Pith. sign in

REVIEW 4 major objections 5 minor 50 references

Cross-Chirality Generalization by Axial Vectors for Hetero-Chiral Protein-Peptide Interaction Design

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read A generative model trained only on L–L complexes can design D-peptide binders for L-protein targets, the paper argues, by injecting axial vectors into the equivariant features and inverting both target and output.

desk verdict First wet-lab-validated generative AI for D-peptide binders, with a clean chirality-injection idea and honest theory; but the theory's key assumption is unverified on the trained net and the wet-lab evidence is thin. read the letter →

arxiv 2602.20176 v3 pith:MOCPZ3WS submitted 2026-02-13 q-bio.BM cs.LG

Ziyi Yang , Zitong Tian , Yinjun Jia , Tianyi Zhang , Jiqing Zheng , Hao Wang , Yubu Su , Juncai He
show 2 more authors
Lei Liu Yanyan Lan
This is my paper
classification q-bio.BMcs.LG
keywords chiralityD-peptidedesignaxialvectorsequivariantneuralnetworkslatentdiffusionprotein-peptideinteractionmirror-imagezero-shotgeneralization
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

This paper tries to establish that a neural peptide generator trained only on homo-chiral (L–L) protein–peptide complexes can, without retraining, design D-peptide binders for natural L-protein targets. The trick is to add axial vector features—quantities that stay unchanged under mirror reflection—to the polar vector features of an E(3)-equivariant network. This breaks inversion symmetry while preserving rotation and translation equivariance, making the model chirality-aware yet stable enough to transfer. The authors implement the idea in PepMirror, a latent diffusion model, and report both in silico gains over existing peptide design tools and a wet-lab binder, D-1412, against CD38 with KD ≈ 10 µM. A sympathetic reader would see this as evidence that cross-chirality design can be achieved zero-shot through the right representation rather than through new training data.

What carries the argument

Axial Feature Injection (AFI): a channel-wise linear mixing of polar vector features v with axial vector features a, written as ev = A^T v + B^T a. The axial channels are built from low-order geometric couplings—u×v, (w·(u×v))w, and (u·v)(u×v)—which remain unchanged under mirror inversion. This changes the network's equivariance class from E(3) to SE(3), making it chirality-aware while preserving rotation and translation equivariance, and the mixing coefficients are what carry the theoretical guarantee.

What would settle it

Compute the correlation matrix C = ⟨v_k, a_l⟩ on actual PepMirror encoder outputs over LNR test residues: if its Frobenius norm is near zero for most residues, the predicted inversion-induced discrepancy should be negligible, contradicting the theorem's applicability. Equivalently, one could initialize the mixing coefficients to be orthogonal and check whether the L/D chirality discrepancy drops to zero.

Watch

Extended reading notes

Core claim

The central claim is that injecting axial features into polar vector features suffices to achieve cross-chirality generalization from homo-chiral L–L training data to hetero-chiral D–L design tasks. The paper proves that without axial features an E(3)-equivariant network cannot distinguish a structure from its mirror image, while with channel-wise mixed axial features the latent-code norm differs between X and −X with high probability, provided the polar–axial correlations are non-degenerate. It further argues that latent codes of enantiomers stay close—so the model still recognizes them as the same residue type—while codes of different residue types separate, which keeps conditional diffusi

Load-bearing premise

The load-bearing premise is that, after training, the polar and axial feature channels keep a non-negligible, non-orthogonal overlap so that inversion changes the latent code; the paper assumes this rather than measuring it directly, and without it the theoretical guarantee collapses.

Editorial extensions

If this is right

  • A model trained exclusively on L–L complexes can be asked to design D–L interfaces through a deterministic inversion of target and output, with no hetero-chiral training samples.
  • Chirality control becomes a representational property: residue-level chirality is preserved at roughly 99.9% after relaxation, unlike E(3)-equivariant baselines that invert the entire structure with the target.
  • The usual L-to-D performance cliff shrinks: PepMirror keeps affinity, success rate, and interface-energy improvement on D-tasks close to its L-task numbers.
  • The pipeline yields a first wet-lab validated generative-AI D-peptide binder: D-1412, a 10-mer against CD38, with KD ≈ 10 µM consistent between kinetic and steady-state fits.

