acceptodds
Under review as a conference paper at ICLR 2027

RNA Foundation Models Use Biology Their Explanations Mostly Miss

Abstract

Identifying which nucleotides drive an RNA model's prediction enables biological discovery beyond prediction alone. Attribution methods provide this insight by scoring each nucleotide's contribution. Their genomic validation has mostly been on task-specific convolutional networks (CNNs) trained from scratch, yet these methods are now applied to pretrained transformer RNA foundation models (FMs). Whether they expose what an RNA FM responds to has not been rigorously tested against labelled biological signal. We use in-silico mutagenesis (ISM), an exhaustive single-substitution scan of the model, to score against the same labels as every method, which we term grounding. ISM shows how much of the labelled signal each model demonstrably uses, a reference every method can be read against. Across six conventional attribution methods, a transformer-specific propagation method (AttnLRP), six RNA FMs (20M–650M parameters) and six RNA tasks, ISM substantially recovers the biological signals on four tasks, yet the conventional methods recover this signal only partially, and recovery varies substantially across FMs of near-identical task performance. We also show that RNA FMs' own substitution responses identify structural base pairs that the conventional methods do not consistently expose. These findings motivate validating each attribution method on the particular fine-tuned model before drawing biological conclusions.

Primary area:
interpretability and explainable AI
Keywords:
attribution methods, RNA FMs

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_YUiEOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This manuscript evaluates the biological grounding and faithfulness of attribution methods applied to RNA foundation models. It compares six conventional attribution methods and AttnLRP across six fine-tuned models and six RNA tasks, using exhaustive single-nucleotide in silico mutagenesis (ISM) as a model-response reference. Attribution and ISM maps are assessed against shared biological annotations and sequence-blind positional baselines, while comprehensiveness and sufficiency are measured on real and dinucleotide-shuffled sequences. The authors report that ISM exceeds the positional baseline on four tasks, whereas conventional attribution methods generally recover less annotated signal and exhibit substantial variation across models. Additional experiments examine the recovery of designed RNA base pairs, comparisons with models trained from scratch, and implementation-specific failures of DeepLIFT. The comparison against each model’s own perturbation responses provides a useful basis for evaluating attribution results, although the broader observation that predictive performance does not guarantee reliable explanations is already established in previous research. Overall, the manuscript presents an RNA-focused empirical evaluation with limited conceptual novelty, whose significance depends on the reliability and practical value of its additional findings.

Critical strengths

Evaluation against model responses and biological annotations

The manuscript compares attribution maps with the same model’s ISM responses against shared biological annotations, allowing poor attribution grounding to be examined alongside evidence of mutation-sensitive model responses. The positional baseline is also a useful control for annotations that can be recovered from their location without sequence-dependent information.

Assessment of structural relationships

The chemical-reactivity experiments extend the evaluation beyond motif localization by examining whether cross-position response maps recover designed RNA base pairs. Matching non-paired controls by distance and partner nucleotide strengthens this comparison by controlling for simple proximity and nucleotide-identity effects.

Concrete implementation diagnostics

Appendix D investigates why the tested DeepLIFT implementations produce results nearly identical to Saliency on five of the six models. The connection between missing propagation rules and attribution outputs provides an actionable diagnostic for researchers applying existing interpretation tools to RNA encoders.

Critical weaknesses

Limited conceptual novelty and insufficiently articulated additional insight

The central concern about the reliability of model explanations has substantial precedent, including Attention is not Explanation (Jain and Wallace, 2019) and Sanity Checks for Saliency Maps (Adebayo et al., 2018). More directly relevant to this manuscript, Majdandzic et al. (2022) demonstrated that genomic models with strong predictive performance can yield substantially different attribution quality, showing that predictive accuracy alone does not guarantee reliable interpretation. Majdandzic et al. (2023) subsequently identified off-simplex gradient noise in one-hot genomic inputs and introduced the correction adopted in this manuscript. Tang et al. (2025) also reported diffuse attribution maps and disagreement across pretrained genomic language models, although their attribution analysis used frozen language models and downstream CNNs. These studies do not establish all of the manuscript’s specific findings for fine-tuned RNA models, but they provide direct precedents for several of its central observations. The manuscript acknowledges this literature, yet should more clearly identify which findings from the ISM-referenced comparisons and CRP experiments extend existing knowledge and explain their practical or scientific significance. As currently presented, the main conclusions largely reinforce established concerns, while the additional insight provided by the RNA-specific empirical extension remains insufficiently clear.

