Table 2.
Recently proposed deep learning methods for PPI prediction.
| Method | Year | Main learning structure | Sources of input feature | Encoding method | Combining method |
|---|---|---|---|---|---|
| DeepPPI [35] | 2017 | Multilayer Perceptron | Protein sequences | Seven sequence-based features (like amino acid composition) | Concatenation |
| DPPI [33] | 2018 | Convolutional Neural Networks | Protein sequences | Protein position specific scoring matrices (PSSM) derived by PSI-BLAST | Element-wise multiplication |
| DeepFE-PPI [36] | 2019 | Multilayer Perceptron | Protein sequences | Pre-trained model embedding (Word2vec [76]) | Concatenation |
| PIPR [34] | 2019 | Bidirectional Gated Recurrent Unit and Convolutional Neural Networks | Protein sequences | Pre-trained model embedding (Skip-Gram [56]) and the similarity of electrostaticity and hydrophobicity among amino acids | Element-wise multiplication |
| S-VGAE [51] | 2020 | Graph Convolutional Neural Networks | Protein sequences and topology information of PPI networks | Conjoint triad (CT) method | Concatenation |
| Liu’s work [77] | 2020 | Graph Convolutional Neural Networks | Protein sequences and topology information of PPI networks | One-hot encoding | Concatenation |
| DeepViral [66] | 2021 | Word2Vec model and Convolutional Neural Networks | Protein sequences, phenotypes associated with human genes and pathogens, and the Gene Ontology annotations of human proteins | DL2Vec embedding model [67] and one hot encoding | Dot product |
| FSNN-LGBM [52] | 2021 | Multilayer Perceptron | Protein sequences | pseudo amino acid composition (PseAAC) and conjoint triad (CT) methods | Element-wise multiplication |
| TransPPI [55] | 2021 | Convolutional Neural Networks | Protein sequences | Protein position specific scoring matrices (PSSM) derived by PSI-BLAST | Concatenation |
| DeepTrio [40] | 2021 | Convolutional Neural Networks | Protein sequences | Trainable symbol lexicon embedding | Element-wise addition |
| FSFDW [78] | 2021 | Skip-Gram (Deepwalk) | Protein sequences and topology information of PPI networks | Sequence-based features selected by Louvain method and Term variance | Element-wise multiplication |
| NXTfusion [68] | 2021 | Multilayer Perceptron | Protein-Protein, Protein-Domain, Protein-Tissue and Protein-Disease relations | One-hot encoding | Bilinear transformation |
| MTT [57] | 2021 | Multilayer Perceptron | Protein sequences | Pre-trained model embedding (UniReo [58]) | Element-wise multiplication |
| CAMP [79] | 2021 | Convolutional Neural Networks and Self-attention | Protein sequences, secondary structures, polarity, and hydropathy properties | Protein position specific scoring matrices (PSSM) calculated by PSI-BLAST and trainable symbol lexicon embedding | Concatenation |
| D-SCRIPT [60] | 2021 | Broadcast subtraction and multiplication, and Convolutional Neural Networks | Protein sequences | Pre-trained model embedding (Bepler and Berger’ work [61]) | Broadcast subtraction and broadcast multiplication |
| TAGPPI [63] | 2022 | Convolutional Neural Networks and Graph attention networks | Protein sequences and structures | Pre-trained model embedding (SeqVec [64]) | Concatenation |