A benchmarking framework for comparative evaluation of low-complexity region detection tools in the human proteome Authors: Anirjit Chatterjee, Nagarjun Vijay Published: 2026 ABSTRACT Low-complexity regions (LCRs) are compositionally biased segments of proteins that play critical roles in molecular recognition, structural flexibility, and phase separation. Yet, their accurate detection remains challenging due to methodological variability among computational tools. In this study, we conducted a comprehensive benchmarking of eight widely used LCR detection methods (under multiple parameter settings) across the Homo sapiens proteome. A modular computational framework was developed to systematically compare LCR characteristics, including residue-centric analyses such as length distribution and coverage percentage. Protein-centric analyses included compositional bias, amino acid composition, and Shannon entropy. Consensus analyses revealed that regions detected by multiple tools were typically longer, more repetitive, and compositionally purer, suggesting stronger structural or functional relevance. Jaccard similarity matrices revealed distinct clustering patterns among algorithms based on shared detection principles. Additionally, entropy and purity analyses highlighted fundamental differences in sequence complexity captured by each tool. Together, these results provide a unified, reproducible framework for evaluating LCR detection performance and offer practical guidelines for reliable annotation of low-complexity regions in proteome-scale studies.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-48432-3. Introduction In recent years, low-complexity regions (LCRs) have gained recognition as significant drivers of the genome1–3 and important contributors to protein evolution and function. While the precise definition of LCRs remains a topic of debate, they are broadly accepted as sequences characterised by repetitive amino acid patterns (e.g., ‘ATLGGGGALTTT’ or ‘LKSSSQQQQSSSK’). Historically referred to as Compositionally Biased Regions (CBRs)4, Cryptic Repeats5 and Tandem Repeats6, these regions are now commonly termed LCRs. Defined by their reduced amino acid diversity, LCRs range from simple single amino acid repeats to stretches with limited residue types7–9. Found in 18–20% of human proteins and more prevalent in mammals10,11, LCRs serve as evolutionary "tuning knobs," fostering adaptation and diversity12. At a more granular level, the evolution of low-complexity regions is often driven by nucleotide-level mechanisms like replication slippage and repeat instability13. Previously overlooked due to their intrinsically disordered nature and labelled as part of the “Dark Proteome”, LCRs are now recognised for their roles in phenotypic variation and disease mechanisms. For example, the polyQ:polyA repeat ratio in RUNX2 influences vertebrate facial morphology14, while expanded polyQ tracts within LCRs contribute to neurodegenerative disorders such as Huntington’s disease (HD) and spinal and bulbar muscular atrophy (SBMA)15. Functionally, LCRs are pivotal in biological processes like cell signalling16, RNA processing17, transcription regulation12,18,19, DNA repair, and neurogenesis11. Structurally, LCRs regulate protein solubility and folding, as seen in prion-like LCRs20, and play roles in RNA binding and phase separation through motifs like RGG/RG21. They can form labile cross-β polymers and hydrogel droplets22, highlighting their versatility in molecular assembly and cellular organisation. In Plasmodium falciparum, longer proteins have been shown to be preferentially enriched in LCRs23. To advance the study of LCRs, various detection methods have been developed in recent times, each employing unique algorithms. However, the diversity of these tools necessitates benchmarking to compare and analyse their performance24. Previous efforts to evaluate and compare low-complexity region (LCR) detection methodologies have addressed multiple complementary aspects, including formalising operational definitions of LCRs25, systematically reviewing tandem repeat detection algorithms26 and developing integrative platforms such as PlaToLoCo—the first web-based meta-server that unifies outputs from diverse LCR detection methods to enable comparative visualisation and interpretation of protein low-complexity landscapes27. Recently, a database called LCRAnnotationsDB has been developed to include unified functional and structural information from a large number of publicly available protein databases28. This study evaluates eight LCR detection methods—fLPS, fLPS 2.0, AlcoR, Dotplot, SEG, LCRFinder, T-REKS and XSTREAM—selected for their availability as command-line tools and representation of diverse detection approaches. Previously, SEG was widely used for masking LCRs in BLAST searches29, while at present, fLPS and its updated version, fLPS 2.0, focus on identifying compositionally biased regions30,31. AlcoR employs bidirectional compression for mapping and visualising LCRs32, LCRFinder examines LCRs from an evolutionary perspective33, and Dotplot leverages dimensionality reduction to analyse LCR sequence space9. T-REKS and XSTREAM both use algorithms to detect the presence of Tandem Repeats within sequences34,35. This study analyses 20,610 sequences across the Homo sapiens proteome9 using these methods, offering insights into their detection capabilities. Various parameters were taken into account to analyse the differences and similarities between the detections by these tools. We also analysed the consensus of the results across the detection methods. Efforts were also made to define the boundaries of the Low-Complexity space25, using disordered regions as background and taking into account sequences detected by the method consensus. Additionally, plots of sequence complexity landscapes were analysed across tools to study the variations in complexity. Evaluation metrics were utilised to study the performance of the methods against various parameters. This research represents a comprehensive benchmarking effort in the field, aiming to encourage the development of more robust studies and methodologies for LCR analysis. Methodology Selection of LCR detection tools for benchmarking A diverse set of computational tools was selected to comprehensively benchmark the detection of low-complexity regions (LCRs) in protein sequences. The tools were grouped into three broad categories based on their underlying methodologies and usage in the literature. The first category, popular tools, included SEG and fLPS (along with its updated version, fLPS 2.0), which are widely used in LCR studies due to their established reliability and computational efficiency. The second category comprises tools designed specifically for tandem repeat detection, including XSTREAM and T-REKS, which identify repetitive motifs that often characterise certain types of LCRs. The third group encompassed novel and emerging methods, such as a custom Dotplot-based approach, LCRFinder, and AlcoR. These tools were chosen for their innovative algorithms and potential to provide complementary insights. AlcoR, in particular, utilises alignment-free techniques for LCR simulation, mapping, and visualisation. Tool selection was informed by recent literature, including methodologies described in Lee et al.9 and Silva et al.32, ensuring the inclusion of both traditional and state-of-the-art approaches for robust and comparative analysis. Tools used SEG is a computational tool developed by Wootton and Federhen (1993)29 to identify low-complexity regions (LCRs) in amino acid sequences. It operates by calculating the local sequence complexity within a sliding window using an information-theoretic measure of sequence entropy. Regions with lower-than-expected complexity—where a limited set of residues recur frequently—are masked or flagged as low-complexity. This approach helps distinguish compositionally biased or repetitive regions from more diverse segments of the sequence. SEG is widely used in sequence analysis pipelines, especially in database searches, to reduce spurious alignments caused by non-informative, repetitive regions. fLPS (fast Lowest Probability Subsequence) is a computational tool designed to detect compositionally biased or low-complexity regions (LCRs) in protein sequences with high speed and statistical rigour30. It identifies segments where the amino acid composition deviates significantly from what would be expected under a random model, using a probabilistic framework to estimate the likelihood of observing such biases. Unlike earlier tools such as SEG, fLPS emphasises statistical significance rather than purely entropy-based measures, allowing for more sensitive and precise detection of compositional biases across diverse proteomes. fLPS 2.0 is an improved version that enhances both the accuracy and efficiency of LCR detection31. It introduces refined background models and better handling of mixed or weakly biased regions, enabling more nuanced identification of low-complexity segments. The updated algorithm also supports larger datasets and provides greater flexibility in parameter tuning, making it suitable for large-scale proteome analyses and comparative studies of sequence bias. AlcoR (Algorithm for Low-Complexity Region detection) is a computational tool developed to identify low-complexity and compositionally biased regions in protein sequences through a statistically grounded, alignment-free approach32. It quantifies amino acid repetitiveness and local compositional bias by analysing the recurrence and arrangement of short sequence motifs. Unlike earlier methods such as SEG or fLPS, AlcoR integrates both compositional and positional information, allowing it to detect a wider range of repetitive or biased regions, including subtle or irregular patterns. The algorithm is optimised for large-scale proteome analyses, offering high speed, adjustable sensitivity, and compatibility with downstream characterisation of detected low-complexity regions. T-REKS (Tandem Repeats Extraction and Analysis Tool) is a computational method designed to identify tandemly repeated motifs within protein sequences34. It detects repeats by clustering identical or highly similar short subsequences (k-mers) and refining their alignment to define repeat boundaries and consensus motifs. T-REKS combines sequence similarity searches with statistical evaluation to distinguish true tandem repeats from random similarities, allowing accurate detection of both perfect and imperfect repeats. The tool is particularly effective in analysing proteins with modular or repetitive architectures, such as surface proteins and structural proteins and identifies short tandem repeats which are part of low-complexity domains, and is frequently used in large-scale proteome analyses to study sequence repetitiveness and evolutionary divergence. XSTREAM (eXtendedSTRing Repeat detection program) is a versatile algorithm developed to detect and characterise repetitive sequence motifs in proteins and nucleic acids35. It identifies both tandem and interspersed repeats, accommodating imperfect, variable-length, and nested repeat structures. XSTREAM uses pattern matching combined with flexible scoring and alignment parameters to recognise repeats that may evolve through mutation, insertion, or deletion. The tool reports detailed information about each repeat family, including consensus motifs, copy number, and sequence divergence. Its strength lies in its ability to detect complex and degenerate repeat architectures that are often missed by simpler repeat-finding algorithms, making it valuable for studying protein domains, low-complexity regions, and the evolutionary dynamics of repetitive sequences. Dotplot analysis is a graphical method used to visualise sequence self-similarity and repetitive patterns within proteins or nucleic acids. In this approach, a sequence is compared against itself (self-dotplot) or another sequence by plotting matching residues as dots on a two-dimensional grid—one axis representing the query sequence and the other the target9. Continuous diagonal lines indicate regions of similarity or internal repeats, while scattered dots suggest randomness or compositional variation. When applied to proteins, dotplots effectively reveal low-complexity regions (LCRs), tandem repeats, and internal duplications that might not be easily detected by linear algorithms. This method provides an intuitive, alignment-free visualisation of repetitive organisation and has become a fundamental step in studying sequence complexity and structural modularity across proteomes. LCRFinder is a computational method developed to identify low-complexity regions (LCRs) in protein sequences based on local amino acid composition bias and sequence repetitiveness33. It employs a statistical framework that scans protein sequences to detect short segments where specific amino acids or motifs occur more frequently than expected under a random distribution. Unlike earlier tools that rely primarily on entropy or window-based masking, LCRFinder integrates compositional enrichment analysis with pattern clustering, allowing it to sensitively detect both simple and complex forms of low-complexity. Selection of the dataset For this study, the human proteome was obtained from the UniProt database, ensuring a high-quality and comprehensive