Inconsistent definitions of positional attribution scores

Section 2.1 and Appendix G describe positional scores obtained by gauge-centering the attribution map and selecting the observed nucleotide channel. However, Appendix F specifies channel summation for Integrated Gradients and GradientSHAP, with task-dependent exceptions, while Table 9 retains an earlier positional score for all ten models. These reductions are not mathematically interchangeable and can change position rankings, affecting both biological grounding and deletion-based faithfulness. Because comparative attribution performance is central to the manuscript, the authors should specify the exact reduction underlying each result, justify the exceptions, and recompute affected tables using consistent, method-appropriate formulations. The sensitivity of the PAS results and the comparisons in Table 9 to these choices is particularly relevant.

Questions for authors

Which specific findings from the ISM-referenced evaluation and CRP experiments provide additional insight beyond existing genomic attribution studies?

Could you clarify how these findings change the interpretation or use of attribution methods for RNA foundation models?

3 comments

We thank the reviewer for their time and their astute observations

W1/Q1/Q2: Added insight

We agree that attribution unreliability is established. Our additional finding concerns base-pair relationships in a fixed RNA model’s predictions: its substitution responses recovered designed pairs, while several conventional methods remained near chance. On RNA-FM, fine-tuned to predict chemical reactivity without base-pair targets, Saliency and DeepLIFT scored 0.495 and Input × Gradient 0.517 for base-pair recovery (AUROC), while the same model’s ISM responses scored 0.719 and AttnLRP 0.816 (submitted Table 14; Table 19 in the revised paper), with the checkpoint, annotated pairs and matched non-pairs held fixed.

The additional insight is that several conventional methods can miss pairing-related dependencies expressed in the same model’s predictions. This identifies an attribution failure that could otherwise be mistaken for an absence of structural sensitivity. Practitioners should therefore validate an uninformative structural map against the same model’s substitution responses and annotations. ISM is a reference, not an upper bound.

The ISM-referenced evaluation adds a second finding. Across the four tasks on which ISM exceeds the positional floor, the attribution methods recover a median 60% of the same model’s ISM grounding and 10% of its gain over the floor, so low grounding cannot be explained solely by an absence of model response to the annotated signal.

Action: We have clarified this contribution and its practical consequences in the Introduction, Related Work and Sections 4.3 and 5.

W2: Position scores

We thank the reviewer for identifying this inconsistency. The revised paper states one rule (Section 3.1, submitted Section 2.1; Appendices C, F and G) and applies it in every analysis, so no task-dependent exception remains:

  • IG: retain the uniform reference and sum raw contributions before centring. Since input and reference each sum to one, a gradient component shared across letters cancels in this sum.
  • GradientSHAP: centre each sampled gradient before multiplication by the noisy displacement, which need not sum to zero, then average across draws and sum the raw channels. Centring the final map instead would make its channel sum zero.
  • Other methods and ISM: the centred observed-base score.

Submitted IG/GradientSHAP results retained the earlier centred observed-base score in submitted Tables 2/3/9 (Tables 4/7/14 in the revised paper), SPL/m6A, CRP faithfulness and seed-0 PAS IG grounding. Other main BBP/MRL/PAS results used raw sums. GradientSHAP gradient centring was already present in those main analyses and in the CRP pair maps, but not in the named tables. We recomputed the affected results using the original checkpoints.

Table R2. PAS IG grounding (AUPRC) on the 197 annotated sequences of the 400-sequence draw; the positional floor is 0.138.

