Pith. sign in

REVIEW 2 major objections 4 minor 64 references

Mask prior-guided denoising diffusion improves inverse protein folding

T0 review · 2 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read A discrete diffusion model with mask-prior refinement sets a new state of the art for inverse protein folding, reaching 61.03% median sequence recovery on CATH 4.2 and the best AlphaFold2 foldability scores without external knowledge.

desk verdict MapDiff is a real empirical advance in inverse folding, but the headline SOTA gap partly reflects a 50-pass Monte-Carlo inference budget that baselines don't get. read the letter →

arxiv 2412.07815 v2 pith:ZWOPIO73 submitted 2024-12-10 q-bio.BM cs.LG

classification q-bio.BMcs.LG
keywords inverseproteinfoldingdiscretedenoisingdiffusionmask-priorpre-trainingsequencedesignequivariantgraphneuralnetworkinvariantpointattentionrecoveryfoldability
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

The paper claims that inverse protein folding — generating amino acid sequences that fold into a given 3D backbone — is better formulated as a discrete denoising diffusion problem than as a one-shot or autoregressive decoding problem. It introduces MapDiff, a diffusion model that starts from noise over the twenty amino acids and iteratively removes it, conditioned on the backbone structure, using a two-step denoising network. The network first predicts the whole sequence, then masks the low-confidence residues (typically flexible or disordered regions) and refines them with a pre-trained masked sequence designer, so residue interactions carry information to the uncertain positions. Across four benchmarks, MapDiff reports median sequence recovery of 61.03% on CATH 4.2 and 60.86% on CATH 4.3, roughly seven points above previous state of the art, and the best AlphaFold2-refolded similarity (TM-score 88.77%, RMSD 2.57 Å), all without external knowledge such as language models.

What carries the argument

The load-bearing object is the mask-prior-guided denoising network $\phi_\theta$, a two-stage denoiser invoked at every reverse-diffusion step. In the first stage, a global-aware equivariant graph neural network (EGNN) maps the noisy sequence and the backbone residue graph to a clean-sequence prediction; in the second, an entropy-based mask with a mask-ratio adapter that scales with the diffusion noise level $\beta_t$ blanks out the low-confidence residues, and an invariant point attention (IPA) network — pre-trained with BERT-style masked language modelling on the same CATH training data — refines them, the two predictions being blended by entropy-weighted logits. Around this denoiser, the paper builds a discrete denoising diffusion model with transition matrices $Q_t = (1-\beta_t)I + \beta_t M$, where $M$ is either the uniform distribution over the 20 amino acids or their marginal distribution in the training data, a cosine noise schedule over 500 steps, and a reverse posterior $q(\mathbf{x}_{t-1} \mid \mathbf{x}_t, \hat{\mathbf{x}}_0)$ computed from the network's predicted clean distribution $\hat{\mathbf{x}}_0$; discrete DDIM skips steps for speed, and Monte-Carlo dropout averages stochastic forward passes to reduce sampling uncertainty.

What would settle it

Express a sample of MapDiff-designed sequences — starting with the 1NI8, 2HKY, and 2P0X test cases — and determine their folded structures experimentally; if a large fraction fails to adopt the target folds, the foldability claim is falsified even though the in silico AlphaFold2 metrics look strong.

Watch

Extended reading notes

Core claim

The central discovery is that discrete denoising diffusion with a mask-prior-guided denoising network is a strong generative formulation for inverse protein folding. Concretely, the paper shows that a diffusion process on the 20 amino acid types, conditioned on the backbone residue graph, can be reversed by a network that alternates between a global-equivariant graph predictor (the base sequence predictor) and a pre-trained invariant point attention network that refines exactly the residues the base predictor is least confident about, chosen by an entropy-based mask whose ratio adapts to the noise level of the denoising step. The paper reports that this combination yields median recovery 61.03% and perplexity 3.46 on CATH 4.2 (best prior: 52.63% recovery), similar margins on CATH 4.3, the best zero-shot recovery on TS50 and PDB2022, and the best foldability statistics when the designed sequences are refolded with AlphaFold2, even for sequences with low recovery. The message is that uncertain, low-confidence positions — disordered loops in particular — are better handled by iterative refinement guided by the confidently predicted context than by single-pass or autoregressive prediction, and that this can be achieved without any external knowledge source.