Reading between the lines

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

  • Editorial extension: the same AFI plug-in should transfer to other E(3)-equivariant structure generators; the prerequisite is only that their latent codes cluster by residue type, which can be checked with the paper's distance-heatmap test.
  • The theory implies a tuning knob: the chirality discrepancy scales with polar–axial correlation, so designers could deliberately shape mixing coefficients to maximize effective rank rather than relying on generic initialization.
  • The stability argument suggests a stress test: evaluate per-residue chirality consistency on achiral or nearly achiral positions, such as glycine, and on residues where the 'enantiomer closer than different types' assumption may fail; those are the likely first failure modes.
  • If the mechanism generalizes, mirror-image display could be largely replaced by in silico inversion for many therapeutic targets, making D-peptide discovery faster and cheaper—an inference beyond the paper's own evidence.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes AFI (Axial Feature Injection), a lightweight modification that injects axial vector features into the polar vector features of an E(3)-equivariant encoder, making the model SE(3)-equivariant and chirality-aware. The authors integrate AFI into a latent diffusion peptide design framework (PepMirror) and argue that, by two-step spatial inversion, the model can generate D-peptide binders for L-protein targets despite being trained only on homo-chiral L–L complexes. The theoretical component proves a generic lower bound on the inversion-induced latent discrepancy under random Gaussian mixing coefficients and a non-degenerate polar–axial correlation assumption. The empirical component reports latent-space analyses, in silico comparisons against several peptide design baselines, and a wet-lab validation identifying one D-peptide binder (D-1412) against CD38 with KD ≈ 10 µM.

Significance. If the claims hold, this is the first wet-lab validated generative-AI system for de novo D-peptide binder design, and AFI is a simple, practical mechanism for adding chirality awareness to efficient E(3)-equivariant backbones. The paper has notable strengths: the two-step inversion protocol does not train on hetero-chiral D–L data, so the setting is genuinely zero-shot; the theoretical derivation is explicit rather than a black-box claim; the in silico evaluation covers a broad set of baselines and includes a careful diagnosis of Rosetta's enantiomer inconsistency, with Vina justified as an alternative; and code is released. The wet-lab result, while modest, is a plausible proof-of-concept. However, some load-bearing theoretical assumptions are not verified on the trained model, and the empirical evidence is partly single-seed and limited to one binder.

