{"paper_id":"91ad7bc6-c703-4bc4-a88b-38948aa8f3ce","body_text":"1\nRapid and Interpretable Protein Contact Map Prediction \nUsing a Pattern-Matching Strategy\nAysima Hacisuleyman1* and Dirk Fasshauer1\n1Department of Computational Biology, University of Lausanne, CH-1015 Lausanne, Switzerland\n* Corresponding author: aisima.chatzisouleiman@unil.ch\nAbstract\nProtein sequence determines the structure, function, and dynamics of a protein. In recent years, \nenormous progress has been made in translating sequence information into structural information \nusing machine learning approaches. However, because of the underlying methodology, it is an \nimmense computational challenge to extract this information from the ever-increasing number of \nsequences.  In the present study, we show that it is possible to create two-dimensional contact maps \nfrom sequences, for which only a few exemplary structures are available on a laptop without the \nneed for GPUs or high-performance computing clusters. This is achieved by using a pattern matching \napproach. The resulting contact maps largely reflect the interactions in the three-dimensional \nstructures. The validity of our method was tested on the 25 protein domains, with abundant \nstructural data, achieving correlations of 0.73-0.94 between predicted and experimental contact \nmaps. To demonstrate broader applicability, we further validated our approach on 7,599 poorly \nannotated sequences using homologous structural templates, achieving a mean F1-score of 0.609 ± \n0.095 and mean accuracy of 0.954 ± 0.036 when compared against high-confidence AlphaFold \nstructures. These results demonstrate that our pattern matching approach maintains robust \nperformance even when relying on a small number of structural templates.\nKeywords:\npattern matching, contact maps prediction, sequence-structure relationship, structural templates, \nhomologous structures\n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.08.674800doi: bioRxiv preprint \n\n2\nAbbreviations:\nGPU; graphics processing unit, MSA; multiple sequence alignment, ROC; Receiver Operating \nCharacteristic, AUC; Area Under the curve, MCC; Matthews Correlation Coefficient, pLDDT; \npredicted Local Distance Difference Test \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.08.674800doi: bioRxiv preprint \n\n3\nIntroduction\nDespite the rapid growth of sequenced proteins, experimentally determined structures remain \navailable for only a small fraction, creating a gap between sequence data and structural knowledge. \nThis imbalance limits our ability to infer function, understand molecular mechanisms, and explore \nprotein dynamics at scale. Machine learning-based structure prediction methods, including \nAlphaFold21, RoseTTAFold2 and ESMfold3, have transformed structural biology, yet they require \nsubstantial computational resources that limit their application in high-throughput analyses.\nContact maps, representing pairwise spatial relationships between amino acid residues, \nprovide a simplified but informative representation of protein structure. These two-dimensional \nmatrices capture essential structural features while reducing computational complexity. \nHere, we present a pattern-matching approach that leverages existing structural repositories to \nrapidly generate contact maps for query sequences. Rather than predicting full three-dimensional \nstructures, the method identifies conserved contact patterns from homologous proteins and maps \nthem onto query sequences. By integrating multiple structural templates, including alternative \nconformations, our approach captures dynamic features of proteins and can identify conserved \nmotifs even when sequence similarity is limited. By analyzing contact patterns across multiple \nconformational states, one can gain insights into protein dynamics that might otherwise require \nextensive molecular dynamics simulations.\nThis strategy offers several advantages: it reduces computational cost, enables high-\nthroughput analyses, and provides biologically meaningful predictions for proteins with limited \nstructural information. To benchmark the method, we analyzed 25 well-characterized protein \ndomains, comparing their predicted contact maps with experimentally determined reference \nstructures. We further evaluated its applicability on poorly annotated sequences lacking \ncomprehensive structural coverage, demonstrating that evolutionarily related homologs can still \nprovide informative patterns for contact prediction.