One RNA design policy now serves many backbones: pairing MCC climbs from 0.61 to 0.73
Synopsis
Structure-conditioned RNA inverse design previously either trained a separate policy per target backbone or relied on offline preference optimization, leaving online post-training of a reusable backbone-conditioned model largely unexplored. MORAD maps six design measurements to bounded desirability scores, combines them with a weighted geometric mean, and uses group-relative normalization to drive a forward-process diffusion update. On 153 evaluation targets it raises GDT-TS from 0.3300 to 0.3827 and pairing MCC from 0.6110 to 0.7295, while cutting normalized ensemble defect from 0.3906 to 0.2830.
Interpretation
A single set of policy weights now serves many target backbones instead of one policy per backbone. Earlier reinforcement learning routes either optimized a separate policy for each target backbone or used offline preference optimization, so online post-training of a reusable backbone-conditioned model was largely unexplored. Across 153 evaluation targets, one checkpoint is compared on tertiary structure, secondary structure and thermodynamics, and MORAD attains the best mean on 14 of the 16 metrics with a preferred direction.
Training first has the policy sample a group of candidate sequences for each target, which a scorer then rewards. The reward no longer scores tertiary structure alone: six measurements spanning tertiary geometry, base pairing, ensemble behavior and sequence composition are each mapped to a bounded desirability score and combined by a weighted geometric mean into one scalar. Group statistics turn that reward into candidate weights for a positive/negative denoising-branch update; structural terms carry 85% of the nominal weight while the three auxiliary terms still directly affect candidate ranking.
The effect of reward design is isolated in a controlled comparison. A control policy rewards tertiary structure only, sharing the same pretrained RIDE, data, optimizer and schedule with the MORAD policy, so the reward is the only difference. The multi-objective reward leads on 10 of the 11 metrics with a preferred direction, including GDT-TS 0.3627 to 0.3694 and MCC 0.6612 to 0.6978; the single exception is MFE-structure frequency, 0.3847 against 0.3880.
Perspective
The experiments evaluate MORAD on a single 10.2M-parameter RIDE backbone, demonstrating the framework on a compact model; the same modular interfaces accommodate larger structure-conditioned backbones. Reward scoring and tertiary evaluation both use the RhoFold+ predictor, consistent with current RNA design reinforcement learning practice, and experimental structural validation remains a separate step beyond this computational study. The reward covers equilibrium folding properties, leaving kinetics and function-specific objectives as future additions.
The evaluation set is built separately from the training set and no normalized evaluation sequence exactly matches a training sequence, yet each evaluation target's maximum sequence identity to the training set has a median of 0.62, so residual similarity is retained rather than excluded. Pairing evaluation uses base pairs extracted from native 3D structures and is available for 137 targets; the other 16 short RNAs have no pairing reference. The reported model is the last checkpoint, with no validation split used for checkpoint selection, and the held-out evaluations every ten updates give 40 observations from a separate run.