EVA (Evolutionary Versatile Architect) is a generative RNA foundation model trained on OpenRNA v1, a curated atlas of 114 million full-length RNA sequences spanning all domains of life. Built on a 1.4B-parameter decoder-only Transformer with a Mixture-of-Experts (MoE) backbone and an 8,192-token context window, EVA unifies RNA sequence scoring and controllable design within a single framework.
You should consider EVA for the reasons as follows:
| 🔓 | Fully Open-Sourced | All training data, model weights, finetuning & training & inference codes and details are publicly released — full transparency for the community to reproduce, build upon, and extend |
| 📏 | 8x Larger Context Window | 8,192-token context window vs. ~1,024 in prior RNA models — enabling full-length RNA processing without truncation or information loss |
| 🗄️ | 7x More Training Data | Trained on OpenRNA v1 — 114M full-length RNA sequences across all domains of life (Eukaryota, Bacteria, Archaea, Viruses), 7.7x larger than the RNAcentral v25.0 |
| 🏆 | SOTA Fitness Prediction | Achieves state-of-the-art zero-shot fitness prediction on ncRNA and mRNA, no task-specific fine-tuning needed |
| 🎯 | 10x+ RNA Generation Accuracy | Over 10x improvement in RNA generation accuracy at both sequence and structure level compared to prior methods |
| 🧬 | 11 RNA Types Supported | Controllable generation across 11 RNA classes (mRNA, tRNA, rRNA, lncRNA, snRNA, snoRNA, miRNA, and more) conditioned on RNA type and taxonomic lineage |
| ⚙️ | Novel Architecture, Capabilities & Training | 1.4B-parameter MoE decoder-only Transformer with dual generation paradigms — CLM for de novo design & GLM for domain redesign — unified in a single model |
| Module | Path | Description |
|---|---|---|
| Model Architecture | eva/ |
MoE Transformer backbone (modeling, attention, MoE routing, causal LM head) |
| Lineage Tokenizer | eva/lineage_tokenizer.py |
Tokenizer for taxonomic lineage strings |
| Generation CLI | tools/generate.py |
Entry point for CLM and GLM sequence generation |
| Scoring CLI | tools/predict.py |
Entry point for log-likelihood scoring |
| Directed Evolution | tools/directed_evolution.py |
In-silico directed evolution pipeline |
| Generators | tools/utils/generators/ |
CLM autoregressive and GLM span-infilling implementations |
| Scorers | tools/utils/scorers/ |
Sequence scoring logic and batch workers |
| Condition Control | tools/utils/conditions/ |
RNA type and taxonomic lineage conditioning |
| Model Loader & Sampler | tools/utils/model/ |
Checkpoint loading and sampling strategies |
| Config Templates | config/ |
YAML configuration examples for generation, scoring, and directed evolution |
Click below to expand the Table of Contents and explore each section in detail.
Table of Contents
- Quick Start
- Condition Control
- Generation
- Scoring
- Directed Evolution
- Batch Processing with YAML
- Input/Output Formats
- Data Availability
- Citation
- License
From Zenodo (link coming soon):
# Download all split archives (part000-part011), then:
cat eva_latest.tar.gz.part* > eva_latest.tar.gz
docker load -i eva_latest.tar.gzOr build from Dockerfile:
cd /data/yanjie_huang/eva/EVA1/docker
docker build -t eva:latest .docker run --gpus all -it eva:latest bashAll model files and output data stay inside the container.
huggingface-cli download GENTEL-Lab/EVA --local-dir ./checkpointpython /eva/tools/generate.py \
--checkpoint ./checkpoint \
--format clm \
--rna_type mRNA \
--taxid 9606 \
--num_seqs 10 \
--output ./output/demo.faThis generates 10 human mRNA sequences and saves them to ./output/demo.fa.
