Pith. sign in

REVIEW 3 major objections 7 minor 1 cited by

All-atom inverse protein folding through discrete flow matching

T0 review · 3 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read ADFLIP is an inverse-folding model that claims the highest sequence-recovery rates to date on protein complexes with ligands, nucleotides, and metal ions, and it handles dynamic complexes by sampling across structure ensembles.

desk verdict New all-atom inverse folding model with solid core results and an undefined guidance algorithm; worth reviewing with heavy revision. read the letter →

arxiv 2507.14156 v1 pith:FP477FVJ submitted 2025-07-04 q-bio.BM cs.AI

Kai Yi , Kiarash Jamali , Sjors H. W. Scheres This is my paper · ORCID
classification q-bio.BMcs.AI
keywords inverseproteinfoldingdiscreteflowmatchingall-atomdesignprotein-ligandcomplexesnucleotide-bindingproteinsmetal-bindingsitesstructureensemblestraining-freeguidance
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 for complexes with non-protein chemistry is best treated as discrete flow matching over amino-acid tokens, and that this choice, together with progressive side-chain packing, outperforms existing all-atom models. The proposed model, ADFLIP, starts with all sequence positions masked and moves through a categorical flow toward the target sequence, conditioned on the full structure of the complex, including ligands, nucleotides, and metal ions. The paper reports higher sequence recovery at interaction sites than LigandMPNN for all three types of complexes, and shows that the same model can design sequences for dynamic complexes by averaging its predictions over NMR structure ensembles. If these results hold, ADFLIP provides one framework for designing sequences across chemically diverse and conformationally mobile assemblies, which is exactly where earlier inverse-folding methods struggle.

What carries the argument

The central object is the conditional categorical flow $p_{t|1}(s_t|s_1)=\mathrm{Cat}(t\delta\{s_1,s_t\}+(1-t)\delta\{m,s_t\})$, which interpolates from a fully masked sequence at $t=0$ to the data sequence at $t=1$. The rate matrix used in sampling is $R_t(s_t,j)=\mathbb{E}_{p_{1|t}(s_1|s_t)}[\delta\{s_1,j\}\delta\{s_t,m\}/(1-t)]$, and the denoising distribution $p_{1|t}$ is approximated by a multi-scale GNN whose atom nodes and residue nodes exchange information, with a context block that passes messages from non-protein atoms to protein residues. Side chains are packed by a separate network and fed back as structural context after each sampling step. The training-free guidance mechanism reweights the predicted distribution using a pseudo-gradient $-\nabla_{s_1}\|y-\hat{y}\|^2$ computed from a pre-trained regressor applied to the denoiser's probability vector.

What would settle it

Run Algorithm 3 on the 210 single-ligand complexes with the guidance term omitted but identical sampling and evaluation; if the affinity-gain rate stays near 58.1%, the guidance mechanism is not carrying the reported effect. Equivalently, check whether DSMBind can be called on the denoiser's probability vector $p(\hat{s}_1)$ rather than on a concrete sequence with packed side chains; if it cannot, Algorithm 3 as written is not executable.

Watch

Extended reading notes

Core claim

ADFLIP defines a discrete flow over 21 states (20 amino acids plus a mask token) with conditional flow $p_{t|1}(s_t|s_1)=\mathrm{Cat}(t\delta\{s_1,s_t\}+(1-t)\delta\{m,s_t\})$, and learns a denoising distribution $p_{1|t}(s_1|s_t,x,\chi_t)$ with a multi-scale graph neural network. Sampling follows Euler steps over the implied rate matrix, and each newly sampled amino acid has its side chains packed and added back as conditioning context for later steps. On the LigandMPNN-derived test sets, ADFLIP recovers 62.19% of ligand-interacting residues versus LigandMPNN's 59.21%, 50.21% versus 46.14% for nucleotide-interacting residues, and 75.73% versus 69.31% for metal-binding residues, with improved foldability on all three. On a new NMR ensemble dataset, averaging the denoiser's predictions over multiple conformations raises recovery by 8.6% for ligand interactions and 5.8% for nucleotide interactions compared with using a single conformation. Guiding sampling with a pre-trained binding-affinity regressor (DSMBind) raises the fraction of designs predicted to bind better than wild type from 41.9% to 58.1% while keeping foldability above 90%.