Load-bearing premise

Every reverse step treats the denoising network's predicted clean sequence as if it were the true one when computing the next less-noisy sequence, so a systematic bias in that prediction is fed back and can accumulate across the 500-step trajectory.

Editorial extensions

If this is right

  • Structure-only designers can match or beat models that lean on protein language models and extra training data.
  • Iterative denoising improves exactly the hard cases: coils, bends, and disordered loops, where the reported margins over prior methods are largest.
  • Low-confidence residues should be treated as refinement targets; removing the IPA refinement step costs 4.47% recovery in the paper's ablation.
  • DDIM skipping with Monte-Carlo dropout keeps sequence quality high while generation is accelerated, so speed need not be sacrificed.

Reading between the lines

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

  • The entropy-mask-plus-refinement pattern is portable: if the gains are real, the same base-predictor-then-refine-uncertain-positions design could transfer to antibody loop grafting, small-molecule docking, or RNA design, where confidence also varies sharply across positions.
  • A testable extension the authors leave implicit is feeding predicted (AlphaFold2-generated) backbones as conditioning input; the mask-ratio adapter might then reveal whether iterative masking also absorbs errors in the structure itself, not only in the sequence.
  • The contribution of the confidence signal itself, as opposed to the refinement module, could be isolated by comparing entropy-based masking with random masking at equal mask ratios — a control that would make the mechanism fully transparent.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The manuscript proposes MapDiff, a discrete denoising diffusion model for inverse protein folding. The method formulates sequence generation as iterative denoising of a categorical amino acid sequence conditioned on a backbone structure, using an EGNN-based base predictor, an entropy-based masking strategy with a mask-ratio adapter, and a pre-trained IPA masked-sequence designer for refinement. Inference combines discrete DDIM step skipping with Monte-Carlo dropout. On CATH 4.2, CATH 4.3, TS50, and PDB2022, the authors report recovery, perplexity, and NSSR improvements over reproduced baselines, plus AlphaFold2 foldability metrics on the CATH 4.2 test set. The paper includes ablations, sensitivity analyses, and qualitative structure comparisons.

Significance. MapDiff is a plausible and well-engineered contribution: the benchmark suite is broad, the main baselines are reproduced, the code and data are publicly archived, and the ablation study decomposes the architecture. The reported AlphaFold2 foldability analysis is an appropriate in-silico check rather than an overclaim. If the headline margins survive matched-inference and uncertainty quantification, the method would be a new state of the art for structure-only inverse folding. However, the central empirical claim is currently supported only by point estimates computed under a substantially larger inference budget than the baselines, so the significance is conditional on the additional comparisons requested below.

major comments (2)
  1. [Implementation setup; Table 1; Supplementary S7/Fig. 3a] The headline comparisons in Table 1 are not made at matched inference cost. MapDiff's reported results use 50 Monte-Carlo dropout samples (C = 50 in Algorithm 2), whereas ProteinMPNN, PiFold, LM-Design, and GRADE-IF are evaluated with their default single-pass settings. The paper's own sensitivity analysis (Supplementary Fig. 3a) shows that recovery 'substantially improves' with the number of Monte-Carlo samples and stabilizes only around 20 samples, so the 61.03% versus 52.63% margin in Table 1 cannot be attributed to the learned denoising architecture alone. Please report MapDiff with a single forward pass, with matched numbers of stochastic samples for stochastic baselines, and at a matched total inference budget such as wall-clock time or FLOPs per protein. Until such a comparison is provided, the 'substantially outperforms' claim is not established.
  2. [Table 1 and Table 2] All headline results are single point estimates. None of the recovery, perplexity, NSSR, or foldability comparisons in Tables 1 and 2 include error bars, confidence intervals, or significance tests, although both MapDiff's diffusion sampling and the baselines' decoding are stochastic. Because the central claim is a quantitative margin (e.g., 7.74% recovery on CATH 4.2 and 6.33% NSSR62 on TS50), the authors should report variability across multiple sampling runs or bootstrap resampling over test proteins, and check whether the margins are statistically significant.
minor comments (4)
  1. [Algorithm 2 and 'DDIM with Monte-Carlo dropout'] The text describes mean-pooling the stochastic logits at each denoising step, but Algorithm 2 as written runs C independent full denoising trajectories and averages only the final-step predictions p^m_0. Please clarify which procedure was actually used, since this affects the interpretation of C and of the sensitivity analysis in Supplementary Fig. 3a.
  2. [Table 3] The checkmark rows in the ablation table are inconsistent with the accompanying text: the text identifies variants 2 and 4 as removing global context and coordinate updating, respectively, but the table's row entries do not clearly show those removals. Please make the component-to-variant mapping explicit so the reader can verify which module deletion produced each result.
  3. [Supplementary S3, Eq. (32)] The notation [I_k Q_t]^T appears without defining I_k; if this is the identity matrix, the expression should be stated more transparently as x_t Q_t^T, and the typo should be corrected.
  4. [Methods, Eq. (8) and Supplementary S3] The reverse posterior substitutes the network prediction x̂_0 for the true clean x_0 in the exact posterior. This is a standard diffusion-model approximation and is acceptable for the empirical claim, but the risk of accumulating bias over the 500-step trajectory should be acknowledged explicitly, since Eq. (8) is presented as an exact derivation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MapDiff's performance claims are empirical and benchmarked against external baselines, and its generative equations are standard derivations rather than restatements of its inputs.

