Category	Name	Core algorithms	Similarity measures	Databases	Numbers^a	Cross-validation	Case study	Source code	Published date
Feature generation-based	GCNCDA [42]	FastGCN and Forest PA	Disease semantic similarity, GIP kernel similarity	CircR2Disease, MeSH	739 circRNA–disease associations (661 circRNAs and 100 diseases)	5-CV	Breast cancer, glioma, colorectal cancer	https://github.com/look0012/GCNCDA/	2020.06
	GANCDA [43]	Generative adversarial network and logistic model tree	Disease semantic similarity, GIP kernel similarity	CircR2Disease, MeSH	739 circRNA-disease associations (661 circRNAs and 100 diseases)	5-CV	Gastric cancer, colorectal cancer, breast cancer	Unavailable	2020.08
	DWNCPCDA [44]	DeepWalk, network consistency projection	CircRNA and disease topological similarity	CircR2Disease	650 circRNA-disease associations (585 circRNAs and 88 diseases)	5-CV	Hepatocellular carcinoma, lung cancer	https://github.com/ghli16/DWNCPCDA	2021.02
	NSL2CD [45]	DeepWalk and adaptive subspace learning	Disease semantic similarity, circRNA function similarity, GIP kernel similarity	CircR2Disease, MeSH	649 circRNA–disease associations (589 circRNAs and 88 diseases)	5-CV	Acute myeloid leukemia, lung cancer, breast cancer, etc.	Unavailable	2021.05
	GATCDA [46]	Graph attention network	Disease symptom, network, entropy similarity, circRNA network entropy similarity,	CircR2Disease, CircAtlas 2.0, Circ2Disease, CircRNADisease, starBase v2.0, DisGeNET	768 circRNA–disease associations (624 circRNAs and 102 diseases)	5-CV	Bladder cancer, diabetes retinopathy, rheumatoid arthritis	Unavailable	2021.06
	AE-RF [47]	Deep autoencoder, random forest	Disease semantic similarity, circRNA function similarity, GIP kernel similarity	CircR2Disease, DO	650 circRNA–disease associations (585 circRNAs and 88 diseases)	5-CV, 10-CV	Breast cancer, colorectal cancer, lung cancer	https://github.com/Deepthi-K523/AE-RF	2020.11
	IMS-CDA [48]	Stacked autoencoder, rotation forest	Disease semantic similarity, disease and circRNA Jaccard similarity, GIP kernel similarity	CircR2Disease, MeSH	739 circRNA–disease associations (661 circRNAs and 100 diseases)	5-CV	Cardiovascular disease, glioma, intracranial aneurysms	https://github.com/look0012/IMS-CDA/	2021.11
	Yang’s method [49]	accelerated attributed network embedding, stacked autoencoder, XGBoost	Disease semantic similarity, circRNA expression profile similarity, GIP kernel similarity	CircR2Disease, MeSH, exoRBase	605 circRNA–disease Associations (545 circRNAs and 75 diseases)	5-CV, 10-CV	Esophageal squamous cell carcinoma, gastric cancer	Unavailable	2021.08
Type discrimination-based	DeepDCR [50]	24 meta-path-based, deep forest	Disease symptom similarity, miRNA functional similarity, circRNA co-expression similarity	Circ2Traits, miRTarBase, circBase, miRbase, GENCODE, TargetScan, miR2Disease, HMDD	17,961 circRNAs, 469 miRNAs, and 248 diseases	5-CV	Breast neoplasms, lung neoplasms, hepatocellular carcinoma, etc.	https:// github.com/xzenglab/DeepDCR	2020.10
Hybrid-based	MSFCNN [51]	Convolutional neural networks	Disease semantic, symptom, Lin, PSB, Resnik, SemFunSim similarity, circRNA sequence, regulatory, expression similarity, GIP kernel similarity,	CircR2Disease, CircBank, HMDD v3.0	325 circRNAs, 53 diseases and 3175 miRNAs	5-CV	Acute myeloid leukemia, breast cancer, colorectal cancer, etc.	Unavailable	2020.10
	Wang’s method [52]	Convolutional neural network, extreme learning machine	Disease semantic similarity, GIP kernel similarity	CircR2Disease, MeSH	739 circRNA–disease associations (661 circRNAs and 100 diseases)	5-CV	Glioma, heart disease, cervical cancer, etc.	https://github.com/look0012/circRNA-Disease-association.	2020.07
	CDASOR [53]	Convolutional neural networks, recurrent neural networks	NA	CircR2Disease, Circ2Disease, circAtlas, LncRNAdisease, MNDR, circBase, UMLS, OMIM, circBase, DO	Dataset1: 754 circRNA–disease associations (630 circRNAs and 87 diseases), Dataset2: 2723 circRNA-disease associations (1998 circRNAs and 150 diseases)	5-CV	Heart disease, colorectal cancer, hypoxia, etc.	https://github.com/BioinformaticsCSU/CDASOR	2021.05