Load-bearing premise

The load-bearing premise is that a scoring model built to read complete sequences and packed structures can steer sampling when its error is turned into a gradient on a partly masked probability distribution, even though the paper does not define what a gradient means for discrete amino-acid choices; if that step is not implementable, the guidance-based affinity results do not stand.

Editorial extensions

If this is right

  • On the LigandMPNN-derived test sets, ADFLIP outperforms LigandMPNN in sequence recovery for ligand-, nucleotide-, and metal-interacting residues, with the largest absolute gain at metal-binding sites.
  • Using the full NMR ensemble as conditioning improves recovery over a single selected conformation by 8.6% for ligands and 5.8% for nucleotides, with no fine-tuning.
  • Designed sequences from ADFLIP refold to the target structure more often than LigandMPNN's on all three complex types, reaching 100% foldability on the small-molecule ligand test set.
  • The training-free guidance scheme increases the fraction of designed protein-ligand sequences whose predicted binding affinity beats wild type from 41.9% to 58.1% while keeping foldability above 90%.

Reading between the lines

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

  • The ensemble-averaging recipe is a retraining-free way to condition generation on conformational diversity, and could plausibly transfer to discrete diffusion models of RNA or small molecules, where structural ensembles also matter.
  • The guidance results are only as strong as the plug-in regressor, since DSMBind is used as-is; the same framework might target solubility, expression, or immunogenicity if a suitable predictor exists, but the discrete-gradient issue would need to be resolved first.
  • Because recovery is measured only at residues within 5 Å of non-protein atoms, the reported gains characterize interaction sites rather than global sequence design; whole-protein recovery on the same complexes could behave differently.
  • A direct experimental test would be to synthesize ADFLIP-guided designs and measure binding affinity in vitro, since the affinity-gain numbers are predictions from DSMBind and the paper notes that no experimental validation was performed.
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 / 7 minor

Summary. ADFLIP is a generative model for inverse protein folding that uses discrete flow matching on amino acid tokens, conditioned on all-atom complex structures including ligands, nucleotides, and metal ions. The model uses a multi-scale GNN denoiser, progressively packs predicted side chains during sampling, supports ensemble conditioning over multiple NMR conformers, and proposes a training-free guidance scheme that couples a pre-trained regressor into the flow. The paper reports improved recovery rates over LigandMPNN for all three complex types (Tables 1-3), improved foldability metrics (Tables 2 and 4), and a DSMBind-guided affinity experiment (Table 5). The authors release code and provide a detailed list of PDB IDs.

Significance. The recovery-rate results, if reproducible, are a useful advance: they extend inverse folding to complexes with non-protein chemistry and to multi-state ensembles, and the progressive side-chain conditioning is a sensible way to exploit all-atom context. The paper follows the LigandMPNN curation protocol and reports code and data lists, which is commendable for reproducibility. The central weakness is the guidance contribution: Algorithm 3 in Appendix C is not mathematically well-defined, and the affinity evaluation in Section 4.3 uses the same predictor for guidance and evaluation. The main benchmark results are independent of the guidance issue, but the paper's abstract and discussion present guidance as one of its three key contributions, so this needs to be fixed rather than treated as optional.