CheckpointSubmittedCorrected
RNA-FM0.0870.073
SpliceBERT0.1480.146
ERNIE-RNA0.0610.064
OmniGenome0.0830.092
RiNALMo0.0630.064
AIDO.RNA0.0810.077

The median changes 0.082 → 0.075; SpliceBERT remains the sole above-floor point estimate. Recomputation changes GPU platform and OmniGenome’s IG steps 32 → 64 (others retain 32).

Across 168 combinations (BBP/MRL/PAS/CRP), median grounding/ISM remains 60%; the median share of ISM’s gain above the floor changes 11% → 10% (10.604% → 10.272%); and 59/168 remain at or below the floor. Paired grounding tests against matched ISM remain significant (Holm p ≤ 2.5×10⁻⁵).

Submitted Table 9 (Table 14):

  • Restricting the submitted maps and score to the 400-sequence draw moved the six-method grounding medians by at most 0.03 AUPRC.
  • On identical maps, checkpoints and sequences, the corrected score moved them by −0.007 to +0.028 (Table 5): all 18 CNN–FM comparisons per task favour CNNs under both scores (3 CNNs × 6 FMs; each model’s median grounding over the six conventional methods; BBP and MRL separately). Individual ranks still change: OmniGenome overtakes RNA-FM on BBP, and individual IG/GradientSHAP values change by −0.034 to +0.271.
  • A paired comparison separating gradient centring from the position score (Appendix F.1, Table 6) confirms these comparisons in every setting.
  • Higher CNN comprehensiveness remains 18/18 task–method pairs; lower sufficiency changes 15/18 → 18/18 (submitted exceptions: BBP IG, MRL IG and MRL GradientSHAP). These faithfulness comparisons include scoring, estimator and execution changes; the same-map isolation applies to grounding.
  • CRP pair-recovery AUROC is unchanged; CRP deletion comprehensiveness changes for IG (0.998 → 0.972) and GradientSHAP (1.111 → 1.134).

Action: We have clarified the scoring rules, removed the exceptions and recomputed the affected comparisons, with headline intervals in Appendix A and sensitivity analyses in Appendix F.1.

Thank you to the authors for their detailed rebuttal. The corrections and additional analyses address my methodological concerns, and I now consider the manuscript methodologically sound.

However, my assessment of its contribution remains unchanged. Given the limitations of current RNA foundation models and the established concerns about attribution reliability discussed in my original review, I find the main conclusions largely unsurprising. The manuscript provides further empirical evidence using existing evaluation techniques, but the rebuttal does not convince me that the RNA-specific findings offer substantial additional scientific or practical insight. The additional analyses strengthen the evidence for the reported phenomena without substantially changing my assessment of their novelty or significance. Overall, while my methodological concerns have been addressed, I continue to find the manuscript’s contribution limited.

AuthorsOpenReview ↗

Respnose to Reviewer YUiE

Thank you for confirming that the methodological concerns have been addressed. We appreciate your clarification that the remaining concern is the significance of the contribution. We agree that attribution unreliability is established. We would like to clarify two practical findings that our previous response to you did not make explicit.

  • First, we identify an actionable implementation failure and test an intervention. In the evaluated implementations, DeepLIFT applies no propagation rule to relevant custom activations or function calls in five of the six RNA models, producing Saliency-equivalent maps. Applying the missing activation rescale rule increases median comprehensiveness from 0.57 to 0.81, with improvements in 24 of 30 checkpoint–task settings (Appendix D.1). This gives practitioners a specific failure to check and evidence about the effect of correcting it. Completeness remains unsatisfied, so we do not claim a general DeepLIFT repair or that this improvement establishes biological grounding.

  • Second, we test how practitioners can select an attribution method under a limited ISM budget. Separately for each of 36 checkpoint–task combinations, we select a method using 50 sequences and evaluate its agreement with full ISM on disjoint held-out data, repeating the procedure 500 times. CRP splits hold out whole target structures. The selected method is the held-out best by ISM agreement in a median 96.3% of draws across combinations (range 40.6–100%; Section 4.4, Appendix N). This supports a measured starting budget, with explicit limitations, for checkpoint-specific method selection. It does not certify biological correctness, which requires a separate annotation-based evaluation.

