Pith. sign in

REVIEW 5 major objections 10 minor 1 cited by

ProtInvTree: Deliberate Protein Inverse Folding with Reward-guided Tree Search

T0 review · 5 major / 10 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read ProtInvTree claims that protein inverse folding can be solved by deliberate, reward-guided tree search rather than one-pass prediction, and that doing so unlocks the latent capability of a frozen protein language model.

desk verdict First MCTS-for-inverse-folding paper, worth a serious look; scTM gains are partly circular, so the verdict rests on independent RMSD evidence that needs verification. read the letter →

arxiv 2506.00925 v1 pith:2DYGWRPI submitted 2025-06-01 q-bio.BM cs.CVcs.LG

Mengdi Liu , Xiaoxue Cheng , Zhangyang Gao , Hong Chang , Cheng Tan , Shiguang Shan , Xilin Chen This is my paper
classification q-bio.BMcs.CVcs.LG
keywords proteininversefoldingMonteCarlotreesearchtest-timescalingreward-guidedlanguagemodelstructuralconsistencysequencediversity
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

ProtInvTree claims that protein inverse folding—choosing an amino-acid sequence that folds into a given backbone—can be treated as a deliberate planning problem rather than a one-pass prediction. It wraps a frozen protein language model (ESM-3) in a Monte Carlo tree search that tries many partial sequences, scores them by how well the folded candidate matches the target, and backtracks toward better designs. The paper argues this is the first reward-guided tree-search framework for inverse folding and that it outperforms existing autoregressive, one-shot, and iterative-refinement methods on the CATH 4.2, CATH 4.3, and TS45 benchmarks. The practical stake is that a large pretrained model can be made substantially better at test time, without retraining, by spending more computation on search.

What carries the argument

The load-bearing object is a search tree whose nodes are partially denoised sequences conditioned on the target backbone. Four Monte Carlo tree search operations carry the argument: selection by a UCT score that balances exploitation against exploration, expansion by sampling candidate next sequences from the frozen protein language model, evaluation by fast jumpy denoising followed by an ESMFold-based TM-score reward, and backpropagation of that reward to ancestor nodes. The focus-and-grounding action mechanism splits each step into choosing which positions to modify and then generating the residue types at those positions, while jumpy denoising approximates a full rollout with a single DDIM-style jump so intermediate states can be scored cheaply. Together these components let the framework explore diverse trajectories while exploiting the highest-scoring candidates.

What would settle it

Design sequences for a set of CATH test backbones with ProtInvTree, synthesize them, and measure folding in vitro; if a substantial fraction of high-reward sequences fail to fold to the target structure, the ESMFold-based reward is not a valid proxy. A cheaper non-biological check is to rescore the same designed sequences with independent structure predictors and see whether the consistency advantage survives outside ESMFold.

Watch

Extended reading notes

Core claim

The central discovery is that branching, reward-guided search over partially generated sequences is enough to push a frozen ESM-3 model past state-of-the-art inverse-folding baselines. ProtInvTree reformulates iterative denoising as a tree-based Markov decision process; at each step a focus-and-grounding action picks which positions to rewrite and which residues to write, and a jumpy denoising step quickly estimates the structural reward of the resulting candidate using ESMFold-based similarity. Search, not training, is what unlocks the model: with zero trainable parameters, ProtInvTree raises sc-TM-score to 0.881 on CATH 4.2, 0.835 on CATH 4.3, and 0.724 on TS45, while also producing sequences more diverse and more distant from the native sequence than baselines at matched structural consistency. The authors interpret this as evidence that inverse folding is one-to-many and that explicit exploration of multiple trajectories finds structurally consistent solutions that the ground-truth-recovery objective misses.

Load-bearing premise

The whole pipeline assumes that ESMFold's predicted structures are a faithful stand-in for real folding, because both the search reward and the reported sc-TM-score come from ESMFold, and the paper reports no wet-lab validation.

Editorial extensions

If this is right

  • A frozen protein language model can be improved substantially at inference time: the same ESM-3 weights give sc-TM-scores roughly 8–18% higher under tree search than in direct iterative decoding on CATH 4.2.
  • Inverse folding can move beyond recovering the native sequence; the search returns diverse, novel sequences that still satisfy the structural reward.
  • Test-time computation acts as a scaling axis: increasing the number of candidate expansions or the planning depth raises structural consistency, up to a saturation point, with no retraining.
  • Search-based design is Pareto-optimal in the consistency–diversity and consistency–novelty planes compared with strong baselines.
  • Methods trained to maximize sequence recovery tend to cluster near the native sequence, while deliberate search reaches distinct local optima in sequence space.

