Filtered neighbors threading.

We present a knowledge-based threading scoring function that exploits the information about protein structure contained in residue packing/neighbor preferences. The proposed algorithm eliminates the stereochemically improbable physical contacts for each possible sequence-to-structure alignment. We use this algorithm to "filter" the score of the sequence-to-structure alignment. Filtering is dynamic, in the sense that the set of neighbor pairs contributing to the alignment score varies during threading. Whether or not a neighbor pair contributes to the score depends on the threaded amino acids. We use a detailed structure description that encodes amino acid side-chain rotamer and physical contact preferences but does not imprint the fold model with the native sequence or native physical contacts. We discretize this description to collect accurate statistics for the scoring function generation. We use the original detailed description for the neighbor filtering. On average, the filtered neighbors threading (FNT) method predicts the sequence-to-structure alignment twice as accurately as does the "standard" unfiltered neighbors threading. For the set of threadings tested by the PHDthreader method, the FNT gives predictions with a sequence-to-structure alignment accuracy of 46.9%, which amounts to a 74% improvement in alignment sensitivity compared with PHDthreader predictions. These results show that reduction of noise from the observed neighbor pair preferences by filtering leads to noticeable improvements in the predicted sequence-to-structure alignments.