Model Overview¶
RNAZoo includes 23 RNA deep learning models across 6 tracks. Each model runs in its own Docker container with baked-in weights.
All models at a glance¶
| Model | Track | Task | Input | Output | Device | License |
|---|---|---|---|---|---|---|
| RiboNN | Translation | TE prediction (82 cell types) | Tab-separated (UTR+CDS) | TSV with TE per cell type | CPU/GPU | Apache 2.0 |
| Riboformer | Translation | Codon-level ribosome density | WIG + FASTA + GFF3 | Density predictions | CPU/GPU | Upstream |
| RiboTIE | Translation | ORF detection from ribo-seq | FASTA + GTF + BAM | GTF + CSV | CPU/GPU | Upstream |
| TranslationAI | Translation | TIS/TTS/ORF prediction | FASTA (mRNA) | TIS/TTS/ORF text files | CPU/GPU | AGPL-3.0 + CC BY-NC 4.0 |
| Saluki | Translation | mRNA half-life | FASTA (case=UTR/CDS) | NumPy array | CPU/GPU | Apache 2.0 |
| CodonTransformer | Translation | Codon optimization | FASTA (protein) | FASTA (DNA) | CPU/GPU | Apache 2.0 |
| RNA-FM | Foundation | RNA embeddings (640-d) | FASTA (RNA) | NumPy (N x 640) | CPU/GPU | MIT |
| RiNALMo | Foundation | RNA embeddings (1280-d) | FASTA (RNA) | NumPy (N x 1280) | CPU/GPU | Apache 2.0 |
| ERNIE-RNA | Foundation | Structure-aware embeddings (768-d) | FASTA (RNA) | NumPy (N x 768) | CPU/GPU | MIT |
| Orthrus | Foundation | Mamba mRNA embeddings (3 variants: 256–512-d) | FASTA (RNA) | NumPy (N x D) | GPU only | MIT |
| HydraRNA | Foundation | Hybrid Hydra-SSM + MHA full-length RNA embeddings (1024-d) | FASTA (RNA, ≤10K nt) | NumPy (N x 1024) | GPU only | MIT |
| RNAErnie | Foundation | Motif-aware RNA embeddings (768-d) | FASTA (RNA) | NumPy (N x 768) | CPU/GPU | Apache-2.0 (HF port) |
| PlantRNA-FM | Foundation | Plant-only RNA embeddings (480-d) | FASTA (RNA) | NumPy (N x 480) | CPU/GPU | MIT |
| CaLM | Foundation | Codon-level RNA embeddings (768-d) | FASTA (CDS, codon-aligned) | NumPy (N x 768) | CPU/GPU | BSD-3-Clause |
| mRNABERT | Foundation | Hybrid UTR/CDS mRNA embeddings (768-d) | FASTA (mRNA, auto-ORF) | NumPy (N x 768) | CPU/GPU | Apache-2.0 |
| RhoFold | Structure | 3D structure prediction | FASTA (RNA) | PDB + CT | CPU/GPU | Apache 2.0 |
| SPOT-RNA | Structure | 2D structure + pseudoknots | FASTA (RNA) | bpseq + CT + prob + dot-bracket | CPU/GPU | MPL-2.0 |
| DRfold2 | Structure (Tier 2) | Single-seq ab initio 3D | FASTA (RNA) | PDB | GPU only | MIT |
| Pangolin | Splicing | Tissue-specific variant-effect splice scores | VCF/CSV + reference FASTA + gffutils DB | Annotated VCF/CSV | CPU/GPU | GPL-3.0 |
| SpliceAI | Splicing | Variant-effect splicing predictions | VCF + reference FASTA + annotation | Annotated VCF (4-class delta) | CPU/GPU | PolyForm Strict + CC BY-NC 4.0 |
| SpliceBERT | Splicing | Vertebrate primary-RNA embeddings (512-d) | FASTA (RNA) | NumPy (N x 512) | CPU/GPU | BSD-3-Clause |
| MultiRM | Modification | 12 RNA modification types | FASTA (RNA, min 51 nt) | TSV (probabilities + p-values) | CPU/GPU | MIT |
| UTR-LM | mRNA Design | MRL / TE / expression level | FASTA (5'UTR DNA) | TSV (predictions) | CPU/GPU | GPL-3.0 |
By track¶
Translation (6 models)¶
Models for predicting translation efficiency, ribosome profiling, ORF detection, mRNA stability, and codon optimization.
