Hidden Markov Models for Protein Domain Annotation

Introduction

Protein domain annotation is a foundational task in functional genomics. Domains are conserved structural and functional units that recur across diverse protein families [1]. Hidden Markov models (HMMs) provide a probabilistic framework for representing the sequence variability within domain families [2]. Profile HMMs, in particular, capture position-specific amino acid substitution probabilities, insertion, and deletion states, enabling sensitive detection of remote homologs [1, 3]. This article reviews the theoretical underpinnings, major database resources, advanced fitting methods, and applications of HMMs for protein domain annotation, with an emphasis on veterinary and pathogen proteomes.

The fundamental architecture of a profile HMM consists of match states (M), insert states (I), and delete states (D) arranged in a linear chain [1]. Each match state emits amino acids with position-specific probabilities derived from a curated multiple sequence alignment (MSA) [4, 5]. Transition probabilities between states model the likelihood of insertions and deletions relative to the consensus. The Viterbi algorithm and forward-backward algorithms are used to align a query sequence to the HMM and to compute the probability of the sequence given the model [2, 6]. Recent vectorized implementations of the Viterbi algorithm, as in HH-suite, have reduced runtime by a factor of 4 compared to earlier versions [6].

Principal HMM Database Resources

Several public databases maintain libraries of profile HMMs for protein domain annotation (Table 1).

Table 1. Major HMM-based protein domain databases.

Pfam is the most widely used resource, containing thousands of domain families each represented by a profile HMM built from a seed alignment [4, 7]. SUPERFAMILY provides assignments at the SCOP superfamily level, linking sequence to structure and function [8]. TIGRFAMs specializes in bacterial and archaeal subsystems, with equivalog models that permit precise functional annotation transfer [9]. VIRify offers a dedicated collection of viral profile HMMs for detecting and classifying viral sequences in metagenomic data [10].

HMM Construction and Parameterization

A profile HMM is constructed from a multiple sequence alignment of representative members of a domain family [4, 5]. The alignment defines the length and conserved columns of the domain. The HMM parameters consist of emission probabilities for each match state (20 amino acid frequencies) and transition probabilities among M, I, and D states [1]. These probabilities are typically estimated using maximum likelihood or Bayesian methods with pseudocounts to avoid zero frequencies [2].

The HMMER software suite implements profile HMM construction and search [7, 11]. The latest HMMER web server provides an updated interface for sequence searches against HMM libraries [11]. For remote homology detection, HH-suite implements HMM-HMM alignment, which is more sensitive than sequence-HMM comparisons [6]. HHsearch and HHblits use pairwise alignment of profile HMMs to detect distant evolutionary relationships [6].

Advanced Methods for Divergent Organisms

Standard profile HMMs may lack sensitivity when applied to organisms with biased amino acid composition or highly divergent sequences [12]. For example, Plasmodium falciparum, the causative agent of malaria in humans and a model for apicomplexan parasites, has an AT-rich genome and biased amino acid usage that reduces HMM detection efficiency [12, 13]. Several strategies have been developed to address this limitation.

Terrapon et al. compared approaches for fitting HMMs to a target species [12]. They proposed two methods that learn global correction rules to adjust match state emission distributions for the entire HMM library, enabling detection of domain families previously absent in the target proteome [12]. A complementary method, Co-Occurrence Domain Discovery (CODD), exploits the tendency of domains to co-occur with specific partner domains [14, 13]. Using HMM-HMM comparisons combined with CODD, Ghouila et al. identified 901 and 1098 new domain occurrences in Plasmodium and Leishmania proteomes, respectively, at an estimated false discovery rate of 5% [14].

Kingdom-specific HMM libraries offer another route to improved coverage. Alam et al. constructed Fungal Pfam (FPfam) using sequences from 30 fungal genomes and demonstrated increased coverage and higher bitscores compared to general Pfam models [15]. They recommended using both general and kingdom-specific libraries together for optimal annotation [15].