	AE-DNN [54]	Autoencoder, deep neural networks	Disease semantic similarity, circRNA sequence similarity, GIP kernel similarity	CircR2Disease, circBase	445 circRNA–disease associations (389 circRNAs and 61 diseases)	5-CV, 10-CV	Glioma, gastric cancer, liver cancer	Unavailable	2020.10
	SGANRDA [55]	Generative adversarial network, extreme learning machine	Disease semantic similarity, circRNA sequence similarity, GIP kernel similarity	CircR2Disease, MeSH	739 circRNA–disease associations (661 circRNAs and 100 diseases)	5-CV, LOOCV	Cervical cancer, colorectal cancer, breast cancer, etc.	https://github.com/look0012/SGANRDA/	2021.11
	DMFCDA [56]	Deep matrix factorization, multi-layer neural networks	NA	CircR2Disease, LncRNADisease, circBase, deepBase, UMLS, OMIM, NCBI	Dataset1: 619 circRNA–disease associations (556 circRNAs and 80 diseases), Dataset2: 744 circRNA-disease associations (632 circRNAs and 89 diseases)	5-CV, LOOCV	Colorectal cancer, hepatocellular carcinoma, lung cancer, etc.	https://github.com/bioinfomaticsCSU/DMFCDA	2021.04
	DMFMSF [57]	Weighted K-nearest known neighbor，singular value decomposition, deep matrix factorization	Disease semantic, hamming profile similarity, GIP kernel similarity	CircR2Disease, LncRNADisease, MeSH	Dataset1: 619 circRNA–disease associations (556 circRNAs and 80 diseases), Dataset2: 744 circRNA–disease associations (632 circRNAs and 89 diseases)	5-CV, LOOCV	Hepatocellular carcinoma, breast cancer, acute myeloid leukemia	https://github.com/lisusu6/DMFMSF	2021.07
	DMFCNNCD [58]	Deep matrix factorization, convolution neural network	Disease semantic similarity, circRNA function similarity	CircR2Disease, DO,	650 circRNA–disease associations (585 circRNAs and 88 diseases)	5-CV	Lung cancer, colorectal cancer, hepatocellular carcinoma	Unavailable	2021.12
	Thosini’s method [59]	Graph convolution network	CircRNA sequence similarity, GIP kernel similarity	CircR2Disease, circBase	About 900 circRNA–disease associations (more than 700 circRNAs and more than 100 diseases)	5-CV	Stomach cancer	Unavailable	2021.08
	CRPGCN [60]	Random walk with restart, principal component analysis, graph convolutional network	Disease semantic similarity, circRNA sequence, gene similarity, GIP kernel similarity	CircR2Disease, MeSH	533 circRNAs and 89 diseases	2-CV, 5-CV, 10-CV	Breast cancer	https://github.com/KajiMaCN/CRPGCN/	2021.11
	RGCNCDA [61]	Random walk with restart, principal component analysis, relational graph convolutional networks	Disease semantic similarity, circRNA function similarity, GIP kernel similarity, miRNA expression similarity	CircR2Disease, HMDD v3.0, ENCORI, MeSH,	650 circRNA–disease associations (585 circRNAs and 88 diseases), 11,824 miRNA-disease associations, 2293 circRNA-miRNA associations (68 circRNAs and 581 miRNAs)	5-CV	Gastric cancer, coronary artery disease, bladder cancer, etc.	Unavailable	2022.02
	IGNSCDA [62]	Graph convolutional network, multi-layer perceptron	CircRNA GIP kernel similarity, circRNA expression profile similarity	CircR2Disease, exoRBase	612 circRNA–disease associations (533 circRNAs and 89 diseases)	5-CV	Bladder cancer	Unavailable	2021.09
	KGANCDA [63]	Graph attention network and Multiple layer perceptron	NA	circR2Cancer, lncRNASNP2, LncRNADisease, circad, circRNADisease	Dataset1: 514 circRNAs, 62 diseases, 564 miRNAs, 573 lncRNAs Dataset2: 330 circRNAs, 79 diseases, 245 miRNAs, 297 lncRNAs,	5-CV	Colorectal cancer	https://github.com/lanbiolab/KGANCDA	2022.04
	GMNN2CD [64]	Graph Markov neural network	Disease semantic similarity, GIP kernel similarity	CircR2Disease, Circ2Disease, circRNA–disease, circAtlas, CircFunBase, MeSH	Dataset1: 533 circRNA–disease associations (612 circRNAs and 89 diseases) Dataset2: 589 circRNA-disease associations (649 circRNAs and 88 diseases) Dataset3: 330 circRNA–disease associations (354 circRNAs and 48 diseases) Dataset4: 848 circRNA–disease associations (930 circRNAs and 110 diseases) Dataset5: 2597 circRNA–disease associations (2984 circRNAs and 67 diseases)	5-CV	Breast cancer, cervical cancer	https://github.com/nmt315320/ GMNN2CD.git	2022.02