These practical results complement the structural finding discussed previously. For RNA-FM fine-tuned on chemical reactivity without base-pair targets, Saliency gives near-chance pair recovery (AUROC 0.495), while the same checkpoint yields 0.719 with ISM and 0.816 with AttnLRP on the same evaluation. Thus, limitations of the underlying model cannot by themselves explain the failed Saliency map. The practical consequence is to check the explanation method before using that map to dismiss structural sensitivity in the model’s predictions.

We recognise that these are empirical contributions built using existing techniques. Their value lies in identifying a specific implementation failure, testing an intervention, and evaluating an affordable method-selection procedure alongside the RNA structural analysis. We would appreciate reconsideration of the manuscript’s practical significance on this more specific basis.

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

This work tests attribution methods on several tasks across RNA foundation models. They show that different attribution methods vary in effectiveness across task types, and across models even when model performance is similar. Additionally, signal can be present in these foundation models even when conventional attribution methods do not detect anything. The authors show that for the structural task, conventional attribution methods miss structure that in-silico mutagenesis and a method specifically designed for transformers reveals.

Critical strengths

Using ISM as a baseline vs faster attribution methods is very sound experimental design and allows us to compare what changing the input explicitly changes in model output with predicted attributional changes. This kind of approach could be used for other similar interpretability, model introspection, or perturbation prediction works.

The structural task shows the strongest point in the work, that RNA foundation models can contain information that is missed by conventional attribution methods. Attention-aware attribution methods (AttnLRP for example) match ISM much more closely than conventional/convolutional methods.

Critical weaknesses

It is noted that classical attribution methods designed for CNN architectures are not often effective on transformer based foundation models, however the study varies both the architecture and the scale of the models. It would be useful as well to evaluate these methods on foundation scale models with convolutional architectures such as HyenaDNA.

Much of the presentation is difficult to understand for me. Some terms are used critically throughout the work that are not clearly defined before they are used, and I don't believe are standard/used in the standard way (Grounding, Faithfulness for example). Additionally a task is not included in table 1 without explanation, the structure task is not well defined, and the results in figure 1 are presented well before they are discussed or the methods are introduced.

Questions for authors

How do conventional + bespoke attribution methods work on convolutional foundation models?

3 comments

We thank the reviewer for reviewing our work and for their helpful criticisms.

W1/Q1: Conventional and bespoke attribution on convolutional foundation models

We evaluated two DNA-pretrained HyenaDNA models (two layers, 1.6M parameters; eight layers, 6.6M) on BBP and MRL. Both use convolutions without attention. Alongside the six conventional methods and ISM, we evaluated adapted LRP, using the normalisation, activation and multiplication rules from our AttnLRP implementation.

Table R1. Held-out prediction and annotated-site recovery. Prediction is BBP AUROC or MRL Pearson’s r; the remaining columns are AUPRC. Ranges span fine-tuning seeds 0 and 1 for every method, including adapted LRP. The best conventional method is GradientSHAP on BBP and Integrated Gradients on MRL.

TaskHyenaDNAPredictionBest conventionalAdapted LRPISM
BBP2 layers, 1.6M0.882–0.8830.420.57–0.640.54–0.56
BBP8 layers, 6.6M0.880–0.8930.470.560.68–0.71
MRL2 layers, 1.6M†0.806–0.8080.92–0.970.81–0.960.91–0.92
MRL8 layers, 6.6M0.907–0.9110.74–0.800.83–0.840.79–0.81

† The two-layer model predicts MRL less accurately (r = 0.806–0.808) than the eight-layer model (0.907–0.911), so its high grounding should be read with that in mind.

  • On BBP, all six conventional methods scored below ISM at both sizes, while adapted LRP exceeded the best conventional method.
  • On MRL, Integrated Gradients approached ISM in the eight-layer model.
  • ISM outperformed adapted LRP on both faithfulness measures (higher comprehensiveness, lower sufficiency) in the eight-layer HyenaDNA model, for BBP and MRL in each of seeds 0 and 1 (Tables 22 and 23).

