Pith. sign in

REVIEW 4 major objections 5 minor 40 references

DisProtEdit: Exploring Disentangled Representations for Multi-Attribute Protein Editing

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read DisProtEdit claims that separating structure from function in a protein's latent representation lets text prompts edit either attribute alone or both together, with up to 61.7% success on simultaneous edits.

desk verdict Useful disentanglement framework and dataset for text-guided protein editing, but the oracle-based success rates are not yet convincing. read the letter →

arxiv 2506.14853 v1 pith:LW4THN4Z submitted 2025-06-17 q-bio.QM cs.LG

Max Ku , Sun Sun , Hongyu Guo , Wenhu Chen This is my paper
classification q-bio.QMcs.LG
keywords controllableproteineditingdisentangledrepresentationlearningtext-guideddesignmaximummeandiscrepancyalignmentanduniformitystructuralfunctionalsemanticsmulti-attribute
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

DisProtEdit aims to show that a protein's structure and function can be treated as two separate semantic channels in a learned representation, so that changing a text description alters one attribute without disturbing the other. To do this, the authors align protein sequences with dual text descriptions, one structural and one functional, under alignment, uniformity, and a disentanglement loss based on maximum mean discrepancy. On a newly built multi-attribute benchmark of 196 sequences, the model reaches a both-hit success rate of up to 61.7% when structural and functional text inputs are edited together. If the claim holds, text-guided protein editing becomes more modular and interpretable than joint-embedding alternatives.

What carries the argument

The load-bearing mechanism is a dual-channel, angularly structured latent space. Each protein sequence is aligned to the concatenation of a structural text embedding and a functional text embedding, while separate maximum mean discrepancy (MMD) terms, a kernel-based distance between distributions, pull the two text subspaces toward independent angular priors with radii $\cos\varphi$ and $\sin\varphi$ for $\varphi\in[0,\pi/2]$. At edit time, a masked spherical linear interpolation blends only the half of the embedding whose text description changed, and the unedited half is retained. The angular separation is what lets one attribute be edited without disturbing the other.

What would settle it

Apply the same editing prompts to a control that performs generic destabilizing mutations or edits the functional half without any text conditioning; if that control matches the reported both-hit success rates (up to 61.7%), the disentangled text conditioning is not responsible for the edits. A second check is to take outputs of 'structure edited, function preserved' prompts and test them with an independent functional assay or holdout predictor; if success does not track the intended attribute changes, the oracle-based evaluation is not measuring true editing.

Watch

Extended reading notes

Core claim

DisProtEdit's central claim is that structural and functional semantics can be explicitly separated in a shared protein-text embedding, and that this separation enables compositional editing from natural language. The editing procedure takes the original protein embedding, interpolates it with a new text-derived embedding in only the subspace whose description was modified, and decodes an edited sequence. The reported evidence is that single-attribute edits outperform the contrastive baseline on several structure targets, and that combined structure-function prompts reach a both-hit success rate of up to 61.7%, for example increasing alpha-helix content while lowering Pin1 stability. The learned representations also transfer to property prediction, with the best variant reaching 0.5373 Spearman correlation on fluorescence and 0.8258 on stability in the four tasks the paper evaluates.

Load-bearing premise

The results stand on the assumption that the pretrained oracle predictors used to score edited sequences (secondary-structure classifiers and the Villin and Pin1 stability predictors) give accurate, unbiased measurements of the intended attributes for model-generated sequences; the paper's own limitation section says these oracles 'may be noisy or biased, especially for out-of-distribution sequences,' and the multi-attribute benchmark has only 196 sequences.

Editorial extensions

If this is right

  • Protein engineers could specify edits as sentences rather than mutation lists, changing structure or function in isolation.
  • Multi-attribute editing becomes compositional: one prompt per attribute, applied to the relevant subspace, with up to 61.7% both-hit success on the paper's benchmark.
  • The learned representation transfers to property prediction without task-specific alignment, matching or beating joint-embedding baselines on several tasks.
  • Replacing contrastive negatives with alignment and uniformity avoids false-negative sampling issues and still produces cross-modal alignment in a single training stage.

