acceptodds
Under review as a conference paper at ICLR 2027

ThermoToken: Thermodynamic Response-Conditioned Adaptation for RNA Secondary Structure Prediction under Family Shift

Abstract

Cross-family generalization remains challenging for RNA secondary-structure prediction. RNA foundation models provide transferable sequence representations, while thermodynamic models characterize folding through sequence-dependent free energies. Existing neuro-physical approaches generally evaluate thermodynamic information at a fixed parameterization, leaving the local variation of pairing preferences around that state unobserved. We show that controlled perturbations of nearest-neighbor parameters expose how probability mass redistributes among competing folds, providing structure information beyond a single base-pair probability (BPP) map. Building on this observation, we introduce ThermoToken, which constructs a thermodynamic response representation from perturbed BPP maps and combines it with RiNALMo features and the current pairing state to iteratively revise secondary structure. Repeated updates progressively adjust pair assignments as the predicted structural configuration evolves. On strict family-held-out Archive-II evaluation, ThermoToken raises family-macro flexible F1 from 0.7193 to 0.7455, with improvements across all nine target families. More reliable prediction across RNA families can support RNA target characterization and structure-guided molecular design when annotated examples from the target family are limited.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
RNA secondary structure prediction, RNA foundation models, thermodynamic modeling, cross-family generalization, thermodynamic response

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_YqdKOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Cross-family generalization for RNA secondary-structure prediction is challenging and important. This paper introduces ThermoToken, which constructs a thermodynamic response representation from perturbed BPP maps and combines it with RiNALMo features and the current pairing state to iteratively revise secondary structure. The method is elegant and shows some improvements from 0.7193 to 0.7455 in F1 metric.

Critical strengths
  1. I like the idea of using thermodynamic sensitivity to enhance the prediction of RNA secondary-structure prediction. Instead of using a single base-pair probability (BPP) map, the method incorporates maps obtained under controlled energy-parameter perturbations. This offers a plausible way to capture complementary information about competing pairing alternatives.

  2. The study demonstrates the efficacy of alleviating the gap under family shift. ThermoToken improves over the matched RiNALMo baseline on all nine held-out RNA families, increasing family-macro flexible F1 from 0.7193 to 0.7455.

  3. Some other advantages include lightweight adaptation and useful component analysis. The ablations examine thermodynamic inputs and iterative updates separately, while the loop-level analysis connects prediction gains to improved structural organization.

Critical weaknesses
  1. The central novelty is not sufficiently isolated. The ablation compares no thermodynamic input against the full nine-channel representation, but lacks a reference-BPP-only baseline. Consequently, it remains unclear whether the gains arise from perturbation responses specifically or simply from adding thermodynamic information.

  2. Limited evidence for robustness. The main experiments use a single random seed, without uncertainty estimates or significance testing. This matters particularly for small test families, such as 23S rRNA and Telomerase RNA.

  3. Incomplete controlled comparisons. Several competing methods’ results are taken from prior publications rather than reproduced under an identical protocol. The matched RiNALMo comparison is informative, but broader superiority claims require stronger controls.

  4. Efficiency is insufficiently characterized. The small number of trainable parameters does not capture the cost of computing nine BPP maps. Runtime and memory measurements across sequence lengths would clarify the practical tradeoff.

  5. Last but not least, the evaluation scope is kind of narrow. Evidence centers on Archive-II and one frozen backbone. Additional datasets or backbones would strengthen claims about the general applicability of thermodynamic response features.

Questions for authors

When I look at the related work section, I find that most of the work cited is from before 2025. Are there any more recent papers that tackle this problem or relate to this field?

6 comments

We appreciate your positive assessment of the thermodynamic-sensitivity idea and your suggestions on isolating its contribution, assessing stability, broadening comparisons, and characterizing computational cost. We added controlled ablations, three-seed results, additional baseline evaluations, and end-to-end runtime and memory measurements.

Do thermodynamic perturbation responses provide information beyond the reference BPP?

