Identifying errors in sequence alignment to improve protein comparative modelling

The difference between the number of known protein sequences and the number of protein structures is vast and comparative modelling offers a way to bridge this gap. Misalignment between target and parent is the largest cause of error in comparative modelling and we define SSMAs (Sequence-Structure MisAlignments) as regions where sequence and structural alignments do not agree. We find that most SSMAs are short (< 10 residues) and that there is a strong preference for starting and finishing an SSMA in an unstructured region or a turn. Neural networks were trained to identify regions of sequence likely to be mis-aligned, first using single sequences to predict ‘alignability’ of homologues with ≤ 35% sequence identity and then combining predictions for single sequences to predict SSMAs in an alignment of two sequences. Predictions of SSMAs in single sequences had positive predictive values up to 89.1% (MCC=0.798) while the alignment predictions had positive predictive values 92.9% (MCC=0.648). In combination with a program to permute alignments, these networks were applied to comparative modelling of sequences previously submitted to CASP5. The average RMSD of these models improved by some 37% illustrating that the method is likely to be extremely valuable in improving alignment for comparative modelling.