## Feature Request: RNA Secondary Structure Doublet Models (RNA6/RNA7/RNA16) in IQ-TREE 2/3 and ModelFinder

[apselvatti-crypto]

Summary: Requesting the implementation of RNA secondary structure doublet models (analogous to RAxML's -A S16/S6/S7 and IQ-TREE 1.x's RNA6A/RNA16A etc.) in IQ-TREE 2/3, including support for ModelFinder model selection over these models.

---

Background and motivation

RNA doublet models treat paired bases in rRNA stem regions as a single character with 16 states (the 16 possible dinucleotides: AA, AC, ..., UU), correctly capturing the co-evolutionary dynamics of compensatory mutations that maintain Watson-Crick base pairing. This is well-established theory (Tillier & Collins 1995; Savill et al. 2001) and has direct practical relevance: standard 4-state nucleotide models applied to stem regions violate the independence assumption between paired sites, leading to biased branch length estimates and potentially incorrect topologies.

The workflow I am currently using (and which I believe is somewhat discussed to in the March 2025 RAxML mailing list thread as "RNA models and bootstrap values of trees inferred with RNA models") involves:

1. Aligning rRNA sequences with cmalign (Infernal) using Rfam covariance models (RF01960 for 12S, RF02013 for 16S), which produces a Stockholm alignment with a #=GC SS_cons annotation
2. Parsing the SS_cons with a custom Python script to identify stem pairs and unpaired loop positions
3. Reordering stem columns so that each pair (i, j) is physically consecutive in the output FASTA — a prerequisite for doublet models
4. Creating a partitioned alignment with a NEXUS partition file specifying GTR+G for loops and a doublet model for stems

The external partitioning is already done. The missing piece is that IQ-TREE 2.3.6 does not recognise RNA6A, RNA16A, or any RNA doublet model name — the error is ERROR: File not found RNA16A, indicating these models were not available in current IQ-TREE versions (2 and 3 tested).

---

What exists and what is missing

A quick survey found the RNA models are implemented in some programs but not in others:

Software	RNA doublet models	Notes
IQ-TREE 1.6.x	✓ RNA6A/B, RNA7A–F, RNA16A/B/C	Not maintained; no MFP+MERGE
IQ-TREE 2.x / 3.x	✗	Models dropped; --seqtype has no RNA option
RAxML 8	✓ S6A–E, S7A–F, S16, S16A/B via -A+-S	Standard RAxML only, not RAxML-NG
RAxML-NG	✗	Not implemented

The IQ-TREE 2/3 paper (Minh et al. 2020, MBE) does not mention RNA doublet models among the new features, and I can confirm empirically that they are absent from 2.3.6.

---

Proposed implementation

I would like to propose two levels of implementation, ordered by complexity:

Level 1 (minimal) — Re-expose RNA doublet model names:
Re-port the RNA6/RNA7/RNA16 model family from IQ-TREE 1.x as named substitution models. The partitioning into stems/loops is already handled externally. The user provides a pre-reordered FASTA where each stem pair occupies two consecutive columns, and specifies the model by name:

iqtree2 -s stems_reordered.fasta -m RNA16A+G -B 1000 -T AUTO

This requires the model names to be registered in the model factory and the alignment reader to treat consecutive column pairs as doublet states — analogous to how codon models read three consecutive columns as a single unit.

Level 2 (complete, but likely more complex) — Secondary structure file support + ModelFinder integration:
Support a secondary structure annotation file (similar to RAxML's -S flag, accepting dot-bracket or Rfam Stockholm SS_cons notation), which automatically handles the internal partitioning:

iqtree2 -s rna_alignment.fasta -S secondary_structure.txt \
        -m TEST --mset GTR,RNA6A,RNA7A,RNA16A,RNA16B,RNA16C \
        -B 1000 -T AUTO

The structure file defines which columns are paired; IQ-TREE handles the rest internally, including bootstrapping stem pairs as units (as Stamatakis confirmed RAxML does in the March 2025 mailing list exchange).

---

Context and urgency

A recent discussion between leading developers in the field highlights how timely this is. The external partitioning step is already being handled by the community — what remains is IQ-TREE recognising the model family, which is the natural place for this given IQ-TREE's role as the dominant ML phylogenetics framework. For rRNA phylogenomics, implementing doublet models is the single most impactful modelling improvement available, as it corrects a formal violation of the site-independence assumption rather than simply adding parameters.

---

Thank you for considering this. I am happy to provide test alignments, the partition-preparation script, or any other material that would help with implementation.

Kind regards

Alexandre Pedro Selvatti (Alex)