Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Improving Protein Sequence Design through Designability Preference Optimization

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Fine-tuning a sequence model with residue-level designability preferences triples its in silico success rate on enzyme scaffolds.

desk verdict A genuinely new residue-level DPO variant with a clean ablation, but the headline gains are in silico and partly self-confirming because pLDDT is both reward and metric; still worth refereeing. read the letter →

arxiv 2506.00297 v1 pith:GG4G6HLC submitted 2025-05-30 cs.LG cs.AIq-bio.BM

classification cs.LGcs.AIq-bio.BM
keywords proteinsequencedesigndesignabilitydirectpreferenceoptimizationResiDPOpLDDTinversefoldingenzymeLigandMPNN
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

Protein sequence design models are typically trained to recover native sequences, but the design problem that matters is whether a generated sequence folds into the intended backbone. This paper argues that this gap can be closed by treating designability as a preference optimization problem, using AlphaFold2's per-residue pLDDT confidence as the reward signal. The authors propose ResiDPO, which decomposes the DPO loss into a residue-level preference term for positions where the preferred sequence improves local confidence, and a residue-level constraint term that holds the model to the reference distribution at already-confident positions. Fine-tuning LigandMPNN this way yields EnhancedMPNN, which raises the in silico design success rate on a five-enzyme benchmark from 6.56% to 17.57% and more than doubles it on a binder benchmark. If valid, this indicates that designability itself, not sequence recovery, can be directly optimized, potentially cutting the computational cost of de novo enzyme and binder design.

What carries the argument

The central object is the ResiDPO loss, an adaptation of Direct Preference Optimization that exploits the fixed length of protein sequences to assign rewards per residue. For each backbone, candidate sequences are scored by AlphaFold2's per-residue pLDDT; the loss then splits into Residue-level Preference Learning (RPL), which acts on positions where the preferred sequence exceeds the dispreferred one by a margin $\alpha$, and Residue-level Constraint Learning (RCL), which applies a KL divergence penalty at positions where the preferred sequence has pLDDT above $\beta$ and the reference model assigns probability above $\gamma$. This decomposition carries the argument: it avoids the conflicting gradients of sequence-level DPO and lets the model improve weak local regions while preserving already-reliable ones.

What would settle it

Re-run the enzyme and binder benchmarks with a success oracle that the model never saw during training, such as a completely different structure predictor or wet-lab folding and activity measurements of the designed sequences, and check whether EnhancedMPNN retains its roughly three-fold advantage over LigandMPNN; if the advantage disappears or inverts, the reported designability gain is an artifact of optimizing the AlphaFold2 confidence metric.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that residue-level preference optimization can substantially improve the designability of protein sequences without sacrificing the model's existing strengths. The central result is the ResiDPO objective: for a preferred sequence $y_w$ and a dispreferred sequence $y_l$ generated for the same backbone, the loss applies a preference term only at the residue set $I = \{i : \mathrm{pLDDT}(y_w,i) - \mathrm{pLDDT}(y_l,i) > \alpha\}$, and a KL-constraint term at the residue set $J$ where the preferred sequence already has high pLDDT and the reference model is confident. This decoupling lets the model raise the probability of residues that improve predicted local structure while keeping confident regions close to the reference distribution. Using this objective to fine-tune LigandMPNN produces EnhancedMPNN, which achieves an average sequence success rate of 17.57% on the five-enzyme RFdiffusion2 benchmark under the criteria pLDDT > 80 and C$_{\alpha}$ RMSD < 1.5 Å, compared with 6.56% for LigandMPNN; the binder benchmark success rate rises from 7.07% to 16.07%.

Load-bearing premise

That AlphaFold2's pLDDT score is a valid stand-in for whether a designed sequence will actually fold to its target shape in the lab, since pLDDT is used as both the training reward and the main evaluation criterion.