dataset for downstream analyses. A multi-protein FASTA file containing 20,610 protein sequences was selected. To maintain dataset integrity and biological relevance, only one representative protein per gene was retained, excluding all isoforms. This rigorous filtering ensured minimal redundancy and high annotation quality, providing a robust foundation for the analysis of low-complexity regions (LCRs). The dataset selection strategy follows the guidelines established by Lee et al.9. Planning the methodology The proteome was analysed for LCR detection by the selected tools. Each tool gave outputs in separate formats. To ensure a uniform framework for benchmarking diverse LCR detection tools, all tool outputs were standardised into a common BED-like format (1-based), using custom Python scripts. This modification allowed direct comparison of predicted low-complexity regions (LCRs) across tools, independent of their native output formats. Each BED entry captured the essential parameters: protein identifier, start and end positions of the detected LCR. Data aggregation, processing, and visualisation To systematically evaluate and compare the performance of multiple low-complexity region (LCR) detection tools, categorised output files were collected across five major analytical dimensions: LCR length distribution, LCR coverage percentage, LCR counts per protein, amino acid composition, and Shannon entropy diversity. Each dataset was generated using distinct LCR detection tools and operating modes and stored as tab-separated values (TSV) files. Custom Python scripts (in GitHub repository) were used to process the raw outputs from each tool. These scripts calculated LCR lengths from start and end coordinates, determined the percentage of each protein sequence covered by LCRs, quantified amino acid compositions, and computed Shannon entropy values to evaluate compositional diversity. For comparative analyses, LCRs were assigned to fixed length bins (0–10, 10–20, 20–50, 50–100, 100–200, and > 200 amino acids), while LCR coverage percentages were categorized into 0–20%, 20–40%, 40–60%, 60–80%, and 80–100% bins. LCR counts per protein were grouped into 0, 1–5, 5–10, 10–15, and 15+ categories. Amino acid proportions were normalised to total residue counts, and entropy distributions were computed to assess the complexity of the detected regions. All categorised datasets were subsequently processed and visualised in R (version ≥ 4.0.0)36 using the tidyverse37, ggplot238, viridis39, reshape240, dplyr41, and tidyr42 packages. Data from all analyses were harmonised by extracting tool identifiers from filenames using regular expressions, merging data frames iteratively, and converting categorical variables into ordered factors to ensure consistent axis ordering across plots. Visualisation was performed using ggplot2 and complementary packages including ggpattern, Polychrome, and patchwork. Heatmaps were generated to display LCR length and coverage distributions, where colour gradients represented frequency intensities. Stacked bar plots illustrated LCR counts per protein and amino acid compositions across tools, while boxplots summarised the Shannon entropy distributions. Additional composite visualisations integrated motif distribution, entropy variation, and sequence purity metrics, enabling a comprehensive, multi-dimensional comparison of the sensitivity, selectivity, and compositional characteristics of the evaluated LCR detection methods. Analysis of common LCR properties across detection tools To characterise the shared properties of low-complexity regions (LCRs) identified by different computational tools, a comparative overlap analysis was performed using BEDTools (v2.30.0)43. The bedtools multiinter command was employed to identify protein coordinate overlaps of LCRs detected across multiple methods, followed by stratification of these intersected regions based on the number of tools (along with their configurations) detecting them (ranging from 1 to 13). Using the bedtools getfasta command, amino acid sequences corresponding to each intersected region were extracted from the Homo sapiens reference proteome FASTA file for downstream compositional and motif analyses. Custom Python scripts were used to determine the most frequently occurring peptide motifs within these shared regions. Each region was annotated with its dominant motif and classified into one of four categories—Monopeptide, Dipeptide, Tripeptide, or Other—based on regular expression–based pattern matching (peptide_count.py). The most abundant peptide repeats were summarised, and their proportions were visualised in R (version ≥ 4.0.0) using a custom graphical workflow implemented with the ggplot2, ggpattern, dplyr, Polychrome44, and patchwork45 packages. The visualisation pipeline consisted of three main components:Motif Composition Plot—A stacked bar plot displaying the relative abundance of dominant peptide motifs across regions detected by varying numbers of tools. Distinct fill patterns (solid, stripe, crosshatch) were used to differentiate monopeptides, dipeptides, and tripeptides, respectively.Entropy Analysis—Shannon’s Diversity Index was calculated for each region to quantify compositional heterogeneity. Boxplots were generated using color-blind–friendly palettes to depict entropy distributions across consensus levels, with outliers excluded for clarity.Purity Analysis—Purity is defined as the ratio of the number of residues of the primary amino acid to the total number of residues in the region46. Purity was computed to measure the dominance of single amino acids within regions. Corresponding boxplots were produced in parallel with the entropy distributions for direct comparison. Motif Composition Plot—A stacked bar plot displaying the relative abundance of dominant peptide motifs across regions detected by varying numbers of tools. Distinct fill patterns (solid, stripe, crosshatch) were used to differentiate monopeptides, dipeptides, and tripeptides, respectively. Motif Composition Plot—A stacked bar plot displaying the relative abundance of dominant peptide motifs across regions detected by varying numbers of tools. Distinct fill patterns (solid, stripe, crosshatch) were used to differentiate monopeptides, dipeptides, and tripeptides, respectively. Entropy Analysis—Shannon’s Diversity Index was calculated for each region to quantify compositional heterogeneity. Boxplots were generated using color-blind–friendly palettes to depict entropy distributions across consensus levels, with outliers excluded for clarity. Entropy Analysis—Shannon’s Diversity Index was calculated for each region to quantify compositional heterogeneity. Boxplots were generated using color-blind–friendly palettes to depict entropy distributions across consensus levels, with outliers excluded for clarity. Purity Analysis—Purity is defined as the ratio of the number of residues of the primary amino acid to the total number of residues in the region46. Purity was computed to measure the dominance of single amino acids within regions. Corresponding boxplots were produced in parallel with the entropy distributions for direct comparison. Purity Analysis—Purity is defined as the ratio of the number of residues of the primary amino acid to the total number of residues in the region46. Purity was computed to measure the dominance of single amino acids within regions. Corresponding boxplots were produced in parallel with the entropy distributions for direct comparison. All three visual components were combined into a composite multi-panel figure using the patchwork package, enabling simultaneous interpretation of motif proportions, entropy variability, and compositional purity across increasing tool consensus levels. Together, these analyses provided a unified view of how increasing agreement among LCR detection tools correlates with compositional simplicity, reduced entropy, and enrichment for highly repetitive peptide motifs, thus revealing the convergent tendencies of algorithms toward strongly biased low-complexity regions. Visualisation of low-complexity (LC) diagrams According to Mier et. al.25, the low-complexity (LC) diagram represents protein sequence regions in a two-dimensional space defined by compositional bias (y-axis; percentage of the most frequent amino acid) and repetitiveness (x-axis; mutation distance from a perfect repeat). Perfect repeats lie at low x-values (high periodicity), while homorepeats occupy the top of the diagram due to maximal compositional bias. Unbiased, typical protein regions cluster in the bottom-right, characterised by low compositional bias and low repetitiveness. Intermediate sequences fall between these extremes, reflecting varying degrees of repetition and amino acid dominance. The diagram thus provides a conceptual framework to distinguish low-complexity regions from structured, compositionally balanced sequences. LC regions identified at different consensus levels (k = number of methods) were aggregated across output files following numeric validation and removal of incomplete entries. Kernel density estimation was applied to generate two-dimensional contour representations of LC distributions, allowing direct comparison of density patterns across consensus tiers. Contours were coloured by consensus level, with line widths scaled to reflect tier order. All density plots were rendered using fixed 0–100% axes to ensure consistent scaling across methods. (plot1.r). To quantify shifts in central tendency and dispersion across consensus tiers, median mutation percentage and compositional dominance were computed for each tier, together with interquartile ranges (IQRs). These summary statistics were visualised as centroid points with orthogonal IQR crossbars. Centroids corresponding to successive consensus tiers were connected by directional arrows to illustrate systematic transitions in LC composition with increasing consensus stringency. This representation captures both variability and directional trends while remaining robust to outliers. (plot2.r). To assess cumulative occupancy of LC space, mutation percentage and compositional dominance values were discretised into fixed 2% × 2% bins. For each bin, the highest consensus tier observed among LC regions falling within that bin was recorded. This information was visualised as a tile-based heatmap, where colour intensity reflects the maximum level of cross-method support present in each region of sequence-complexity space. Bin geometry, coordinate limits, and aspect ratios were held constant to preserve spatial interpretability. (plot3.r). Finally, tier-specific LC density patterns were examined by generating independent binned density maps for each consensus tier using identical bin sizes and axis limits. Within each tier, the number of LC regions per bin was counted and visualised as faceted heatmaps. A square-root transformation was applied to bin counts to stabilise variance and improve visual contrast between sparse and dense regions. This small-multiple approach enables direct comparison of how LC landscape occupancy changes as consensus stringency increases. (plot4.r). Visualisation of entity retention across purity levels To assess and compare the performance of multiple computational methods in retaining entities across varying purity thresholds, a custom workflow was implemented in R (version ≥ 4.0.0) using the ggplot2, dplyr, tidyr, and Polychrome packages. Tab-delimited data containing the proportion of entities retained by each method at different purity levels were imported and grouped by detection method. For each method, line plots modelled the relationship between purity level and entity retention. A distinct colour palette was generated using the Polychrome package to ensure clear visual distinction among methods. The processed data were visualised using ggplot2 as a line plot, with purity level on the x-axis and the proportion of entities retained on the y-axis. Each method was represented by a unique colour-coded line and corresponding points. Jaccard similarity and overlap analysis between LCR detection tools To evaluate and compare the degree of overlap and similarity among various computational methods for low-complexity region (LCR) detection, a correlation-style heatmap was constructed using custom R scripts. All analyses were conducted in R (version ≥ 4.0.0) with the corrplot47, tidyverse and reshape2 packages for data manipulation and visualisation. A Jaccard similarity matrix was derived from actual LCR coordinate overlaps encompassing eight methods, including AlcoR (Mode 1 and Mode 2), SEG (default, intermediate, and strict modes), fLPS (default and strict), fLPS 2.0 (default and strict), Dotplot, LCRFinder, T-REKS, and XSTREAM. Pairwise Jaccard coefficients were computed using the bedtools jaccard command, and the resulting similarity file (human_jaccard.tsv) was imported into R, formatted as a numeric matrix, and harmonised with the overlap dataset. A figure was then produced that depicted the Jaccard similarity matrix with a diverging white–red colour scheme to represent low-to-high similarity levels. Estimating the boundaries of low-complexity (LC) sequence space Low-complexity sequence space has been defined with transient boundaries25, but is not defined by a single physical rule but emerges from reduced sequence diversity, compositional bias, and tolerance to mutation. To