CaCoFold-R3D

Structural RNAs do their work in three dimensions, and the geometry that packs their canonical helices together comes largely from short recurrent elements in loop regions — kink-turns, sarcin-ricin loops, GNRA tetraloops, C-loops — held by non-Watson–Crick interactions. Conventional secondary-structure folding ignores these motifs entirely; motif-detection tools find them one at a time, outside the context of the fold they help create. CaCoFold-R3D, from Aayush Karan and Elena Rivas at Harvard University, predicts both together: a single joint model outputs the canonical helices, including pseudoknots, and the 3D motifs that decorate every loop between them.

Unlike most of the catalog, this is not a neural network. CaCoFold-R3D is a probabilistic grammar. An array of stochastic context-free grammars (SCFGs) describes the structural architecture of each motif class, and profile hidden Markov models — the R3D-HMMs — describe sequence variability within a motif, with match, insert, and delete states allowing observed length variation. The grammar and HMM parameters are estimated once from a curated descriptor library of known motifs and then applied unchanged to any new RNA alignment, the probabilistic-model analogue of running inference from a frozen checkpoint.

The insight that makes it work is evolutionary. RNA 3D motifs mostly do not covary, so covariation cannot find them directly. Canonical helices do covary reliably, and R-scape's covariation statistics pin those helices down; because motifs must occupy the loops the helices delimit, covariation constrains motif placement indirectly. The authors summarize the design as everything, everywhere, all-at-once: an almost-exhaustive motif library, any non-helical loop including higher-order junctions, one grammar and one parse.

#Key Features

• Joint secondary-structure and motif parse: Helices, pseudoknots, and 3D motifs are produced by one CYK parse of one grammar, rather than by folding first and annotating afterwards.
• Almost-exhaustive motif library: The descriptor file defines 51 distinct motifs, which the software expands internally into SCFGs for 96 motif variants covering hairpin loops, bulges, internal loops, three- and four-way junctions, and multiloop branch segments.
• Covariation as a placement constraint: On GNRA tetraloop and kink-turn benchmarks, adding covariation from the alignment lifts weighted sensitivity from 84.5% to 95.4% and sharply reduces false positives per sequence.
• Editable, not retrained: New motifs are added by writing them into a plain-text descriptor file, so the method extends to novel motif discovery without any retraining step.
• Fast enough for whole-database scans: 98% of Rfam families (4,079 of 4,178) complete end-to-end in under 60 seconds on an Apple M3 Max.

#Technical Details

The base RBGJ3J4 grammar contributes 12 nonterminals and the R3D motif grammars contribute 96, and the all-at-once CYK parse scales as O(L³ × M) in alignment length L and nonterminal count M. Running CaCoFold-R3D over every Rfam v15.0 seed alignment detects 2,124 motifs, of which 1,460 have covariation support — at least one covarying base pair in a bounding helix — spread across 591 families; the eukaryotic small and large ribosomal subunits alone account for 45 and 62 supported motifs. Column-permuted negative alignments, which destroy covariation while preserving base composition, yield 121 covariation-supported motifs, implying an estimated 8.3% false discovery rate for supported predictions versus 25.4% for unsupported ones. The two largest structures take 32 minutes (eukaryotic SSU) and 2.9 hours (eukaryotic LSU). The method ships in the R-scape package from version 2.5.7, invoked as R-scape -s --cacofold --r3d, written in C with BSD 3-clause terms in its LICENSE file — GitHub's license detector reports "Other" because that grant sits alongside notices for the bundled HMMER, Easel, FastTree, and R2R dependencies.

#Applications

The direct use is annotating structural RNA alignments: an Rfam curator, a riboswitch biologist, or anyone with a homolog alignment gets a consensus structure in which the non-canonical interactions are named rather than left as unpaired loop. Because those all-residue interactions are the constraints all-atom predictors need, the predictions are also intended as input to deep-learning 3D pipelines such as AlphaFold 3 and RoseTTAFold All-Atom, which already consume Rfam alignments. Motif annotation is likewise a starting point for treating structured RNA loops as small-molecule drug targets.

#Impact

Published in Nature Methods in October 2025 after a December 2024 preprint, CaCoFold-R3D changes what alignment-based motif prediction can cover. The closest earlier methods, RMDetect and BayesPairing2, model a limited set of motifs one at a time, require a secondary structure to be supplied or predicted first, and are expensive enough that BayesPairing2 has to window RNAs longer than 300 nucleotides; CaCoFold-R3D handles the largest rRNAs and uses one family-independent parameterization across all of Rfam. Predicted structures for the full Rfam database are distributed with the paper. The limits are inherent to the approach: it needs an alignment with genuine covariation signal, since without it sensitivity and precision both drop; it can only find motifs already in the descriptor library; and it is distributed as a C program built with configure/make, not a Python package or a hosted service.