acceptodds
Under review as a conference paper at ICLR 2027

Compositional Representation Learning for RNA Modifications via Chemical Transformation Equivariance

Abstract

Predicting nanopore responses for RNA modifications with limited or no response labels requires transferring knowledge across chemical types and sequence contexts. Recurring chemical transformations provide a basis for this sharing: they act across nucleoside substrates and compose along valid chemical paths. We formulate **Chemical Transformation Equivariance (CTE)**, under which valid chemical transformations are approximately preserved by corresponding actions on molecular representations. **RiboCTE** learns these actions in a shared latent subspace from precursor–product relations and valid paths, using frozen molecular embeddings as prediction targets. The same actions transform sequence-conditioned states, and a shared decoder predicts changes in nanopore substitution-plus-deletion rates relative to canonical controls. We introduce **RiboModBench** to evaluate cross-modification response prediction alongside chemical-action transfer and composition. RiboCTE achieves the highest mean response-ranking performance among the compared methods in modification/context holdouts and external evaluation. It reaches Spearman correlations of when both modifications and sequence contexts are held out and on external modifications without source response labels. Shared chemical actions thus provide an effective inductive bias for RNA modification representation learning. Code and RiboModBench are available at https://anonymous.4open.science/r/cte-7B11.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
RNA modification, chemical transformation equivariance, compositional generalization, action-conditioned prediction, molecular representation learning

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_ebisOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RiboCTE, a framework for learning composable chemical transformation operators to predict nanopore responses for RNA modifications with limited supervision. The method learns shared transformation operators from molecular precursor–product relations and chemical paths, and transfers them to sequence-conditioned representations for downstream response prediction. The authors also introduce RiboModBench, which evaluates cross-modification generalization, chemical-action composition, few-shot adaptation, and external transfer. Experiments on ONT and Curlcake datasets show that RiboCTE achieves higher Spearman correlations than the compared baselines, particularly under modification and sequence-context distribution shifts.

Critical strengths

Modeling RNA modifications through shared chemical transformation operators is well motivated. Exploiting recurring chemical transformations across different nucleoside substrates provides a meaningful inductive bias for transferring knowledge to modifications with limited response labels.

Critical weaknesses
  • All eight RNA modifications in the main response prediction benchmark involve only single-step chemical transformations (Appendix B, Table 6). Multi-step composition is evaluated separately through molecular embedding retrieval in Track B, rather than downstream nanopore response prediction. Furthermore, Table 2 shows that the action-conditioned MLP achieves an MRR of 0.962 on edge-unseen composition, compared with 0.981 for RiboCTE, while both methods achieve an MRR of 1.000 on edge-seen composition. It remains unclear whether the learned compositional operators provide benefits for downstream response prediction beyond those obtained from modeling individual chemical transformations.

  • In Appendix G (Table 7), permuting the learned action offsets while preserving their original norms still achieves Spearman correlations of 0.204 under LOMO and 0.149 under Double-OOD, compared with 0.216 and 0.184 for the original RiboCTE model. The relatively small performance difference under LOMO raises questions about how much of the improvement comes from the specific learned chemical action assignments rather than the shared architecture or other chemical information. Moreover, the experiment uses only one permutation without reporting variability across different random assignments, making it difficult to assess the robustness of this result.

Questions for authors
  • Table 2 evaluates multi-step chemical composition through molecular embedding retrieval, while the main downstream response prediction benchmark involves only single-step transformations (Appendix B, Table 6). Could the authors clarify whether the learned compositional operators provide additional benefits for downstream nanopore response prediction? Is there any evidence that the method generalizes to response prediction involving multi-step chemical modifications?

  • Table 3 shows that removing path supervision reduces the Spearman correlation from 0.216 to 0.063 under LOMO and from 0.184 to 0.055 under Double-OOD. Could the authors explain why path supervision has such a large impact on single-step response prediction? How can this improvement be distinguished from the general benefits of additional chemical supervision rather than compositional learning itself?

