-6296(02)00126-1 doi: 10.1016/s0167-6296(02)00126-1
    [2] M. A. Yildirim, K. Goh, M. E. Cusick, A. Barabas, M. Vidal, Drug-target network, Nat. Biotechnol., 25 (2007), 1119–1126. https://doi.org/10.1038/nbt1338 doi: 10.1038/nbt1338
    [3] T. T. Ashburn, K. B. Thor, Drug repositioning: identifying and developing new uses for existing drugs, Nat. Rev. Drug Discov., 3 (2004), 673–683. https://doi.org/10.1038/nrd1468 doi: 10.1038/nrd1468
    [4] F. Cheng, C. Liu, J. Jiang, W. Lu, W. Li, G. Liu, et al., Prediction of drug-target interactions and drug repositioning via network-based inference, PLoS Comput. Biol., 8 (2012), 1002503. https://doi.org/10.1371/journal.pcbi.1002503 doi: 10.1371/journal.pcbi.1002503
    [5] Y. Yamanishi, M. Kotera, M. Kanehisa, S. Goto, Drug-target interaction prediction from chemical, genomic and pharmacological data in an integrated framework, Bioinformatics, 26 (2010), 246–254. https://doi.org/10.1093/bioinformatics/btq176 doi: 10.1093/bioinformatics/btq176
    [6] A. C. Cheng, R. G. Coleman, K. T. Smyth, Q. Cao, P. Soulard, D. R. Caffrey, et al., Structure-based maximal affinity model predicts small-molecule druggability, Nat. Biotechnol., 25 (2007), 71–75. https://doi.org/10.1038/nbt1273 doi: 10.1038/nbt1273
    [7] H. Alonso, A. A. Bliznyuk, J. E. Gready, Combining docking and molecular dynamic simulations in drug design, Med. Res. Rev., 26 (2006), 531–568. https://doi.org/10.1002/med.20067 doi: 10.1002/med.20067
    [8] K. Wang, J. Sun, S. Zhou, C. Wan, S. Qin, C. Li, et al., Prediction of drug–target interactions for drug repositioning only based on genomic expression similarity, PLoS Comput. Biol., 9 (2013), e1003315. https://doi.org/10.1371/journal.pcbi.1003315 doi: 10.1371/journal.pcbi.1003315
    [9] A.Tropsha, Best practices for QSAR model development, validation, and exploitation, Mol. Inform., 29(2010), 476–488. http://doi.org/10.1002/minf.201000061 doi: 10.1002/minf.201000061
    [10] D. Rognan, Chemogenomic approaches to rational drug design, Br. J. Pharmacol., 152 (2007), 38–52. https://doi.org/10.1038/sj.bjp.0707307 doi: 10.1038/sj.bjp.0707307
    [11] A. C. A. Nascimento, R. B. C. Prudêncio, I. G. Cost, A multiple kernel learning algorithm for drug-target interaction prediction, BMC Bioinformatics, 17 (2016), 1–16, https://doi.org/10.1186/s12859-016-0890-3 doi: 10.1186/s12859-016-0890-3
    [12] T. He, M. Heidemeyer, F. Ban, A. Cherkasov, M. Ester, SimBoost: a read-across approach for predicting drug–target binding affinities using gradient boosting machines, J. Cheminform., 9 (2017), 1–14. https://doi.org/10.1186/s13321-017-0209-z doi: 10.1186/s13321-017-0209-z
    [13] K. Y. Gao, A. Fokoue, H. Luo, A. Iyengar, S. Dey, P. Zhang, Interpretable drug target prediction using deep neural representation, in IJCAI International Joint Conferences on Artificial Intelligence Organization, (2018), 3371–3377. https://doi.org/10.24963/ijcai.2018/468
    [14] M. Tsubaki, K. Tomii, J. Sese, Compound–protein interaction prediction with end-to-end learning of neural networks for graphs and sequences, Bioinformatics 35 (2019), 309–318. https://doi.org/10.1093/bioinformatics/bty535 doi: 10.1093/bioinformatics/bty535
    [15] Y. Luo, X. Zhao, J. Zhou, J. Yang, Y. Zhang, W. Kuang, et al., A network integration approach for drug-target interaction prediction and computational drug repositioning from heterogeneous information, Nat. Commun., 8 (2017), 573. https://doi.org/10.1038/s41467-017-00680-8 doi: 10.1038/s41467-017-00680-8