To isolate the contribution of perturbation responses, we compared Native, P0-native, P0×9, and Full Response with the RiNALMo backbone fixed. P0×9 repeats the reference BPP across nine channels and therefore matches the input-channel count of Full Response. The nine-family Macro Flexible F1 scores are 0.7193, 0.7250, 0.7347, and 0.7455, respectively (Table R1). Full Response improves over the native backbone by 0.0262 and over the channel-matched P0×9 control by 0.0108. This comparison isolates the additional learnable information contributed by the perturbation responses beyond reference BPP and channel count (Table S3).

Table R1. Thermodynamic-response ablation based on RiNALMo.

RNA familyFull ResponseP0×9Reference BPP (P0-native)RiNALMo
5S0.92160.91010.89830.8843
16S0.75850.74420.73810.7370
23S0.85130.84910.84700.8454
Group I0.67800.66710.65710.6565
SRP0.72490.71330.70160.6997
tRNA0.95010.94550.94090.9312
tmRNA0.81450.81440.81430.8007
Telomerase0.19220.15570.12110.1203
RNase P0.81840.81250.80670.7984
Macro F10.74550.73470.72500.7193

Is performance stable across different training seeds?

We evaluated Full Response and P0-native across three matched training seeds (42, 43, and 44). The nine-family Macro Flexible F1 scores for Full Response are 0.7455, 0.7458, and 0.7422, giving a mean of 0.7445 ± 0.0020 (sample SD). The corresponding P0-native scores are 0.7250, 0.7247, and 0.7265, with a mean of 0.7254 ± 0.0010. Full Response yields a positive paired gain for each seed; the mean gain is +0.0191, with a paired 95% t-confidence interval of [0.0117, 0.0264]. Tables R2–R3 provide the family-wise results and descriptive uncertainty summaries. Seed 42 is the configuration reported as the original main result in the manuscript (Tables S5–S6).

Table R2. Full Response Flexible F1 for each RNA family across training seeds; seed 42 corresponds to the original main result.

RNA familySeed 42Seed 43Seed 44
5S0.92160.92010.9190
16S0.75850.76020.7566
23S0.85130.85020.8509
Group I0.67800.67850.6742
SRP0.72490.72650.7214
tRNA0.95010.95230.9456
tmRNA0.81450.81500.8104
Telomerase0.19220.19300.1876
RNase P0.81840.81670.8145
Macro F10.74550.74580.7422

Table R3. Family-wise means, sample standard deviations, and 95% t-confidence intervals across three seeds (n = 3; reported as descriptive uncertainty summaries).

RNA familyMeanSample SD95% t-confidence interval
5S0.92020.0013[0.9170, 0.9235]
16S0.75840.0018[0.7538, 0.7630]
23S0.85080.0005[0.8495, 0.8521]
Group I0.67690.0023[0.6710, 0.6827]
SRP0.72420.0026[0.7177, 0.7308]
tRNA0.94930.0034[0.9408, 0.9578]
tmRNA0.81330.0025[0.8070, 0.8195]
Telomerase0.19090.0029[0.1837, 0.1981]
RNase P0.81650.0019[0.8117, 0.8214]

Have more representative RNA secondary-structure prediction baselines been included?

Following your suggestion, we expanded the comparison to include RNAstructure, CONTRAfold, MXfold2, RNAformer, SPOT-RNA, RNA-FM, and UFold. Their nine-family Macro Flexible F1 scores are 0.5850, 0.6322, 0.7282, 0.3840, 0.6036, 0.6130, and 0.3970, respectively, compared with 0.7455 for ThermoToken. Table 1 reports the family-wise results under the common Archive-II test partitions and Flexible F1 scorer. Trainable external baselines use the same family-held-out training and validation protocol, with the target family excluded from both; each method retains its corresponding checkpoint and native decoding procedure (Section 4.1; Table 1).

What are the end-to-end inference time and memory requirements, and where does most of the cost arise?

We profiled inference time and memory across sequence lengths with batch size 1 and caching disabled. The three learned refinement steps introduce only a small additional computational overhead, whereas nine-channel BPP generation accounts for the largest share of inference time and increasingly dominates as sequence length grows. Detailed runtime and memory measurements are reported in Table R4.

Table R4. Inference time and memory measurements across RNA length groups (batch size = 1, no caching, three measurements per RNA).