ground this abstract concept biologically, we framed LC space using experimentally supported disorder as a reference. We used DisProt48–50 annotations, which represent experimentally validated intrinsically disordered regions, and missing residues in PDB structures, which indicate regions that fail to adopt stable conformations in structural experiments. These datasets capture functional and structural manifestations of disorder, respectively, and provide biologically meaningful anchors for interpreting low-complexity sequence patterns. Such disordered regions served as background. Each sequence segment was represented using two intuitive features: mutation percentage, which reflects how freely a region can change without disrupting its role, and percentage of the most frequent amino acid, which reflects how compositionally biased the sequence is. Together, these features summarise how information-rich or information-poor a sequence region is. In the first approach, the empirical LC boundary was estimated using binning (plot_LC_boundary_1.r). LC space was discretised into a regular two-dimensional grid defined over mutation percentage (x-axis) and amino-acid dominance (y-axis), using bins of width Δ = 2%. Each sequence segment with coordinates (x,y) was assigned to a bin\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{{{\mathrm{bin}}}} = \left\lfloor {x/\Delta } \right\rfloor,y_{{{\mathrm{bin}}}} = \left\lfloor {y/\Delta } \right\rfloor$$\end{document} For each bin (i,j), we counted the number of LC segments nij+ and background segments nij−​. The probability that a segment in that bin belongs to an LC region was estimated using a smoothed frequency ratio:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P\left( {{\mathrm{LC}}\,|\, i, \, j} \right) \, = \, \left( {n_{ij}^{ + } + a} \right) \, / \, \left( {n_{ij}^{ + } + n_{ij}^{ - } + { 2}a} \right)$$\end{document}with a = 1. This small correction prevents unstable estimates in sparsely populated bins and reflects prior uncertainty rather than assuming zero probability. To evaluate how well these probabilities separate LC from non-LC segments, we performed receiver operating characteristic (ROC) analysis on held-out data (actual coordinates of the IDRs and the missing residues). For a probability threshold t, predictions were defined as LC if P ≥ t. The true-positive rate (TPR) and false-positive rate (FPR) were computed as\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{TPR}}(t) \, = {\text{ TP }}/ \, \left( {{\text{TP }} + {\text{ FN}}} \right)$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{FPR}}(t) \, = {\text{ FP }}/ \, \left( {{\text{FP }} + {\text{ TN}}} \right)$$\end{document} The optimal threshold t* was selected by maximising Youden’s J statistic,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$J\left( t \right) = {\text{ TPR}}\left( t \right) \, - {\text{ FPR}}\left( t \right),$$\end{document}which balances sensitivity and specificity. The LC boundary was then defined as the contour in feature space where P(LC) = t*. This contour represents a confidence boundary, not a strict biological delimiter. The second approach estimates LC boundaries using a smooth probabilistic surface rather than discrete bins using a generalised additive model (GAM) (plot_LC_boundary_2.r). We modelled the probability that a segment belongs to an LC region as a continuous function of mutation percentage x and amino-acid dominance y:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y_{i}\sim {\mathrm{Bernoulli}}\left( {p_{i} } \right),{\text{ log}}\left( {p_{i} /\left( {{1 } - p_{i} } \right)} \right) \, = f\left( {x_{i} , \, y_{i} } \right),$$\end{document}where Yi indicates LC status and f(x,y) is a smooth two-dimensional function learned from the data. This function was represented using penalised regression splines, which allow the surface to adapt to data while discouraging unnecessary complexity. Smoothness penalties ensure that the inferred LC probability varies gradually across sequence space, consistent with biological expectations of gradual transitions rather than sharp cutoffs. Predicted probabilities \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{p}_i$$\end{document} were evaluated using ROC analysis as above, and an optimal threshold t* was selected via Youden’s J statistic. To visualise the LC boundary, probabilities were evaluated over a regular grid covering the domain [0,100] × [0,100], yielding a posterior probability surface\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{p}\left( {x,y} \right) \, = P\left( {{\mathrm{LC}}|x,y} \right).$$\end{document} The LC boundary was extracted as the contour where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{p}$$\end{document}(x,y) = t*. Together, these complementary approaches show that LC regions do not occupy a sharply defined territory in sequence space. Instead, LC space emerges as a probability envelope, shaped by entropy reduction, compositional bias, and biological disorder constraints. DisProt annotations and missing PDB residues provide biological grounding, while probabilistic modelling reveals that confidence in low-complexity increases gradually rather than abruptly. The explicit mathematical formulation underscores that LC boundaries are inherently fuzzy and context-dependent, reflecting a continuum of sequence information rather than a binary state. Visualisation of the most frequent amino acid and mutation percentage To explore the interplay between amino acid composition, periodicity, and structure in protein regions, a two-dimensional (2D) diagram plotting compositional bias (y-axis) versus repetitiveness (x-axis) was constructed, according to the model presented by Lee et al. 20229. The analysis has focused on sequence regions of 10–50 residues, a length optimal for balancing statistical robustness and biological relevance. For each region, two complexity metrics were computed: (1) the frequency of the most abundant amino acid as a proxy for compositional bias, and (2) the proportion of residues that must be mutated to convert the region into a perfect repeat, representing repetitiveness. Sample sequences were analysed to demonstrate these relationships, such as ACDEFEGEIE, where the highest frequency amino acid (E, 40%) sets y, and periodicity is estimated by calculating minimal mutational distance to a perfect repeat (e.g., EEEEEEEEEE, 60%). The complexity triangle represents a continuous sequence landscape defined by amino acid compositional dominance and mutational distance to perfect repetition. Sequences with uniform amino acid usage (e.g., globular proteins) cluster at the lower-right of the diagram, while perfect and homorepeat sequences occupy the left side. Intermediate sequences (e.g., direpeats, tripeats) span between these extremes. Utilising this model, we mapped selected proteins with low-complexity regions (LCRs), and unbiased regions into the diagram to visualise how their sequence characteristics relate to their potential structural tendencies and LCR classification. The two-dimensional diagram was broadly divided into 3 segments:LCR segment (X < 50 and Y > 50)LTR segment (X < 50 and Y < 50)HCR segment (X > 50 and Y < 50) LCR segment (X < 50 and Y > 50) LCR segment (X < 50 and Y > 50) LTR segment (X < 50 and Y < 50) LTR segment (X < 50 and Y < 50) HCR segment (X > 50 and Y < 50) HCR segment (X > 50 and Y < 50) The exact boundaries between regions of various complexity are difficult and ambiguous to identify (as attempted in the above subsection). For operational classification, this space was subdivided into low-complexity (LCR), long tandem repeats (LTR), and high-complexity (HCR) regions using 50% thresholds along both axes. These thresholds were chosen for interpretability rather than as strict biological cutoffs. A dominance value of 50% separates sequences primarily governed by a small subset of residues from those with more evenly distributed amino acid usage, consistent with classical entropy- and composition-based definitions of low-complexity51. Similarly, a mutation requirement of 50% distinguishes sequences that are close to repetitive or biased patterns from those requiring extensive substitutions to reach such states, aligning with prior treatments of repeat degeneracy and compositional bias52,53. It is to be emphasised that these boundaries are heuristic and intended for comparative contextualisation rather than strict biological classification. A Python-based computational pipeline was developed to evaluate the compositional complexity of low-complexity regions (LCRs) in protein sequences and to detect them using a sliding-window approach. The workflow accepts a protein FASTA file and a corresponding BED file as input to extract LCR coordinates, employing a sequence window size of 20 amino acids with a sliding window of 10 amino acids (using seq_parser.py). For each extracted region, Shannon entropy was calculated to quantify sequence diversity, while amino acid frequencies were analysed to identify the most abundant residue and its relative contribution. Repetitive sequence motifs (k-mers) of varying lengths were detected, and the minimal mutation percentage required to achieve perfect periodicity was estimated. The resulting LCRs were visualised as density plots with repetitiveness and compositional bias represented on the x- and y-axes, respectively (using complexity_plot.py), and all computed parameters—including entropy, dominant residue, k-mer composition, and mutation metrics—were compiled into a structured tab-delimited output file. Annotating “positives” and “negatives” To quantify the agreement between low-complexity regions (LCRs) detected by different methods, we computed overlap-based metrics using a residue-level comparison framework. Reference LCRs were obtained from the resulting sequences generated from the pipeline in the above subsection (only those regions which had Mutation % < 50 and Most Frequent Amino Acid % > 50, were annotated as LCRs), whereas predicted LCRs were taken from the result bed files of the various methods used. Importantly, this reference set does not represent experimentally validated ground truth, but rather a consistent internal definition used for comparative benchmarking. All protein sequences were treated as linear residue coordinates. Each residue was assigned a binary label indicating whether it belonged to an LCR according to the prediction set, the reference set, or both. A residue was considered “positive” if it fell within an annotated LCR interval and “negative” otherwise. Overlap categories were defined as follows: residues identified as LCRs by both the reference definition and a given method were considered overlapping detections (“true positive” for this study); residues identified only by the reference definition were considered reference-specific (“false negative” for this study); residues identified only by the method were considered method-specific (“false positive” for this study); and residues identified by neither were considered background (“true negative” for this study) (Table 1). A custom bash script (conf_mat.sh) was used to perform this comparison.Table 1Definition of classification categories used for evaluating LCR detection methods. Regions are classified based on their agreement with the operational definition of low-complexity regions (LCRs) proposed by Mier et al. and their identification by a given method. True positives (TP) represent regions defined as LCRs and correctly detected by the method, false positives (FP) are non-LCR regions incorrectly identified as LCRs, false negatives (FN) are defined LCRs missed by the method, and true negatives (TN) are non-LCR regions correctly not detected.CategoryAs per the definition of Mier et al.25Detected by methodTrue positive (TP)YesYesFalse positive (FP)NoYesFalse negative (FN)YesNoTrue negative (TN)NoNo Table 1Definition of classification categories used for evaluating LCR detection methods. Regions are classified based on their agreement with the operational definition of low-complexity regions (LCRs) proposed by Mier et al. and their identification by a given method. True positives (TP) represent regions defined as LCRs and correctly detected by the method, false positives (FP) are non-LCR regions incorrectly identified as LCRs, false negatives (FN) are defined LCRs missed by the method, and true negatives (TN) are non-LCR regions correctly not detected.CategoryAs per the definition of Mier et al.25Detected by methodTrue positive (TP)YesYesFalse positive (FP)NoYesFalse negative (FN)YesNoTrue negative (TN)NoNo Definition of classification categories used for evaluating LCR detection methods. Regions are classified based on their agreement with the operational definition of low-complexity regions (LCRs) proposed by Mier et al. and their identification by a given method. True positives (TP) represent regions defined as LCRs and correctly detected by the method, false positives (FP) are non-LCR regions incorrectly identified as LCRs, false negatives (FN) are defined LCRs missed by the method, and true negatives (TN) are non-LCR regions correctly not detected. These categories were subsequently used to compute true positive rate (TPR) and false positive rate (FPR) as operational measures of agreement between methods and the reference definition. It is important to emphasise that the reference LCR set used in this