MORAD-RNA/MORAD-Targets

MORAD Targets

RNA 3D backbone targets used to train and evaluate MORAD: Reinforcement Learning with Multi-Objective Rewards for RNA Inverse Design.

Paper · Project page · Code · Model

Splits

File	Targets	Length (nt)	Used for
data/train.pt	527	27–258 (median 75)	MORAD training
data/test.pt	153	15–186 (median 65)	evaluation

The test set is built separately from the training set. It comprises 71 RNA3DB entries, drawn from RNA3DB components that contain no training target, and 82 entries curated from the RCSB PDB. No test sequence exactly matches a training sequence. When the native sequences are folded with RhoFold+, the median GDT-TS is 0.6400 on the training set and 0.6014 on the test set.

Format

Each file is a torch.saved dictionary with a targets list. Every target holds

Field	Content
metadata	target_id, native sequence, length, structure_hash and source identifiers
ref_backbone_coords	backbone coordinates (P, C4′, N1/N9), [L, 3, 3]
ref_c4p_coords	C4′ coordinates, [L, 3]
mask_coords	coordinate-validity mask, [L]
raw_record	backbone features consumed by RIDE and the C1′ trace

import torch

test = torch.load("data/test.pt", map_location="cpu", weights_only=False)
target = test["targets"][0]
print(target["metadata"]["target_id"], target["metadata"]["length"])

Download

huggingface-cli download MORAD-RNA/MORAD-Targets --repo-type dataset --local-dir MORAD-Targets

Inside the code repository, bash scripts/download_assets.sh data places the files where the configs expect them. SHA-256 checksums are listed in SHA256SUMS.

Citation

@misc{tang2026morad,
  title         = {{MORAD}: Reinforcement Learning with Multi-Objective Rewards for {RNA} Inverse Design},
  author        = {Tang, Jixin and Guo, Ji and Zhao, Jun},
  year          = {2026},
  eprint        = {2610.13305},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2610.13305}
}