Reading between the lines

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

  • Because the reward and the headline metric are both computed with ESMFold, the measured gains could partly reflect the search exploiting ESMFold's biases; an independent structure predictor or wet-lab folding assay would test whether the designed sequences are genuinely foldable.
  • The same focus-and-grounding plus jumpy-denoiser recipe could transfer to other structure-conditioned generation problems, such as RNA or small-molecule inverse design, where a fast surrogate reward is available.
  • If the reward model is swapped or ensembled across several structure predictors, the framework's diversity–consistency frontier might shift, suggesting that reward quality, not search alone, sets the upper bound.
  • The paper's own admitted lack of experimental validation leaves open whether high sc-TM-score translates to function in the lab.
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

5 major / 10 minor

Summary. ProtInvTree frames protein inverse folding as a tree-based Markov decision process and applies Monte Carlo tree search on top of a frozen ESM-3 policy. A two-stage focus-and-grounding action selects positions and generates residues, and a jumpy denoising rollout estimates node values via ESMFold-based TM-score rewards. The paper claims state-of-the-art sc-TMscore and RMSD on CATH 4.2, CATH 4.3, and TS45, and Pareto-optimal diversity-novelty trade-offs, with test-time scaling in search depth and breadth.

Significance. If the results were supported by independent evidence, the paper would demonstrate a useful direction: test-time search over a frozen protein language model can improve structural consistency and sequence diversity without retraining, addressing the one-to-many nature of inverse folding. The authors also deserve credit for explicitly acknowledging in Section 5.2 that sc-TMscore gains are expected because the same quantity is optimized by the reward, and for defining a clean MCTS formulation with reproducible algorithmic steps. However, the significance is conditional: the headline evaluations are computed with the same ESMFold predictor that is used inside the reward loop, no independent structure predictor is used, and no uncertainty estimates, code release, or wet-lab validation are provided. As presented, the empirical evidence does not establish that the designed sequences are genuinely more structurally consistent than baseline designs.

