FIRE-pro seeks to identify protein motifs whose pattern of presence and absence across all amino acid sequences is highly informative about the behavior profile for the corresponding proteins. The algorithm takes as input a user-specified protein behavior profile listing a quantitative measurement or discrete attribute of every protein (e.g., half-life or nuclear localization). Presented is a schematic example using discrete localization data. Here, knowing whether the motif is present or absent in the amino acid sequence provides significant information regarding the behavior of the protein. For each candidate motif (e.g., “KRK”), FIRE-pro calculates the correlation between the motif profile and the protein behavior profile using mutual information. Motifs that maximize the mutual information are ultimately selected for further characterization.

(A) P-value heatmap of motifs found in Cdc28-interacting proteins. Columns correspond to classes of proteins and rows correspond to predicted motifs. The yellow color-map indicates over-representation of a motif in a given class; significant over-representation (p<0.05 after Bonferroni correction) is highlighted using red frames. Similarly, the blue color-map and blue frames indicate under-representation. For each motif, we indicate 1) position-weight matrix (PWM) representation, 2) mutual information (MI) value, 3) z-score associated with the MI value, 4) robustness score ranging from 0 to 10/10. (B) Motif interaction heat map. Columns/rows correspond to motifs. Light-colored boxes represent co-occurring motifs and “+” signs represent spatial co-localization. Dark-colored boxes represent motif co-avoidance. Values represent information (in bits). (C) Auto-generated enrichment analysis table. For each motif, we indicate 1) the presence of a position bias, 2) GO enrichment, 3) domain enrichment, 4) domain overlap score indicating the positional overlap between the motif and the most enriched domain. (D) Positional bias of “SP.[RK]”. For every motif, a histogram is automatically generated showing the distribution of motif instance positions, normalized to protein length. Upper row, histogram of motif instance positions in Cdc28-interacting proteins (“Targets”). Lower row, histogram of positions in proteins that do not interact with Cdc28 (“Other”).

The data [44] were clustered into six distinct localization patterns, each represented by a column of the matrix: nucleus, mitochondria, cell periphery, nucleus & cytoplasm, cytoplasm, and ER (see Figure S2 and Text S1). (A) P-value heatmap of motifs uncovered in the analysis. The top motif, “K[KRP].K”, matches the well-known nuclear localization signal (NLS). The hydrophobic motifs found to be enriched in ER proteins may suggest the existence of signals within stretches of hydrophobic residues. Enrichment analysis for the motifs can be found in Figure S3. (B) Linear sequence position bias of a mitochondrial motif corresponding to the “RFYS” consensus sequence for the N-terminal mitochondrial signal peptide cleavage site [47], [48]. Comparing motif positions in mitochondrial proteins against non-mitochondrial proteins reveals strong N-terminal enrichment.

Half-life data for ∼3,750 yeast proteins [50] were sorted and binned into ten equally populated classes, with the shortest half-life proteins comprising the left-most column and longest half-life proteins comprising the right-most column. The range of half-lives in each bin in minutes is indicated below the heatmap. Four motifs were found to be informative of half-life, all of which are associated with short half-life. The heatmap shows a gradual transition from over- to under-representation of each motif across the ten bins. The top motif shows an association with protein kinase domains though it does not overlap with the domain, while the bottom three motifs may represent protein kinase domain signatures (see Figure S4 for functional enrichment analysis).