full rationale

The paper's central claims are benchmark results (sequence recovery, NSSR, AlphaFold2 foldability) obtained by training on CATH 4.2/4.3 and evaluating on held-out CATH test sets plus independent TS50 and PDB2022 datasets; no test-label information is used to set parameters, and model selection is explicitly based on validation recovery. The discrete diffusion posterior (Eq. 8) and its DDIM variant (Eq. 9) are derived from Bayes' rule in Supplementary S3, with the network prediction x̂0 trained by cross-entropy, so the sampling distribution is not definitionally equal to its input. The mask-prior pre-training, entropy-based masking, mask-ratio adapter, and entropy-weighted logit combination (Eq. 21) form an internal refinement and gating scheme rather than a self-referential prediction, and the ablation study tests each component's contribution. No load-bearing self-citation appears: DDIM, Monte-Carlo dropout, discrete DDIM, EGNN, IPA, and entropy masking are attributed to external prior work, and the only self-citation is the code archive. The 50-pass Monte-Carlo dropout inference with baselines evaluated at default single-pass settings raises a benchmark-fairness concern (Supplementary Fig. 3a shows recovery improves with sample count), but this is not a circularity: the reported numbers characterize the method's chosen stochastic ensemble, not a quantity forced to equal a fitted input by construction.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

MapDiff introduces no new physical or biological entities; it is a machine-learning architecture. The central claim depends on standard discrete diffusion mathematics plus several domain assumptions about entropy-based refinement, independent per-residue noise, dataset splitting, and AlphaFold2 as a foldability proxy. The hand-set constants in the mask-ratio adapter and the inference hyperparameters are not fitted to test labels but do affect the reported gains.

free parameters (4)
  • mask ratio adapter minimum ratio m = 0.4
    Hand-set in Eq (17); controls the minimum fraction of residues masked for IPA refinement in the two-step denoising network.
  • mask ratio adapter deviation sigma = 0.2
    Hand-set in Eq (17); shapes how the mask ratio varies with the diffusion noise level beta_t.
  • Monte-Carlo forward passes C = 50
    Chosen for sampling inference; sensitivity analysis in Supp. Fig. 3a shows performance stabilizes around 20 samples.
  • DDIM skip steps k = 100
    Chosen for accelerated sampling; sensitivity is studied in Supp. Fig. 3b.
assumptions (6)
  • domain assumption The reverse posterior q(x_{t-1}|x_t, x̂0) in Eq (8), using the network's point estimate x̂0, is a valid approximation for sampling the denoising trajectory.
    All discrete diffusion sampling relies on this approximation; errors in x̂0 can compound over iterations. MC dropout mitigates but does not eliminate this.
  • domain assumption Entropy-based masking of high-entropy residues identifies residues whose identities are best refined by the IPA network using sequence context.
    Used in Eq (16) and the two-step denoising network; if the entropy signal does not localize correctable residues, the refinement step would not help.
  • domain assumption The CATH topology-based split provides a valid test of generalization to unseen folds and sequences.
    Standard protocol, but it does not exclude sequence-level similarity or homologous fragments between train and test.
  • domain assumption AlphaFold2-refolded structural similarity is a meaningful in silico proxy for foldability.
    Explicitly acknowledged in Results; high AlphaFold2 confidence does not guarantee experimental folding.
  • standard math Discrete transition matrices Q_t of uniform or marginal form yield a valid stationary prior and tractable closed-form posterior (Eqs 1-8).
    Follows Austin et al. and DiGress; proofs are sketched in Supp. S3.
  • domain assumption The forward noise is applied independently to each residue (Eqs 4-5 and 7), so residue correlations are only captured by the learned denoiser.
    Common in discrete diffusion for sequences; if residue correlations are essential during the diffusion process, the independent Markov chain may limit generation quality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mask prior-guided denoising diffusion improves inverse protein folding." pith.science (2026). https://pith.science/paper/ZWOPIO73