  • In Appendix G (Table 7), permuting the learned action offsets still achieves Spearman correlations of 0.204 under LOMO and 0.149 under Double-OOD, compared with 0.216 and 0.184 for the original model. Could the authors clarify how much of the improvement comes from the specific chemical action assignments? Are the reported differences consistent across multiple random permutations?

3 comments

Dear Reviewer ebis,

Thank you for focusing on the connection between composition, chemical supervision, and downstream transfer!

W1 & Q1: Multi-step composition and downstream response prediction

The connection between composition and response prediction has two components: learning primitive actions jointly from chemical paths, and reusing composed actions for a multi-step response target. The ONT path-control experiment establishes the first component. The external m1Ψ comparison addresses the second, with Track B supplying the complementary evaluation of chemical endpoint recovery under relation and composition holdouts.

To directly assess downstream transfer to a multi-step target, we evaluate the source-unseen Curlcake modification m1Ψ, whose shortest-supported canonical-parent path in our chemical graph is U→Ψ→m1ΨU \rightarrow \Psi \rightarrow m^1\Psi. Both constituent actions receive chemical supervision; among the two path products, only the intermediate Ψ has ONT response supervision. RiboCTE composes the learned primitive actions from U to predict the final modification's response. Response fitting and model selection use ONT data only, without m1Ψ response labels.

Mol-JEPA provides a molecular endpoint baseline, and the exact-endpoint control retains RiboCTE's molecular subspace and response architecture while supplying the product representation directly.

Predictor on Curlcake m1ΨSpearman ρ\rho
Mol-JEPA0.154
Exact endpoints0.198
RiboCTE0.236

RiboCTE improves Spearman correlation by 0.082 over Mol-JEPA and 0.038 over exact endpoints. This result extends the downstream evidence beyond the single-step ONT cohort to the two-step m1Ψ target, connecting the reuse of composed chemical actions to response prediction. The exact-endpoint comparison specifically addresses the value of the shared-action representation when product structures are available.

The corresponding Track B analysis measures endpoint recovery for held-out chemical relations and compositions, as detailed in our response to LZjN. Together, the two analyses distinguish chemical composition fidelity from response transfer using composed actions.

W1 & Q2: Path supervision versus additional chemical supervision

For the additive operators, define the residual of an edge as

The residual of the corresponding path endpoint is exactly ∑trt\sum_t r_t. Thus, the full path term is

\mathcal L_{\mathrm{path}} = \mathbb E_\pi \left\\|\sum_{t\in\pi}r_t\right\\|_2^2.

Expanding the squared norm gives

\left\\|\sum_t r_t\right\\|_2^2 =\sum_t\\|r_t\\|_2^2+2\sum_{s<t}\langle r_s,r_t\rangle.

The cross-edge terms couple the estimation of primitive actions, so the learned offsets can change even when downstream queries use only one action.

To distinguish these joint constraints from repeated exposure to the same relations, the matched control uses

Lreplay=Ledge+λpathEπ∑t∈π∣rt∣22,\mathcal L_{\mathrm{replay}} = \mathcal L_{\mathrm{edge}} + \lambda_{\mathrm{path}} \mathbb E_\pi \sum_{t\in\pi}\\|r_t\\|_2^2,

where Ledge\mathcal L_{\mathrm{edge}} is the single-step prediction loss. Full path training and replay use the same sampled paths, constituent relations, model, selection protocol, and optimization budget. Replay retains the individual squared-residual terms while removing their cross-edge coupling.

Chemical-training objectiveLOMO ρ\rhoDouble-OOD ρ\rho
Edge-only, original no-path ablation0.063 ± 0.0180.055 ± 0.009
Exposure-matched edge replay0.100 ± 0.0620.097 ± 0.064
Full path supervision0.216 ± 0.0330.184 ± 0.036

The full objective improves over replay by 0.116 under LOMO and 0.087 under Double-OOD. These results support the value of jointly constraining cumulative endpoint residuals along valid paths beyond repeating the constituent-edge supervision. The comparison links the cross-edge constraints in compositional chemical training to improved transfer of primitive actions on the single-step response task.