    [16] R. S. Olayan, H. Ashoor, V. B. Bajic, DDR: efficient computational method to predict drug–target interactions using graph mining and machine learning approaches, Bioinformatics, 34 (2018), 1164–1173. https://doi.org/10.1093/bioinformatics/btx731 doi: 10.1093/bioinformatics/btx731
    [17] W. Zhang, W. Lin, D. Zhang, S. Wang, J. Shi, Y. Niu, Recent advances in the machine learning-based drug-target interaction prediction, Curr. Drug Metab., 20 (2019), 194–202. https://doi.org/10.2174/1389200219666180821094047 doi: 10.2174/1389200219666180821094047
    [18] M. Thafar, A. B. Raies, S. Albaradei, M. Essack, V. B. Bajic, Comparison study of computational prediction tools for drug–target binding affinities, Front. Chem., 7 (2019), 782. https://doi.org/10.3389/fchem.2019.00782 doi: 10.3389/fchem.2019.00782
    [19] S. Anusuya, M. Kesherwani, K. V. Priya, A. Vimala, G. Shanmugam, D. Velmurugan, et al., Drug–target interactions: prediction methods and applications, Curr. Protein Pept. Sci., 19 (2018), 537–561. https://doi.org/10.2174/1389203718666161108091609 doi: 10.2174/1389203718666161108091609
    [20] K. Sachdev, M. K. Gupta, A comprehensive review of feature based methods for drug target interaction prediction, J Biomed Inform., 93 (2019), 103159. https://doi.org/10.1016/j.jbi.2019.103159 doi: 10.1016/j.jbi.2019.103159
    [21] W. Chen, L. Kurgan, Survey of similarity-based prediction of drug-protein interactions, Curr. Med. Chem., 27 (2020), 5856–5886. https://doi.org/10.2174/0929867326666190808154841 doi: 10.2174/0929867326666190808154841
    [22] K. Abbasi, P. Razzaghi, A. Poso, S. Ghanbari-Ara, A. Masoudi-Nejad, Deep learning in drug target interaction prediction, Curr. Med. Chem., 28 (2021), 2100–2113. https://doi.org/10.2174/0929867327666200907141016 doi: 10.2174/0929867327666200907141016
    [23] Q. Bai, J. Ma, S. Liu, T. Xu, A. J. Banegas-Luna, H. Pérez-Sánchez, et al., WADDAICA: a webserver for aiding protein drug design by artificial intelligence and classical algorithm, Comput. Struct. Biotechnol. J., 19 (2021), 3573–3579. https://doi.org/10.1016/j.csbj.2021.06.017 doi: 10.1016/j.csbj.2021.06.017
    [24] Q. Bai, S. Tan, T. Xu, H. Liu, J. Huang, X. Yao, MolAICal: a soft tool for 3D drug design of protein targets by artificial intelligence and classical algorithm, Brief. Bioinform., 22 (2021), 161. https://doi.org/10.1093/bib/bbaa161 doi: 10.1093/bib/bbaa161
    [25] Q. Bai, S. Liu, Y. Tian, T. Xu, A. J. Banegas‐Luna, H. Pérez‐Sánchez, et al., Application advances of deep learning methods for de novo drug design and molecular dynamics simulation, Wiley Interdiscip. Rev.: Comput. Mol. Sci., 12 (2022), 1581. https://doi.org/10.1002/wcms.1581 doi: 10.1002/wcms.1581
    [26] M. Koutrouli, E. Karatzas, D. Paez-Espino, G. A. Pavlopoulos, A guide to conquer the biological network era using graph theory, Front. Bioeng. Biotechnol., 8 (2020), 34. http://doi.org/10.3389/fbioe.2020.00034 doi: 10.3389/fbioe.2020.00034
    [27] Y. Tabei, M. Kotera, R. Sawada, Y. Yamanishi, Network-based characterization of drug-protein interaction signatures with a space-efficient approach, BMC Syst. Biol., 39 (2019). https://doi.org/10.1186/s12918-019-0691-1 doi: 10.1186/s12918-019-0691-1
    [28] W. Wang, H. Lv, Y. Zhao, D. Liu, Y. Wang, Y. Zhang, DLS: A link prediction method based on network local structure for predicting drug-protein interactions, Front. Bioeng. Biotechnol., 8 (2020). https://doi.org/10.3389/fbioe.2020.00330 doi: 10.3389/fbioe.2020.00330