major comments (3)
  1. [Section 3.3, Algorithm 3] The training-free guidance procedure is not well-defined as written. In Algorithm 3, the line 'Compute \hat y by p(\hat s1): \hat y = h_\phi(p(\hat s1))' passes the denoiser's probability vector directly into a regressor that requires a concrete amino acid sequence and all-atom structure; DSMBind cannot accept a distribution as input. The next line, 'Approximate p(y|s1) \approx -\nabla_{s1} ||y - \hat y||^2', treats a likelihood as a negative gradient of a squared error and differentiates with respect to a discrete categorical variable, so the object on the right is not defined. The subsequent sampling step 'Sample s1 \sim p(\hat s1) p(y|\hat s1)' uses an unnormalized product and is therefore not a valid categorical distribution. Moreover, DSMBind needs side-chain coordinates, but side chains are only generated after s1 is sampled (by g(s1, x_n)), so the regressor cannot be evaluated at the point where guidance is applied. As a result, the affinity-gain results in Table 5 are not reproducible from the description. Please either specify a concrete continuous relaxation (e.g., Gumbel-softmax or straight-through gradient on logits, with a proper likelihood-ratio weighting and normalization), give the actual implementation used for the reported numbers, or remove the guidance claim from the abstract and Section 4.3.
  2. [Section 4.3, Table 5] The guidance experiment is self-referential: DSMBind is used both to steer sampling and to evaluate the resulting sequences' affinity. A generator optimized against a fixed regressor can exploit that regressor's idiosyncrasies, so the reported increase from 41.9% to 58.1% in 'Affinity gain' may reflect reward hacking rather than a true affinity improvement. This is particularly relevant because the paper states in its Impact Statement that no experimental validation was performed. Please add an evaluation with an independent affinity predictor or a docking/scoring protocol, and report per-complex distributions and standard errors for the guided and unguided conditions so that 58.1% vs 41.9% can be assessed statistically.
  3. [Tables 1-4] The central 'state-of-the-art' claim rests on recovery-rate margins that are small relative to the sampling noise. For example, Table 1 shows ADFLIP at 62.19% vs LigandMPNN at 59.21% recovery for small-molecule complexes, and Table 3 shows 50.08% vs 40.58% for the NMR ligand set, but no error bars, confidence intervals, or significance tests are reported for any of the recovery or foldability numbers, and only 10 sequences are sampled per complex. Without uncertainty quantification, it is not possible to tell whether the reported improvements are robust. Please add per-complex or bootstrap confidence intervals and paired tests (the baselines are evaluated on the same test set), and consider whether the term 'state-of-the-art' should be qualified to 'state-of-the-art among the compared methods'.
minor comments (7)
  1. [Section 4.2, Table 4] The text states that using the ensemble reduces the average RMSD 'from 9.11 to 7.61 Å', but Table 4 lists 9.10 Å for ADFLIP (Single) and 7.21 Å for ADFLIP (Multiple). Please correct the numbers so the text matches the table.
  2. [Section 4.1, Table 2] The text says 'both methods achieve high pLDDT scores (bigger than 95%)', but Table 2 reports pLDDT 90.6 for ADFLIP on small molecules and 84.7/87.5 for nucleotide complexes. Please correct this claim or clarify which subset of complexes it refers to.
  3. [Appendices B-C] The algorithms depend on the purity threshold τ, maximum iterations K, time step Δt, and guidance target y, but no values are given in the paper. Please report the settings used for Tables 1-5, or point to a configuration file in the code repository.
  4. [Section 3.1, Algorithm 1] Algorithm 1 samples a concrete sequence s1 from p(\hat s1) and then computes the rate matrix R_t as an expectation over p_{1|t}(s1|s_t); please clarify whether the rate matrix uses the sampled s1 or the full predicted distribution, since this affects the variance and the correctness of the flow matching update.
  5. [Section 4.1] The text reports PiFold nucleotide recovery as 38.1%, but Table 1 lists 38.41%; please make the numbers consistent.
  6. [Section 4.3] The text says the generation target was a 10% improvement over wild-type affinity, but Table 5 reports 'Affinity gain' as the fraction exceeding wild-type affinity, which is a different quantity; please define the metric precisely and report both the fraction meeting the 10% target and the mean/median affinity gain.
  7. [Table 2] The metal-ion row of Table 2 contains formatting artifacts ('0 .79', '97 .85', '83 .7%'); please fix the spacing in the table.

Circularity Check

1 steps flagged · score 6.0 of 10

The affinity-guidance result is self-referential: DSMBind steers the generated sequences and also defines the measured 'affinity gain', so Table 5's improvement partly reports optimization against its own objective; the core inverse-folding benchmarks are independent.

  1. self definitional [Section 4.3 'Guidance by binding affinity', Table 5; Algorithm 3 in Appendix C]
    "First, we used DSMBind to establish baseline binding affinities for the wild-type sequences. ... For each structure, we generated 10 different sequences and evaluated their predicted binding affinities using DSMBind. // Algorithm 3: 'Compute ŷ by p(ŝ1): ŷ = h_φ(p(ŝ1))', 'Approximate p(y|s1) ≈ −∇_{s1} ||y − ŷ||^2', 'Sample s1 ∼ p(ŝ1)p(y|ŝ1)'"

