acceptodds
Under review as a conference paper at ICLR 2027

RiboUnmix: Learning Shared Translational Dynamics from Biased and Noisy Ribo-seq Measurements

Abstract

Ribosome profiling (Ribo-seq) is used to study translation dynamics by measuring how ribosomes are distributed along mRNA sequences, but the resulting occupancy profiles also reflect experiment-specific distortions and stochastic measurement variability. Models trained to predict these profiles from mRNA sequences can also learn these distortions, so accurate profile prediction alone does not establish recovery of the underlying biological behavior. We ask whether combining Ribo-seq datasets affected by different experimental conditions can enable us to reveal shared sequence-dependent patterns in the underlying ribosome distribution. We introduce RiboUnmix, a probabilistic multi-dataset framework that jointly learns from multiple Ribo-seq datasets collected in different experiments. RiboUnmix represents each expected measured profile as a shared sequence dependent signal modulated by a dataset-specific multiplicative factor, while a negative-binomial observation model captures variability across individual Ribo-seq replicates. To evaluate recovery of the shared profile and dataset-specific effects, we construct a controlled synthetic Ribo-seq benchmark combining programmed translation kinetics, ribosome traffic, stochastic count sampling, and multiple sequence-dependent experimental distortions. The programmed kinetics and injected distortions provide known targets for evaluating the inferred shared profile and dataset-specific effects separately. Both inferred components show high correlation with their synthetic targets, showing RiboUnmix's ability to separate shared kinetic patterns from dataset-specific effects under controlled conditions. Across four organism-specific real-data benchmarks, RiboUnmix outperforms the evaluated sequence-to-profile baselines in predicting measured Ribo-seq profiles. The model independently trained on separate subsets of 114 HEK-derived datasets recovers concordant shared profiles for holdout transcripts, remarking reproducibility of the shared signal. Complementary experiments varying the number and composition of training datasets further show that the inferred shared representation remains substantially stable as the experimental evidence base changes. RiboUnmix turns variation across experiments into evidence for reproducible sequence-dependent patterns of ribosome occupancy, providing a foundation for biological hypothesis generation from diverse Ribo-seq datasets.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
RNA, Deep Learning, Probabilistic Modeling, Ribosome Profile

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_VvH7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces a framework to learn ribosome profiles from multiple Ribo-seq datasets, called RiboUnmix. The idea is that Ribo-seq measurements contain both biological information and experimental biases. The authors propose to separate these two components by learning a shared sequence-dependent profile and dataset-specific corrections. The model uses two bidirectional GRUs and a negative-binomial observation model. It is evaluated on synthetic data, 114 HEK-derived datasets, and four prediction benchmarks.

Critical strengths
  1. the paper is well written and makes the biological concepts clear also for not experts.
  2. the research question of the paper is interesting and clear.
  3. Using multiple experiments to separate shared information from dataset-specific biases is a good idea. I also like that the model learns these corrections directly from data, without requiring predefined bias features. This approach could also be useful for other biological applications.
Critical weaknesses
  1. I think that the shared profile does not necessarily represent the true biological ribosome occupancy. If several datasets share the same systematic bias this bias may become part of the shared profile instead of being removed. Therefore, good reproducibility across models does not necessarily mean biological correctness. Some independent biological validation would make the results stronger.

  2. Figure 13 suggests that the learned profile is already close to its reference target with few datasets. Adding more datasets seems to makes the reference target closer to the true occupancy because different biases cancel each other; so, part of the improvement may come from averaging different measurements rather than from the proposed model. A comparison with a simple consensus baseline, such as a model trained on the weighted geometric mean of normalized profiles, is important to understand better this.

  3. The prediction benchmarks show good performance, but they are based on individual datasets and do not directly demonstrate the benefit of the proposed multi-dataset decomposition. Comparisons with simpler approaches, such as pooled training, averaging predictions from separate models, or learning a consensus profile, would help isolate the contribution.

  4. I think that the general idea is related to existing approaches in multi-task learning, batch-effect correction, and shared representation learning. I think the main novelty is the application of this approach to Ribo-seq; for this reason, I am not sure that this paper is a good fit for this venue, since the contribution seems more focused on a specific biological application than on methodological novelty.

  5. The synthetic datasets reuse the same ribosome trajectories, and observations are not fully independent.

Questions for authors

See weaknesses.

2 comments

We thank the reviewer for a detailed and technically precise review, and for appreciating the clarity of the paper, the research question, and the idea of learning dataset-specific corrections directly from data without predefined bias features.

