Trimming of AlphaFold Predicted Protein Structures for Functional Characterization

The recent advances in protein structure prediction performance by AlphaFold have vastly expanded the reach of protein structure analysis in biological research, but the surfaces of proteins represent unique challenges.Often, exposed loop or terminus regions are highly flexible and represent structural ensembles rather than a single defined structure, but AlphaFold sometimes predicts such regions to have a protruding "stringy" structure.These predicted structural elements are not functionally meaningful and may affect downstream analysis of these predictions such as structural alignment, surface analysis or ligand binding prediction through docking.We therefore developed a heuristic approach to remove such sub-structures automatically.Residues were trimmed from a structure if every nearby residue in 3D space, as defined by an adjustable distance cut-off parameter, is also close in the amino acid sequence, as quantified by whether or not it is inside an adjustable amino acid cut-off (e.g. if two residues are adjacent in the sequence, their distance is one).The trimmed predicted structures were then compared to the original AlphaFold predictions and the crystal structures.Trimming enhanced the agreement between the surface characteristics of the predicted structures and experimentally determined structures.Furthermore, trimming improved the process and the outcome of protein-ligand docking with Autodock Vina, as evidenced by the observations that (1) trimming decreased the bounding box volumes, leading to a reduction in out-of-memory errors and (2) the extraneous structures did not "distract" ligands to bind to structures that do not exist in stable form.