    The same regressor h_φ (DSMBind) is used on both sides of the experiment: Algorithm 3 steers sampling by the pseudo-gradient −∇_{s1}||y − ŷ||^2, moving the sequence toward what DSMBind scores closer to the target, and Section 4.3 then defines the reported 'affinity gain' as the fraction of designed sequences whose DSMBind-predicted affinity exceeds the wild-type DSMBind score. Success is therefore measured by the very scoring function that was used as the generator's reward. The 41.9% to 58.1% improvement is an expected consequence of optimizing against the evaluation metric and does not provide independent evidence that the designs have higher binding affinity; it is self-referential rather than a prediction about a held-out measure.

full rationale

The central inverse-folding claims (sequence recovery in Tables 1 and 3, foldability in Tables 2 and 4) are evaluated against external baselines such as LigandMPNN, ProteinMPNN, and PiFold, and against the external structure predictor Chai-1, so those claims are not circular. The multi-state ensemble experiment uses the same model with different inputs but evaluates with external Chai-1; no circularity is present there. The only self-referential step is in Section 4.3 and Algorithm 3: DSMBind is both the guidance regressor used to steer sampling and the evaluation metric used to measure 'affinity gain'. A generated sequence counts as a success if DSMBind's predicted affinity exceeds the wild-type DSMBind score, and the sampler is explicitly pushed along −∇_{s1}||y − ŷ||^2 computed from the same DSMBind. Hence the 41.9% to 58.1% improvement partly measures the optimizer's success on its own reward, not an independent validation of binding affinity. This is compounded by a correctness gap: Algorithm 3 is not well-defined as written, because h_φ receives a probability vector rather than a concrete sequence with side chains, and the pseudo-gradient on the discrete variable s1 is not a likelihood. That gap is a reproducibility concern rather than a circular equivalence. No load-bearing self-citation or imported uniqueness theorem was found. The score of 6 reflects that one prediction — guidance improving binding affinity — reduces by construction to its evaluation input, while the primary state-of-the-art inverse-folding result remains externally benchmarked.

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

The central claims rest on domain assumptions about the generative framework, the accuracy of external side-chain packing and structure/affinity predictors, and the validity of uniform averaging over conformational states. None of these is proved in the paper; they are adopted from prior work or heuristics.

free parameters (4)
  • purity threshold tau
    Hyperparameter in adaptive sampling (Algorithm 2); positions with denoiser probability above tau are sampled early. Value is not reported.
  • maximum iterations K
    Stopping criterion for adaptive sampling (Algorithm 2). Not reported.
  • time step Delta-t
    Step size in Euler sampling of the discrete flow (Equation 1). Not reported.
  • guidance target y = 10% improvement over wild-type predicted affinity
    Section 4.3: the target for guided generation was set to 10% above DSMBind's prediction for the wild-type sequence. This choice affects the guided results.
assumptions (6)
  • domain assumption A mask-interpolation discrete flow can model the one-to-many mapping from structure to sequence
    The whole method is built on the conditional flow p_{t|1} between mask and data, following Campbell et al. 2024. It is assumed that the denoiser can learn this mapping well enough to generate functional sequences.
  • domain assumption The denoiser network f_theta can accurately approximate p_{1|t}(s1|s_t, x, chi_t)
    Section 3.1: the training objective is cross-entropy on this conditional; the model's performance depends on network capacity and training data.
  • domain assumption Predicted side chains from PIPPACK are accurate enough to serve as conditioning context
    Section 3.1 and Algorithm 1: side chains for sampled positions are taken from an external packing network; if these are systematically wrong, the conditioning signal is corrupted. The paper does not report an ablation.
  • domain assumption Chai-1 without MSA is a valid proxy to measure foldability of designed sequences
    Section 4.1 foldability evaluation: TM-score >= 0.5 cutoff is used. This is an external predictor, and the cutoff is arbitrary.
  • domain assumption DSMBind's predicted binding affinity is a meaningful objective and metric
    Section 4.3: DSMBind is used both to guide generation and to evaluate affinity gain. No experimental validation is provided.
  • domain assumption Averaging denoiser probabilities over an ensemble of conformations produces a valid sampling distribution for multi-state design
    Algorithm 1: p(s1) = 1/N sum of p_n(s1). This linear average assumes all states contribute equally and ignores state-specific confidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of All-atom inverse protein folding through discrete flow matching." pith.science (2026). https://pith.science/paper/FP477FVJ

@misc{pith2026250714156,
  author       = {Pith},
  title        = {Pith review of: All-atom inverse protein folding through discrete flow matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FP477FVJ}},
  note         = {Machine review of arXiv:2507.14156}
}
read the original abstract