    [29] Z. Wu, W. Li, G. Liu, Y. Tang, Network-based methods for prediction of drug-target interactions, Front. Pharmacol., 9 (2018), 1134. https://doi.org/10.3389/fphar.2018.01134 doi: 10.3389/fphar.2018.01134
    [30] K. Abbas, A. Abbasi, S. Dong, L. Niu, L. Yu, B. Chen, et al., Application of network link prediction in drug discovery, BMC Bioinformatics, 187 (2021). https://doi.org/10.1186/s12859-021-04082-y doi: 10.1186/s12859-021-04082-y
    [31] M. A. Thafar, R. S. Olayan, H. Ashoor, S. Albaradei, V. B. Bajic, X. Gao, et al., DTiGEMS+: drug–target interaction prediction using graph embedding, graph mining, and similarity-based techniques, J. Cheminform., 12 (2020), 1–17. https://doi.org/10.1186/s13321-020-00447-2 doi: 10.1186/s13321-020-00447-2
    [32] L. Wang, Z. You, X. Chen, S. Xia, F. Liu, X. Yan, et al., Computational methods for the prediction of drug– target interactions from drug fingerprints and protein sequences by stacked auto-encoder deep neural network, in International Symposium on Bioinformatics Research and Applications, 46–58. https://doi.org/10.1007/978-3-319-59575-7_5
    [33] B. Agyemang, W. P. Wu, M. Y. Kpiebaareh, Z. Lei, E. Nanor, L. Chen, Multi-view self-attention for interpretable drug–target interaction prediction, J. Biomed. Inform., 110 (2020), 103547. https://doi.org/10.1016/j.jbi.2020.103547 doi: 10.1016/j.jbi.2020.103547
    [34] P. Zhang, Z. Wei, C. Che, B. Jin, DeepMGT-DTI: Transformer network incorporating multilayer graph information for Drug-Target interaction prediction, Comput. Biol. Med., 142 (2022), 105214. https://doi.org/10.1016/j.compbiomed.2022.105214 doi: 10.1016/j.compbiomed.2022.105214
    [35] C. Chen, H. Shi, Z Jiang, A. Salhi, R. Chen, X. Cui, et al., DNN-DTIs: Improved drug-target interactions prediction using XGBoost feature selection and deep neural network, Comput. Biol. Med., 136 (2021), 104676. https://doi.org/10.1016/j.compbiomed.2021.104676 doi: 10.1016/j.compbiomed.2021.104676
    [36] X. Zeng, S. Zhu, W. Lu, Z. Liu, J. Huang, Y Zhou, et al., Target identification among known drugs by deep learning from heterogeneous networks, Chem. Sci., 11 (2020), 1775–1797. https://doi.org/10.1039/C9SC04336E doi: 10.1039/C9SC04336E
    [37] B. Wang, A. M. Mezlini, F. Demir, M. Fiume, Z. Tu, M. Brudno, et al., Similarity network fusion for aggregating data types on a genomic scale, Nat. Methods, 11 (2014), 333–337. https://doi.org/10.1038/nmeth.2810 doi: 10.1038/nmeth.2810
    [38] F. Xia, J. Liu, H. Nie, Y. Fu, L. Wan, X. Kong, Random walks: A review of algorithms and applications, IEEE Trans. Emerging Top. Comput. Intell., 4 (2019), 95–107. https://doi.org/10.1109/TETCI.2019.2952908 doi: 10.1109/TETCI.2019.2952908
    [39] A. Ghulam, X. Lei, M. Guo, C. Bian, Disease-pathway association prediction based on random walks with restart and PageRank, IEEE Access, 8 (2020), 72021–72038. https://doi.org/10.1109/ACCESS.2020.2987071 doi: 10.1109/ACCESS.2020.2987071
    [40] S. Li, L. Jiang, J. Tang, N. Gao, F. Guo, Kernel fusion method for detecting cancer subtypes via selecting relevant expression data, Front. Genet., 11 (2020), 979. https://doi.org/10.3389/fgene.2020.00979 doi: 10.3389/fgene.2020.00979
