Genomics & Single-Cell · GENTEL Lab, Shanghai Jiao Tong University
La-Ribo
Generates de novo RNA sequences and heavy-atom 3D structures jointly from a specified sequence length.
What La-Ribo is
La-Ribo is a computational model developed by the GENTEL Lab at Shanghai Jiao Tong University for the simultaneous generation of RNA primary sequences and three-dimensional heavy-atom coordinates. Rather than solving sequence design for a pre-specified backbone or predicting a structure from a fixed sequence, it treats sequence and geometry co-design as a single unified generation task.
The model addresses the limitation of decoupled RNA design pipelines by generating an unconstrained structural scaffold and residue identity in one integrated workflow. It operates unconditionally, requiring only the desired nucleotide sequence length as input to generate plausible complete RNA models.
Researchers can use it to explore RNA structural space de novo across lengths typically ranging from 40 to 150 nucleotides, providing candidate conformations and sequences for downstream analysis, independent refolding, and computational screening.
How it works
La-Ribo uses a 12-layer geometry–latent flow matching architecture to jointly sample a coarse phosphate–sugar–base backbone alongside a 16-dimensional continuous latent vector per residue. A shared, scaffold-conditioned autoencoder then decodes this joint latent representation into discrete nucleotide identities and full heavy-atom 3D coordinates in a single forward pass.
Generation uses numerical integration over 200 Euler steps with exponential moving average (EMA) weights, calibrated using a scaffold power of 1.25 and a latent power of 2.0. Two model variants are provided: La-Ribo-Tri, which integrates triangle pair updates every two layers using equivariant kernels, and La-Ribo-Base, which omits pair updates to achieve faster sampling times.
Key facts
- Architecture
- 12-layer flow matching model with a shared scaffold-conditioned autoencoder
- Latent dimension
- 16 dimensions per residue
- Model variants
- La-Ribo-Base and La-Ribo-Tri
- Default sampling steps
- 200 Euler steps
- Benchmark length range
- 40 to 150 nucleotides
- Inputs
- Sequence length, random seed, model variant
- Outputs
- Heavy-atom PDB coordinates, FASTA sequences, and summary metrics
- Hardware requirements
- 8 GB GPU memory (NVIDIA CUDA) or CPU
What you can use it for
- De novo RNA co-design
- Generate unconstrained RNA sequences and matching heavy-atom structural models simultaneously by providing an intended nucleotide length.
- Structural space exploration
- Sample structural topologies across lengths between 40 and 150 nucleotides to produce diverse backbone arrangements and compositions.
- Library generation for in silico screening
- Produce large ensembles of RNA sequence-structure pairs for validation with independent secondary and tertiary structure prediction tools.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on On-demand GPU
Runs on our servers with on-demand compute. A first run needs time to load the model; active capacity can be reused and scales down when idle.Run this model
La-Ribo
Strengths and limits
Good at
- Generates both sequence identities and heavy-atom coordinates together in a single pass without needing a fixed backbone template.
- Provides a choice between a triangle-updated architecture (La-Ribo-Tri) and an accelerated variant (La-Ribo-Base).
- Outputs complete heavy-atom PDB coordinates alongside physical metrics such as radius of gyration, GC content, and mean C4'–C4' spacing.
- Runs on standard hardware with an 8 GB GPU memory footprint, with support for both CPU and CUDA execution.
Watch out for
- Generation is strictly unconditional; it cannot accept target folds, binding motifs, target ligands, or functional constraints.
- Does not predict per-residue confidence scores or pLDDT values, setting the output PDB B-factor column to zero.
- Sampled sequences may reproduce recognisable sequence motifs from the training sets (bpRNA and RNAStrAlign), requiring novelty verification.
- Generated models must be refolded and validated using external tools, as the generation step does not evaluate designability.
- Identical seeds across CPU and CUDA environments do not produce identical atomic coordinates due to backend differences in random-number generation.
Inputs and outputs
You provide
Length (nucleotides)
The paper benchmarks 40–150 nt and the released training structures span 16–252 nt. Outside that range the result says so.
Samplesoptional
Independent RNAs from one batch. Every sequence is listed; the viewer shows the first, because the model offers no score to rank them by. A batch too large to finish in the time the run allows is refused before it starts, with the number that would fit.
Variantoptional
Seedoptional
The same seed and length return the same batch. Change it for a different one.
You get back
Heavy-atom coordinates for the first sample. The B-factor column is zero: the model predicts no confidence.
Each sample's sequence, GC content, radius of gyration and mean C4'–C4' spacing.
Every sampled sequence as FASTA, with what the model can and cannot be asked for.
Licence
Commercial use permitted · MIT
MIT on both halves, read on 2026-10-29: the model card declares `license: mit` and its text says the release is under the MIT License, and the code repository carries an MIT LICENSE file (GENTEL-lab/La-Ribo). The repository is not gated. The 10,631 structures published beside the weights are trRosettaRNA2 predictions and are not used here — only the weights and the sampling code are.
Questions
Can I use La-Ribo commercially?
Commercial use permitted. MIT on both halves, read on 2026-10-29: the model card declares `license: mit` and its text says the release is under the MIT License, and the code repository carries an MIT LICENSE file (GENTEL-lab/La-Ribo). The repository is not gated. The 10,631 structures published beside the weights are trRosettaRNA2 predictions and are not used here — only the weights and the sampling code are.
How do I run La-Ribo without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 1× A100 40GB, typically in 15s, and a run costs about 28 credits. Visitors without an account get 1,000 credits a month.
Can I condition generation on a specific secondary structure or functional motif?
No. The model is strictly unconditional, taking only sequence length as input. Specifying a length associated with an aptamer or riboswitch produces an RNA of that length, but not one engineered to bind or switch.
What is the difference between La-Ribo-Base and La-Ribo-Tri?
La-Ribo-Tri applies triangle multiplication pair updates every two layers and is the default variant. La-Ribo-Base omits these pair updates, making it significantly faster to run.
Does the output PDB file include confidence estimates like pLDDT?
No. The model does not estimate confidence, and the B-factor column in the output PDB file is set to zero.
How should I assess the quality of generated RNA candidates?
You should check sequences for novelty against training databases such as bpRNA and RNAStrAlign, and perform independent refolding and designability checks using external tertiary structure prediction software.
How to cite it
Ma, Runze, Will Hua, and Shuangjia Zheng. "La-Ribo: RNA Co-Design via Geometry-Latent Flow Matching." arXiv preprint arXiv:2610.12236 (2026).
Written from La-Ribo's own documentation (model card, code, paper). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.