\n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.08.674800doi: bioRxiv preprint \n\n4\nMethods\nSequence and Structure Retrieval\nWe prepared two complementary datasets for method evaluation. The primary benchmark \nconsisted of 25 well-characterized protein domains from InterPro4 (Table 1). We retrieved all \navailable sequences and curated PDB structures for each entry to ensure high-quality annotation \nand capture conformational diversity. The second dataset comprised of 7,599 poorly annotated \nproteins from UniProt’s5 TrEMBL6 collection (Table 2, Table S2). Unreviewed sequences were filtered \nfor low annotation scores(annotation scores of 1), sequence diversity, and high-confidence \npredicted structures (with average pLDDT ≥ 80). JackHMMER7 was used to identify homologous PDB \nstructures, applying additional filters to exclude sequences with excessive (>1,500) or insufficient \n(<50) structural hits. After filtering, the dataset included proteins with diverse secondary structure \ncompositions representative of the general protein population. Specifics of the selection process are \nexplained and shown in the Zenodo repository. \nFor both datasets, available PDB structures were filtered by resolution (≤3.0 Å), and all NMR \nstructures were retained to preserve conformational variability. For domains lacking experimental \nstructures, AlphaFoldDB can be used as a complementary structural source.\nContact Pattern Identification\nContact patterns were defined as spatial arrangements of up to five amino acid residues whose \ncenters of mass are located within a maximum distance of 8.0 Å from each other. For each structure, \nall pairwise residue distances were computed, and clusters meeting this criterion were extracted as \npatterns. Each pattern was encoded using a simple notation: residues involved in contacts were \nrepresented by uppercase letters, and intervening residues by lowercase letters. Patterns were \nstored in a domain- or sequence-specific library for subsequent alignment.\nPattern Alignment and Contact Map Generation\nPredicted contact maps were generated by aligning patterns from the library to query sequences \nusing MATLAB’s(MATLAB R2023b) localalign function, with an increased gap-opening penalty of 10 \nto favor precise motif matching. When a pattern is successfully aligned to the query sequence, \ncorresponding elements of the N × N contact matrix and their symmetric counterparts were \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.08.674800doi: bioRxiv preprint \n\n5\nincremented by 1. This process was repeated for all patterns in the domain or sequence-specific \npattern library, cumulatively building the contact map through pattern alignment on the query \nsequence.\nBenchmarking and Validation\nA two-tier benchmarking strategy was employed:\n1. Curated Domains (25 InterPro Domains):\nFor each of the 25 protein domains in our benchmark set (Table 1), we compared mean contact \nmaps generated by our pattern matching approach with mean reference contact maps derived from \nexperimentally determined structures.To address the substantial size difference between prediction \ndatasets and experimental datasets, we employed a bootstrap resampling framework with 1,000 \niterations. Multiple sequence alignments were generated using Clustal Omega8 to establish \ncorrespondence between predicted and experimental structures.\nThe statistical analysis implemented a dual-metric approach evaluating both full alignments and \nfiltered regions, with Pearson correlation coefficients and bias metrics including mean absolute error \nand systematic deviation calculated across all aligned positions for full alignment analysis. In parallel, \na filtering strategy addressed alignment quality differences by eliminating MSA columns with high \ngap content, calculating per-position coverage for both datasets and applying an intersection \nstrategy to identify positions where both achieved at least 30% coverage for unbiased comparison. \nStatistical significance was established through bootstrap distributions enabling calculation of 95% \nconfidence intervals and paired t-tests, with p-values computed for both comparison scenarios. The \nanalysis extended to distance-dependent characterization examining sequence separations of 1, 5, \n10, 15, 20, and 25 residues, and regional classification grouping contacts into short-range (|i-j| < 12), \nmedium-range (12 ≤ |i-j| < 24), and long-range (|i-j| ≥ 24) categories. Performance evaluation \nencompassed comprehensive classification metrics including ROC curves with AUC values, precision-\nrecall curves, and F1 scores at optimal thresholds, computed for both full and filtered analyses, \nproviding statistically robust evidence for comparing datasets of vastly different sizes while \naccounting for varying data quality.