PMC

Modeling of local best path selection model for analysis of error probability in case of a fixed window size. (a) Expression of SNP within window size. (b) Expression of an indel within window size. (c) The scoring system of the alignment.

Accuracy of membrane protein sequence alignments/homology models obtained from different sequence alignment methods as a function of sequence identity. Results are given for (A) the whole protein and (B) the transmembrane regions. The average AL0 score is given over all alignments/models within a window of 10% sequence identity, and error bars indicate the standard deviation over that window. Numbers correspond to the number of alignments in each window and apply to both plots. Abbreviations: seq-seq, sequence-to-sequence alignment; seq-profile, sequence-to-profile alignment; and MSA, multiple sequence alignment. The two HMAP labels indicate profile-to-profile alignments.

(a) Four algorithms learn to mimic a linear function: ‘shallow' learning (light grey); reinforcement learning (dark grey); backprop (black); and feedback alignment (green). NSE is normalized squared error (see Methods). (b) The angle between modulator vectors prescribed by feedback alignment and backprop, δ_FA and δ_BP, decreases. Error bars are two s.d.'s for a sliding window of 10 examples.

Illustrative schematic of SWAMP masking. (A) Unmasked alignment with two potential data errors, short stretches of divergent sequence possibly caused by sequencing errors, and longer stretches possibly caused by exon splicing. (B) Application of a two-step filtering masks data errors; a shorter window of 3 AA with a maximum of 2 substitutions effectively masks possible sequence error (orange), and a longer window of 15 AA with a maximum of 10 substitutions effectively masks longer alignment errors (blue). (C) The ‘Interscan’ option further masks the data if short stretches of sequence are surrounded by masked data and the sum of the masked data either side of the sequence exceeds the length of the interceding unmasked data (purple).