@misc{pith2026241207815,
  author       = {Pith},
  title        = {Pith review of: Mask prior-guided denoising diffusion improves inverse protein folding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZWOPIO73}},
  note         = {Machine review of arXiv:2412.07815}
}
read the original abstract

Inverse protein folding generates valid amino acid sequences that can fold into a desired protein structure, with recent deep-learning advances showing strong potential and competitive performance. However, challenges remain, such as predicting elements with high structural uncertainty, including disordered regions. To tackle such low-confidence residue prediction, we propose a Mask-prior-guided denoising Diffusion (MapDiff) framework that accurately captures both structural information and residue interactions for inverse protein folding. MapDiff is a discrete diffusion probabilistic model that iteratively generates amino acid sequences with reduced noise, conditioned on a given protein backbone. To incorporate structural information and residue interactions, we develop a graph-based denoising network with a mask-prior pre-training strategy. Moreover, in the generative process, we combine the denoising diffusion implicit model with Monte-Carlo dropout to reduce uncertainty. Evaluation on four challenging sequence design benchmarks shows that MapDiff substantially outperforms state-of-the-art methods. Furthermore, the in silico sequences generated by MapDiff closely resemble the physico-chemical and structural characteristics of native proteins across different protein families and architectures.

Figures

Figures reproduced from arXiv: 2412.07815 by the authors.

Figure 1
Figure 1. Mask-prior-guided denoising diffusion (MapDiff) for inverse protein folding. (a) The mask-prior pre-training stage randomly masks residues within the AA sequence and pre-trains an invariant point attention (IPA) network with the masked sequence and the 3D backbone structure to learn prior structural and sequence knowledge, using BERT-like masked language modelling objectives. (b) The mask-prior-guided denoising netw… view at source ↗
Figure 2
Figure 2. Model performance comparison and sensitivity analysis across different scenarios on the CATH datasets. (a) NSSR scores for MapDiff and baseline methods1, 13, 29, 30 on the full test sets and the short and single-chain protein subsets for four different BLOSUM matrices and no BLOSUM matrix. (b) Sum-normalized confusion matrices for MapDiff (left) and LM-Design (right) predictions, and the softmax-normalized native BL… view at source ↗
Figure 3
Figure 3. Comparison of three refolded structures (left) and the respective model-designed sequences (right) for proteins with PDB IDs 1NI8, 2HKY, and 2P0X. (a) Refolded tertiary structure visualization of the sequences designed by three models MapDiff (red), GRADE-IF (orange) and LM-Design (blue). The refolded structures are generated by AlphaFold2 and superposed against the ground-truth structures (purple). For each model a… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

64 extracted references · 61 canonical work pages

  1. [1]

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

  2. [2]

    Høie, M. H. et al. Antifold: Improved antibody structure-based design using inverse folding. Bioinforma. Adv. 5, vbae202 (2025)

  3. [3]

    Alford, R. F. et al. The rosetta all-atom energy function for macromolecular modeling and design. J. Chem. Theory Comput. 13, 3031–3048 (2017)

  4. [4]

    E., Arnold, F

    Wu, Z., Johnston, K. E., Arnold, F. H. & Yang, K. K. Protein sequence design with deep generative models. Curr. Opin. Chem. Biol. 65, 18–27 (2021)

  5. [5]