\n2. Poorly Annotated Sequences (7,599 Unreviewed Proteins):\nPredicted contact maps were validated against high-confidence AlphaFold models. Standard \nclassification metrics were computed for each sequence, including accuracy, precision, recall, \nspecificity, F1-score, and Matthews Correlation Coefficient (MCC). Population-level performance was \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.08.674800doi: bioRxiv preprint \n\n6\nsummarized by mean and standard deviation across all sequences, and best/worst performers were \nidentified to illustrate range of applicability.\nThis dual-validation framework allows robust assessment of method performance across datasets \nwith vastly different structural coverage, ensuring applicability to both well-characterized protein \nfamilies and poorly annotated sequences with limited structural information.\n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.08.674800doi: bioRxiv preprint \n\n7\nResults\nApplication of the pattern-matching approach yielded contact maps that closely recapitulated the \ncharacteristic residue–residue networks of known protein folds. Across diverse protein families, the \nmethod reliably recovered conserved interaction patterns that could be directly compared with \nexperimentally determined contact maps. This approach builds upon the successful framework \ndemonstrated by Bradley, Kim, and Berger (2002)9 for identifying structural motifs that represent \nspecific protein fold characteristics. As illustrated in Figure 1, this procedure reconstructs contact \nnetworks that align well with structural reference data.\nFigure 1: Example of a contact pattern in the Grb2 SH2 domain.\nFour residues (Ala70, Leu74, Leu84, and Ser98; shown as purple sticks) are located within 8.0 Å of each other, illustrating \nhow spatially proximal residues are defined as a contact pattern (PDB ID: 6ICG)10). Contact patterns, defined as clusters of \nup to five residues within 8.0 Å, are aligned to query sequences. The example shows a sample pattern \n\"AeenLskqrhdgafLireresapgdfslS\" aligning against a query sequence from the human P56-LCK Tyrosine Kinase SH2 domain \n(PDB ID: 1LKK)11. Uppercase letters denote residues involved in contacts, lowercase letters represent intervening residues. \nThe alignment begins at position 16 of the query sequence, with pattern residues A, L, L, and S corresponding to positions \n16, 21, 31, and 45 in the query sequence. When a pattern aligns to a sequence, contacts are mapped to the corresponding \npositions in the contact matrix (C(16,21)=1, C(16,31)=1, C(16,45)=1, C(21,31)=1, C(21,45)=1, C(31,45)=1, along with \nsymmetric counterparts). This process is repeated for all patterns in the library to generate the full predicted contact map. \nTo assess the performance of our approach, we applied it to two complementary datasets. \nThe primary benchmark consisted of 25 well-characterized protein domains from InterPro4(Table 1). \nThese domains are supported by well-curated sets of  PDB structures, often capturing distinct \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.08.674800doi: bioRxiv preprint \n\n8\nconformational states, and thus provide a stringent test set with comprehensive structural \nannotation.\nAs a second benchmark, we assembled a dataset of sequences from UniProt’s5 TrEMBL6 \ncollection. Unlike the curated InterPro domains, these proteins lack extensive experimental \nannotation. This dataset therefore evaluates the applicability of the method under conditions where \nstructural information is sparse, reflecting a more realistic scenario for newly identified proteins.\nPerformance on well-characterized protein domains\nWe first benchmarked our approach on 25 InterPro domains representing diverse architectures and \nsupported by abundant structural data (Table 1). These domains span 20–250 residues and cover a \nwide range of secondary structure compositions, from loop-rich EGF-like domains (76.6% loops) to \nhighly ordered spectrin repeats (9.0% loops). The large number of available experimental structures \nper domain (on average 473, up to 2,047 for the ubiquitin family) provided a robust basis for \nstatistical evaluation.\nAcross all 25 families, the pattern-matching method showed robust performance, with \ncorrelation coefficients between predicted and reference contact maps ranging from 0.735 to 0.942. \nClassification accuracy was consistently high, with ROC AUC scores between 0.865 and 0.995, and F1 \nscores between 0.664 and 0.931. Thus, even in families with more modest correlations, \ndiscrimination between true and false contacts remained excellent (AUC >0.9 in all cases).\nFamily-specific differences were observed. Top-performing domains such as PF01023 (S-\n100), PF00505, PF00435 (Spectrin repeat), PF00001 (7TM receptor), and PF00249 consistently \nachieved correlations above 0.92 and F1 scores above 0.91, reflecting strong evolutionary \nconstraints. By contrast, families such as PF00089 (Trypsin) showed lower correlations (0.735) but \nstill maintained strong classification accuracy (AUC = 0.910).