W1: The shared profile is not necessarily biological occupancy. We agree, and we will state this explicitly in the abstract and introduction rather than only in Appendices B.4 and H.7. This is a mathematical limitation of the problem itself rather than of the model. When the only information available is a set of datasets, a bias shared by all of them is indistinguishable from biological signal, and no method can separate the two without external information. Our goal is to reach the maximum recovery possible in this setting, where nothing is known except the availability of multiple heterogeneous datasets, and to characterize exactly what can and cannot be recovered (Appendix B). We fully agree that independent biological validation would substantially increase the value of the results, and we will add a comparison of the shared profile with known sequence-dependent pausing motifs.

W2 and W3: Contribution of the model beyond averaging, and multi-dataset comparisons. The reviewer's reading of Figure 13 is correct: the learned profile tracks its reference target closely, and recovery of the simulated profile improves as complementary biases cancel across datasets. We see this cancellation as the intended mechanism; what the model contributes is performing it from incomplete, noisy observations and generalizing it to transcripts never seen in training. We agree this must be demonstrated directly, and that the single-dataset benchmarks in Section 7 only show that the sequence encoder is a competitive predictor. We will therefore add the following baselines in both the synthetic and HEK multi-dataset experiments:

  • a sequence model trained on the arithmetic mean of normalized profiles across datasets;
  • a sequence model trained on the weighted geometric mean of normalized profiles, as the reviewer suggests;
  • RiboUnmix with the correction term deactivated (γ = 1), which corresponds to pooled training without dataset-specific corrections.

These comparisons isolate the contribution of the dataset-specific corrections from the effect of combining datasets.

W4: Relation to existing approaches and venue fit. We agree that the method shares ideas with multi-task learning, batch-effect correction and shared representation learning, and we will discuss these connections explicitly. However, disentangling a shared term from multiplicative bias terms is not identifiable in general, and RiboUnmix relies on a specific mathematical construction to make it so: fixed-reference two-way centering of the corrections, with formal identification results (Section 4.3, Appendix B). Both components are predicted from sequence and must generalize to unseen inputs. The formulation is not specific to Ribo-seq, and applies to other settings where multiple biased measurements of a shared sequence-dependent signal are available, as the reviewer also notes in the strengths.

W5: Dependence in the synthetic data. We agree (Appendices E.3 and F.1). We will extend the synthetic benchmark with independent trajectories and independent count sampling for each condition, and evaluate recovery on a held-out test set not used for checkpoint selection.

We thank the reviewer again; these comments identify precisely the analyses needed to strengthen the paper's claims.

Thank you for your response. I appreciate that you response and that you plan to add new experiments.

However, I decided to keep my original score for these reasons:

  1. Missing baselines (my Weaknesses 2 and 3). I appreciate that you plan to add comparisons with arithmetic and geometric mean baselines and pooled training. However, these results are not yet provided, and I think they are important to understand the real contribution of the proposed method beyond simple averaging.

  2. Synthetic data (my Weakness 5). You acknowledge the dependence between synthetic observations, but the experiments with independent trajectories and an independent test set are still missing.

  3. Novelty (my Weakness 4). Your explanation helps clarify the contribution but I still have some doubts about the methodological novelty compared to existing shared representation and batch-effect correction approaches.

Several important concerns still need experimental evidence.

Reviewer_dwDTOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This manuscript introduces RiboUnmix, a computation method to learn relationships between sequence and ribosome occupancy. It is focused on the challenge of combining multiple ribo-seq data sets that may have different experimental biases or artifacts. It does this by learning both a shared profiles and dataset-specific corrections. Because it learns dataset-specific biases, RiboUnmix does not require pre-specifying the biases that are specific to individual datasets.

The method is evaluated on several data sets and against several reasonable baselines.

Critical strengths
  1. The method is evaluated on several datasets, including simulated data where there is known ground truth for both the true signal and data-set specific variation.

  2. The experiments show better empirical results on synthetic data and several real data set collections from different organisms.

  3. The manuscript tackles an interesting problem and has interesting methodological novelty to separately model joint signal and data-set-specifc variation. The modeling choices seem reasonable.

Critical weaknesses
  1. Though there is insight behind individual modeling choices, in terms of big picture and conceptual innovations, the methodology is somewhat incremental (in the sense that such unmixing approaches are not that novel, even if this is the first to apply that broad idea to ribo-seq profile estimation).
Questions for authors
  1. Could you provide in the appendix a description of how each of the baseline methods were run? I did not see code to run the baselines in the anonymized git repo (though I may have missed it). For example, Seq2ribo’s performance seems lower than what was reported in that paper.
2 comments

We thank the reviewer for the positive assessment of our evaluation design and modeling choices, and for the question about the baselines, which has helped us improve the benchmark.

Q1: Baseline reproducibility. Comparing sequence-to-profile models fairly is difficult. The published baselines differ in nearly every part of their pipelines: transcript selection, count normalization, masking of terminal codons, handling of zero-count positions, loss function, training regime and checkpoint criterion. No single setting is neutral for all of them. As described in Appendix I.3, we therefore ran every baseline under three settings:

  • Native: each model's own published preprocessing and training procedure.
  • Matched–U: a common training cohort and target preparation, with uniform transcript weights.
  • Matched–W: the same, with transcript reliability weights that combine read depth (70%) and positional coverage (30%), normalized within each dataset (Appendix H.1).