Stratified Statistics and Error Rate Control

Domain annotation involves multiple hypothesis testing across many query proteins and many HMMs. E-values have traditionally been used to assess the significance of matches [16]. Ochoa et al. formally showed that stratifying statistical tests by domain family and controlling the local false discovery rate (lFDR) per stratum yields the most predictions at a given overall FDR threshold [16]. They developed the first FDR-estimating algorithms for domain prediction and demonstrated that stratified q-value thresholds substantially outperform E-values [16].

Applications in Veterinary and Pathogen Proteomics

HMM-based domain annotation has direct applications in veterinary medicine, particularly for characterizing proteins from animal pathogens.

Circumsporozoite protein (CSP) of Plasmodium species. Parikesit et al. used SUPERFAMILY HMMs to annotate domains in CSP from P. vivax, P. malariae, and P. knowlesi [17]. They identified a single C-terminal domain belonging to the TSP-1 type 1 repeat family with high reliability, suggesting a conserved functional role in sporozoite invasion of hepatocytes [17]. Sequence clustering indicated that P. knowlesi and P. vivax CSP are more similar to each other than to P. malariae, potentially reflecting different infection modes [17]. CSP is a promising vaccine target, and accurate domain annotation supports rational antigen design [17].

Viral metagenomics. The VIRify pipeline uses virus-specific profile HMMs to detect and classify viral contigs from metagenomic assemblies [10]. It includes manually curated HMMs targeting prokaryotic and eukaryotic viral taxa, achieving taxonomic classification accuracy exceeding 95.5% at genus and family ranks [10]. This tool is valuable for identifying novel viruses in veterinary samples, such as those from livestock or wildlife.

Fungal pathogens. Kingdom-specific HMM libraries like FPfam improve domain coverage for fungal genomes, including plant and animal pathogens such as Ustilago maydis and Aspergillus species [15]. Improved annotation of secreted proteins, cell wall modifying enzymes, and virulence factors can aid in understanding pathogenicity mechanisms and identifying drug targets [15].

Glycoside hydrolases. Rossi et al. evaluated the performance of HMM profiles for classifying glycoside hydrolases (GHs) according to the CAZy standard [18]. Their meta-analysis showed that HMM profiles recovered 65% of matches to CAZy families, compared to 61% using dbCAN HMMs, indicating that HMM-based classification is useful but requires further refinement for fully automated annotation [18].

G-protein coupled receptors (GPCRs). Kostiou et al. developed GprotPRED, a profile HMM-based method for annotating Gα, Gβ, and Gγ subunits of G-proteins, and applied it to proteomes [19]. Opsin-specific HMMs have been used to survey the taxonomic distribution of opsins across 260 metazoan species, providing insight into the evolution of vision and circadian regulation [20].

Integration with Other Computational Methods

HMMs are often combined with other approaches to improve domain detection. Interaction profile HMMs model contact patterns between residues within domains, enabling more accurate prediction of interaction interfaces [21]. Partial label training algorithms, such as that developed by Li et al., allow HMMs to be trained using sequences with incomplete annotation, which is common in large-scale functional genomics [22]. HMMeta uses a separate HMM for each Gene Ontology term, applying data augmentation to balance sparse functional classes [23].

Machine learning classifiers trained on HMM score vectors have been employed for remote homology detection [24]. Fisher kernel methods and HMM combining scores (e.g., independent and dependent probability scores) have been shown to improve discrimination between homologous and non-homologous sequences [24]. For detecting remote homologs, the TEDLH approach uses domain HMMs with a novel scoring scheme that integrates secondary structure information [25].

The HMM framework also supports the prediction of complex domain architectures. Uricaru et al. introduced a new type of HMM that models the sequential arrangement of multiple domains within a protein, allowing simultaneous prediction of domain composition and order [26].

Workflow for HMM-Based Domain Annotation

The typical workflow for annotating protein domains using HMMs is depicted in Figure 1.

flowchart TD
    A[Query protein sequence], > B[Collect curated MSAs of domain families]
    B, > C[Build profile HMMs for each family]
    C, > D[Search HMM library against query using HMMER or HH-suite]
    D, > E[Apply significance thresholds (E-value, q-value, lFDR)]
    E, > F{Significant match?}
    F, Yes, > G[Assign domain family to query position]
    F, No, > H[Optional: use HMM-HMM or co-occurrence methods]
    H, > I[Re-evaluate with adjusted thresholds]
    I, > G
    G, > J[Consolidate domain architectures]
    J, > K[Functional annotation transfer]

Figure 1. General workflow for HMM-based protein domain annotation. Starting from a query sequence, profile HMMs constructed from curated MSAs are searched against the query. Significant matches are identified using appropriate statistical thresholds. For divergent sequences, additional methods such as HMM-HMM alignment or co-occurrence detection may be applied. Domain architectures are assembled and functional annotations transferred.