study is based on an operational definition of low sequence complexity derived from compositional constraints, consistent with frameworks such as that of Mier et al.25. As such, the terms “true positive” and “false positive” are used here strictly in relation to agreement with this definition and do not imply biological validity or functional relevance. A region may satisfy criteria for low-complexity without necessarily having a defined biological role. These metrics should therefore be interpreted as relative measures of concordance rather than indicators of biological validity. Visualisation of tool performance across multiple LCR-associated parameters To comprehensively evaluate the performance of various low-complexity region (LCR) detection tools across multiple quantitative biological parameters, we developed a modular visualisation pipeline in R (v4.0.0 or later). The script incorporated the ggplot2, dplyr, tidyr, cowplot54, patchwork, and readr55 libraries to generate comparative line plots for both true positive rate (TPR) and false positive rate (FPR) across distinct feature categories, which were defined as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text{TPR }} = {\text{ TP }}/ \, \left( {{\text{TP }} + {\text{ FN}}} \right)$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text{FPR }} = {\text{ FP }}/ \, \left( {{\text{FP }} + {\text{ TN}}} \right)$$\end{document} Four key LCR-related parameters were analysed: gene length, LCR number per gene, percentage of gene length covered by LCRs (LCR coverage), and the entropy ratio between LCRs and entire genes. For each parameter, precomputed TPR and FPR values for 8 different computational tools (with 13 parameter settings) were loaded from tab-delimited files and visualised using consistent formatting. Gene length categories (1–10) [sorted by length and divided into 10 bins of approximately equal size], LCR number (1–6), and LCR coverage per cent bins (5%, 10%, 15%, 20%, > 20%) were treated as ordered categorical variables. Similarly, LCR: Gene entropy ratios (0.2–1.0) were modelled as ordered factors. All plots employed a shared colour palette to maintain consistent tool representation. A common minimalist theme was defined and applied across all figures to ensure visual uniformity and clarity. For each parameter, two line plots were generated—one for TPR and one for FPR—where tool performance was plotted against the relevant parameter category. The eight individual plots were grouped into four panels (TPR and FPR for each parameter) and arranged in a 2 × 2 grid using the patchwork and cowplot packages. Quantitative comparison of LCR detection tools based on mean TPR and FPR across multiple parameters To summarise the overall performance of 8 computational methods (with 13 modes) in detecting low-complexity regions (LCRs), we calculated and visualised the mean true positive rate (TPR) and false positive rate (FPR) under four different biological stratifications: gene length, number of LCRs per gene, LCR coverage percentage, and entropy ratio between LCR and gene. All analyses and visualisations were performed in R (v4.0.0 or later) using the packages ggplot2, dplyr, tidyr, and patchwork. Preprocessed tab-delimited data files were read from respective directories corresponding to each stratification. Within each dataset, performance metrics (TPR and FPR) were grouped by tool, and the mean values were calculated. Each parameter yielded a pair of grouped bars per tool—one for mean TPR and one for mean FPR—to visually separate the metrics. The tools were ordered along the x-axis by descending value for each plot to enhance interpretability. Custom axis themes and font sizes were applied through additional formatting. Quantification and visualisation of structurally disordered regions To project and compare the structurally disordered regions in the Low-Complexity sequence space, the entire dataset of experimentally verified disordered regions was selected from DisProt48 and those regions corresponding to Homo sapiens were selected (along with the coordinates of the disordered regions) into a bed file. These were further classified based on the experimental methods used to study the disordered regions. Based on the complexity triangle, each region was plotted into the Low-Complexity sequence space, with each method being represented separately by distinctly coloured points. Furthermore, protein regions lacking atomic coordinates in experimentally determined structures were used as structural evidence consistent with conformational disorder. Missing-residue annotations were obtained from UniProt–PDB residue mapping files by identifying residues present in the UniProt canonical sequence but absent from corresponding PDB atomic coordinates. Missing residues were interpreted as conservative structural indicators compatible with disorder, acknowledging alternative explanations such as experimental limitations, crystal packing effects, or transient structural states. Similar to the previous method, these segments were plotted into the Low-Complexity sequence space as a heatmap. Results LCR detection methods are different from each other Variation in LCR length distribution Figure 1A demonstrates unique detection patterns influenced by the design of each tool. This analysis is residue-centric, focusing on the length distribution of individual low-complexity segments irrespective of their protein context. The fLPS clan (fLPS, fLPS strict, fLPS 2.0, fLPS 2.0 strict) detects the largest number of LCRs, particularly in larger categories (50–200 + amino acids), with over 60,000 counts in the 200 + category. This is consistent with fLPS’s emphasis on detecting large compositional biases. Conversely, AlcoR (Mode 1 and Mode 2) is less selective, focusing on LCRs of lengths 10–50 residues, as expected with its statistical bias toward local significance. Dotplot also focuses on shorter LCRs, but prefers 10–50 residue regions associated with self-similarity over extended bias. SEG family (SEG, SEG intermediate, SEG strict) are biased toward finding extremely short regions (0–20 residues), in accordance with its entropy-based methodology for masking very uniform segments. LCRFinder takes this even further, with the majority of ultra-short motifs less than 10 residues. In contrast, T-REKS and XSTREAM, designed for periodic tandem repeats, identify fewer LCRs, primarily between 10–100 residues. In general, the distribution profiles demonstrate how algorithmic variation affects detection.Fig. 1Comparative overview of Low-Complexity Region (LCR) detection across multiple methods. (A) Shows a heatmap of counts across LCR length bins detected by each method across the proteome, with the y-axis listing the methods and the x-axis showing the LCR length bins [in bins of 0–10, 10–20, 20–50, 50–100, 100–200 and above 200]. (B) Represents a heatmap of counts in percentage coverage bins (LCR coverage of the total gene length) [in bins of 0–20, 20–40, 40–60, 60–80 and 80–100 per cent]. (C) Shows a stacked barplot which compares the number of LCRs detected per protein by each method [in categories of 0, 1–5, 5–10, 10–15 and above 15]. (D) depicts a stacked barplot of the overall amino acid composition of LCRs detected by each method. (E) Shows boxplot which compare the overall Shannon Entropy of LCRs detected by each method. Comparative overview of Low-Complexity Region (LCR) detection across multiple methods. (A) Shows a heatmap of counts across LCR length bins detected by each method across the proteome, with the y-axis listing the methods and the x-axis showing the LCR length bins [in bins of 0–10, 10–20, 20–50, 50–100, 100–200 and above 200]. (B) Represents a heatmap of counts in percentage coverage bins (LCR coverage of the total gene length) [in bins of 0–20, 20–40, 40–60, 60–80 and 80–100 per cent]. (C) Shows a stacked barplot which compares the number of LCRs detected per protein by each method [in categories of 0, 1–5, 5–10, 10–15 and above 15]. (D) depicts a stacked barplot of the overall amino acid composition of LCRs detected by each method. (E) Shows boxplot which compare the overall Shannon Entropy of LCRs detected by each method. Variation in LCR coverage Figure 1B reveals clear patterns among various LCR detection tools based on the fraction of residues within a protein annotated as low-complexity, providing a residue-centric view normalised to gene length. The fLPS set of tools (fLPS, fLPS strict, fLPS 2.0, and fLPS 2.0 strict) strongly identifies numerous proteins with 80–100% LCR coverage, indicating that these algorithms are very sensitive to long compositional biases and tend to mark large or whole proteins as low-complexity. Whereas programs such as AlcoR (Mode 1 and Mode 2) and Dotplot primarily find LCRs spanning merely 0–40% of protein sequences, the reason is the conservative nature that targets local high-scoring, low-complexity regions more than extended domains. The SEG family demonstrates mixed behaviour: SEG and SEG strict preferentially identify short LCRs (0–20% coverage), whereas SEG intermediate identifies proteins in the 20–40% range as a result of its loose entropy threshold. Similarly, LCRFinder prefers finding LCRs that span less than 40% of the protein, in accordance with its emphasis on small, repetitive motifs. T-REKS and XSTREAM also prefer to identify structured repeats, primarily in the 0–20% coverage bin. Tools either prefer broad coverage, such as fLPS, or localised identification, such as SEG, AlcoR, and Dotplot. Variation in LCR counts To complement residue-level analyses, we next examined LCR counts per gene, a protein-centric metric that captures how frequently low-complexity regions occur within individual proteins, independent of their total length or coverage. The outcomes were highly variable among methods, as seen in Fig. 1C. Some tools, like fLPS and fLPS 2.0, including their “strict” variants, found numerous genes with very high LCR counts, most notably in the 10–15 and 15 + bins. This indicates high sensitivity toward LCR-abundant genes but simultaneously brings the prospect of higher false positives. Dotplot exhibited modest capacity to detect multiple LCR-bearing genes, representing high detection and probably higher specificity than the fLPS-based approaches. SEG identified fewer numbers of highly LCR abundant genes, rather showing well-distributed cases among the 1–5 and 5–10 ranges, representing a more cautious and sensitive detection pattern. SEG strict and SEG intermediate were more restrictive, with fewer LCR detections in total. LCRFinder and T-REKS had intermediate sensitivity, primarily detecting genes with 1–5 LCRs. AlcoR (both modes) and XSTREAM were the most conservative, with the majority of genes being assigned to “0 LCRs” or "1–5 LCRs," suggesting low sensitivity. Overall, these results align with previous findings: very sensitive methods such as fLPS variants identify multiple LCRs per gene, whereas Dotplot, SEG, AlcoR, and XSTREAM emphasise specificity. Amino acid proportion variability To determine the amino acid composition biases intrinsic to various LCR detection approaches, we compared the relative amino acid compositions of predicted LCRs from all tools, as seen in Fig. 1D. Extensive compositional profile differences were noted, mirroring each approach’s inherent detection strategy and selectivity. AlcoR (both Mode 1 and Mode 2) identified LCRs most heavily enriched in a limited set of amino acids, specifically lysine (K) [9%] and leucine (L) [19%], showing an apparently strong preference for highly biased homopolymeric or oligomeric areas. XSTREAM was heavily enriched for lysine (K) [42%] and serine (S) [14%]. Likewise, SEG strict also had a highly discriminatory profile, with LCRs most heavily dominated by proline (P) [17%] and glutamic acid (E) (16%) in line with its highly stringent entropy thresholds aimed at isolating the most extreme compositional simplicity areas. Conversely, fLPS, fLPS strict, and the fLPS 2.0 variants presented wider distributions of amino acids, which included a greater diversity of residues like serine (S) [14%], glutamic acid (E) [11%], leucine (L) [21%] (for fLPS and fLPS 2.0) and proline (P) [18%] (for the stricter variants). The trend indicates an ability to identify more diverse LCR types, including flexible and structurally disordered regions. Dotplot also showed relatively balanced composition, though being moderately enriched with small, compact residues such as L [14%], S [15%], and P [15%], consistent with their sensitivity for disorder-prone, composition-biased motifs. SEG (intermediate and default mode) showed enrichment for P [13%], E [11%], and G [8%]. LCRFinder and T-REKS exhibited an intermediate profile, preferring a lesser, though more heterogeneous, range of amino acids than the heavily selective approaches. In general, these findings suggest that although very strict tools like AlcoR and XSTREAM are very good at identifying sharply biased, straightforward LCRs, more general-purpose tools like fLPS variants and Dotplot identify a greater variety of LCR types. Shannon entropy variation Figure 1E shows the compositional complexity of the putative low-complexity regions (LCRs) by comparing their Shannon entropy distributions. For the standard 20–amino-acid alphabet, the theoretical maximum Shannon entropy is log₂(20) ≈ 4.32, corresponding to uniform amino acid