Pith. sign in

REVIEW 4 major objections 6 minor 48 references

From Sentences to Sequences: Rethinking Languages in Biological System

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper shows that biomolecule sequences should be generated in stochastic order rather than left-to-right, treating 3D structure as the semantic content, and demonstrates with the RiFold model that this improves RNA inverse folding.

desk verdict A useful RNA inverse folding model with a real empirical contribution, but the energy metric as written is internally contradictory and must be fixed before the structural claims can be trusted. read the letter →

arxiv 2507.00953 v2 pith:RBLXGHJG submitted 2025-07-01 q-bio.BM cs.AI

Ke Liu , Shuaike Shen , Hao Chen This is my paper
classification q-bio.BMcs.AI
keywords inversefoldingstochastic-ordergenerationRNAdesignproteinstructurerecoverybiologicallanguagemodelingTM-scoreRiFold
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Biomolecules carry their meaning in three-dimensional structure, not in their linear sequence, so carrying over NLP's left-to-right generation and token-matching evaluation into biology is a mismatch. The paper argues that because residues and bases interact over long ranges—base pairing in RNA, hydrogen bonds in proteins—sequence tokens should be generated in a stochastic order, placing the most confident positions first rather than following the 1D chain. It tests this on inverse folding (designing a sequence that folds into a given structure) and reports that its model RiFold beats prior state of the art on RNA, with larger gains when scoring by structural recovery (TM-score, RMSD, energy) than by sequence recovery (NSR, Macro-F1). The paper also documents that sequence recovery and structure recovery are correlated but not aligned, so high token match does not guarantee that the designed molecule folds correctly. If correct, this reframes how biological language models should be built and evaluated.

What carries the argument

The central mechanism is stochastic-order autoregressive decoding: the model is an encoder-decoder over the target structure graph, and the decoder chooses which position to fill next based on the confidence of the currently unmasked positions, so tokens that interact in 3D (base pairs, contacting residues) tend to be generated in nearby decoding steps. The second piece is the structure recovery evaluation pipeline: candidate sequences are folded by ESMFold or E2EFold and compared to the target structure by TM-score, RMSD, and energy. The paper's RNA model RiFold pairs a structure featurizer (dihedral angles, distances, directions) with this decoding scheme and beam search.

What would settle it

Take a set of RNA inverse-folding targets with experimentally determined structures (e.g., RNA-Puzzles blind predictions or crystal structures), generate candidate sequences with RiFold and RDesign, and fold them experimentally (e.g., by SHAPE probing or crystallography). If RiFold's advantage over RDesign in TM-score computed by E2EFold disappears when structures are measured experimentally—or if high TM-score sequences fail to fold—the stochastic-order advantage and the structure-recovery claims would be revealed as artifacts of the in-silico evaluator.

Watch

Extended reading notes

Core claim

The central claim is that the working definition of 'language' in biomolecular modeling should be reversed: the 3D structure is the semantics, the sequence is the surface form. From this, the paper derives two prescriptions. Generation should be stochastic-order: at each step, emit the token at the position with highest confidence, since spatially close but sequence-distant tokens (e.g., base-paired nucleotides) must be co-decided; sequential left-to-right decoding locks the model into local context and misses those interactions. Evaluation should be structure-based: fold the predicted sequence with ESMFold (protein) or E2EFold (RNA), then compare the predicted structure to the target using TM-score, RMSD, and energy, rather than counting token overlap. Empirically, the paper claims stochastic-order RiFold outperforms RDesign on RNA inverse folding, and that among protein inverse-folding models the ranking by structure recovery differs from the ranking by sequence recovery, with perplexity correlating strongly (-0.90) with structure recovery while NSR correlates only moderately (0.63).

Load-bearing premise

The strongest results depend on ESMFold and E2EFold predicting the real folded structures of designed sequences; if those predictors are biased toward or against certain sequences, the reported TM-scores, RMSDs, and energy values could be artifacts rather than true folding behavior.