Frequently Asked Questions

What is a profile hidden Markov model in the context of protein domain annotation?

A profile HMM is a probabilistic model that represents the consensus sequence of a protein domain family, capturing position-specific amino acid preferences and allowing insertions and deletions relative to the consensus [1].

Which major databases provide HMM-based protein domain annotations?

The major databases are Pfam, SUPERFAMILY, TIGRFAMs, and VIRify, each with distinct foci and HMM construction strategies [9, 8, 4, 10].

How are HMMs built from multiple sequence alignments?

HMMs are built by extracting position-specific emission probabilities and transition probabilities from a curated MSA, typically using maximum likelihood estimation with pseudocounts [1, 2].

What are the limitations of standard HMMs for divergent organisms?

Standard HMMs may lack sensitivity due to biased amino acid composition and sequence divergence, as seen in Plasmodium falciparum [12, 13].

How can domain detection be improved for divergent species?

Methods include HMM fitting with global correction rules, HMM-HMM comparisons, co-occurrence domain discovery (CODD), and kingdom-specific HMM libraries [12, 15, 14, 6].

What statistical measures are used to assess domain annotation significance?

E-values have been traditional, but stratified q-values and local false discovery rates (lFDR) offer improved control of false positives across multiple tests [16].

Can HMMs be used to predict domain architectures?

Yes, advanced HMMs can model the arrangement of multiple domains in a protein, predicting both domain composition and order [26].

How are HMM-based domain annotations applied in veterinary virology?

They are used to annotate viral proteins, such as the circumsporozoite protein of Plasmodium species and viral contigs in metagenomics, supporting vaccine design and pathogen characterization [17, 10].

Conclusion

Hidden Markov models remain a cornerstone of protein domain annotation. The combination of profile HMMs with advanced statistical controls, HMM-HMM alignment, co-occurrence analysis, and species-specific fitting extends their utility to highly divergent proteomes, including those of veterinary pathogens. Ongoing developments in software acceleration, partial label training, and stratified error rate estimation continue to enhance the sensitivity and specificity of domain discovery. Integration with other computational methods, such as protein language models and structural prediction, promises to further advance our ability to functionally characterize proteins from any organism.

References

[1] Eddy SR. Profile hidden Markov models. Bioinform. 1998. https://www.semanticscholar.org/paper/50edb17bb311757206a60801a25dd56ca2b342dd

[2] Krogh A. Hidden Markov models in computational biology. Applications to protein modeling. Journal of Molecular Biology. 1993. https://www.semanticscholar.org/paper/5d28fc1a4027d23cc9e4ad8555361d48940e9be8

[3] Karplus K, Barrett C, Hughey R. Hidden Markov models for detecting remote protein homologies. Bioinform. 1998. https://www.semanticscholar.org/paper/af852603b9c0d14da8d650483f54ea2038fd22

[4] Bateman A, Birney E, Cerruti L, et al. The Pfam protein families database. Nucleic Acids Research. 2002. https://www.semanticscholar.org/paper/7abfe35d1dba4a12dbe177a8fbce690402cc43a5

[5] Mistry J, Finn RD. Pfam: a domain-centric method for analyzing proteins and proteomes. Methods in molecular biology. 2007. https://www.semanticscholar.org/paper/6200415c2dfd7a154c7cd7bfad7025a8a45240a8

[6] Steinegger M, Meier M, Mirdita M, et al. HH-suite3 for fast remote homology detection and deep protein annotation. BMC Bioinformatics. 2019. https://www.semanticscholar.org/paper/e1e5929fd511548d4ff081444b64da87c9c9dbe3