\nSystematic biases in contact probability calibration varied by family, with small domains (25–\n30 residues; e.g., PF00400, PF00096, PF00036) tending to under-predict contacts, while larger \ndomains (83–250 residues; e.g., PF00089, PF00085, PF00017) showed slight over-prediction. These \ntrends suggest that structural constraints and domain size jointly shape predictive behavior.\nContact range analysis revealed that short-range contacts (≤5 residues) were predicted most \naccurately (r = 0.772–0.929), while medium-range (5–15 residues) and long-range (>15 residues) \ncontacts showed greater variability across families. Notably, the EF-hand domain (PF00036) achieved \nexceptional long-range prediction (r = 0.946), whereas the ligand-gated ion channel domain \n(PF00060) performed less well (r = 0.424).\nSecondary structure composition strongly influenced predictive accuracy. Domains with \nhigher loop content consistently exhibited reduced correlations: the loop-rich EGF-like domain \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.08.674800doi: bioRxiv preprint \n\n9\n(PF00008, 76.6% loops) reached r = 0.836, below the overall mean of 0.867. In contrast, the spectrin \nrepeat (PF00435, 9.0% loops) achieved the highest correlation (r = 0.932). By contrast, domain size \nitself showed no significant correlation with accuracy, indicating that loop content, rather than \nlength, is the primary determinant of performance.\nRepresentative examples illustrate the range of outcomes. The S-100 domain (PF01023; r = \n0.942, AUC = 0.972; Figure 2) exemplifies high accuracy, the RNA recognition motif (PF00076; r = \n0.872, AUC = 0.988; Figure 3) shows intermediate performance, and trypsin (PF00089; r = 0.735, AUC \n= 0.911; Figure 4) represents lower correlations but still strong classification. Additional examples \nare provided in Figures S1–S50, with detailed metrics in Supplementary Text Files S1–S26.\nOverall, these results demonstrate that the pattern-matching approach achieves highly \nreliable contact prediction across structurally diverse protein families, with performance primarily \ninfluenced by secondary structure composition rather than domain length.\nTable 1. Structural composition and contact prediction performance for 25 well-characterized protein \ndomains\nDomain \nID \nDomain Name\nNumber of \nStructures \nAverage \nloop % \nAverage \nalpha helix % \nAverage \nbeta sheet \n% \nAverage \ndomain size\nNumber of \ncollected\nsequences\nBootstrap \nCorrelation\nROC AUC\nPF01023 S-100 449 34.57 64.44 0.11 43.5 2,275 0.942 ± 0.001 0.9722\nPF00505\nHMG (high \nmobility group) \nbox\n128 26.44 56.7 0 58.6 13,319 0.938 ± 0.006 0.9834\nPF00435 Spectrin repeat 43 9 45.72 0.16 105.6 10,983 0.932 ± 0.008 0.9946\nPF00001\n7 \ntransmembrane \nreceptor \n(rhodopsin \nfamily)\n170 20.28 73.03 1.58 257.4 16,565 0.928 ± 0.002 0.9327\nPF00249\nMyb-like DNA-\nbinding\n73 25.1 45.97 0 40.7 11,506 0.924 ± 0.007 0.9729\nPF00240 Ubiquitin family 2,047 42.74 20.47 21.94 67.9 8,195 0.903 ± 0.004 0.9907\nPF00060\nLigand-gated ion \nchannel\n623 29.32 42.39 4.45 268.9 9,743 0.902 ± 0.003 0.9436\nPF00036 EF hand 213 35.82 62.61 0 26.5 3,900 0.901 ± 0.002 0.8752\nPF00023 Ankyrin repeat 627 41.16 50.07 0.29 31.3 12,582 0.893 ± 0.002 0.9222\nPF00373\nFERM central \ndomain\n183 38.12 45.89 12.73 104.5 7,525 0.892 ± 0.012 0.9936\nPF00313\n‘Cold-shock’ \nDNA-binding\n123 49.09 4.88 40.96 58.3 17,078 0.890 ± 0.003 0.9864\nPF00085 Thioredoxin 679 37.38 37.35 22.35 95.4 10,868 0.875 ± 0.003 0.9884\n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.08.674800doi: bioRxiv preprint \n\n10\nPF00076\nRNA recognition \nmotif\n853 42.53 26.47 21.78 60.2 11,651 0.872 ± 0.012 0.9876\nPF00017 SH2 587 44.78 23.05 28.78 69.1 13,676 0.872 ± 0.005 0.9862\nPF00018 SH3 365 49.41 4.09 35.49 42.7 8,770 0.858 ± 0.003 0.9902\nPF00595 PDZ 416 52.6 15 26.82 73.9 12,265 0.850 ± 0.005 0.9787\nPF00186 DHFR 1,029 47.85 22.46 29.23 159 8,825 0.844 ± 0.001 0.9738\nPF00008 EGF-like domain 317 76.6 4.66 11.82 16.7 19,808 0.836 ± 0.003 0.9362\nPF00069 Protein kinase 144 44.14 36.52 16.37 260.5 14,752 0.832 ± 0.007 0.9132\nPF00270\nDEAD/DEAH box \nhelicase\n413 36.88 43.76 14.66 147.3 11,339 0.831 ± 0.002 0.9666\nPF00041\nFibronectin type \nIII \n464 51.3 1.01 42.33 81.1 11,283 0.816 ± 0.007 0.9714\nPF00096\nZinc finger, C2H2 \ntype\n145 49.02 40.99 4.31 22.8 10,241 0.813 ± 0.005 0.878\nPF00481\nProtein \nphosphatase 2C\n109 33.68 32.67 22.61 195 8,497 0.809 ± 0.006 0.8649\nPF00400\nWD domain, G-\nbeta repeat\n913 41.23 0.97 53.41 34.2 11,623 0.797 ± 0.008 0.8869\nPF00089 Trypsin 717 57.07 7.82 34.54 210.4 12,664 0.735 ± 0.003 0.9105\nFigure 2: Example of a high-quality contact prediction for protein family PF01023 (S-100 domain).