major comments (4)
  1. [§3.3, Theorem 3.1 / Appendix B.3, Assumption B.4] Theorem B.8's lower bound depends critically on Assumption B.4 (||C||_F ≥ τ, reff ≥ r0), but this assumption is never measured on the trained PepMirror encoder. The paper itself notes in Eqs. (6)–(7) that when polar–axial orthogonality (v·a = 0) holds, the discrepancy vanishes, and the cross and commutator axial features (Algorithm 1) are by construction orthogonal to their source polar channels. The proof is a generic statement over Gaussian mixing coefficients at initialization (Assumption B.2); it does not cover trained coefficients. The sentence in Appendix B.3 saying that 'learned parameters fall into the discrepancy-inducing regime with high probability' is therefore unsupported. Please verify B.4 empirically on the trained model (report C, reff, and Δ per channel/layer), or explicitly reframe the theorem as an initialization/architecture-level guarantee rather than a guarantee abo
  2. [§3.4, Eq. (14)] Eq. (14), d(X,X') > d(X,−X), is an assumption, not a consequence of the architecture. The only support offered is the indirect Tanimoto shape-similarity analogy in Fig. 2, but amino-acid shape similarity is not the trained latent metric, and the step from Eq. (14) to Eq. (15) additionally assumes that φ preserves the ordering of embedding distances. Fig. 4 provides direct empirical evidence for this ordering in the trained encoder, but it is single-model and mean-pooled. Please state explicitly that Eq. (14) is an empirical assumption and add direct verification on the trained encoder, e.g., report within-type L/D distances versus between-type distances with error bars.
  3. [§4.2.3, Tables 1–2; §C.4] All in silico results are single-seed point estimates (random seed 12) with no error bars or significance tests. The central empirical claims—PepMirror consistently outperforms baselines on D-peptide design and has the smallest L-to-D degradation—rely on differences that could be within run-to-run variability (e.g., Table 2 D-task Avg: −4.15 vs −3.38; IMP D: 63.44 vs 50.54). Please report results over multiple seeds as mean ± std, or otherwise quantify variability. This is important because the paper's headline is a cross-chirality generalization advantage, not merely a new architecture.
  4. [§4.3 Wet-lab validation] The claim of 'efficacy in wet-lab validation' rests on a single peptide (D-1412, KD ≈ 10 µM) out of 12 tested candidates, with no control confirming that binding requires the D-stereochemistry. For example, the L-enantiomer of D-1412 could be tested as a negative control, or chiral purity could be established by chiral HPLC/optical rotation. Without such a control, the observed signal is consistent with a non-specific or chirality-independent interaction. Please add an enantiomer control or temper the claim to 'a D-peptide binder candidate with micromolar affinity was identified, but stereochemical specificity remains to be confirmed.'
minor comments (5)
  1. [Throughout] Typos: 'immunogenecity' (Section 1) should be 'immunogenicity'; 'chiraliity' (§4.2.2); 'Ramachadran' (Fig. S7 caption and §A.3); 'devided' (§C.6); Table 2 caption contains 'of of generated'.
  2. [Figure 3] The box plots summarize distances across samples but do not show the distribution across channels/layers or across seeds. Reporting these would directly address whether the nonzero discrepancy is concentrated in a few channels or is a robust effect.
  3. [§3.3 vs Algorithm 1] The theoretical notation uses channel-wise mixing vectors A_k, B_k, while Algorithm 1 implements a single unbiased Linear layer applied to the concatenated polar–axial channels. Please clarify the exact correspondence between the two descriptions, including how the linear layer's rows relate to A_k and B_k.
  4. [§4.2.2 and Appendix A.4] The Rosetta enantiomer inconsistency table (Table S4) is important because it justifies the choice of Vina over Rosetta as the primary interface-affinity metric. Consider moving a concise version into the main text or summarizing the numbers there.
  5. [§3.4, Eq. (16)] The diffusion SDE in Eq. (16) is written in forward time, but the proof in Appendix B.4 couples two solutions backwards from time T. The notation is understandable, but a one-sentence clarification that the SDE is run backwards in practice would help.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central theory is a conditional theorem, and the D-binder pipeline is validated against external wet-lab data.

full rationale

The derivation chain is not circular. The two-step inversion protocol (Section 3.1, Gb = P(fθ(P(Gt)))) is a geometric identity, not a fitted prediction, and no hetero-chiral D–L data are used in training. Theorem 3.1 / Theorem B.8 derives chirality-induced latent discrepancy from explicitly stated Assumptions B.2–B.4; the proof is a genuine probabilistic derivation, and the paper explicitly acknowledges the A=B=I orthogonal case (Eqs. 6–7) where the effect vanishes, so the theorem is conditional rather than tautological. The stability argument in Section 3.4 relies on the explicit assumption Eq. (14), which is not derived from the model; this is a robustness gap, not a circular reduction, because the paper does not claim to derive Eq. (14) from the target conclusion. The latent-space experiments and wet-lab CD38 binder validation provide independent, external evidence, and the framework citations (UniMoMo, EPT, D-Flow) are prior work by other author groups rather than self-citations that carry the argument. The paper itself concedes the theory is not end-to-end, stating it focuses on 'the feature-mixing mechanism under simplifying assumptions, rather than providing an end-to-end theory of the trained network.' Unverified assumptions such as B.4 and Eq. (14) are correctness risks, not circularity.

