acceptodds
Under review as a conference paper at ICLR 2027

Zero-Shot RNA Modification Site Prediction with Molecular Side Information

Abstract

RNA modification site prediction can prioritize candidate loci for experimental validation, but site annotations are highly uneven across modification types. A modification can be chemically characterized before a sufficiently large transcriptome-wide site map is available because profiling depends on modification-specific assays and curation. Conventional sequence predictors treat modification identity as a fixed supervised label, so they cannot use this known molecular identity when the corresponding site labels are absent. We address this gap by augmenting MultiRM with standardized annotations of complete modified nucleosides and using them as molecular side information in a shared RNA-modification compatibility framework. The resulting modification embedding conditions RNA-context selection and final scoring, allowing one model to evaluate a held-out modification without a target-specific output head. Across twelve provenance-aware inductive leave-one-modification-out tasks and stricter source-separated and grouped-substructure holdouts, molecular side information improves zero-shot ranking. We further evaluate transfer to ac4C, which is not included among the twelve modification types in MultiRM. Transfer is strongest when training includes modifications with related molecular substructures and becomes less important as direct target labels accumulate. The framework therefore supports candidate-site prioritization for unseen modification types and for modifications with few annotated sites when their molecular identity is known. Resources are available at: https://anonymous.4open.science/status/RNA_modification-1178.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
RNA modification prediction, chemical conditioning, zero-shot learning, RNA representation learning, molecular representation, target-conditioned prediction

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_RgHbOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work addresses zero‑shot RNA‑modification‑site prediction by leveraging molecular side‑information of modified nucleosides. Conventional multi‑task RNA‑modification predictors rely on modification‑specific output heads and cannot score unseen modification types without target‑specific labelled training data. The authors augment the MultiRM benchmark with standardized molecular descriptors (Morgan fingerprints, physicochemical scalars, parent‑base information) for each modified ribonucleoside and design a target‑conditioned compatibility framework. The model uses modification embedding to condition cross‑attention‑based RNA context aggregation as well as final site‑level scoring, avoiding modification‑dedicated output heads, so it can evaluate held‑out modification types purely from their molecular descriptors.

Extensive inductive leave‑one‑modification‑out (LOMO) evaluations are conducted with provenance‑aware leakage prevention. The authors further adopt stricter source‑separated and grouped‑substructure hold‑out protocols, plus external zero‑shot transfer tests on ac4C which is entirely absent from MultiRM. Controlled ablations indicate that molecular substructure fingerprints deliver most transferable gain, while physicochemical scalar features contribute little. Target‑conditioned cross‑attention for RNA context selection brings extra improvements beyond feeding molecular descriptors only into the final scorer. Performance gain correlates positively with chemical similarity between the held‑out modification and training‑set modifications. As target‑specific labelled data increases, the benefit from molecular side‑information diminishes, showing this mechanism is most valuable under zero‑shot or low‑label regimes. Gains are preserved when swapping the BiLSTM encoder for the stronger pretrained AIDO.RNA backbone.

Critical strengths
  1. This paper puts forward a practical zero‑shot setting for RNA‑modification‑site prediction: predicting sites for unseen modification types given only their molecular chemical structure, without any site labels of that modification for training. It fills a gap compared with existing multi‑type RNA modification predictors which require target‑specific supervision.

  2. Provenance‑aware inductive LOMO is carefully designed to eliminate data leakage; the authors remove samples originating exclusively from the held‑out modification, instead of only masking labels. Multiple hold‑out variants (source‑separated, grouped‑substructure removal) and external real‑world test on ac4C strengthen empirical evidence. Multiple control baselines (parent‑only descriptors, permuted fingerprints, nearest‑target transfer, independent pooling) cleanly isolate contributions of each component.

  3. Systematic ablations decompose contributions from molecular substructure fingerprints, physicochemical features, target‑conditioned context aggregation. The paper characterizes how performance gain changes with chemical similarity and with increasing amounts of target‑specific labels. Replacing the RNA encoder backbone verifies that the zero‑shot gain is not an artifact of weak sequence encoders.

Critical weaknesses
  1. The macro AUC gain brought by full molecular descriptors over parent‑only baseline is around +0.019 under LOMO setting. Although statistically significant via bootstrap resampling, absolute improvement is limited. Two held‑out modifications (m5C, inosine) even show negative gains. The authors should further discuss practical utility given this modest uplift.