GCNCDA: A new method for predicting circRNA-disease associations based on Graph Convolutional Network Algorithm

https://github.com/look0012/GCNCDA/ 不可用

Materials and methods

Method overview

pcbi.1007568.g006.jpg

根据疾病语义相似性网络和circRNA-disease邻接矩阵构建疾病语义相似性矩阵和疾病高斯相互作用相似性矩阵（Gaussian interaction profile ，GIP）。

然后根据circRNA相似性网络和circRNA-disease邻接矩阵构建circRNA GIP相似性矩阵。3.
疾病和circRAN相似性矩阵融合。
FastGCN 提取高层特征，生成特征
ForestPA 分类器进行分类

Benchmark dataset

使用circR2Disease数据库作为benchmark，它包含661 circRNAs，100 diseases，和739 circRNA-disease联系。为了平衡，这里选择739 个负样本作为全部的负样本。

Construction of CircRNA similarity model

circRNA c(i) 的向量 V(c(i)) 为100维，表示和100个疾病的关系，相关设为1 否则为0。

circRNA c(i) 和circRNA c(j) 的GIP核心相似性 GC(c(i),c(j))

GC(c(i),c(j)) = exp( -\theta_c ||V(c(i)) - V(c(j)) ||^2)

\theta_c = \frac{1}{n}\sum_{i=1}^n||V(c(i))||^2

Construction of disease similarity model

疾病的GIP相似性构建 Gd(d(i),d(j)) 和 circRNA 构建方式一样

疾病的语义相似性用 MesSH数据库构建。它使用DAG反映不同疾病之间的关系。一个疾病d 机构用 DAG_d=(d,N_d,E_d) 表示 ,N_d表示所有和d有关系的节点，包括d ，E_d 表示这些病之间的关系。对于 DAG_d 内的疾病 s ，它的贡献值 D_d(s) 计算如下

{\left\{\begin{matrix} D_d(s) = 1 && if\ s=d \\ D_d(s) = max\{\mu \cdot D_d(s')|s' \in children\ of\ s \} && if \ s \ne d \end{matrix} \right. }

这里是层级的关系，个人理解，d 值为1 所有和 d 上游（父节点）为 0.5 ，然后上游的上游为0.5*0.5，依次进行。

\mu 表示疾病和它的子疾病的贡献因子，这里取0.5。把所有疾病 d 相关疾病的值加起来得到它们的语义值

DV(d) = \sum _{s\in N_d} D_d(s)

根据DAG中疾病的层级结构关系疾病d(i) 和 d(j) 之间的语义相似性 SV_1(d(i),d(j))

SV_1(d(i),d(j))= \frac{\sum_{s\in N_d(i) \cap N_d(j)}(N_{d(i)}(s) + N_{d(j)}(s))}{DV(d(i)) + DV(d(j))}

