SKEMPI 2.0: An updated benchmark of changes in protein-protein binding energy, kinetics and thermodynamics upon mutation

Justina Jankauskaite; Brian Jiménez-García; Justas Dapkunas; Juan Fernández-Recio; Iain H. Moal

In total 7085 entries were collected, summarised in Table I and Figure 1A . These data were derived from the literature and consequently, while encompassing a broad range of residues, proteins, interactions and systems, are biased toward the interests and capabilities of the research community. These biases are evident in the composition of the database according to parameters shown in Figure 1 . The ∆∆G values span a large range, but mostly fall within -3 to 7 kcal.mol −1 (Figure 1B) , for both biophysical and technical reasons (see Section III-C). Almost three quarters of the data correspond to single point mutations, and more than half of those are mutations to alanine ( Figure 1C ). Charge swap mutations and mutations between aromatic residues are also over-represented. Most single point mutations are located at the binding site, and most of those are at the core of the interface. Similarly, most double mutations are both in the binding site, and most of those are both in the core. By far the most popular methods for measuring binding affinity were surface plasmon resonance and spectroscopic methods such as fluorescence. Large biases towards specific interactions and classes of interaction are also present, such as early studies into protease inhibition and immunological interactions such as antibody-antigen complexes, the recognition of peptides presented on cell surfaces by T-cell receptors, cytokine signalling and the complement system. Indeed, almost half of the data corresponds to protease-inhibitor, antibody-antigen and pMHC/TCR interactions alone. While many interactions within these classes share common binding sites or homologous binding sites (Figure 2 ), there are also connections between these groups, for instance via inhibitory antibodies which bind to a protease active site, or due to common binding regions of proteins in the immuoglobulin superfamily, such as antibodies, TCRs, MHCs and β-2 microglobulin. Also present in the data are smaller clusters of shared and homologous interactions, such as the Raseffector cluster. These relationships are noted in the database and may be useful for avoiding overfitting when developing models or for validation and estimating generalisation error, as described previously [47] .