Editorial extensions

If this is right

  • EnhancedMPNN raises the average sequence design success rate on the five-enzyme benchmark from 6.56% to 17.57%, so roughly 2.7 times fewer designed sequences are needed to obtain one predicted to fold to the target.
  • The backbone success rate rises from 19.74% to 40.34%, meaning many backbones that LigandMPNN could not design successfully become designable with EnhancedMPNN.
  • On the binder benchmark, success rises from 7.07% to 16.07%, showing the preference signal transfers to multichain interfaces despite training on monomers.
  • ResiDPO reaches comparable validation pLDDT accuracy with roughly 1,000 training samples that DPO needs 19,000 samples to match, indicating the residue-level supervision is substantially more data-efficient.

Reading between the lines

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

  • Editorial inference: because the paper optimizes and evaluates with the same AlphaFold2 pLDDT oracle, the reported gain may be partly an artifact of overfitting to that oracle; re-scoring with an independent structure predictor would test this.
  • Editorial inference: the residue-level decoupling should transfer to other per-residue feedback signals such as predicted binding energy or experimental stability measurements, but the paper does not demonstrate this.
  • Editorial inference: the observed compositional shift toward charged residues and away from alanines suggests the model is resolving sequence-structure ambiguity; a testable extension would be checking whether these designs express and remain soluble in vitro.
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

3 major / 4 minor

Summary. The paper proposes ResiDPO, a residue-level variant of Direct Preference Optimization (DPO) for protein sequence design. The method uses AlphaFold2 per-residue pLDDT scores as a reward signal, decoupling the DPO loss into residue-level preference learning (RPL) and residue-level constraint learning (RCL). The authors fine-tune LigandMPNN to obtain EnhancedMPNN and report a nearly 3-fold improvement in in silico design success rate on an RFdiffusion2 enzyme benchmark (6.56% to 17.57%) and a roughly 2.3-fold improvement on a binder benchmark (7.07% to 16.07%), using success criteria based on AF2-predicted pLDDT, Cα RMSD, and PAE. The paper also introduces a curated dataset (PDB-D), presents ablations of sampling strategies and hyperparameters, and analyzes residue substitution trends. All benchmark results are consistently described as in silico, and the paper explicitly defers experimental validation to future work.

Significance. If the reported gains hold under independent evaluation, ResiDPO would be a practically useful and conceptually clean way to align inverse-folding models with a structure-confidence signal, while the architecture-agnostic formulation and the data-efficiency analysis (Fig. 5) add value beyond the specific benchmark numbers. The paper also makes a clear, falsifiable prediction: ResiDPO-designed sequences should show higher success under an independent structure predictor or in experimental folding assays. The main weakness is that the evaluation metric (AF2 pLDDT) is the same signal used as the training reward, so the central designability claim is not yet independently established. The authors are appropriately cautious in labeling their results as in silico, which is a strength.

major comments (3)
  1. [§4.3.1, §3.1, §3.3] The main evaluation criterion (pLDDT > 80 and Cα RMSD < 1.5 Å) is computed with AlphaFold2, whose per-residue pLDDT is also the reward signal used to construct preference pairs (§3.2) and to define the RPL and RCL losses (Eqs. 2–5). The reported 17.57% vs. 6.56% gain therefore conflates improved designability with improved alignment to this particular AF2 checkpoint. The supporting correlation in Fig. 6 is computed on ResiDPO-designed sequences, i.e., on the same distribution the model was trained to optimize, so it does not break the circularity. Section 5 explicitly defers experimental characterization. I request at least one evaluation with an independent oracle not used during training (for example, ESMFold, AlphaFold3, or a different AF2 checkpoint) and, ideally, experimental expression/folding data, to support the title-level claim of improved designability.
  2. [§4.1, Table 1, Figs. 3 and 5] The 'pLDDT Accuracy' metric is the basis for the main ablations and the data-efficiency analysis, but it is only described as 'the correlation between the model's output likelihood for a validation sequence and the sequence's actual pLDDT score.' No formula is given, and no correlation type (Pearson, Spearman, Kendall), aggregation level (per-residue or per-sequence), or normalization is specified. Without a precise definition, the reported percentages (e.g., 62.11% vs. 66.08%) cannot be interpreted, compared, or reproduced. Please define the metric formally, state its range, and report confidence intervals or error bars for the table and figure values.
  3. [§4.3] The headline success rates are point estimates without confidence intervals, despite the benchmark comprising 5 enzymes × 1,000 backbones × 8 sequences and 5 targets × 100 backbones × 8 sequences. Appendix A.2 reports t-tests on pLDDT distributions, but not on the sequence- or backbone-level success rates that form the central claim. Please provide bootstrap or exact binomial confidence intervals for the success rates, and per-enzyme and per-target breakdowns, which are currently only summarized verbally ('consistent across all five enzymes').
minor comments (4)
  1. [§3.3.1] When I=∅, the text says the method 'effectively' applies standard DPO, but Eq. (2) with I={1..L} lacks the π_ref terms that appear in Eq. (1). It is a reference-free sequence-level preference loss, not standard DPO. Either add the reference-model terms in this fallback case or rephrase the claim.
  2. [§4.1] The paper states that the PDB-D dataset 'will be released' but gives no repository link, release timeline, or mention of code/model weights. For a methods paper, a concrete release plan for code, data, and EnhancedMPNN weights would greatly aid reproducibility.
  3. [Eqs. (2) and (4)] The notation y_i^w|x and y_i^l|x would be clearer if the conditioning on the backbone x were written explicitly in the equations themselves, since the current notation could be misread as unconditional residue probabilities.
  4. [§4.5] The interpretation that EnhancedMPNN 'reduces the ambiguity in the sequence-structure mapping' is presented as an explanation of the observed AF2 confidence shift, but no direct measure of sequence-structure ambiguity is provided. Consider labeling this as a hypothesis rather than a conclusion.