major comments (5)
  1. [§4.4 Eq. (11); Algorithm 1 line 12; Appendix B Eq. (12)] The reward inside the tree search and the evaluation metric are both computed with ESMFold, so the comparison in Tables 1-3 is partly circular. Eq. (11) defines the reward as TMScore(f(x̃_T), c) with f being ESMFold per Section 5.1, while the reported sc-TMscore in Appendix B Eq. (12) is TMScore(f(x̃), f(x)) with the same f, and Section 5.2 concedes the sc-TM gain is "expected" from reward optimization. The RMSD numbers are not an independent check either: Eq. (13) computes RMSD between f(x̃) and f(x) using the same ESMFold structures. Because the search actively selects sequences that satisfy ESMFold's structural biases while baselines do not receive this selection pressure, the tables do not demonstrate genuine improvement in foldability. The authors should re-evaluate all methods (or at least the final candidate sets) with a structure predictor not used in the reward loop, such as AlphaFold2/3, and report sc-TM and RMSD under that protocol.
  2. [§4.4 Eq. (11) vs. Algorithm 1 line 12] The reward definition is inconsistent between the main text and the algorithm box, and the difference is load-bearing. Eq. (11) defines R(s_t, a_t) = TMScore(f(x̃_T), c), comparing against the target backbone structure c, whereas Algorithm 1 line 12 computes r = TMScore(f(x̃_T), f(x_gt)), comparing against ESMFold's predicted structure of the ground-truth sequence. These are different reference objects, and the Algorithm 1 version makes the optimization objective essentially identical to the sc-TMscore metric of Eq. (12). The manuscript does not state which reward definition was used in the experiments, nor does it quantify how the choice affects the search behavior or the reported results.
  3. [Tables 1-3; §5.1; Figure 3] No measure of uncertainty is reported anywhere in the empirical section. Tables 1-3 present single point estimates without error bars, number of independent seeds, the number Z of designed sequences per protein (Algorithm 1 output), or the rule used to aggregate per-protein values into the displayed averages. Figure 3 describes curves as produced under "different sampling temperatures," but the temperature range and the number of sequences sampled per protein are not specified. Under these conditions, the RMSD advantage over KWDesign in Table 1 (1.513 vs. 1.566 on All) cannot be distinguished from run-to-run variation. The authors should report means and variances over repeated runs with a fully specified sampling protocol.
  4. [§5.4; Figure 5] The paper's central practical claim is a compute-quality trade-off via test-time scaling, but the compute side is never quantified. Section 5.4 only states that "the average time consumed also rises to some extent," with no wall-clock times, GPU-hours, or number of ESMFold calls per protein. Since each reward evaluation in Algorithm 1 (lines 10-13) requires at least one ESMFold structure prediction, the M×K search budget corresponds to a large, unquantified inference cost. Without cost figures, a reader cannot judge whether the sc-TM improvements in Figure 5 justify the added computation, nor compare ProtInvTree's efficiency with that of iterative-refinement baselines.
  5. [§4.3, §4.4; Appendix C, Table 4] The two advertised methodological components are not validated by adequate ablations. The focus operation is implemented, per Section 4.3, as "a random subset of sequence positions," and Appendix C (Table 4) shows that replacing random selection with autoregressive or entropy-based strategies changes sc-TMscore by at most 0.011 (0.870-0.881); the grounding mechanism is not ablated separately. The jumpy denoising estimator of Eq. (10) is used to evaluate every tree node, but no comparison against full-rollout evaluation is provided to show that the node values driving UCT selection (Eq. (3)) are faithful. Without these ablations, it is unclear which of the proposed components actually contributes to the reported results.
minor comments (10)
  1. [§4.1] The phrase "Markov decision process (MPD)" should read "MDP".
  2. [Figure 4] The quadrant labels contain typos: "Hign" should be "High" and "Noverty" should be "Novelty".
  3. [Figure 2 caption] The caption spells "Evalution" and should be "Evaluation".
  4. [§2.1, reference [50]] The method name "BridegIF" appears to be a misspelling of "BridgeIF".
  5. [§4.2, Evaluation paragraph] The text says node values are computed "As described in Equation 12," but Eq. (12) in Appendix B defines the sc-TMscore evaluation metric; the intended reference is presumably the reward definition in Eq. (11) together with the rollout procedure.
  6. [Appendix B, Diversity definition] The expression "2/N M(M−1)" contains an extraneous N in the denominator; after evaluating the double sum the prefactor should be 2/(M(M−1)).
  7. [§5.1, Implementation Details] The thresholds and schedules are incompletely specified: the reward threshold tau, the rollout count m for jumpy denoising, the cosine-schedule parameters for K_t, and the sampling temperature are not reported, although M=50 and w=0.01 are given.
  8. [Table 3] The model column writes "ProteinInvTree" in one row while the rest of the paper uses "ProtInvTree"; the naming should be consistent.
  9. [§1 and §2.2] The claims "the first work to apply a tree-search framework to structure-based protein sequence design" and "the first to extend the paradigm of test-time scaling to protein language models" are broader than what the paper establishes; the related-work discussion does not engage with existing MCTS-based generative design methods in protein or molecule design, which would be needed to support these priority claims.
  10. [Figure 6] The displayed values "r=0.99" and "r=0.98" are not defined in the caption; it should be stated whether these are jumpy-denoising rewards, final sc-TMscores, or some other quantity.

Circularity Check

1 steps flagged · score 7.0 of 10

The MCTS reward is the same ESMFold-based TM-score as the reported sc-TM metric, so the headline sc-TM improvements are expected by construction; RMSD and diversity provide the only independent support.

  1. self definitional [Section 4.4 Eq. (11); Algorithm 1 line 12; Appendix B Eq. (12); Section 5.2]
    "Eq. (11): R(st,at) = TMScore(f(˜xT), c). Algorithm 1 line 12: Compute reward: r = TMScore(f(˜xT), f(x_gt)). Appendix B Eq. (12): sc-TMScore = TMScore(f(˜x), f(x)). Section 5.2: "While the improvement in scTMscore is expected due to its use as the reward function during search...""