Reading between the lines

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

  • The strong asymmetry between increase and decrease edits (destabilization succeeds far more often than stabilization) suggests some reported edits may reflect generic sequence disruption rather than targeted attribute control, a possibility the benchmark's oracle scoring does not rule out.
  • The paper never checks that a structural edit preserves function or vice versa, so the cleanest test of disentanglement is still open: measure cross-attribute preservation with independent predictors or assays on the released benchmark.
  • If the disentanglement recipe generalizes, the same dual-channel scheme could be applied to other paired protein properties such as stability versus binding specificity or localization versus catalytic activity.
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 introduces DisProtEdit, a framework for controllable protein editing that learns disentangled representations of structural and functional properties from dual-channel text descriptions. The method uses alignment and uniformity losses to align protein and text embeddings, an MMD-based loss to encourage separation between structural and functional subspaces, and a T5 decoder to reconstruct sequences from edited latents. The authors also release SwissProtDis, a ~540k-entry dataset of protein sequences paired with LLM-decomposed structural and functional descriptions, and a 196-sample multi-attribute editing benchmark. Experiments report editing success rates for single and multi-attribute edits (both-hit up to 61.7%) and competitive TAPE benchmark results. The paper claims modular, interpretable, and controllable editing via text-guided latent interpolation.

Significance. If the claims hold, DisProtEdit would be a valuable contribution: it provides a large dual-channel dataset, a representation-learning objective tailored to protein structure/function disentanglement, and a text-interface editing mechanism that goes beyond joint-embedding baselines. The release of SwissProtDis and the multi-attribute benchmark are useful resources for the community. However, the central claims of disentanglement and controllability are not adequately supported by the current evaluation. The editing metric checks only the sign of oracle-score differences for targeted attributes, and the paper explicitly acknowledges that structural edits were not checked for functional preservation and that oracles may be biased for out-of-distribution sequences. With a small benchmark (N=196) and no confidence intervals or oracle validation, the reported success rates, especially the headline 61.7%, are fragile and may reflect oracle bias or generic degradation rather than genuine disentangled control. The TAPE results also show high sensitivity to the disentanglement weight λD (e.g., fluorescence Spearman 0.16 at λD=0.5 vs 0.54 at λD=1.0), raising concerns about robustness.