W2 & Q3: Action assignments and repeated permutations

The norm-preserving permutation retains the learned subspace, canonical-parent coordinates, cross-substrate action sharing, and each receiving action's original magnitude. The response decoder is then refitted. A consistent action code can still be used across substrates, so useful performance after permutation is compatible with the sharing mechanism. The experiment measures the additional contribution of the native action-direction correspondence.

We repeat the original control over 10 independent random assignments chosen without using response labels, each using the same three response-training seeds and folds. Each assignment permutes offsets across all 41 chemical actions. For each assignment, scores are first averaged over training seeds; the permuted entries below report the mean ± SD across assignments.

Response-exposure groupLOMO: nativeLOMO: permutedDouble-OOD: nativeDouble-OOD: permuted
Four 2′-O-methylated targets0.2800.201 ± 0.1060.2400.199 ± 0.085
I, Ψ\Psi, m5C, and m6A0.1530.000 ± 0.1400.128-0.007 ± 0.144
All eight targets0.2160.101 ± 0.1070.1840.096 ± 0.094

Native offsets outperform the permuted control in 90% of assignments under LOMO and 70% under Double-OOD.

The native assignment improves over the permutation mean by 0.115 under LOMO and 0.088 under Double-OOD, and exceeds most sampled assignments. Chemically aligned directions therefore provide additional predictive value beyond the sharing, subspace, and magnitude information retained by the control. The original single-assignment results, 0.204 and 0.149, remain part of the reported evidence; the repeated comparison places that draw alongside the wider assignment variability.

The group rows further distinguish transfer with and without same-action source response support. Together with the response-trained-vector comparison and exposure-matched replay, this analysis provides complementary evidence for the contributions of action sharing, native chemical alignment, and joint path constraints to response transfer.

The direct downstream multi-step evidence is the external m1Ψ case; Track B separately evaluates chemical-action transfer and composition.

Dear Reviewer ebis,

Thank you again for your time and thoughtful review. As the discussion closes tomorrow, November 18 (AoE), we would appreciate hearing whether our rebuttal addresses your concerns. We would be happy to clarify any remaining points before the deadline.

Best, The authors

Reviewer_cCaROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper predicts how chemical RNA modifications change nanopore basecalling errors in different sequence contexts, including for modifications with no measured responses. The idea is that the same chemical edit, such as 2′-O-methylation, appears on several nucleosides and can be chained along reaction paths. RiboCTE learns each edit as a shared offset in a molecular embedding space using known reaction data. It then reuses these offsets to predict the response for a new modification in a given context. The authors also release a benchmark, RiboModBench, covering held-out modifications, held-out contexts, few-shot adaptation and an external dataset. RiboCTE gives the best Spearman correlation across these settings, including 0.184 when both modifications and contexts are held out and 0.268 on external modifications not seen in training.

Critical strengths
  • The evaluation is careful. Double-OOD removes test contexts from all training modifications, and model selection never touches test data. The authors also include controls that could have hurt them, such as using the true product embedding instead of the predicted one. With so few labelled modifications, this protocol is what makes the results credible.

  • The external result on unseen modifications (0.268 against 0.129 for the best baseline) is the most interesting finding. It holds under a change in both modification type and dataset.