EVA supports conditioning on RNA type and species/lineage for both generation (generate.py) and scoring (predict.py). These conditions can be used independently or combined.
| RNA Type | Description |
|---|---|
| mRNA | Messenger RNA - carries genetic information from DNA to ribosomes |
| tRNA | Transfer RNA - brings amino acids to the ribosome during translation |
| rRNA | Ribosomal RNA - forms the core of the ribosome structure |
| miRNA | MicroRNA - regulates gene expression |
| lncRNA | Long non-coding RNA - various regulatory functions |
| circRNA | Circular RNA - circularized RNA molecules |
| snoRNA | Small nucleolar RNA - modifies other RNAs |
| snRNA | Small nuclear RNA - involved in splicing |
| piRNA | PIWI-interacting RNA - silences transposons |
| sRNA | Small RNA - general category for small RNA molecules |
| viral_RNA | RNA from viruses |
Species can be specified in three ways: --taxid, --species, or --lineage (Greengenes format).
Common species:
| TaxID | Species |
|---|---|
| 9606 | Homo sapiens (Human) |
| 10090 | Mus musculus (Mouse) |
| 10116 | Rattus norvegicus (Rat) |
| 7227 | Drosophila melanogaster (Fruit fly) |
| 6239 | Caenorhabditis elegans (Nematode) |
| 3702 | Arabidopsis thaliana (Plant) |
| 4932 | Saccharomyces cerevisiae (Yeast) |
| 562 | Escherichia coli (Bacteria) |
10x Higher Modeling Accuracy Than Ever Before — At Both Sequence and Structure Level.
RNA landscape modeling comparison |
Species-specific modeling comparison |
CLM (Causal Language Model) generates RNA sequences autoregressively from left to right. This is the primary generation mode in EVA.
Generate sequences without any biological constraints:
python /eva/tools/generate.py \
--checkpoint /path/to/model \
--format clm \
--num_seqs 1000 \
--output /output/unconditional.faEVA supports conditioning on RNA type, species (via TaxID, species name, or lineage string), or both. See Condition Control for the full list of supported RNA types and species.
# RNA type only
python /eva/tools/generate.py \
--checkpoint /path/to/model \
--format clm \
--rna_type mRNA \
--num_seqs 1000 \
--output /output/mrna.fa
# Species only (via TaxID)
python /eva/tools/generate.py \
--checkpoint /path/to/model \
--format clm \
--taxid 9606 \
--num_seqs 1000 \
--output /output/human.fa
# Both RNA type and species
python /eva/tools/generate.py \
--checkpoint /path/to/model \
--format clm \
--rna_type mRNA \
--taxid 9606 \
--num_seqs 1000 \
--output /output/human_mrna.faSpecies can also be specified via --species homo_sapiens or --lineage "D__Eukaryota;P__Chordata;..." in Greengenes format.
Extend existing sequences in either direction. Use --split_ratio (fraction) or --split_pos (exact position) to control the split point.
Forward (extend 3' end):
python /eva/tools/generate.py \
--checkpoint /path/to/model \
--format clm \
--input /input/partial_seq.fa \
--direction forward \
--split_ratio 0.5 \
--num_seqs 5 \
--output /output/continuation.faReverse (extend 5' end):
python /eva/tools/generate.py \
--checkpoint /path/to/model \
--format clm \
--input /input/partial_seq.fa \
--direction reverse \
--split_pos 699 \
--num_seqs 20 \
--output /output/reverse_continuation.faAdd --output_details to include prompt, ground truth, and generated content in the output.
GLM (General Language Model) performs span infilling — it masks a region within an existing sequence and generates what should fill the gap based on surrounding context. Like CLM, GLM supports both unconditional and conditional generation.
