🧬 biology

Cross-species ncRNA annotation using synteny-constrained embedding similarity

The paper introduces CURIA, a novel pipeline that combines synteny-constrained genome alignment with RiNALMo RNA foundation model embeddings to improve cross-species non-coding RNA annotation by identifying localized embedding-supported regions that complement conventional sequence conservation metrics.

Original authors: Bogdan M. Kirilenko

Published 2026-07-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Bogdan M. Kirilenko

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine the human body as a massive, bustling library. Inside, there are tens of thousands of books (genes) that tell our cells how to build proteins, the bricks and mortar of life. For decades, scientists have been great at finding the "instruction manuals" for these protein books across different species. If you find a manual for building a heart in a human, you can usually find a very similar one in a mouse or a cow because the text is almost identical.

But hidden in the library are thousands of other books that don't give instructions for building things. These are called non-coding RNAs (ncRNAs). They act as the library's catalog system, the security guards, or the sticky notes in the margins—performing various regulatory roles to manage the library. However, for many of these transcripts, their specific biological functions remain a mystery. These "management books" are also incredibly messy; their text changes wildly between species, their page numbers are hard to pin down, and they often look nothing like each other even if they perform the same job. Trying to find the mouse version of a human management book by just reading the words is like trying to find a specific recipe in a foreign language by only looking at the font style—it's nearly impossible.

This is where a new tool called CURIA comes in. Think of it as a super-smart detective that doesn't rely on pre-existing cheat sheets—it doesn't look for known patterns (motifs), doesn't use lists of known RNA families, and doesn't need to know what the books actually do or what their physical shapes look like. Instead, it learns the patterns of the language itself and looks at the "neighborhood" where the book is found. Instead of getting confused by the messy text, CURIA uses a special kind of "mathematical fingerprint" (called an embedding) to see if two books feel similar, even if they look different. It also checks the "street address" (synteny) to make sure the book is in the right part of the library.

The researcher behind CURIA tested this detective on 19 different animal genomes, from our closest relatives like monkeys to distant cousins like kangaroos and whales. He wanted to see if he could find the mouse or cow versions of human "management books" that had been lost in translation.

Here is what he found:

1. The "Short and Sweet" Success
For the tiny, compact books (like microRNAs), CURIA worked surprisingly well. When comparing human books to mouse books, about 52.4% of the predicted matches landed right on top of a known mouse book. For the most famous, well-named books, the success rate jumped to 83.6%. This suggests that for small, simple management notes, the new "fingerprint" method is a great way to find their cousins in other animals.

2. The "Long and Messy" Puzzle
The real magic happened with the long, complicated books (called lncRNAs). These are often tens of thousands of letters long and very different between species. CURIA realized that you can't compare the whole book at once. Instead, it broke them down into small "islands" or chunks. It found that while the whole book might look different, specific small islands within them often looked very similar.

In a detailed look at humans versus mice and cows, these "islands" often matched up even when the surrounding text didn't. It's like finding that two different editions of a novel have the exact same three paragraphs in the middle, even if the rest of the story is rewritten. This suggests that evolution keeps these specific small chunks because they are important, even if the rest of the book changes.

3. The "Super-Finder" List
The author took this a step further. He looked for "islands" that appeared in almost all 19 animals he tested. He found 1,693 human long books that had these special, matching islands in at least 17 of the 19 other species. He even created a "super-stringent" list of just 170 books where the matches were extremely strong. These are the most promising candidates for important biological functions that have been preserved for millions of years.

What This Means (and What It Doesn't)
The author is careful to say that this isn't a magic wand that solves everything. He didn't prove that these matching islands definitely do the same job in every animal, nor did he prove they are the same "book" in an evolutionary sense. He simply showed that this new method can find candidates—places where the math says, "Hey, these look related!"

He also found that for the tiny books, the new method didn't add much more than just looking at the text itself. But for the long, messy books, the new "fingerprint" method found connections that old text-matching tools missed completely. It's like having a new pair of glasses that lets you see patterns in the fog that were invisible before.

In short, CURIA suggests that we can find the hidden connections between species' genetic management systems by looking at the shape of the story and the neighborhood, not just the words. It is a proof of concept that opens the door to understanding how these complex, non-coding parts of our DNA have evolved, one small island at a time.

Technical Summary: Cross-species ncRNA annotation using synteny-constrained embedding similarity

Problem Statement

Reference-based annotation transfer is well-established for protein-coding genes but remains challenging for non-coding RNAs (ncRNAs). Unlike coding genes, ncRNAs often exhibit weak primary-sequence conservation, ambiguous transcript boundaries, and heterogeneous annotations across species. Evolutionary conservation in ncRNAs may be restricted to short motifs or localized domains rather than the entire transcript, and whole-locus comparisons often dilute these localized signals. Existing methods, such as covariance models (Infernal/Rfam) or structure-aware alignments (LocARNA), rely on pre-defined family models or known sequences, making them less suitable for transferring annotations for reference loci lacking curated models or for identifying conserved subregions within large, multi-exon loci.

Methodology: The CURIA Pipeline

The author presents CURIA, a computational pipeline that frames cross-species ncRNA correspondence as synteny-constrained matching in RNA embedding space. The pipeline combines genome alignment chains with representations from the RiNALMo RNA foundation model.