usage; lower values indicate increasing compositional bias. Higher entropy indicates greater diversity of amino acid usage in LCRs, whereas lower entropy indicates areas characterised by a small subset of residues. Interestingly, AlcoR (Mode 1 and Mode 2) and XSTREAM had high median entropy values ranging from 3.0 to 3.8. This indicates that these algorithms are quite permissive, frequently detecting LCRs with significant amino acid variation even though they are labelled as low-complexity, thus covering a larger range of sequence types. Dotplot and SEG strict, however, showed much lower median entropy values, around 1.5 to 2.0, reflecting strong selection bias towards highly compositionally simple regions composed of a limited number of amino acids. SEG default and intermediate modes, and LCRFinder, exhibited intermediate entropy profiles, finding a combination of strongly biased and moderately heterogeneous LCRs. The fLPS suite of tools (fLPS, fLPS strict, fLPS 2.0, and fLPS 2.0 strict) disclosed wide entropy distributions with comparatively high medians (3.0–3.5), indicative of their design sensitivity to both narrowly and widely biased low-complexity sequences. T-REKS had moderate entropy values and strong variation, reflecting flexible detection by sequence periodicity. Overall, these results establish that SEG strict and Dotplot are very stringent and prefer plain compositional patterns, while AlcoR, XSTREAM, and the fLPS variants use more general definitions of low-complexity, covering a larger set of sequence complexities. The detection consensus of the independent methods hints towards regions of homorepeats We conducted a comprehensive analysis of low-complexity regions (LCRs) by examining how the compositional characteristics of the regions commonly detected across tools varied with the number of independent computational tools that identified them, thereby assessing the relationship between detection consensus and sequence properties (Fig. 2). This analysis revealed a striking continuum in both sequence diversity and complexity. LCRs identified by one or more limited tools (in most cases 1–3) displayed extensive heterogeneity of amino acid composition, with often a wide variety of residues and intricate motif types. Such regions had higher Shannon entropy and lower purity values of sequences, which correspond to no strong residue usage and a larger level of randomness or compositional richness. Such features indicate that these LCRs reside outside the canonical definition of low-complexity, possibly including context-specific or structure-sensitive sequences. The more tools identified a given LCR, the more significant the shift in compositional features became. The incidence of simple, compositionally skewed motifs—especially homopolymers like polyalanine (A), polyleucine (L), and polyserine (S), and simple dipeptide and tripeptide repeats like AA, SS, and GG—increased markedly. LCRs consistently detected by 10 or more tools had overall low entropy and highly pure values (median values near 0.8–1.0), indicating a strong enrichment for homorepetitive and compositionally extreme sequences. This pattern was particularly widespread in areas identified by all 8 methods (with 13 modes), which predominantly represented typical LCRs with little sequence complexity. The convergent identification of such areas probably owes to a mix of algorithmic bias—given that most tools are tuned to identify strong compositional skew—and the biological importance of such simple, repetitive motifs in protein function, aggregation, or phase separation. Significantly, this gradient highlights the value of methodological heterogeneity in LCR detection: whereas high-consensus regions can be genuine biological hot spots of compositional bias, a strict consensus-based approach will discard compositionally dense or unusual LCRs. Therefore, combining findings from both strict and relaxed detection thresholds from different algorithms allows a more inclusive capture of the LCR landscape, encompassing both highly repetitive and subtly biased sequences with functional significance.Fig. 2Compositional and sequence complexity characteristics of LCRs stratified by detection consensus. (A) Shows a stacked bar plot showing the proportional contribution of peptide motifs across LCRs grouped by the number of methods (n) detecting each region (consensus level, n = 1–13). Motifs include mono-, di-, and tri-peptide compositions, with colours representing specific amino acids or repeating sequence patterns and hatching indicating peptide type. Regions detected by fewer methods are enriched in diverse and heterogeneous motif compositions, whereas higher-consensus regions show increased dominance of specific low-complexity motifs, reflecting stronger compositional bias. (B) Depicts boxplots of Shannon entropy for LCRs stratified by consensus level. Lower-consensus regions (detected by fewer methods) exhibit higher entropy, indicating greater sequence heterogeneity, while regions detected by multiple methods show progressively reduced entropy, consistent with increased repetitiveness and compositional constraint. (C) Depicts boxplots of compositional purity across the same consensus groups. Purity increases with the number of detecting methods, indicating that high-consensus LCRs are more compositionally homogeneous and dominated by fewer amino acid types. Compositional and sequence complexity characteristics of LCRs stratified by detection consensus. (A) Shows a stacked bar plot showing the proportional contribution of peptide motifs across LCRs grouped by the number of methods (n) detecting each region (consensus level, n = 1–13). Motifs include mono-, di-, and tri-peptide compositions, with colours representing specific amino acids or repeating sequence patterns and hatching indicating peptide type. Regions detected by fewer methods are enriched in diverse and heterogeneous motif compositions, whereas higher-consensus regions show increased dominance of specific low-complexity motifs, reflecting stronger compositional bias. (B) Depicts boxplots of Shannon entropy for LCRs stratified by consensus level. Lower-consensus regions (detected by fewer methods) exhibit higher entropy, indicating greater sequence heterogeneity, while regions detected by multiple methods show progressively reduced entropy, consistent with increased repetitiveness and compositional constraint. (C) Depicts boxplots of compositional purity across the same consensus groups. Purity increases with the number of detecting methods, indicating that high-consensus LCRs are more compositionally homogeneous and dominated by fewer amino acid types. Low-complexity sequence space is progressively constrained by method agreement In Fig. 2, a pattern of decrease in entropy and increase in sequence purity is observed, with an increase in the consensus of the number of methods used. Observing this trend, efforts were made to visualise the LC sequence space by increasing the consensus of the number of methods used (where k = number of methods). In Suppl. Figure 1, density contours reveal that all k tiers lie along a common diagonal manifold in LC space, but increasing k progressively sharpens this distribution. Low-consensus regions occupy diffuse, multimodal densities, whereas high-consensus regions collapse into compact peaks. Thus, consensus does not redefine low-complexity, but increasingly restricts it to a narrow, well-defined subset. In Suppl. Figure 2, centroid trajectories across k values demonstrate a coherent directional flow: increasing k corresponds to increasing mutation tolerance and decreasing compositional dominance. Variability is high at low k, reflecting method-specific definitions, but narrows substantially at high k, indicating convergence toward a shared LC regime. This monotonic progression implies that consensus effectively filters out borderline cases, retaining only regions with strong, method-independent low-complexity signatures. In Suppl. Figure 3, the cumulative consensus map shows that the highest k values do not occur at extremes of compositional bias, but instead cluster in intermediate regions combining moderate dominance with elevated mutation rates. Regions with low mutation and low dominance remain largely unoccupied, while extreme repeats show limited consensus. This pattern indicates that methodological agreement is highest for LC regions balancing compositional bias and sequence variability, rather than for strictly repetitive motifs. In Suppl. Figure 4, stratification by consensus tier (k) reveals that agreement among LC detection methods is not uniform across sequence space. Low k tiers (1–3) span broad regions with moderate mutation and substantial amino-acid dominance, indicating that many methods independently flag compositionally biased segments. As k increases, the occupied space contracts and shifts toward lower mutation percentages and higher dominance. High-consensus regions (k ≥ 9) are rare and confined to highly homogeneous sequences, suggesting that only LC regions are robustly recognised across algorithms. The retention rate of LCRs predicted by independent methods drops with increasing sequence purity The purity-retention profiles in Fig. 3 indicate unique behaviours for each LCR detection tool with increasing compositional bias (purity). LCRFinder is one of the most uniform tools at all purity levels, retaining high values even at very high purity limits. Its performance curve is one of gradual decline, indicating strong sensitivity to highly repetitive and homopolymeric areas. Dotplot also retains relatively well at moderate to high purity levels but starts to fall around 40 to 50% purity, suggesting good but slightly less sensitivity than LCRFinder. SEG strict exhibits excellent retention across the entire purity range, even better than most tools at high purity levels, and is therefore especially well-suited to identifying highly compositionally biased regions. Its intermediate and default forms (SEG intermediate and SEG) have similar patterns but with drop-offs in retention occurring earlier, indicating a compromise between sensitivity and specificity. T-REKS has a moderate percentage of sequences retained at low-to-mid purity levels but begins a consistent decline after 25% purity, indicating that it prefers moderately biased areas over very pure homorepeats. The fLPS family of programs displays a definite hierarchy according to stringency of parameters. fLPS and fLPS 2.0 have similar retention profiles, trapping a wide range of LCRs with moderate purity, but both fall off steeply at purities > 50–60%. Their more stringent versions (fLPS strict and fLPS 2.0 strict) have significantly lower retention at lower purities and fall off steeply, suggesting their stringent bias towards highly repetitive, homopolymeric sequences. This indicates how the same algorithm structure can be adjusted to move the detection spectrum towards either broader or narrower compositional definitions. AlcoR (Mode 1 and Mode 2), on the other hand, maintains a high percentage of LCRs only at low purity levels, with a sharp fall beyond 25–30% purity, and practically no sequences preserved beyond 50% purity. This trend indicates that AlcoR is best suited for identifying complex, compositionally heterogeneous LCRs and is less effective at finding simple homorepeats. Likewise, XSTREAM detects a wide range of low purity LCRs but loses the majority of detections with increasing purity. Its curve highlights an extremely strong bias toward identifying complex, heterogeneous regions. Overall, every tool has a distinct response to compositional purity increase. Tools such as LCRFinder and SEG strict are very efficient in identifying plain, repetitive LCRs. Some others, such as AlcoR and XSTREAM, perform well to pick up complex, low-purity areas.Fig. 3Purity distribution of LCRs detected by different computational methods. This figure shows the variation in LCR purity across detection methods. The x-axis represents the purity level, defined as the percentage of the most dominant amino acid within each LCR, while the y-axis denotes the proportion of total LCRs retained at or above each purity threshold. Each line corresponds to a different method, as indicated by the colour-coded legend. Steeper declines indicate methods that detect fewer highly pure (compositionally biased) regions, whereas gradual slopes reflect identification of more homogeneous LCRs. This comparison emphasises how algorithms vary in their sensitivity to amino acid repetitiveness. Purity distribution of LCRs detected by different computational methods. This figure shows the variation in LCR purity across detection methods. The x-axis represents the purity level, defined as the percentage of the most dominant amino acid within each LCR, while the y-axis denotes the proportion of total LCRs retained at or above each purity threshold. Each line corresponds to a different method, as indicated by the colour-coded legend. Steeper declines indicate methods that detect fewer highly pure (compositionally biased) regions, whereas gradual slopes reflect identification of more homogeneous LCRs. This comparison emphasises how algorithms vary in their sensitivity to amino acid repetitiveness. XSTREAM was also run with different -m parameters (1 to 5), on the human proteome. However, the results are nearly identical for -m values