Assumptions & free parameters 2 free parameters · 7 assumptions · 0 invented entities

The central claim rests less on new physics than on a set of explicitly stated but largely unverified distributional and separation assumptions. The proven mathematics covers generic random initialization and bounded activations, not the specific trained PepMirror network. The wet-lab result is a single micromolar binder, which is a weak anchor for the phrase 'efficacy in wet-lab validation.'

free parameters (2)
  • Channel-mixing coefficients A_k, B_k = Learned during training; no explicit values given
    The AFI mixing in Eq. (2) introduces learned matrices. Theorem B.8 requires them to behave like random sub-Gaussian draws with non-degenerate polar–axial correlation to guarantee a chirality discrepancy; the actual trained values are not checked against Assumption B.4.
  • Wet-lab candidate filtering thresholds = absBSA>400, 0.20<relBSA<0.85, vina<-4.0, BuriedHBonds<10, ShapeComplementarity>0.65, ElectrostaticComplementarity>0.65,
    These hand-set thresholds in Appendix C.6 select the 12 (or 6) candidates for synthesis from 5,000 generated peptides. They influence the reported wet-lab hit rate and are not derived from any external principle.
assumptions (7)
  • domain assumption Assumption B.2: mixing coefficients A_k, B_k are independent Gaussian/sub-Gaussian
    Used in Lemma B.6 to obtain the high-probability lower bound on the bilinear form; central to Theorem 3.1 but not verified for trained parameters.
  • domain assumption Assumption B.3: bounded polar and axial vector features (∥v_i,:∥_F ≤ S_v, ∥a_i,:∥_F ≤ S_a)
    Required to upper-bound the denominator in Theorem B.8; plausible for normalized network features but not checked numerically.
  • domain assumption Assumption B.4: non-degenerate polar–axial correlation ∥C∥_F ≥ τ and efficient rank reff ≥ r0
    This is the load-bearing premise: if the polar and axial channels are orthogonal, the paper itself shows the chirality discrepancy can vanish (A=B=I example in Section 3.3). The proof gives a generic random-initialization guarantee, not a verification on the trained PepMirror.
  • domain assumption Eq. (14): inter-type embedding distance exceeds inversion distance, d(X,X′) > d(X,−X)
    Needed to conclude that L/D forms of the same amino acid form tight clusters separated from other types. Support is indirect, via Tanimoto shape similarity of amino acid conformers, not the model embedding itself.
  • domain assumption Assumption 3.3: Lipschitz drift of the diffusion model in state and condition
    Required for the Wasserstein stability theorem (Theorem 3.4) that links close latent codes to close generated distributions. Standard but unverified for the trained model.
  • domain assumption Coercivity of φ in the norm coordinate (Corollary B.9)
    Needed to convert the vector-norm discrepancy into a scalar latent-code discrepancy. Remark B.10 argues genericity, but it is not demonstrated for the trained network.
  • standard math Standard SO(3) representation theory, Clebsch–Gordan decomposition, Hanson–Wright and Grönwall inequalities
    These are standard mathematical tools used to construct the axial features and prove the theorems; no controversy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cross-Chirality Generalization by Axial Vectors for Hetero-Chiral Protein-Peptide Interaction Design." pith.science (2026). https://pith.science/paper/MOCPZ3WS

@misc{pith2026260220176,
  author       = {Pith},
  title        = {Pith review of: Cross-Chirality Generalization by Axial Vectors for Hetero-Chiral Protein-Peptide Interaction Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MOCPZ3WS}},
  note         = {Machine review of arXiv:2602.20176}
}
abstract

D-peptide binders targeting L-proteins have promising therapeutic potential. Despite rapid advances in machine learning-based target-conditioned peptide design, generating D-peptide binders remains largely unexplored. In this work, we show that by injecting axial features to $E(3)$-equivariant (polar) vector features, it is feasible to achieve cross-chirality generalization from homo-chiral (L--L) training data to hetero-chiral (D--L) design tasks. By implementing this method within a latent diffusion model, we achieved D-peptide binder design that not only outperforms existing tools in \textit{in silico} benchmarks, but also demonstrates efficacy in wet-lab validation. To our knowledge, our approach represents the first wet-lab validated generative AI for the \textit{de novo} design of D-peptide binders, offering new perspectives on handling chirality in protein design. Codes are available at https://github.com/YZY010418/PepMirror

Figures

Figures reproduced from arXiv: 2602.20176 by the authors.