The recent breakthrough of AlphaFold3 in modeling complex biomolecular interactions, including those between proteins and ligands, nucleotides, or metal ions, creates new opportunities for protein design. In so-called inverse protein folding, the objective is to find a sequence of amino acids that adopts a target protein structure. Many inverse folding methods struggle to predict sequences for complexes that contain non-protein components, and perform poorly with complexes that adopt multiple structural states. To address these challenges, we present ADFLIP (All-atom Discrete FLow matching Inverse Protein folding), a generative model based on discrete flow-matching for designing protein sequences conditioned on all-atom structural contexts. ADFLIP progressively incorporates predicted amino acid side chains as structural context during sequence generation and enables the design of dynamic protein complexes through ensemble sampling across multiple structural states. Furthermore, ADFLIP implements training-free classifier guidance sampling, which allows the incorporation of arbitrary pre-trained models to optimise the designed sequence for desired protein properties. We evaluated the performance of ADFLIP on protein complexes with small-molecule ligands, nucleotides, or metal ions, including dynamic complexes for which structure ensembles were determined by nuclear magnetic resonance (NMR). Our model achieves state-of-the-art performance in single-structure and multi-structure inverse folding tasks, demonstrating excellent potential for all-atom protein design. The code is available at https://github.com/ykiiiiii/ADFLIP.

Figures

Figures reproduced from arXiv: 2507.14156 by the authors.

Figure 1
Figure 1. Overview of ADFLIP. ADFLIP is a flow matching generative framework for protein sequence design. Starting from t = 0, it progressively generates the protein sequence through sequential denoising using only the protein backbone and non-protein structure as inputs. The process begins with all amino acid positions masked (X) and gradually unmasks them to reveal the sequence. As amino acids are sampled, their correspondi… view at source ↗
Figure 2
Figure 2. Architecture of the ADFLIP denoiser. a, The information flow between residue features and atom features. Updating residue features from atom features involves averaging over all atoms in a residue; updating atom features from residue features requires scattering the same residue feature over all atoms in that residue. b, Schematic showing the main blocks of the graph neural network (GNN) encoder. c, Schematic showin… view at source ↗
Figure 3
Figure 3. Comparison of the inverse folding of three structures with nucleotides and ligands generated with ProteinMPNN, Lig￾andMPNN, and ADFLIP. The sequences are refolded (coloured) with Chai-1 (Chai Discovery, 2024) and compared to the starting structure (black outline). a, Comparison for a helix–loop–helix transcriptional activator protein (PDB ID: 1AM9, see Parraga et al., 1998), showing that ADFLIP’s integration of nucl… view at source ↗

Discussion (0). Continue with ORCID 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. UMA-Inverse: Ligand-Conditioned Protein Inverse Folding with a Distogram-Supervised Dense Pair Encoder

    q-bio.BM 2026-07 conditional novelty 5.5 of 10

    A ~3.3M-parameter dense PairMixer inverse folder trails LigandMPNN on interface recovery while propagating ligand signal to distal residues where local graphs lose it.

Reference graph

Works this paper leans on

49 extracted references · 33 canonical work pages · cited by 1 Pith paper

  1. [1]

    write newline

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

  2. [2]

    J., Bambrick, J., et al

    Abramson, J., Adler, J., Dunger, J., Evans, R., Green, T., Pritzel, A., Ronneberger, O., Willmore, L., Ballard, A. J., Bambrick, J., et al. Accurate structure prediction of biomolecular interactions with alphafold 3. Nature, pp.\ 1--3, 2024

    2024

  3. [3]

    D., Ho, J., Tarlow, D., and Van Den Berg, R

    Austin, J., Johnson, D. D., Ho, J., Tarlow, D., and Van Den Berg, R. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 34: 0 17981--17993, 2021

    2021

  4. [4]

    R., Watson, J

    Bennett, N. R., Watson, J. L., Ragotte, R. J., Borst, A. J., See, D. L., Weidle, C., Biswas, R., Shrock, E. L., Leung, P. J., Huang, B., et al. Atomically accurate de novo design of single-domain antibodies. bioRxiv, 2024

    work page 2024

  5. [5]