    The reward that guides MCTS selection and backpropagation is the same ESMFold-based TM-score (with reference f(x_gt) in Algorithm 1, or the target structure c in Eq. 11) that is reported as the sc-TMscore evaluation metric in Tables 1-3. The search therefore directly optimizes the metric used to claim structural consistency, so the observed sc-TMscore advantage over baselines is expected by construction rather than an independent confirmation. The paper itself concedes that the scTMscore gain is expected because it is the reward function. RMSD remains a less directly optimized structural metric, but the sc-TMscore result cannot be counted as independent evidence.

full rationale

The central circular step is explicit: the reward function in Section 4.4 (Eq. 11) and Algorithm 1 (line 12) is the same ESMFold-based TM-score as the sc-TMscore evaluation metric defined in Appendix B (Eq. 12). Thus the headline claim of superior structural consistency, as measured by sc-TMscore, is partly tautological: the method selects sequences that score well on the very metric used for benchmarking, while the baselines do not receive the same selection pressure. The paper honestly acknowledges this in Section 5.2, where it says the improvement in scTMscore is expected due to its use as the reward function. The RMSD results and diversity/novelty analyses provide independent content and are consistent with the method's effectiveness, which prevents the paper from being fully circular. However, because the abstract and central claim emphasize structural consistency and the primary metric for that claim is the optimized reward, the circularity is substantial. No load-bearing self-citation or uniqueness-import circularity was found; self-citations to KWDesign, AlphaDesign, and PiFold are baseline comparisons only. Overall score: 7.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical entities. The central claim rests on pretrained models (ESM-3, ESMFold), a set of manually chosen search hyperparameters, and self-consistency assumptions. No code is shipped, so many implementation details remain unobservable.

free parameters (7)
  • UCT exploration weight w = 0.01
    Manual hyperparameter balancing exploitation and exploration in Eq. 3; no sensitivity analysis across values is reported.
  • Max MCTS iterations M = 50
    Termination condition in Section 4.2 and Algorithm 1; default chosen without systematic study of interaction with other hyperparameters.
  • Expansion number K = not explicitly stated
    Number of child candidates per node; Figure 5 varies it from 2 to 6 but the default used in Tables 1-3 is not specified.
  • Planning depth T = not explicitly stated
    Maximum tree depth; Figure 5 varies depth but the default in main results is not given.
  • Cosine schedule for stepwise selection numbers Kt = not specified
    Section 4.3 says Kt follows a cosine schedule but gives no formula or explicit values.
  • Reward threshold tau = not specified
    Termination and acceptance criterion in Algorithm 1; value never reported.
  • Rollout count m for jumpy denoising evaluation = not specified
    Section 4.4 says samples m fully generated sequences for node value but m is never stated.
assumptions (4)
  • domain assumption ESM-3 provides a usable inverse-folding policy without task-specific fine-tuning.
    Section 5.1 chooses ESM-3 as policy model because it directly supports inverse folding; the search quality depends on this.
  • domain assumption ESMFold structure predictions are accurate enough that TM-score against ESMFold measures structural consistency.
    Eq. 11 and Eq. 12 use f(xtilde) and f(x) from ESMFold; the entire reward and evaluation rest on this.
  • domain assumption Self-consistent TM-score computed with the same folding model cancels the model's inductive bias.
    Appendix B argues this cancellation, but the same predictor is used as reward and metric, which makes the cancellation argument circular.
  • domain assumption A shallow MCTS with UCT and limited rollouts finds globally optimal sequences rather than locally optimal ones.
    Section 4.2 and the analysis in Section 5.4 assume search depth and width improve quality, with no guarantee of global optimality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ProtInvTree: Deliberate Protein Inverse Folding with Reward-guided Tree Search." pith.science (2026). https://pith.science/paper/2DYGWRPI

@misc{pith2026250600925,
  author       = {Pith},
  title        = {Pith review of: ProtInvTree: Deliberate Protein Inverse Folding with Reward-guided Tree Search},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2DYGWRPI}},
  note         = {Machine review of arXiv:2506.00925}
}
read the original abstract

Designing protein sequences that fold into a target 3D structure, known as protein inverse folding, is a fundamental challenge in protein engineering. While recent deep learning methods have achieved impressive performance by recovering native sequences, they often overlook the one-to-many nature of the problem: multiple diverse sequences can fold into the same structure. This motivates the need for a generative model capable of designing diverse sequences while preserving structural consistency. To address this trade-off, we introduce ProtInvTree, the first reward-guided tree-search framework for protein inverse folding. ProtInvTree reformulates sequence generation as a deliberate, step-wise decision-making process, enabling the exploration of multiple design paths and exploitation of promising candidates through self-evaluation, lookahead, and backtracking. We propose a two-stage focus-and-grounding action mechanism that decouples position selection and residue generation. To efficiently evaluate intermediate states, we introduce a jumpy denoising strategy that avoids full rollouts. Built upon pretrained protein language models, ProtInvTree supports flexible test-time scaling by expanding the search depth and breadth without retraining. Empirically, ProtInvTree outperforms state-of-the-art baselines across multiple benchmarks, generating structurally consistent yet diverse sequences, including those far from the native ground truth.