Figure 1
Figure 1. We use chiral-sensitive model PepMirror to design D-peptides by flipping. PepMirror is a latent diffusion model using AFI-EPT, which injects axial vector features to learn the chirality. Axial vectors are invariant under the spatial inversion, and we give three direct constructions. The commutator feature (third) captures higher frequency information of the angle between u, v. where H(X) ∈ R N×K denotes the E(3)-inv… view at source ↗
Figure 2
Figure 2. The max-pooled pair-wise Tanimoto shape similarity between L/D amino acids. The similarities between an amino acid and its enantiomer are among the highest compared with similaries between different amino acids. Because and similarities between the same amino acid or between ”D-Gly” and ”L-Gly” are 1.0 by definition, we excluded these entries in the heatmap. Diffusion stability We establish a continuity theorem on c… view at source ↗
Figure 3
Figure 3. Latent-code distances between each amino acid and its in￾verted counterpart (X vs. −X) for L- and D-forms across different encoder variants. All and Peptide refer to UniMoMo without AFI trained on different datasets (see section 4), and the other three are equipped with AFI based on different axial features. Distances are summarized as box plots over all amino acids. Encoders equipped with AFI exhibit a non-negligib… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Left: t-SNE of 20 types of amino acids including both L and D chirality. As t-SNE cannot keep distance, we plot the heatmap (right) of mean pairwise latent-code distances among 40 amino-acid classes including 20 L amino acids and 20 D amino acids. The three diagonals a…
Figure 5
Figure 5. Figure 5: The identified D-peptide binder against CD38. Left: Complex structure of D-1412 and CD38 generated by PepMirror (cross), where multiple interactions can be identified. Middle: Stacked curves of association and dissociation under different concentrations with kinetic fi…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 4 linked inside Pith

  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]

    Adams, K., Pattanaik, L., and Coley, C. W. Learning 3d representations of molecular chirality with invariance to bond rotations. arXiv preprint arXiv:2110.04383, 2021

    arXiv 2021

  3. [3]

    Blackmond, D. G. The origin of biological homochirality. Cold Spring Harbor perspectives in biology, 2 0 (5): 0 a002147, 2010

    2010

  4. [4]

    S., Jude, K

    Cao, L., Coventry, B., Goreshnik, I., Huang, B., Sheffler, W., Park, J. S., Jude, K. M., Markovi \'c , I., Kadam, R. U., Verschueren, K. H., et al. Design of protein-binding proteins from the target structure alone. Nature, 605 0 (7910): 0 551--560, 2022

    2022

  5. [5]

    Blocking of the pd-1/pd-l1 interaction by ad-peptide antagonist for cancer immunotherapy

    Chang, H.-N., Liu, B.-Y., Qi, Y.-K., Zhou, Y., Chen, Y.-P., Pan, K.-M., Li, W.-W., Zhou, X.-M., Ma, W.-W., Fu, C.-Y., et al. Blocking of the pd-1/pd-l1 interaction by ad-peptide antagonist for cancer immunotherapy. Angewandte Chemie International Edition, 54 0 (40): 0 11760--11764, 2015

    2015

  6. [6]

    P., and Geiger, A

    Coors, B., Condurache, A. P., and Geiger, A. Spherenet: Learning spherical representations for detection and classification in omnidirectional images. In Proceedings of the European conference on computer vision (ECCV), pp.\ 518--533, 2018

    2018

  7. [7]

    Pretrained graph transformers for chirality-related tasks

    Du, Y., Chen, J., and Schwaller, P. Pretrained graph transformers for chirality-related tasks. In AI for Accelerated Materials Design-Vienna, 2024

    2024

  8. [8]

    findr: A web server for in silico d-peptide ligand identification

    Engel, H., Guischard, F., Krause, F., Nandy, J., Kaas, P., Hoefflin, N., Koehn, M., Kilb, N., Voigt, K., Wolf, S., et al. findr: A web server for in silico d-peptide ligand identification. Synthetic and Systems Biotechnology, 6 0 (4): 0 402--413, 2021

    2021

Show all 50 references
  1. [9]

    Foster, M. S. Rotations in 3d, so (3) , and su (2) , version 2.1. Lecture notes, Rice University, 2021

    2021

  2. [10]