Editorial extensions

If this is right

  • Benchmarks that rank inverse folding models by native sequence recovery will mis-rank them relative to structural fidelity; perplexity is a better fast proxy than NSR.
  • Stochastic-order decoding should be applied beyond RNA to any biomolecule design task where long-range 3D contacts matter, including protein binder and enzyme design.
  • The reported gap between sequence and structure recovery means models optimized for one objective will not automatically optimize the other; evaluation should always fold and compare structures.
  • RiFold's gains concentrate on RNA with canonical base pairs, supporting the mechanistic story that the decoding order is what recovers long-range pairing.
  • Beam search over both decoding positions and token types further improves RiFold, so decoding-order search is a usable axis of scaling.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If structure recovery is the true objective, then training objectives should also become structure-aware; the paper evaluates structure post hoc but trains with token-level likelihood, so a structure-aware loss is a natural extension.
  • The strong correlation between perplexity and structure recovery suggests perplexity could be used as a cheap filter during training or early stopping, with full folding evaluation only on the survivors.
  • The stochastic-order advantage may transfer to other structured outputs where order is arbitrary, such as set generation or molecular graph generation, not just biopolymers.
  • The RNA results could be tested experimentally: sequences designed by RiFold on known aptamer or ribozyme structures could be synthesized and their folding verified by crystallography or SHAPE-seq, which would separate evaluator bias from true folding fidelity.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper argues that biological sequences are not well modeled by the usual NLP assumptions: their semantics are grounded in 3D structure, and they contain long-range inter-token dependencies. On that basis it advocates structure-aware evaluation (TM-score, RMSD, energy) for inverse folding, and proposes RiFold, a confidence-ordered autoregressive RNA inverse folding model that is compared against RDesign and other baselines. The paper reports improved sequence recovery and Macro-F1 over RDesign, improved structure recovery on RNA, and a correlation analysis on proteins suggesting that sequence and structure recovery are related but not identical. The manuscript includes code and an appendix with ablations and additional results.

Significance. If the empirical claims are correct, the paper is a useful contribution: it provides a concrete test of decoding-order flexibility for inverse folding, advocates evaluation metrics that are arguably more meaningful than token-level recovery, and releases code. The RNA results are consistent with the intuition that fixed left-to-right decoding is suboptimal for structure-conditioned sequence design, and the comparison of structure versus sequence recovery is a valuable observation for the field. However, the significance is conditional: the structural evaluation, which anchors the central claim, currently contains a direct contradiction in the energy metric and lacks pass-rate statistics for RNA. These issues must be resolved before the paper's main conclusions can be accepted.

major comments (4)
  1. [Section 4.3, Table 4, Figure 5] The paper states in Section 4.3 that 'lower energy indicates a more stable (and thus more plausible) structure', but Table 4 labels the metric 'Energy ↑' and Figure 5 says 'Larger is better'. RiFold's mean energy (5.8757) is higher than RDesign's (5.7646), so the two conventions lead to opposite conclusions about whether RiFold improves energy. The sentence in Section 5.1.2 that '60.22% of the sequences predicted by RiFold achieved a lower energy than RDesign' is consistent with lower-is-better for a majority, but then the mean and median values in Table 4 are in the wrong direction, and the 'Energy ↑' header is incorrect. If instead larger-is-better is the intended convention, then Section 4.3's definition and the '60.22% lower energy' sentence are wrong. The authors must state the energy convention unambiguously, correct the table/figure/text accordingly, and re-examine the claim in Section 5.1.2 that RiFold outperforms RDesign in 'all three metrics'. The units 'Energy(log-)' in Table 4 should also be defined.
  2. [Section 4.3, Section 5.1.2, Table 4] Section 4.3 defines successful structure recovery as TM-score > 0.5 and RMSD < 2 Å, but for RNA the paper reports only mean and median TM-score and RMSD values, with no pass-rate statistics. The reported means (e.g., TM-score 0.258 and RMSD 12.06 Å for RiFold) are far from the defined success thresholds. Without the fraction of sequences satisfying the stated criteria, the reader cannot determine whether the claimed improvement represents meaningful structural recovery or a shift among structures that all fail the criterion. Please report the pass rates defined in Section 4.3 for both RiFold and RDesign, and include interval estimates or paired statistics for the structure metrics rather than only group means.
  3. [Section 6, Section A.2.2] The structural evaluation pipeline relies on ESMFold and E2EFold, and Section 6 acknowledges that these predictors 'may introduce biases or noise'. This acknowledgment is important because the paper's central claim is about structure recovery, not just sequence recovery. In particular, the RNA energy is reported after E2EFold's restrained AMBER minimization (Section A.2.2), so it is not an independent physical measurement. I ask the authors to add a calibration or control experiment, for example folding native sequences with the same predictors and reporting the TM-score/RMSD/energy distributions, or comparing all methods under an alternative folding tool, to show that the reported differences are not artifacts of predictor bias favoring RiFold over RDesign.
  4. [Section 3.1, Algorithm 1, Section 5.1.2] The paper's headline claim is that 'stochastic-order generation' works better than sequential-order generation, but the method actually selects positions by maximum confidence: Algorithm 1 and Section 5.1.2 describe argmax-based position selection, with beam search over starting positions, and no stochastic sampling of the generation order is used. The formal definition in Section 3.1 permits any permutation of positions, so 'stochastic' is a misnomer for the implemented procedure. Either rename the method and the claim (e.g., 'confidence-ordered' or 'permutation-order' decoding), or add and evaluate a genuinely stochastic-order variant. As written, the evidence supports a deterministic adaptive order, not the stated stochastic-order claim.
minor comments (6)
  1. [Section 5.1.2] The sentence 'the average confidence for RDesign is only 0.4356, which means that RiFold is not certain about its prediction' should refer to RDesign, not RiFold.
  2. [Table 2] Several 'All' Macro-F1 values are impossible as weighted averages of the subgroup means; for example, SeqRNN (h=128) has subgroup means 17.22, 17.20, and 8.44 but 'All' is 17.74, and SeqRNN (h=256) has subgroup means 12.54, 13.64, and 8.85 but 'All' is 13.64. Please verify the aggregation.
  3. [Table 10] For beam width 1, the 'All' NSR of 53.52 is lower than all three subgroup values (57.07, 56.27, 61.80), which cannot be a weighted average; check whether these numbers were aggregated incorrectly.
  4. [Table 5 caption] The caption uses 'RIFOLD' in uppercase while the text uses 'RiFold'; make the capitalization consistent.
  5. [Section 1] The phrase 'to access semantic fidelity' should be 'to assess semantic fidelity'.
  6. [Table 4 caption] The metric label 'Energy(log-), RMSD(Å)' is not self-explanatory; define the logarithmic transformation and the units, and state whether higher or lower energy is better consistently with the text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's empirical claims are benchmarked against external datasets and external structure predictors (ESMFold, E2EFold), with no self-citation chain or fitted-input-as-prediction step.