Fill in a masked region without any biological constraints:
python /eva/tools/generate.py \
--checkpoint /path/to/model \
--format glm \
--input /input/sequences.fa \
--span_ratio 0.1 \
--num_seqs 5 \
--output /output/glm_output.faCondition on RNA type and/or species to generate biologically consistent infills:
python /eva/tools/generate.py \
--checkpoint /path/to/model \
--format glm \
--input /input/sequences.fa \
--rna_type mRNA \
--taxid 9606 \
--span_ratio 0.2 \
--num_seqs 5 \
--output /output/glm_conditional.fa| Parameter | Description | Example |
|---|---|---|
--span_length |
Fixed number of nucleotides to mask | --span_length 20 |
--span_ratio |
Fraction of sequence to mask | --span_ratio 0.1 |
--span_position |
Where to place the span: random or specific index |
--span_position 100 |
--span_id |
Which span token to use: random or 0-49 |
--span_id 0 |
| Parameter | Description | Recommended Range |
|---|---|---|
--temperature |
Controls randomness. Lower = more deterministic, higher = more diverse | 0.1 - 1.5 |
--top_k |
Only consider the top k most likely nucleotides at each position | 10 - 100 |
--top_p |
Nucleus sampling — consider smallest set of nucleotides whose cumulative probability exceeds p | 0.8 - 0.95 |
Example with all sampling parameters:
python /eva/tools/generate.py \
--checkpoint /path/to/model \
--format clm \
--temperature 0.8 \
--top_k 50 \
--top_p 0.9 \
--num_seqs 100 \
--output /output/sampled.faEVA achieves state-of-the-art zero-shot fitness prediction across both mRNA and ncRNA benchmarks, substantially outperforming existing RNA, codon, and DNA language models — including models with up to 40B parameters. This strong sequence-level understanding directly translates into high-quality generation: EVA's learned distribution faithfully captures the evolutionary constraints of natural RNA, enabling it to produce biologically realistic sequences out of the box.
|
|
Evaluate how well a given sequence fits the model's learned distribution by computing its log-likelihood. Higher (less negative) scores indicate more probable sequences.
Score RNA sequences and compute per-sequence log-likelihood:
python /eva/tools/predict.py \
--checkpoint /path/to/model \
--input /input/sequences.fa \
--output /output/scores.jsonSupports --rna_type and --taxid conditioning, same as generation.
Score protein sequences by reverse-translating them to RNA first:
python /eva/tools/predict.py \
--checkpoint /path/to/model \
--input /input/proteins.fa \
--output /output/protein_scores.json \
--mode protein \
--codon_optimization first--codon_optimization options: first (first codon in table) or most_frequent (most common codon for the species).
EVA supports in-silico directed evolution — an iterative optimization pipeline that improves a given RNA sequence through cycles of mutation, LLM-based fitness scoring, and structural stability evaluation. Starting from an existing RNA sequence, EVA applies point mutations at specified positions, scores each mutant by log-likelihood and Minimum Free Energy (MFE via LinearFold), and uses simulated annealing with dynamic beam search to balance exploration and exploitation. Over multiple rounds the temperature cools to converge on optimized sequences, and the top N candidates ranked by combined LLM + MFE score are returned.
|
|
|
|
python /eva/tools/directed_evolution.py \
--checkpoint /path/to/model \
--input /input/sequence.fa \
--output /output/evolved.fa \
--rna_type mRNA \
--taxid 9606 \
--iterations 10 \
--mutations 5 \
--beam_width 10 \
--output_count 5| Parameter | Description | Default |
|---|---|---|
--iterations |
Number of evolution cycles | 10 |
--mutations |
Total mutations per iteration | 2 |
--beam_width |
Beam search width — keeps top candidates after each position | 10 |
--output_count |
Number of final sequences to output | 5 |
--T_init / --T_min |
Simulated annealing temperature range | 1.0 → 0.01 |
--cooling_rate |
Temperature decay per iteration | 0.95 |
--mutate_positions |
Specific positions to mutate (0-based, comma-separated) | Random |
--mutate_range |
Range of positions to mutate (e.g., 0-100) |
Entire sequence |
Define multiple tasks in a single YAML config file. The defaults section sets shared parameters, which individual tasks can override.
