# Chapter 62. Comparative RNA Structure Inference and Covariation Evidence

## Scope Note

This chapter explains how homologous sequences are used to infer and test conserved RNA secondary structure. Its primary subjects are structure-aware alignment as an inference input, covariation statistics, compensatory substitutions as structural evidence, covariance-model principles, comparative secondary-structure inference, and inference benchmarking. RNA-family origins and the limits of evolutionary claims hand off to [Chapter 12](chapter1011.md). Database-scale homology search, Rfam curation, family classification, thresholds, record assignment, and annotation pipelines hand off to [Chapter 140](chapter1127.md). Short bridges explain where those upstream and downstream operations meet structural inference without repeating their full workflows.

## Executive Summary

RNA structure can be inferred not only from one sequence, one energy model, or one experiment, but also from evolution. If a stem in an RNA is functionally important, mutations that disrupt a base pair tend to be selected against, while paired changes that restore Watson-Crick or wobble pairing can be tolerated. A G-C pair in one species may become A-U, G-U, or C-G in another species while preserving the helix. This pattern is called covariation. When both sides of a pair change in a way that preserves pairing, the pattern is called a compensatory mutation or compensatory substitution. Comparative RNA structure methods try to convert such evolutionary patterns into secondary-structure evidence, family models, and genome annotations.

Comparative analysis begins with multiple sequence alignment. A multiple sequence alignment places homologous residues from related RNAs into columns so that sequence positions can be compared across species, strains, or paralogs. For protein-coding genes, alignment often follows codon or amino acid conservation. For structured RNAs, alignment is harder because insertions, deletions, and substitutions may preserve structure more than primary sequence. A riboswitch aptamer, an RNase P RNA domain, or a viral structured element can conserve paired columns while allowing loops and peripheral segments to vary. A good alignment for structured RNA therefore reflects both homology and plausible structural correspondence. A poor alignment can create false covariation, erase real covariation, or make a conserved structural element appear absent.

Covariation is powerful because it is a natural perturbation experiment repeated through evolution. A predicted helix supported by independent compensatory changes across many lineages is stronger evidence than a helix predicted only by a nearest-neighbor energy model. However, covariation is not simple correlation. Shared ancestry, phylogenetic structure, base composition, alignment uncertainty, sequencing errors, and repeated lineage-specific pressures can all create apparent dependencies between columns. Modern covariation analysis therefore asks whether paired columns covary more than expected after accounting for conservation, phylogeny, and alignment features. R-scape-style analysis is designed for this statistical question, although this chapter uses final reference item notes for R-scape because verified R-scape bibliographic metadata are not present in the current chapter reference file.

Covariance models provide a probabilistic representation of conserved RNA sequence and secondary structure. A profile stochastic context-free grammar contains states for paired and unpaired consensus positions, insertions, deletions, and transitions through a structural architecture. Paired states emit nucleotide pairs rather than treating the two columns independently. This makes the model useful for evaluating whether a sequence is compatible with a family-level structural hypothesis. The model's evidential value depends on the seed alignment and consensus structure; it cannot rescue a biased alignment, establish expression, or prove biological function.

Comparative methods are also central to benchmarking. Many RNA structure predictors, RNA language models, and noncoding RNA classifiers are trained or evaluated on sequences that are homologous to sequences in their test sets. If related RNAs are split randomly between training and testing, the model may appear to generalize because it recognizes family-specific signatures rather than because it has learned transferable structural principles. Homology-aware benchmarking uses family-level splits, nonredundant clustering, taxonomic separation, structure-class separation, and explicit reporting of sequence identity or covariance-model overlap. RNAconTest illustrates the need to compare noncoding RNA multiple sequence alignment tools by structural consistency rather than by sequence agreement alone (Wright 2020), and recent multiple-sequence-alignment-based RNA language modeling work shows both the promise and the benchmarking sensitivity of using aligned homologs for structure inference (Zhang Y et al. 2024).

Comparative RNA structure analysis remains most reliable when it is integrated with other evidence. Covariation can support a conserved helix; chemical probing can test whether positions are accessible in a given condition; mutational rescue can test whether pairing matters for function; high-resolution structural biology can show atomic contacts; and genetics can show phenotype. The strongest claims about RNA structure combine these evidence classes. The weakest claims overinterpret conservation, use an uncertain alignment, ignore phylogeny, or call every predicted conserved stem a functional structural element without experimental or comparative support.