full rationale

The central claims are (1) stochastic-order generation outperforms sequential-order on inverse folding, (2) structure-based metrics better capture biological 'semantics', and (3) sequence recovery and structure recovery are related but not consistent. None of these reduces to its inputs by construction. RiFold is compared with RDesign, PiFold, ProteinMPNN, and others on held-out splits of RNAsolo, RNA-Puzzles, and CATH; the improvement is evaluated with external folding tools (ESMFold_v1, E2EFold) and standard metrics (TM-score, RMSD, NSR, Macro-F1), not with quantities fitted from the model's own outputs. The ablation study (Table 5) isolates stochastic-order decoding and beam search, providing an internal but non-circular check. The paper contains no load-bearing self-citation: the reference list includes no prior work by the present authors that is invoked to justify the main claims. The only relevant caveats are external to circularity: the structure-prediction tools may bias structure recovery (explicitly acknowledged in Section 6), and the energy metric direction is internally inconsistent (Section 4.3 says lower energy is more stable, while Table 4 and Figure 5 label Energy with an up arrow and 'Larger is better'). That inconsistency is a correctness/consistency concern, not a circularity, because the energy values come from an external predictor and the comparison is not definitionally forced. Treating 3D structure as 'semantics' is a framing/definition used to motivate the metric choice, not a result derived from that definition.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The paper is an empirical machine-learning study, so it adds no free parameters beyond decoding hyperparameters and relies on domain assumptions about structure-as-semantics, folding-tool accuracy, threshold conventions, and dataset representativeness. No new entities (particles, forces, dimensions) are introduced.

free parameters (1)
  • Beam width w = 4 for RNA (tested 1, 3, 5)
    Decoding hyperparameter searched on validation; larger beams generally improve RiFold, so the reported gains partly depend on this hand-chosen value.
assumptions (4)
  • domain assumption The 3D structure of a biomolecule is the semantic content of its sequence, so structural similarity metrics measure semantic fidelity.
    Section 4.2 and the evaluation pipeline treat folded structure as ground-truth semantics; this is a physically motivated premise, not a proven equivalence.
  • domain assumption ESMFold and E2EFold produce sufficiently accurate structures for predicted sequences to serve as evaluation oracles.
    Section A.2.2 uses these tools to fold predicted sequences; Section 6 states they may introduce bias or noise.
  • domain assumption TM-score > 0.5 and RMSD < 2 angstrom are standard thresholds for successful structure recovery.
    Section 4.3 defines success criteria; these thresholds are conventional but arbitrary.
  • domain assumption Inverse folding datasets and splits (RNAsolo, RNA-Puzzles, CATH) are representative and follow prior work.
    Sections 5.1.1 and 5.2.1 invoke existing dataset conventions; representatives is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Sentences to Sequences: Rethinking Languages in Biological System." pith.science (2026). https://pith.science/paper/RBLXGHJG