[7] Coggill P, Finn RD, Bateman A. Identifying Protein Domains with the Pfam Database. Current Protocols in Bioinformatics. 2008. https://www.semanticscholar.org/paper/07f810154a6a4281b0fca6a8f3b3128b1eb44617

[8] Wilson D, Pethica R, Zhou Y, et al. SUPERFAMILY, sophisticated comparative genomics, data mining, visualization and phylogeny. Nucleic Acids Res. 2008. https://www.semanticscholar.org/paper/79623f573e440084f4edb1cd040e66852671273f

[9] Haft DH, Selengut J, Richter RA, et al. TIGRFAMs and Genome Properties in 2013. Nucleic Acids Res. 2012. https://www.semanticscholar.org/paper/3ee7d214192130a46316a59c75723de936545431

[10] Rangel-Pineros G, Almeida A, Beracochea M, et al. VIRify: An integrated detection, annotation and taxonomic classification pipeline using virus-specific protein profile hidden Markov models. bioRxiv. 2022. https://www.semanticscholar.org/paper/8e105f5432de2df5d759a81a7f6c65aab1fe12c9

[11] Rajković A, Beracochea M, Rogers AB, et al. HMMER web server: 2026 update. Nucleic Acids Res. 2026. https://pubmed.ncbi.nlm.nih.gov/42037125/ *** Disclaimer: This article is for educational and informational purposes only. It is not intended to substitute for professional veterinary advice, diagnosis, treatment, or regulatory guidance. Always consult a licensed veterinarian or qualified specialist regarding animal health, disease diagnosis, and therapeutic decisions.

[12] Terrapon N, Gascuel O, Maréchal E, et al. Fitting hidden Markov models of protein domains to a target species: application to Plasmodium falciparum. BMC Bioinformatics. 2012. https://www.semanticscholar.org/paper/7f8b827ee9ab1f9bc66722824f1f85d4c1db6d2e

[13] Terrapon N. Recherche de domaines protéiques divergents à l'aide de modèles de Markov cachés : application à Plasmodium falciparum. 2010. https://www.semanticscholar.org/paper/a9554467fd9def15dbcd5ce21442b9f340c234f8

[14] Ghouila A, Florent I, Guerfali F, et al. Identification of Divergent Protein Domains by Combining HMM-HMM Comparisons and Co-Occurrence Detection. PLoS ONE. 2014. https://www.semanticscholar.org/paper/a2a54bf5d7b5df945cc359bb4a4660a2be896c41

[15] Alam I, Hubbard S, Oliver S, et al. A kingdom-specific protein domain HMM library for improved annotation of fungal genomes. BMC Genomics. 2007. https://www.semanticscholar.org/paper/6fc9e1953f0d6733ef46f5d427768272451cd97a

[16] Ochoa A, Storey JD, Llinás M, et al. Beyond the E-Value: Stratified Statistics for Protein Domain Prediction. PLoS Comput Biol. 2014. https://www.semanticscholar.org/paper/b3863b45c13153ed580b55ee052171332c410ea6

[17] Parikesit AA, Utomo D, Karimah N. Protein Domain Annotation of Plasmodium spp. Circumsporozoite Protein (CSP) Using Hidden Markov Model-based Tools. 2018. https://www.semanticscholar.org/paper/f2335cd14b70f6bf7fca1efed79688c4479a0980

[18] Rossi M, Mello B, Schrago CG. Performance of Hidden Markov Models in Recovering the Standard Classification of Glycoside Hydrolases. Evolutionary bioinformatics online. 2017. https://www.semanticscholar.org/paper/efb36d5cc4b206b0ba7923264a3e5fa79e743c7e

[19] Kostiou V, Theodoropoulou MC, Hamodrakas S. GprotPRED: Annotation of Gα, Gβ and Gγ subunits of G-proteins using profile Hidden Markov Models (pHMMs) and application to proteomes. Biochimica et Biophysica Acta. 2016. https://www.semanticscholar.org/paper/b8cf3efed64a1713eb80ac1c91265f00d51aa530