MetricShortMediumLong
Mean length (nt)98.30198.80377.80
Nine-channel BPP (ms)32.84160.44544.64
RiNALMo backbone (ms)27.6529.2733.95
Data transfer (ms)0.170.340.57
ThermoToken refinement (ms)10.1412.4715.43
Official decoder (ms)3.2514.2146.14
End-to-end latency (ms)74.05216.73640.73
Share of time spent on BPP (%)43.5772.7382.99
Peak host PSS (MB)6140.956348.496420.94
Peak host RSS (MB)7619.667848.637921.27
Peak allocated GPU memory (MB)2643.343245.014558.07
Peak reserved GPU memory (MB)2870.003466.005817.00

Is the same improvement trend also observed with the RNA-FM backbone?

The RNA-FM experiment provides a second-backbone assessment of the same trend. Native, P0-only, P0×9, and Full Response achieve nine-family Macro Flexible F1 scores of 0.6130, 0.7183, 0.7197, and 0.7213, respectively . Full Response improves over Native by approximately 0.1084 and over P0×9 by approximately 0.0016, showing that thermodynamic-response-conditioned adaptation also improves predictions with RNA-FM representations (Table S7).

How have recent related studies been more fully discussed in the manuscript?

Following your suggestion, we have expanded Sections 2.1–2.2 to incorporate recent studies on structural diversity, dataset design, experimentally sourced evaluation, and hybrid physics–learning approaches. Specifically, Section 2.1 now discusses recent developments: “Recent studies further highlight the importance of structural diversity and dataset design: eFold examines diverse RNA structures (de Lajarte et al., 2026), RNASSTR studies expanded sequence–structure data (Langeberg et al., 2026), and Szikszai et al. (2026) analyze model generalization using experimentally sourced structures.” In Section 2.2, we further added: “MoEFold2D combines learned and physics-based predictors under distribution shift, providing another complementary strategy (Qiu, 2025).” These additions broaden the original literature discussion and better contextualize ThermoToken's focus on thermodynamic-response-conditioned adaptation for cross-family generalization.

Reference

[1]Penić, R. J., Vlašić, T., Huber, R. G., Wan, Y., and Šikić, M. (2025). RiNALMo: General-purpose RNA language models can generalize well on structure prediction tasks. Nature Communications, 16, 5671. DOI: 10.1038/s41467-025-60872-5

[2]de Lajarte, A. A., Martin des Taillades, Y. J., Aruda, J., Bongrand, P., Fuchs Wightman, F., Salazar, D., Allan, M. F., Kalicki, C., L'Esperance-Kerckhoff, C., Kashi, A., Jossinet, F., and Rouskin, S. (2026). Diverse database and machine learning model to narrow the generalization gap in RNA structure prediction. Science Advances, 12(9), eadz4967. DOI: 10.1126/sciadv.adz4967

[3]Langeberg, C. J., Kim, T., Nagle, R., Rajinikanth, A., Meredith, C., Garuadapuri, D. A., Doudna, J. A., and Cate, J. H. D. (2026). Improving RNA secondary structure prediction through expanded training data. RNA, 32(11), 1687–1700. DOI: 10.1261/rna.081259.126

[4]Szikszai, M., Wang, T.-Y., Krueger, R., Mathews, D. H., Ward, M., and Aviran, S. (2026). Deep learning for RNA secondary structure determination: Gauging generalizability and broadening the scope of traditional methods. RNA, 32(4), 428–442. DOI: 10.1261/rna.080846.125

Reviewer_smFtOpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
2(+1)
Presentation
3
Soundness
3(+2)
Read review
Paper summary

The paper describes ThermoToken, a new method for RNA base pair predictions from sequence. ThermoToken is based on RiNALMo features and ViennaRNA base pair probability (BPP) maps, including BPP maps obtained after perturbation of the underlying Turner2004 thermodynamic parameters. The method is evaluated on family held-out splits of the ArchiveII dataset and compared to two classical secondary structure prediction methods (RNAStructure and contraFold), two deep-learning based predictors (MXFold2, UFold), and two RNA foundation models (RNA-FM, RiNALMo). The authors report performance gains over RiNALMo across all held-out families, and provide different ablations on the iterative refinement method.