DAG中不同疾病的数量也会影响其语义相似性

D'_d(s) = - log(\frac{num(DAGs(s))}{num(diseases)}

第二个语义相似性模型 SV_2(d(i),d(j))

SV_2(d(i),d(j))= \frac{\sum_{s\in N'_d(i) \cap N'_d(j)}(N'_{d(i)}(s) + N'_{d(j)}(s))}{DV(d(i)) + DV(d(j))}

实际计算后发现，由于661种circRNA和100种disease中只有很少一部分有关联，所以计算得到的矩阵十分稀疏，最后得到的相似度普遍过于接近，作用不大。

Multi-source data fusion

DSim(d(i),d(j)) = {\left\{\begin{matrix} \frac{SV_1(d(i),d(j)) + SV_2(d(i),d(j))}{2}&& if\ d(i)\ and\ d(j)\ has \ senamntic\ similarity \\ GD(d(i),d(j)) && otherwise \end{matrix} \right. }

RSim(c(i),c(j)) = GD(c(i),c(j))

FV(c(i),d(j)) = [RSim(i),DSim(j)]

Rsim(i) 表示 Rsim矩阵中第 i 行，也就是 c(i) 和所有其他circRNA之间相似性形成的向量，DSim(j)同理。

Feature extraction by fast learning with Graph Convolutional Networks

Prediction by forest PA classifier

Inferring Potential CircRNA–Disease Associations via Deep Autoencoder‑Based Classification

https://github.com/Deepthi-K523/AE-RF

Materials and Methods

Datasets

CircR2Disease 数据库

相似性计算方面大差不差

Method

采用autoencoder和随机森林分类器去推理联系，一对circRNA disease 构建一个特征向量，所有的对构建的特征向量Q，Q输入到编码器中，获得一个低维的向量Q‘，再通过解码器尽量的还原成Q，然后用Q’获得的向量预测。

Autoencoder‑Based Feature Selection

Random Forest‑Based Association Prediction

RGCNCDA: Relational graph convolutional network improves circRNA-disease association prediction by incorporating microRNAs

Materials and methods

Data source and preprocessing

CircRNA-disease association

CircR2Disease

MiRNA-disease association

HMDD v3.0 有 32281条 miRNA-disease 其中 1206 miRNAs 和 893 diseases 过滤后选用了11,824条

CircRNA-miRNA interaction

ENCORI 中 2293条circRNA-miRNA 其中 68 of 13839 circRNAs and 581 of 642 miRNAs.

然后构建邻接矩阵 CD, CM, MD 代表关系

CD circRNA-disease associations
CM circRNA-miRNA interactions
MD miRNA-disease associations

Disease similarity network

Disease semantic similarity

同 1

D_d(t) = {\left\{\begin{matrix} 1 && if\ t=d \\ max\{\mu \cdot D_d(t')|t' \in children\ of\ t \} && if \ t \ne d \end{matrix} \right. }

S(d) = \sum _{t\in N_d} D_d(t)

S_s(d(i),d(j))= \frac{\sum_{t\in N_{d_i} \cap N_{d_j} }(D_{{d_i} }(t) + D_{{d_j} }(t))}{S(d_i) + S(d_j)}

Disease GAS similarity

同 1

S^d_{GAS}(d_i,d_j) = exp(-\lambda_dR(d_i) - R(d_j)^2)

\lambda_d = \frac{1}{\frac{1}{n}\sum_{i=1}^nR(d_i)^2}

Disease fusional similarity network

同 1

S^d(d_i,d_j) = {\left\{\begin{matrix} S_s(d_i,d_j)&& if\ S_s(d_i,d_j) \ne 0 \\ S^d_{GAS}(d_i,d_j) && otherwise \end{matrix} \right. }

融合的相似性分数作为疾病相似性网络边的权重。

CircRNA similarity network

CircRNA functional similarity

疾病 d 和疾病集合 D={d_1,d_2,...,d_n} 之间的语义相似度

SS(d,D) = \max_{1\le t \le n} (S_d(d,d_t))

这里 S_d 在上文中未出现过，文章的意思应该是上文中的 S_s .

circRNA c_i 和 circRNA c_j 之间的功能相似性

S_f(c_i,c_j) = \frac{\sum _{1\le i\le |D_i|}SS(d_i,D_j) + \sum_{1\le j \le |D_j|}SS(d_j,D_i)}{|D_i| + |D_j|}