major comments (4)
  1. [Section 4.2, Eq. (10)] The editing success metric only counts whether the oracle-predicted scores for the targeted attributes moved in the instructed direction; it does not check whether untargeted attributes (e.g., structure when function is edited) are preserved. This is exactly the property that the 'disentangled' and 'controllable' claims require. The paper itself states in Section 5 that structural edits were not evaluated for functional preservation and vice versa. As a result, the central claim that DisProtEdit enables modular, disentangled editing is not supported by the presented evidence. Please add a preservation metric (e.g., fraction of untargeted attributes unchanged) or explicitly weaken the claims to 'targeted-attribute editing without evidence of preservation'.
  2. [Section 3.1, Eqs. (5), (7), (9)] The disentanglement loss and the editing mask rely on a fixed partition of the latent space into structural and functional halves, but the paper does not specify how Z_f and Z_s are extracted from the encoders or how the mask m in Eq. (9) selects the structural versus functional half. The alignment loss in Eq. (5) uses concat(z_ts, z_tf), but the order of concatenation is arbitrary and not tied to any loss term that would enforce 'the first half encodes structural semantics and the second half encodes functional semantics'. The MMD terms in Eq. (7) match Z_f and Z_s to priors X1 and X2, but the relationship between these embeddings and the halves of z_p is not defined. Without this specification, the editing mechanism and the claim of disentangled control are underspecified and cannot be reproduced.
  3. [Section 4.2, Tables 1 and 2] The evaluation relies on pretrained oracle predictors (secondary-structure classifiers, Villin/Pin1 stability predictors) applied to model-generated sequences that are likely out-of-distribution for those oracles. The authors acknowledge in Section 5 that these oracles 'may be noisy or biased, especially for out-of-distribution sequences,' but no calibration or validation of the oracles on generated sequences is provided. The observed asymmetry—decrease edits succeed at very high rates (−Villin up to 100%, −Pin1 up to 98.44%) while increase edits succeed rarely (+Villin ≤10.94%, +Pin1 ≤14.06%)—is consistent with generic destabilization that the authors attribute to ProteinDT but do not rule out for DisProtEdit. With only 196 samples in the multi-attribute benchmark and no confidence intervals or significance tests, the both-hit numbers (e.g., 61.7% for +α-helices, −Pin1) cannot be distinguished from oracle-driven noise. Please report confidence intervals, perform oracle validation on a held-out set of generated sequences with known properties, or re-frame the quantitative claims as preliminary.
  4. [Appendix D, Table 6] The example 'structural' descriptions in SwissProtDis often contain family/annotation information rather than structural properties, e.g., 'This protein belongs to the asfivirus MGF 100 family' or 'suggesting it shares structural characteristics common to this family.' If these are representative, the two text channels are not cleanly separated, and the disentanglement target may be mis-specified. This directly affects the premise of the method. Please provide a quantitative analysis of channel separation (e.g., overlap in keywords) and clarify how the LLM decomposition is validated.
minor comments (5)
  1. [General] Throughout the tables and text, 'α-helice' and 'β-sheet' should be 'α-helices' and 'β-sheets' for grammatical consistency; also 'DisProtEdit' vs 'DisProtE- dit' spacing in the abstract should be fixed.
  2. [Section 3.1, Eq. (8)] The loss notation is inconsistent: LA is defined but Eq. (8) uses LE = LA + λU LU + λD LD. Clarify whether the reconstruction loss is included in LA or trained separately; the decoder training setup in Appendix A says the decoder is trained separately, so the total training loss for the encoders should be stated explicitly.
  3. [Section 4.3, Table 3] The fluorescence results are highly sensitive to λD (0.1614 at λD=0.5 vs 0.5373 at λD=1.0). This instability suggests that the reported superiority of DisProtEdit on regression tasks may not be robust; please discuss or provide variance estimates.
  4. [References] The citation for T5 (Raffel et al., 2023) should be the original 2020 paper (or a later archival version); the current citation is misleading.
  5. [Section 5] The limitations section is candid and helpful, but the strength of the claims in the abstract and conclusion (e.g., 'explicitly separates semantic factors') is not aligned with the limitations. Please temper the language to match the evidence.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: DisProtEdit is an empirical training framework whose editing claims are checked against external pretrained oracles and external benchmarks; the disentanglement objective makes separation partly by construction, but no prediction reduces to a fitted input or a self-citation chain.

full rationale

DisProtEdit is a representation-learning method, not a derivation, and I found no load-bearing circular step. The abstract's claim that the model 'explicitly separates semantic factors' is implemented directly through Eq. 7 (MMD to independent angular priors) and Eq. 8, so the separation visualized in Figures 2-4 is partly by construction; however, this is an architectural design choice, not a result claimed to be derived from the data. The headline editing numbers (Tables 1-2) are computed by Eq. 10 using pretrained oracle predictors for secondary structure and Villin/Pin1 stability; these oracles are external to the training objective, and the paper's Section 5 openly warns that they 'may be noisy or biased, especially for out-of-distribution sequences.' That is a validity limitation, not circularity. The TAPE results (Table 3) are likewise external benchmarks. Hyperparameters λU and λD are ablated, not fitted to reproduce the evaluation outcomes; reporting the best 'up to 61.7%' is selection over configurations, not a fitted parameter renamed as a prediction. The only self-citation is the ProteinDT baseline (Liu et al. 2023, with overlapping author H. Guo), which is used as a comparison baseline rather than as a load-bearing premise; no uniqueness theorem or prior result is imported to force the method. The paper's own limitations (no preservation check for cross-attribute edits, oracle bias, decoder overfitting) reduce confidence in the biological claims but do not make the derivation circular. Under the stated rubric, this is 'no significant circularity' with a minor non-load-bearing self-citation.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central method relies on several domain assumptions that are not independently validated: the LLM decomposition is treated as ground truth, oracle predictors define editing success, and the angular MMD loss is assumed to produce true disentanglement. These assumptions, rather than a derivation, carry the central claim.