Figures

Figures reproduced from arXiv: 2506.00925 by the authors.

Figure 1
Figure 1. (A) Schematic illustration on various approaches of structure-based protein design. Each node denotes an intermediate sequence conditioned on the target structure, progressively expanding toward full generation. See the complete framework and details in Figures 2. (B) Landscape of scTMscore across the sequence space with respect to the target structure. The ground truth sequence is marked, and candidate sequences ar… view at source ↗
Figure 2
Figure 2. The framework of ProtInvTree. (a) The four steps of reward-guided tree search— [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Pareto comparison of structural consistency (sc-TMScore) against diversity (left) and [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Conceptual illustration of the difference between diversity and novelty of the generated [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Test-time scaling laws analysis of our ProtInvTree under [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 7
Figure 7. Figure 7: Structural align￾ment visualization. Case Study. To facilitate understanding of the entire workflow of our proposed ProtInvTree, we visualize a reward-guided search tree in [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. HD-Prot: A Protein Language Model for Joint Sequence-Structure Modeling with Continuous Structure Tokens

    cs.CE 2025-12 conditional novelty 6.0 of 10

    HD-Prot shows that a protein language model can jointly generate sequences and structures using continuous structure tokens instead of quantized tokens, reaching competitive performance on four protein design tasks.

Reference graph

Works this paper leans on

50 extracted references · 26 canonical work pages · cited by 1 Pith paper

  1. [1]

    Accurate structure prediction of biomolecular interactions with alphafold 3.Nature, pages 1–3, 2024

    Josh Abramson, Jonas Adler, Jack Dunger, Richard Evans, Tim Green, Alexander Pritzel, Olaf Ronneberger, Lindsay Willmore, Andrew J Ballard, Joshua Bambrick, et al. Accurate structure prediction of biomolecular interactions with alphafold 3.Nature, pages 1–3, 2024

    2024

  2. [2]

    Accurate prediction of protein structures and interactions using a three-track neural network.Science, 373(6557):871–876, 2021

    Minkyung Baek, Frank DiMaio, Ivan Anishchenko, Justas Dauparas, Sergey Ovchinnikov, Gyu Rie Lee, Jue Wang, Qian Cong, Lisa N Kinch, R Dustin Schaeffer, et al. Accurate prediction of protein structures and interactions using a three-track neural network.Science, 373(6557):871–876, 2021

    2021

  3. [3]

    Large language monkeys: Scaling inference compute with repeated sampling.arXiv preprint arXiv:2407.21787, 2024

    Bradley Brown, Jordan Juravsky, Ryan Ehrlich, Ronald Clark, Quoc V Le, Christopher Ré, and Azalia Mirhoseini. Large language monkeys: Scaling inference compute with repeated sampling.arXiv preprint arXiv:2407.21787, 2024

    arXiv 2024

  4. [4]

    Protein design: a hierarchic approach.Science, 270(5238):935– 941, 1995

    James W Bryson, Stephen F Betz, Helen S Lu, Daniel J Suich, Hongxing X Zhou, Karyn T O’Neil, and William F DeGrado. Protein design: a hierarchic approach.Science, 270(5238):935– 941, 1995

    work page 1995

  5. [5]

    How root-mean-square distance (rmsd) values depend on the resolution of protein structures that are compared.Applied Crystallography, 36(1):125–128, 2003

    Oliviero Carugo. How root-mean-square distance (rmsd) values depend on the resolution of protein structures that are compared.Applied Crystallography, 36(1):125–128, 2003

    work page 2003

  6. [6]

    Sparks of function by de novo protein design

    Alexander E Chu, Tianyu Lu, and Po-Ssu Huang. Sparks of function by de novo protein design. Nature Biotechnology, 42(2):203–215, 2024

    work page 2024

  7. [7]

    Efficient selectivity and backup operators in monte-carlo tree search

    Rémi Coulom. Efficient selectivity and backup operators in monte-carlo tree search. In International conference on computers and games, pages 72–83. Springer, 2006

    2006

  8. [8]

    Robust deep learning–based protein sequence design using proteinmpnn.Science, 378(6615):49–56, 2022