    [41] R. Qin, L. Duan, H. Zheng, J. Li-Ling, K. Song, Y. Zhang, An ontology-independent representation learning for similar disease detection based on multi-layer similarity network, IEEE/ACM Trans. Comput. Biol. Bioinform., 18 (2021), 183–193. https://doi.org/10.1109/tcbb.2019.2941475 doi: 10.1109/tcbb.2019.2941475
    [42] H. I. Lee, H. Nam, Identification of drug-target interaction by a random walk with restart method on an interactome network, BMC Bioinformatics, 208 (2018), 9–18. https://doi.org/10.1186/s12859-018-2199-x doi: 10.1186/s12859-018-2199-x
    [43] T. N. Jarada, J. G. Rokne, R. Alhajj, SNF-NN: computational method to predict drug-disease interactions using similarity network fusion and neural networks, BMC Bioinformatics, 28 (2021), 1–20. https://doi.org/10.1186/s12859-020-03950-3 doi: 10.1186/s12859-020-03950-3
    [44] D. V. Anand, Q. Xu, J. Wee, K. Xia, T. C. Sum, Topological feature engineering for machine learning based halide perovskite materials design, Npj Comput. Mater., 203 (2022), 1–8. https://doi.org/10.1038/s41524-022-00883-8 doi: 10.1038/s41524-022-00883-8
    [45] M. Wang, Z. Cang, G. W. Wei, A Topology-based network tree for the prediction of protein-protein binding affinity changes following mutation, Nat. Mach. Intell., 2 (2020), 116–123. https://doi.org/10.1038/s42256-020-0149-6 doi: 10.1038/s42256-020-0149-6
    [46] S. Fan, J. Zhu, X. Han, C. Shi, L. Hu, B. Ma, et al., Metapath-guided heterogeneous graph neural network for intent recommendation, in Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, (2019), 2478–2486. https://doi.org/10.1145/3292500.3330673
    [47] X. Fu, J. Zhang, Z. Meng, I. King, MAGNN: Metapath aggregated graph neural network for heterogeneous graph embedding, in Proceedings of The Web Conference, (2020), 2331–234. https://doi.org/10.1145/3366423.3380297
    [48] Z. Lu, Y. Wang, M. Zeng, M. Li, HNEDTI: Prediction of drug-target interaction based on heterogeneous network embedding, in 2019 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2019,211–214. https://doi.org/10.1109/BIBM47256.2019.8983181
    [49] M. Samizadeh, B. Minaei-Bidgoli, Drug-target interaction prediction by Metapath2vec node embedding in heterogeneous network of interactions, Int. J. Artif. Intell. Tools, 29 (2020), 2050001. https://doi.org/10.1142/S0218213020500013 doi: 10.1142/S0218213020500013
    [50] X. Liu, H. Feng, J. Wu, K. Xia, Persistent spectral hypergraph-based machine learning (PSH-ML) for protein-ligand binding affinity prediction, Brief. Bioinform., 22 (2021), 127. https://doi.org/10.1093/bib/bbab127 doi: 10.1093/bib/bbab127
    [51] X. Liu, X. Wang, J. Wu, K. Xia, Hypergraph-based persistent cohomology (HPC) for molecular representations in drug design, Brief. Bioinform., 22 (2021), 411. https://doi.org/10.1093/bib/bbaa411 doi: 10.1093/bib/bbaa411
    [52] G. A. Petsko, Guilt by association, Genome Biol., 104 (2009), 1–3. https://doi.org/10.1186/gb-2009-10-4-104 doi: 10.1186/gb-2009-10-4-104
    [53] H. Wang, H. Zheng, J. Wang, C. Wang, F. Wu, Integrating omics data with a multiplex network-based approach for the identification of cancer subtypes, IEEE Trans. Nanobioscience, 15 (2016), 335–342. https://doi.org/10.1109/tnb.2016.2556640 doi: 10.1109/tnb.2016.2556640
    [54] D. S. Wishart, Y. D. Feunang, A. C. Guo, E. J. Lo, A. Marcu, J. R. Grant, et al., DrugBank 5.0: a major update to the DrugBank database for 2018, Nucleic Acids Res., 46 (2018), 1074–1082. https://doi.org/10.1093/nar/gkx1037 doi: 10.1093/nar/gkx1037