These comparisons do not isolate architecture from scale or pretraining, or establish results for RNA decoder or hybrid models.

Action: We have added the two-size, two-seed comparison, predictive performance, adapted-LRP details and faithfulness results to Section 4.3 and Appendix M.1 (Tables 21–23).

W2: Presentation

We thank the reviewer for identifying these presentation issues. We specifically agree that grounding and faithfulness need clearer introductions. The revised paper now defines both terms before either is used:

  • Grounding (Abstract, page 1, lines 17–19): “We evaluate grounding, the recovery of annotated sites or base pairs, by scoring attribution maps against biological annotations.”
  • Grounding and faithfulness (Introduction, page 1, lines 46–48): “Grounding asks whether a map ranks annotated sites or base pairs highly. Faithfulness asks whether perturbing the positions ranked highly changes this model’s output. A map can score well on the second test while recovering little of the annotated signal.”
  • How each is measured (Section 4, pages 4–5): grounding is the AUPRC between a method’s attributions and the labelled signals (AUROC for CRP). Faithfulness is measured by comprehensiveness and sufficiency: “Higher comprehensiveness and lower sufficiency are better (Appendix A.1).” Appendix A.1 gives the exact definition of both measures.

The structure task. Section 3.3 now states: “In CRP, the model predicts chemical reactivity at each nucleotide; base pairs are not training targets and serve only to evaluate explanations (Section 4.3).” Section 4.3 then explains the evaluation before the structure figure (page 8, lines 382–385): “To test for pairing, we substitute the base at position i, read the change in predicted reactivity at every other position j, and ask whether the resulting L×L map ranks the designed base pairs above matched non-pairs (AUROC).”

The structure figure. The figure (submitted Figure 1) is now Figure 4 and appears in Section 4.3, after this explanation. Its caption states that the maps are “shown for AIDO.RNA on one design” and gives the AUROC of each panel, including Saliency 0.50, ISM 0.78 and AttnLRP 0.79.

The task missing from Table 1. Section 4 now states which tasks each analysis uses (page 4, lines 211–214): “The aggregates therefore cover BBP, MRL, PAS and CRP. Table 1 also shows SPL, which the models predict well (macro-F1 0.859–0.869) but whose annotations lie at fixed offsets from the queried site (floor 1.00); the m6A control (Section 4.1) is reported in Table 20.” m6A is the task absent from Table 1. It is a motif-presence control on which neither ISM nor any method exceeds the positional floor (0.83). Table 20 reports SPL and m6A for every method and model.

Action: We have added the definitions, explained the CRP task before its figure, moved the figure to Section 4.3 and clarified its caption, and stated which tasks each analysis uses and why.

Thank you for your detailed response and for addressing my critiques. I think that the updated presentation and additional experiments improve the work, and I will raise my score.

AuthorsOpenReview ↗

Response to Reviewer NPxx

Thank you for your reassessment and for recognizing the improvements from the additional experiments and revised presentation. We are glad that these changes have addressed the critiques raised in your original review.

Could you clarify whether any substantive reservations remain about the revised manuscript? If so, we would appreciate knowing the specific issue so that we can address it directly during the discussion.

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

This paper systematically evaluates popular attribution methods for pre‑trained Transformer‑based RNA foundation models (FMs). While attribution tools were originally validated on scratch‑trained CNN genomic models, their reliability for RNA Transformers has not been rigorously benchmarked. The authors use in‑silico mutagenesis (ISM) as a model‑internal reference: ISM exhaustively perturbs each nucleotide to measure how the model itself responds to biological signals, rather than treating ISM as ground‑truth annotation. They further introduce a positional floor baseline to separate true biological signal from positional statistical bias in annotations.

Experiments cover six RNA foundation models (20 M–650 M parameters), seven attribution approaches (Saliency, Input×Gradient, SmoothGrad, GradientSHAP, Integrated Gradients, DeepLIFT, and Transformer‑aware AttnLRP), plus six RNA tasks including functional prediction and RNA secondary‑structure base‑pair recovery.

