acceptodds
Withdrawn from ICLR 2027

RNAOmni: An Autonomous Modeling Agent for RNA Downstream Tasks

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

As a foundational molecule of life, RNA property prediction and functional identification represent a vital frontier at the intersection of AI and biology. However, developing high-performance models for downstream RNA tasks still relies heavily on labor-intensive iterations of model design, validation, and refinement. Given the growing demand for task-specific modeling, domain experts possess rich data but lack engineering bandwidth, while AI researchers have strong modeling capabilities yet struggle to address the diverse domain-specific needs. Existing AutoML methods partially alleviate this burden but remain restricted to hyperparameter tuning within fixed search spaces, and LLM-based agents depend on human-provided blueprints, neither of which achieves full end-to-end automation from data to model. To address these limitations, we introduce RNAOmni, an end-to-end autonomous modeling agent for RNA prediction tasks that mimics expert decision-making by jointly exploring data processing strategies, feature representations, model backbones, and hyperparameters, starting solely from raw data. Furthermore, RNAOmni supports both Direct and Thinking modes, incorporating an attribution probe within the latter to systematically identify modeling weaknesses and guide efficient exploration. Across 17 RNA downstream tasks, RNAOmni achieves average relative improvements of 59.50% over AutoML methods and 18.4% over AI research agents. Its generated solutions consistently surpass domain-specific RNA models and further enhance the utility of RNA foundation models. These results demonstrate that RNAOmni enables a new paradigm for RNA downstream modeling by shifting the field from expert-dependent engineering toward autonomous, data-driven model development. Code will be available upon acceptance.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI research agent, autonomous modeling, RNA modeling, bioinformatics

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_DKJfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes RNAOmni, an Agent for downstream RNA prediction tasks, which can perform data processing, feature selection, model design, and hyper-parameter optimization. The method incorporates Attribution Probes to guide model iteration based on historical experimental results. Across 17 RNA tasks, it achieved average performance improvements of 59.5% and 18.4% compared to AutoML and the AI Research Agent, respectively.

Critical strengths
  1. It automates the process from raw data to model building and achieves good predictive performance across multiple RNA-related tasks.

  2. Comprehensive experiments. By comparing the method with various AutoML and AI agents, they demonstrate the applicability of RNAOmni.

Critical weaknesses
  1. Limited innovation: The core framework primarily combines existing techniques for automatic feature engineering, model search, and feedback optimization. The technical innovation of Attribution Probe remains insufficient.

  2. Key modules are not sufficiently validated. The main ablation experiments for Attribution Probe were conducted only on the T6 task, which is not yet sufficient to demonstrate the mechanism’s general effectiveness across different RNA tasks.

Questions for authors
  1. Can authors compare the performance differences with and without the Attribution Probe across all 17 tasks?

  2. Do the different methods use the same computational budget and feature encoding resources? Specifically, the feature encoding overhead that was not accounted for in the runtime comparisons.

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

This work proposes an agentic system that automates RNA prediction pipelines by jointly optimizing data preparation, feature representations, model architecture, and hyperparameters. It further proposes to use an attribution probe to propose improvements based on feedback from previous attempts. Experiments on 17 downstream tasks show its advantage on accuracy-efficiency trade-offs over AutoML systems and previous AI research agents.

Critical strengths
  1. The motivation is clear, and the method seems reasonable: the iterative "evaluate-decide-revert" process can be a useful searching strategy for the joint configuration, and the definition of the attribution probe with the bottleneck-to-action mapping covers common failure modes in the modeling workflow. This is the major strength I identified with the paper.
  2. The experiments are comprehensive: 17 downstream tasks, comparison with AutoML, AI research agent methods, and RNA-specific predictors, showing significant improvements in many settings.
Critical weaknesses
  1. It is not clear to me how the attribution probe identifies the bottleneck. Section 2.3 gives the definition of the probe and the bottleneck-to-action mapping, but the exact method to identify the bottleneck was not specified. Further, improving from feedback is a common practice in agentic systems, and the results (e.g. the comparision in Fig. 3 between history guided Thinking with Direct mode) does not establish how the attribution mechanism is particularly effective or novel compared to ordinary "improving from feedback". This affects my assessment on the paper's novelty, so I lean toward rejection.
  2. The experiments need more careful control of different agents. In particular, RNAOmni can search over different representations like RNA-FM and Evo2 embeddings, while it is not clear what features are assessible to the baseline methods.
Questions for authors
  1. Clarification: Can you provide more details on the probe, like its input and prompts? How does it differ from ordinary improving from feedback?
  2. Clarification: What features/embeddings did the AutoML methods and the other agents use in your experiments?
Reviewer_qxyCOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses RNA modeling with an agentic approach. The agent builds models from raw data by searching over data sampling, features, model heads, and hyperparameters. It combines traditional features with foundation model features (RNA-FM, RiNALMo, Evo, Evo 2). The main stated methodological contribution is an attribution probe, which identifies why trials succeed or fail and uses that to guide the next step. The paper evaluates the method on 17 RNA tasks against AutoML methods, LLM research agents, and published RNA foundation model results. It also includes ablation, and shares detailed logs, including failed runs.

Critical strengths
  1. Practical and relevant problem. The paper addresses a useful problem in automating RNA modeling. The feature library combining classical and foundation model representations is a practical component that could benefit domain scientists.

  2. Broad evaluation and transparency. The paper evaluates 17 tasks and provides detailed trial logs, including failed runs. This transparency is valuable for understanding the system's behavior and supporting reproducibility.

Critical weaknesses
  1. Novelty remains unclear. The proposed search loop is similar to existing agents such as autoresearch (https://github.com/karpathy/autoresearch), AIDE (https://arxiv.org/abs/2502.13138), and MLE-STAR (https://arxiv.org/abs/2506.15692), which are not discussed. The attribution probe appears to be the main new component, but its mechanism is not clearly explained or isolated in the ablation. Overall, the methodological contribution beyond existing approaches remains unclear.

  2. Evaluation limitations. The paper does not state whether baselines had access to the foundation model embeddings that RNAOmni uses, making the comparisons difficult to interpret. Some reported baseline results also come from different datasets or splits. The improvement over the strongest baseline is smaller than the headline claim. Additionally, large gaps between validation and test performance raise concerns about possible overfitting. It remains unclear how much of the improvement comes from the agent itself.

Questions for authors
  1. How does your search loop differ from autoresearch, AIDE, and MLE-STAR?
  2. How does the attribution probe decide the cause of a trial? Can you share the prompts or rules?
  3. Did the baselines get the same foundation model embeddings as RNAOmni?
  4. Why do your UTR data sizes and splits differ from the cited papers?
  5. Why are validation scores much higher than test scores?

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

Related papers

Discussion (0)

Sign in to comment.