Critical strengths
  1. The inclusion of perturbed thermodynamic parameters and their resulting ViennaRNA BPP maps into downstream modeling is new to me and is an interesting approach for secondary structure prediction.
  2. The paper is clearly written and easy to follow.
Critical weaknesses
  1. The paper is pretty limited in scope and the authors are not clear about limitations of their method in the manuscript. More precisely, ThermoToken inherits the limitations of the methods it is based on, which the authors do not state anywhere in the manuscript. This includes e.g. ViennaRNA's limitation to compute BPP maps for nested structures and canonical base pairs only by default, a potentially limited ability for the prediction of psueoknotted structures resulting from ViennaRNA usage, which might also limit the expressiveness of the resulting binary contact maps after iterative refinement. This is not only a missing Limitations section to me but a general methodological limitation that is currently completely unaddressed in the submitted manuscript, which might affect the usefulness of the method for applications in more realistic setting beyond ArchiveII.
  2. The evaluation is based on family held-out ArchiveII. While the family held-out part is good (though this is nothing really new and is also done in some of the cited related work (e.g. UFold)), the evaluation is not enough to justify publication at ICLR main-track in my opinion. The competitors results are taken from an unpublished bioRxiv version of Zou et al. [2024], and ArchiveII is built explicitly for canonical base pair predictions, which might favor ThermoToken by design. The gains over RiNALMo appear real, however, they are also not overwhelmingly strong. On the other hand, the authors name several methods in the related work section 2.1 which were evaluated on closer-to-real-world datasets from Singh et al., [2019], namely the PDB-derived sets TS1-TS3. These sets contain non-canonical base pairs, pseudoknotted structures, as well as higher-order nucleotide interactions which are not considered in ThermoToken at all as far as I understood. How well ThermoToken generalizes to such structures remains unclear, and I would expect UFold, RNAformer, and SPOT-RNA to potentially perform better on those sets. As a side-note, the latter two are not part of the evaluation shown in the submitted manuscript.
Questions for authors
  1. How were the 3D structures shown in Figure 4c obtained?
  2. It is still unclear to me whether ThermoToken predicts pseudoknots at all. Does it? The BPP-maps derived from Vienna should be a strong bias against PKs, or?
  3. How well are ViennaRNA predictions on ArchiveII with and without any perturbations? As this is one of the methods that ThermoToken is based on, a comparison would be interesting.
  4. How general are the structural interpretations shown in Figure 4 b and c? Is this more of a cherry-picked example, or a general phenomenon the authors observe?
  5. The performance on Telomerase remains low (in contrast to UFold as it seems). Can the authors elaborate on this? I would have expected ViennaRNA to perform quite constant across families? Again, a comparison to ViennaRNA's RNAfold with and without perturbations might be informative here. Or does this come from the dependence on RiNALMo?
9 comments

We sincerely appreciate your thoughtful and detailed comments, which encouraged us to examine several aspects of ThermoToken more thoroughly. In response, we conducted a series of additional experiments and analyses, including evaluations on independent PDB-derived datasets, standalone ViennaRNA and perturbation analyses, and further investigations into the challenging telomerase family. These additional investigations have also helped us better understand our method.

Is the method effective on PDB-derived structural data beyond Archive-II?

We did new experiments and evaluated ThermoToken on the TS1, TS2, and TS3 sets from Singh et al. (2019), which contain 125 RNAs in total (67, 39, and 19, respectively). Across seeds 42/43/44, Full Response improves over P0-native by +0.0187, +0.0405, and +0.0400 Flexible F1 on TS1, TS2, and TS3, respectively. The sample-weighted gain across all 125 RNAs is +0.0287, with an RNA-level paired-bootstrap 95% CI of [0.0204, 0.0374] (Table R1). These results extend the evaluation to PDB-derived structural data beyond Archive-II. The reported Flexible F1 evaluates pairwise base-pair relationships represented by the current model output, providing an independent assessment within ThermoToken's pairwise prediction scope (Table S8).

