fLPS 2.0: rapid annotation of compositionally-biased regions in biological sequences

Implementation

The fLPS 2.0 package is written in standard C. The name ‘fLPS’ stands for fast LPS, where LPS stands for ‘Low Probability Subsequences’. The package comprises the source code and executables compiled for MacOSX and Linux. There is the fLPS program itself, plus two accessory programs: CompositionMaker, which can be used to calculate background residue compositions; and DomainFilter, which is used to either excise or mask previously annotated domains (such as those with known protein structure, see section immediately below). Each of the programs works on input files of any size in standard FASTA format. The package is available at https://github.com/pmharrison/flps2 and in File S1.

Algorithm and new added features

The program fLPS works through a process of binomial probability (P-value) minimization, as described in detail previously (Harrison, 2017). There are four main steps that are summarized in Fig. 1 at the top of the figure: (i) QUICK SCAN; (ii) MINIMIZE; (iii) MERGE; (iv) OUTPUT. At the end of the process, single-, and multiple-residue LPSs, are output if they are below the user-specified P-value threshold, or default threshold. Biased regions are labelled with a bias signature which is a list of the biasing residues in order of bias precedence delimited with curly brackets. At each of these stages, efficiency measures are taken to avoid or delay probability calculations unless/until they are necessary (Harrison, 2017).

The following are the main new options added to the fLPS code:

1. Precision of the calculation: In the initial QUICK SCAN step, by default (the ‘–z fast’ option), windows with a P-value below the baseline threshold of 0.001 are considered. Also, the windowing along the sequence proceeds with a step size = 3 residues (Fig. 1). This means that some regions that are made from biases for a larger number of residue types might be missed; also, short regions with a milder bias that might have biological significance could sometimes be overlooked. Therefore, options for the base-line precision of the program have been added. If ‘–z medium’ is specified, the base-line P-value threshold is set to 0.01, with a windowing step size = 2. For the most precise option ‘–z thorough’, the base-line P-value is 0.1 and the step size = 1 (Fig. 1). However, these latter two options can produce a huge amount of output for larger databases, so they should be applied to such databases with caution.

2. DNA analysis: DNA sequences can be specified using the –n option. By default, each of the four bases A, G, C and T has equal background probability.

3. Domain filtering: Using the DomainFilter accessory program, previously annotated domains can be filtered in either of two ways, i.e., either ‘excised’ or ‘masked’. When ‘excised’ is specified, DomainFilter outputs shorter sequences, with the domain sequences removed. The ‘masked’ option outputs sequences with the domains masked with Xs. The positions of the excised or masked domains are labelled on the name line of the sequences in the FASTA-format file. When the FASTA-format output file from DomainFilter is used as input for fLPS bias annotations, the domain positions appear in the fLPS output if the ”-option –D is specified.

4. Restriction lists: With the –r option, the user can specify a subset of residue types, e.g., only negatively-charged amino acids (E, D), or the six-membered aromatics (F, Y, W).

Further option additions include: An option (−O) option to specify a prefix for a unique output filename that also contains the parameters used in running the fLPS program; a ‘–o oneline’ output option, wherein the results for each sequence are listed in a single-line summary; a –k option to ignore the unknown residues in calculations (‘X’ for proteins and ‘N’ for DNA). The output has also been updated to include further new features. A calculation of the enrichment of the biasing residues in the output LPSs has been added, which is the proportion of biasing residues in the LPS divided by the total expected background frequencies of the biasing residues. To enable quicker characterization of bias trends in a data set, ‘bias class’ labels are now featured for both protein and DNA sequences. For proteins, these labels are derived from the Taylor amino-acid classification Venn diagram, with some additional categories (Taylor, 1986). The applicable class label that has the smallest membership is picked, when assigning these. For DNA, these labels represent the eight possible compositional biases: {A}—{T}, {G}—{C}, {AT}, {GC}, {AC}—{GT}, {AG}—{CT}, {ATC}—{ATG} and {ACG}—{CGT} (discussed below).