    SE (3)- Transformers : 3D Roto-Translation Equivariant Attention Networks

    Fuchs, F., Worrall, D., Fischer, V., and Welling, M. SE (3)- Transformers : 3D Roto-Translation Equivariant Attention Networks . In Advances in Neural Information Processing Systems , volume 33, pp.\ 1970--1981. Curran Associates, Inc., 2020

    1970

  3. [11]

    and Harris, J

    Fulton, W. and Harris, J. Representation Theory , volume 129 of Graduate Texts in Mathematics . Springer New York, New York, NY, 2004. ISBN 978-3-540-00539-1 978-1-4612-0979-9

    2004

  4. [12]

    Chienn: Embracing molecular chirality with graph neural networks

    Gai \'n ski, P., Koziarski, M., Tabor, J., and \'S mieja, M. Chienn: Embracing molecular chirality with graph neural networks. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pp.\ 36--52. Springer, 2023

    2023

  5. [13]

    A., Wang, K

    Garton, M., Nim, S., Stone, T. A., Wang, K. E., Deber, C. M., and Kim, P. M. Method to generate highly stable d-amino acid analogs of bioactive helical peptides using a mirror image of the entire pdb. Proceedings of the National Academy of Sciences, 115 0 (7): 0 1505--1510, 2018

    2018

  6. [14]

    and Smidt, T

    Geiger, M. and Smidt, T. E3nn: Euclidean Neural Networks , 2022

    2022

  7. [15]

    Hall, B. C. Lie Groups , Lie Algebras , and Representations : An Elementary Introduction , volume 222 of Graduate Texts in Mathematics . Springer International Publishing, Cham, 2015. ISBN 978-3-319-13466-6 978-3-319-13467-3

    2015

  8. [16]

    A survey of geometric graph neural networks: Data structures, models and applications

    Han, J., Cen, J., Wu, L., Li, Z., Kong, X., Jiao, R., Yu, Z., Xu, T., Wu, F., Wang, Z., et al. A survey of geometric graph neural networks: Data structures, models and applications. Frontiers of Computer Science, 19 0 (11): 0 1911375, 2025

    2025

  9. [17]

    Classical Electrodynamics

    Jackson, J. Classical Electrodynamics. Wiley, 2021. ISBN 978-1-119-77076-3

    2021

  10. [18]

    Equivariant pretrained transformer for unified geometric learning on multi-domain 3d molecules

    Jiao, R., Kong, X., Yu, Z., Huang, W., and Liu, Y. Equivariant pretrained transformer for unified geometric learning on multi-domain 3d molecules. In ICLR 2024 Workshop on Generative and Experimental Perspectives for Biomolecular Design, 2024

    2024

  11. [19]

    J., and Dror, R

    Jing, B., Eismann, S., Suriana, P., Townshend, R. J., and Dror, R. Learning from protein structure with geometric vector perceptrons. arXiv preprint arXiv:2009.01411, 2020

    2009 arXiv

  12. [20]

    Highly accurate protein structure prediction with AlphaFold

    Jumper, J., Evans, R., Pritzel, A., et al. Highly accurate protein structure prediction with AlphaFold . Nature, 596: 0 583--589, 2021

    2021

  13. [21]

    Kingma, D. P. and Welling, M. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013

    2013 arXiv

  14. [22]

    O., Willett, E., Martinkus, K., Kleinhenz, J., Makowski, E., Watkins, A., Gligorijevic, V., Bonneau, R., and Saremi, S

    Kirchmeyer, M., Pinheiro, P. O., Willett, E., Martinkus, K., Kleinhenz, J., Makowski, E., Watkins, A., Gligorijevic, V., Bonneau, R., and Saremi, S. Unified all-atom molecule generation with neural fields. In The 39th Annual Conference on Neural Information Processing Systems, 2025

    2025

  15. [23]

    Full-atom peptide design with geometric latent diffusion

    Kong, X., Jia, Y., Huang, W., and Liu, Y. Full-atom peptide design with geometric latent diffusion. Advances in Neural Information Processing Systems, 37: 0 74808--74839, 2025 a

    2025

  16. [24]