\nThis case illustrates a very good prediction outcome. (A) Mean predicted contact map generated by our pattern-matching \napproach, showing contact probabilities between residue pairs (yellow: >0.8 probability; dark: low or no probability). (B) \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.08.674800doi: bioRxiv preprint \n\n11\nExperimental reference contact map derived from crystal structures for comparison. (C) Correlation analysis between \npredicted and experimental contact probabilities (r = 0.943). The red line shows the linear fit; the dashed black line \nindicates perfect correlation (y = x). (D) Comparison between full alignment analysis (blue, r = 0.552) and filtered high-\ncoverage regions (red, r = 0.942), highlighting a 70.7% improvement with filtering. (E) ROC curve with area under the curve \n(AUC) of 0.972, indicating near-perfect discrimination between true and non-contacts. The dashed diagonal indicates \nrandom performance. (F) Summary of key metrics: best F1-score (0.931), ROC AUC (0.972), filtered correlation (0.943), and \nfull alignment correlation (0.552).\nFigure 3: Example of a moderate contact prediction for protein family PF00076 (RNA recognition motif).\nThis case illustrates a typical outcome for more challenging domains. (A) Predicted contact map from our pattern-matching \napproach, showing residue–residue contact probabilities (yellow/red: 0.4–0.8 probability; dark: low or no probability). (B) \nExperimental reference contact map derived from crystal structures, displaying a sparser contact distribution compared to \nwell-conserved domains. (C) Correlation between predicted and experimental contact probabilities (r = 0.886). (D) \nComparison of full alignment analysis (blue, r = 0.626) versus filtered high-coverage regions (red, r = 0.872), showing a \n39.5% improvement with filtering. (E) ROC curve with area under the curve (AUC) of 0.988, indicating strong discriminative \npower even for this challenging family. (F) Summary of key metrics: best F1-score (0.817), ROC AUC (0.988), filtered \ncorrelation (0.872), and full alignment correlation (0.626).\n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.08.674800doi: bioRxiv preprint \n\n12\nFigure 4: Example of a lower-quality contact prediction for protein family PF00089 (Trypsin).\nThis case illustrates a poorer outcome, typical for highly diverse protein families. (A) Predicted contact map from our \npattern-matching approach, showing residue–residue contact probabilities (yellow/red: 0.4–0.8; dark: low or none). (B) \nExperimental reference contact map derived from crystal structures. (C) Correlation between predicted and experimental \ncontact probabilities (r = 0.738). The red line shows the linear fit; the dashed black line indicates perfect correlation (y = x). \n(D) Comparison of full alignment analysis (blue, r = 0.417) versus filtered high-coverage regions (red, r = 0.735), showing a \n76.2% improvement with filtering. (E) ROC curve with area under the curve (AUC) of 0.911, indicating limited but still \nsignificant discriminative ability. The dashed diagonal indicates random performance. (F) Summary of key metrics: best F1-\nscore (0.664), ROC AUC (0.911), filtered correlation (0.735), and full alignment correlation (0.417).\nPerformance on poorly annotated sequences\nTo test the generalizability of our approach, we next evaluated 7,599 poorly annotated sequences, \nwhich lack curated experimental annotation and can only be represented by reference AlphaFold \nmodels. The method maintained robust predictive performance: over 90% of sequences achieved \naccuracies above 0.90 (Fig. 5A), and precision–recall analyses confirmed balanced classification with \nrelatively few outliers (Fig. 5B–C). Additional measures, including specificity and Matthews \nCorrelation Coefficient (MCC), consistently supported the reliability of contact predictions across \ndiverse protein families (Fig. 5D–F).\nPrediction accuracy was strongly affected by secondary structure content. Proteins \ndominated by well-ordered helices or sheets showed near-perfect predictions—for example, \nA2AIM4, with only 1% loop residues, achieved accuracy = 0.995 and MCC = 0.919. By contrast, highly \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.08.674800doi: bioRxiv preprint \n\n13\ndisordered proteins performed more poorly: Q7M4S4, composed entirely of loops, reached only \n0.550 accuracy and MCC = 0.379. Representative examples of high-, medium-, and low-accuracy \ncases are shown in Figures 6–8. Overall, these results highlight that loop-rich or intrinsically \ndisordered regions are the main challenge for contact prediction, whereas structured domains yield \nexcellent agreement with reference models.