Can large scale structural modeling be applied to peptide screening and deorphanization?

Discussion/ConclusionsPresent analyses rely on the assumption that the scoring of models generated using Alphafold2 is able to discriminate models containing different entities.Note that peptides are free to bind to any place on the protein surface. For the GPCR deorphanization, varied receptors are confronted with a given peptide, and the results are rather satisfactory, consistently with the results in [2]. For the GPCR peptide screening, different peptides are confronted with a given receptor.Here again, the results are encouraging. For the PEPscan candidate validation, the results are more ambiguous.In some cases, no pose of the peptide targets the protein-protein interface.Furthermore, all peptides tend to bind the same location on protein surface (see Figure 3), which is possibly not the desired one.Overall, the specificity of the protein-peptide modeling is at stake, and our results suggest that it is really poor. When peptides are confronted with the same protein, it is possible the scores lose effectiveness due to the smaller size of the peptides compared to their target protein. Ranking heterogeneity is possibly an indicator of the poor performance of the scoring. Why such a difference between PEPscan candidate validation and GPCR peptide screening ? Different hypotheses under investigation are: All peptides confronted with GPCRs are naturally occurring peptides whereas PEPscan candidates might exhibit different properties and correspond to a class of peptides not learnt by AlphaFold2. PEPscan peptides that correspond to protein fragments could be more hydrophobic than natural peptides selected during evolution, and thus be more sticky.It is however surprizing that in some cases such peptides do not target the protein-protein interface which should be more hydrophobic.Overall our results suggest the need to develop peptide-protein specific scores to discriminate the true binders from the incorrect ones.Another challenge is that of protein-peptide binding specificity, but this supposes to go back to AlphaFold2 internal prediction scheme, especially to increase the diversity of the poses.