    [55] D. Weininger, SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules, J. Chem. Inf. Comput. Sci, 28 (1988), 31–36. https://doi.org/10.1021/ci00057a005 doi: 10.1021/ci00057a005
    [56] C. Steinbeck, Y. Han, S. Kuhn, O. Horlacher, E. Luttmann, E. Willighagen, Recent developments of the chemistry development kit (CDK)-an opensource java library for chemo-and bioinformatics, Curr. Pharm. Des., 12 (2006), 2111–2120. https://doi.org/10.1021/ci025584y doi: 10.1021/ci025584y
    [57] A. Rácz, D. Bajusz, K. Héberger, Life beyond the Tanimoto coefficient: similarity measures for interaction fingerprints, J. Cheminform., 10 (2018), 1–12. https://doi.org/10.1186/s13321-018-0302-y doi: 10.1186/s13321-018-0302-y
    [58] M. Kuhn, M. Campillos, I. Letunic, L. J. Jensen, P. Bork, A side effect resource to capture phenotypic effects of drugs, Mol. Syst. Biol., 6 (2010), 343. https://doi.org/10.1038/msb.2009.98 doi: 10.1038/msb.2009.98
    [59] F. Zhu, TTD, 2009. Available from: https://bio.tools/ttd
    [60] UniProt consortium, Uniprot, 2023.Available from: http://www.uniprot.org/
    [61] T. F. Smith, M. S. Waterman, Identification of common molecular subsequences, J Mol. Biol., 147 (1981), 195–197. https://doi.org/10.1016/0022-2836(81)90087-5 doi: 10.1016/0022-2836(81)90087-5
    [62] A. P. Davis, C. J. Grondin, R. J. Johnson, D. Sciaky, B. L. King, R. McMorran, et al., The Comparative toxicogenomics database: update 2017, Nucleic Acids Res., 45 (2017), 972–978. https://doi.org/10.1093/nar/gkw838 doi: 10.1093/nar/gkw838
    [63] Gene Ontology, The Gene Ontology Resource, 2023. Available from: http://www.geneontology.org/
    [64] J. A. Blake, M. Dolan, H. Drabkin, D. P. Hill, L. Ni, D.Sitnikov, et al., Gene Ontology annotations and resources, Nucleic Acids Res., 41 (2013), 530–535. https://doi.org/10.1093/nar/gks1050 doi: 10.1093/nar/gks1050
    [65] G. Yu, F. Li, Y. Qin, X. Bo, Y. Wu, S. Wang, GOSemSim: an R package for measuring semantic similarity among GO terms and gene products, Bioinformatics, 26 (2010), 976–978. https://doi.org/10.1093/bioinformatics/btq064 doi: 10.1093/bioinformatics/btq064
    [66] S. S. Deepika, T. V. Geetha, A meta-learning framework using representation learning to predict drug-drug interaction, J. Biomed. Inform., 84 (2018), 136–147. https://doi.org/10.1016/j.jbi.2018.06.015 doi: 10.1016/j.jbi.2018.06.015
    [67] L. Huang, H. Luo, S. Li, F. Wu, J. Wang, Drug-drug similarity measure and its applications, Brief. Bioinform., 22 (2021), 265. https://doi.org/10.1093/bib/bbaa265 doi: 10.1093/bib/bbaa265
    [68] T. van Laarhoven, S. B. Nabuurs, E. Marchiori, Gaussian interaction profile kernels for predicting drug-target interaction, Bioinformatics, 27 (2011), 3036–3043. https://doi.org/10.1093/bioinformatics/btr500 doi: 10.1093/bioinformatics/btr500
    [69] H. Ye, K. Tang, L. Yang, Z. Cao, Y. Li, Study of drug function based on similarity of pathway fingerprint, Protein Cell, 3 (2012), 132–139. https://doi.org/10.1007/s13238-012-2011-z doi: 10.1007/s13238-012-2011-z
    [70] M. Kanehisa, M. Furumichi, M. Tanabe, Y. Sato, K. Morishima, KEGG: new perspectives on genomes, pathways, diseases and drugs, Nucleic Acids Res., 45 (2017), 353–361. https://doi.org/10.1093/nar/gkw1092 doi: 10.1093/nar/gkw1092
    [71] L. R. Sancho, BioCor: Functional similarities. Available from: https://llrs.github.io/BioCor/
    [72] A. Herbert, MaxCluster, 2023. Available from: http://www.sbg.bio.ic.ac.uk/~maxcluster/