checkpoint: /path/to/model
output_dir: ./output
defaults:
temperature: 1.0
top_k: 50
max_length: 8192
batch_size: 1
tasks:
- name: unconditional
mode: generation
format: clm
num_seqs: 1000
- name: human_mrna
mode: generation
format: clm
rna_type: mRNA
taxid: "9606"
lineage: "D__Eukaryota;P__Chordata;C__Mammalia;O__Primates;F__Hominidae;G__Homo;S__Homo sapiens"
num_seqs: 1000
- name: glm_infill
mode: generation
format: glm
input: ./input/seqs.fa
span_ratio: 0.1
num_seqs: 5checkpoint: /path/to/model
output_dir: ./scores
defaults:
batch_size: 128
tasks:
- name: score_basic
mode: scoring
input: ./input/seqs.fa
- name: score_human_mrna
mode: scoring
input: ./input/seqs.fa
rna_type: mRNA
taxid: "9606"
normalize: true
exclude_special_tokens: true
- name: score_protein
mode: scoring
input: ./input/proteins.fa
scoring_mode: protein
codon_optimization: firstpython /eva/tools/generate.py --config config.yaml # Run all tasks
python /eva/tools/generate.py --config config.yaml --task name # Run specific task
python /eva/tools/generate.py --config config.yaml --device cuda:1 # Override device>sequence_id_1
AUGCGCUAUGCGCUAUGCGCUAUGCGCUAUGCGCUAUGCGCUAUGCGCU
>sequence_id_2
AUGAAAAUGCGGCCGCAUUACGUAAACGGCCGCAAAUGUUUCCGGCAAA
>unconditional_0
AUGCGCUAUGCGCUAUGCGCUAUGCGCUAUGCGCUAUGCGCUAUGCGCU
With --output_details (GLM / Continuation):
>test_seq_sample0_forward_split50
PROMPT: AUGCGCUAUGCGCUAUGCG
GROUND_TRUTH: CU AUGCGCUAUGCG
GENERATED: CU AAUGCGCUAGCG
FULL_SEQ: AUGCGCUAUGCGCUAAUGCGCUAGCG
{
"scores": [
{
"header": "seq1",
"sequence": "AUGGCCGUAGU...",
"length": 67,
"log_likelihood": -1.25
}
]
}Higher (less negative) log_likelihood = better sequence.
Output is in FASTA format with scores in the header:
>candidate_1_score=-125.4321_best
AUGC... (evolved sequence)
>candidate_2_score=-128.1056
AUGC... (second best)
Some large files are not included in this repository due to size constraints. They can be downloaded from the sources listed below.
| Resource | Description | Size | Source |
|---|---|---|---|
| Model Checkpoint | EVA 1.4B (MoE) pre-trained weights | HuggingFace | |
Docker Image (eva_latest.tar.gz) |
Pre-built EVA runtime environment | Zenodo | |
| Training Data | OpenRNA v1 — 114M full-length RNA sequences | HuggingFace |
| Resource | Description | Size | Source |
|---|---|---|---|
| mRNA Codon Optimization | Optimized sequences from 9 vaccine targets (5 mRNA: linear, HIV, PR8, RABV, VZV; 4 circRNA: CAR-T, RABV-G, SARS-Group-I, SARS-T4) using EVA (homo/nolineage), CodonFM-1B, Evo2-1B, GemoRNA, and random baseline | 26 KB | Zenodo |
| mRNA 6 Species | Generated & natural mRNA sequences for 6 species (Homo sapiens, Mus musculus, Rattus norvegicus, Cricetulus griseus, Drosophila melanogaster, Caenorhabditis elegans) with MFE, GC content, k-mer, loop-helix, pair metrics | 38 MB | Zenodo |
| GeneRRNA Generation | EVA-generated rRNA sequences benchmarked against GeneRRNA database (100k natural, 1k sampled) with k-mer, loop-helix, pair, merged MFE-GC metrics | 13 MB | Zenodo |
| Generated RNA Sequences | 15 RNA types (mRNA, lncRNA, circRNA, tRNA, rRNA, miRNA, piRNA, sRNA, snRNA, snoRNA, scaRNA, vault_RNA, Y_RNA, ribozyme, viral) generated by EVA with matched natural controls (5k each) and MFE, GC, k-mer, loop-helix, pair metrics | 156 MB | Zenodo |
If you find EVA useful in your research, please cite:
@article{huang2026eva,
title={EVA: A Generative Foundation Model for Universal RNA Modeling and Design},
author={Huang, Yanjie and Lyu, Guangye and others},
journal={TODO},
year={2026},
url={TODO}
}This project is licensed under the Apache License 2.0. See LICENSE for details.