Partner-aware prediction of interacting residues in protein-protein complexes from sequence data

Abstract

Computational prediction of residues that participate in protein-protein interactions is a difficult task, and state of the art methods have shown only limited success in this arena. One possible problem with these methods is that they try to predict interacting residues without incorporating information about the partner protein, although it is unclear how much partner information could enhance prediction performance. To address this issue, the two following comparisons are of crucial significance: (a) comparison between the predictability of inter-protein residue pairs, i.e., predicting exactly which residue pairs interact with each other given two protein sequences; this can be achieved by either combining conventional single-protein predictions or making predictions using a new model trained directly on the residue pairs, and the performance of these two approaches may be compared: (b) comparison between the predictability of the interacting residues in a single protein (irrespective of the partner residue or protein) from conventional methods and predictions converted from the pair-wise trained model. Using these two streams of training and validation procedures and employing similar two-stage neural networks, we showed that the models trained on pair-wise contacts outperformed the partner-unaware models in predicting both interacting pairs and interacting single-protein residues. Prediction performance decreased with the size of the conformational change upon complex formation; this trend is similar to docking, even though no structural information was used in our prediction. An example application that predicts two partner-specific interfaces of a protein was shown to be effective, highlighting the potential of the proposed approach. Finally, a preliminary attempt was made to score docking decoy poses using prediction of interacting residue pairs; this analysis produced an encouraging result.

Figure 1. Residue pair and single-residue contact data preparation from ligand/receptor complexes (an example with a dimeric ligand complexed with a dimeric receptor is shown here).

For each of the 124 complexes, the data sets were prepared by pairing residues from the ligand and receptor chains for a pair-wise prediction (as shown in the second part of the illustration). Single-residue data were also prepared for the whole complex. However, the residues were not encoded as pairs; they were taken from individual chains and partner information was discarded, and the contact data for all the chains were pooled together to obtain whole-complex data. In one training cycle, the contact and feature data from all but one of the complexes were used for training, and the left-out complex data were then used to evaluate prediction performance. Performance scores were calculated for one complex in one training cycle. The obtained set of 124 scores was then averaged to obtain an overall performance score.

Figure 2. Overall prediction of interacting pairs of residues in two stages.

Figure 3. Sequence-based residue-pair contact propensities (natural logarithmic values) in a protein-protein interface.

Each plot corresponds to interface propensity of a residue with all of the 20 possible partner residues. Single residue propensity values for the target residue are shown by a horizontal dashed line. See Table 1 for comments and Table S1 for additional details.

Figure 4. Binding site predictions mapped to the three-dimensional structure of Acetylcholinesterase in complex with Toxin F-VII Fasciculin-2 (PDB ID: 1MAH, chains A and F respectively in red and blue color cartoons).

The left and right images were drawn from the top-scoring 20 predictions from single-protein trained models (solid red) and pair-wise trained models (solid green), respectively. Many false positive cases observed in the single-protein trained model were eliminated in the pair-wise model. (The false positive rate in the selected 20 residues is 75% and 50% with an overall AUC of ROC being 60% and 82%, respectively. Predictions are made from the models trained by excluding this complex from the training data.)

Figure 5. Relationship between prediction performance and the RMSD between bound and unbound complexes.

Even though there are few data points in high RMSD category making the statistical point only suggestive in nature, poorer prediction performance for complexes undergoing large conformational change is consistent with the arguments in the discussion.

Figure 6. Partner-specific prediction of two interfaces for the beta subunit of the guanine nucleotide binding protein (transducin beta chain 1) (PDB ID: 3PSC, chain B is shown as the blue cartoon).

The two partners are shown in transparent colors (chain A, which is the beta-adrenergic receptor kinase-1, is shown in red, and chain G, which is the guanine nucleotide binding protein's gamma-2 subunit, is shown in green). Predictions from the pair-wise model for each partner chain were converted into single chain predictions and displayed on chain B. Common binding sites, predicted with both partners, were removed and residues exclusively predicted with each partner are shown in the corresponding partner color. Out of the 30 top-scoring residues after removing common predictions, most residues have been assigned to the correct partner.