    [73] National Science Foundation, US Department of Energy, National Cancer Institute, National Institute of Allergy and Infectious Diseases, National Institute of General Medical Sciences of the National Institutes of Health, RCSB Protein Data Bank (RCSB PDB), 2023. Available from: https://www.rcsb.org/
    [74] I. Kufareva, R. Abagyan, Methods of protein structure comparison, Homol. Model., 857 (2012), 231–257. https://doi.org/10.1007/978-1-61779-588-6_10 doi: 10.1007/978-1-61779-588-6_10
    [75] X. Wu, E. Pang, K. Lin, Z. M. Pei. Improving the measurement of semantic similarity between gene ontology terms and gene products: Insights from an edge- and IC-Based hybrid method, PLoS ONE, 8 (2013), 66745. https://doi.org/10.1371/journal.pone.0066745 doi: 10.1371/journal.pone.0066745
    [76] M. Shatnawi, Review of recent protein-protein interaction techniques, in Emerging Trends in Computational Biology, Bioinformatics and Systems Biology, (2015), 99–121. http://doi.org/10.1016/B978-0-12-802508-6.00006-5
    [77] C. Menichelli, O. Gascuel, L. Bréhélin, Improving pairwise comparison of protein sequences with domain co-occurrence, PLoS Comput. Biol., 14 (2018), 1005889. https://doi.org/10.1371/journal.pcbi.1005889 doi: 10.1371/journal.pcbi.1005889
    [78] R. D. Finn, A. Bateman, J. Clements, P. Coggil, R. Y. Eberhardt, S. R. Eddy, et al., Pfam: the protein families database, Nucleic Acids Res, 42 (2014), 222–230. https://doi.org/10.1093/nar/gkt1223 doi: 10.1093/nar/gkt1223
    [79] Y. M. Omar, P. Plapper, A survey of information entropy metrics for complex networks, Entropy, 22 (2020), 1417. https://doi.org/10.3390/e22121417 doi: 10.3390/e22121417
    [80] A. P. Riascos, J. L. Mateos, Random walks on weighted networks: Exploring local and non-local navigation strategies, 2019, preprint, arXiv: 1901.05609.
    [81] Chris McCormick, Word2vec tutorial-the skip-gram model, 2016. Available from: http://mccormickml.com/2016/04/19/word2vec-tutorial-the-skip-gram-model/
    [82] B. Wang, A. Wang, F. Chen, Y. Wang, C. C. J. Kuo, Evaluating word embedding models: methods and experimental results, APSIPA Trans. Signal Inf. Process., 8 (2019). https://doi.org/10.1017/ATSIP.2019.12 doi: 10.1017/ATSIP.2019.12
    [83] Y. Li, M. Yang, Z. Zhang, A survey of multi-view representation learning, IEEE Trans. Knowl. Data Eng., 31 (2018), 1863–1883. https://doi.org/10.1109/TKDE.2018.2872063 doi: 10.1109/TKDE.2018.2872063
    [84] K. Sun, L. Wang, B. Xu, W. Zhao, S. W. Teng, F. Xia, Network representation learning: From traditional feature learning to deep learning, IEEE Access, 8 (2020), 205600–205617. https://doi.org/10.1016/j.jfds.2017.05.001 doi: 10.1016/j.jfds.2017.05.001
    [85] A. Grover, J. Leskovec, node2vec: Scalable feature learning for networks, in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, (2016), 855–864. https://doi.org/10.1145/2939672.2939754
    [86] Y. Dong, N. V. Chawla, A. Swami, metapath2vec: Scalable representation learning for heterogeneous networks, in Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, (2017), 135–144. https://doi.org/10.1145/3097983.3098036
    [87] B. Perozzi, R. Al-Rfou, S. Skiena, Deepwalk: Online learning of social representations, in Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, (2014), 701–710. https://doi.org/10.1145/2623330.2623732
    [88] P. Velickovic, G. Cucurull, A. Casanova, A. Romero, P. Liò, Y. Bengio, Graph attention networks, in ICLR, (2018).