Critical strengths

The research addresses a critically important problem and delivers strong cautionary insights for interpretable AI in genomics. RNA foundation models are increasingly deployed with attribution methods for biological discovery, yet the community largely reuses attribution tools originally developed for CNN‑based models. This paper systematically demonstrates that for RNA Transformer foundation models with comparable predictive performance, different attribution methods produce drastically different outputs. A method that works well on model A cannot be assumed to transfer to another model.

Critical weaknesses
  1. ISM incurs substantial computational overhead and is impractical for large‑scale real‑world usage. The paper lacks discussion of lightweight alternative solutions. ISM requires 3L+13L+1 forward passes per sequence, leading to prohibitive costs for long RNA sequences. While the authors note practitioners may perform ISM validation on a small subset of samples, they do not discuss how to verify attribution reliability under limited compute budgets.

  2. All experiments are restricted to BERT‑style encoder‑only RNA foundation models. Decoder‑based and hybrid‑architecture RNA models already exist. Whether the conclusions generalize to other architectures is only mentioned as future work, without exploratory pilot experiments.

  3. The CRP base‑pairing task relies on synthetic RNA sequences; results are not validated on natural RNA, which casts doubt on real‑world applicability of these findings.

  4. The paper observes that DeepLIFT degrades to behave equivalently to Saliency on 5 out of 6 RNA foundation models. Although the phenomenon is identified, further diagnosis is absent: which Transformer components cause the breakdown of DeepLIFT propagation rules remains unexplored.

  5. Faithfulness metrics (comprehensiveness / sufficiency) can only tell good methods apart, yet have limited power to judge the quality of per‑sample explanations. Statistical results are reported, but there is a lack of concrete illustrative case examples to demonstrate this limitation intuitively.

Questions for authors
  1. ISM incurs substantial computational overhead and is impractical for large‑scale real‑world usage. The paper lacks discussion of lightweight alternative solutions. ISM requires 3L+13L+1 forward passes per sequence, leading to prohibitive costs for long RNA sequences. While the authors note practitioners may perform ISM validation on a small subset of samples, they do not discuss how to verify attribution reliability under limited compute budgets.

  2. All experiments are restricted to BERT‑style encoder‑only RNA foundation models. Decoder‑based and hybrid‑architecture RNA models already exist. Whether the conclusions generalize to other architectures is only mentioned as future work, without exploratory pilot experiments.

  3. The CRP base‑pairing task relies on synthetic RNA sequences; results are not validated on natural RNA, which casts doubt on real‑world applicability of these findings.

  4. The paper observes that DeepLIFT degrades to behave equivalently to Saliency on 5 out of 6 RNA foundation models. Although the phenomenon is identified, further diagnosis is absent: which Transformer components cause the breakdown of DeepLIFT propagation rules remains unexplored.

  5. Faithfulness metrics (comprehensiveness / sufficiency) can only tell good methods apart, yet have limited power to judge the quality of per‑sample explanations. Statistical results are reported, but there is a lack of concrete illustrative case examples to demonstrate this limitation intuitively.

1 comment

We thank the reviewer for their care and feedback when reviewing our work

W1/Q1: ISM cost

We agree that practical guidance is needed. For each of the 36 checkpoint–task combinations, we drew 50 or 100 sequences 500 times, selected the method that agreed best with full ISM (median per-sequence Spearman correlation: agreement with ISM, not biological grounding) and tested the choice on disjoint held-out sequences. CRP draws held out whole target structures, so a parent design and its mutants stayed together.