## Concept Inventory

- **Multiple sequence alignment:** an inferred arrangement of homologous sequences into columns intended to represent corresponding ancestral or structural positions.
- **Structure-aware alignment:** an alignment that evaluates conserved stems, loops, motifs, insertions, deletions, and domain architecture alongside primary-sequence similarity.
- **Covariation:** coordinated variation between alignment columns; for RNA structure, the relevant signal is excess pairing-compatible change after background dependencies are considered.
- **Compensatory substitution:** change on both sides of a proposed pair that restores or preserves compatible pairing.
- **Compatible substitution:** a one-sided change that retains pairing potential, such as G-C to G-U.
- **Covariation statistic:** a numerical measure of dependence between columns, interpreted relative to a null model and the effective number of independent substitutions.
- **Consensus secondary structure:** a family-level pairing model inferred from aligned sequences and supporting evidence.
- **Covariance model:** a probabilistic model that represents sequence conservation and consensus secondary structure with paired and unpaired emission states.
- **Profile stochastic context-free grammar:** a formal model whose state topology represents nested pairing and family-specific insertions, deletions, and conserved positions.
- **Phylogenetic correction:** control for shared ancestry so closely related descendants do not count as independent substitution events.
- **Homology-aware benchmarking:** evaluation that separates related sequences or families according to the intended generalization claim.

## What to Know Before Reading This Chapter

The reader should know that RNA secondary structure is a map of intramolecular base pairs, usually including Watson-Crick G-C and A-U pairs and wobble G-U pairs. Chapters [3](chapter1003.md), [4](chapter1004.md), [60](chapter1055.md), and [61](chapter1056.md) explain the thermodynamic and probabilistic vocabulary of folding from a single sequence. This chapter adds a different source of information: homologous sequences. The central question changes from "Which structure is favorable for this sequence?" to "Which structural features have been preserved while homologous sequences diverged?"

The reader should also distinguish sequence conservation from structural conservation. Sequence conservation means that the same nucleotide or motif is retained across homologs. Structural conservation means that a base pair, helix, loop architecture, or domain organization is retained even when the exact nucleotides change. A tRNA acceptor stem, for example, can preserve helical pairing across extensive sequence variation. A protein-binding RNA motif may conserve specific nucleotides because a protein recognizes their identities, while neighboring stems conserve pairing rather than identity. Comparative RNA methods are useful precisely because they can separate these two modes of conservation.

The running examples in this chapter are tRNAs, ribosomal RNAs, riboswitches, Rfam family alignments, influenza viral RNA structure, and viral or cellular cap-independent translation elements. tRNAs illustrate a highly conserved and easily recognized RNA fold. Ribosomal RNAs illustrate deep evolutionary conservation, enormous alignments, and the historical power of comparative analysis. Riboswitches show how conserved aptamer structures can be discovered across bacteria. Influenza and other viral RNAs show the promise and difficulty of comparative structure analysis in fast-evolving genomes, where selection can preserve RNA elements but phylogeny, coding constraints, and recombination can confound interpretation.

Finally, the reader should remember that a computational method can make a strong hypothesis without proving biological function. A covariance model hit can indicate homology; a covarying pair can support a helix; a conserved helix can suggest structural constraint. None of those outputs alone proves that the RNA element regulates translation, controls decay, binds a ligand, or affects infection. Functional claims require additional evidence from genetics, biochemistry, structure probing, structural biology, or cellular assays.

## 62.1. Structure-aware multiple sequence alignment as an inference input

Multiple sequence alignment is the entry point for comparative RNA structure analysis. In its simplest form, a multiple sequence alignment places homologous nucleotides into columns. If column 20 in one bacterial RNase P RNA and column 20 in another bacterial RNase P RNA derive from the same ancestral position, comparing those columns is meaningful. If the columns have been misaligned, later calculations will compare unrelated positions and may infer spurious conservation or covariation. Alignment quality therefore limits every downstream comparative claim.

**Table 62.1. Alignment Strategies as Inputs to Structure Inference.** Alignment strategies contribute different combinations of sequence, profile, and structural information to comparative inference; poorly aligned homologous positions can create false conservation or covariance even when overall similarity appears plausible.

| Strategy | Information used | Appropriate inference setting | Main failure mode | Structural check |
|---|---|---|---|---|
| **Sequence-dominant alignment** | Nucleotide similarity and gap costs | Close homologs with limited structural divergence | Shifts stems or variable loops as identity falls | Inspect pairing compatibility and gap placement |
| **Structure-constrained alignment** | Sequence plus proposed paired and unpaired regions | Families with a defensible structural prior | Propagates an incorrect input structure | Compare alternative constraints and independent evidence |
| **Joint sequence-structure inference** | Sequence, candidate structures, and comparative support | Divergent families with coupled alignment and structure uncertainty | Converges on a self-reinforcing local solution | Perturb alignment, sampling, and structure hypotheses |
| **Expert-curated alignment** | Comparative evidence, motifs, experiments, and domain knowledge | High-value family consensus inference | Subjective decisions and limited scalability | Preserve edits and test inter-curator or method agreement |

