BindProfX: Assessing Mutation-Induced Binding Affinity Change by Protein Interface Profiles with Pseudo-Counts

“Since the number of collected sequences is often much lower than that required for stable Boltzmann statistics, here it is necessary to make corrections to the amino acid probability in the interface multiple structural alignment (iMSA) according to amino acid substitutionmatrix [12].”

“To calculate the BindProfX score, we first constructed the complex structures by overlaying the unbound monomer structure to the template complex structure provided by ZEMu, using the TM-align structural alignment program [22].”

“118), which was based on FoldX potential [14] using the complex structure refined from MacroMoleculeBuilder (MMB) simulations after the structural-alignment overlay [21].”

“One of themostly usedmethods is the log-odd likelihood score [9] that is equivalent to the averageof the substitution scales from theDayhoff [10] or BLOSUM [11] matrix between target amino acid and all amino acids at the corresponding position in the multiple interface alignments; this approach was taken in the former…”

“As expected, the difference between the training and test becomes smaller in the mutation-level cross-validation compared to the protein-level ones, probably due to the effect of homologous correlation between test and training samples, which is consistent with the observation made on the cross-validation of monomer-c…”

“In addition to the physics-based approach of FoldX, we reported in Table 1 the results obtained from two statistical potentials: BeAtMuSiC [19] and Dcomplex [20].”

“org/1 coupling effect from individual residues could further complicate the problem [3].”

“As expected, the difference between the training and test becomes smaller in the mutation-level cross-validation compared to the protein-level ones, probably due to the effect of homologous correlation between test and training samples, which is consistent with the observation made on the cross-validation of monomer-c…”

“2 presents the correlation between experimental ΔΔG data and that predicted by the combined evolutionary and FoldX scores, for the 1131 single-point mutation samples in SKEMPI.”

“approaches [5,6] for mutation calculations.”

“However, predicting the binding affinity change upon amino acid mutations remains a challenge for physical force field due to the lack of accurate methods describing the interactions [2].”

“We also found several mutations in the ZEMu dataset that have inconsistent amino acid type to that in the PDB structure (e.g., " Immunoglobulin FC/ Fragment B of protein A " in 2jwd_3dz8_1fc2 has a single mutation " YC133W " in ZEMu, but residue 133 of chain C in the structure is actually W, not Y); these mutations ha…”

“with low-resolution complex structure built by protein docking [5,23] and multi-chain threading approaches [6,24].”

“BindProf program, although BindProf used a different interface template structure library frommultiple-chain threading approach that consists of ~55,000 protein– protein complexes [16].”

“We also listed in the figure the correlation coefficient data when we use the Gribskov log-odds profile score [9], that is, using the log-odds difference between the wild-type and mutant amino acid to calculate the free-energy changes, which has been adopted in the first version of BindProf [4].”

“Interface alignment and structural analog search are performed by the I-align program [15], which is built on IS score:”

“Our NIL is derived from the PIFACE library [7] that consists of 130,209 protein dimer interfaces extracted from the PDB.”

“with low-resolution complex structure built by protein docking [5,23] and multi-chain threading approaches [6,24].”

“The application of the BindProfX scoring approach to guide the evolutionary-based protein design [27] on protein–protein interactions is under progress.”

“How to build a connection between the interface profiles and the protein binding affinity change upon mutations is, however, a question, especially given the limit number of known interactions in the current structure library [7,8].”

“Ona test set of 104mutationswith only unbound monomer structures available [21], BindProfX uses the complex structure docked from structural overlay of monomers onto the homologous complexes and achieved a correlation 0.”

“A number of previous methods have been tested on this dataset, such as BindProfX (Xiong et al., 2017), Profilescore (Lensink and Wodak, 2013), BeAtMuSiC (Lensink and Wodak, 2013), SAMMBE (Petukh et al.”