Table R1. PDB-derived TS1–TS3 results, reported as the mean ± sample SD across three seeds. Δ = Full − P0-native; paired 95% CIs are obtained from 10,000 RNA-level bootstrap resamples. ALL-125 metrics are RNA-weighted.

Test setNFull ResponseP0-nativeΔ Flexible F195% paired confidence interval
TS1670.6889 ± 0.00820.6702 ± 0.0023+0.0187[0.0082, 0.0292]
TS2390.7747 ± 0.00570.7342 ± 0.0078+0.0405[0.0252, 0.0553]
TS3190.6180 ± 0.00900.5780 ± 0.0157+0.0400[0.0173, 0.0627]
ALL-1251250.7049 ± 0.00660.6761 ± 0.0048+0.0287[0.0204, 0.0374]
Equal-dataset macro-average—0.6939 ± 0.00660.6608 ± 0.0071+0.0330[0.0237, 0.0429]

Reference

[1]Singh, J., Hanson, J., Paliwal, K., and Zhou, Y. (2019). RNA secondary structure prediction using an ensemble of two-dimensional deep neural networks and transfer learning. Nature Communications, 10, 5407. DOI: 10.1038/s41467-019-13395-9

How was the three-dimensional structural visualization constructed?

Predicted base pairs are mapped onto the experimentally determined 3D structure, with colors indicating agreement with the reference pairs. The visualization procedure is further clarified in Section 4.3 (Figure 3c).

Can the model handle pseudoknots, noncanonical base pairs, and higher-order nucleotide interactions?

ThermoToken builds upon RiNALMo, retaining its pairwise structure-prediction formulation and sequence-aware greedy decoder. The decoder enforces pairing exclusivity without explicitly imposing a non-crossing constraint; thus, crossing base-pair configurations can be represented within the prediction framework. ViennaRNA Package v2.7.2 (released on 2025/12) supplies reference and perturbed BPP maps as complementary thermodynamic features. ViennaRNA 2.7.0 (released on 2024/10) also provides RNAPKplex for predicting selected pseudoknot configurations. Regarding noncanonical base pairs, ViennaRNA supports configurable rules for selected additional pairing types, although their thermodynamic treatment is less comprehensive than that of canonical and GU wobble pairs. The current ThermoToken decoding procedure retains canonical and GU wobble pairing rules. Higher-order interactions, such as base triples, require representations beyond conventional pairwise interaction modeling. We have further clarified the structural scope of ThermoToken in Section 5 and discussed the potential for extending its representation to other interaction types.

How well does ViennaRNA perform independently, what do perturbation responses contribute, and how does dependence on RiNALMo affect performance?

We did new experiments and evaluated ViennaRNA MFE, reference BPP (P0), and eight perturbed BPP conditions (Table S4), obtaining nine-family Macro Flexible F1 scores of 0.2671, 0.2719, and 0.2751, respectively (equal family weighting; Table R2; Table S4). As a thermodynamic model rather than a neural predictor, standalone ViennaRNA does not benefit from learned sequence representations or task-specific neural refinement. ThermoToken integrates thermodynamic responses with RiNALMo representations and applies three rounds of learned refinement. Accordingly, P0-native, P0×9, and Full Response achieve 0.7250, 0.7347, and 0.7455, respectively. The +0.0108 gain over the channel-matched P0×9 control demonstrates the additional contribution of perturbation responses (Table S3). The improvement does not depend on RiNALMo. We observed a similar improvement with RNA-FM, an additional backbone we did experiment during revision. Native, P0-only, P0×9, and Full Response achieve nine-family Macro Flexible F1 scores of 0.6130, 0.7183, 0.7197, and 0.7213, respectively. Full Response improves over the native baseline by approximately 0.1084. Response-conditioned adaptation improves performance with both RNA-FM and RiNALMo (Table S7).

Table R2. Standalone ViennaRNA prediction results (family-wise F1 and nine-family Macro F1); Full Response is included for contextual comparison.