\nWe also examined whether the number of structural homologs in the PDB influences \nprediction accuracy. Surprisingly, performance showed a negative correlation on the number of PDB \nhit counts with MCC and F1 scores, the correlation values are -0.164 and -0.178 respectively; Fig. 9B–\nC). Grouping sequences by homolog abundance revealed slightly declining MCC and F1 scores as hit \ncounts increased (Fig. 9D). Thus, adding more distant homologs does not improve performance and \nmay in fact dilute the signal. Tables 3 and 4 list the best and worst performers, further illustrating \nthat accuracy is not driven by homolog availability.\nThese findings demonstrate that the pattern-matching approach remains effective for poorly \nannotated proteins, even when only predicted structures are available. Performance is primarily \ndetermined by secondary structure composition rather than dataset curation or homolog \nabundance. The comparable results between curated domains and unreviewed sequences suggest \nthat the essential structural motifs captured by the 8.0 Å contact patterns are conserved across wide \nevolutionary distances, supporting the biological relevance of this representation for newly \ndiscovered proteins, including archaeal candidates.\n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.08.674800doi: bioRxiv preprint \n\n14\nFigure 5: Comprehensive binary classification performance analysis for 7,599 poorly annotated sequences \nusing homologous structural templates.\n(A) Accuracy distribution across all analyzed sequences, showing consistently high performance with most sequences \nachieving >95% accuracy. The narrow distribution (mean: 0.954 ± 0.036) demonstrates robust classification performance \nacross diverse protein families and sizes (20-2,017 residues). (B) Precision-recall relationship colored by F1-score, revealing \nthe method's characteristic high sensitivity (recall) with moderate precision. The color gradient from blue to yellow \nindicates F1-scores ranging from 0.10 to 0.90, with the majority of sequences achieving F1-scores between 0.5-0.8. The \nbroad recall range (0.2-1.0) reflects varying contact density across different protein architectures. (C) F1-score distribution \nshowing a right-skewed distribution with mean performance of 0.609 ± 0.095, indicating that most sequences achieve \nbalanced precision-recall performance with relatively few poorly performing outliers. (D) Matthews Correlation Coefficient \n(MCC) distribution demonstrating substantial agreement between predicted and AlphaFold reference contact maps, with \nmean MCC of 0.617 ± 0.086. The distribution peak around 0.6-0.7 indicates reliable binary classification performance \nacross the dataset. (E) Recall versus specificity analysis showing the method's high true positive rate (mean recall: 0.820) \ncombined with excellent specificity (mean: 0.959), indicating effective identification of true contacts with minimal false \npositive rates. The concentrated distribution in the upper-right quadrant demonstrates consistent performance across \ndiverse protein families. (F) Summary of average performance metrics across all sequences, with MCC values normalized \nfor comparison. The high specificity (0.959) and accuracy (0.954) demonstrate the method's reliability for practical contact \nprediction applications, while the moderate precision (0.512) reflects the challenging nature of contact prediction for \ndiverse protein families using homologous templates. The balanced F1-score (0.609) and substantial MCC (0.617) indicate \nmeaningful structural information extraction across the entire dataset of poorly annotated sequences.\n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.08.674800doi: bioRxiv preprint \n\n15\nTable 2. Structural composition and prediction performance for poorly annotated sequences\nValue\nNumber of sequences analysed 7,599\nAverage loop % 43.11\nAverage alpha helix % 39.16\nAverage beta sheet % 17.74\nAverage protein size 338.93\nMean Accuracy 0.954 ± 0.036\nMean Precision 0.512 ± 0.135\nMean Recall 0.820 ± 0.141\nMean Specificity 0.959 ± 0.041\nMean F1-score 0.609 ± 0.095\nMean MCC 0.617 ± 0.086\nFigure 6: Comparison of predicted and AlphaFold contact maps for protein A0A3S5H5C4 (60S ribosomal \nprotein L10 from Leishmania donovani).\n(A) Predicted contact map. (B) Reference contact map derived from the AlphaFold structure. The prediction shows high \naccuracy (96.8%) and specificity (98.8%), with moderate precision (69.8%) and recall (57.2%), yielding an F1-score of 0.629 \nand MCC of 0.615. Performance corresponds to 1,239 true positives, 536 false positives, 42,666 true negatives, and 928 \nfalse negatives out of 45,369 residue pairs. (C) AlphaFold structure shown in cartoon representation; the loop content of \nthis protein is 50.7%.