  2. When structurally‑related modification groups are fully removed (strict substructure hold‑out), performance gain shrinks to merely +0.003. This suggests the method cannot deliver reliable predictions for completely novel modification families with no related sub‑structures present in training corpus. The practical boundary of generalization should be highlighted more prominently in the main text.

  3. It is natural to ask: can off‑the‑box RNA LLMs perform comparable zero‑shot modification‑site prediction via prompting with modification chemical information? The absence of such baseline weakens the demonstration of the proposed specialized framework’s necessity.

  4. The work uses fixed RDKit‑computed Morgan fingerprints as molecular representation. It remains untested whether alternative molecular featurization choices change zero‑shot transfer performance.

Questions for authors
  1. The macro AUC gain brought by full molecular descriptors over parent‑only baseline is around +0.019 under LOMO setting. Although statistically significant via bootstrap resampling, absolute improvement is limited. Two held‑out modifications (m5C, inosine) even show negative gains. The authors should further discuss practical utility given this modest uplift.

  2. When structurally‑related modification groups are fully removed (strict substructure hold‑out), performance gain shrinks to merely +0.003. This suggests the method cannot deliver reliable predictions for completely novel modification families with no related sub‑structures present in training corpus. The practical boundary of generalization should be highlighted more prominently in the main text.

  3. It is natural to ask: can off‑the‑box RNA LLMs perform comparable zero‑shot modification‑site prediction via prompting with modification chemical information? The absence of such baseline weakens the demonstration of the proposed specialized framework’s necessity.

  4. The work uses fixed RDKit‑computed Morgan fingerprints as molecular representation. It remains untested whether alternative molecular featurization choices change zero‑shot transfer performance.

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

The paper addresses a structural limitation of RNA modification site predictors. Multi-task models such as MultiRM identify each modification through a dedicated output coordinate, so a modification with no site labels has no output to score it, even when its chemistry is fully characterised. The authors reformulate the task as a conditional one. Each modification is represented by a deterministic descriptor of the complete modified ribonucleoside (a 2,048-bit Morgan fingerprint, five physicochemical scalars and a parent-base one-hot), encoded by a shared MLP into an embedding that both queries a multi-head cross-attention over the RNA window and enters a bilinear compatibility score against the centre and context states. All parameters are shared across targets, so a held-out modification is scored from its descriptor alone. The authors augment the MultiRM benchmark with standardised molecular annotations and reconstructed example provenance, which enables a provenance-aware inductive leave-one-modification-out protocol in which windows admitted only through the held-out target are removed rather than label-masked. Evaluation covers twelve LOMO folds with matched controls (target-agnostic, parent-only, permuted descriptors, nearest-target transfer, query-independent pooling), study/transcript-separated and grouped-substructure holdouts, external transfer to ac4C from two independent human datasets, a descriptor ablation, a supervision-level sweep, and an encoder substitution. The headline is +0.019 macro AUC_b over parent-only conditioning, with a bootstrap CI of [+0.010, +0.029] over twelve targets.

Critical strengths

• The motivation is a good one. There is usually a long gap between knowing a modification’s chemical structure and having any idea where it sits in the transcriptome, often decades, because each one needs its own assay built around its own chemistry. The structures are free and sitting in public databases; the maps are the expensive part. Using what you have to say something about what you do not is the right instinct, and if it worked well it would let people design experiments on modifications nobody has mapped yet instead of waiting for an assay. • The controls are thorough and well chosen. The authors ran the ones that could have sunk their own result. They fed the model scrambled descriptors to check that the gain is not just extra input capacity, and they removed from training exactly the chemistry the method leans on hardest, reporting what came out either way. They also report what a looser evaluation protocol would have given them alongside the stricter number they actually use. I am willing to take the reported effect at face value because of this, which is not always the case with results this size. • The statistics are set up sensibly. Each held-out modification counts as one observation rather than each training run. That is the right call for a claim about transferring to new modifications, and it means the confidence interval measures the thing the paper is about.

Critical weaknesses

• The effect is small, and it leans heavily on a handful of closely related modifications. Four of the twelve held-out targets are chemically near-identical to each other, and between them they account for more than half the total gain. The method is essentially interpolating: a new modification does well when something chemically similar was in training, and much less well otherwise. The authors test this themselves by pulling those four out of training, and the advantage largely disappears. That is a narrower claim than the abstract makes, namely that chemical descriptors transfer between close relatives and not much further. It matters because the motivating use case is a modification nobody has mapped, which is exactly the case most likely to have no close relative available.