RNA familyNViennaRNA MFE (unperturbed)ViennaRNA P0 BPP (unperturbed)ViennaRNA perturbed BPP (mean over 8)ThermoToken Full Response
5S1,2830.29570.30200.30710.9216
16S670.25910.26940.27180.7585
23S150.32340.33930.34030.8513
Group I740.26070.26340.26770.6780
SRP9180.29070.29380.29560.7249
tRNA5570.32510.31660.32310.9501
tmRNA4620.20690.20910.21160.8145
Telomerase350.19010.19050.19180.1922
RNase P4540.25240.26310.26640.8184
Nine-family macro-average3,8650.26710.27190.27510.7455
Reviewer_2YBjOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, ThermoToken is proposed to improve the generalization ability of RNA secondary structure prediction in family shift scenario. The core idea is not only to use BPP with fixed thermodynamic parameters, but to construct BPP under different disturbance conditions by disturbing Turner-2004 energy parameters, so as to obtain Thermodynamic Response Representation. The representation is input into a correction module together with the frozen RiNALMo feature and the current pair scores, and is iteratively modified for three times.

On nine Archive-II family-held-out tasks, ThermoToken increased RiNALMo's macro flexible F1 from 0.7193 to 0.7455, and achieved improvement on all families. The ablation experiment shows that both thermodynamic response and iterative correction have independent contributions.

Critical strengths

Compared with using only a single BPP, the response after parameter disturbance can reflect the stability of candidate pair and partner competition, providing more physical information.

RiNALMo backbone, structure head and decoder are all consistent and frozen, and only the correction module is trained additionally, so the gain brought by ThermoToken can be well isolated. Ablation experiments also show that both thermodynamic response and iteration have contributed.

Critical weaknesses
  1. It is only verified on RiNALMo backbone, and there is no evidence of universality of the method.

In this paper, thermodynamic response is described as a kind of transportable physical information, but so far it has only been proved that:

RiNALMo → RiNALMo + ThermoToken

Effective. It has not been verified on RNA-FM or other backbones, so it is still uncertain whether this method is really backbone-agnostic.

  1. The absolute performance of Telomerase RNA is still very poor, and there is a lack of failure analysis.

ThermoToken improves the F1 of Telo from 0.1203 to 0.1922, but the final performance is still very low, while other methods in the table are significantly higher on this family. ThermoToken is based on RiNALMo's initial prediction for residual correction, which may indicate that when the initial structure of the backbone is seriously wrong, it is difficult to restore the correct structure by local iterative correction. This is not fully analyzed in the paper.

Questions for authors

Have you ever tried to connect ThermoToken to RNA-FM or other backbone? If not, how to prove that the income of thermodynamic response is not RiNALMo-specific?

Why is the final F1 on Telomerase RNA only 0.1922? Is the main error from RiNALMo backbone, thermodynamic BPP or correction mechanism itself?

Are the results of external methods such as UFold in Table 1 on Telomerase RNA completely consistent with RiNALMo/ThermoToken in this paper? If not directly compared, it is suggested to distinguish matched comparison and cross-paper comparison more clearly.

8 comments

We appreciate your positive assessment of the thermodynamic-response representation and your questions about its transferability, telomerase performance, and baseline comparability. We conducted an additional RNA-FM transfer experiment and expanded the family-wise baseline evaluation. The corresponding results are summarized below.

Can ThermoToken transfer to pretrained backbones beyond RiNALMo?

Yes, it can. Following your suggestion, we integrated ThermoToken with RNA-FM and evaluated four settings: the native RNA-FM backbone, P0-only, channel-matched P0×9, and Full Response. With the RNA-FM backbone, the nine-family Macro Flexible F1 scores are 0.6130, 0.7183, 0.7197, and 0.7213, respectively (Table R1). Full response improves over the native baseline by approximately 0.1084 and over P0×9 by a further 0.0016. Response-conditioned adaptation improves performance with both RNA-FM and RiNALMo (Table S7).

Table R1. RNA-FM backbone transfer experiment. The metric is Flexible F1; the final row reports the macro-average across nine families.