The distinctive challenge for structured RNAs is that homologous structure may be more conserved than exact sequence. In a protein-coding alignment, codons and translated amino acids can guide correspondence. In a structured noncoding RNA, the alignment must often respect stems, loops, junctions, and insertions. A stem may contain different Watson-Crick pairs in different taxa. A loop may tolerate insertions. A peripheral domain may be present in one clade and absent in another. A structural alignment should place paired positions so that the pairing pattern can be compared, even when simple nucleotide identity would suggest a different arrangement.

![Figure 62.1. Structure-Aware Alignment Versus Sequence-Only Alignment](../assets/figures/chapter1057_figure1.png)

**Figure 62.1. Structure-Aware Alignment Versus Sequence-Only Alignment.** Pairing-compatible changes become interpretable only after homologous stem positions are aligned correctly.

Consider a small riboswitch aptamer. The ligand-binding pocket may require a few conserved nucleotides, while the surrounding helices only need to maintain base pairing. If one homolog contains a G-C pair and another contains an A-U pair at the corresponding stem position, a sequence-only alignment might treat the positions as weakly conserved. A structure-aware alignment recognizes that both sequences preserve the same stem. If the alignment incorrectly shifts one side of the stem, the apparent covariation disappears. If it aligns unrelated positions because both happen to contain G and C, it may create false support for a helix.

Several information sources can guide a structured RNA alignment. Primary sequence similarity is useful when homologs are close. Known secondary structure is useful when a family has curated stems and loops. Thermodynamic predictions can help align uncharacterized homologs but should not dominate when comparative evidence says otherwise. Conserved motifs, such as riboswitch ligand-contact nucleotides or tRNA identity elements, help anchor alignment. Genomic context can help when neighboring genes or synteny support orthology. For viral RNAs, coding frame, protein constraints, conserved untranslated regions, and recombination history may all affect alignment.

The alignment problem becomes harder as sequence divergence increases. With too little divergence, most columns are identical and there is not enough variation to detect covariation. With too much divergence, homology may be uncertain and insertions or deletions can obscure correspondence. Comparative RNA analysis is strongest in the intermediate zone: enough variation to observe compensatory changes, but enough homology to align the sequences confidently. This is why a large number of sequences is not sufficient by itself. One hundred nearly identical isolates may carry less covariation information than twenty moderately diverged orthologs.

The evidence basis for alignment quality can be tested in several ways. One can ask whether known structural motifs align consistently, whether predicted paired columns show compatible base pairs across the family, whether independent alignment tools agree, and whether removing highly similar sequences changes the result. RNAconTest was developed to compare noncoding RNA multiple sequence alignment tools based on structural consistency rather than only sequence similarity, emphasizing that structured RNA alignment has its own validation target (Wright 2020). General reviews of multiple sequence alignment also emphasize that alignment has evolved from a simple string-matching problem into a structure- and function-aware inference problem (Zhang C et al. 2024).

Modern models can use aligned homologs as input features. Multiple-sequence-alignment-based RNA language models attempt to learn structural or functional information from homologous sequence patterns, including covariation-like signals (Zhang Y et al. 2024). This is promising because an alignment contains evolutionary constraints that a single sequence lacks. It also raises the same alignment and homology-control issues that affect classical methods. If the alignment is wrong, the model may learn artifacts. If the benchmark includes close homologs in training and test sets, the model may appear to infer structure when it is mainly recognizing family membership.

Do not overgeneralize: a multiple sequence alignment is not a neutral observation. It is an inferred object. The choice of sequences, alignment algorithm, structural constraints, trimming rules, and reference model can all change the apparent evidence. A reader should treat an alignment as part of the evidence chain, not as raw data.

## 62.2. Covariation statistics and compensatory substitutions as structural evidence

Covariation is the comparative signature most closely associated with RNA secondary structure. The basic mechanism is straightforward. A base pair contributes to an RNA fold. A mutation at one side of the pair can weaken or disrupt the fold. If a second mutation at the other side restores compatible pairing, the RNA can preserve structure despite sequence change. Across evolution, homologs may therefore show paired columns in which G-C, C-G, A-U, U-A, G-U, and U-G occur in patterns that are richer than expected from independent column conservation.