@misc{pith2026250700953,
  author       = {Pith},
  title        = {Pith review of: From Sentences to Sequences: Rethinking Languages in Biological System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RBLXGHJG}},
  note         = {Machine review of arXiv:2507.00953}
}
read the original abstract

The paradigm of large language models in natural language processing (NLP) has also shown promise in modeling biological languages, including proteins, RNA, and DNA. Both the auto-regressive generation paradigm and evaluation metrics have been transferred from NLP to biological sequence modeling. However, the intrinsic structural correlations in natural and biological languages differ fundamentally. Therefore, we revisit the notion of language in biological systems to better understand how NLP successes can be effectively translated to biological domains. By treating the 3D structure of biomolecules as the semantic content of a sentence and accounting for the strong correlations between residues or bases, we highlight the importance of structural evaluation and demonstrate the applicability of the auto-regressive paradigm in biological language modeling. Code can be found at \href{https://github.com/zjuKeLiu/RiFold}{github.com/zjuKeLiu/RiFold}

Figures

Figures reproduced from arXiv: 2507.00953 by the authors.

Figure 1
Figure 1. Structures of protein and RNA. (a) protein structure and sequence. One protein structure corresponds to [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. In sequential-order generation, tokens are generated from left to right and tokens are allowed to be decoded [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Replacing words with synonyms will not lead [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Evaluation workflow for biomolecule inverse [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Structure recovery and energy comparison between RiFold and RDesign. C3 and C4 indicate the results [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: The correlation between structure and sequence recovery. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: The detail architecture of RiFold. Width Macro-F1 (×100) ↑ Sequence Recovery (%) ↑ Short Medium Long All Short Medium Long All 1 40.87 45.85 42.62 43.64 41.90 45.76 43.72 42.86 3 41.15 45.97 42.37 43.68 42.07 45.90 43.43 43.02 5 41.22 46.36 42.44 43.87 42.45 46.24 43.5…
Figure 8
Figure 8. Figure 8: The correlation between structure and sequence recovery. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 31 canonical work pages

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  2. [2]

    Bartosz Adamczyk, Maciej Antczak, and Marta Szachniuk. 2022. Rnasolo: a repository of cleaned pdb-derived rna 3d structures. Bioinformatics, 38(14):3668--3670

    work page 2022

  3. [3]

    Satanjeev Banerjee and Alon Lavie. 2005. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65--72

    2005

  4. [4]

    Aadyot Bhatnagar, Sarthak Jain, Joel Beazer, Samuel C Curran, Alexander M Hoffnagle, Kyle Ching, Michael Martyn, Stephen Nayfach, Jeffrey A Ruffolo, and Ali Madani. 2025. Scaling unlocks broader generation and deeper functional understanding of proteins. bioRxiv, pages 2025--04

    work page 2025

  5. [5]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, and 1 others. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877--1901

    2020

  6. [6]

    Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. 2020. Generative pretraining from pixels. In International conference on machine learning, pages 1691--1703. PMLR

    work page 2020

  7. [7]

    Justas Dauparas, Ivan Anishchenko, Nathaniel Bennett, Hua Bai, Robert J Ragotte, Lukas F Milles, Basile IM Wicky, Alexis Courbet, Rob J de Haas, Neville Bethel, and 1 others. 2022. Robust deep learning--based protein sequence design using proteinmpnn. Science, 378(6615):49--56

    work page 2022

  8. [8]

    Noelia Ferruz, Steffen Schmidt, and Birte H \"o cker. 2022. Protgpt2 is a deep unsupervised language model for protein design. Nature communications, 13(1):4348

    2022

Show all 48 references
  1. [9]

    Zhangyang Gao, Cheng Tan, Xingran Chen, Yijie Zhang, Jun Xia, Siyuan Li, and Stan Z Li. 2023 a . Kw-design: Pushing the limit of protein deign via knowledge refinement. In The Twelfth International Conference on Learning Representations