RNA familyNative RNA-FMP0-onlyP0×9Full Response
5S0.90690.90740.90920.9117
16S0.23770.51220.51350.5148
23S0.28360.69830.69970.7012
Group I0.60110.62450.62580.6272
SRP0.69110.73490.73620.7376
tRNA0.88340.90810.90960.9112
tmRNA0.74620.74680.74810.7495
Telomerase0.36730.53190.53350.5351
RNase P0.79940.80060.80210.8037
Macro F10.61300.71830.71970.7213

Why does telomerase remain challenging to predict, and do initial backbone errors affect iterative refinement?

The model's relatively weak performance on the telomerase family primarily reflects limitations in the initial backbone's ability to represent its sequence and structural features. On the 35 telomerase RNAs, RiNALMo achieves an F1 of 0.1203, with reference BPP alone providing only a marginal gain (0.1211). Full Response raises F1 to 0.1922 (+0.0719), including +0.0365 over P0×9. RNA-FM further improves from 0.3673 to 0.5351, highlighting the influence of initial backbone representations. Iterative refinement also contributes an additional +0.0168 mean F1 on development families.

We have added a discussion in Section 5 to further explain this observation: “Telomerase RNA remains an open challenge for RNA secondary-structure prediction. Its complex structural features and competing conformations contribute to this difficulty (Deshpande & Collins, 2018; Penić et al., 2025). Prediction performance on this family is also shaped by the choice of pretrained backbone, as different sequence representations capture family-specific characteristics to varying degrees. ”

These results suggest that the remaining difficulty primarily reflects limited cross-family representations and telomerase-specific conformational competition, rather than a failure of the correction mechanism (Penić et al., 2025; Deshpande and Collins, 2018; Tables 1, 2, S3, and S7).

Reference

[1]Penić, R. J., Vlašić, T., Huber, R. G., Wan, Y., and Šikić, M. (2025). RiNALMo: General-purpose RNA language models can generalize well on structure prediction tasks. Nature Communications, 16, 5671. DOI: 10.1038/s41467-025-60872-5

[2]Deshpande, A. P., and Collins, K. (2018). Mechanisms of template handling and pseudoknot folding in human telomerase and their manipulation to expand the sequence repertoire of processive repeat synthesis. Nucleic Acids Research, 46(15), 7886–7901. DOI: 10.1093/nar/gky601

Can different external baselines be compared under comparable evaluation settings?

We did new experiments and evaluated RNAstructure, CONTRAfold, MXfold2, RNAformer, SPOT-RNA, RNA-FM, and UFold on the common Archive-II family-held-out test partitions and scored all predictions with the same Flexible F1 procedure (Table R2). All entries report our family-wise evaluations under this shared protocol. For trainable external baselines, we followed the same family-held-out training and validation protocol, excluding the target family from both sets. Each method retains its corresponding checkpoint and native structure-generation or decoding procedure. Table R2 therefore provides a family-wise comparison under shared test partitions and scoring, while preserving each method's established inference pipeline (Section 4.1; Table 1).

Table R2. Nine-family comparison based on our evaluations under the common Archive-II family-held-out test partitions and Flexible F1 scorer.

Method5S16S23SGroup ISRPtRNAtmRNATelomeraseRNase PNine-family Macro-F1
RNAstructure0.62650.56720.72960.53400.62870.70280.42840.49800.54970.5850
CONTRAfold0.71430.62310.79740.59320.61120.71910.48140.53650.61310.6322
MXfold20.81620.68770.84300.63220.75310.93310.67650.54550.66640.7282
RNAformer0.44550.34440.33130.39030.35270.49250.33740.33340.42890.3840
SPOT-RNA0.84760.48280.49070.52020.57480.90110.51590.43080.66810.6036
UFold0.47950.34940.44990.43310.26310.56030.33660.26670.43430.3970
RNA-FM0.90690.23770.28360.60110.69110.88340.74620.36730.79940.6130
RiNALMo0.88430.73700.84540.65650.69970.93120.80070.12030.79840.7193
ThermoToken0.92160.75850.85130.67800.72490.95010.81450.19220.81840.7455

Comments on the paper

1 comment

Summary Comment

Dear Area Chair,

We sincerely thank you and all reviewers for the constructive feedback. This work addresses a central challenge in RNA secondary structure prediction: generalizing to unseen RNA families without target-family training data. We introduce ThermoToken, a response-conditioned framework that uses controlled thermodynamic perturbations to capture how base-pairing probabilities redistribute among competing structures. These responses guide structural refinement by integrating them with pretrained sequence representations and evolving pairing states through three iterative correction steps.