Circularity Check

1 steps flagged · score 6.0 of 10

AF2 pLDDT is both the training reward and the main success threshold, so the reported 3x 'designability' gain is partly by construction.

  1. self definitional [Section 3.1 (pLDDT as preference signal), Section 4.3.1 (pLDDT>80 success criterion), Appendix A.1 (pLDDT adopted as designability proxy)]
    "For a quantitative measure of designability, we utilize the predicted Local Distance Difference Test (pLDDT) score derived from AlphaFold2 (AF2). The pLDDT score serves as a proxy for folding accuracy and stability ... a sequence yw with a higher pLDDT score is considered preferable ... We assessed design success based on the criteria of pLDDT > 80 and Cα RMSD < 1.5 Å."

    The quantity being optimized is the same AF2 pLDDT score used as the main evaluation threshold. ResiDPO's residue masks I and J in Eqs. (2)-(5) are defined by per-residue pLDDT differences and pLDDT>80, and the final benchmark success criterion is pLDDT>80. Therefore the pLDDT component of the reported success-rate improvement is the training objective itself evaluated on held-out backbones: optimizing pLDDT raises pLDDT by construction. The additional RMSD threshold comes from the same AF2 prediction and is correlated with pLDDT; the paper's supporting correlation (Fig. 6) is computed on ResiDPO-designed sequences, i.e., the optimized distribution.

full rationale

ResiDPO is a genuine algorithmic contribution: the residue-level DPO variant is well defined, and the comparison against LigandMPNN and DPO on the RFdiffusion2 benchmark backbones is an honest empirical measurement. However, the central quantity called designability is operationalized through AF2 pLDDT. Section 3.1 defines the preference signal as pLDDT; Eqs. (2)-(5) build the loss entirely from per-residue pLDDT, with the RCL mask J using the same threshold (pLDDT>80) that later defines success; Section 4.3.1 scores 'design success' as pLDDT>80 and Cα RMSD<1.5 Å. A model trained to maximize pLDDT will, by construction, increase the pLDDT component of the evaluation metric. The RMSD criterion adds independent content, but it is computed from the same AF2 prediction and is correlated with pLDDT; Fig. 6 shows that correlation on ResiDPO-designed sequences, i.e., on the distribution being optimized. Section 5 explicitly defers experimental characterization ('Experimental characterization of EnhancedMPNN-generated sequences will reveal whether the enhanced in silico foldability translates to greater control over structure and higher design success rates'), so no external anchor is provided. This is partial circularity, not fraud: the held-out benchmark and baseline comparisons are real, and the gain in pLDDT is a legitimate generalization result; but as evidence of improved physical designability, the claim reduces largely to 'optimizing the proxy raises the proxy.' No load-bearing self-citation chain was found.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the pLDDT proxy, the representativeness of RFdiffusion backbones, and the base-model coverage of the sequence space; these are domain assumptions rather than mathematical axioms. Hyperparameters were selected by validation-set pLDDT accuracy, which is itself derived from the same AF2 signal used in evaluation.