    2023

  2. [10]

    Zhangyang Gao, Cheng Tan, and Stan Z Li. 2022. Alphadesign: A graph protein design method and benchmark on alphafolddb. arXiv preprint arXiv:2202.01079

    2022 arXiv

  3. [11]

    Zhangyang Gao, Cheng Tan, and Stan Z. Li. 2023 b . https://openreview.net/forum?id=oMsN9TYwJ0j Pifold: Toward effective and efficient protein inverse folding . In The Eleventh International Conference on Learning Representations

    2023

  4. [12]

    Zhangyang Gao, Cheng Tan, Yijie Zhang, Xingran Chen, Lirong Wu, and Stan Z Li. 2024. Proteininvbench: Benchmarking protein inverse folding on diverse tasks, models, and metrics. Advances in Neural Information Processing Systems, 36

    2024

  5. [13]

    Thomas Hayes, Roshan Rao, Halil Akin, Nicholas J Sofroniew, Deniz Oktay, Zeming Lin, Robert Verkuil, Vincent Q Tran, Jonathan Deaton, Marius Wiggert, and 1 others. 2025. Simulating 500 million years of evolution with a language model. Science, page eads0018

    2025

  6. [14]

    John Ingraham, Vikas Garg, Regina Barzilay, and Tommi Jaakkola. 2019. Generative models for graph-based protein design. Advances in neural information processing systems, 32

    2019

  7. [15]

    Bowen Jing, Stephan Eismann, Patricia Suriana, Raphael Townshend, and Ron Dror. 2020. Learning from protein structure with geometric vector perceptrons. Int. Conf. on Learning Representations

    2020

  8. [16]

    Maria U Johansson, Vincent Zoete, Olivier Michielin, and Nicolas Guex. 2012. Defining and searching for structural motifs using deepview/swiss-pdbviewer. BMC bioinformatics, 13:1--11

    2012

  9. [17]

    Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74--81

    2004

  10. [18]

    Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Nikita Smetanin, Robert Verkuil, Ori Kabeli, Yaniv Shmueli, and 1 others. 2023 a . Evolutionary-scale prediction of atomic-level protein structure with a language model. Science, 379(6637):1123--1130

    2023

  11. [19]

    Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Nikita Smetanin, Robert Verkuil, Ori Kabeli, Yaniv Shmueli, and 1 others. 2023 b . Evolutionary-scale prediction of atomic-level protein structure with a language model. Science, 379(6637):1123--1130

    2023

  12. [20]

    Nono SC Merleau and Matteo Smerlak. 2022. arnaque: an evolutionary algorithm for inverse pseudoknotted rna folding inspired by l \'e vy flights. BMC bioinformatics, 23(1):335

    2022

  13. [21]

    Zhichao Miao, Ryszard W Adamiak, Maciej Antczak, Micha J Boniecki, Janusz Bujnicki, Shi-Jie Chen, Clarence Yu Cheng, Yi Cheng, Fang-Chieh Chou, Rhiju Das, and 1 others. 2020. Rna-puzzles round iv: 3d structure predictions of four ribozymes and two aptamers. Rna, 26(8):982--995

    2020

  14. [22]

    Erik Nijkamp, Jeffrey A Ruffolo, Eli N Weinstein, Nikhil Naik, and Ali Madani. 2023. Progen2: exploring the boundaries of protein language models. Cell systems, 14(11):968--978

    2023

  15. [23]

    Christine A Orengo, Alex D Michie, Susan Jones, David T Jones, Mark B Swindells, and Janet M Thornton. 1997. Cath--a hierarchic classification of protein domain structures. Structure, 5(8):1093--1109

    1997

  16. [24]

    C Nick Pace, Hailong Fu, Katrina Lee Fryar, John Landua, Saul R Trevino, David Schell, Richard L Thurlkill, Satoshi Imura, J Martin Scholtz, Ketan Gajiwala, and 1 others. 2014 a . Contribution of hydrogen bonds to protein stability. Protein Science, 23(5):652--661

    2014

  17. [25]

    C Nick Pace, Hailong Fu, Katrina Lee Fryar, John Landua, Saul R Trevino, David Schell, Richard L Thurlkill, Satoshi Imura, J Martin Scholtz, Ketan Gajiwala, and 1 others. 2014 b . Contribution of hydrogen bonds to protein stability. Protein Science, 23(5):652--661

    2014

  18. [26]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311--318

    2002

  19. [27]