D_i,D_j 表示和 c_i, c_j 相关的疾病集合。

CircRNA GAS similarity

S^c_{GAS}(c_i,c_j) = exp(-\lambda_cR(c_i) - R(c_j)^2)

\lambda_c = \frac{1}{\frac{1}{n}\sum_{i=1}^nR(c_i)^2}

CircRNA fusional similarity network

S^c(c_i,c_j) = {\left\{\begin{matrix} S_f(c_i,c_j)&& if\ S_f(c_i,c_j) \ne 0 \\ S^c_{GAS}(c_i,c_j) && otherwise \end{matrix} \right. }

Similarity calculation of miRNAs

计算基于这样的假设：具有相似表达模式的 miRNA 也具有相似的功能或生物学途径

miRNA i 和 miRNA j 之间的相似性可以通过miRNA表达谱之间的Pearson correlation coefficient (PCC) 计算

S^m(m_i,m_j) = \frac{\sum_{k=1}^n X_kY_k- \frac{\sum_{k=1}^nX_k\sum_{k=1}^nY_k}{n}}{\sqrt{\sum_{k=1}^nX_k^2 - \frac{\sum_{k=1}^nX_k^2}{n}}\sqrt{\sum_{k=1}^nY_k^2 - \frac{\sum_{k=1}^nY_k^2}{n}}}

X, Y 作为 m_i,m_j 的基因表达谱，每个表达谱包括n列从不同类型的人类组织测量的表达值

Global heterogeneous network

G= \begin{bmatrix} S^c & CM & CD \\ CM^T& S^m &MD \\ CD^T& MD^T & S^d \end{bmatrix}

RGCNCDA method

重启随机游走算法（Random Walk with Restart，RWR）和主成成分分析（principal component analysis PCA) 获取节点的特征输入到 RGCNCDA。然后使用R-GCN编码器和DistMult解码器构建一个预测模型。带有Laplacian regularization term 的损失函数计算最后的分数。

relational graph convolutional networks (R-GCNs) 关系图神经网络

DistMult 模型用作衡量可信度的评分函数。

Experiments and results

一样的

Case study

GMNN2CD: identification of circRNA–disease associations based on variational inference and graph Markov neural networks

https://github.com/nmt315320/GMNN2CD

Materials and methods

Datasets

用了5个数据集 CircR2Disease , Circ2Disease , circRNA–disease , circAtlas and CircFunBase

Disease semantic similarity 1

同1

Disease semantic similarity 2

同1

Gaussian interaction profile kernel similarity for disease

同1

Gaussian interaction profile kernel similarity for circRNA

同1

Comprehensive similarity of multisource data fusion

同1

circRNA-disease association prediction

GCNCDA: A new method for predicting circRNA-disease associations based on Graph Convolutional Network Algorithm

Materials and methods

Method overview

Benchmark dataset

Construction of CircRNA similarity model

Construction of disease similarity model

Multi-source data fusion

Feature extraction by fast learning with Graph Convolutional Networks

Prediction by forest PA classifier

Inferring Potential CircRNA–Disease Associations via Deep Autoencoder‑Based Classification

Materials and Methods

Datasets

Method

Autoencoder‑Based Feature Selection

Random Forest‑Based Association Prediction

RGCNCDA: Relational graph convolutional network improves circRNA-disease association prediction by incorporating microRNAs

Materials and methods

Data source and preprocessing

CircRNA-disease association

MiRNA-disease association

CircRNA-miRNA interaction

Disease similarity network

Disease semantic similarity

Disease GAS similarity

Disease fusional similarity network

CircRNA similarity network

CircRNA functional similarity

CircRNA GAS similarity

CircRNA fusional similarity network

Similarity calculation of miRNAs

Global heterogeneous network

RGCNCDA method

Experiments and results

Case study

GMNN2CD: identification of circRNA–disease associations based on variational inference and graph Markov neural networks

Materials and methods

Datasets

Disease semantic similarity 1

Disease semantic similarity 2

Gaussian interaction profile kernel similarity for disease

Gaussian interaction profile kernel similarity for circRNA

Comprehensive similarity of multisource data fusion

Graph Markov neural networks