Table 9 reports the training cohorts and selected checkpoints, and Table 10 reports results under all three settings. Table 1 uses Matched–W, so all models are trained on the same transcripts with the same weights. Even so, model-specific losses and checkpoint criteria remain, which limits how far differences can be attributed to architecture alone (Appendix I.4). We will add all baseline scripts, configurations and preprocessing code to the anonymous repository, and a step-by-step description of each run to Appendix I.

Seq2Ribo. We agree that its performance in our benchmark is lower than in the original publication. One possible contributing factor is checkpoint selection: Seq2Ribo selects by minimum validation loss, whereas most other baselines use validation Pearson correlation. Our evaluation protocol also differs (datasets, positive-count positions only, five-codon trimming at each end). To obtain a closer comparison, we will report Seq2Ribo with Pearson-based checkpoint selection and verify that we reproduce the published results with the authors' original configuration, updating Tables 1 and 10.

W1: Conceptual novelty. We respectfully disagree that the approach is incremental. Existing approaches to bias in Ribo-seq rely on prior knowledge:

  • Choros (Mok et al., 2023) removes biases specified in advance through predefined features.
  • Riboformer (Shao et al., 2024) requires a measured reference profile as an anchor at inference time.
  • Riboclette (Nallapareddy et al., 2025) models known biological perturbations rather than technical effects.
  • scVI (Lopez et al., 2018) and similar batch-correction models work at gene level and infer latent factors for each observed sample, not codon-resolution profiles for unseen sequences.

RiboUnmix asks a different question: what can be recovered when nothing is known about the biases and no reference is available, only multiple heterogeneous datasets? The separation relies only on the model's inductive bias. A sequence-derived shared profile must explain all datasets at once, while dataset-specific corrections capture how each departs from it. This gives a reference-free estimate of what is recoverable, and characterizes what is not, such as biases common to all datasets (Appendix B.4). It also yields interpretable differences between datasets, and generalizes to unseen transcripts. To our knowledge, this problem has not been posed or evaluated for Ribo-seq, either with synthetic ground truth or across 114 real datasets. We will make this framing more explicit in the revision.

We would be grateful to know whether additional evidence, such as validating the shared profile against known pausing motifs, would address the reviewer's remaining concerns.

Thank you for the response. I understand your argument regarding novelty, but continue to disagree that this is substantially innovative in ML. My review stands as is.

Reviewer_CvfYOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Ribosome profiling (Ribo-seq) measures ribosome occupancy along RNA sequences, and hence permits (indirect) study of protein translation dynamics. A typical problem in understanding Ribo-seq data, which is shared by other kinds of sequencing data, is that of technical/observation/measurement noise: there is a difference between what is measured and the underlying biological signal. As is well known, there can even be systematic biases, like experiment-specific batch effects. Effective analysis of Ribo-seq data must do something to grapple with technical noise, for example by estimating it so that it can be adjusted for.

The authors introduce RiboUnmix, a specific method for estimating ribosome occupancy that accounts for measurement noise and dataset-specific biases. They present their method, evaluate it on some synthetic data, and then apply it to certain Ribo-seq datasets.

Critical strengths

The authors consider an interesting problem. To my knowledge, characterizing Ribo-seq data in a way that reduces bias is not well-studied.

The authors' central approach seems to make sense and is grounded is simple probabilistic models. They apply their method to reasonably-constructed synthetic data and to a large number of real datasets.

Critical weaknesses

The paper's biggest problem is that it is difficult to understand. Because this problem isn't as well-studied as other ML-relevant biology problems (although I'm not that familiar with it, so I'm not very confident in this claim), I think the authors should spend more space explaining how Ribo-seq works and what is difficult about analyzing such data. It is also critically important to explain better how their work builds on or differs from existing related work. Related work is mentioned in Sec. 2, but without a detailed familiarity of those papers, it was difficult to follow the discussion of the model formulation in Sec. 4. It's not obvious to me why the authors made various choices and this should be explained better. Where possible, when introducing model components, the authors should reference existing models and explain why the choice made is identical/different/better. Schematics would help. Fig. 1 is not that clear and is kind of ugly (and obviously AI-generated).

(I think the authors should engage more with existing work on estimating technical noise, most of which does not appear in ML conference papers. It appears highly relevant.)

Also, I found the mathematical exposition unnecessarily dense and confusing. The idea is fairly simple (include multipliers that account for bias, for example) and should be explained more simply. If it is true, as the authors claim, that they used an LLM to "review mathematical exposition for clarity and consistency", I suggest they adjust their approach. Sec. 3 should be conceptually simple but is quite dense.

