Table 3.
Summary of recent studies regarding the prediction of lncRNA–protein interaction. It is important to acknowledge that each study utilized diverse datasets, cross-validation methods, and simulation settings to assess accuracy, thus rendering direct comparisons potentially inconclusive. The best accuracy was selected if the model was assessed with various datasets.
| Ref. | Methods | Accuracy | Merits | Disadvantages |
|---|---|---|---|---|
| [58] | BiHo–GNN, bipartite graph embedding based on GNN | AUC: 0.950 AUPR: 0.899 |
High AUC and recall, outperforms existing methods | Not specified |
| [59] | Capsule–LPI, multimodal features, multichannel capsule network framework | AUC: 0.951 AUPR: 0.932 |
Superior performance, integration of multimodal features | Absence of detailed evaluation for each feature |
| [60] | DeepLPI, interactions between lncRNAs and protein isoforms with the hybrid framework of deep neural networks | AUC: 0.866 AUPR: 0.703 |
Use of isoforms, application of multiple instance learning | Lower performance metrics compared to other methods |
| [61] | DFRPI, deep autoencoder and marginal Fisher analysis, random forest-based predictor | AUC: 0.906 | Constructing a discriminative feature space, high precision | Necessity to generate a reasonable and effective feature space |
| [62] | EnANNDeep, an ensemble-based framework with an adaptive k-nearest neighbor classifier and deep models | AUC: 0.916 AUPR: 0.905 |
Incorporates multiple source features, performs well in cross-validations | May produce prediction bias with single dataset evaluation |
| [63] | iEssLnc, graph neural network with meta-path-guided random walks on the lncRNA–protein interaction network | AUC: 0.912 AUPR: 0.921 |
Provides quantitative essentiality scores for lncRNA genes | Specific to essential lncRNA genes, not general lncRNA–protein interactions |
| [64] | LGFC–CNN, deep learning-based prediction combining raw sequence composition, hand-designed, and structure features | AUC: 0.976 AUPR: 0.970 |
Multiple-feature integration, highly accurate performance | Not specified |
| [65] | LPI–CSFFR, a feature fusion method based on CNN with feature reuse and serial fusion | AUC: 0.879 | Integrates diverse features of lncRNAs and proteins, high accuracy | Requires complex feature fusion |
| [66] | LPI–deepGBDT, multiple-layer deep framework based on gradient boosting decision trees | AUC: 0.9073 AUPR: 0.8849 |
Uses diverse biological information of lncRNAs and proteins | Limited application for new lncRNAs or proteins |
| [67] | Deep learning framework with dual-net neural architecture, LPI–DLDN | AUC: 0.911 AUPR: 0.898 |
Best average AUC and AUPR, outperforms six other LPI prediction methods | Requires dimension reduction for feature concatenation |
| [68] | LPI–HyADBS, AdaBoost-based feature selection, combined with DNN, XGBoost, C-SVM | AUC: 0.851 AUPR: 0.841 |
Hybrid approach integrates multiple classifiers, surpasses six other models | Requires complex integration of classifiers |
| [69] | LPICGAE, combined graph autoencoders | AUC: 0.974 Acc.: 0.985 |
Low-dimensional representations, outperforms six other computational methods | May need alternate loss minimization for optimal results |
| [70] | PRPI–SC, ensemble deep learning model using stacked denoising autoencoder and CNN | Acc.: 0.889 AUC: 0.950 |
Predicts plant LPIs, generalizes well beyond plant data | Only reports accuracy for plant data |
| [71] | RLF–LPI, AE–ResLSTM with fuzzy decision | Acc.: 0.921 AUC: 0.980 |
Potential for high performance due to the use of AE–ResLSTM and fuzzy decision | Not specified |