of 2 to 5. Very similar sharp drops in entity retention are observed when the sequence purity is gradually increased (Suppl. Figure 9). Only at -m = 1, almost 38% of the entities are retained, even at 100% sequence purity. [At default mode, the -m parameter is set at 5.] The fLPS and SEG groups provide sensitivity tunability so that flexible deployment may be based on target LCR properties. These biases in individual tools highlight the utility of integrative methods, making use of complementary strengths to derive a more comprehensive LCR picture. The similarity between the methods is dependent on their underlying algorithms Figure 4 illustrates several top LCR detection algorithms which were compared in terms of their pairwise residue-level overlap using the scores in a Jaccard Matrix, providing insight into their algorithmic foundations and detection preferences. The conventional SEG method, which identifies regions of low sequence complexity through computation of Shannon entropy in moving windows, shared a moderate degree of overlap (represented by the Jaccard scores) with compositionally skewed tools: 0.1235 with fLPS 2.0 strict, 0.0981 with fLPS 2.0, and 0.0509 with AlcoR. However, its overlap with repeat-detecting tools like T-REKS (0.1399) and Dotplot (0.3706) was quite high, suggesting that SEG could also identify repetitive regions if they possess less complex compositions. fLPS and its updated version, fLPS 2.0 (strict modes enabled), showed very high reciprocal overlap (0.9462 between fLPS 2.0 and fLPS, and 0.6728 between fLPS2 and fLPS 2.0 strict), reflecting their amino acid compositional bias sensitivity. The strict modes, designed to suppress false positives, as expected, overlapped less with the rest of the tools (e.g., fLPS and SEG: 0.099, vs. fLPS 2.0 strict and SEG: 0.1235), reflecting that parameter optimisation influences region identification. AlcoR and AlcoR 2 (AlcoR operated in 2nd mode), from statistical profiles of amino acid usage, presented high inter-method concordance (0.7236), thus validating method reproducibility. Crossmatches with other methods, such as fLPS 2.0 (0.1031) and T-REKS (0.1179), further suggest that AlcoR methods detect both repetitive and compositional components to a limited extent. The Dotplot technique, which identifies self-similarity between sequences through comparisons of overlap windows, showed minimal overlap with most composition-based tools: 0.0346 with AlcoR, 0.0478 with fLPS, and 0.0621 with fLPS 2.0 strict. However, it did have significant overlap with SEG (0.3706) and LCRFinder (0.2557), demonstrating its capacity to identify visually separable repeat-like structures that are also compositionally simple. LCRFinder, detecting highly frequent motifs (with mismatches), was closest in its top overlap to Dotplot (0.2557) and SEG (0.1853), but considerably less so with algorithms like fLPS 2.0 strict (0.0375) or AlcoR (0.0299), indicating its emphasis on detecting local repetitive motifs more than compositionally skewed patterns. T-REKS, as an optimised tool for tandem repeat detection by self-alignment and periodicity verification, exhibited maximum overlap with XSTREAM (0.1469) (also a program for finding repeats), followed by moderate overlap with SEG (0.1399) and Dotplot (0.1241). Overlap of the latter with fLPS (0.0424) and AlcoR (0.1179), however, was relatively low, validating its structural rather than compositional repetition target. XSTREAM, employing suffix trees and pattern matching to identify imperfect, complex repeats, overlapped very little with composition-biased methods such as SEG strict (0.0054) and fLPS 2.0 strict (0.0184), but was most similar to T-REKS (0.1469), which indicates its repeat-oriented nature. Finally, in the stricter forms of SEG (e.g., SEG strict), overlaps were systematically reduced, as found between SEG strict vs. SEG (0.0559) and SEG strict vs. fLPS2 (0.0082), which affirms that strict complexity constraints merely permit the discovery of the most well-defined LCRs. These results evidence clear partitioning of algorithms based on detection principles. Composition-biased tools (fLPS, fLPS 2.0h, SEG) exhibit high mutual overlap due to their bias for amino acid frequency. Tools like Dotplot, T-REKS, XSTREAM, and LCRFinder, which target repetitive or periodic motifs, exhibit relatively low overlap with composition-biased tools. Dotplot and LCRFinder overlap reasonably, but XSTREAM is low against nearly everything except T-REKS, and even that overlap is still modest. This divergence illustrates the importance of tool selection according to biological context: if the goal is to detect statistically enriched amino acid regions, tools like fLPS or fLPS 2.0 are best; to identify repetitive, periodic or motif-based, T-REKS or XSTREAM are better.Fig. 4Pairwise overlap and similarity between LCR detection methods. It presents the Jaccard similarity index for each pair of methods, calculated as the intersection divided by the union of detected LCR sets. The colour gradient (white to red) denotes increasing similarity, with diagonal values (1.0) representing self-comparisons. Together, this quantifiesthe extent of agreement and divergence among the evaluated detection algorithms. Pairwise overlap and similarity between LCR detection methods. It presents the Jaccard similarity index for each pair of methods, calculated as the intersection divided by the union of detected LCR sets. The colour gradient (white to red) denotes increasing similarity, with diagonal values (1.0) representing self-comparisons. Together, this quantifiesthe extent of agreement and divergence among the evaluated detection algorithms. True LCR boundaries are fundamentally fuzzy and ambiguous Suppl. Figure 5 and Suppl. Figure 6 present probabilistic reconstructions of low-complexity (LC) space across increasing consensus stringency (k = 7–12). In both figures, each panel shows a posterior probability surface P(LCR∣bin) in mutation–composition space, with black contours indicating ROC-optimised decision boundaries. Across both datasets, increasing k produces a systematic contraction of the LC-permissive region toward the low-mutation, high-compositional-bias corner of the space. This trend reflects the fact that higher consensus thresholds preferentially retain sequence segments that are consistently identified as low-complexity by many independent methods, thereby enriching for highly repetitive sequences. Correspondingly, the ROC-optimised probability threshold decreases monotonically with k, indicating that increasingly confident LC assignments emerge from progressively smaller regions of sequence space. Importantly, the inferred boundaries are not sharp demarcations, but fuzzy, graded transition zones. The posterior landscapes exhibit smooth probability gradients rather than step-like separations, underscoring that low-complexity is not a binary sequence property. Instead, LC behaviour arises from a continuum of sequence architectures governed by compositional entropy and mutational tolerance. Regions near the diagonal boundary—where mutation rate and amino-acid dominance trade off—retain intermediate posterior probabilities and shift gradually with k, reflecting ambiguity rather than misclassification. Comparison between PDB-based (Suppl. Figure 5) and DisProt-based (Suppl. Figure 6) maps further highlights that boundary placement depends on the underlying biological reference. DisProt-derived boundaries extend further into higher-mutation regimes, consistent with the presence of experimentally validated disorder that tolerates greater sequence variability, whereas UniProt-derived boundaries are more restrictive. The GAM (Generalized Additive Model) posterior maps derived from DisProt- and PDB-anchored models (Suppl. Figure 7; Suppl. Figure 8) provide a unified view of how low-complexity (LC) space is structured by sequence entropy and compositional bias. Across all panels (A–F), LC propensity is organised along a curved boundary in the two-dimensional space defined by mutation percentage and dominance of the most frequent amino acid. This boundary is not arbitrary: it reflects a tradeoff whereby increasing mutational heterogeneity necessitates reduced compositional bias, and vice versa. Regions combining high mutation with high dominance are systematically excluded, indicating sequence architectures that are rarely compatible with either curated regions missing in PDB structures or experimentally validated disorder (DisProt). A key trend emerges across panels as the minimum method consensus constraint (k_min) increases. At lower k values (A–C), the LC space is expansive and irregular, encompassing short, compositionally mixed regions that satisfy permissive disorder criteria. As k increases (D–F), the LC space contracts progressively, and the decision boundary becomes smoother and more monotonic. This shrinkage indicates that LC regions detected with consensus across more methods must satisfy stricter entropic constraints, filtering out marginal or noisy configurations and retaining only sequence architectures with robust disorder signatures. Notably, DisProt-derived boundaries are consistently more restrictive than PDB-derived ones, underscoring the tighter constraints imposed by experimental validation relative to missing residue-based curation. Together, these figures demonstrate that low-complexity (LC) boundaries are not sharp decision surfaces but probabilistic envelopes embedded in sequence-information space, whose shape depends on both reference background sequences and scale. Apparent LC space emerges gradually as compositional bias strengthens and effective sequence entropy decreases, rather than at a single, universal cutoff. At the same time, LC space is intrinsically scale-dependent: increasing the minimum method consensus constraint (k) systematically contracts this space and smooths its geometry, reflecting progressively stricter entropic requirements for regions detected across a greater number of methods. The convergence toward stable, smooth boundaries at higher k does not resolve LC into a uniquely defined class, but instead highlights that “true” LCR boundaries remain fundamentally fuzzy and ambiguous. Collectively, these results support an information-theoretic view of intrinsic disorder in which allowable LC architectures occupy a constrained, annotation-dependent envelope governed by the balance between entropy and compositional bias, rather than a fixed or discrete notion of low-complexity. LCR complexity profiles of the methods show varying degrees of clustering To further evaluate the behaviour of various low-complexity region (LCR) detection tools, we analysed the two-dimensional distribution of percentage mutation and percentage most common amino acid inside predicted LCRs, employing the Low-Complexity triangle model depicted in Fig. 5. This triangle divides sequence space into areas corresponding to homorepeats (top), long imperfect repeats (bottom left), low-complexity regions (centre), and high-complexity sequences (bottom right). The heatmaps of density for each of the tools demonstrate their respective detection biases and sensitivities in this compositional and mutational landscape.Fig. 5Relationship between amino acid composition and mutation percentage in LCRs across detection methods. Each panel corresponds to one detection method and illustrates the joint distribution of amino acid composition bias and sequence mutation percentage. The x-axis represents mutation percentage (fraction of variable residues within the LCR), and the y-axis shows the percentage of the most frequent amino acid. The colour gradient indicates the number of LCRs within each bin, as shown by the colour key. The diagonal trend visible across most methods indicates that higher amino acid dominance is typically associated with lower mutation percentages, reflecting stronger compositional homogeneity in repetitive regions. Relationship between amino acid composition and mutation percentage in LCRs across detection methods. Each panel corresponds to one detection method and illustrates the joint distribution of amino acid composition bias and sequence mutation percentage. The x-axis represents mutation percentage (fraction of variable residues within the LCR), and the y-axis shows the percentage of the most frequent amino acid. The colour gradient indicates the number of LCRs within each bin, as shown by the colour key. The diagonal trend visible across most methods indicates that higher amino acid dominance is typically associated with lower mutation percentages, reflecting stronger compositional homogeneity in repetitive regions. AlcoR and AlcoR2 both showed high concentrations of dark red signal in the bottom-right corner of the triangle, which represents sequences with high rates of mutation and comparatively low amino acid bias. These trends show that both tools often predict relatively noisy and complex areas as LCRs. Their intrinsic algorithms, which take compositional skew and k-mer frequency metrics into account, seem to be geared towards detecting “fuzzy” LCRs—regions that are far from perfect repeats but exhibit faint compositional anomalies. Although this renders AlcoR and AlcoR2 very sensitive, it also implies a liberal detection mechanism that might sacrifice specificity for the sake of coverage, detecting areas that other methods would filter