The strongest pattern is a compensatory substitution. Imagine that one lineage has G at column i and C at column j, while another lineage has A at column i and U at column j. The pair changes from G-C to A-U, but pairing is retained. A compatible single-sided change is weaker but still informative: G-C to G-U preserves a wobble pair without changing both positions. Noncanonical pairs can also be conserved in specific RNA motifs, but they require more care because simple Watson-Crick compatibility scoring may miss them or overcount weak interactions.

Covariation is not merely two columns having high mutual information. Mutual information can be high because of shared ancestry, compositional bias, limited sampling, or clade-specific insertions. If all sequences in one bacterial clade have G-C and all sequences in another clade have A-U, the pattern might reflect one historical substitution event rather than many independent compensatory events. Counting every descendant sequence as independent would exaggerate support. Phylogenetic correction reduces this problem by considering the evolutionary tree or by down-weighting redundant sequences.

Alignment uncertainty is a second major artifact. A misaligned loop can place unrelated positions into the same columns. A shifted stem can generate an apparent compensatory pattern because two unrelated local motifs happen to vary together. Sequence errors, low-quality assemblies, paralog mixing, and pseudogene fragments can also create misleading column patterns. For viral RNA, recombination and reassortment can make a single tree poor for the whole alignment. For coding viral RNA, amino acid constraints and codon structure can create nucleotide correlations that are not caused by RNA base pairing.

Structural conservation is broader than covariation. A helix can be conserved even if no compensatory substitutions are observed, because the family may be too conserved or the paired nucleotides may be constrained by sequence-specific recognition. Conversely, a region can show covariation without a stable universal helix if two subfamilies have different local structures. Comparative methods therefore need to distinguish absence of evidence from evidence of absence. Lack of covariation does not disprove a helix in a low-divergence alignment. Strong covariation in a well-sampled, well-aligned family is positive evidence for pairing.

The classic experimental logic of compensatory mutation can also be reproduced in the laboratory. A designed mutation that disrupts a predicted base pair should weaken structure or function. A compensatory rescue mutation that restores pairing should restore the phenotype or biochemical behavior. This is stronger causal evidence than natural covariation alone because the experiment directly tests the proposed pair. Recent mutation-design tools for assessing RNA structure, such as SMDesigner, illustrate how computational design can help choose sequence changes that distinguish structural hypotheses while controlling for other sequence effects (Hou et al. 2025). Natural covariation suggests which pairs matter; designed perturbation can test whether they matter in the chosen system.

Influenza A virus provides a useful boundary case. Viral RNAs often contain overlapping constraints: coding sequence, packaging signals, replication signals, RNA structure, and host interactions. Gultyaev et al. analyzed structural constraints in the evolution of the influenza A virus NP segment and used comparative evidence to identify RNA structural features under constraint (Gultyaev et al. 2014). Such examples show why comparative RNA analysis is valuable in viral genomes, but they also show why interpretation must account for protein-coding constraints, lineage structure, and segment-specific biology.

R-scape-style analysis addresses a specific statistical question: do the proposed base pairs in an alignment show more covariation than expected under a background model that accounts for conservation and phylogeny? A pair annotated in a consensus structure may be plausible for thermodynamic or historical reasons, but statistical testing asks whether the alignment itself contains significant covariation for that pair. A positive result is strong comparative support. A negative result can mean the pair is wrong, the alignment is too conserved, the sample is too small, the family is poorly aligned, or the pair is structurally conserved without enough substitutions to reach significance (Rivas et al. 2017).

![Figure 62.2. Evidence Ladder for a Conserved RNA Helix](../assets/figures/chapter1057_figure2.png)

**Figure 62.2. Evidence Ladder for a Conserved RNA Helix.** Covariation supports conserved pairing, while causal function and cellular occupancy require additional evidence.

![Figure 62.4. True Structural Covariation Versus Apparent Dependence](../assets/figures/chapter1057_figure4.png)

**Figure 62.4. True Structural Covariation Versus Apparent Dependence.** Pairing-compatible changes support a conserved helix only when the alignment is defensible and the apparent dependence survives phylogenetic, redundancy, composition, and alignment-sensitivity checks. A conserved pair with too few independent changes is low power, not evidence that the helix is absent.

> **Box 62.1. How to Read a Covariation Result**
>
> - Strong support: pairing-compatible substitutions in a reliable alignment remain significant after phylogenetic and redundancy controls.
> - Low-power result: conserved or sparsely sampled columns may lack enough independent substitutions to test the pair.
> - Artifact-prone result: alignment shifts, shared ancestry, composition, coding constraints, or recombination can create column dependence.

