RNAGenScape

This is the official repository for the NeurIPS 2026 paper
RNAGenScape: property-guided, optimized generation of mRNA sequences with manifold Langevin dynamics.

Please raise issues here.

Why would we emphasize "on-manifold"?

Overview of the method

Conceptual comparison

Environment

Requires uv and Python >= 3.12.

cd /path/to/RNAGenScape

# Create / sync the virtualenv from the lockfile
uv sync --python 3.12

# Activate
source .venv/bin/activate

Cluster note (Misha / PyG)

torch-scatter is built from source on clusters with older glibc. Load a newer GCC only while building, then unload it before running Python (leaving GCC loaded can break the torch_scatter ABI):

module load GCC/12.2.0 CUDA/12.2.2
UV_HTTP_TIMEOUT=600 CXX=$(which g++) CC=$(which gcc) uv sync --python 3.12
module unload GCC

Pretrained UTR-LM weights

Oracle training for UTRLM starts from the official pre-trained checkpoint (not domain-specific fine-tuned versions). Download:

mkdir -p checkpoints/utrlm
curl -L -o checkpoints/utrlm/utrlm_pretrained_siss_ep93.pkl \
  "https://raw.githubusercontent.com/a96123155/UTR-LM/main/Model/Pretrained/ESM2SISS_FS4.1_fiveSpeciesCao_6layers_16heads_128embedsize_4096batchToks_lr1e-05_supervisedweight1.0_structureweight1.0_MLMLossMin_epoch93.pkl"

Source: a96123155/UTR-LM.

RhoFold (optional folding confidence)

pLDDT evaluation uses the public ml4bio/RhoFold code and Hugging Face weights. Nothing under lab-private GPFS is required.

# One-time setup (clones into external_src/, downloads checkpoint, installs Bio/etc.)
bash bash/setup_rhofold.sh
# If RhoFold code and weights are already installed, only sync dependencies:
# uv sync --extra rhofold

# Or let eval auto-download the checkpoint once the source tree exists:
# Requires saved generations for the specified experiment.
python src/evaluation/eval_rhofold.py --dataset OpenVaccine --model DiffAb --experiment pos_guided

Defaults:

• Code: external_src/RhoFold
• Weights: external_src/RhoFold_pretrained/RhoFold_pretrained.pt

Override with RHOFOLD_DIR / RHOFOLD_CKPT or --rhofold_dir / --ckpt if you already have a local install.

Data

Dataset loaders read the following paths relative to data/:

Dataset	Input file(s)	Sequence / label columns	Sequence length
OpenVaccine	OpenVaccine/train.csv	sequence / reactivity_mean	107
Zebrafish (proprietary; not included)	Zebrafish/MPRA_mean_translation_2hpf_pa_Fish5UTR.csv	sequence / translation	124
RibosomeLoading	RibosomeLoading/MRL_Random50Nuc_SynthesisLibrary_Sample/4.10_train_data_GSM3130438_egfp_pseudo_2.csv and 4.10_test_data_GSM3130438_egfp_pseudo_2.csv in the same directory	utr / rl	50

OpenVaccine and the RibosomeLoading train/test files are tracked. The large RibosomeLoading training CSV uses Git LFS; install Git LFS and retrieve the data after cloning:

git lfs install
git lfs pull

The Zebrafish CSV must be supplied separately at the path above. data/ is ignored for additional untracked files; this does not exclude data files already tracked by Git.

OpenVaccine and Zebrafish use seeded 80% / 10% / 10% train/validation/test splits. RibosomeLoading uses the predefined train/test files (260,000 / 20,000 rows), with 15% of the training file reserved for validation. Label normalization is fitted on training labels and applied to validation and test labels.

Experiments

The workflow for each dataset is:

1. Oracle. Train or load a separate property predictor. Use validation for checkpoint selection and the held-out test set for evaluation. For RNAGenScape, this oracle evaluates generated sequences; the OAE's own property head guides generation.
2. Train the OAE. Optimize the training split jointly for property prediction and sequence reconstruction. Validation selects the checkpoint and controls early stopping; it is not used for gradient updates.
3. Prepare the manifold projector. Encode training sequences with the trained OAE. Optionally augment these latents with SUGAR (sugar_w > 0; default is off). Train a DAE projector, or cache the reference latents for a kNN projector. kNN has no trainable parameters but still requires this preparation step.
4. Generate. Encode unseen test sequences, run property-guided Langevin updates in latent space, apply the selected projector after every step when projection is enabled, and decode the final latents to sequences.
5. Evaluate. Score start and generated sequences with the oracle held fixed during evaluation. Report property changes, Hamming distances, novelty, held-out distances, and optional folding confidence.

The default OAE is a deterministic residual convolutional autoencoder with 16 positional pooling bins and a 128-dimensional continuous latent vector. It has a sequence decoder and a property regression head, both operating on the same latent vector. Training minimizes reg_w * MSE + recon_w * CE, with defaults reg_w=1 and recon_w=5. Checkpoint selection maximizes 0.5 * (val_pearson + val_spearman) + val_token_acc.

SUGAR augments the latent reference set for projector preparation; it does not train the OAE. Reconstruction is measured using both token accuracy and exact-sequence accuracy.

Citation

If you use RNAGenScape, please cite the paper:

@inproceedings{liao2026rnagenscape,
  title={RNAGenScape: Property-Guided, Optimized Generation of mRNA Sequences with Manifold Langevin Dynamics},
  author={Liao, Danqi and Liu, Chen and Sun, Xingzhi and Tang, Di{\'e} and Wang, Haochen and Youlten, Scott and Gopinath, Srikar Krishna and Lee, Haejeong and Strayer, Ethan C and Giraldez, Antonio J and others},
  booktitle={Advances in neural information processing systems},
  year={2026},
}