out as not having enough compositional or structural simplicity. Dotplot, on the other hand, showed a steeper diagonal pattern in the centre of the triangle. This represents an emphasis on moderately biased, moderately mutated sequence regions—typically regions with classical LCRs. The algorithm’s use of visual self-similarity heuristics causes it to be more discriminating in its detection profile than AlcoR and to have very little detection at high-complexity sequence space. It can avoid overprediction without sacrificing the detection of many low-complexity patterns that are varied in nature. The fLPS family, comprising fLPS, fLPS strict, fLPS 2.0, and fLPS 2.0 strict, showed universal detection, with darker areas stretching from the top (high amino acid bias, i.e., homorepeats) towards the centre. They are very good at detecting both highly biased homopolymeric tracts and moderately imperfect LCRs, reflecting their compositional filtering and scoring flexibility. But the fact that the signal is seen in regions of intermediate mutation implies that the default (non-strict) variants of fLPS are, to some extent, permissive. Their strict counterparts, by contrast, limit predictions to more compositionally extreme and less mutated parts, raising specificity at the risk of excluding important biological variation. SEG and its variants offered a wide range of detection. The default SEG tool exhibited dark signals from the lower-left to the triangle’s centre, indicating a good ability to detect long imperfect repeats and biased low-complexity sequences. SEG’s entropy-based scoring accommodates this flexibility. The intermediate and strict versions progressively lowered coverage, with SEG strict narrowly examining areas of high compositional bias and low mutation—largely homorepeats—exhibiting greater stringency at the cost of breadth. LCRFinder exhibited moderate dark red signals along the left-hand side of the triangle, indicating a bias towards sequences with moderate amino acid bias and relatively low mutation. This trend indicates that LCRFinder balances compositional bias and structural complexity. Its detection approach, which incorporates Shannon entropy with expected-vs-observed residue frequencies, enables it to spot a variety of biased and partially repetitive areas without generating false positives in highly complex or disordered regions—showing improved specificity over AlcoR. T-REKS, employing a k-means clustering strategy optimised for tandem repeat detection, was found to be located primarily in the central part of the triangle, with a slight bias toward sequences with moderate bias and low-to-moderate mutation. It steers clear of highly complex and highly biased areas, as designed to identify structured, periodic LCRs. T-REKS is slightly more permissive of mutation than LCRFinder, though it still efficiently eliminates high-complexity noise. XSTREAM showed the most selective profile, with scattered but localised dark red signals restricted to areas ofhigh mutation and low-dominance. This suggests a strong preference for long, complex and compositionally heterogeneous tandem arrays. This makes it a good choice for identifying long tandem repeats. Overall, these complexity heatmaps highlight that LCR detection is strongly tool-dependent. Compositional tools like fLPS, AlcoR, and SEG are optimally suited for the detection of amino acid enrichment or disorder-driven regions, whereas structure-aware tools like T-REKS and XSTREAM emphasise periodicity and sequence regularity. Hybrid methods like LCRFinder provide greater flexibility by combining compositional and statistical aspects. Further, in all the tools, strict or conservative modes always sacrifice sensitivity for accuracy, possibly eliminating noisy but biologically relevant LCRs. These findings underscore the need to choose a suitable detection strategy depending on the particular biological question or sequence context: there is no one tool that can catch all of the diversity of low-complexity sequence space. The methods exhibit varying degrees of sensitivity and specificity to LCR detection Figure 6 illustrates performance trends (TPR and FPR) of different LCR detection tools over four variables: gene length (A), number of LCRs per gene (B), LCR coverage percentage (C), and LCR: Gene Entropy Ratio (D). Many common trends of increase and decrease can be noted. Across all the panels, fLPS and fLPS 2.0 (strict variants) exhibit a consistently higher TPR as the gene length, number of LCRs, coverage, and entropy ratio increase, indicating stronger agreement with the reference definition under increasingly pronounced sequence features. Their FPRs also significantly increase, particularly in Panels A and D, reflecting greater divergence from the reference in more complex sequence contexts. Dotplot and SEG (intermediate and default) have constantly high TPRs on all variables, save for slight fluctuations. Their FPRs are relatively low and constant, but increase with increased length and higher entropy (Panels A and D). They all prove stable over various conditions, with Dotplot producing an overall stability of performance that is optimal. SEG strict is unique in demonstrating a relatively flat TPR trend, being lower than its more relaxed variants but with an exceptionally low and stable FPR under all conditions. The trend supports its application as a high-specificity tool, particularly when minimising false positives. LCRFinder and T-REKS exhibit modest increases in TPR with growing LCR features, while keeping FPR low or moderately increasing, indicating balanced but less stringent detection. By contrast, AlcoR (both modes) and XSTREAM exhibit flat or slightly decreasing TPRs under most conditions, reflecting poor sensitivity despite increasingly prominent LCR features. Their FPRs are low, indicating conservative detection that presumably results in false negatives rather than true specificity. In general, most tools exhibit a positive correlation between TPR and rising LCR-related features, but few (specifically, Dotplot and SEG variants) are able to restrain the increase in FPR, and hence are more suitable for precise LCR detection in a wide range of genomic contexts. It is important to emphasise that these metrics quantify agreement with the operational reference definition rather than absolute detection of biologically validated LCRs. Therefore, higher TPR reflects closer concordance with this definition, not necessarily greater biological relevance.Fig. 6Evaluation of LCR detection performance across sequence and compositional parameters. This figure assesses the performance of different tools in detecting internally defined reference LCR set under varying conditions. True positive rate (TPR; left subpanels) and false positive rate (FPR; right subpanels) are plotted against four parameters (TPR/FPR reflect agreement with an operational internal reference, not biological truth): (A) gene length categories, (B) number of LCRs per gene, (C) LCR coverage percentage, and (D) the LCR-to-gene entropy ratio. The x-axes represent increasing parameter values, while the y-axes show detection rates. Distinct colours denote individual tools, as indicated in the legend. Higher TPRs indicate more sensitive detection, whereas lower FPRs reflect higher specificity. These comparisons reveal how each method’s performance varies with sequence length, LCR density, coverage, and compositional complexity. Evaluation of LCR detection performance across sequence and compositional parameters. This figure assesses the performance of different tools in detecting internally defined reference LCR set under varying conditions. True positive rate (TPR; left subpanels) and false positive rate (FPR; right subpanels) are plotted against four parameters (TPR/FPR reflect agreement with an operational internal reference, not biological truth): (A) gene length categories, (B) number of LCRs per gene, (C) LCR coverage percentage, and (D) the LCR-to-gene entropy ratio. The x-axes represent increasing parameter values, while the y-axes show detection rates. Distinct colours denote individual tools, as indicated in the legend. Higher TPRs indicate more sensitive detection, whereas lower FPRs reflect higher specificity. These comparisons reveal how each method’s performance varies with sequence length, LCR density, coverage, and compositional complexity. A complete performance comparison of an extensive number of low-complexity region (LCR) detection tools (as shown in Fig. 7)—based on an average true positive rate (TPR) and false positive rate (FPR) across a broad set of conditions—reveals that different sensitivity–specificity trade-offs exist between methods. Comparisons were made using a variety of parameters, including gene length, LCR per gene, coverage percentage of the LCRs, and LCR: Gene Entropy Ratio, and hence constituted a rigorous test of the consistency of each methodology under varying genomic circumstances. Among all the tools, fLPS and fLPS 2.0 (both of their strict modes) consistently registered the highest TPRs and exhibited their natural strength for detecting LCRs even at tough and heterogeneous cases. Due to this, they are incredibly efficient in obtaining real LCRs under any typical conditions. But this sensitivity comes with a cost: they also possessed some of the highest FPRs, particularly when gene length, LCR density, and entropy ratio were increasing. This means that while these programs are sensitive to the detection of diverse LCRs, they would overpredict, reducing their specificity by falsely labelling non-LCR regions. SEG, especially default and intermediate modes, and Dotplot, respectively, showed the best possible balance of sensitivity and specificity. Both tools produced high TPRs with comparatively low FPRs, reflecting efficient actual LCR detection with very few false positives. Surprisingly, Dotplot ranked as the best-performing tool overall, consistently outperforming all other methods by rendering correct detection with the least number of false positives. SEG strict, while possessing lower TPRs, had one of the lowest FPRs across all conditions and, therefore, is an excellent choice for high-specificity use where false positives must remain low. LCRFinder and T-REKS were moderately sensitive and had low false positives, indicating a conservative but quite reasonable performance. They both do well in obtaining a good trade-off for overall LCR detection, especially when there is a desire for precision along with some inclusivity. Conversely, AlcoR (default and alternative mode) and XSTREAM all reported low TPRs with low FPRs, likely due to their extremely conservative detection thresholds. This is in agreement with a tendency to under-detect true LCRs and hence be suboptimal for complete LCR identification, particularly in high LCR diversity or compositional complexity datasets. Together, this comparison indicates that there is no single tool that is better all around. These methods, like fLPS and fLPS 2.0, are appropriate for thorough and sensitive searches but are prone to generating false positives. SEG and Dotplot are more balanced and stable in their performance and are top contenders as efficient, general-purpose LCR detectors. If high specificity is required, SEG strict is the most precise method. Ultimately, the choice should be made based on biological context and tolerance for false positives, as performance can be extremely variable as a function of complexity and type of sequence data.Fig. 7Comparative summary of average detection performance metrics across tools. This figure presents the mean true positive rate (TPR, cyan) and mean false positive rate (FPR, pink) for all detection methods across four parameter categories (TPR/FPR reflect agreement with an operational internal reference, not biological truth). (A) shows the effect of gene length, (B) depicts the effect of the number of LCRs per gene, (C) shows the effect of LCR coverage percentage, and (D) represents the effect of the LCR-to-gene entropy ratio. The x-axis lists all detection tools, and the y-axis indicates normalised rate values on a 0–1 scale. This integrated summary provides a direct comparison of each method’s overall sensitivity and specificity across multiple structural and compositional parameters. Comparative summary of average detection performance metrics across tools. This figure presents the mean true positive rate (TPR, cyan) and mean false positive rate (FPR, pink) for all detection methods across four parameter categories (TPR/FPR reflect agreement with an operational internal reference, not biological truth). (A) shows the effect of gene length, (B) depicts the effect of the number of LCRs per gene, (C) shows the effect of LCR coverage percentage, and (D) represents the effect of the LCR-to-gene entropy ratio. The x-axis lists all detection tools, and the y-axis indicates normalised rate values on a 0–1 scale. This integrated summary provides a direct comparison of each method’s overall sensitivity and specificity across multiple structural and compositional parameters. Decoupling sequence low-complexity from intrinsic disorder in protein sequence space To further examine the sequence characteristics of structurally disordered regions, we projected DisProt-annotated regions and PDB segments lacking resolved coordinates onto the same mutation percentage versus most