\n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.08.674800doi: bioRxiv preprint \n\n16\nFigure 7: Comparison of predicted and AlphaFold contact maps for protein Q9PS57 (Glutathione transferase \nisoenzyme III from Bufo bufo).\n(A) Predicted contact map. (B) Reference contact map derived from the AlphaFold structure. The prediction shows an \naccuracy of 87.5% and specificity of 89.2%, with moderate precision (68.8%) and recall (81.5%), yielding an F1-score of \n0.746 and MCC of 0.668. Performance corresponds to 238 true positives, 108 false positives, 896 true negatives, and 54 \nfalse negatives out of 1,296 residue pairs. (C) AlphaFold structure shown in cartoon representation; the loop content of this \nprotein is 44.4%.\nFigure 8: Comparison of predicted and AlphaFold contact maps for protein M0R1V7 (Ubiquitin A-52 residue \nribosomal protein fusion product 1 from Homo sapiens).\n(A) Predicted contact map. (B) Reference contact map derived from the AlphaFold structure. The prediction shows an \naccuracy of 41.7% and specificity of 32.0%, with low precision (19.6%) and recall (1.0%), yielding an F1-score of 0.328 and \nMCC of 0.251. Performance corresponds to 565 true positives, 2,314 false positives, 1,090 true negatives, and 0 false \nnegatives out of 3,969 residue pairs. (C) AlphaFold structure shown in cartoon representation; the loop content of this \nprotein is 46.0%.\n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.08.674800doi: bioRxiv preprint \n\n17\nFigure 9: Performance Analysis of Protein Function Prediction Model Relative to PDB Hit Lengths.\n(A) Distribution of PDB hit counts. (B) Scatter plot revealing weak negative correlation (r = -0.164) between PDB hit count \nand MCC. (C) Similar negative correlation (r = -0.178) between PDB hit count and F1-score. (D) Mean MCC values and F1-\nscores are grouped by PDB hit counts. \nTable 3. Top 10 sequences ranked by MCC and F1-score\nTop 10 performers MCC F1-score Accuracy PDB Hit count\nA0A151DYG7 0.927 0.939 0.978 83\nA7XZE4 0.923 0.925 0.995 62\nA0A4X1VW25 0.92 0.923 0.995 62\nA2AIM4 0.919 0.922 0.995 62\nA0A1Y7VNT9 0.911 0.933 0.964 288\nD0VX28 0.895 0.907 0.978 68\nQ6CM20 0.892 0.897 0.99 249\nA0A140LIU3 0.886 0.906 0.963 68\nQ6CPN9 0.885 0.893 0.985 241\nA0A804RMZ3 0.878 0.881 0.992 89\nTable 4. Bottom 10 sequences ranked by MCC and F1-score\n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.08.674800doi: bioRxiv preprint \n\n18\nBottom 10 Performers MCC F1-score Accuracy PDB Hit count\nQ9PST8 0.166 0.277 0.285 890\n F5GZ39 0.186 0.297 0.314 872\nM0R1V7 0.251 0.328 0.417 881\nA0A0P0W7F3 0.299 0.31 0.606 807\nC4M760 0.309 0.343 0.528 898\nQ9UMG3 0.346 0.528 0.515 177\nA0A498MMR3 0.351 0.232 0.991 201\nQ0JBH4 0.365 0.364 0.695 856\nD3YVJ8 0.372 0.36 0.716 894\nQ7M4S4 0.379 0.545 0.55 612\n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.08.674800doi: bioRxiv preprint \n\n19\nDiscussion\nUnderstanding the protein structure is essential for explaining the biological function, yet the gap \nbetween sequenced proteins and experimentally determined structures continues to widen. In this \nstudy, we developed a pattern-matching approach that uses existing structural data to rapidly \ngenerate contact maps from sequence information alone. By validating the method on both well-\ncharacterized protein domains and poorly annotated sequences, we demonstrate its versatility, \nrobustness, and practical utility across a wide range of proteins.\nOur approach accurately predicts contacts across diverse protein families, capturing both \nshort- and long-range interactions. Prediction performance is influenced by protein architecture, \nsecondary structure composition, and evolutionary constraints, with well-structured domains \nyielding the most reliable results and loop-rich or intrinsically disordered proteins remain more \nchallenging, highlighting the method’s dependence on conserved structural motifs for optimal \naccuracy.\nThe method also performs robustly for poorly annotated sequences, using distant homologs \nto generate meaningful contact predictions. This demonstrates its applicability to metagenomic \ndatasets and proteins from organisms with limited structural coverage. Examples of such \nuncharacterized proteins include archaeal sequences, where our predictions show good agreement \nwith AlphaFold reference structures (supplementary contact maps provided). Interestingly, \nincreasing the number of structural homologs does not necessarily improve prediction quality, \nsuggesting that a modest set of representative templates are sufficient to capture essential \nstructural patterns.