free parameters (4)
  • lambda_U (uniformity loss weight) = 0.2
    Chosen by hand after observing training stability; increasing lambda_U beyond 0.2 causes unstable optimization (Section 4.1).
  • lambda_D (disentanglement loss weight) = 1.0
    Chosen by hand; the paper reports a sweep from 0 to 5.0 and selects 1.0 as the best balance (Section 4.1).
  • alpha (slerp interpolation factor) = 0.9
    Fixed at 0.9 in Eq. 9 for editing interpolation; no ablation reported.
  • r1^2, r2^2 (angular decomposition radii) = 0.5, 0.5
    Set to equal split between structural and functional subspaces (Section 4.1).
assumptions (5)
  • domain assumption The concatenation of structural and functional text embeddings can approximate the protein embedding (Eq. 5).
    This alignment loss assumes structure and function text together capture the protein's semantics, which is a strong simplification given structure and function are coupled in biology.
  • ad hoc to paper MMD matching to angular priors enforces disentanglement between structure and function embeddings (Eq. 7).
    This is a design choice; the paper does not show theoretically that this objective yields independent subspaces.
  • domain assumption GPT-4o decomposition of SwissProt annotations produces accurate, non-overlapping structure/function descriptions.
    The dataset is built automatically; the paper acknowledges the LLM may introduce hallucinated or biologically imprecise content (Section 5).
  • domain assumption Oracle predictors (e.g., for alpha helices, Villin, Pin1) provide valid measures of the intended attributes for edited sequences.
    The paper states 'these may be noisy or biased, especially for out-of-distribution sequences' (Section 5).
  • standard math Standard mathematical tools (alignment/uniformity, MMD, slerp) behave as expected in this setting.
    These are established; the paper describes them in Section 2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DisProtEdit: Exploring Disentangled Representations for Multi-Attribute Protein Editing." pith.science (2026). https://pith.science/paper/LW4THN4Z

@misc{pith2026250614853,
  author       = {Pith},
  title        = {Pith review of: DisProtEdit: Exploring Disentangled Representations for Multi-Attribute Protein Editing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LW4THN4Z}},
  note         = {Machine review of arXiv:2506.14853}
}
read the original abstract

We introduce DisProtEdit, a controllable protein editing framework that leverages dual-channel natural language supervision to learn disentangled representations of structural and functional properties. Unlike prior approaches that rely on joint holistic embeddings, DisProtEdit explicitly separates semantic factors, enabling modular and interpretable control. To support this, we construct SwissProtDis, a large-scale multimodal dataset where each protein sequence is paired with two textual descriptions, one for structure and one for function, automatically decomposed using a large language model. DisProtEdit aligns protein and text embeddings using alignment and uniformity objectives, while a disentanglement loss promotes independence between structural and functional semantics. At inference time, protein editing is performed by modifying one or both text inputs and decoding from the updated latent representation. Experiments on protein editing and representation learning benchmarks demonstrate that DisProtEdit performs competitively with existing methods while providing improved interpretability and controllability. On a newly constructed multi-attribute editing benchmark, the model achieves a both-hit success rate of up to 61.7%, highlighting its effectiveness in coordinating simultaneous structural and functional edits.

Figures

Figures reproduced from arXiv: 2506.14853 by the authors.

Figure 1
Figure 1. Overview of the DisProtEdit framework. (a) During joint training, proteins and their corresponding structural and functional text descriptions are encoded into a modality-aligned embedding space using alignment, uniformity, and disentanglement objectives. (b) A decoder is trained to reconstruct sequences from latent representations. (c) Protein editing is performed via interpolation between the original embedding an… view at source ↗
Figure 2
Figure 2. UMAP visualizations of text and protein embeddings under different training strategies. Each visualization sampled 500 pairs of data. Each point represents a text (yellow) or protein (blue) embedding. Lines connect paired structural/functional text embeddings and their corresponding protein embedding, illustrat￾ing the degree of cross-modal alignment. (a) Random projection baseline. (b) Contrastive learning shows a … view at source ↗
Figure 3
Figure 3. Effect of the uniformity loss weight λU on cross-modal alignment. The plots compare embedding distributions of functional text and structural text representations under varying values. 𝜆U=0.2, 𝜆𝐷=0 𝜆U=0.2, 𝜆𝐷=0.1 𝜆U=0.2, 𝜆𝐷=0.5 𝜆U=0.2, 𝜆𝐷=0.8 𝜆U=0.2, 𝜆𝐷=1.0 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Effect of the disentanglement loss weight λD on cross-modal alignment, comparing embedding distributions of functional text and structural text representations under varying values. varying the disentanglement loss weight λD on the geom￾etry of structural and functiona…
Figure 5
Figure 5. Figure 5: Qualitative visualization of structure–function pro￾tein edit samples. (a, c) Original protein sequences with their corresponding structural and functional attributes. (b, d) Edited proteins generated by DisProtEdit in response to compositional prompts: (b) increase al…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 14 canonical work pages