    Unimomo: Unified generative modeling of 3d molecules for de novo binder design

    Kong, X., Zhang, Z., Zhang, Z., Jiao, R., Ma, J., Huang, W., Liu, K., and Liu, Y. Unimomo: Unified generative modeling of 3d molecules for de novo binder design. In The 42nd International Conference on Machine Learning, 2025 b

    2025

  17. [25]

    B., Tran, H

    Kremsmayr, T., Aljnabi, A., Blanco-Canosa, J. B., Tran, H. N., Emidio, N. B., and Muttenthaler, M. On the utility of chemical strategies to improve peptide gut stability. Journal of medicinal chemistry, 65 0 (8): 0 6191--6206, 2022

    2022

  18. [26]

    J., Jin, Y., and Luk, L

    Lander, A. J., Jin, Y., and Luk, L. Y. D-peptide and d-protein technology: recent advances, challenges, and opportunities. ChemBioChem, 24 0 (4): 0 e202200537, 2023

    2023

  19. [27]

    Levin, D. A. and Peres, Y. Markov chains and mixing times, volume 107. American Mathematical Soc., 2017

    2017

  20. [28]

    Joint design of protein surface and backbone using a diffusion bridge model

    Li, G., Zhao, X., Wu, F., and Laue, S. Joint design of protein surface and backbone using a diffusion bridge model. In The 39th Annual Conference on Neural Information Processing Systems, 2025 a

    2025

  21. [29]

    Full-atom peptide design based on multi-modal flow matching

    Li, J., Cheng, C., Wu, Z., Guo, R., Luo, S., Ren, Z., Peng, J., and Ma, J. Full-atom peptide design based on multi-modal flow matching. In Proceedings of the 41st International Conference on Machine Learning (ICML 2024), pp.\ 27615–27640, 2024

    2024

  22. [30]

    E2Former : An Efficient and Equivariant Transformer with Linear-Scaling Tensor Products

    Li, Y., Huang, L., Ding, Z., Wei, X., Wang, C., Yang, H., Wang, Z., Liu, C., Shi, Y., Jin, P., Qin, T., Gerstein, M., and Zhang, J. E2Former : An Efficient and Equivariant Transformer with Linear-Scaling Tensor Products . In The Thirty-ninth Annual Conference on Neural Informa...

    2025

  23. [31]

    Lin, H., Zhang, O., Zhao, H., Jiang, D., Wu, L., Liu, Z., Huang, Y., and Li, S. Z. PPFlow : target-aware peptide design with torsional flow matching. In Proceedings of the 41st International Conference on Machine Learning, pp.\ 30510–30528, 2024

    2024

  24. [32]

    A., Martinez, C., Kasavajhala, K., Wickstrom, L., Hauser, K

    Maier, J. A., Martinez, C., Kasavajhala, K., Wickstrom, L., Hauser, K. E., and Simmerling, C. ff14sb: improving the accuracy of protein side chain and backbone parameters from ff99sb. Journal of chemical theory and computation, 11 0 (8): 0 3696--3713, 2015

    2015

  25. [33]

    and Cheng, J

    Morehead, A. and Cheng, J. Geometry-complete perceptron networks for 3d molecular graphs. Bioinformatics, 40 0 (2): 0 btae087, 2024

    2024

  26. [34]

    J., Gal, Y., Sander, C., and Marks, D

    Notin, P., Rollins, N. J., Gal, Y., Sander, C., and Marks, D. Machine learning for functional protein design. Nature Biotechnology, 42: 0 216--228, 2024

    2024

  27. [35]

    Memoires sur la relation qui peut exister entre la forme crystalline et al composition chimique, et sur la cause de la polarization rotatoire

    Pasteur, L. Memoires sur la relation qui peut exister entre la forme crystalline et al composition chimique, et sur la cause de la polarization rotatoire. Compt. rend., 26: 0 535--538, 1848

  28. [36]

    Atom-level generative foundation model for molecular interaction with pockets

    Peng, X., Guo, F., Guo, R., Sun, J., Guan, J., Jia, Y., Xu, Y., Huang, Y., Zhang, M., Peng, J., Wang, X., Han, C., Wang, Z., and Ma, J. Atom-level generative foundation model for molecular interaction with pockets. bioRxiv, 2025. bioRxiv:2024.10.17.618827