Critical weaknesses
  • It is unclear what drives the gains. Four of the eight test modifications (Am, Cm, Gm, Um) are the same edit on different bases, and the model uses one shared offset for it. So when Am is held out, the model has already seen its edit through Cm, Gm and Um. The other four modifications use edits never seen in training. The macro average mixes these two cases, and Figure 3 shows only the easy one. Sharing would also explain why RiboCTE beats the true product embeddings (0.216 against 0.162): the true embeddings differ, while the shared offset gives one common signal. I cannot tell whether the gain comes from learned chemistry or just from tying together modifications with the same edit.
  • The evidence is thin. There are only eight test modifications, the margins are a few hundredths, and Tables 1, 3 and 4 have no error bars or per-modification results. Most methods have nMAE above 1, but no trivial baseline is reported. Removing path supervision drops performance below the context-only baseline (0.063 against 0.132), which is odd when every test modification is a single edit. Finally, the chemistry track in Table 2 is saturated, with R@5 of 1.000 for nearly every method.
Questions for authors
  • Can you report results per modification, separating the four 2′-O-methylated ones from the rest? Does RiboCTE still win on I, Ψ, m5C and m6A?
  • How does a simpler model compare? Such as one learned vector per edit type, with the same context encoder and decoder?
  • Most methods in Table 1 have nMAE above 1. How does this compare with a trivial baseline that ignores the modification?
  • Why does removing path supervision fall below context-only?
  • Can you add seed variation or confidence intervals to Tables 1, 3 and 4?
3 comments

Dear Reviewer cCaR,

Thank you for recognizing the evaluation protocol and external-transfer results!

W1 & Q1: Per-modification results and same-action response exposure

We separate Am, Cm, Gm, and Um from I, pseudouridine (Ψ\Psi), m5C, and m6A. In the first group, the held-out action has response supervision on three other substrates. In the second, no other source modification supplies response labels for the same action. All five action types represented by the eight ONT targets are available during chemical pretraining, so the distinction concerns source response supervision.

The table reports mean Spearman correlation ± SD across the three training seeds. HAE-E and Morgan are the strongest aggregate LOMO and Double-OOD baselines in Table 1, respectively.

Modification / groupLOMO: RiboCTELOMO: HAE-EDouble-OOD: RiboCTEDouble-OOD: Morgan
Am0.441 ± 0.0140.254 ± 0.0540.311 ± 0.0870.331 ± 0.031
Cm0.083 ± 0.0200.079 ± 0.0050.105 ± 0.016-0.138 ± 0.019
Gm0.233 ± 0.0080.082 ± 0.0410.211 ± 0.0070.125 ± 0.025
Um0.364 ± 0.0180.351 ± 0.0300.334 ± 0.0230.294 ± 0.025
I0.235 ± 0.0640.207 ± 0.1100.166 ± 0.1150.276 ± 0.010
Ψ\Psi0.335 ± 0.1350.414 ± 0.0300.367 ± 0.0430.262 ± 0.052
m5C0.272 ± 0.0600.103 ± 0.0260.213 ± 0.095-0.146 ± 0.057
m6A-0.231 ± 0.024-0.117 ± 0.018-0.235 ± 0.028-0.225 ± 0.023
Four 2′-O-methylated targets0.280 ± 0.0070.192 ± 0.0190.240 ± 0.0320.153 ± 0.012
I, Ψ\Psi, m5C, and m6A0.153 ± 0.0600.152 ± 0.0130.128 ± 0.0520.042 ± 0.024
All eight targets0.216 ± 0.0330.172 ± 0.0160.184 ± 0.0360.098 ± 0.013

The clearest gain without same-action source response supervision appears under Double-OOD: RiboCTE reaches 0.128 versus 0.042, with positive differences for Ψ\Psi and m5C. Under LOMO, RiboCTE has higher means for I and m5C than the displayed comparator. The 2′-O-methylated group benefits in both settings. Together, these results identify the conditions in which sharing and chemistry-informed transfer improve prediction.

W1 & Q2: A simple learned vector per edit

We compare against one response-trained vector per typed edit, using the same frozen encoders and molecular subspace, and the same context-adapter and shared difference-decoder architectures. The edit vectors are independently initialized, and both the vectors and response modules are fitted only from source response labels under the matched response-training protocol. The edit vectors receive no chemical edge/path supervision.