    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, et al. Robust deep learning–based protein sequence design using proteinmpnn.Science, 378(6615):49–56, 2022

    2022

Show all 50 references
  1. [9]

    Zhangyang Gao, Cheng Tan, Xingran Chen, Yijie Zhang, Jun Xia, Siyuan Li, and Stan Z. Li. KW-design: Pushing the limit of protein design via knowledge refinement. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  2. [10]

    Alphadesign: A graph protein design method and benchmark on alphafolddb.arXiv preprint arXiv:2202.01079, 2022

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

    2022 arXiv

  3. [11]

    Zhangyang Gao, Cheng Tan, and Stan Z. Li. Pifold: Toward effective and efficient protein inverse folding. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  4. [12]

    Zhangyang Gao, Cheng Tan, Yijie Zhang, Xingran Chen, Lirong Wu, and Stan Z. Li. Pro- teininvbench: Benchmarking protein inverse folding on diverse tasks, models, and metrics. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023

    2023

  5. [13]

    Uniif: Unified molecule inverse folding.Advances in Neural Information Processing Systems, 37:135843–135860, 2024

    Zhangyang Gao, Jue Wang, Cheng Tan, Lirong Wu, Yufei Huang, Siyuan Li, Zhirui Ye, and Stan Z Li. Uniif: Unified molecule inverse folding.Advances in Neural Information Processing Systems, 37:135843–135860, 2024

    2024

  6. [14]

    Protein remote homology detection and structural alignment using deep learning

    Tymor Hamamsy, James T Morton, Robert Blackwell, Daniel Berenberg, Nicholas Carriero, Vladimir Gligorijevic, Charlie EM Strauss, Julia Koehler Leman, Kyunghyun Cho, and Richard Bonneau. Protein remote homology detection and structural alignment using deep learning. Nature biot...

    2023

  7. [15]

    Reasoning with language model is planning with world model.arXiv preprint arXiv:2305.14992, 2023

    Shibo Hao, Yi Gu, Haodi Ma, Joshua Jiahua Hong, Zhen Wang, Daisy Zhe Wang, and Zhit- ing Hu. Reasoning with language model is planning with world model.arXiv preprint arXiv:2305.14992, 2023

    2023 arXiv

  8. [16]

    Simulating 500 million years of evolution with a language model.Science, page eads0018, 2025

    Thomas Hayes, Roshan Rao, Halil Akin, Nicholas J Sofroniew, Deniz Oktay, Zeming Lin, Robert Verkuil, Vincent Q Tran, Jonathan Deaton, Marius Wiggert, et al. Simulating 500 million years of evolution with a language model.Science, page eads0018, 2025. 10

    2025

  9. [17]

    Learning inverse folding from millions of predicted structures

    Chloe Hsu, Robert Verkuil, Jason Liu, Zeming Lin, Brian Hie, Tom Sercu, Adam Lerer, and Alexander Rives. Learning inverse folding from millions of predicted structures. InInternational conference on machine learning, pages 8946–8970. PMLR, 2022

    2022

  10. [18]

    The coming of age of de novo protein design

    Po-Ssu Huang, Scott E Boyken, and David Baker. The coming of age of de novo protein design. Nature, 537(7620):320–327, 2016

    2016

  11. [19]

    Generative models for graph-based protein design.Advances in neural information processing systems, 32, 2019

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

    2019

  12. [20]

    Learning from protein structure with geometric vector perceptrons

    Bowen Jing, Stephan Eismann, Patricia Suriana, Raphael John Lamarre Townshend, and Ron Dror. Learning from protein structure with geometric vector perceptrons. InInternational Conference on Learning Representations

  13. [21]

    Learning from protein structure with geometric vector perceptrons

    Bowen Jing, Stephan Eismann, Patricia Suriana, Raphael John Lamarre Townshend, and Ron Dror. Learning from protein structure with geometric vector perceptrons. InInternational Conference on Learning Representations, 2021

    2021

  14. [22]

    Representativeness revisited: Attribute substitution in intuitive judgment.Heuristics and biases: The psychology of intuitive judgment, 49(49-81):74, 2002

    Daniel Kahneman, Shane Frederick, et al. Representativeness revisited: Attribute substitution in intuitive judgment.Heuristics and biases: The psychology of intuitive judgment, 49(49-81):74, 2002

    2002

  15. [23]

    Mindstar: Enhancing math reasoning in pre-trained llms at inference time.arXiv preprint arXiv:2405.16265, 2024

