Table 1.
Summary of scoring functions used to score or classify the homodimer interfaces of the benchmark dataset, and to compute consensus scores.
Column 1 lists individual groups (PI name -group). Column 2 includes a high-level description of the methods with literature references and links to servers provided whenever appropriate. Column 3 lists more detailed descriptions of the scores. Additional information can be found in the reports of individual groups collated in the Supplementary Material.
| Group | Feature type / source | Brief scoresFeature description |
|---|---|---|
| Fernandez-Recio | CCharPPI features [58] ---- Integrated scores I_pyDock_Desolv_VDW [59] |
- Residue contact/step potentials - Residue distance-dependent potentials - Atomic contact/step potentials - Atomic distance-dependent potentials - Statistical potential constituent terms - Interface packing: F_NIPacking - Composite scoring functions - Solvation energy functions - Hydrogen bonding - Van der Waals and electrostatics --------- pyDock [59]: Desolvation, electrostatics and Van derWaals scores |
| MOBI | Three descriptors for each interface (F_shape, F_hydro and F_tails) and one integrated score (I_shape_hydro_tails) [60,61] (see MOBI group Report in Supplementary Material) | - I=F_shape+F_hydro(Ftails) - F_shape (Sum of # of hits/residues) - F_hydro (Frac. of surface hydrophobicity) - F_tails (Y/N of chain ends) |
| Venclovas | Multiple Voronoi tessellation-derived features computed using the Voronota software (https://kliment-olechnovic.github.io/voronota/) [21,62] | - Voronoi tessellation-derived interface contact areas, - Solvent contact areas, - All the contact areas. - Voronoi tessellation-derived volumes. - VoroMQA energy values (of interface contacts, solvent contacts, all the contacts) - VoroMQA-light and VoroMQA-dark scores . |
| Wolfson | All-atoms scores [63] --------- Deep-Learning scores [64] |
- F_SOAP: interaction score (all atoms) - F_FireDockScore (all atoms) - F_Network_binding_0 - F_Network_binding_1 (DL NN, P_residues in interface) - F_Network_full (DL, NN, P_residues in contact) --------- - I_Proba_Consensus (integrated score) |
| Zou | [65,66] | - ITScorePP (atomic-level, statistical potentials) --------- - DLScoreBC (DL/CNN model for interface prediction) |
| Bonvin | Two different classifiers [22,67] Scoring method used in HADDOCK [68] DeepRank-GNN [69]: The PPI interfaces were converted into residue graphs and each node was assigned PSSM information only (i.e. 20 features per node). |
- PRODIGY-CRYSTAL [67]: RF classifier (residue contacts; residue contacts per amino acid type, contact density/interface, trained on the MANY dataset [12]) --------- - DeepRank-GNN (DL/GNN, trained using PSSM features only on the MANY dataset [12]) --------- - HADDOCK score and its raw components [68] (not trained as a classifier) |
| Furman | Combined docking and local refinement with RosettaDock, InterfaceAnalyzer protocol (multiple features) [70,71], RosettaCommon* | - I_sc: interface score (‘Interaction energy_1’) - dG_cross/dSASA* (‘Interfaces energy_ 2’) - sc_value: shape complementarity - fa_intra_sol_xover4: intra-residue LK solvation - dG_separated/dSASA: binding energy of separated components/unit interface area - fa_dun: Internal energy of sidechain rotamers - dSASA_polar: polar components of interface area - fa_intra_rep: Lennard-Jones repulsive between atoms in the same residue |
| Oliva | COCOMAPS/CONSRANK features [72] and BSA calculated with NACCESS (http://www.bioinf.manchester.ac.uk/naccess/) ----------- Integrated scores: I_NN_all, I_NN_sel, I_RF_all, I_RF_sel [32] (see Oliva Group Report in Supplementary Material) |
- Residue-residue contact stats, including the total number of contacts at the interface, the number of contacts per physico-chemical class of amino acids involved in the contacts (polar, apolar, aliphatic, aromatic and charged) and the relative fraction over the total number of contacts per complex. - BSA upon complex formation plus the polar and apolar components calculated by NACCESS [73] and FreeSASA [74] --------- - Probability for a dimer to be physiological and predicted class (TRUE/FALSE) from Neural Network(NN)- and RF-based classifiers, using 148 (_all) and 42 selected (_sel) features |
| Kihara | A range of scores and potential including scores published by the Kihara group and other groups (See Kihara Group report in Supplementary Material). | Examples of scores: - DFIRE: all-atom statistical potential [75] - GOAP : all-atom statistical potential [75] - Dove: DL (3D CNN) model [76] - GNN-DOVE : Graph Neural Networks Model [77] - ITScore : knowledge-based scoring function [65] - PhysicsScore: physics-based score in Multi-LZerd [78] - RosettaInterfaceEnergy : Interface Energy of Rosetta Energy Function [79] - VoroMQA [21] |
| Casadio | ISPRED4 predictor of protein interaction sites (https://ispred4.biocomp.unibo.it) [80] | - Support vector machines (SVMs) and grammatical-restrained hidden conditional random fields (GRHCRFs) incorporating 46 features: - Sequence profiles (MSA) - Residue physical-chemical properties: - PSICOV: intra-chain coevolution scores (Jones et al. 2012) - Interface residue propensity - Difference between predicted and observed solvent exposure - Structural/geometric features: secondary structure, DPX, CX (computed using PSAIA (Mihel et al. 2008)) |
| SWISS-MODEL | QSQE score from SWISS-MODEL[46,81] | - QSQE: composite score (values 0–1); ML(SVM)-based (interface conservation, structural clustering, and other template features); depends on availability of templates in the SWISS-MODEL template library (not trained as a classifier but to rank templates for modelling). |
| Guerois | Scores from InterEvDock [82] | - SPPh.10seq and SPPh.10seq.normsize: novel version of SOAP-PP [63] using coevolutionary information at atomic level, using information from a set of 10 homologous complexes (.normsize: score normalized by interface size). - IESh.10 seq and IESh.10seq.normsize: same as above but using InterEvScore [83] as a base scoring function instead of SOAP-PP |
| Correia | DL MaSIF model [84] | Integrated score, combining chemical (electrostatics, H-bonds, hydrophobicity) and geometric (shape and curvature) features, in a surface descriptor for the interacting surface patch of each protein. --------- Computed the following quantities: - Descriptor Distance Score: complementarity of the 2-interacting surface patches. - Neural Network Alignment Score (0–1): calculates the alignment quality between the interacting surface patches. |