  1. [1]

    J., Bambrick, J., Bodenstein, S

    Abramson, J., Adler, J., Dunger, J., Evans, R., Green, T., Pritzel, A., Ronneberger, O., Willmore, L., Ballard, A. J., Bambrick, J., Bodenstein, S. W., Evans, D. A., Hung, C.-C., O’Neill, M., Reiman, D., Tunyasuvunakool, K., Wu, Z., Žemgulytė, A., Arvaniti, E., Beattie, C., Bertolli, O., Bridgland, A., Cherepanov, A., Congreve, M., Cowen-Rivers, A. I., Co...

    2024

  2. [2]

    Scibert: Pretrained language model for scientific text

    Beltagy, I., Lo, K., and Cohan, A. Scibert: Pretrained language model for scientific text. In EMNLP, 2019

    work page 2019

  3. [3]

    Proteinbert: a universal deep-learning model of protein sequence and function

    Brandes, N., Ofer, D., Peleg, Y., Rappoport, N., and Linial, M. Proteinbert: a universal deep-learning model of protein sequence and function. Bioinformatics, 38 0 (8): 0 2102--2110, 02 2022. ISSN 1367-4803. doi:10.1093/bioinformatics/btac020. URL https://doi.org/10.1093/bioinformatics/btac020

    doi:10.1093/bioinformatics/btac020 2022

  4. [4]

    and Lowe, D

    Broomhead, D. and Lowe, D. Multivariable functional interpolation and adaptive networks. Complex Systems, 2: 0 321--355, 1988

    work page 1988

  5. [5]

    A simple framework for contrastive learning of visual representations, 2020

    Chen, T., Kornblith, S., Norouzi, M., and Hinton, G. A simple framework for contrastive learning of visual representations, 2020. URL https://arxiv.org/abs/2002.05709

    arXiv 2020

  6. [6]

    Consortium, T. U. Uniprot: the universal protein knowledgebase in 2025. Nucleic Acids Research, 53 0 (D1): 0 D609--D617, 11 2024. ISSN 1362-4962. doi:10.1093/nar/gkae1010. URL https://doi.org/10.1093/nar/gkae1010

    doi:10.1093/nar/gkae1010 2025

  7. [7]

    Toward de novo protein design from natural language

    Dai, F., Fan, Y., Su, J., Wang, C., Han, C., Zhou, X., Liu, J., Qian, H., Wang, S., Zeng, A., Wang, Y., and Yuan, F. Toward de novo protein design from natural language. bioRxiv, 2025. doi:10.1101/2024.08.01.606258. URL https://www.biorxiv.org/content/early/2025/01/24/2024.08.01.606258

    doi:10.1101/2024.08.01.606258 2025

  8. [8]

    Prottrans: Towards cracking the language of life s code through self-supervised deep learning and high performance computing

    Elnaggar, A., Heinzinger, M., Dallago, C., Rehawi, G., Wang, Y., Jones, L., Gibbs, T., Feher, T., Angerer, C., Steinegger, M., BHOWMIK, D., and Rost, B. Prottrans: Towards cracking the language of life s code through self-supervised deep learning and high performance computing. bioRxiv, 2020. doi:10.1101/2020.07.12.199554. URL https://www.biorxiv.org/cont...

    doi:10.1101/2020.07.12.199554 2020

Show all 40 references
  1. [9]

    Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context, 2024

    Gemini-Team. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context, 2024

    2024

  2. [10]

    M., Rasch, M

    Gretton, A., Borgwardt, K. M., Rasch, M. J., Sch\" o lkopf, B., and Smola, A. A kernel two-sample test. J. Mach. Learn. Res., 13 0 (null): 0 723–773, March 2012. ISSN 1532-4435