free parameters (6)
  • alpha (residue pLDDT margin for RPL) = 10
    Residues are included in preference learning only when pLDDT(y_w,i) - pLDDT(y_l,i) > alpha; value selected by ablation on validation pLDDT accuracy.
  • beta (pLDDT threshold for RCL) = 80
    Residues are preserved when pLDDT(y_w,j) > beta; chosen based on the heuristic that pLDDT > 80 indicates sufficient designability, and described as correlated with gamma.
  • gamma (reference model confidence threshold) = 0.5
    Residues are preserved when the reference model's probability exceeds gamma; tuned by search.
  • lambda (RCL loss weight) = 0.01
    Balances the RPL preference loss and RCL constraint loss; selected by ablation.
  • delta (pLDDT difference threshold for preference pairs) = 10
    Only pairs with pLDDT difference greater than delta are used; chosen over 30 for a larger and more diverse training set.
  • DPO beta (KL regularization coefficient in Eq. 1) = not stated
    Standard DPO KL strength; set via grid search, but the exact value is not reported in the text.
assumptions (3)
  • domain assumption AlphaFold2 pLDDT is a valid proxy for protein designability and correlates with experimental design success.
    Invoked in Section 3.1 ('pLDDT score serves as a proxy for folding accuracy and stability') and used both as training reward and evaluation success criterion. If this correlation is weak, the reported improvements may not reflect actual foldability.
  • domain assumption RFDiffusion and RFdiffusion2 generated backbones are representative targets for evaluating designability.
    The enzyme and binder benchmarks use these generated backbones (Section 4.3, Appendix A.5), and success is judged against them. If the backbones are not representative of real design targets, the benchmark numbers may not generalize.
  • domain assumption LigandMPNN as reference model provides a base distribution with sufficient coverage of candidate sequences.
    Preference pairs are sampled with LigandMPNN at temperature 1.0 (Section 4.1). If the base model never proposes high-pLDDT sequences, DPO cannot learn to favor them; the paper notes the scarcity of high-pLDDT generations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Protein Sequence Design through Designability Preference Optimization." pith.science (2026). https://pith.science/paper/GG4G6HLC

@misc{pith2026250600297,
  author       = {Pith},
  title        = {Pith review of: Improving Protein Sequence Design through Designability Preference Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GG4G6HLC}},
  note         = {Machine review of arXiv:2506.00297}
}
read the original abstract

Protein sequence design methods have demonstrated strong performance in sequence generation for de novo protein design. However, as the training objective was sequence recovery, it does not guarantee designability--the likelihood that a designed sequence folds into the desired structure. To bridge this gap, we redefine the training objective by steering sequence generation toward high designability. To do this, we integrate Direct Preference Optimization (DPO), using AlphaFold pLDDT scores as the preference signal, which significantly improves the in silico design success rate. To further refine sequence generation at a finer, residue-level granularity, we introduce Residue-level Designability Preference Optimization (ResiDPO), which applies residue-level structural rewards and decouples optimization across residues. This enables direct improvement in designability while preserving regions that already perform well. Using a curated dataset with residue-level annotations, we fine-tune LigandMPNN with ResiDPO to obtain EnhancedMPNN, which achieves a nearly 3-fold increase in in silico design success rate (from 6.56% to 17.57%) on a challenging enzyme design benchmark.

Figures

Figures reproduced from arXiv: 2506.00297 by the authors.