- RiboNN — Multi-task TE prediction across 82 human cell types from mRNA sequence
- Riboformer — Refine codon-level ribosome densities from ribo-seq data
- RiboTIE — Detect translated ORFs from ribo-seq + genomic sequence
- TranslationAI — Identify translation initiation/termination sites and ORFs
- Saluki — Predict mRNA half-life from sequence (50-model ensemble)
- CodonTransformer — Optimize codon usage for 164 organisms
RNA Foundation Models (9 models)¶
General-purpose RNA language models that produce embeddings for downstream tasks.
- RNA-FM — 99M params, 640-d embeddings, max 1022 nt (MIT)
- RiNALMo — 650M params, 1280-d embeddings, no hard length limit (Apache 2.0)
- ERNIE-RNA — 86M params, 768-d embeddings, structure-aware attention (MIT)
- Orthrus — Mamba SSM (~10M params), 256–512-d embeddings (3 4-track variants), unbounded input (linear memory), GPU only (MIT)
- HydraRNA — Hybrid Hydra-SSM + MHA (~84M params), 1024-d embeddings, ≤10K nt, GPU only (MIT)
- RNAErnie — 12-layer transformer, 768-d embeddings, motif-aware MLM, max 2046 nt (Apache-2.0 HF port)
- PlantRNA-FM — 35M-param ESM transformer, 480-d embeddings, plant-only training (1124 species), max 1024 nt (MIT)
- CaLM — 12-layer codon-level transformer (~86M params), 768-d embeddings, max 1024 codons (~3 kb) (BSD-3-Clause)
- mRNABERT — 12-layer MosaicBERT (~86M params) with hybrid UTR/CDS tokenization + ALiBi extrapolation, 768-d embeddings, max 1024 tokens (Apache-2.0)
RNA Structure (3 models)¶
Secondary and 3D structure prediction from sequence.
- RhoFold — Full-atom 3D structure prediction (PDB output), single-sequence mode
- SPOT-RNA — 2D structure with pseudoknots, 5-model TF ensemble
- DRfold2 — Single-seq ab initio 3D (Tier 2), composite LM + 4-model ensemble + Arena refinement; novel-fold wins vs RhoFold/RoseTTAFoldNA, GPU only
Splicing (3 models)¶
- Pangolin — Variant-effect splicing predictor trained on 4 tissues (heart/liver/brain/testis) across human + mouse + rhesus + rat; complements SpliceAI with tissue-aware scores (GPL-3.0)
- SpliceAI — 32-layer dilated CNN, ±5 kb context per variant; 4-class delta scores (acceptor/donor gain/loss) on annotated genes (PolyForm Strict + CC BY-NC 4.0; non-commercial only)
- SpliceBERT — 6-layer/512-d BERT pretrained on 2M+ vertebrate primary RNAs (72 species), 1024 nt context (BSD-3-Clause)
RNA Modification (1 model)¶
- MultiRM — Predicts 12 RNA modification types per position (m6A, m5C, pseudouridine, Am, Cm, Gm, Um, m1A, m5U, m6Am, m7G, A-to-I editing)
mRNA Design (1 model)¶
- UTR-LM — Predicts mean ribosome loading, translation efficiency, or expression level from 5'UTR sequences
Fine-tuning support¶
Some models can be fine-tuned on your own data. The 9 foundation models share a generic head trainer (linear / MLP / XGBoost, regression or classification) — see the Fine Tuning guide for the full how-to. Three other models have custom training pipelines on their own pages.
| Model | Fine-tuning | Details |
|---|---|---|
| 10 foundation models | Generic head trainer | Frozen backbone + linear / MLP / XGBoost head on user (sequence, label) data. Regression or classification (auto-detected). Optional precomputed-embeddings shortcut. See the Fine Tuning guide. |
| RiboNN | Transfer learning | Freeze pretrained conv layers, train head on user TE data; use saved checkpoint via --ribonn_checkpoint |
| UTR-LM | Full backbone | Train ESM2 backbone + head on user MRL/TE/EL data; use saved checkpoint for prediction |
| RiboTIE | Built-in | Automatically fine-tunes on user ribo-seq BAMs before ORF prediction |