    A continuous time framework for discrete denoising models

    Campbell, A., Benton, J., De Bortoli, V., Rainforth, T., Deligiannidis, G., and Doucet, A. A continuous time framework for discrete denoising models. Advances in Neural Information Processing Systems, 35: 0 28266--28279, 2022

    work page 2022

  6. [6]

    Generative flows on discrete state-spaces: Enabling multimodal flows with applications to protein co-design

    Campbell, A., Yim, J., Barzilay, R., Rainforth, T., and Jaakkola, T. Generative flows on discrete state-spaces: Enabling multimodal flows with applications to protein co-design. In Forty-first International Conference on Machine Learning, 2024

    work page 2024

  7. [7]

    Chai-1: Decoding the molecular interactions of life

    Chai Discovery . Chai-1: Decoding the molecular interactions of life. bioRxiv, 2024. doi:10.1101/2024.10.10.615955. URL https://www.biorxiv.org/content/early/2024/10/11/2024.10.10.615955

    doi:10.1101/2024.10.10.615955 2024

  8. [8]

    J., Milles, L

    Dauparas, J., Anishchenko, I., Bennett, N., Bai, H., Ragotte, R. J., Milles, L. F., Wicky, B. I., Courbet, A., de Haas, R. J., Bethel, N., et al. Robust deep learning--based protein sequence design using proteinmpnn. Science, 378 0 (6615): 0 49--56, 2022

    work page 2022

Show all 49 references
  1. [9]

    R., Pecoraro, R., An, L., Anishchenko, I., Glasscock, C., and Baker, D

    Dauparas, J., Lee, G. R., Pecoraro, R., An, L., Anishchenko, I., Glasscock, C., and Baker, D. Atomic context-conditioned protein sequence design using ligandmpnn. Biorxiv, pp.\ 2023--12, 2023

    2023

  2. [10]

    Visual feature extraction by a multilayered network of analog threshold elements

    Fukushima, K. Visual feature extraction by a multilayered network of analog threshold elements. IEEE Transactions on Systems Science and Cybernetics, 5 0 (4): 0 322--333, 1969

    1969

  3. [11]

    Gao, Z., Tan, C., Chen, X., Zhang, Y., Xia, J., Li, S., and Li, S. Z. Kw-design: Pushing the limit of protein design via knowledge refinement. In The Twelfth International Conference on Learning Representations, 2023 a

    2023

  4. [12]

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

    2023

  5. [13]

    T., Synnaeve, G., Adi, Y., and Lipman, Y

    Gat, I., Remez, T., Shaul, N., Kreuk, F., Chen, R. T., Synnaeve, G., Adi, Y., and Lipman, Y. Discrete flow matching. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  6. [14]

    J., Goreshnik, I., Coventry, B., Bera, A

    Gl \"o gl, M., Krishnakumar, A., Ragotte, R. J., Goreshnik, I., Coventry, B., Bera, A. K., Kang, A., Joyce, E., Ahn, G., Huang, B., et al. Target-conditioned diffusion generates potent tnfr superfamily antagonists and agonists. Science, 386 0 (6726): 0 1154--1161, 2024

    2024

  7. [15]

    and Gimpel, K

    Hendrycks, D. and Gimpel, K. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016

    2016 arXiv

  8. [16]

    Argmax flows and multinomial diffusion: Learning categorical distributions

    Hoogeboom, E., Nielsen, D., Jaini, P., Forr \'e , P., and Welling, M. Argmax flows and multinomial diffusion: Learning categorical distributions. Advances in Neural Information Processing Systems, 34: 0 12454--12465, 2021

    2021

  9. [17]

    Learning inverse folding from millions of predicted structures

    Hsu, C., Verkuil, R., Liu, J., Lin, Z., Hie, B., Sercu, T., Lerer, A., and Rives, A. Learning inverse folding from millions of predicted structures. ICML, 2022. doi:10.1101/2022.04.10.487779. URL https://www.biorxiv.org/content/early/2022/04/10/2022.04.10.487779

    2022 doi

  10. [18]

    Protein-nucleic acid complex modeling with frame averaging transformer

    Huang, T., Song, Z., Ying, R., and Jin, W. Protein-nucleic acid complex modeling with frame averaging transformer. arXiv preprint arXiv:2406.09586, 2024

    2024 arXiv

  11. [19]

    Generative models for graph-based protein design

    Ingraham, J., Garg, V., Barzilay, R., and Jaakkola, T. Generative models for graph-based protein design. Advances in neural information processing systems, 32, 2019