Figure 1
Figure 1. ResiDPO enables precise protein sequence optimization by enabling residue-level des [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Design success rates for LigandMPNN, DPO-finetuned LigandMPNN, and ResiDPO [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Ablations on the hyperparameters of ResiDPO. The marker size N in panels a and d [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Comparison of amino acid distributions generated by different methods on natural back [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: ResiDPO achieves good per￾formance with only a thousand training samples. Beyond enhancing sequence designability, ResiDPO of￾fers promise for optimizing other functional properties, particularly those constrained by the high cost of wet-lab experiments and data acquis…
Figure 6
Figure 6. Figure 6: Correlation between pLDDT and Cα RMSD of designed sequences from ResiDPO. A Appendix A.1 pLDDT is a good designability proxy. Protein designability quantifies how well a designed sequence folds into a target structure. Common metrics for assessing designability rely on…
Figure 7
Figure 7. Figure 7: Comparison of the pLDDT distribution of predicted sequence from different methods on [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Comparison of the PAE distribution of predicted sequence from different methods on the [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: Comparative structural analysis of designed proteins. AlphaFold2 predictions of sequences [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: Confusion matrices showing residue substitutions induced by EnhancedMPNN on natural [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]
Figure 11
Figure 11. Figure 11: Distribution of pLDDT difference (δ) for the PDB-D training set. Most structures exhibit a pLDDT difference ranging from 5 to 15. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_11.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. Preference-based Antibody Expression Ranking: Scaling with Large-scale Weak Supervision

    cs.LG 2026-06 conditional novelty 5.0 of 10

    A two-stage recipe — camelid-sequence continual pretraining plus DPO-style preference fine-tuning with weak positive pairs — improves antibody expression ranking over supervised baselines on an internal 1254-sequence ...

Reference graph

Works this paper leans on

47 extracted references · 45 canonical work pages · cited by 1 Pith paper

  1. [1]

    Boyken, and David Baker

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

  2. [2]

    Watson, David Juergens, Nathaniel R

    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, Basile I. M. Wicky, Nikita Hanikel, Samuel J. Pellock, Alexis Courbet, William Sheffler, Jue Wang, Preetham Venkatesh, Isaac Sappington, Susana Vázquez Torres, Anna Lauko, Valentin De Bort...

  3. [3]

    Mackessy, Stefanie K

    Susana Vázquez Torres, Melisa Benard Valle, Stephen P. Mackessy, Stefanie K. Menzies, Nicholas R. Casewell, Shirin Ahmadi, Nick J. Burlet, Edin Muratspahi ´c, Isaac Sappington, Max D. Overath, Esperanza Rivera-de-Torre, Jann Ledergerber, Andreas H. Laustsen, Kim Boddum, Asim K. Bera, Alex Kang, Evans Brackenbrough, Iara A. Cardoso, Edouard P. Crittenden, ...

  4. [4]

    Woodbury, Woody Ahern, Indrek Kalvet, Nikita Hanikel, Saman Salike, Samuel J

    Donghyo Kim, Seth M. Woodbury, Woody Ahern, Indrek Kalvet, Nikita Hanikel, Saman Salike, Samuel J. Pellock, Anna Lauko, Donald Hilvert, and David Baker. Computational Design of Metallohydrolases, November 2024

  5. [5]

    Dauparas, I

    J. Dauparas, I. Anishchenko, N. Bennett, H. Bai, R. J. Ragotte, L. F. Milles, B. I. M. Wicky, A. Courbet, R. J. de Haas, N. Bethel, P. J. Y . Leung, T. F. Huddy, S. Pellock, D. Tischer, F. Chan, B. Koepnick, H. Nguyen, A. Kang, B. Sankaran, A. K. Bera, N. P. King, and D. Baker. Robust deep learning–based protein sequence design using ProteinMPNN.Science, ...

  6. [6]

    Generative Models for Graph-Based Protein Design

    John Ingraham, Vikas Garg, Regina Barzilay, and Tommi Jaakkola. Generative Models for Graph-Based Protein Design. InNeurIPS, volume 32. Curran Associates, Inc., 2019

  7. [7]

    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. InICLR, October 2020

  8. [8]

    Rotamer-free protein sequence design based on deep learning and self-consistency.Nature Computational Science, 2(7):451–462, July 2022

    Yufeng Liu, Lu Zhang, Weilun Wang, Min Zhu, Chenchen Wang, Fudong Li, Jiahai Zhang, Houqiang Li, Quan Chen, and Haiyan Liu. Rotamer-free protein sequence design based on deep learning and self-consistency.Nature Computational Science, 2(7):451–462, July 2022