    [89] G. Douzas, F. Bacao, F. Last, Improving imbalanced learning through a heuristic oversampling method based on k-means and SMOTE, Inf. Sci., 465 (2018), 1–20. https://doi.org/10.1016/j.ins.2018.06.056 doi: 10.1016/j.ins.2018.06.056
    [90] A. Creswell, T. White, V. Dumoulin, K. Arulkumaran, B. Sengupta. A. A. Bharath, Generative adversarial networks: An overview, IEEE Signal Proc. Mag., 35 (2018), 53–65. https://doi.org/10.1109/MSP.2017.2765202 doi: 10.1109/MSP.2017.2765202
    [91] J. Peng, J. Li, X. Shang, A learning-based method for drug-target interaction prediction based on feature representation learning and deep neural network, BMC Bioinformatics, 21 (2020), 1–13. https://doi.org/10.1186/s12859-020-03677-1 doi: 10.1186/s12859-020-03677-1
    [92] Z. Zheng, N. Q. K. Le, M. C. H. Chua, MaskDNA-PGD: An innovative deep learning model for detecting DNA methylation by integrating mask sequences and adversarial PGD training as a data augmentation method, Chemom. Intell. Lab. Syst., 232 (2023), 104715. https://doi.org/10.1016/j.chemolab.2022.104715 doi: 10.1016/j.chemolab.2022.104715
    [93] J. N. Sua, S. Y. Lim, M. H. Yulius, X. Su, E. K. Y. Yapp, N. Q. K. Le, et al., Incorporating convolutional neural networks and sequence graph transform for identifying multilabel protein Lysine PTM sites, Chemom. Intell. Lab. Syst., 206 (2020), 104171. https://doi.org/10.1016/j.chemolab.2020.104171 doi: 10.1016/j.chemolab.2020.104171
    [94] Y. Yamanishi, M. Araki, A. Gutteridge, W. Honda, M. Kanehisa, Prediction of drug-target interaction networks from the integration of chemical and genomic spaces, Bioinformatics, 24 (2008), 232–240. https://doi.org/10.1093/bioinformatics/btn162 doi: 10.1093/bioinformatics/btn162
    [95] Q. Kuang, X. Xu, R. Li, Y. Dong, Y. Li, Z. Huang, et al., An eigenvalue transformation technique for predicting drug-target interaction, Sci. Rep., 5 (2015), 13867. https://doi.org/10.1038/srep13867 doi: 10.1038/srep13867
    [96] J. Davis, M. Goadrich, The relationship between precision-recall and ROC Curves, in Proceedings of the 23rd international conference on machine learning, (2006), 233–240. https://doi.org/10.1145/1143844.1143874
    [97] T. Nguyen, T. H. Nguyen, T. N. H. Pham, N. T. Huy, M. V. Bay, M. Q. Pham, et al., Autodockvina adopts more accurate binding poses but autodock4 forms better binding affinity, J. Chem. Inf. Model., 60 (2019), 204–211. https://doi.org/10.1021/acs.jcim.9b00778 doi: 10.1021/acs.jcim.9b00778
    [98] PunMed, National Library of MedicineNational Center for Biotechnology Information, 2023. Available from: https://pubmed.ncbi.nlm.nih.gov/
    [99] European Molecular Biology Laboratory, ChEMBL, 2023. Available from: https://www.ebi.ac.uk/chembl/
    [100] N. Rohani, C. Eslahchi, Drug-drug interaction predicting by neural network using integrated similarity, Sci. Rep., 9 (2019), 1–11. https://doi.org/10.1038/s41598-019-50121-3 doi: 10.1038/s41598-019-50121-3
    [101] R. Zhou, Z. Lu, H. Luo, J. Xiang, M. Zeng, M. Li, NEDD: a network embedding based method for predicting drug-disease associations, BMC Bioinformatics, 387 (2020), 1–12. https://doi.org/10.1186/s12859-020-03682-4 doi: 10.1186/s12859-020-03682-4
  • This article has been cited by:

    1. JI-HUAN HE, QIAN YANG, CHUN-HUI HE, HAI-BIN LI, EERDUN BUHE, PULL-IN STABILITY OF A FRACTAL MEMS SYSTEM AND ITS PULL-IN PLATEAU, 2022, 30, 0218-348X, 10.1142/S0218348X22501857
  • Reader Comments
  • © 2023 the Author(s), licensee AIMS Press. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0)
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Metrics

Article views(4030) PDF downloads(229) Cited by(4)

Figures and Tables

Figures(14)  /  Tables(10)

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return

Catalog