We evaluate this control on Am, Cm, Gm, and Um, where the held-out action has response supervision on other substrates and its vector can be learned directly.

Predictor on the four 2′-O-methylated targetsLOMO ρ\rhoDouble-OOD ρ\rho
Response-trained shared edit vector0.252 ± 0.0100.203 ± 0.039
RiboCTE0.280 ± 0.0070.240 ± 0.032

The observed LOMO advantage is 0.028 with action tying, the molecular frame, and the response architecture held fixed. This supports the additional value of chemically trained offsets within the shared molecular frame. The Double-OOD gain is 0.037, extending this benefit to unseen contexts. The four targets without same-action response labels are evaluated through the separate exposure-group analysis.

The exact-endpoint comparison addresses a related question: whether shared transformations are a useful response representation when product structures are already available. Sharing can pool information across substrates and suppress substrate-specific variation in endpoint embeddings.

W2 & Q3: Trivial baselines and the meaning of nMAE

Our metric is

nMAEm=MAEmIQR(Ym,Qm)+ϵ\mathrm{nMAE}_{m} = \frac{\mathrm{MAE}_{m}}{\mathrm{IQR}(Y_{m,Q_m}) + \epsilon}

An nMAE of one means that MAE is approximately the query profile's interquartile range. We use zero-effect prediction, Δy^=0\widehat{\Delta y}=0, and a source-only constant fitted to minimize modification-macro MAE as explicit references. Both use the same query cohorts and scoring rule; the constant is fitted using source responses alone.

PredictorLOMO nMAEDouble-OOD nMAE
Zero-effect1.1801.231
Source-only constant1.1151.160
RiboCTE0.9831.029

RiboCTE reduces macro nMAE by 11.8% under LOMO and 11.3% under Double-OOD relative to the stronger constant predictor. This comparison supports numerical response accuracy alongside contextual ranking. Under the benchmark's scoring convention, both constant predictors have Spearman correlation zero; the learned context-only model provides the complementary sequence-based reference.

W2 & Q4: Why path supervision affects single-step response prediction

Path supervision jointly constrains the primitive offsets and shared subspace that are subsequently reused for single-step targets. To separate this effect from additional exposure to chemical relations, we compare full path supervision with an exposure-matched edge-replay control. It samples the same training paths but replaces the squared cumulative path residual with the sum of the constituent squared edge residuals, retaining the same action exposure and optimization budget.

Predictor / chemical-training objectiveLOMO ρ\rhoDouble-OOD ρ\rho
Context-only0.132 ± 0.0290.085 ± 0.012
Edge-only, original no-path ablation0.063 ± 0.0180.055 ± 0.009
Exposure-matched edge replay0.100 ± 0.0620.097 ± 0.064
Full path supervision0.216 ± 0.0330.184 ± 0.036

Full path supervision improves over exposure-matched replay by 0.116 under LOMO and 0.087 under Double-OOD. Matching constituent-edge exposure therefore retains a substantial benefit from joint path constraints. These constraints shape the primitive offsets and shared subspace during chemical training, enabling the same actions to transfer more effectively to single-step response targets.

The no-path model and context-only model have different prediction interfaces. Context-only directly predicts from the RNA representation; the no-path model must express the response through a frozen chemical offset and shared difference decoder, without a free context-only branch.

W2 & Q5: Seed variation

The original results average three independent training seeds. The per-modification and path-control tables above report the available mean ± SD values. Each aggregate applies the manuscript's fold and modification averaging within a seed before computing SD across seeds. For the primary comparison, RiboCTE reaches 0.216 ± 0.033 versus 0.172 ± 0.016 for HAE-E under LOMO, and 0.184 ± 0.036 versus 0.098 ± 0.013 for Morgan under Double-OOD.

A leave-one-target-out calculation from the displayed rounded means retains a positive overall advantage after removing any single target: approximately 0.025–0.068 under LOMO against HAE-E and 0.048–0.115 under Double-OOD against Morgan. Thus, the overall improvement is preserved across these changes in target composition. Seed variation and target-wise sensitivity address complementary aspects of repeatability and the distribution of the gain across modifications.