Do not overgeneralize: covariation is evidence for conserved pairing, not automatically evidence for a specific three-dimensional contact, ligand-binding function, regulatory mechanism, or disease relevance. The evidence ladder moves from covariation to secondary-structure hypothesis, then to functional tests and structural or biochemical validation.

## 62.3. Covariance models and comparative secondary-structure inference

A covariance model is a probabilistic description of an RNA family's conserved sequence and nested secondary structure. Unpaired consensus positions emit single nucleotides, whereas paired states jointly emit the two nucleotides assigned to opposite sides of a helix. Insert, delete, bifurcation, and transition states represent family-specific variation around that structural core. The model therefore assigns probability to a sequence in the context of an explicit structural architecture rather than treating every aligned column independently (Eddy et al. 1994).

The paired emission distribution captures more than a preference for canonical pairs. If a proposed stem contains G-C, A-U, and G-U combinations across homologs, a paired state can reward the observed family-specific distribution and penalize combinations inconsistent with it. A sequence-only profile may recognize conserved ligand-contact nucleotides but miss support distributed across coordinated changes. Conversely, a covariance model can overvalue the wrong structure if the input alignment places unrelated residues into paired columns. The model formalizes comparative evidence; it does not make the evidence independent of its input assumptions.

Comparative secondary-structure inference begins by treating alignment and structure as coupled unknowns. A preliminary alignment proposes homologous columns. Sequence conservation, thermodynamic plausibility, known motifs, and candidate pairing define possible consensus structures. Covariation statistics then test which proposed pairs receive comparative support. The alignment or structure may be revised when a stem requires implausible gaps, incompatible pairs, or a clade-specific shift. This cycle is inference, not mere formatting: different alignment and structural hypotheses can explain the same sequences, and the analyst must compare their consequences.

A consensus structure is a family-level model, not a claim that every molecule adopts one rigid conformation. Conserved cores can coexist with variable peripheral helices, alternative structures, ligand-dependent switches, or lineage-specific insertions. A single model may be appropriate when one nested architecture dominates. Subfamily models or explicitly conditional structures may be needed when pairing changes systematically among clades. The structural claim must specify whether a pair is universal, prevalent, clade-specific, or only weakly supported.

Covariance models commonly use profile stochastic context-free grammars because nested base pairs can be represented recursively. Standard formulations are less natural for crossing interactions such as pseudoknots and for ensembles of alternative conformations. They also summarize family variation through estimated emission and transition probabilities, which can be unstable when the alignment is small, redundant, or taxonomically biased. Effective sequence number and phylogenetic breadth therefore matter more than raw sequence count.

Model scores answer a compatibility question: how well does a sequence fit the modeled sequence-and-structure distribution relative to a background model? They do not directly answer whether the RNA is expressed, whether the proposed structure forms in a cell, or whether the structure is functional. Likewise, significant covariation supports conserved pairing in the evaluated alignment but does not validate every untested pair in the consensus structure. Comparative inference should preserve pair-level support rather than treating the entire model as uniformly proven.

Rfam seed alignments and Infernal-style covariance-model searches provide important inputs and downstream applications for these principles. Their operational search, thresholding, curation, classification, and annotation workflows are owned by [Chapter 140](chapter1127.md). Here the boundary is the structural inference itself: which alignment columns correspond, which pairs are supported, how paired emissions represent that evidence, and how uncertainty in the input changes the inferred consensus structure.

The main failure modes follow directly from this logic. An incorrect seed alignment can create a confident but false paired state. A consensus chosen from well-studied family members can underrepresent divergent subfamilies. An overly broad model can merge structurally distinct groups, while an overly narrow model can mistake family-specific detail for a universal constraint. These failures should be tested by alignment perturbation, taxonomic subsampling, alternative consensus structures, pair-level covariation analysis, and comparison with probing or high-resolution structures when available.

## 62.4. Benchmarking comparative structure inference and generalization

Benchmarking is where comparative information often becomes a source of hidden error. A benchmark is supposed to estimate how well a method will perform on new cases. For RNA structure and ncRNA annotation, "new" must be defined carefully. A sequence can be new as a file entry while still being homologous to training data. A model trained on many members of a riboswitch family may perform well on another member of the same family because the family grammar is familiar. That performance is useful for annotation, but it is not evidence that the model can infer the structure of an unrelated RNA family.