    Jikun Kang, Xin Zhe Li, Xi Chen, Amirreza Kazemi, and Boxing Chen. Mindstar: Enhancing math reasoning in pre-trained llms at inference time.arXiv preprint arXiv:2405.16265, 2024

    2024 arXiv

  16. [24]

    A new age in protein design empowered by deep learning.Cell Systems, 14(11):925–939, 2023

    Hamed Khakzad, Ilia Igashov, Arne Schneuing, Casper Goverde, Michael Bronstein, and Bruno Correia. A new age in protein design empowered by deep learning.Cell Systems, 14(11):925–939, 2023

    2023

  17. [25]

    Bandit based monte-carlo planning

    Levente Kocsis and Csaba Szepesvári. Bandit based monte-carlo planning. In Johannes Fürnkranz, Tobias Scheffer, and Myra Spiliopoulou, editors,Machine Learning: ECML 2006, 17th European Conference on Machine Learning, Berlin, Germany, September 18-22, 2006, Proceedings, volume...

    2006

  18. [26]

    Language models of protein sequences at the scale of evolution enable accurate structure prediction.BioRxiv, 2022:500902, 2022

    Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Allan dos Santos Costa, Maryam Fazel-Zarandi, Tom Sercu, Sal Candido, et al. Language models of protein sequences at the scale of evolution enable accurate structure prediction.BioRxiv, 2022:500902, 2022

    2022

  19. [27]

    De novo protein design using geometric vector field networks

    Weian Mao, Muzhi Zhu, Zheng Sun, Shuaike Shen, Lin Yuanbo Wu, Hao Chen, and Chunhua Shen. De novo protein design using geometric vector field networks. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  20. [28]

    Increasing sequence diversity with flexible backbone protein design: the complete redesign of a protein hydrophobic core.Structure, 20(6):1086–1096, 2012

    Grant S Murphy, Jeffrey L Mills, Michael J Miley, Mischa Machius, Thomas Szyperski, and Brian Kuhlman. Increasing sequence diversity with flexible backbone protein design: the complete redesign of a protein hydrophobic core.Structure, 20(6):1086–1096, 2012

    2012

  21. [29]

    Cath–a hierarchic classification of protein domain structures.Structure, 5(8):1093–1109, 1997

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

    1997

  22. [30]

    Mutual reasoning makes smaller llms stronger problem-solvers.arXiv preprint arXiv:2408.06195, 2024

    Zhenting Qi, Mingyuan Ma, Jiahang Xu, Li Lyna Zhang, Fan Yang, and Mao Yang. Mutual reasoning makes smaller llms stronger problem-solvers.arXiv preprint arXiv:2408.06195, 2024

    2024 arXiv

  23. [31]

    Fast uncovering of protein sequence diversity from structure

    LUCA Silva, Barthelemy Meynard-Piganeau, Carlo Lucibello, Christoph Feinauer, et al. Fast uncovering of protein sequence diversity from structure. InThe Thirteenth International Conference on Learning Representations. (seleziona...), 2025

    2025

  24. [32]

    Mas- tering the game of go with deep neural networks and tree search.nature, 529(7587):484–489, 2016

    David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driess- che, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mas- tering the game of go with deep neural networks and tree search.nature, 529(7587):484–489...

    2016

  25. [33]

    The empirical case for two systems of reasoning.Psychological bulletin, 119(1):3, 1996

    Steven A Sloman. The empirical case for two systems of reasoning.Psychological bulletin, 119(1):3, 1996

    1996

  26. [34]

    Scaling llm test-time compute opti- mally can be more effective than scaling model parameters.arXiv preprint arXiv:2408.03314, 2024

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. Scaling llm test-time compute opti- mally can be more effective than scaling model parameters.arXiv preprint arXiv:2408.03314, 2024

    2024 arXiv

  27. [35]

    Psychology Press, 1999

    Keith E Stanovich.Who is rational?: Studies of individual differences in reasoning. Psychology Press, 1999

    1999

  28. [36]

    Generative de novo protein design with global context.arXiv preprint arXiv:2204.10673, 2022

    Cheng Tan, Zhangyang Gao, Jun Xia, Bozhen Hu, and Stan Z Li. Generative de novo protein design with global context.arXiv preprint arXiv:2204.10673, 2022

    2022 arXiv

  29. [37]