    & Zhou, Y

    Li, Z., Yang, Y ., Faraggi, E., Zhan, J. & Zhou, Y . Direct prediction of profiles of sequences compatible with a protein structure by neural networks with fragment-based local and energy-based nonlocal profiles. Proteins: Struct. Funct. Bioinforma. 82, 2565–2573 (2014)

  6. [6]

    O’Connell, J. et al. Spin2: Predicting sequence profiles from protein structures using deep neural networks. Proteins: Struct. Funct. Bioinforma. 86, 629–633 (2018)

  7. [7]

    Anand, N. et al. Protein sequence design with a learned potential. Nat. Commun. 13, 746 (2022)

  8. [8]

    & Daggett, V

    Towse, C.-L. & Daggett, V . When a domain is not a domain, and why it is important to properly filter proteins in databases: conflicting definitions and fold classification systems for structural domains make filtering of such databases imperative. Bioessays 34, 1060–1069 (2012)

Show all 64 references
  1. [9]

    Li, B., Tian, J., Zhang, Z., Feng, H. & Li, X. Multitask non-autoregressive model for human motion prediction. IEEE Transactions on Image Process. 30, 2562–2574 (2020)

  2. [10]

    & Odobez, J.-M

    Martínez-González, A., Villamizar, M. & Odobez, J.-M. Pose transformers (POTR): Human motion prediction with non-autoregressive transformers. In 2021 IEEE/CVF International Conference on Computer Vision Workshops , 2276–2284 (IEEE, 2021)

  3. [11]

    Starr, T. N. & Thornton, J. W. Epistasis in protein evolution. Protein Sci. 25, 1204–1218 (2016)

  4. [12]

    Anytime sampling for autoregressive models via ordered autoencoding

    Xu, Y .et al. Anytime sampling for autoregressive models via ordered autoencoding. In International Conference on Learning Representations (2021)

  5. [13]

    Gao, Z., Tan, C. & Li, S. Z. Pifold: Toward effective and efficient protein inverse folding. In International Conference on Learning Representations (2023)

  6. [14]

    & Garton, M

    Lyu, S., Sowlati-Hashjin, S. & Garton, M. Variational autoencoder for design of synthetic viral vector serotypes. Nat. Mach. Intell. 6, 1–14 (2024)

  7. [15]

    Jumper, J. et al. Highly accurate protein structure prediction with alphafold. Nature 596, 583–589 (2021). 15/24

  8. [16]

    Abramson, J. et al. Accurate structure prediction of biomolecular interactions with alphafold 3. Nature 630, 493–500 (2024)

  9. [17]

    Watson, J. L. et al. De novo design of protein structure and function with rfdiffusion. Nature 620, 1089–1100 (2023)

  10. [18]

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

  11. [19]

    & Jaakkola, T

    Jing, B., Corso, G., Chang, J., Barzilay, R. & Jaakkola, T. Torsional diffusion for molecular conformer generation. Adv. Neural Inf. Process. Syst. 35, 24240–24253 (2022)

  12. [20]

    Schneuing, A. et al. Structure-based drug design with equivariant diffusion models. Nat. Comput. Sci. 4, 899–909 (2024)

  13. [21]

    & Ermon, S

    Song, J., Meng, C. & Ermon, S. Denoising diffusion implicit models. In International Conference on Learning Representations (2021)

  14. [22]

    & Ghahramani, Z

    Gal, Y . & Ghahramani, Z. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In International Conference on Machine Learning , 1050–1059 (PMLR, 2016)

  15. [23]

    G., Hoogeboom, E

    Satorras, V . G., Hoogeboom, E. & Welling, M. E (n) equivariant graph neural networks. InInternational Conference on Machine Learning, 9323–9332 (PMLR, 2021)

  16. [24]

    Zhou, X. et al. Prorefiner: an entropy-based refining strategy for inverse protein folding with global graph attention. Nat. Commun. 14, 7434 (2023)

  17. [25]

    Hsu, C. et al. Learning inverse folding from millions of predicted structures. In International Conference on Machine Learning, 8946–8970 (PMLR, 2022)

  18. [26]

    & Jaakkola, T

    Ingraham, J., Garg, V ., Barzilay, R. & Jaakkola, T. Generative models for graph-based protein design.Adv. Neural Inf. Process. Syst. 32, 15820–15831 (2019)

