TABLE 2.
Non-exhaustive list of deep learning architectures for protein-ligand binding affinity prediction and their performance on the CASF-2016 scoring benchmark (if available). MLPs are included regardless of the number of hidden layers. Some methods are described in multiple publications and the ones referenced in this table are the ones where the model has been evaluated on the PDBbind Core set 2016/CASF-2016 set (or the original publication, if this evaluation is not available). The best result (the highest Pearson’s r) is reported. Different publications might use slightly different custom variations of the CASF-2016 benchmark and the overlap between training and test sets might be taken into account in different ways. We refer the reader to the original publications for details, but we also report the number, N, of systems in the test set to outline possible differences. RMSEs are expressed in pK units.
| Model | References | Architecture | Pearson’s r | RMSE | N |
|---|---|---|---|---|---|
| — | Artemenko (2008) | MPL | — | — | — |
| NNScore 2.0 | Durrant and McCammon (2011b) | MPL | — | — | — |
| BgN- & BsN-Score | Ashtawy and Mahapatra (2015) | MPL | — | — | — |
| DLscore | Hassan et al. (2018) | MPL | — | — | — |
| PLEC-NN | Wójcikowski et al. (2018) | MLP | 0.82 | — | 290 |
| Pair | Zhu et al. (2020) | MLP | 0.75 | 1.44 | 285 |
| AEScore | Meli et al. (2021) | MLP | 0.83 | 1.22 | 285 |
| TopologyNet | Cang and Wei (2017) | CNN | 0.81 | 1.34 | 290 |
| Kdeep | Jiménez et al. (2018) | CNN | 0.82 | 1.27 | 290 |
| Pafnucy | Stepniewska-Dziubinska et al. (2018) | CNN | 0.78 | 1.42 | 290 |
| 1D2D-CNN | Cang et al. (2018) | CNN | 0.85 | 1.21 | 290 |
| DeepAtom | Li et al. (2019c) | CNN | 0.81 | 1.32 | 290 |
| OnionNet | Zheng et al. (2019) | CNN | 0.82 | 1.28 | 290 |
| Gnina | Francoeur et al. (2020) | CNN | 0.80 | 1.37 | 280 |
| RosENet | Hassan-Harrirou et al. (2020) | CNN | 0.82 | 1.24 | |
| AK-Score | Kwon et al. (2020) | CNN | 0.81 | — | 285 |
| LigityScore1D | Azzopardi and Ebejer (2021) | CNN | 0.74 | 1.46 | 285 |
| OnionNet-2 | Wang et al. (2021d) | CNN | 0.86 | 1.16 | 285 |
| SE-OnionNet | Wang et al. (2021b) | CNN | 0.83 | — | 285 |
| ACNN | Gomes et al. (2017) | GNN | — | — | — |
| PotentialNet | Feinberg et al. (2018) | GNN | — | — | — |
| graphDelta | Karlov et al. (2020) | GNN | 0.87 | 1.05 | 285 |
| SIGN | Li et al. (2021c) | GNN | 0.80 | 1.32 | 290 |
| InteractionGraphNet | Jiang et al. (2021) | GNN | 0.84 | 1.22 | 262 |
| GraphBAR | Son and Kim (2021) | GNN | 0.78 | 1.41 | 290 |
| PLIG/GATNet | Moesser et al. (2022) | GNN | 0.84 | 1.22 | 272 |
| PIGNet | Moon et al. (2022) | GNN | 0.76 | — | 283 |
| — | Berishvili et al. (2019) | CNN/RNN | — | — | — |
| FAST | Jones et al. (2021) | CNN + GNN | 0.81 | 1.31 | 290 |
| BAPA | Seo et al. (2021) | CNN + ATT | 0.82 | 1.30 | 285 |
| PointTransformer | Wang et al. (2021c) | CNN + ATT | 0.85 | 1.19 | 285 |