    Global-context aware generative protein design

    Cheng Tan, Zhangyang Gao, Jun Xia, Bozhen Hu, and Stan Z Li. Global-context aware generative protein design. InICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2023

    2023

  30. [38]

    Q*: Improving multi-step reasoning for llms with deliberative planning, 2024

    Chaojie Wang, Yanchen Deng, Zhiyi Lv, Shuicheng Yan, and An Bo. Q*: Improving multi-step reasoning for llms with deliberative planning, 2024

    2024

  31. [39]

    Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  32. [40]

    High-resolution de novo structure prediction from primary sequence.BioRxiv, pages 2022–07, 2022

    Ruidong Wu, Fan Ding, Rui Wang, Rui Shen, Xiwen Zhang, Shitong Luo, Chenpeng Su, Zuofan Wu, Qi Xie, Bonnie Berger, et al. High-resolution de novo structure prediction from primary sequence.BioRxiv, pages 2022–07, 2022

    2022

  33. [41]

    An empirical analysis of compute-optimal inference for problem-solving with language models.arXiv preprint arXiv:2408.00724, 2024

    Yangzhen Wu, Zhiqing Sun, Shanda Li, Sean Welleck, and Yiming Yang. An empirical analysis of compute-optimal inference for problem-solving with language models.arXiv preprint arXiv:2408.00724, 2024

    2024 arXiv

  34. [42]

    Fast non-autoregressive inverse folding with discrete diffusion.arXiv preprint arXiv:2312.02447, 2023

    John J Yang, Jason Yim, Regina Barzilay, and Tommi Jaakkola. Fast non-autoregressive inverse folding with discrete diffusion.arXiv preprint arXiv:2312.02447, 2023

    2023 arXiv

  35. [43]

    Tree of thoughts: Deliberate problem solving with large language models.Ad- vances in Neural Information Processing Systems, 36, 2024

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models.Ad- vances in Neural Information Processing Systems, 36, 2024

    2024

  36. [44]

    Graph denoising diffusion for inverse protein folding

    Kai Yi, Bingxin Zhou, Yiqing Shen, Pietro Lio, and Yu Guang Wang. Graph denoising diffusion for inverse protein folding. InThirty-seventh Conference on Neural Information Processing Systems, 2023

    2023

  37. [45]

    Inverse protein folding problem: designing polymer sequences

    Kaizhi Yue and Ken A Dill. Inverse protein folding problem: designing polymer sequences. Proceedings of the National Academy of Sciences, 89(9):4163–4167, 1992

    1992

  38. [46]

    Accessing gpt-4 level mathematical olympiad solutions via monte carlo tree self-refine with llama-3 8b.arXiv preprint arXiv:2406.07394, 2024

    Di Zhang, Jiatong Li, Xiaoshui Huang, Dongzhan Zhou, Yuqiang Li, and Wanli Ouyang. Accessing gpt-4 level mathematical olympiad solutions via monte carlo tree self-refine with llama-3 8b.arXiv preprint arXiv:2406.07394, 2024

    2024 arXiv

  39. [47]

    Tm-align: a protein structure alignment algorithm based on the tm-score.Nucleic acids research, 33(7):2302–2309, 2005

    Yang Zhang and Jeffrey Skolnick. Tm-align: a protein structure alignment algorithm based on the tm-score.Nucleic acids research, 33(7):2302–2309, 2005

    2005

  40. [48]

    Structure- informed language models are protein designers

    Zaixiang Zheng, Yifan Deng, Dongyu Xue, Yi Zhou, Fei Ye, and Quanquan Gu. Structure- informed language models are protein designers. InInternational Conference on Machine Learning, pages 42317–42338. PMLR, 2023

    2023

  41. [49]

    Generative ai for controllable protein sequence design: A survey

    Yiheng Zhu, Zitai Kong, Jialu Wu, Weize Liu, Yuqiang Han, Mingze Yin, Hongxia Xu, Chang- Yu Hsieh, and Tingjun Hou. Generative ai for controllable protein sequence design: A survey. arXiv preprint arXiv:2402.10516, 2024

    2024 arXiv

  42. [50]

    Bridge-if: Learning inverse protein folding with markov bridges

    Yiheng Zhu, Jialu Wu, Qiuyi Li, Jiahuan Yan, Mingze Yin, Wei Wu, Mingyang Li, Jieping Ye, Zheng Wang, and Jian Wu. Bridge-if: Learning inverse protein folding with markov bridges. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 12 A Algorithms ...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.