Abstract
Inflammation is a critical component of our immune system when correctly functioning, but issues with this pathway can cause numerous diseases, creating a need for anti-inflammatory therapies. Viruses, particularly poxviruses, have been identified as a potential source of these therapies as they produce immune modulating proteins that allow them to evade the immune system. Of particular interest to this group is the chemokine binding protein (CKBP) from orf virus (ORFV), a sheep pathogen that is also capable of human infection. This CKBP has been shown to interact with 14 human chemokines and 4 mouse chemokines, with an additional five human chemokines and one mouse chemokine that do not interact. Previous work studying the binding recognition mechanism led to crystal structures of ORFV CKBP in isolation and in complex with hCCL2, hCCL3, and hCCL7 being obtained. While these structures were important for understanding how these three chemokines are recognised, it was unclear if this recognition mechanism would generalise to other interacting chemokines. Additionally, it was unknown if there were any structural features common to non-interacting chemokines that prevent the interaction.
To address these outstanding questions, AlphaFold2 was used to predict the structures of the ORFV CKBP:chemokine complexes that had not been experimentally determined. The interfaces of these complexes were subsequently analysed using PISA, MAPPIS, and molecular dynamics software, the results of which were used to create a chemokine sequence alignment. From this, a consensus sequence was derived and used to make predict if a chemokine could interact with ORFV CKBP. This was done using custom, but reasonably simple. Python scripts, the outcomes of which helped to inform which interface elements were important for the interaction. The first of these, CIWI, used categorical variables to make the consensus then scored unseen sequences using a simple binary scoring system. Following on from predictions, nine chemokines were tested for binding by surface plasmon resonance (SPR), including four for which no prediction could be made. This was very successful, with all five predictions being proven correct and interaction behaviours determined for the four chemokines without predictions.
However, including these experimental results as part of an expanded sequence alignment had proved less successful, as interactors and non-interactors could not be distinguished.
A successor script, SCOUT, was written to use a more advanced consensus generation approach, with the inclusion of continuous variables and a more complex scoring system, the results of which were fed into a multi-layer perceptron algorithm. This approach was more successful at separating known binders and non-binders, with only two mistakes, but its ability to predict novel behaviour has yet to be tested. Additionally, analysis of the weighting given to each variable by SCOUT gave insight into what features the algorithm was considering when making predictions, which then informed the proposal of an interaction mechanism.
Using the results from all these methods, a two-stage interaction model was proposed. In this model, the many negatively charged residues of ORFV CKBP attract the positively chemokine to form the initial contacts. A mixture of hydrophobic interactions and hydrogen bonding between the chemokine N-tail and ORFV CKBP β8 strand then creates an anchor that stabilises the complex.
Further computational approaches were subsequently employed to investigate elements preventing binding and what mutations could be employed to counteract them. This began by modelling a selection of non-binding chemokines in the correct interface orientation, from which different clashes could be identified qualitatively by visual inspection, and quantitatively by PyClash, a custom Python script. Following this, FoldX was employed to simulate the effects of large libraries of interface mutations to both find problematic regions and potential solutions by identifying mutations that help stabilise the complex. Using information from all these approaches, mutation sets of interest were modelled in AlphaFold2, with any that resulted in a correct complex going on to be validated experimentally. This approach produced five mutant sets across four chemokines, three of which proved effective after testing by SPR.