    2012

  3. [11]

    R., Maire, M., and Khademi, M

    Huynh, T., Kornblith, S., Walter, M. R., Maire, M., and Khademi, M. Boosting contrastive self-supervised learning with false negative cancellation, 2022. URL https://arxiv.org/abs/2011.11765

    2022 arXiv

  4. [12]

    B., Baranov, M., Costello, Z., Barber, K

    Ingraham, J. B., Baranov, M., Costello, Z., Barber, K. W., Wang, W., Ismail, A., Frappier, V., Lord, D. M., Ng-Thow-Hing, C., Van Vlack, E. R., Tie, S., Xue, V., Cowles, S. C., Leung, A., Rodrigues, J. a. V., Morales-Perez, C. L., Ayoub, A. M., Green, R., Puentes, K., Oplinger...

    2023 doi

  5. [13]

    Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R., Z \' dek, A., Potapenko, A., Bridgland, A., Meyer, C., Kohl, S. A. A., Ballard, A. J., Cowie, A., Romera-Paredes, B., Nikolov, S., Jain, R., Adler, J., Back, T., Peter...

    2021

  6. [14]

    Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization, 2017. URL https://arxiv.org/abs/1412.6980

    2017 arXiv

  7. [15]

    Li, T., Guo, H., Grazioli, F., Gerstein, M., and Min, M. R. Disentangled wasserstein autoencoder for t-cell receptor engineering, 2023. URL https://arxiv.org/abs/2210.08171

    2023 arXiv

  8. [16]

    Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning

    Liang, W., Zhang, Y., Kwon, Y., Yeung, S., and Zou, J. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. In NeurIPS, 2022. URL https://openreview.net/forum?id=S7Evzt9uit3

    2022

  9. [17]

    Pre-training molecular graph representation with 3d geometry

    Liu, S., Wang, H., Liu, W., Lasenby, J., Guo, H., and Tang, J. Pre-training molecular graph representation with 3d geometry. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=xQUe1pOKPam

    2022

  10. [18]

    A text-guided protein design framework

    Liu, S., Li, Y., Li, Z., Gitter, A., Zhu, Y., Lu, J., Xu, Z., Nie, W., Ramanathan, A., Xiao, C., Tang, J., Guo, H., and Anandkumar, A. A text-guided protein design framework. arXiv preprint arXiv:2302.04611, 2023

    2023 arXiv

  11. [19]

    Causal effect inference with deep latent-variable models, 2017 a

    Louizos, C., Shalit, U., Mooij, J., Sontag, D., Zemel, R., and Welling, M. Causal effect inference with deep latent-variable models, 2017 a . URL https://arxiv.org/abs/1705.08821

    2017 arXiv

  12. [20]

    The variational fair autoencoder, 2017 b

    Louizos, C., Swersky, K., Li, Y., Welling, M., and Zemel, R. The variational fair autoencoder, 2017 b . URL https://arxiv.org/abs/1511.00830

    2017 arXiv

  13. [21]

    Y.-C., Yuan, L., and Tian, Y

    Lv, L., Lin, Z., Li, H., Liu, Y., Cui, J., Chen, C. Y.-C., Yuan, L., and Tian, Y. Prollama: A protein large language model for multi-task protein language processing. arXiv preprint arXiv:2402.16445, 2024

    2024 arXiv

  14. [22]

    Ma, Z., Fan, C., Wang, Z., Chen, Z., Lin, X., Li, Y., Feng, S., Zhang, J., Cao, Z., and Gao, Y. Q. Prottex: Structure-in-context reasoning and editing of proteins with large language models, 2025. URL https://arxiv.org/abs/2503.08179

    2025 arXiv

  15. [23]

    S., Anand, N., Eguchi, R

    Madani, A., McCann, B., Naik, N., Keskar, N. S., Anand, N., Eguchi, R. R., Huang, P.-S., and Socher, R. Progen: Language modeling for protein generation, 2020. URL https://arxiv.org/abs/2004.03497

    2020 arXiv

  16. [24]

    Mathieu, E., Rainforth, T., Siddharth, N., and Teh, Y. W. Disentangling disentanglement in variational autoencoders, 2019. URL https://arxiv.org/abs/1812.02833