Questions for authors
  1. Four of the twelve held-out modifications are chemically near-identical to one another, and the bootstrap currently treats them as four independent observations. What does the interval look like if they are collapsed into a single unit, leaving nine rather than twelve? The grouped holdout answers a related question by removing them from training, but this one is about how the evidence is counted. It is the result that would most affect my confidence in the headline number.
  2. For the external ac4C evaluation, the negatives are positions with reasonable coverage and no published site call. Since ac4C is present at low levels, some of those are presumably real sites that simply were not detected, which would make your numbers look worse than they are. Do you have a sense of how large that fraction might be, or how the precision figures move under a stricter coverage requirement?
Reviewer_MvFsOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper reformulates RNA modification site prediction as a shared compatibility task conditioned on molecular descriptors of the complete modified ribonucleoside. It combines a frozen MultiRM 3-mer embedding table and BiLSTM RNA encoder with Morgan fingerprints, physicochemical descriptors, parent-base identity, target-conditioned cross-attention, and a shared scorer. The main evaluation uses provenance-aware inductive leave-one-modification-out (LOMO) folds for 12 modification types, plus source-separated and grouped-substructure holdouts. Transfer to ac4C is tested on two external datasets without ac4C site-label training or tuning. Reported macro targeted AUC rises from 0.628 for parent-only conditioning to 0.647 with full descriptors; the target-level paired-bootstrap 95% interval for the gain is [+0.010, +0.029]. The framework supports candidate-site prioritization for unseen modification types and for modifications with few annotated sites when their molecular identity is known.

Critical strengths
  1. The zero-shot protocol is carefully controlled. The provenance-aware LOMO setup removes examples attributable only to the held-out target, masks its labels, and removes exact locus/window overlap. Source-separated and grouped-substructure holdouts further probe whether the results survive stricter transfer settings.

  2. Parent-only, permuted-descriptor, nearest-target, and independent-pooling controls help test whether the improvement depends on meaningful target descriptors and target-conditioned RNA context, rather than simply adding a target-specific output head. This makes useful matched comparisons.

  3. The ac4C evaluation uses two datasets without ac4C-specific training or model selection. The AIDO.RNA experiment also retains a positive molecular-conditioning gain (+0.021 versus +0.023 with the scratch BiLSTM), suggesting the effect is not solely compensation for a weak RNA encoder.

Critical weaknesses
  1. There is limited evidence under strict transfer. The gain drops from +0.019 AUC in the primary LOMO setting to +0.003 in the strict substructure holdout, without a confidence interval. This weakens the evidence for robust transfer to chemically less-related targets.

  2. The shared encoder learns from the other 11 modifications, and fingerprint-plus-parent achieves 0.645 AUC versus 0.647 for the full descriptor. Thus, the distinct contribution of molecular descriptors—and of the five scalar features in particular—remains unclear.

  3. Practical evaluation is pretty limited. The main benchmark uses constructed negatives and reports AUC but not AUPRC or precision/recall at relevant operating points, limiting assessment of candidate-site prioritization.

  4. The ac4C evaluation covers two datasets and may still reflect assay or positive-call ascertainment differences. Sample counts and uncertainty after filtering, as well as uncertainty for the strict holdout, would strengthen the evidence.

Questions for authors
  1. How sensitive are the main results to alternative negative-site sampling strategies and to the coverage criteria used to define eligible negatives? Can you report AUPRC or precision/recall at relevant operating points for the 12-target benchmark?

  2. Can you report external ac4C sample counts and uncertainty after both leakage filters, and clarify how assay-specific ascertainment affects interpretation?

  3. The primary full-minus-parent gain is +0.019 AUCb, whereas the strict substructure-holdout gain is only +0.003. Can you provide uncertainty for the strict holdout and assess whether this gain is distinguishable from noise?

  4. How much of the gain comes from molecular descriptors beyond sequence–target relationships learned from the other eleven modifications? Given that fingerprint-plus-parent reaches 0.645 versus 0.647 for the full descriptor, do the scalar physicochemical features contribute meaningfully, and could the learned mapping exploit substructure overlap with training targets?

  5. Table 5 shows that the molecular-information advantage shrinks from +0.019 at 0% target labels to +0.002 at 100%. Is the intended practical regime therefore zero- or very-low-label prediction, and how does the external ac4C setting reflect that use case?

  6. With AIDO.RNA, the gain remains +0.021 compared with +0.023 for the BiLSTM. Can you clarify the matched fold/protocol details and discuss what this result implies about the role of RNA encoder quality?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.