gCoSRNA

Multi-way junctions — where three or more helices meet — set the global architecture of most structured RNAs, and the interaction that most strongly determines their geometry is coaxial stacking, in which two adjacent stems align end to end into a single quasi-continuous helix. Knowing which stems stack collapses a large space of junction geometries into a handful, making stacking a valuable constraint for RNA 3D modeling. Yet the one publicly reported tool built for it, Junction-Explorer (2012), trains a separate classifier per junction order, covers only three- and four-way junctions, and its web server no longer returns results.

gCoSRNA, from the RNA folding group at Wuhan Textile University, replaces per-topology models with one classifier. It decomposes any n-way junction into "pseudo two-way junctions" — each pair of adjacent stems plus the unpaired nucleotides connecting them — scores every pair, then reassembles the pairwise calls into a global stacking configuration with no junction-type classification step. Because two-way motifs such as bulges and internal loops are abundant in the Protein Data Bank, the decomposition both multiplies the effective training data and makes the model topology-independent: it transfers to five-, six-, and seven-way junctions despite almost no high-order RNA-only examples in training.

The design is deliberately classical — a random forest over twelve hand-crafted sequence and thermodynamic features rather than a deep network — and gCoSRNA is a component for RNA tertiary structure pipelines, not a structure predictor itself. It complements end-to-end deep-learning modeling such as AlphaFold 3 and trRosettaRNA, which CASP15 and CASP16 showed has not yet displaced physics-based and fragment-assembly methods for RNA. The preprint has not been peer reviewed.

#Key Features

• Topology-independent decomposition: One binary classifier handles junctions of two to seven branches by reducing each to adjacent stem pairs, removing the need for order-specific models or structurally resolved homologs.
• Fixed pretrained checkpoint: The repository ships a serialized model (gCoSRNA_model.pkl) with separate training and prediction scripts, so inference runs on a sequence and dot-bracket structure without retraining.
• Interpretable features: The twelve features are dominated by coaxial stacking free energy between the stems, connecting loop length, number of intervening stems, and helix stability, all from Turner nearest-neighbor parameters.
• Held-out benchmarking: An independent test set of CASP15/16 and RNA-Puzzles targets is filtered to under 80% sequence identity with the training data.
• Model selection: Predicted stacking also filters and ranks candidate 3D models.

#Technical Details

The training corpus starts from 2,534 RNA-only PDB entries, reduced by CD-HIT at 80% identity to 681 representative molecules, parsed by DSSR into junctions and clustered again at the junction level. The non-redundant set holds 1,642 two-way (subsampled to 1,457), 312 three-way, 231 four-way, 87 five-way, and 18 six-way-or-higher junctions, split 85/15 into training and Test Set I; 45 six-way and 49 seven-way junctions from non-RNA-only structures were added to the test side. Bayesian optimization under 10-fold cross-validation yielded a 562-tree forest of maximum depth 24 with a positive class weight of 1.09, calling pairs stacked above probability 0.42 and resolving conflicts in favor of the highest-scoring partner.

Cross-validation on pseudo two-way junctions reached an AUC of 0.97. On Test Set I, exact-configuration accuracy was 0.94 for two-way, 0.88 for three-way, and 0.85 for four-way junctions, with pairwise accuracy of 0.92 on five-way and 0.95 across the 94 six- and seven-way cases. On the CASP/RNA-Puzzles set, accuracy was 0.86 on 80 two-way and 0.86 on 29 four-way junctions but 0.56 on the 9 three-way junctions. Averaged over three- and four-way junctions, gCoSRNA scored 0.79 against 0.71 for a Junction-Explorer re-implementation retrained on the new dataset and 0.64 for the original.

#Applications

The practical target is RNA 3D structure modeling, where predicted stacking can enter a physics-based or fragment-assembly energy function as a constraint, or rank candidate models after the fact. On three- and four-way junction targets from CASP15/16 and RNA-Puzzles, submitted models spanned roughly 4–50 Å RMSD with a mean near 27 Å; restricting to models with secondary-structure F1 above 0.8 lowered the mean to about 17 Å, and additionally requiring agreement with gCoSRNA's predicted stacking brought it to about 10 Å — an inexpensive filter for groups modeling riboswitches, ribozymes, and other junction-rich RNAs.

#Impact

gCoSRNA is a focused contribution rather than a general-purpose foundation model, but it covers junction orders beyond four-way and revives a capability that had effectively disappeared with the Junction-Explorer server. Its limitations are concrete: stacked pairs outnumber non-stacked ones roughly sixfold in the PDB, biasing predictions toward stacking, and only about one-third of non-stacked two-way pairs in Test Set I were classified correctly. The hand-crafted features constrain scalability, and training on RNA-only structures leaves RNA–protein complexes and ligand-induced conformational shifts under-represented — the authors' explanation for the weak three-way result on the CASP/RNA-Puzzles set. Code and the trained model are released under GPL-3.0.