Homology-aware benchmarking begins by asking what kind of generalization is being tested. If the goal is to annotate additional members of known families, then within-family generalization is appropriate. The benchmark should hold out sequences from known families and test whether the method recovers them. If the goal is to discover new RNA families or predict structures for unrelated RNAs, then the benchmark must separate families, clans, structural classes, or taxonomic groups. A single random train-test split cannot answer all these questions.

**Table 62.2. Benchmark Splits and Structural Generalization Claims.** Random, family-held-out, structure-held-out, and taxonomically separated benchmark splits support different generalization claims; leakage across homologous families or structures inflates apparent performance.

| Split type | Separation unit | Supported claim | Unsupported claim |
|---|---|---|---|
| **Random sequence split** | Individual records | Held-out records from sampled distributions | Independence from homologous training examples |
| **Identity-cluster split** | Sequences above a chosen identity threshold | Transfer beyond near duplicates | Absence of remote homolog leakage |
| **Family or clan split** | Curated related groups | Generalization beyond withheld families or broader groups | Transfer to all structural classes |
| **Taxonomic split** | Held-out taxa | Performance across selected organism groups | Separation of taxonomy from RNA-class composition |
| **Structure-class split** | Folds or structural architectures | Transfer to unseen structural classes | Within-class calibration or condition-specific folding |
| **Time split** | Records after a release cutoff | Prospective performance under database growth | Protection from retrospective annotation changes |

The most common failure is homolog leakage. Homolog leakage occurs when close relatives, near-duplicate sequences, or family members appear in both training and test sets. The method can then use family-specific sequence motifs, alignment patterns, or model memory to succeed. This problem is acute for deep learning models, RNA language models, and classifiers trained on public databases because those databases contain many related sequences. Multiple-sequence-alignment-based RNA language models explicitly use homologous sequences, so their evaluation must make clear whether the test measures use of available homologs, extrapolation to new family members, or transfer to unrelated families (Zhang Y et al. 2024).

There are several practical controls. Sequence-identity clustering removes near duplicates, but identity alone is not enough because remote homologs can share structure with low sequence identity. Family-level splitting holds out entire Rfam families or equivalent curated groups. Clan-level splitting is stricter when related families belong to a larger structural or evolutionary group. Taxonomic splitting can test whether a model trained on bacterial RNAs generalizes to archaeal or eukaryotic homologs, but it may confound taxonomy with RNA class. Structure-class splitting holds out all members of a structural type, such as tRNA-like folds or riboswitch aptamers, when testing broad structural transfer. The right split depends on the scientific claim.

Negative examples also require care. In ncRNA discovery, a random genomic sequence is not necessarily a true negative; it may contain an unannotated RNA or a structured regulatory element. A shuffled sequence may destroy composition or dinucleotide properties. A coding sequence may contain real RNA structures while also encoding protein. A benchmark that treats all unannotated regions as negatives can penalize real discoveries. A benchmark that uses unrealistic shuffled negatives can inflate performance. The most informative negative sets are matched to the intended use: same genome composition, same length distribution, same annotation uncertainty, and explicit exclusion of known or suspected structured elements where possible.

Homology-aware evaluation should also track alignment availability. Some methods require an alignment, some search for homologs internally, and some operate on single sequences. Comparing them without controlling inputs can be misleading. A method given a curated Rfam alignment has more information than a single-sequence thermodynamic predictor. A method that builds an alignment from a large database may outperform in well-sampled families and fail in orphan RNAs. Benchmark reports should therefore state whether homologs were provided, searched, filtered, or withheld.

RNAconTest is a useful example because it evaluates noncoding RNA multiple sequence alignment tools using structural consistency (Wright 2020). That framing recognizes that, for structured RNAs, the relevant output is not only column identity but preservation of structural relationships. Similarly, a structure predictor benchmark should not rely only on global base-pair F1 score when the biological question concerns a specific conserved helix, ligand-binding pocket, or viral regulatory element. Local structural accuracy, uncertainty calibration, and family-level transfer can be more informative than a single aggregate score.

Generalization failures are not only computational. They also affect biological interpretation. A covariance model trained on bacterial riboswitches may miss eukaryotic structural analogs that bind similar metabolites but are not homologous. A model trained on canonical tRNAs may misclassify tRNA-derived fragments or unusual mitochondrial tRNAs. A viral RNA predictor trained on one virus family may learn coding-composition constraints rather than RNA structural rules. Benchmarking should expose these limits instead of hiding them behind high average performance.