  • With 50 sequences the selected method was the held-out best in a median 96.3% of draws (range 40.6–100%), and with 100 in 99.5% (43.2–100%). Both minima are RNA-FM on SPL, whose leading methods are nearly tied: a wrong choice cost 0.005 there (best method’s held-out Spearman agreement minus the selected method’s, averaged over wrong choices only). The largest average in any combination, 0.100 (RiNALMo on BBP), came from one wrong choice in 500 draws, not a near tie.
  • One random substitution per position (redrawn each draw), about one-third of the model evaluations, gave 95.5% (43.0–100%) with 50 sequences.
  • Fifty sequences are therefore a reasonable starting budget, not a guarantee, on these 41–206-nt tasks, to be enlarged when the leading methods are close.
  • Full ISM on 50 sequences takes an estimated 11 s (SpliceBERT) to 4.4 min (AIDO.RNA) at 206 nt, and 5.5 min to 2.2 h at 1,022 nt, on one GH200.

Action: Added as Section 4.4 and Appendix N (Tables 24–26).

W2: Encoders only

We decided upon BERT-style encoders to represent the current approaches to RNA FMs; however, we agree that this scope limits generalisation. We therefore fine-tuned HyenaDNA, a DNA-pretrained model that reads left to right using convolutions without attention, with our training recipe and two seeds.

On the eight-layer model (held-out BBP AUROC 0.880–0.893; MRL r 0.907–0.911), all six conventional methods score below ISM on BBP (AUPRC: median 0.31–0.32, best 0.47, ISM 0.68–0.71), while Integrated Gradients approaches ISM on MRL (0.74–0.80 against 0.79–0.81). HyenaDNA predicted CRP less accurately than the RNA models, so we do not use that task. This is an initial test beyond bidirectional encoders; RNA decoder and hybrid models remain to be evaluated.

Action: Added to Section 4.3 and Appendix M.1 (Tables 21–23; orange text, shared with Reviewer NPxx).

W3: Synthetic RNA

Nine of the 20 CRP target structures originate from natural RNAs, although most of the evaluated sequences are redesigns. We compared each method with ISM on the same designs, separately by target origin. Two natural-origin targets failed the 20-pair minimum and the engineered guide RNA was set aside, leaving 17 targets and 123 of the 131 scored designs (34 of 36 for the costly methods).

At the median across six models, every method achieved lower base-pair AUROC than ISM in both groups: the difference ranged from −0.111 (AttnLRP) to −0.250 (Saliency/DeepLIFT) for natural-origin targets and from −0.119 to −0.237 for designed targets, and every 95% interval (2,000 resamples of whole targets) lay below zero. These results concern natural-origin target structures; they do not validate performance on native RNA sequences.

Action: Added to Section 4.3 and Appendix J.7 (Table 9).

W4: DeepLIFT

We agree that a clearer diagnosis was needed.

  • Appendix D’s hook checks found no DeepLIFT rules on the custom activations or function calls in five models, yielding Saliency-equivalent maps. In ERNIE-RNA, tangermeme recognised the standard GELU in the pairing-bias projection; Captum 0.9.0 did not.
  • On 30 checkpoints (six models × five tasks; 400 examples per task), applying DeepLIFT’s activation rescale rule increased median comprehensiveness from 0.57 to 0.81, improving 24/30 comparisons.
  • Contributions still did not sum to the prediction change, and the tested attention-softmax rule caused numerical instability.
  • Practitioners should verify which rules their implementation applies and check explanation quality and completeness after adapting it.

Action: Added to Section 5 and Appendix D.1 (Tables 2 and 3); the findings concern the tested implementations.

W5: Example

For Saliency on one OmniGenome checkpoint, two held-out BBP sequences (IDs 511 and 532) contain the identical six-base QKI motif, and ISM ranks all six motif bases first in both. Their comprehensiveness is nearly identical (0.746 and 0.744; sufficiency 0.328 and 0.440), yet motif recovery (AUPRC) is 0.804 and 0.154. We selected the largest motif-recovery difference among eligible pairs with the same method, checkpoint and motif type, and comprehensiveness within 0.05. Comprehensiveness alone therefore did not establish an individual explanation’s biological grounding; Figure 3 (submitted Figure 4) gives the broader result (median within-method Spearman’s ρ = 0.27).

Action: Added to Section 4.2 and Appendix O (Figure 14).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.