RIME

RNA-RNA interactions underpin much of post-transcriptional regulation, yet predicting which long transcripts will pair — and where — remains difficult. Classical thermodynamics-based tools scale poorly to long molecules and often miss the sequence-level features that govern real interactions. RIME (RNA Interaction prediction with eMbeddings) addresses this gap with a sequence-based deep learning framework developed in Gian Gaetano Tartaglia's group at the Italian Institute of Technology (IIT) and posted to bioRxiv in February 2025.

The accompanying study highlights the role of low-complexity repeats (LCRs) in RNA regulation, showing that LCRs enable thermodynamically stable interactions with multiple partners and act as hubs in RNA-RNA interaction networks. RIME operationalizes these insights by learning to score interactions directly from sequence rather than relying solely on folding free energies, allowing it to generalize to previously unseen RNA pairs.

RIME is distributed as an open framework with a fixed pretrained checkpoint and a public web server, making long-RNA interaction prediction accessible without model retraining.

#Key Features

• Sequence-based prediction: Predicts interactions between two input RNA molecules directly from sequence, sidestepping the scaling limits of thermodynamics-only methods on long transcripts.
• Language model embeddings: Builds on the Nucleotide Transformer 2.5B multi-species model to represent RNA, transferring pretrained genomic representations to the interaction task.
• Windowed scoring: Computes prediction scores across 200x200 nucleotide windows with a 100-nucleotide step, producing interpretable interaction maps and heatmaps rather than a single score.
• Ready-to-run distribution: Ships a fixed RIMEfull checkpoint plus a wrapper script and a public web server, enabling zero-shot application to new RNA pairs.

#Technical Details

RIME uses embeddings from the Nucleotide Transformer 2.5B multi-species model (roughly a 10 GB dependency that is downloaded automatically) as its input representation, and applies the fixed RIMEfull checkpoint trained on a Psoralen-based crosslinking dataset of RNA-RNA contacts. Inference tiles each query-target pair into 200x200 nucleotide windows advanced by 100 nucleotides, emitting a BEDPE file of scored interaction windows together with heatmap visualizations; a single query-target pair runs in about two minutes on an NVIDIA Quadro RTX 5000 GPU. The code is released under a CC BY-NC 4.0 license for academic and research use, and the work is a preprint that has not yet completed peer review.

#Applications

RIME is intended for RNA biologists studying regulatory networks, competing-endogenous-RNA relationships, and the roles of long non-coding RNAs. By taking two sequences and returning localized interaction scores, it lets researchers screen candidate RNA-RNA pairs, prioritize interacting regions for experimental validation such as crosslinking assays, and investigate how low-complexity repeats mediate multi-partner binding — all through either the code or the hosted web server.

#Impact

By pairing genomic language model embeddings with a windowed scoring scheme, RIME shows that pretrained representations can capture long-range RNA-RNA interaction signals that are hard to recover from folding energetics alone. Its emphasis on low-complexity repeats as interaction hubs offers a mechanistic lens on RNA regulatory networks, and the availability of a fixed checkpoint and public server lowers the barrier for the RNA community to adopt sequence-based interaction prediction. As a non-commercially licensed preprint, its broader validation will continue as the work moves through peer review.