    2019

  12. [20]

    Unsupervised protein-ligand binding energy prediction via neural euler's rotation equation

    Jin, W., Sarkizova, S., Chen, X., Hacohen, N., and Uhler, C. Unsupervised protein-ligand binding energy prediction via neural euler's rotation equation. Advances in Neural Information Processing Systems, 36, 2024

    2024

  13. [21]

    K., Jamasb, A

    Joshi, C. K., Jamasb, A. R., Vi \ n as, R., Harris, C., Mathis, S., Morehead, A., Anand, R., and Li \`o , P. grnade: Geometric deep learning for 3d rna inverse design. bioRxiv, 2024

    2024

  14. [22]

    Highly accurate protein structure prediction with alphafold

    Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R., Z \' dek, A., Potapenko, A., et al. Highly accurate protein structure prediction with alphafold. nature, 596 0 (7873): 0 583--589, 2021

    2021

  15. [23]

    J., Anischanka, I., Sumida, K

    Lauko, A., Pellock, S. J., Anischanka, I., Sumida, K. H., Juergens, D., Ahern, W., Shida, A., Hunt, A., Kalvet, I., Norn, C., et al. Computational design of serine hydrolases. bioRxiv, 2024

    2024

  16. [24]

    and Zhang, Y

    Mukherjee, S. and Zhang, Y. Mm-align: a quick algorithm for aligning multiple-chain protein complex structures using iterative dynamic programming. Nucleic acids research, 37 0 (11): 0 e83--e83, 2009

    2009

  17. [25]

    Glide: Towards photorealistic image generation and editing with text-guided diffusion models

    Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., and Chen, M. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021

    2021 arXiv

  18. [26]

    S., Lunegova, D

    Nikolaev, A. S., Lunegova, D. A., Raevskii, R. I., Shishkin, P. E., Remeeva, A. A., Ge, B., Maksimov, E. G., Gushchin, I. Y., and Sluchanko, N. N. Re-engineering of a carotenoid-binding protein based on nmr structure. Protein Science, 33 0 (12): 0 e5216, 2024

    2024

  19. [27]

    and Jin, W

    Nori, D. and Jin, W. Rnaflow: Rna structure & sequence design via inverse folding-based flow matching. In Forty-first International Conference on Machine Learning, 2024

    2024

  20. [28]

    Spin2: Predicting sequence profiles from protein structures using deep neural networks

    O'Connell, J., Li, Z., Hanson, J., Heffernan, R., Lyons, J., Paliwal, K., Dehzangi, A., Yang, Y., and Zhou, Y. Spin2: Predicting sequence profiles from protein structures using deep neural networks. Proteins: Structure, Function, and Bioinformatics, 86 0 (6): 0 629--633, 2018

    2018

  21. [29]

    Bindcraft: one-shot design of functional protein binders

    Pacesa, M., Nickel, L., Schellhaas, C., Schmidt, J., Pyatova, E., Kissling, L., Barendse, P., Choudhury, J., Kapoor, S., Alcaraz-Serna, A., et al. Bindcraft: one-shot design of functional protein binders. bioRxiv, pp.\ 2024--09, 2024

    2024

  22. [30]

    Parraga, A., Bellsolell, L., Ferre-D'Amare, A., and Burley, S. K. Co-crystal structure of sterol regulatory element binding protein 1a at 2.3 resolution. Structure, 6 0 (5): 0 661--672, 1998

    1998

  23. [31]

    and Xie, S

    Peebles, W. and Xie, S. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 4195--4205, 2023

    2023

  24. [32]

    Z., Bezemek, Z., Patel, S., Rector-Brooks, J., Yao, S., Tong, A., and Chatterjee, P

    Peng, F. Z., Bezemek, Z., Patel, S., Rector-Brooks, J., Yao, S., Tong, A., and Chatterjee, P. Path planning for masked diffusion model sampling. arXiv preprint arXiv:2502.03540, 2025

    2025

  25. [33]

    J., and Lipman, Y

    Puny, O., Atzmon, M., Ben-Hamu, H., Misra, I., Grover, A., Smith, E. J., and Lipman, Y. Frame averaging for invariant and equivariant network design. arXiv preprint arXiv:2110.03336, 2021

    2021 arXiv

  26. [34]