Recent reviews of deep learning in RNA structure studies emphasize the promise of data-driven methods but also the importance of data representation, benchmark construction, and independent validation (Yu et al. 2022). The lesson for comparative methods is not that machine learning is unreliable; it is that evolutionary relatedness must be an explicit variable in evaluation. A model that excels within known families is valuable for annotation. A model that transfers across unrelated families is valuable for discovery. Confusing these two achievements produces overstated claims.

Do not overgeneralize: "held-out test set" does not automatically mean "independent biological test." For RNA structure and ncRNA discovery, independence must be defined at the level of sequence identity, family, structural class, taxonomy, and data source.

## 62.5. Comparative structure inference for ncRNAs and viral RNAs

Noncoding RNA discovery is one of the major successes of comparative RNA analysis. Many functional ncRNAs do not encode proteins and may lack long open reading frames, strong expression in every condition, or obvious sequence motifs. Comparative structure can reveal them because selection preserves a fold, a processing signal, a guide region, or a protein-binding architecture. Riboswitches, small nucleolar RNAs, RNase P RNA, SRP RNA, spliceosomal snRNAs, and many bacterial small RNAs have all benefited from comparative annotation and family modeling.

For structural inference, the candidate source is less important than the evidential question asked of the aligned homologs. The analyst tests whether proposed stems are supported by pairing-compatible substitutions, whether conserved nucleotides define a pocket or interaction surface, whether gaps and insertions respect a plausible architecture, and whether the signal persists after redundant sequences or uncertain regions are removed. Candidate collection, database search, family assignment, and annotation-record production are operational tasks owned by [Chapter 140](chapter1127.md). This chapter begins once a defensible homolog set is evaluated for conserved structure.

Riboswitch discovery illustrates the evidence ladder. A conserved structured element upstream of metabolic genes is a candidate. Covariation in stems supports a conserved fold. Conserved nucleotides in an aptamer pocket suggest ligand recognition. Genomic association with genes in the relevant metabolic pathway strengthens the hypothesis. Biochemical binding assays, genetic reporter assays, and structural biology can then test ligand binding and regulatory mechanism. Comparative evidence is not the final proof, but it makes the candidate specific enough to test.

Small RNAs and long noncoding RNAs are more difficult. Many small RNAs are lineage-specific, have short seed regions, or act through imperfect base pairing with mRNA targets. Long noncoding RNAs may conserve only local motifs, expression patterns, or synteny rather than a global secondary structure. A full-length covariance model may be inappropriate if only a small domain is conserved. Comparative analysis can still be valuable, but claims should specify the conserved unit: a hairpin, a binding motif, a splice-associated structure, an antisense overlap, or a processed small RNA. Treating an entire long transcript as a conserved structured RNA because one local element is conserved is an overgeneralization.

Viral RNA analysis adds both opportunity and difficulty. RNA viruses evolve rapidly, so related isolates can provide abundant variation. Viral genomes also use RNA structures for replication, translation, frameshifting, packaging, immune evasion, and host adaptation. At the same time, viral genomes often encode proteins, overlap reading frames, recombine, reassort, and experience strong host-specific selection. A conserved nucleotide in a viral coding region may be conserved because of amino acid constraint, codon usage, RNA structure, packaging, replication, or immune pressure. Comparative claims must separate these possibilities when possible.

Influenza A virus NP segment analysis provides one concrete example of using comparative sequence information to infer RNA structural constraints in a coding viral segment (Gultyaev et al. 2014). Such work must account for coding constraints because a synonymous change is not always neutral: it can alter codon usage, RNA folding, splicing signals, packaging signals, or host immune recognition. The best viral comparative analyses combine synonymous-site analysis, structural prediction, covariation, reverse genetics, probing, and fitness assays.

Cap-independent translation elements provide another comparative discovery context. Weingarten-Gabbay et al. used comparative genetics to systematically discover cap-independent translation sequences in human and viral genomes (Weingarten-Gabbay et al. 2016). The broader lesson is that comparative signals can guide functional discovery, but the output of a screen must still be interpreted through mechanistic assays. A sequence that promotes cap-independent translation in a reporter context may depend on RNA structure, RNA-protein binding, cryptic promoter activity, transcript abundance, or context-specific translation factors. Comparative conservation helps prioritize candidates; it does not by itself define the mechanism.

High-throughput structure probing now provides an important complement to comparative analysis. Reviews of RNA structure and in vivo probing emphasize that RNA structures participate in many RNA functions and that probing can report condition-specific folding in cells (Cao et al. 2024; Mitchell et al. 2019). Comparative analysis asks which structural features have been preserved over evolution. Probing asks which positions are reactive or protected in a particular experimental condition. Agreement between covariation and probing strengthens a structural model. Disagreement can be informative: a conserved helix may form only in a different condition, may require a ligand or protein, may be transient, or may be incorrectly aligned or annotated.