    2025

  29. [37]

    Mirror-image protein and peptide drug discovery through mirror-image phage display

    Qi, Y.-K., Zheng, J.-S., and Liu, L. Mirror-image protein and peptide drug discovery through mirror-image phage display. Chem, 10 0 (8): 0 2390--2407, 2024

    2024

  30. [38]

    High-resolution image synthesis with latent diffusion models

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10684--10695, 2022

    2022

  31. [39]

    G., Hoogeboom, E., and Welling, M

    Satorras, V. G., Hoogeboom, E., and Welling, M. E(n) Equivariant Graph Neural Networks , 2022

    2022

  32. [40]

    A., Fechner, N., and Ertl, P

    Schneider, N., Lewis, R. A., Fechner, N., and Ertl, P. Chiral cliffs: Investigating the influence of chirality on binding affinity. ChemMedChem, 13 0 (13): 0 1315--1324, 2018

    2018

  33. [41]

    E., Geiger, M., and Miller, B

    Smidt, T. E., Geiger, M., and Miller, B. K. Finding symmetry breaking order parameters with Euclidean neural networks. Physical Review Research, 3 0 (1): 0 L012002, 2021

    2021

  34. [42]

    Accurate de novo design of heterochiral protein--protein interactions

    Sun, K., Li, S., Zheng, B., Zhu, Y., Wang, T., Liang, M., Yao, Y., Zhang, K., Zhang, J., Li, H., et al. Accurate de novo design of heterochiral protein--protein interactions. Cell Research, 34 0 (12): 0 846--858, 2024

    2024

  35. [43]

    Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds, 2018

    Thomas, N., Smidt, T., Kearnes, S., Yang, L., Li, L., Kohlhoff, K., and Riley, P. Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds, 2018

    2018

  36. [44]

    K., Avraham, O., Ben-Aharon, Z., Khramushin, A., and Schueler-Furman, O

    Tsaban, T., Varga, J. K., Avraham, O., Ben-Aharon, Z., Khramushin, A., and Schueler-Furman, O. Harnessing protein folding neural networks for peptide--protein docking. Nature communications, 13 0 (1): 0 176, 2022

    2022

  37. [45]

    High- Dimensional Probability : An Introduction with Applications in Data Science

    Vershynin, R. High- Dimensional Probability : An Introduction with Applications in Data Science . Cambridge University Press, 2018. ISBN 978-1-108-41519-4

    2018

  38. [46]

    Target-specific de novo peptide binder design with DiffPepBuilder

    Wang, F., Wang, Y., Feng, L., Zhang, C., and Lai, L. Target-specific de novo peptide binder design with DiffPepBuilder . Journal of Chemical Information and Modeling, 64 0 (24): 0 9135--9149, 2024

    2024

  39. [47]

    L., Juergens, D., Bennett, N

    Watson, J. L., Juergens, D., Bennett, N. R., et al. De novo design of protein structure and function with RFdiffusion . Nature, 620: 0 1089--1100, 2023

    2023

  40. [48]

    Wedig, S., Elijo s ius, R., Schran, C., and Schaaf, L. L. Rem3di: Learning smooth, chiral 3d molecular representations from equivariant atomistic foundation models. In NeurIPS 2025 Workshop on Symmetry and Geometry in Neural Representations, 2025

    2025

  41. [49]

    D-Flow : Multi-modality flow matching for D -peptide design

    Wu, F., Jin, S., Tang, X., Xu, J., Gerstein, M., and Zou, J. D-Flow : Multi-modality flow matching for D -peptide design. arXiv preprint arXiv:2411.10618, 2024

    2024 arXiv

  42. [50]

    A novel d-peptide identified by mirror-image phage display blocks tigit/pvr for cancer immunotherapy

    Zhou, X., Zuo, C., Li, W., Shi, W., Zhou, X., Wang, H., Chen, S., Du, J., Chen, G., Zhai, W., et al. A novel d-peptide identified by mirror-image phage display blocks tigit/pvr for cancer immunotherapy. Angewandte Chemie International Edition, 59 0 (35): 0 15114--15118, 2020

    2020

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.