    William Peebles and Saining Xie. 2023. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195--4205

    2023

  20. [28]

    Maja Popovi \'c . 2015. chrf: character n-gram f-score for automatic mt evaluation. In Proceedings of the tenth workshop on statistical machine translation, pages 392--395

    2015

  21. [29]

    Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, and 1 others. 2018. Improving language understanding by generative pre-training

    2018

  22. [30]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, and 1 others. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9

    2019

  23. [31]

    Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie. 2020. Comet: A neural framework for mt evaluation. arXiv preprint arXiv:2009.09025

    2020 arXiv

  24. [32]

    \'A lvaro Rubio-Largo, Nuria Lozano-Garc \' a, Jos \'e M Granado-Criado, and Miguel A Vega-Rodr \' guez. 2023. Solving the rna inverse folding problem through target structure decomposition and multiobjective evolutionary computation. Applied Soft Computing, page 110779

    2023

  25. [33]

    Frederic Runge, Danny Stoll, Stefan Falkner, and Frank Hutter. 2018. Learning to design rna. arXiv preprint arXiv:1812.11951

    2018 arXiv

  26. [34]

    Tao Shen, Zhihang Hu, Zhangzhi Peng, Jiayang Chen, Peng Xiong, Liang Hong, Liangzhen Zheng, Yixuan Wang, Irwin King, Sheng Wang, and 1 others. 2022. E2efold-3d: end-to-end deep learning method for accurate de novo rna 3d structure prediction. arXiv preprint arXiv:2207.01586

    2022 arXiv

  27. [35]

    Matthew Snover, Bonnie Dorr, Richard Schwartz, Linnea Micciulla, and John Makhoul. 2006. A study of translation edit rate with targeted human annotation. In Proceedings of the 7th Conference of the Association for Machine Translation in the Americas: Technical Papers, pages 223--231

    2006

  28. [36]

    M Spencer. 1959. The stereochemistry of deoxyribonucleic acid. ii. hydrogen-bonded pairs of bases. Acta Crystallographica, 12(1):66--71

    1959

  29. [37]

    L Stryer, JL Tymoczko, and JM Berg. 2002. Biochemistry 5th ed freeman. WH and Company, 41

    2002

  30. [38]

    Cheng Tan, Zhangyang Gao, Jun Xia, and Stan Li. 2022. Generative de novo protein design with global context. Proc. Int. Conf. Acoustics, Speech, and Signal Processing

    2022

  31. [39]

    Cheng Tan, Yijie Zhang, Zhangyang Gao, Bozhen Hu, Siyuan Li, Zicheng Liu, and Stan Z. Li. 2024. https://openreview.net/forum?id=RemfXx7ebP RD esign: Hierarchical data-efficient representation learning for tertiary structure-based RNA design . In The Twelfth International Confe...

    2024

  32. [40]

    Ignacio Tinoco Jr and Carlos Bustamante. 1999. How rna folds. Journal of molecular biology, 293(2):271--281

    1999

  33. [41]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30

    2017

  34. [42]

    Joseph L Watson, David Juergens, Nathaniel R Bennett, Brian L Trippe, Jason Yim, Helen E Eisenach, Woody Ahern, Andrew J Borst, Robert J Ragotte, Lukas F Milles, and 1 others. 2023. De novo design of protein structure and function with rfdiffusion. Nature, 620(7976):1089--1100

    2023

  35. [43]

    Xiufeng Yang, Kazuki Yoshizoe, Akito Taneda, and Koji Tsuda. 2017. Rna inverse folding using monte carlo tree search. BMC bioinformatics, 18:1--12

    2017

  36. [44]

    Trippe, Valentin De Bortoli, Emile Mathieu, Arnaud Doucet, Regina Barzilay, and Tommi Jaakkola

    Jason Yim, Brian L. Trippe, Valentin De Bortoli, Emile Mathieu, Arnaud Doucet, Regina Barzilay, and Tommi Jaakkola. 2023. SE (3) diffusion model with application to protein backbone generation. In Proceedings of the 40th International Conference on Machine Learning, volume 202...

    2023

  37. [45]

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2019. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675

    2019 arXiv

  38. [46]

    Zaixiang Zheng, Yifan Deng, Dongyu Xue, Yi Zhou, Fei Ye, and Quanquan Gu. 2023. Structure-informed language models are protein designers. In International conference on machine learning, pages 42317--42338. PMLR

    2023

  39. [47]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  40. [48]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.