Comparative methods also help identify when a structure prediction is likely to be biologically overfit. A single viral sequence may fold into many plausible thermodynamic structures. If only one predicted helix is conserved across related viruses and shows covariation, that helix becomes a stronger candidate for function. Conversely, a long minimum-free-energy structure with no comparative support and no probing support should be treated as a computational hypothesis. Comparative analysis is a filter that prioritizes evolutionarily constrained features for experiments.

The clinical and technological relevance of comparative methods is indirect but important. Viral RNA structure models can guide attenuation strategies, antiviral target discovery, vaccine design, and interpretation of variants. ncRNA family annotation can identify regulatory RNAs in pathogens or microbiomes. tRNA and other structured RNA annotation supports genome annotation and synthetic biology. However, translational use raises the evidence threshold. A conserved viral stem proposed as a drug target requires structural validation, functional perturbation, conservation across clinically relevant lineages, and resistance analysis. Comparative evidence begins the argument; it does not complete it.

## Recent Consensus

The current consensus is that comparative sequence analysis provides some of the strongest evidence for conserved RNA secondary structure when the alignment is reliable, the sequence set has appropriate divergence, and covariation is assessed with statistical and phylogenetic controls. Covariation is especially compelling when compensatory substitutions preserve base pairing across independent lineages. It is weaker when the alignment is uncertain, sequences are too similar, or apparent dependencies can be explained by shared ancestry or compositional bias.

There is also broad agreement that covariance models are useful because they encode a family-level hypothesis about both conserved sequence and nested secondary structure. Their paired emissions preserve evidence that sequence-only profiles omit, but their conclusions remain conditional on alignment quality, representative sampling, consensus-structure choice, and model scope. Search and annotation systems apply these models at scale; their operational treatment belongs to [Chapter 140](chapter1127.md).

A third consensus is that benchmarking must account for homology. Random splits are often inadequate for RNA structure and ncRNA discovery tasks because related RNAs can leak family information between training and test sets. Benchmarks should report sequence redundancy, family membership, alignment availability, database versions, and the intended generalization target. Recent deep-learning reviews and RNA language model studies reinforce that evolutionary information is powerful but can inflate performance estimates if not controlled (Yu et al. 2022; Zhang Y et al. 2024).

Finally, comparative evidence is best interpreted alongside thermodynamic prediction, structure probing, high-resolution structure, mutational rescue, and functional assays. The field no longer treats any single evidence type as sufficient for all claims. A conserved, covarying helix is strong evidence for a secondary-structure feature; a biological mechanism requires additional experiments.

## Open Questions, Controversies, Deprecated Models, and Common Misconceptions

Open questions:

- How can researchers model RNA families with genuine structural heterogeneity? Some RNA families contain conserved cores with variable peripheral domains. Others may switch between conformations or use alternative structures in different lineages. A single consensus secondary structure can be too rigid. Future family models may need to represent subfamily structures, conditional conformations, pseudoknots, protein-bound states, and tertiary motifs more explicitly.

Controversies:

- Another unsettled area is the integration of comparative and machine-learning methods. RNA language models can use homologous sequence information and may learn covariation-like patterns, but benchmark design determines whether performance reflects transferable structure learning or recognition of known family distributions. The field needs better calibration, family-disjoint evaluation, and experimental validation of predictions outside well-curated RNA families.
- R-scape-style covariation testing also has interpretation boundaries. A lack of statistically significant covariation is not proof that a pair is absent. It may reflect low divergence, low power, poor alignment, or sequence-specific conservation. Conversely, significant covariation supports conserved pairing but does not identify the full three-dimensional fold or establish function.

Common misconceptions:

- "Conserved sequence means conserved structure." Some RNA elements conserve sequence because a protein, ribosome, nuclease, guide RNA, or ligand recognizes specific nucleotides. Other elements conserve structure while changing sequence. Both modes can coexist in the same RNA.
- "Covariation proves function." Covariation supports structural conservation. Function requires evidence that the structure affects expression, stability, localization, replication, translation, binding, catalysis, or phenotype in the relevant biological context.
- "More sequences always improve comparative analysis." More sequences help only if they add informative independent variation and are aligned correctly. Many near-identical sequences mostly add redundancy, while many divergent or misannotated sequences can add noise.
- "A covariance model hit is an expressed RNA gene." A covariance model hit is evidence that a genomic sequence resembles a modeled RNA family. Expression, processing, cellular localization, and function require separate evidence.