W2: Saturation in the chemical track

The top-1 comparison in our response to LZjN uses the same edge-unseen queries and 106-state candidate bank. RiboCTE achieves 97.6% versus 94.4% for the action-conditioned MLP, showing more frequent recovery of the correct endpoint in first position.

The aggregate benefit without same-action source response supervision is clearest under Double-OOD; the corresponding LOMO group means are comparable at 0.153 and 0.152.

Dear Reviewer cCaR,

Thank you again for your time and thoughtful review. As the discussion closes tomorrow, November 18 (AoE), we would appreciate hearing whether our rebuttal addresses your concerns. We would be happy to clarify any remaining points before the deadline.

Best, The authors

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

Three contributions have been made in this paper:

  1. a mechanism: authors introduce Chemical Transformation Equivariance, ensuring consistency between valid chemical transformations and learned actions on molecular representations.
  2. a proposed framework/pipeline: authors propose RiboCTE, leveraging a shared chemcial core to map how chemical transformations affect molecular states by predicting product representations from precursor nucleosides and valid reaction paths.
  3. a benchmark for evaluation: to evaluate RiboCTE, authors introduce RiboModBench with two tracks (1) cross-modification response generalization--prediction evaluation and comaprisons using correlation and distance scores against other predictors, and (2) action tansfer and composition diagnositics--action path reconstruction compared with an action-conditioned MLP.

Authors evaluate their proposal under both leave-one-modification-out (lomo) and double ood (both modifications and sequence context) settings.

Critical strengths
  1. authors proposed to study some invariance in representation learning, Chemical Transformation Equivariance (CTE), which is good yet practical in self-supervised learning. Finding a reasonable self-supervision signal from CTE and formulate it into a loss function and training pipeline is even greater.
  2. Evaluation based on LOMO and Double OOD settings demonstrate the strong performance of the proposed RiboCTE, especially when comparing with the existing predictors.The experiment results are solid and significant, especially considering from the perspectives of correlation.
Critical weaknesses
  1. I am not very convinced by the results of track B: action tansfer and composition diagnositics, especially comparing with action-conditioned MLP. Even the proposed method is stronger than MLP under edge-unseen settings, the advantage looks trivial. I guess MRR might not be the best metric for this evaluation. MLP can only return a simple combination of actions, while the proposed methods take care of composition or paths. Again, I am fully aware of the performance gain of the proposed method even in track B. Authors might need to re-interpret the metric: in information retrieval tasks, MRR from 0.96 to 1 is even congratulatory, it is not clear how it works with a 106-state candidate bank.

  2. Evaluating a model on completely "in-the-wild" novel discoveries -- such as a brand-new RNA modification just discovered in a wet lab that has no prior database records -- would be the ultimate test of true generalization. LOMO and double OOD are a necessary comprise for a data-driven evaluation. I'm doubting about the potential leakage risk here, as this approach leverages pre-trained models such as CheMeleon and RINALMo. I am not sure if these two models' pre-training has ever covered some "held-out" data entry.

Questions for authors
  1. if the problem of action-conditioned MLP is returning unconstrained combination, can we just evaluate the rejection rate of its output subject to certain constraints about composition/paths.

  2. Please make a deep-dive on the pre-training datasets/recpies of CheMeleon and RINAMo. Make sure there was no data leakage risk for held-out validation in this study, even the labels or prediction targets were not seen during their pre-training.

2 comments

Dear Reviewer LZjN,

Thank you for recognizing the motivation of Chemical Transformation Equivariance and the response-generalization results!

W1 & Q1: Track B, the action-conditioned MLP, and metric interpretation

The action-conditioned MLP also performs sequential composition. As specified in Appendix F.1, it applies

recursively along the supplied action path, using the same descriptors, permitted edge/path targets, and optimization budget as RiboCTE. The comparison therefore tests the inductive bias of shared additive actions against a more flexible state-dependent predictor.