    Randolph, N. Z. and Kuhlman, B. Invariant point message passing for protein side chain packing. Proteins: Structure, Function, and Bioinformatics, pp.\ 1--14, 2024. doi:10.1002/prot.26705

    2024 doi

  27. [35]

    W., Landthaler, M., Shub, D

    Shen, B. W., Landthaler, M., Shub, D. A., and Stoddard, B. L. Dna binding and cleavage by the hnh homing endonuclease i-hmui. Journal of molecular biology, 342 0 (1): 0 43--56, 2004

    2004

  28. [36]

    P., Kumar, A., Ermon, S., and Poole, B

    Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., and Poole, B. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=PxTIG12RRHS

    2021

  29. [37]

    Surfpro: Functional protein design based on continuous surface

    Song, Z., Huang, T., Li, L., and Jin, W. Surfpro: Functional protein design based on continuous surface. In Forty-first International Conference on Machine Learning, 2024

    2024

  30. [38]

    and S \"o ding, J

    Steinegger, M. and S \"o ding, J. Mmseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nature biotechnology, 35 0 (11): 0 1026--1028, 2017

    2017

  31. [39]

    Saprot: Protein language modeling with structure-aware vocabulary

    Su, J., Han, C., Zhou, Y., Shan, J., Zhou, X., and Yuan, F. Saprot: Protein language modeling with structure-aware vocabulary. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=6MRm3G4NiU

    2024

  32. [40]

    and Ward, T

    Trindler, C. and Ward, T. R. Artificial metalloenzymes. Effects of Nanoconfinement on Catalysis, pp.\ 49--82, 2017

    2017

  33. [41]

    N., Kaiser, L., and Polosukhin, I

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., and Polosukhin, I. Attention is all you need, 2017. URL https://arxiv.org/abs/1706.03762

    2017 arXiv

  34. [42]

    Digress: Discrete denoising diffusion for graph generation

    Vignac, C., Krawczuk, I., Siraudin, A., Wang, B., Cevher, V., and Frossard, P. Digress: Discrete denoising diffusion for graph generation. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=UaAD-Nu86WX

    2023

  35. [43]

    Z., and Qi, Y

    Wang, J., Cao, H., Zhang, J. Z., and Qi, Y. Computational protein design with deep learning neural networks. Scientific reports, 8 0 (1): 0 1--9, 2018

    2018

  36. [44]

    Diffusion language models are versatile protein learners

    Wang, X., Zheng, Z., YE, F., Xue, D., Huang, S., and Gu, Q. Diffusion language models are versatile protein learners. In Forty-first International Conference on Machine Learning, 2024. URL https://openreview.net/forum?id=NUAbSFqyqb

    2024

  37. [45]

    L., Juergens, D., Bennett, N

    Watson, J. L., Juergens, D., Bennett, N. R., Trippe, B. L., Yim, J., Eisenach, H. E., Ahern, W., Borst, A. J., Ragotte, R. J., Milles, L. F., et al. De novo design of protein structure and function with rfdiffusion. Nature, 620 0 (7976): 0 1089--1100, 2023

    2023

  38. [46]

    Graph denoising diffusion for inverse protein folding

    Yi, K., Zhou, B., Shen, Y., Li \`o , P., and Wang, Y. Graph denoising diffusion for inverse protein folding. Advances in Neural Information Processing Systems, 36, 2024

    2024

  39. [47]

    J., Brown, D., Burns-Kurtis, C

    Young, R. J., Brown, D., Burns-Kurtis, C. L., Chan, C., Convery, M. A., Hubbard, J. A., Kelly, H. A., Pateman, A. J., Patikis, A., Senger, S., et al. Selective and dual action orally active inhibitors of thrombin and factor xa. Bioorganic & medicinal chemistry letters, 17 0 (1...

    2007

  40. [48]

    Structure-informed language models are protein designers

    Zheng, Z., Deng, Y., Xue, D., Zhou, Y., Ye, F., and Gu, Q. Structure-informed language models are protein designers. In International conference on machine learning, pp.\ 42317--42338. PMLR, 2023

    2023

  41. [49]

    Bridge-if: Learning inverse protein folding with markov bridges

    Zhu, Y., Wu, J., Li, Q., Yan, J., Yin, M., Wu, W., Li, M., Ye, J., Wang, Z., and Wu, J. Bridge-if: Learning inverse protein folding with markov bridges. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

Pith tools

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