The significance of the work is also unclear. Although the major innovation appears to be accounting for dataset-specific biases, which is useful in multi-dataset analyses, most (all?) of the real data analyses in Sec. 7 involved single datasets. Hence, it's not clear that the extra complexity of the dataset-specific model components is really doing anything here.

Also related to significance, Table 1 claims that the authors' method is better than others, and the numbers indeed look better, but it's not obvious to me how to interpret them. What are big numerical differences here? If the model is simplified (e.g., the dataset-specific part is removed), how does performance change?

A more minor issue on line 53: ".Riboformer" is missing a space.

Questions for authors
  1. Can the authors clarify the primary differences between their approach and existing methods?
  2. Does accounting for dataset-specific biases matter for the Sec. 7 analyses?
  3. How does model performance change when it is simplified? Is the difference substantial?
  4. How important are the differences shown in (e.g.) Table 1? What downstream analyses might they substantially negatively affect?
2 comments

We thank the reviewer for their careful and constructive review. We appreciate that the reviewer has understood both the problem we aim to address and our rationale for first testing the model on synthetic data before applying it to real datasets.

We agree that the presentation can be improved to help readers better understand the scope of the problem and the challenges involved in analysing Ribo-seq data. Below, we address each point and describe the changes we will make.

Clarity and background. We agree that the paper assumes too much familiarity with Ribo-seq. In the revision, we will add a short introduction explaining how Ribo-seq profiles are generated and where systematic distortions arise, including nuclease digestion, ligation and fragment-end preferences, translation inhibitors, and sequencing depth. We will also cite studies investigating these artefacts. Furthermore, we will simplify Section 3. Each predicted profile consists of a dataset-shared component multiplied by a dataset-specific distortion, with additional count noise.

We will move the formal results to the appendix and redesign Figure 1 as a clearer schematic.

Q1: Differences from existing methods.

To the best of our knowledge, no existing method addresses exactly the problem we tackle in this paper. However, we will add a comparison table, to the extent possible, covering methods and models that address technical biases in the data. In brief:

  • Choros (Mok et al., 2023) corrects bias through regression on predefined bias features, such as fragment-end sequences. Consequently, it can only correct for biases that are specified in advance. In contrast, RiboUnmix does not prespecify the form of the bias; instead, it learns dataset-specific multiplicative corrections jointly across datasets.

  • Riboformer (Shao et al., 2024) can correct distortions but requires a measured reference profile as input at inference time. RiboUnmix, in contrast, makes predictions using only the coding sequence.

  • Riboclette (Nallapareddy et al., 2025) models condition-specific biological changes rather than separating technical measurement effects.

  • iXnos, RiboMIMO, and related models predict measured profiles from sequence. However, they do not distinguish between shared and dataset-specific structure, meaning that they may learn reproducible experimental artefacts as if they were biological signals. This motivates our approach.

We will also expand the discussion of the non-ML literature on technical-noise estimation, as suggested by the reviewer, including bias-correction methods developed for RNA-seq and Ribo-seq.

Q2: Do the dataset-specific components matter in Section 7?

The reviewer is correct that the benchmarks in Section 7 each use a single dataset, meaning that the decomposition cannot be applied in that setting. This section was intended only to demonstrate that the sequence model is a competitive predictor of measured profiles. The decomposition itself is evaluated in Sections 5 and 6:

  • Section 5 (synthetic data): The true profile and the injected biases are known, allowing both components to be evaluated directly.

  • Section 6 (114 HEK-derived datasets): Models trained on panels that do not share any studies recover similar shared profiles.

We will make this distinction explicit and present Section 7 as a sanity check rather than as evidence supporting the decomposition.

Q3: Ablations.

We agree that an ablation study removing individual components of RiboUnmix is important to demonstrate their respective contributions to performance. We will extend the article to include this study.

Q4: How meaningful are the differences in Table 1?

The improvements over the strongest baseline, RiboMIMO, are modest: 0.01–0.08 in median transcript-level Pearson correlation, with overlapping 95% confidence intervals in three of the four organisms. We do not claim that these differences are decisive for downstream applications.

The main contribution of the paper is not an improvement in reconstruction accuracy. Rather, it is a framework for separating shared structure from dataset-specific distortions, something that reconstruction accuracy alone cannot assess (Section 3). We will revise the abstract and Section 7 to reflect this distinction and will report paired statistical tests across transcripts.

Minor comment. The missing space on line 153 has been fixed.

We thank the reviewer again for these constructive comments, which will help us substantially improve the paper.

I confess that I still don't totally understand the motivation for this, but I appreciate the authors' honest engagement. These answers are helpful. Giving you the benefit of the doubt, I will slightly raise my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.