Abstract/Summary

Computational protein modelling has increased in public profile following the success of AlphaFold2 at CASP14 in 2020. This led many to proclaim the protein folding problem essentially solved, meaning in silico methods could now fill the sequence-structure gap which had grown since the advent of next generation sequencing techniques. However, proteins which prove problematic to experimental methods like X-ray crystallography and NMR are often multimeric in nature, like trans-membrane proteins or receptor binding interactions and, as the 2020 success was limited to tertiary structures, significant obstacles in quaternary structure elucidation remained. Contemporaneous analysis of assembly modelling showed that atomic contact prediction was a particular weakness and, as model refinement focusses on correcting small errors in atomic positioning, we proposed that a novel refinement method could be realised if full model coordinate files could be successfully submitted and recycled through the AF2 neural network. We present data in this thesis demonstrating that this is possible and that it significantly improved the quality of models including the official AF2 competition models from CASP14. Model quality assessment programs for quaternary structures had been largely absent with modellers relying on various proprietary accuracy estimates and docking scores. ModFOLDdock was conceived to independently evaluate multimeric model quality from any modelling software. Here we show how ModFOLDdock was improved by neural network training using three conceptual target scores and regression analysis leading to a significant increase in predictive performance. Further optimisation of our three unique combinations of distance-based quality measures resulted in the definition of three ModFOLDdock variants, all of which were subsequently highly placed in the CASP15 EMA competition, ranking 2nd for global score, 1st for interface score and 2nd for interface residue score. Evidence is also presented showing that ModFOLDdock outperforms the AlphaFold2 quality measures plDDT and pTM at quality-ranking quaternary structure models.