MORAD-RIDE

The released model of MORAD: Reinforcement Learning with Multi-Objective Rewards for RNA Inverse Design: the pretrained RIDE backbone-conditioned RNA diffusion model, post-trained online for 390 updates with the six-objective MORAD reward. One shared policy designs nucleotide sequences for any target 3D backbone, with no per-target optimization.

Project page · Code · Data

Results on the 153 test targets

Model	GDT-TS ↑	TM-score ↑	RMSD (Å) ↓	Pairing MCC ↑	NED ↓
RiboDiffusion	0.3132	0.2839	10.9215	0.3267	0.6056
gRNAde	0.3591	0.3179	8.8374	0.6734	0.3458
RDesign	0.3291	0.2902	10.3441	0.6566	0.3349
RIDE	0.3300	0.2841	9.7199	0.6110	0.3906
MORAD-RIDE	0.3827	0.3400	7.9196	0.7295	0.2830

_{Structures predicted with RhoFold+ (TM-score on C4′), base pairs with EternaFold, normalized ensemble defect (NED) with ViennaRNA. gRNAde at T = 0.1.}

Files

File	Content
morad_ride.pt	policy weights, {"model": state_dict}, 10.2 M parameters
SHA256SUMS	checksum

Quick start

git clone https://github.com/Gabrile166/MORAD.git && cd MORAD
# set up the environments as described in the README, then
bash scripts/download_assets.sh            # RIDE + RhoFold+ weights, data, this model
bash scripts/evaluate_test153.sh morad     # evaluate on the 153 test targets

import torch

state_dict = torch.load("morad_ride.pt", map_location="cpu")["model"]
# RIDE network definition: configs/ride_policy.yaml in the code repository

Training this model from pretrained RIDE takes about 2 hours on 4 × RTX 3090 (bash scripts/train_morad.sh).

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}
}