Track B measures recovery of the known chemical endpoint in a common frozen molecular space. Top-1 retrieval makes the near-ceiling comparison more interpretable by measuring how often the correct product is ranked first. The comparison uses the same queries, 106-state candidate bank, and group weighting.

PredictorEdge-unseen MRREdge-unseen top-1
Action-conditioned MLP0.96294.4%
RiboCTE0.98197.6%

RiboCTE improves macro top-1 retrieval by 3.2 percentage points, giving a direct interpretation of the MRR advantage. Edge-unseen composition is the informative transfer setting because the constituent relations are withheld and each primitive must transfer from another substrate. Edge-seen composition establishes successful rollout with observed constituent relations.

Both models receive valid action paths and predict embeddings. For the supplied valid paths, the macro top-1 endpoint mismatch rate is 5.6% for the action-conditioned MLP and 2.4% for RiboCTE. Recovering the endpoint specified by the supplied edits directly evaluates chemical consistency. A molecular-validity rejection rate after projection onto the valid candidate bank would instead measure a property already enforced by the bank.

W2 & Q2: Foundation-model pretraining and held-out information

The evaluated use case is response prediction for specified modification chemistry without target response measurements. LOMO excludes all response labels for the target modification; Double-OOD additionally excludes the query flank groups from every source modification. Response fitting, calibration, and model selection use the permitted source data.

Section 3 and Appendix B explicitly permit target structures and outcome-free chemical relations during Track A chemical training. Track B separately withholds chemical relations and compositions. This distinction separates access to a modification's chemistry from access to its measured nanopore response.

We use two frozen checkpoints: CheMeleon chemeleon_mp.pt (release), pretrained to predict deterministic Mordred descriptors from approximately one million PubChem molecules (method), and RiNALMo-Giga rinalmo_giga_pretrained.pt (release), pretrained by masked-language modeling on 36 million non-coding RNA sequences from RNAcentral, nt, Rfam, and Ensembl (method). Neither reported pretraining objective uses nanopore response measurements.

Our comparison with the released CheMeleon corpus identifies exact structure matches for 34/106 benchmark states, including six ONT modifications; these are corpus-level matches, with training-split membership unverified. For RiNALMo, exact historical corpus membership remains unresolved, and canonical 5-mers may occur within longer pretraining sequences. Our evaluation permits chemical structures and sequence priors while withholding target nanopore responses from response fitting, calibration, and model selection.

Dear Reviewer LZjN,

Thank you again for your time and thoughtful review. As the discussion closes tomorrow, November 18 (AoE), we would appreciate hearing whether our rebuttal addresses your concerns. We would be happy to clarify any remaining points before the deadline.

Best, The authors

Comments on the paper

2 comments

Dear Area Chair and Reviewers,

We sincerely thank all three reviewers for their thoughtful feedback and for recognizing the motivation behind Chemical Transformation Equivariance (CTE), the careful generalization protocols, and the promising external-transfer results. We have provided detailed responses and substantial additional analyses addressing the main concerns. We summarize the most important new evidence below.

1. The improvements extend beyond simple action sharing and reflect learned chemical structure.

A central concern was whether RiboCTE benefits merely from tying the same edit across substrates. We introduced a matched baseline with one response-trained vector per chemical edit, retaining the same molecular subspace, context adapter, and response decoder. On the four 2′-O-methylated targets, RiboCTE improves Spearman correlation from 0.252 to 0.280 (LOMO) and 0.203 to 0.240 (Double-OOD), demonstrating benefits beyond action tying alone.

Importantly, we also separately evaluated the four targets without same-action source response supervision (I, Ψ, m5C, and m6A). Under Double-OOD, RiboCTE achieves 0.128 versus 0.042 for the strongest aggregate baseline, although the corresponding LOMO results are comparable. This distinguishes the contributions of cross-substrate response sharing from chemistry-informed transfer.