For better annotation of short low-complexity regions, trimming of LPSs of minimum window length is now employed. That is, if possible, residues are sheared off both ends of the minimum-length LPS if they do not contribute to the bias. This improves the annotation of ∼6–8% of LPSs, in trials on the S. cerevisiae S288C proteome (downloaded from UniProt reference proteomes (Boeckmann et al., 2003)) using a variety of parameters.

Example data

The UniProt canonical reference human and budding yeast (S. cerevisiae strain 288C) proteomes were downloaded from http://www.uniprot.org in January 2021 (Boeckmann et al., 2003). The human proteome was cross-referenced with the InterPro list of domain annotations downloaded from https://www.ebi.ac.uk/interpro/, to make a list of human proteins that contain the RRM RNA-binding domain (Blum et al., 2021).

Human promoter data was obtained from the EPD eukaryotic promoter database (Schmid et al., 2004). These were a set of representative promoters (one per gene) defined by the EPD. Sequences spanning from −999 to +100 around the transcription start site were analysed.

Prion-like regions

Prion-like regions were annotated for the human proteome using the PLAAC program with default parameters (Lancaster et al., 2014).

Using the new domain-filtering and restriction list options of fLPS: application to analysis of human RNA-binding proteins

It is often advantageous to restrict CB annotation to a subset of residues to enable easy counting of different types of bias region. Users are now able to restrict their bias annotation using a ‘restriction list’ specified with the –r option of fLPS (Fig. 2A). Also, it is possible that certain proteins have discontinuous CB regions, i.e., the CB regions may have small, structured domains embedded in them, or they may be comprised of the loop regions within a single protein domain. To enable discovery of such discontinuous CB regions, the DomainFilter program can be used to excise or mask domains or domain parts before using the fLPS program (Fig. 2A). These two options were combined in analyzing the CB regions of human RNA-binding proteins, specifically those containing the RRM RNA-binding domain (Fig. 2). The RRM domain is used in eukaryotes to bind RNA during diverse cellular processes, and is typically associated with intrinsically-disordered regions (Su & Harrison, 2020). After applying DomainFilter to excise Pfam protein domain annotations (Mistry et al., 2021), the main single-residue biases were assessed with an initial run of fLPS (those having > 50 cases); thereafter, a final run of fLPS used a restriction list based on these main single-residue biases to enable better counting of bias types.

There are 206 RRM-domain-containing human proteins in this protein data set. The most common multiple- and single-residue biases involve arginine, serine, proline and glycine, and are associated mostly with ‘mixed’, ‘polar’, ‘small’ and ‘charged’ bias classes; glutamine and asparagine biases, which are associated typically with prion-like domains, are only of middling abundance (Figs. 2B–2D). Indeed, although prion-like domains are often cited as being associated with RNA-binding proteins, in this case they only occur in ∼1 in 6 RRM-containing proteins, as judged by the PLAAC program (Lancaster et al., 2014) (Fig. S1, 32/206 (15.5%) have PLAAC LLR scores ≥ 15.0, and 35/206 (17.0%) have PLAAC PRD scores > 15.0). These PLAAC prion-like regions arise despite only moderate asparagine and glutamine frequencies and are thus substantially dependent on other residues that are common in prion-forming domains, such as tyrosine, glycine and serine, which are common biases in the RRM-containing proteins (Figs. 2B–2D). In Fig. 2E, the human BOLL ‘protein boule-like’ is presented as an example of a discontinuous CB domain around an RRM domain.

Increased precision with the –z option

As described in Methods, the –z option can be used to increase the precision of the initial scanning by the fLPS algorithm for compositional deviations (Fig. 1). Two examples of the effects of this option are illustrated (Fig. 3). Multi-protein bridging factor MBF1 is a transcriptional coactivator that promotes GCN4-dependent transcriptional activity by bridging between the DNA-binding areas of GCN4 and TATA-binding protein. The default fLPS settings detect a mild bias for positively-charged residues {KR}, which becomes a stronger bias comprised of further biasing residues {KRQGANVSP} when the ‘thorough’ option is applied. Also, a region weakly biased for polar residues {TDN} appears (Fig. 3A). These biases are likely linked to DNA and protein interactions within complexes. The second example is the Shadoo protein from human (Fig. 3B). Shadoo is a member of the prion-protein (PrP) family that has demonstrated some neuroprotective behaviour (Westaway et al., 2011). Like PrP, the protein that underlies prion diseases, it contains CB and intrinsically-disordered regions. Here, the major CB annotations are stable when –z thorough is specified, but additional mildly biased regions are detected, one of which corresponds to a signal peptide, the other an area bridging between and intrinsically disordered region and a pro-peptide (Fig. 3B).