  19. [27]

    Jing, B., Eismann, S., Suriana, P., Townshend, R. J. L. & Dror, R. Learning from protein structure with geometric vector perceptrons. In International Conference on Learning Representations (2020)

  20. [28]

    Gao, Z., Tan, C. & Li, S. Z. Alphadesign: A graph protein design method and benchmark on alphafolddb. arXiv preprint arXiv:2202.01079 (2022)

  21. [29]

    Zheng, Z. et al. Structure-informed language models are protein designers. In International Conference on Machine Learning, 42317–42338 (PMLR, 2023)

  22. [30]

    & Wang, Y

    Yi, K., Zhou, B., Shen, Y ., Liò, P. & Wang, Y . Graph denoising diffusion for inverse protein folding.Adv. Neural Inf. Process. Syst. 36, 10238–10257 (2023)

  23. [31]

    Orengo, C. A. et al. Cath–a hierarchic classification of protein domain structures. Structure 5, 1093–1109 (1997)

  24. [32]

    & Merkl, R

    Löffler, P., Schmitz, S., Hupfeld, E., Sterner, R. & Merkl, R. Rosetta: Msf: a modular framework for multi-state computational protein design. PLoS Comput. Biol. 13, e1005600 (2017)

  25. [33]

    & Henikoff, J

    Henikoff, S. & Henikoff, J. G. Amino acid substitution matrices from protein blocks. Proc. Natl. Acad. Sci. 89, 10915–10919 (1992)

  26. [34]

    Vignac, C. et al. Digress: Discrete denoising diffusion for graph generation. In International Conference on Learning Representations (2023)

  27. [35]

    & Sander, C

    Kabsch, W. & Sander, C. Dictionary of protein secondary structure: pattern recognition of hydrogen-bonded and geometrical features. Biopolym. Orig. Res. on Biomol. 22, 2577–2637 (1983)

  28. [36]

    Schrödinger Release 2023-4: Maestro, Schrödinger , LLC, New York, NY, 2023 (2023)

    Schrödinger, LLC. Schrödinger Release 2023-4: Maestro, Schrödinger , LLC, New York, NY, 2023 (2023)

  29. [37]

    Limpert, E., Stahel, W. A. & Abbt, M. Log-normal distributions across the sciences: Keys and clues: On the charms of statistics, and how mechanical models resembling gambling machines offer a link to a handy way to characterize log-normal distributions, which can provide deepe...

  30. [38]

    The h-ns dimerization domain defines a new fold contributing to dna recognition

    Bloch, V .et al. The h-ns dimerization domain defines a new fold contributing to dna recognition. Nat. Struct. & Mol. Biol. 10, 212–218 (2003)

  31. [39]

    The flexible and clustered lysine residues of human ribonuclease 7 are critical for membrane permeability and antimicrobial activity

    Huang, Y .-C.et al. The flexible and clustered lysine residues of human ribonuclease 7 are critical for membrane permeability and antimicrobial activity. J. Biol. Chem. 282, 4626–4633 (2007)

  32. [40]

    Mansy, S. S. et al. Structure and evolutionary analysis of a non-biological atp-binding protein. J. Mol. Biol. 371, 501–513 (2007). 16/24

  33. [41]

    F., Gish, W., Miller, W., Myers, E

    Altschul, S. F., Gish, W., Miller, W., Myers, E. W. & Lipman, D. J. Basic local alignment search tool.J. Mol. Biol. 215, 403–410 (1990)

  34. [42]

    D., Ho, J., Tarlow, D

    Austin, J., Johnson, D. D., Ho, J., Tarlow, D. & Van Den Berg, R. Structured denoising diffusion models in discrete state-spaces. Adv. Neural Inf. Process. Syst. 34, 17981–17993 (2021)

  35. [43]

    Nichol, A. Q. & Dhariwal, P. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, 8162–8171 (PMLR, 2021)

  36. [44]

    Tan, C., Gao, Z., Xia, J., Hu, B. & Li, S. Z. Global-context aware generative protein design. In IEEE International Conference on Acoustics, Speech and Signal Processing , 1–5 (2023)

  37. [45]

    & Toutanova, K

    Devlin, J., Chang, M.-W., Lee, K. & Toutanova, K. BERT: Pre-training of deep bidirectional transformers for language understanding. In Burstein, J., Doran, C. & Solorio, T. (eds.) Proceedings of the 2019 Conference of the North American Chapter of the Association for Computati...