Furthermore, we expanded the action-permutation analysis to 10 independent assignments across 41 chemical actions. Native chemical assignments achieve 0.216 versus 0.101 mean permuted correlation under LOMO, and 0.184 versus 0.096 under Double-OOD. Native assignments outperform 90% and 70% of the sampled permutations, respectively. Together, these controls demonstrate that chemically aligned actions provide predictive information beyond the shared architecture and action magnitudes.

2. Compositional learning benefits downstream prediction, beyond additional chemical supervision.

To isolate the contribution of compositional constraints, we introduced an exposure-matched edge-replay control using the same chemical paths, constituent relations, and optimization budget, but without joint path-endpoint constraints. Full path supervision improves correlation from 0.100 to 0.216 (LOMO) and 0.097 to 0.184 (Double-OOD). This directly supports the value of jointly constraining primitive actions through composition, rather than simply exposing the model to more chemical relations.

We also added a direct downstream evaluation on the two-step modification m1Ψ in the external Curlcake dataset. Without any m1Ψ response labels during training or model selection, RiboCTE composes the learned actions along U → Ψ → m1Ψ and achieves 0.236 Spearman, outperforming both Mol-JEPA (0.154) and the exact-endpoint control (0.198). This provides an initial downstream demonstration of multi-step compositional transfer, complementing the chemical-retrieval evaluations.

For Track B, we additionally clarified that the action-conditioned MLP also performs sequential composition. Under edge-unseen evaluation, RiboCTE improves top-1 endpoint retrieval from 94.4% to 97.6%, making the advantage over this compositional baseline more interpretable than MRR alone.

3. Additional controls strengthen the empirical evidence and its interpretation.

We provided per-modification results, three-seed variability, target-group analyses, and sensitivity to removing individual target modifications. The primary ONT results remain 0.216 ± 0.033 versus 0.172 ± 0.016 under LOMO and 0.184 ± 0.036 versus 0.098 ± 0.013 under Double-OOD, relative to the strongest aggregate baselines.

We also added zero-effect and source-only constant baselines. RiboCTE reduces normalized MAE by 11.8% (LOMO) and 11.3% (Double-OOD) relative to the stronger constant predictor. Beyond ONT, the original external evaluation achieves 0.268 versus 0.129 on modifications absent from source response training, supporting transfer under a simultaneous change in modification identity and experimental dataset.

We hope these additional analyses make the scope and variability of the improvements more transparent, including the remaining difficult individual modifications.

(Cont.)

4. We clarified the precise generalization setting and pretrained-model exposure.

Our evaluation targets modifications with specified chemistry but without target nanopore response labels. Chemical structures, outcome-free relations, and pretrained sequence representations are explicitly permitted; target response labels are withheld from response fitting, calibration, and model selection.

We investigated the pretrained encoders and disclosed the available corpus-overlap evidence, including 34 exact structure matches among 106 benchmark states in the released CheMeleon corpus, with actual training-split membership unverified. RiNALMo's historical sequence overlap likewise remains unresolved. Neither model's reported pretraining objective uses nanopore response labels. We therefore distinguish response-label generalization from the stronger, untested setting of discovering entirely unrecorded chemistry.

In summary, we believe these additional experiments substantially strengthen the central contribution of the paper: chemically grounded, composable transformation operators provide a useful inductive bias for transferring molecular knowledge to context-dependent RNA response prediction. The new controls separately investigate action sharing, chemical alignment, and compositional supervision, while the external multi-step example connects learned composition to downstream prediction.

We acknowledge the limitations of the currently available modification coverage, but believe the combined methodological contribution, rigorous benchmark, and additional evidence address the principal scientific concerns raised during review.

We respectfully hope that the Area Chair and reviewers will consider these clarifications and new results in their final assessment.

Thank you again for your time and consideration.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.