[20] Clarke N, Taylor J. Taxonomic distribution of opsin families inferred from UniProt Reference Proteomes and a suite of opsin-specific hidden Markov models. Frontiers in Ecology and Evolution. 2023. https://www.semanticscholar.org/paper/31854e68f76105954f74f9d79f720ebb6c74cc46

[21] Friedrich T, Pils B, Dandekar T, et al. Modelling interaction sites in protein domains with interaction profile hidden Markov models. Bioinform. 2006. https://www.semanticscholar.org/paper/6b55001ee3c3bd7b4bb11c7cd6ca91e8f1893825

[22] Li J, Lee JY, Liao L. A new algorithm to train hidden Markov models for biological sequences with partial labels. BMC Bioinformatics. 2021. https://www.semanticscholar.org/paper/2edacf96e8d36c80deddbc5205d431c881c272fe

[23] Gbenro S, Hippe K, Cao R. HMMeta: Protein Function Prediction using Hidden Markov Models. Proceedings of the 11th ACM International Conference on Bioinformatics, Computational Biology and Health Informatics. 2020. https://www.semanticscholar.org/paper/ccaee06dae6530cac87151b2f4b144d3facfc90f

[24] Zaki N, Deris S, Illias R. Features Extraction For Protein Homology Detection Using Hidden Markov Models Combining Scores. International Journal of Computational Intelligence and Applications. 2004. https://www.semanticscholar.org/paper/890534ef12d437912b03630b9419e55eb39c1976

[25] Alvarez-Carreño C, Petrov AS, Waman VP, et al. TEDLH: Domain HMMs for sensitive detection of remote homologues. Bioinformatics. 2026. https://pubmed.ncbi.nlm.nih.gov/42398027/

[26] Uricaru R, Bréhélin L, Rivals E. A new type of Hidden Markov Models to predict complex domain architecture in protein sequences. 2007. https://www.semanticscholar.org/paper/55a70d99e846164523f9a9343c625e24a5f25a56

[27] Jablonowski K. Hidden Markov Models for Protein Domain Homology Identification and Analysis. Methods in molecular biology. 2017. https://www.semanticscholar.org/paper/d7ca3ee0b37b8db81bf79133402fc47899670edb

[28] AbuShanab T, Al-Hmouz R. Classification of Structural Protein Domain Based on Hidden Markov Model. 2017. https://www.semanticscholar.org/paper/ea4f115de45c8b15fdf31cc3b84dd8a32c78db47

[29] Scheeff ED, Bourne P, McCammon JA. Multiple alignments of protein structures and their application to sequence annotation with hidden markov models. 2003. https://www.semanticscholar.org/paper/93be0d0aeaddccb9304ca87114e44782c2a16114

[30] Sonnhammer ELL, von Heijne G, Krogh A. A Hidden Markov Model for Predicting Transmembrane Helices in Protein Sequences. Intelligent Systems in Molecular Biology. 1998. https://www.semanticscholar.org/paper/023499b96536b1a0dcb04a48578c5ee2533237bc

[31] Brejová B, Brown DG, Vinař T. Advances in Hidden Markov Models for Sequence Annotation. 2007. https://www.semanticscholar.org/paper/66dbd58f0d376b25ab6a333ca9e8c5eb3d15f3e9

[32] Upadhyay AA. Application of Hidden Markov Model based methods for gaining insights into protein domain evolution and function. 2015. https://www.semanticscholar.org/paper/11c06eb3523854635c12b328f0b14aaffde6b4bd

[33] Wistrand M. Hidden Markov models for remote protein homology detection. 2005. https://www.semanticscholar.org/paper/ccf0d65b4ef4878e3f37e0d2259d0cd47aafa727

[34] Holbrook PG, Geetha V, Beaven MA, et al. Recognizing the Pleckstrin homology domain fold in mammalian phospholipase D using hidden Markov models. FEBS Letters. 1999. https://www.semanticscholar.org/paper/d162d7c6fa482dc8191f69fc3e64cd2d45bc5554

[35] Golod D. The k-best paths in Hidden Markov Models. Algorithms and Applications to Transmembrane Protein Topology Recognition. 2009. https://www.semanticscholar.org/paper/f4e90429e0cf9acf4cbcf3ef69575ed080ab8947