    2019 arXiv

  17. [25]

    Umap: Uniform manifold approximation and projection

    McInnes, L., Healy, J., Saul, N., and Grossberger, L. Umap: Uniform manifold approximation and projection. The Journal of Open Source Software, 3 0 (29): 0 861, 2018

    2018

  18. [26]

    Colabfold: making protein folding accessible to all

    Mirdita, M., Sch \"u tze, K., Moriwaki, Y., Heo, L., Ovchinnikov, S., and Steinegger, M. Colabfold: making protein folding accessible to all. Nature Methods, 19 0 (6): 0 679--682, Jun 2022. ISSN 1548-7105. doi:10.1038/s41592-022-01488-1. URL https://doi.org/10.1038/s41592-022-01488-1

    2022 doi

  19. [27]

    N., Naik, N., and Madani, A

    Nijkamp, E., Ruffolo, J., Weinstein, E. N., Naik, N., and Madani, A. Progen2: Exploring the boundaries of protein language models, 2023. URL https://openreview.net/forum?id=ZOn4HXehSJ6

    2023

  20. [28]

    Gpt-4 technical report, 2023

    OpenAI. Gpt-4 technical report, 2023

    2023

  21. [29]

    A., Zhang, R., and Zhu, J.-Y

    Park, T., Efros, A. A., Zhang, R., and Zhu, J.-Y. Contrastive learning for unpaired image-to-image translation, 2020. URL https://arxiv.org/abs/2007.15651

    2020 arXiv

  22. [30]

    Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., and Liu, P. J. Exploring the limits of transfer learning with a unified text-to-text transformer, 2023. URL https://arxiv.org/abs/1910.10683

    2023 arXiv

  23. [31]

    Rao, R., Bhattacharya, N., Thomas, N., Duan, Y., Chen, X., Canny, J., Abbeel, P., and Song, Y. S. Evaluating protein transfer learning with tape. In Advances in Neural Information Processing Systems, 2019

    2019

  24. [32]

    L., Ma, J., and Fergus, R

    Rives, A., Meier, J., Sercu, T., Goyal, S., Lin, Z., Liu, J., Guo, D., Ott, M., Zitnick, C. L., Ma, J., and Fergus, R. Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences. PNAS, 2019. doi:10.1101/622803. URL https://www....

    2019 doi

  25. [33]

    Contrastive learning with hard negative samples, 2021

    Robinson, J., Chuang, C.-Y., Sra, S., and Jegelka, S. Contrastive learning with hard negative samples, 2021. URL https://arxiv.org/abs/2010.04592

    2021 arXiv

  26. [34]

    Representation learning with contrastive predictive coding, 2019

    van den Oord, A., Li, Y., and Vinyals, O. Representation learning with contrastive predictive coding, 2019. URL https://arxiv.org/abs/1807.03748

    2019 arXiv

  27. [35]

    and Isola, P

    Wang, T. and Isola, P. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pp.\ 9929--9939. PMLR, 2020

    2020

  28. [36]

    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., Wicky, B. I. M., Hanikel, N., Pellock, S. J., Courbet, A., Sheffler, W., Wang, J., Venkatesh, P., Sappington, I., Torres, S. V., Lauko,...

    2023

  29. [37]

    Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation, 2017

    Yan, H., Ding, Y., Li, P., Wang, Q., Xu, Y., and Zuo, W. Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation, 2017. URL https://arxiv.org/abs/1705.00609

    2017 arXiv

  30. [38]

    Multi-modal clip-informed protein editing, 2024

    Yin, M., Zhou, H., Zhu, Y., Lin, M., Wu, Y., Wu, J., Xu, H., Hsieh, C.-Y., Hou, T., Chen, J., and Wu, J. Multi-modal clip-informed protein editing, 2024. URL https://arxiv.org/abs/2407.19296

    2024 arXiv

  31. [39]

    Ontoprotein: Protein pretraining with gene ontology embedding

    Zhang, N., Bi, Z., Liang, X., Cheng, S., Hong, H., Deng, S., Zhang, Q., Lian, J., and Chen, H. Ontoprotein: Protein pretraining with gene ontology embedding. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=yfe1VMYAXa4

    2022

  32. [40]

    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 gl...

Pith tools

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