Table 1.
Statistical information of the datasets. “#” represents the number.
Fig 1.
Distribution of datasets in two application scenarios.
(a) Scenario1: associated cancer ranking for novel queries (b) Scenario2: associated cancer ranking for known queries.
Table 2.
Comparison of different rankers in LTR.
Fig 2.
The impact of parameters of LambdaMART model.
(a), (b), (c), and (d) respectively represent the AUC and NDCG@10 values obtained by AutoEdge-CCP under variations in the Number of Trees, Learning Rate, Number of Threshold Candidates, and Minimum Leaf Support.
Table 3.
Performance comparison of AutoEdge-CCP and other methods in novel circRNA associated cancers prediction.
Table 4.
Performance comparison of AutoEdge-CCP and other methods in novel drug associated cancers prediction.
Fig 3.
Performance of AutoEdge-CCP in multiple scenarios.
(A) ROCk values comparison between AutoEdge-CCP and alternative methods in Scenario1. (B) Overall ROCs for 46 cancers. Median AUROC was shown on the top of each panel. Here, each gray line represents one cancer, the red line represents the median curve, and the light green part represents the region between the 25th and 75th quantiles. (C) Box plot depicting the metric scores of AutoEdge-CCP in Scenario 2. (A-C): left side presents circRNA-cancer association prediction, right side presents drug-cancer association prediction.
Fig 4.
Analysis of the edge features derived from autoGNN.
(A)-(B) Performance comparison under different graph embedding algorithms. (C) Performance comparison between AutoEdge-CCP and models without node feature or edge feature.
Table 5.
Two adaptive GNN framework for autoGNN and the ablated model.
Fig 5.
Heat maps of the similarity matrix for edge embedding.
(a) and (b) represent the edge embedding similarity matrices learned by AutoEdge-CCP for 12 pairs of circRNA-cancer and drug-cancer, respectively. Note: * designates the labeled pairs, and the rest are unlabeled pairs. The abbreviations correspond to the following full names: hsa_circ_0001733 (0001733), hsa_circ_0081161 (0081161), Lung Adenocarcinoma (LA); Head and Neck Squamous Cell Carcinoma (HNSCC), Papillary Thyroid Cancer (PTC), Breast Cancer (BC), Liver Cancer (LC), Multiple Myeloma (MM), Thyroid Cancer (TC), Nasopharyngeal Carcinoma (NPC), Acute Lymphoid Leukemia (ALL), Urinary Bladder Cancer (UBC), Prostatic Cancer (PC); Gastric Cancer (GC).
Table 6.
Top-ranked candidate cancers related to circ-RAD23B predicted by AutoEdge-CCP.
Table 7.
Top-ranked candidate cancers related to NVP-AUY922 predicted by AutoEdge-CCP.
Fig 6.
Visualization of NVP-AUY922 (PubChem CID: 135539077) and binding pockets.
(A) The 3D representations of NVP-AUY922 with the binding pockets of 5E8Y,4ZZJ and 3D3L. (B) The interaction maps of NVP-AUY922 with 5E8Y,4ZZJ and 3D3L.
Table 8.
The molecular binding energy of NVP-AUY922 with human target proteins associated with Esophageal Squamous Cell Carcinoma and Colorectal Cancer.
Fig 7.
The framework of AutoEdge-CCP.
There are four steps: (A). multi-source heterogeneous network construction. Integrating association data encompassing circRNA, drugs, and cancer from the circRic, circR2Cancer, and CTD databases. (B). Attribute feature representation. Extracting cancer, circRNA, and drug attribute features based on similarity calculations. (C). Edge feature representation. AutoGNN explicitly modeling link information to obtain edge features. (D). Query associated cancers ranking. The lambdaMART algorithm transforms the association problem into associated cancer lists ranking for queried circRNA or drug.