acceptodds
Under review as a conference paper at ICLR 2027

Moirain: Multimodal Alignment and Preference Optimization for Zero-Shot Conditional RNA Generation

Abstract

The design of RNA molecules that interact with specific proteins is a critical challenge in experimental and computational biology. Despite recent progress in natural language modeling and deep learning-based protein design, there remains significant room to improve the frequency of successful interactions and the authenticity of generated sequences for functional applications. In this work, we frame conditional RNA sequence generation as a multi-stage alignment problem, introducing Moirain: a suite of models optimized via multimodal supervised fine-tuning (SFT) and Direct Preference Optimization (DPO). Our approach begins with large-scale pretraining on diverse RNA corpora to capture the fundamental grammars of sequence plausibility. To achieve target-specific generation, we employ a multimodal SFT architecture that conditions RNA synthesis on protein structural and sequential features. Finally, we leverage DPO to refine the model using synthetic interaction data: taking advantage of DPO’s unique ability to navigate non-aligned preference spaces, we improve functional fitness without collapsing the learned natural distribution. Extensive evaluation demonstrates that our framework consistently produces novel, diverse, and biologically plausible RNA sequences with superior binding affinities compared to existing baselines.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
RNA Design, Preference Optimization, SFT, Drug Design, LLM

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_oReBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes Moirain, a model that enables target-protein-conditional RNA sequence generation through pretraining, multimodal SFT, and DPO. The model integrates protein sequence and structural information and utilizes RNA–protein binding data to construct preference training signals. In zero-shot generation tasks, the method achieves good performance in predicting binding affinity and sequence diversity.

Critical strengths
  1. This research question has significant practical value. Generating RNA-binding sequences for unknown protein targets has clear biological implications.

  2. The experimental results are promising. On the OOD test set, Hits improved from 0.46 to 0.61 compared to RNA-BAnG, and ablation experiments also demonstrated the effectiveness of DPO.

Critical weaknesses
  1. The method shows limited innovation. It primarily relies on existing techniques such as pre-training, Cross-Attention, SFT, and DPO, and lacks new optimization mechanisms.

  2. Evidence of biological validity is insufficient. The ability to integrate information relies primarily on DeepCLIP predictions, and there is a lack of independent experimental validation. Additionally, some baselines use different sampling budgets, which affects the fairness of the comparison.

Questions for authors
  1. Can the authors use independent RNA–protein binding prediction models or experimental data to validate the binding capacity of the generated sequences?
  2. Can the authors re-compare Moirain with RNAFlow and RFDPoly under the same sampling budget?
  3. Does DPO offer a clear advantage over supervised fine-tuning using binding data directly?
Reviewer_591POpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Moirain, a multimodal framework for zero-shot protein-conditioned RNA generation. The approach has three stages: pretraining an autoregressive RNA language model on large-scale non-coding RNA data, multimodal supervised fine-tuning using protein sequence and structural representations, and Direct Preference Optimization (DPO) with experimentally derived RNA-binding preference pairs. The framework integrates an RNA-BAnG-based protein encoder with cross-attention conditioning and parameter-efficient fine-tuning. The authors evaluate Moirain using the RNA-BAnG benchmark and DeepCLIP-based binding predictions, reporting improved predicted binding scores over several existing methods. Ablation experiments examine the effects of preference optimization, fine-tuning objectives, and cross-attention placement.

Critical strengths
  1. The paper addresses the practically important task of zero-shot protein-conditioned RNA generation. Combining large-scale RNA pretraining, protein–RNA interaction data, and experimentally derived binding preferences provides a coherent framework for exploiting different types of biological supervision.

  2. Moirain improves OOD [email protected] from 0.46 for RNA-BAnG to 0.61, matching RFDPoly on this metric. The ablation studies also demonstrate that preference optimization, the SFT objective, and cross-attention placement affect the reported performance. These results suggest that the proposed training pipeline is practically effective within the evaluated benchmark.

Critical weaknesses
  1. The architecture and training recipe are largely adapted from existing methods: a LLaMA-style RNA language model, the RNA-BAnG protein encoder, standard cross-attention, LoRA, TOFU, and DPO. Uniformly distributing a small number of cross-attention blocks is a sensible implementation choice, but not a clearly new alignment mechanism. The gain from preference optimization is noteworthy (OOD [email protected]: 0.29 without PO versus 0.61 with DPO; Table 2), yet this demonstrates the value of an additional preference-data training stage rather than the necessity of standard DPO. Comparisons with IPO are included, but compute-matched preferred-only SFT or simpler pairwise ranking objectives are missing. These omissions make it hard to isolate a new, generally useful methodological principle beyond the assembly of established components.