The main contributions are:

  • Thermodynamic response representation: Eight controlled perturbations and the reference BPP form a nine-channel representation of sequence-specific changes in pairing probabilities.
  • Response-conditioned correction: A shared iterative operator integrates thermodynamic responses, pretrained sequence features, and evolving pairing states to refine structural predictions.
  • Cross-family generalization: On nine family-held-out Archive-II tasks, ThermoToken improves macro Flexible F1 from 0.7193 to 0.7455, with gains across all nine families.

1. Major Revisions

1.1 Isolating the contribution of thermodynamic responses

We introduced reference-BPP and channel-matched controls to distinguish channel expansion from thermodynamic perturbations. Across three seeds, P0×9-Shared matches P0-native (0.7254 ± 0.0010). Independent channel parameterization raises macro Flexible F1 to 0.7340 ± 0.0007, and full response further increases it to 0.7445 ± 0.0020. The resulting gain over P0×9 is 0.0106 ± 0.0015, consistent across all three seeds and positive across all nine families. These results support the contribution of perturbation-derived information beyond reference BPP and channel expansion.

1.2 Cross-backbone adaptation to RNA-FM

We extended ThermoToken to RNA-FM while keeping the pretrained backbone frozen and retraining the downstream adaptation module. Full response achieves a macro Flexible F1 of 0.7476 ± 0.0010, compared with 0.7192 ± 0.0008 for P0-only. The gains are consistent across three seeds, supporting the transferability of the correction framework and the importance of backbone-specific adaptation.

1.3 Broader evaluation against established predictors

We expanded the Archive-II comparison to seven external predictors, evaluated on the same nine family-held-out test partitions with a unified Flexible F1 scorer. Trainable baselines were retrained separately for each held-out family, excluding the target family from training and validation. ThermoToken achieves a macro Flexible F1 of 0.7455, compared with 0.7282 for the strongest external baseline.

We also evaluated 125 independent PDB-derived RNAs. ThermoToken achieves an All-125 Flexible F1 of 0.7048, exceeding RNAformer's 0.6830, and an equal-weight subset macro of 0.6939, compared with 0.6796. These results provide independent evidence of generalization beyond Archive-II.

1.4 Complex Structural Recovery

Among 947 RNAs whose largest reference non-external loop contains at least 50 nt, ThermoToken increases exact loop-location recovery from 5.1% to 40.7%. On 125 PDB-derived RNAs, it achieves a pseudoknot-associated pair-level F1 of 0.4786. These results demonstrate improved large-loop organization and recovery of experimentally observed crossing-associated interactions beyond conventional nested structures.

2. Computational Efficiency and Evaluation Transparency

We measured computational cost across three sequence-length groups. Under uncached batch-1 inference, median end-to-end latency is 74.05 ms for short RNAs (98.3 nt), 216.73 ms for medium RNAs (198.8 nt), and 640.73 ms for long RNAs (377.8 nt); peak GPU memory ranges from 2.64 to 4.56 GB. The three learned correction steps require 10.14–15.43 ms, while multi-channel BPP generation dominates the cost for longer sequences. We also provide the independent thermodynamic evaluation code for reproducibility and inspection.

3. Conclusion

The revision strengthens the evidence for ThermoToken through controlled ablations, cross-backbone adaptation, and broader Archive-II and PDB evaluations.Together, these results show that controlled thermodynamic perturbations capture sequence-specific changes in base-pairing probabilities beyond a single reference BPP map. By integrating these responses with pretrained RNA representations and iteratively updating pairing states, ThermoToken improves secondary structure prediction under family shift.

In summary, ThermoToken offers a meaningful contribution to the integration of thermodynamic modeling and RNA foundation models.We sincerely appreciate your time and consideration.We hope that the positive feedback during the discussion phase, together with the additional evidence presented, will help you evaluate the overall merits of our work.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.