\nA key advantage of our approach is its computational efficiency and interpretability, \nrequiring minimal resources while revealing which structural motifs drive contact predictions. By \nintegrating patterns from multiple structural states, the method inherently captures conformational \nflexibility, providing insights into dynamic protein behavior and allosteric mechanisms that would \notherwise require extensive molecular dynamics simulations. This capability enables practical \napplications in functional annotation and drug discovery, allowing researchers to identify potential \nbinding sites, infer allosteric pathways with additional methods, and guide structural \ncharacterization for novel proteins, even when experimental structures are unavailable.\nDespite these advantages, the approach has limitations. Prediction accuracy is reduced for \nproteins dominated by loops or disorder, and the current 8.0 Å contact cutoff may not capture all \nfunctionally relevant long-range interactions in very large proteins. Future improvements could \ninvolve adaptive cutoffs or expanded pattern libraries to enhance coverage and accuracy.\n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.08.674800doi: bioRxiv preprint \n\n20\nOur study demonstrates that pattern-based contact map prediction provides a practical \nalternative to full structure prediction, bridging the gap between sequence and structure in a \ncomputationally efficient manner. As structural databases continue to expand, this approach can be \nreadily adapted for exploratory studies in emerging protein families. \nAvailability of the code\nThe data that support the findings of this study are openly available in the Zenodo \nrepository with the link https://zenodo.org/records/17043595.\nAcknowledgments\nThis work was supported by the Swiss National Science Foundation (Grant 310030_219549 \nto D.F.). We thank the Division de Calcul et Soutien à la Recherche of the UNIL for access to \nthe university’s computer infrastructure. We thank all members of the Fasshauer Laboratory \nfor helpful discussions. \nAuthor contributions\nA.H. designed the study, performed the experiments, and analyzed the data; A.H. and D.F. \nwrote the paper.\nCompeting interests\nThe authors declare no competing interests.\nReferences\n(1) Jumper, J.; Evans, R.; Pritzel, A.; Green, T.; Figurnov, M.; Ronneberger, O.; \nTunyasuvunakool, K.; Bates, R.; Žídek, A.; Potapenko, A. Highly accurate protein \nstructure prediction with AlphaFold. Nature 2021, 596 (7873), 583-589.\n(2) Baek, M.; DiMaio, F.; Anishchenko, I.; Dauparas, J.; Ovchinnikov, S.; Lee, G. R.; \nWang, J.; Cong, Q.; Kinch, L. N.; Schaeffer, R. D. Accurate prediction of protein \nstructures and interactions using a three-track neural network. Science 2021, 373 \n(6557), 871-876.\n(3) Lin, Z.; Akin, H.; Rao, R.; Hie, B.; Zhu, Z.; Lu, W.; Smetanin, N.; Verkuil, R.; \nKabeli, O.; Shmueli, Y. Evolutionary-scale prediction of atomic-level protein structure \nwith a language model. Science 2023, 379 (6637), 1123-1130.\n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.08.674800doi: bioRxiv preprint \n\n21\n(4) Blum, M.; Andreeva, A.; Florentino, L. C.; Chuguransky, S. R.; Grego, T.; Hobbs, \nE.; Pinto, B. L.; Orr, A.; Paysan-Lafosse, T.; Ponamareva, I. InterPro: the protein \nsequence classification resource in 2025<? mode longmeta?>. Nucleic Acids \nResearch 2025, 53 (D1), D444-D456.\n(5) Consortium, U. UniProt: the universal protein knowledgebase in 2023. Nucleic \nAcids Research 2023, 51 (D1), D523-D531.\n(6) Bairoch, A.; Apweiler, R. The SWISS-PROT protein sequence database and its \nsupplement TrEMBL in 2000. Nucleic acids research 2000, 28 (1), 45-48.\n(7) Johnson, L. S.; Eddy, S. R.; Portugaly, E. Hidden Markov model speed heuristic \nand iterative HMM search procedure. BMC bioinformatics 2010, 11, 1-8.\n(8) Sievers, F.; Wilm, A.; Dineen, D.; Gibson, T. J.; Karplus, K.; Li, W.; Lopez, R.; \nMcWilliam, H.; Remmert, M.; Söding, J. Fast, scalable generation of high‐quality \nprotein multiple sequence alignments using Clustal Omega. Molecular systems \nbiology 2011, 7 (1).\n(9) Bradley, P.; Kim, P. S.; Berger, B. TRILOGY: Discovery of sequence-structure \npatterns across diverse proteins. In Proceedings of the sixth annual international \nconference on Computational biology, 2002; pp 77-88.\n(10) Hosoe, Y.; Numoto, N.; Inaba, S.; Ogawa, S.; Morii, H.; Abe, R.; Ito, N.; Oda, M. \nStructural and functional properties of Grb2 SH2 dimer in CD28 binding. Biophysics \nand physicobiology 2019, 16, 80-88.\n(11) Tong, L.; Warren, T. C.; King, J.; Betageri, R.; Rose, J.; Jakes, S. Crystal \nstructures of the human p56lckSH2 domain in complex with two short \nphosphotyrosyl peptides at 1.0 Å and 1.8 Å resolution. Journal of molecular biology \n1996, 256 (3), 601-610.\n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted September 15, 2025. ; https://doi.org/10.1101/2025.09.08.674800doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}