Show all 47 references
  1. [9]

    Protein sequence design by deep learning.Nature Computational Science, 2(7):416– 417, July 2022

    Jue Wang. Protein sequence design by deep learning.Nature Computational Science, 2(7):416– 417, July 2022

  2. [10]

    Zhangyang Gao, Cheng Tan, and Stan Z. Li. PiFold: Toward effective and efficient protein inverse folding. InICLR, September 2022

  3. [11]

    Mathews, Carla P

    Namrata Anand, Raphael Eguchi, Irimpan I. Mathews, Carla P. Perez, Alexander Derry, Russ B. Altman, and Po-Ssu Huang. Protein sequence design with a learned potential.Nature Commu- nications, 13(1):746, February 2022

  4. [12]

    Moksh Jain, Emmanuel Bengio, Alex Hernandez-Garcia, Jarrid Rector-Brooks, Bonaventure F. P. Dossou, Chanakya Ajit Ekbote, Jie Fu, Tianyu Zhang, Michael Kilgour, Dinghuai Zhang, Lena Simine, Payel Das, and Yoshua Bengio. Biological Sequence Design with GFlowNets. InProceedings ...

  5. [13]

    ProRefiner: An entropy-based refining strategy for inverse protein folding with global graph attention.Nature Communications, 14(1):7434, November 2023

    Xinyi Zhou, Guangyong Chen, Junjie Ye, Ercheng Wang, Jun Zhang, Cong Mao, Zhanwei Li, Jianye Hao, Xingxu Huang, Jin Tang, and Pheng Ann Heng. ProRefiner: An entropy-based refining strategy for inverse protein folding with global graph attention.Nature Communications, 14(1):743...

  6. [14]

    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. InICML, June 2023

  7. [15]

    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, October 2023

  8. [16]

    Zhangyang Gao, Jue Wang, Cheng Tan, Lirong Wu, Yufei Huang, Siyuan Li, Zhirui Ye, and Stan Z. Li. UniIF: Unified Molecule Inverse Folding. InNeurIPS. arXiv, May 2024

  9. [17]

    PRosettaC: Rosetta Based Modeling of PROTAC Mediated Ternary Complexes.Journal of Chemical Information and Modeling, 60(10):4894–4903, October 2020

    Daniel Zaidman, Jaime Prilusky, and Nir London. PRosettaC: Rosetta Based Modeling of PROTAC Mediated Ternary Complexes.Journal of Chemical Information and Modeling, 60(10):4894–4903, October 2020

  10. [18]

    Nathaniel R. Bennett, Brian Coventry, Inna Goreshnik, Buwei Huang, Aza Allen, Dionne Vafeados, Ying Po Peng, Justas Dauparas, Minkyung Baek, Lance Stewart, Frank DiMaio, Steven De Munck, Savvas N. Savvides, and David Baker. Improving de novo protein binder design with deep lea...

  11. [19]

    ProteinBench: A Holistic Evaluation of Protein Foundation Models

    Fei Ye, Zaixiang Zheng, Dongyu Xue, Yuning Shen, Lihao Wang, Yiming Ma, Yan Wang, Xinyou Wang, Xiangxin Zhou, and Quanquan Gu. ProteinBench: A Holistic Evaluation of Protein Foundation Models. https://arxiv.org/abs/2409.06744v1, September 2024

  12. [20]

    Christiano, Jan Leike, Tom B

    Paul F. Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. InNeurIPS, NIPS’17, pages 4302–4310, Red Hook, NY , USA, December 2017. Curran Associates Inc

  13. [21]

    https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/

    Llama 3.2: Revolutionizing edge AI and vision with open, customizable models. https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/

  14. [22]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning, January 2025

    DeepSeek-AI. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning, January 2025

  15. [23]

    Ab initio characterization of protein molecular dynamics with AI2BMD

    Tong Wang, Xinheng He, Mingyu Li, Yatao Li, Ran Bi, Yusong Wang, Chaoran Cheng, Xiangzhen Shen, Jiawei Meng, He Zhang, Haiguang Liu, Zun Wang, Shaoning Li, Bin Shao, and Tie-Yan Liu. Ab initio characterization of protein molecular dynamics with AI2BMD. Nature, 635(8040):1019–1...