  38. [46]

    Akpinaroglu, D. et al. Structure-conditioned masked language models for protein sequence design generalize beyond the native sequence space. bioRxiv 2023–12 (2023)

  39. [47]

    Berman, H. M. et al. The protein data bank. Nucleic Acids Res. 28, 235–242 (2000)

  40. [48]

    Paszke, A. et al. Automatic differentiation in PyTorch. In NeurIPS 2017 Autodiff Workshop (2017)

  41. [49]

    Cock, P. J. et al. BioPython: freely available python tools for computational molecular biology and bioinformatics. Bioinformatics 25, 1422–1423 (2009)

  42. [50]

    & Lenssen, J

    Fey, M. & Lenssen, J. E. Fast graph representation learning with PyTorch Geometric. InICLR Workshop on Representation Learning on Graphs and Manifolds (2019)

  43. [51]

    Pedregosa, F. et al. Scikit-learn: Machine learning in python. J. Mach. Learn. Res. 12, 2825–2830 (2011)

  44. [52]

    Harris, C. R. et al. Array programming with NumPy. Nature 585, 357–362 (2020)

  45. [53]

    RDKit: Open-source cheminformatics

    Landrum, G. RDKit: Open-source cheminformatics. http://www.rdkit.org (2006)

  46. [54]

    & Söding, J

    Steinegger, M. & Söding, J. MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nat. biotechnology 35, 1026–1028 (2017)

  47. [55]

    & Söding, J

    Mirdita, M., Steinegger, M. & Söding, J. MMseqs2 desktop and local web server app for fast, interactive sequence searches. Bioinformatics 35, 2856–2858 (2019)

  48. [56]

    Mirdita, M. et al. Colabfold: making protein folding accessible to all. Nat. Methods 19, 679–682 (2022)

  49. [57]

    Bai, P. et al. peizhenbai/MapDiff: v1.0.0. Zenodo https://doi.org/10.5281/zenodo.15162932 (2025)

  50. [58]

    Ganea, O.-E. et al. Independent SE(3)-equivariant models for end-to-end rigid protein docking. InInternational Conference on Learning Representations (2022)

  51. [59]

    Yim, J. et al. Se(3) diffusion model with application to protein backbone generation. In International Conference on Machine Learning, 40001–40039 (PMLR, 2023). 17/24 Supplementary Material S1. Closest training structures and sequences In Supplementary Fig. 1, we present the s...

  52. [60]

    Residue graph feature construction The protein is represented as a residue graph G =(X, A, E) to reflect its geometric structure and topological relationships

    =ℎ𝜙(Xaa 𝑡 , Xpos, Xprop, E,𝑡) {Global-aware EGNN} 7: Compute base cross-entropy loss𝐿b =𝐿CE(𝑝( ˆXaa 0), Xaa 0) 8: Second step: masked sequence refinement 9: Compute base entropy{ent𝑏 1,··· ent𝑏 𝑁} =𝐸(𝑝( ˆXaa 0)) 10: Compute mask ratio mr𝑡 = sin 𝜋 2𝛽𝑡𝜎 +𝑚 11: Generate entropy-b...

  53. [61]

    by neural networksℎ𝜙, 𝑓𝜃 and Xaa 𝑡 8: Compute𝑝𝑐 𝑡(Xaa 𝑡−𝑘|Xaa 𝑡 , ˆXaa

  54. [62]

    {DDIM posterior computation} 9: end for 10: end for 11: Compute mean prediction𝑝m 0(Xaa

  55. [63]

    = 1 𝐶 Í𝑝𝑐 0(Xaa 0|Xaa 𝑘, ˆXaa

  56. [64]

    Invariant point attention The description of an invariant point attention (IPA) layer is illustrated in Algorithm 3

    {Monte-Carlo estimation} 12: return Xaa 0 ∼𝑝m 0(Xaa 0) S6. Invariant point attention The description of an invariant point attention (IPA) layer is illustrated in Algorithm 3. IPA plays a crucial role in AlphaFold215 as it determines the refinement process of protein structure...

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.