frequent amino acid percentage space (Fig. 8). DisProt regions occupy a broad triangular envelope similar to that observed for low-complexity regions, but without clear segregation by experimental annotation method. In contrast, missing PDB residues are enriched toward high mutation percentages and low compositional dominance, indicating highly heterogeneous sequence composition. Notably, these distributions only partially overlap with the low-complexity regimes defined earlier, reinforcing that intrinsic disorder and sequence low-complexity capture related but distinct aspects of protein sequence organisation.Fig. 8Visualising the complexity of structurally disordered regions. (A) Scatter plot of DisProt-annotated intrinsically disordered regions showing mutation percentage versus most frequent amino acid percentage. Points are colored by experimental method. Regions from diverse techniques occupy a common triangular envelope, indicating a conserved sequence-complexity constraint independent of experimental methods. (B) Heatmap showing the distribution of PDB regions lacking resolved coordinates projected onto mutation percentage (x-axis) and most frequent amino acid percentage (y-axis). Colour intensity indicates segment counts (log-scaled). Missing residues are enriched at high mutation percentages and low compositional dominance, consistent with highly heterogeneous sequence regions. Visualising the complexity of structurally disordered regions. (A) Scatter plot of DisProt-annotated intrinsically disordered regions showing mutation percentage versus most frequent amino acid percentage. Points are colored by experimental method. Regions from diverse techniques occupy a common triangular envelope, indicating a conserved sequence-complexity constraint independent of experimental methods. (B) Heatmap showing the distribution of PDB regions lacking resolved coordinates projected onto mutation percentage (x-axis) and most frequent amino acid percentage (y-axis). Colour intensity indicates segment counts (log-scaled). Missing residues are enriched at high mutation percentages and low compositional dominance, consistent with highly heterogeneous sequence regions. Discussion Our systematic benchmarking of low-complexity region (LCR) detection tools across the human proteome highlights several important insights with broad implications for protein sequence annotation, functional inference and evolutionary interpretation. First, the clear divergence in length-, coverage- and entropy-based detection profiles between methods underscores that LCRs are not a monolithic class of sequence features, but rather a heterogeneous collection of compartments that differ in size, compositional bias, repetitiveness and biological context. For example, our finding that tools that are composition-bias tuned and tools that are repeat-periodicity tuned capture largely non-overlapping LCR sets echoes that “statistical measures alone cannot capture all structural aspects of LCRs”25. Second, our observation that LCRs common to multiple tools tend to be longer, more repetitive (i.e., lower Shannon entropy) and more consistently detected suggests that detection consensus may by itself be a useful proxy for functional relevance. However, the biological relevance of LCRs is inherently context-dependent and cannot be inferred from sequence complexity alone. Functionally relevant LCRs may include regions involved in phase separation, protein–protein interactions, transcriptional activation, or regulatory processes, and in some cases may overlap with intrinsically disordered regions (IDRs) that enable conformational flexibility or disorder-to-order transitions. In that regard, previous studies have shown that LCRs located at protein termini tend to be associated with increased numbers of binding partners56. Our multi-tool consensus may preferentially enrich for LCRs with stronger structural or interaction-driven constraints, although direct functional validation will be required to establish this link. Importantly, the designation of a region as a low-complexity region is based on its compositional properties rather than its functional role, and regions classified as LCRs may or may not have biological significance. Third, the Jaccard similarity analysis revealed tight clustering of methods that use similar algorithmic lenses (e.g., composition vs. repeat detection), which aligns with previous evolutionary analyses. For instance, it has been shown that LCR conservation depends on functional context and subcellular localisation, suggesting that different algorithms may preferentially detect distinct sequence subclasses57. Fourth, our entropy and purity analyses contribute to the emerging view of LCRs as functional modules rather than being uniformly neutral or non-functional. It has been found that many LCRs within disordered regions may adopt secondary structures (e.g., helices)58. Our results show that lower-entropy (more repetitive) LCRs are preferentially retained across multiple detection methods. This enrichment may reflect a combination of biological constraint and intrinsic sequence properties, but does not imply that low-entropy regions are universally maintained by evolutionary selection. Indeed, such regions can arise through stochastic processes such as replication slippage, and only a subset is likely to acquire or retain functional relevance10. Fifth, the observed association of LCR-containing proteins with lower protein abundance but higher transcript abundance59 offers a useful context for interpreting our coverage-based findings, suggesting that LCRs may influence translational efficiency or post-translational stability rather than transcript production per se. Our work also highlights specific methodological recommendations. A unified reporting framework that includes length-bin distribution, coverage percentage, amino-acid composition, Shannon entropy and consensus frequency (number of tools detecting) offers richer annotation of LCR-calls than simple “present/absent”60. In practical terms, our results suggest that functional studies of LCRs (e.g., phase separation, promiscuous interactions, aggregation) should prioritise LCRs that are long (> 50 aa), have high-coverage within a protein (> 20%), and are detected by multiple independent algorithms—because these appear enriched for lower entropy and higher compositional purity. This prioritisation strategy is consistent with previously observed links between repetitive or compositionally biased sequences and structural or functional properties, while remaining exploratory in nature61. From an evolutionary perspective, our detection-consensus clustering provides a quantitative basis for assessing LCR “families” of detection methods, which may reflect underlying evolutionary modes of LCR origin (e.g., slippage, duplication, expansion). In particular, poor correlation between protein and DNA low-complexity sequences62 suggests that many protein LCRs reflect post-translational or structural constraints that are invisible at the DNA level. Platforms such as PlaToLoCo provide a useful aggregation framework for contextual comparison of low-complexity region (LCR) annotations across multiple detection methods. However, because PlaToLoCo largely integrates outputs from the same underlying algorithms evaluated here, its annotations are not independent and should not be treated as ground truth. Accordingly, such platforms are best viewed as exploratory or comparative resources rather than benchmarks for method validation27. Our analyses further suggest that both structurally unresolved regions in the Protein Data Bank and experimentally validated intrinsically disordered regions are governed by shared sequence-level constraints linking amino-acid composition, mutational tolerance, and conformational flexibility. Regions with high sequence heterogeneity are more likely to lack resolved structural coordinates, consistent with the long-standing view that extensive compositional diversity disfavors stable structure formation51,63. At the same time, the presence of missing residues within compositionally biased segments indicates that low-complexity alone is insufficient to explain structural absence, implicating additional factors such as dynamics and experimental context64,65. Experimentally curated disorder occupies a constrained, method-independent region of sequence-complexity space, supporting the idea that intrinsic disorder reflects regulated and evolutionarily maintained sequence architectures rather than random compositional noise29,66,67. An important observation emerging from this study is the limited overlap between computationally defined low-complexity regions and experimentally characterised intrinsically disordered regions. As shown by the sequence-space distributions of DisProt annotations and unresolved PDB segments (Fig. 8), structurally disordered regions span a broader compositional regime and are enriched in highly heterogeneous sequences that do not consistently fall within the low-complexity criteria used here. This indicates that, while low-complexity and intrinsic disorder may intersect in certain cases, they represent fundamentally distinct sequence properties. Consequently, agreement between LCR detection methods should be interpreted as methodological concordance rather than evidence of biological validity, and caution must be exercised when inferring functional or structural roles from low-complexity annotations alone. In practical terms, this benchmark can be used as a guide for tool selection based on the biological question being asked. When the primary aim is to detect broad compositional bias, including weakly biased or extended low-complexity segments, methods such as fLPS and fLPS 2.0 are likely to provide the most inclusive annotations. When the focus is instead on highly repetitive, compositionally pure, or tandem-repeat-like regions, more conservative or repeat-oriented approaches such as SEG strict, T-REKS, or XSTREAM may be more appropriate. Dotplot and the intermediate SEG settings appear particularly useful when a balance between breadth and selectivity is desired. For studies seeking high-confidence candidate regions for downstream functional analysis, a consensus strategy that prioritises regions detected by multiple independent methods may offer a robust starting point. However, such consensus should be interpreted as methodological concordance rather than biological proof, and candidate LCRs should ideally be followed by orthogonal validation using structural, evolutionary, biochemical, or functional evidence. In conclusion, our benchmarking and visualisation framework advances the annotation of low-complexity regions by moving beyond single-tool outputs to a multi-metric consensus model, comparing and interpreting LCR predictions across methods. Limitations and future directions While the present benchmarking framework provides a comprehensive comparative analysis of existing LCR detection tools, several limitations warrant consideration. First, the study focuses exclusively on the human proteome, and LCR features may vary substantially across phylogenetic lineages, particularly in organisms with extreme GC content or compact genomes. Expanding this framework to include diverse taxa could elucidate evolutionary trends in LCR expansion and constraint. Although we benchmarked 8 major methods, emerging methods (e.g., machine-learning based) might detect additional classes of LCRs, and our human-proteome focus may limit generalisation to other lineages (e.g., prokaryotes, pathogens) where LCR dynamics differ68. One such method is LCD-Composer69, which is a composition-based predictor relying on overlapping underlying principles. Second, our evaluation is based on algorithmic and compositional metrics rather than experimentally validated functional data. Integrating proteomic, structural, or biophysical evidence—such as phase-separation propensity or disorder–order transitions—would strengthen the biological interpretation of computationally predicted LCRs. Third, we note that restricting the dataset to a single representative isoform may underrepresent LCRs that are preferentially introduced or expanded through alternative splicing, particularly in intrinsically disordered terminal or linker regions. However, this choice was made to enable consistent cross-method benchmarking and to minimise redundancy-driven biases. Fourth, missing PDB residues reflect both intrinsic disorder and experimental limitations, and therefore cannot be interpreted as a direct proxy for disorder. Along with that, the analysis remains sequence-centric and does not explicitly account for cellular context, post-translational modifications, or interaction-induced folding, which may modulate disorder in vivo. This calls for sophisticated and well-planned experimental biology approaches. Finally, emerging machine-learning and deep-learning models68,70–72 trained on multi-parameter datasets could capture complex relationships between sequence composition, disorder, and functionality that remain inaccessible to traditional heuristic tools. Future developments should therefore aim toward a hybrid consensus paradigm, combining statistical, compositional, and predictive approaches to achieve more precise, biologically meaningful LCR annotation. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1