  16. [24]

    Senior, Richard Evans, John Jumper, James Kirkpatrick, Laurent Sifre, Tim Green, Chongli Qin, Augustin Žídek, Alexander W

    Andrew W. Senior, Richard Evans, John Jumper, James Kirkpatrick, Laurent Sifre, Tim Green, Chongli Qin, Augustin Žídek, Alexander W. R. Nelson, Alex Bridgland, Hugo Penedones, Stig Petersen, Karen Simonyan, Steve Crossan, Pushmeet Kohli, David T. Jones, David Silver, Koray Kav...

  17. [25]

    Kinch, R

    Minkyung Baek, Frank DiMaio, Ivan Anishchenko, Justas Dauparas, Sergey Ovchinnikov, Gyu Rie Lee, Jue Wang, Qian Cong, Lisa N. Kinch, R. Dustin Schaeffer, Claudia Millán, Hahnbeom Park, Carson Adams, Caleb R. Glassman, Andy DeGiovanni, Jose H. Pereira, Andria V . Rodrigues, Alb...

  18. [26]

    John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ron- neberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, Alex Bridgland, Clemens Meyer, Simon A. A. Kohl, Andrew J. Ballard, Andrew Cowie, Bernardino Romera-Paredes, Sta...

  19. [27]

    Manning, Stefano Ermon, and Chelsea Finn

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D. Manning, Stefano Ermon, and Chelsea Finn. Direct Preference Optimization: Your Language Model is Secretly a Reward Model.Advances in Neural Information Processing Systems, 36:53728–53741, December 2023

  20. [28]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is All you Need. In I. Guyon and U. V . Luxburg and S. Bengio and H. Wallach and R. Fergus and S. Vishwanathan and R. Garnett, editor, NeurIP...

  21. [29]

    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. InICML, pages 8946–8970. PMLR, June 2022

  22. [30]

    Cheng Tan, Zhangyang Gao, Jun Xia, Bozhen Hu, and Stan Z. Li. Global-Context Aware Generative Protein Design. InICASSP, pages 1–5, June 2023

  23. [31]

    Nakou, Sebastian Lindner, Gavin Ayres, Lesley S

    Geraldene Munsamy, Ramiro Illanes-Vicioso, Silvia Funcillo, Ioanna T. Nakou, Sebastian Lindner, Gavin Ayres, Lesley S. Sheehan, Steven Moss, Ulrich Eckhard, Philipp Lorenz, and Noelia Ferruz. Conditional language models enable the efficient design of proficient enzymes, May 2024

  24. [32]

    Accurate and robust protein sequence design with CarbonDesign.Nature Machine Intelligence, 6(5):536–547, May 2024

    Milong Ren, Chungong Yu, Dongbo Bu, and Haicang Zhang. Accurate and robust protein sequence design with CarbonDesign.Nature Machine Intelligence, 6(5):536–547, May 2024

  25. [33]

    Atomic context-conditioned protein sequence design using LigandMPNN.Nature Methods, 22(4):717–723, April 2025

    Justas Dauparas, Gyu Rie Lee, Robert Pecoraro, Linna An, Ivan Anishchenko, Cameron Glasscock, and David Baker. Atomic context-conditioned protein sequence design using LigandMPNN.Nature Methods, 22(4):717–723, April 2025

  26. [34]

    A General Theoretical Paradigm to Understand Learning from Human Preferences

    Mohammad Gheshlaghi Azar, Zhaohan Daniel Guo, Bilal Piot, Remi Munos, Mark Rowland, Michal Valko, and Daniele Calandriello. A General Theoretical Paradigm to Understand Learning from Human Preferences. InProceedings of The 27th International Conference on Artificial Intelligen...