These examples demonstrate three effects of increasing the precision of the initial compositional scanning: (i) mildly biased tracts are detected that can be quite short and that may have biological significance; (ii) further bias detail is sometimes added to CB regions, decreasing the binomial P-value; (iii) tracts with a bias made from several residues and that were previously not detectable (such as the {AWLC} tract in Fig. 1B) become evident. In aggregate, these three effects increase the ability of the program to delineate compositionally-defined domains in proteins. As shown in Table 1 detailing analysis of the S. cerevisiae proteome, a significant number of further multiple-residue CB regions are detected, even for smaller P-value thresholds (such as P ≤ 1e−09). Since one of the tracts in Fig. 1B corresponds to a signal peptide, the correspondence between signal peptide positions in the S. cerevisiae proteome and CB regions was also examined (Table 2). The number of signal peptides corresponding to CB regions increases to a highest value of ∼60% with a P ≤ 1e−03 bias P-value threshold. The results are generally in line with a previous analysis of sequence complexity in signal peptides, where 24% of residues of signal peptides in analyzed data sets were labelled part of low-complexity tracts by SEG (Wong, 2010; Wootton, 1996).

Analysis of DNA sequence

DNA sequences can be analyzed by specifying the –n option; by default, each base is expected with equal background probability. In total, there are 40 different possible biases (Fig. 4A). These can be segregated into eight bias classes for DNA (Fig. 4A). The last two of these bias classes, {ATC}—{ATG} and {ACG}—{CGT}, correspond to strand-specific depletions of single bases, i.e., {ATC}—{ATG} indicates a strand-specific lack of C or G. An example of a DNA CB region from a human promoter is illustrated (Fig. 4B). To illustrate its application to DNA, we used the fLPS program to examine bias trends in a representative data set of human promoters taken from the EPD database (Fig. 4C). Interestingly, the tri-base {ATC}—{ATG} and {ACG}—{CGT} bias classes are almost as prevalent as the two-base {AC}—{GT} bias class.

Examples of running the package

A diverse choice of parameters is possible in running the fLPS 2.0 program. Here are some examples:

1. Annotating low-complexity regions: For the specific task of annotating short CB regions of the sort termed ‘low-complexity’, the following parameters are suitable (with the yeast proteome ‘yeast.fasta’ as an example input file):

   ./fLPS –t1e−5 (or –t1e−6) –m5 –M25 –o long yeast.fasta

2. Analyzing for discontinuous CB domains: Firstly, structured domains are excised from the sequences, then fLPS is run using the –D option:

   ./DomainFilter –D excised yeast.fasta >yeast.Dexcised.fasta

   ./fLPS ”—D yeast.Dexcised.fasta

3. Restricting biases to specific sets: To analyze biases for just six-membered aromatic side-chain amino acids only (F, Y and W), using the yeast proteome background composition:

   ./CompositionMaker yeast.fasta (makes file ‘yeast.fasta.COMPOSITION’)

   ./fLPS –dv –ooneline –c yeast.fasta.COMPOSITION –r FYW yeast.fasta

   Also specified are headers and footers (–d), oneline output format (–o oneline) and verbose behaviour during runtime (–v).

4. Annotating longer biased regions with the thorough option: To find longer biased regions that have compositional skew the following options may be suitable:

   ./fLPS –z thorough –t0.001 –M 1000 yeast.fasta

5. DNA: For DNA, the ”—n option is specified:

   ./fLPS –dn DNA.example.fasta

Here, headers and footers are also output (–d option).