  2. The current ablations do not establish target-specific conditioning: protein-free or shuffled inputs, sequence-only versus structure-aware controls, and cross-target scoring are absent. High diversity (0.96 ± 0.09) indicates dissimilar sequences, not target specificity, nor does it imply random generation. Binding results rely on DeepCLIP predictions. DPO uses experimental binding preferences rather than DeepCLIP labels directly; shared data sources or motifs remain possible, but direct leakage is not demonstrated. An overlap audit and independent computational evaluation would clarify this. Finally, the 3-mer Fidelity distance to the SFT interacting-RNA reference increases from 0.05 (no PO) to 0.29 (DPO; Table 2), revealing a trade-off, not proven mode collapse.

Questions for authors
  1. Beyond the combination of existing architectures and objectives, what is Moirain’s new transferable methodological contribution? Can the authors compare DPO against preferred-only SFT or simpler pairwise ranking objectives, controlling for preference data and computational budget?

  2. How do Hits and target specificity change when the protein condition is removed or shuffled, or when sequence-only input is compared with sequence-plus-structure input? Can the authors provide a cross-target specificity matrix and length/composition-matched random RNA baselines?

  3. Please report whether the DPO preference sets and DeepCLIP evaluators overlap at the level of experiments, proteins/families, or motifs. Can the main gains be checked with an independent computational evaluation (e.g., a separately trained binding predictor and, where appropriate, AF3 complex-interface confidence or validated protein–RNA interface scoring)? Such structure scores should be treated as complementary proxies, not direct binding affinities.

  4. How do the authors reconcile the improvement in predicted binding Hits with the deterioration in 3-mer Fidelity relative to the SFT interacting-RNA reference set? Can they present the performance–Fidelity trade-off across DPO settings without equating high diversity with target specificity?

Reviewer_2syLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Moirain combines natural RNA pretraining, protein-conditioned supervised fine-tuning, and DPO using preference pairs derived from experimentally scored synthetic RNAs. Protein sequence, structure, and confidence features condition the decoder through cross-attention. Evaluation uses DeepCLIP predictions on 59 ID and 12 OOD protein targets, alongside sequence fidelity, novelty, and diversity metrics.

Critical strengths

The curated data and training recipe provide useful computational gains. OOD Hits0.7/0.9 increase from RNA-BAnG's 0.46/0.34 to 0.61/0.51. Moirain matches RFDPoly on several binding columns and has a lower Fidelity distance (0.29 versus 0.52), although RFDPoly has higher Hits0.9 (0.54).

The preference supervision comes from experimental scores, and the paper documents interaction filtering and OOD-related protein exclusions. Architecture, SFT-loss, and IPO ablations are informative. Reported novelty and diversity support sequence variety under the specified search and clustering criteria.

Critical weaknesses

Section 3.3 specifically argues that learning directly from high-scoring synthetic RNAs would be counterproductive, whereas pairwise DPO can emphasize binding motifs while bypassing synthetic background. No PO removes the additional preference-stage supervision as well as the objective. IPO is a useful alternative-objective comparison, but neither comparison tests the stated disadvantage of directly learning from the preferred RNAs. This limits the DPO-specific explanation while leaving the empirical training recipe valuable.

Table 2 shows a binding-fidelity trade-off: OOD Hits0.7 rises from 0.29 without PO to 0.61, while the 3-mer total-variation distance increases from 0.05 to 0.29. This does not prove distributional collapse, synthetic-artifact learning, or loss of biological function. However, novelty against preferred training sequences and the Moirain-Base structural analysis do not by themselves establish selective motif/background learning in the final model. The conclusion already acknowledges reduced structural plausibility and future experimental validation; the remaining concern is the stronger mechanistic interpretation.

Questions for authors
  1. What evidence distinguishes an advantage of pairwise DPO from the benefit of the additional experimentally scored supervision, specifically relative to learning from the same preferred RNAs? A matched-data preferred-only SFT comparison of predicted binding and Fidelity would directly test Section 3.3's motivation.

  2. How should the motif/background interpretation be bounded given the Table 2 fidelity trade-off and Figure 3's focus on Moirain-Base? If the current evidence supports an empirical trade-off, can that be distinguished from demonstrating selective motif extraction?

Comments on the paper

1 comment

We thank the reviewers for their time, effort, and constructive feedback on our submission.

Since the initial submission, our ongoing work on this project has led to substantial revisions that significantly expand our results and alter several of our original conclusions. Because the current version of the manuscript differs markedly from the reviewed draft, we have decided to abstain from submitting a formal rebuttal and instead incorporate your valuable insights into a thoroughly revised version for a future venue.

We greatly appreciate your thoughtful evaluation, which has been very helpful in shaping the next iteration of this work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.