  27. [35]

    Liu, and Xuanhui Wang

    Tianqi Liu, Zhen Qin, Junru Wu, Jiaming Shen, Misha Khalman, Rishabh Joshi, Yao Zhao, Mohammad Saleh, Simon Baumgartner, Jialu Liu, Peter J. Liu, and Xuanhui Wang. LiPO: Listwise Preference Optimization through Learning-to-Rank, May 2024

  28. [36]

    SimPO: Simple Preference Optimization with a Reference-Free Reward

    Yu Meng, Mengzhou Xia, and Danqi Chen. SimPO: Simple Preference Optimization with a Reference-Free Reward. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, November 2024

  29. [37]

    Ryan Park, Darren J. Hsu, C. Brian Roland, Maria Korshunova, Chen Tessler, Shie Mannor, Olivia Viessmann, and Bruno Trentini. Improving Inverse Folding for Peptide Design with Diversity-regularized Direct Preference Optimization, October 2024

  30. [38]

    The Llama 3 Herd of Models, November 2024

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, and et al. The Llama 3 Herd of Models, November 2024

  31. [39]

    Berman, John Westbrook, Zukang Feng, Gary Gilliland, T

    Helen M. Berman, John Westbrook, Zukang Feng, Gary Gilliland, T. N. Bhat, Helge Weissig, Ilya N. Shindyalov, and Philip E. Bourne. The Protein Data Bank.Nucleic Acids Research, 28(1):235–242, January 2000

  32. [40]

    Ballard, Joshua Bambrick, Sebastian W

    Josh Abramson, Jonas Adler, Jack Dunger, Richard Evans, Tim Green, Alexander Pritzel, Olaf Ronneberger, Lindsay Willmore, Andrew J. Ballard, Joshua Bambrick, Sebastian W. Bodenstein, David A. Evans, Chia-Chun Hung, Michael O’Neill, David Reiman, Kathryn Tunyasuvunakool, Zachar...

  33. [41]

    D. P. Kingma and L. J. Ba. Adam: A Method for Stochastic Optimization. InICLR. Ithaca, NY ArXiv, 2015

  34. [42]

    Woodbury, Donghyo Kim, Indrek Kalvet, Yakov Kipnis, Brian Coventry, Han Raut Altae-Tran, Magnus Bauer, Regina Barzilay, Tommi S

    Woody Ahern, Jason Yim, Doug Tischer, Saman Salike, Seth M. Woodbury, Donghyo Kim, Indrek Kalvet, Yakov Kipnis, Brian Coventry, Han Raut Altae-Tran, Magnus Bauer, Regina Barzilay, Tommi S. Jaakkola, Rohith Krishna, and David Baker. Atom level enzyme active site scaffolding usi...

  35. [43]

    Provably Robust DPO: Aligning Language Models with Noisy Feedback

    Sayak Ray Chowdhury, Anush Kini, and Nagarajan Natarajan. Provably Robust DPO: Aligning Language Models with Noisy Feedback. InICML, pages 42258–42274. PMLR, July 2024

  36. [44]

    Liu, and Jialu Liu

    Tianqi Liu, Yao Zhao, Rishabh Joshi, Misha Khalman, Mohammad Saleh, Peter J. Liu, and Jialu Liu. Statistical Rejection Sampling Improves Preference Optimization. InICLR, October 2023

  37. [45]

    Model Alignment as Prospect Theoretic Optimization

    Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. Model Alignment as Prospect Theoretic Optimization. InICML, pages 12634–12651. PMLR, July 2024

  38. [46]

    Noise Contrastive Alignment of Language Models with Explicit Rewards.NeurIPS, 37:117784–117812, December 2024

    Huayu Chen, Guande He, Lifan Yuan, Ganqu Cui, Hang Su, and Jun Zhu. Noise Contrastive Alignment of Language Models with Explicit Rewards.NeurIPS, 37:117784–117812, December 2024

  39. [47]

    designability

    Yue Wu, Zhiqing Sun, Huizhuo Yuan, Kaixuan Ji, Yiming Yang, and Quanquan Gu. Self-Play Preference Optimization for Language Model Alignment. InICLR, January 2025. 13 30 40 50 60 70 80 90 100 pLDDT 0 10 20 30 40RMSD Figure 6: Correlation between pLDDT and CαRMSD of designed seq...

Pith tools

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