Core Workflow

Input & Preprocessing: The pipeline accepts a reference genome with ncRNA annotations (BED12), a query genome, and pairwise genome alignment chains (UCSC chain format). Isoforms at the same locus are merged into a "union exon model" to define the search space.
Candidate Region Prediction: Using a chain-based classifier adapted from the TOGA orthology framework, CURIA projects reference loci onto the query genome. This step uses features like synteny gene count, exonic occupancy, and flank coverage to identify coarse candidate regions without reconstructing full transcripts.
Branch Routing: Loci are processed based on geometry rather than biotype:
- Compact Branch: Single-exon loci $\le$ 256 nt are treated as complete regions.
- Localized Subregion Branch: Longer single-exon loci (>256 nt) and all multi-exon loci are scanned for "islands" (detector-positive subregions).
Embedding Generation: Sequences are represented using per-nucleotide embeddings from the RiNALMo foundation model (650M parameters). Two task-specific PCA projections are applied:
- $k=16$ for position-level matching.
- $k=64$ for localized subregion detection.
Detection and Matching:
- Short ncRNAs: Candidate windows in the query genome are compared to the reference using Maximum Mean Discrepancy (MMD) with an RBF kernel on the 16-dimensional embeddings. The best window is refined via local boundary perturbation.
- Long/Multi-exon ncRNAs: A logistic regression detector identifies "islands" (localized candidate regions in embedding space that are suitable for cross-species matching) in the reference. These islands are projected to the query genome via alignment chains. Query islands are detected similarly. Finally, reference and query islands are matched using Smith–Waterman local alignment over a per-token cosine-similarity matrix (derived from the 16D embeddings).
Scoring: Matches are scored by an embedding-SW distance $d = 1/(1+s)$ , where $s$ is the alignment score. Lower $d$ indicates stronger matches. Collinearity constraints are applied to filter assignments.

Key Contributions

Formulation: The task of cross-species ncRNA annotation is redefined as synteny-constrained matching within RNA embedding space, avoiding the need for pre-existing family models or end-to-end transcript conservation.
Pipeline Development: Introduction of CURIA, which integrates chain-based locus projection with localized embedding-based comparison.
Dual-Strategy Approach: Development of complementary strategies for compact loci (MMD-based global comparison) and extended loci (localized island detection and Smith–Waterman matching).
Genome-Scale Application: Demonstration of the approach across 19 mammalian genomes, identifying recurrent candidate correspondences that conventional sequence metrics fail to fully explain.

Results

The pipeline was evaluated on 19 human-to-query genome pairs (ranging from primates to marsupials), using human (hg38) as the reference.

Compact ncRNAs: In the human–mouse comparison, 52.4% of predicted short-ncRNA loci overlapped an annotated exon. MMD scores were strongly anticorrelated with nucleotide identity. However, adding MMD to a sequence-only baseline provided only a marginal improvement in predicting annotation support ( $\Delta$ ROC-AUC = +0.0064), suggesting that for compact ncRNAs, embedding similarity largely tracks conventional sequence metrics.
Localized Subregions (Islands): For longer loci, the pipeline identified 27,784 reference islands. In human–mouse comparisons, 50.1% of accepted query-island matches overlapped an annotated exon.
Recurrent Cores: A multi-species screen identified 1,693 human lncRNA loci containing recurrent embedding-supported regions in at least 17 of 19 query genomes. A stricter subset of 170 loci was identified using a distance threshold of $d \le 0.02$ .
Complementarity of Embeddings: In localized subregion matching, embedding-based scores provided ranking information beyond nucleotide identity, local alignment, k-mer similarity, and composition. For low-sequence-similarity pairs, embedding-SW substantially outperformed nucleotide Smith–Waterman in ranking assigned partners (ROC-AUC 0.798 vs. 0.615 for human–mouse).
Structural Analysis: Exploratory ViennaRNA comparisons showed heterogeneous structural agreement. While some matches (e.g., RNU6-7) showed near-identical pairing, others showed weak or boundary-dependent agreement, indicating that strong embedding correspondence does not uniformly imply conserved secondary structure.

Significance and Claims

The author frames this work as a proof of concept rather than a production-ready annotation framework. The primary significance lies in demonstrating that synteny-constrained, embedding-based annotation can recover informative cross-species ncRNA correspondences at a genome scale, particularly in cases where primary-sequence conservation is weak or heterogeneous.

Key claims include:

Localized Conservation: The results show that detectable cross-species correspondence can be concentrated in localized subregions (islands) rather than distributed across the entire annotated locus.
Complementary Signal: Localized RNA embeddings provide a comparative signal that complements conventional sequence conservation. This is particularly evident in the ranking of candidate correspondences for long ncRNAs where sequence similarity is low.
Feasibility: Restricting comparison to chain-defined candidate loci makes the computationally intensive task of genome-wide embedding scanning feasible.
Limitations: The author explicitly states that the method does not establish evolutionary orthology, biological function, or stable RNA structure. The results are candidate correspondences supported jointly by synteny and embedding similarity, requiring further validation. The approach depends heavily on the quality of genome alignment chains and the specific representations learned by the RNA foundation model.