Pith. sign in

REVIEW 4 major objections 4 minor 44 references

A Variational Perspective on Generative Protein Fitness Optimization

T0 review · 4 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read A variational framework that embeds protein sequences in a continuous latent space and steers a flow-matching prior with a fitness predictor, reporting the highest fitness scores on AAV and GFP benchmarks.

desk verdict Good method paper with a real evaluation problem: Table 8 suggests the oracle overestimates off-manifold VAE-decoded sequences, so the SOTA claims should not be taken at face value. read the letter →

arxiv 2501.19200 v2 pith:66V5OT6E submitted 2025-01-31 cs.LG

Lea Bogensperger , Dominik Narnhofer , Ahmed Allam , Konrad Schindler , Michael Krauthammer This is my paper · ORCID
classification cs.LG
keywords proteinfitnessoptimizationflowmatchingvariationalautoencoderclassifierguidancelatentspacegenerativemodelAAVGFPlimited-datadesign
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Protein fitness optimization searches a huge, discrete, mostly low-fitness space of amino-acid sequences for variants with better stability, binding, or activity. VLGPO moves the search into a continuous latent space learned by a variational autoencoder, models the distribution of variants there with flow matching, and uses a fitness predictor's gradient to push sampling toward high-fitness regions. The paper reports that this guided latent-space sampler achieves the highest median normalized fitness among all compared methods on the medium and hard AAV and GFP benchmarks, which are designed to mimic limited-data protein engineering. If correct, VLGPO provides a modular, purely computational way to propose candidate variants worth testing in the lab.

What carries the argument

The load-bearing mechanism is manifold-constrained classifier guidance in latent space, where flow matching is a generative modeling scheme that learns a velocity field transporting a simple noise distribution to the data distribution. At each step of the flow integration, the method first proposes the next latent point, then uses the flow model to estimate the denoised sequence that this point would decode to; the fitness predictor's squared error against the target fitness is differentiated with respect to the current latent point, so the guidance gradient is computed on a valid decoded sequence while being applied to a point on the learned latent manifold. This step is what prevents the sampling trajectory from drifting off the manifold, and the paper's ablation shows it improves fitness on most tasks. The surrounding components—a variational autoencoder for the latent embedding, a flow-matching prior, and a convolutional fitness predictor—are interchangeable modules.

What would settle it

Synthesize the top-ranked sequences VLGPO produces for each of the four tasks, measure their fitness in the lab, and compare with the oracle's predictions and with the training-set median; if the measured median is not above the best baseline's measured median, the central claim fails. A cheaper in-silico falsifier: remove all VAE-decoded sequences that fail reconstruction (AAV medium reconstruction accuracy is 80.4%) and re-evaluate; if the fitness gain vanishes, the gain is an artifact of invalid decodes.

Watch

Extended reading notes

Core claim

The paper's central claim is that protein fitness optimization should be done in a compressed continuous latent space rather than directly on discrete amino-acid tokens. VLGPO trains a variational autoencoder to map sequences into that space and reconstruct them, trains a flow-matching model to generate new points from the latent distribution, and guides the generative trajectory with a fitness predictor. The guidance step is manifold-constrained: at every integration step the algorithm estimates the denoised endpoint, decodes it, evaluates the predictor's error there, and backpropagates that error to the current latent point, so the sample stays on the learned manifold. With this recipe, the paper reports the best median normalized fitness on all four benchmark tasks—AAV medium and hard, GFP medium and hard—using either the plain or the graph-smoothed predictor.

Load-bearing premise

The whole evaluation assumes that a computer model trained on all known mutations correctly measures the fitness of newly generated sequences; if that model is wrong for sequences unlike the training data, the reported improvements may not appear in real proteins.

Editorial extensions

If this is right

  • On the four benchmark tasks, VLGPO reaches the highest median normalized fitness of all compared methods, including the two methods that share its predictors: GWG and GGS.
  • Because the guidance gradient is evaluated through the decoder at the estimated denoised endpoint, any differentiable fitness predictor, any decoder, and any flow prior can be swapped in, so the same sampler can be pointed at new proteins or new fitness definitions.
  • The ablation quantifies the value of the manifold constraint: removing it lowers fitness on most tasks, for example GFP medium drops from 0.87 to 0.81 with the plain predictor and AAV hard from 0.51 to 0.47.
  • Classifier guidance beats a directly fitness-conditioned flow posterior, with the largest gap on GFP hard, where the training fitness range is [0.0, 0.1]; this indicates explicit predictor gradients are what allow extrapolation beyond the observed fitness values.

Reading between the lines

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

  • Beyond the paper, the same mechanism should transfer to latent spaces produced by protein language models or to other discrete design spaces such as RNA and DNA, because the guidance step only needs a differentiable decoder and a differentiable predictor; a direct test would replace the VAE and measure fitness gain versus reconstruction fidelity.
  • The reported numbers are oracle evaluations, not wet-lab measurements, and the top-128 sequences are selected by the same predictor that steers sampling; an experimental pass could therefore shrink the measured advantage, since oracle and predictor share systematic errors.
  • Given the AAV medium VAE reconstructs only 80.4% of validation sequences, some generated latents decode to invalid proteins; filtering on reconstruction confidence before oracle evaluation would reveal how much of the reported gain survives stricter validity checks, which the paper does not report.
  • A cheap experiment the paper does not run: vary the number of decoded samples and the top-k cutoff, and plot oracle fitness against diversity; the fixed 512-sample, 128-selection choice may not be the operating point that maximizes useful, diverse candidates.
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 / 4 minor

Summary. This paper introduces VLGPO, a latent-space generative approach to protein fitness optimization. A VAE embeds protein sequences into a continuous latent space; a flow-matching model learns the latent distribution of sequence variants; at sampling time, the flow ODE is guided by gradients of a fitness predictor, using a manifold-constraint update to avoid off-manifold drift, in order to generate sequences with high predicted fitness. The method is evaluated on two benchmarks (AAV and GFP) in medium/hard limited-data regimes from Kirjner et al. (2023), using an in-silico oracle trained on the full DMS data. The paper reports that VLGPO outperforms a range of baselines, including GWG and GGS, which use the same fitness predictors, and claims state-of-the-art performance.

Significance. If the reported gains hold up, VLGPO would be a useful contribution: it shows that a continuous latent space plus a flow-matching prior can be effectively guided by off-the-shelf fitness predictors, with fewer discrete-sampling complications than Gibbs-based methods. The code is available, the ablations (Tables 5-6) support the benefit of the manifold-constrained gradient, and the comparison against GWG/GGS with the same predictor isolates the effect of the proposed sampling scheme. However, the evaluation relies entirely on an in-silico oracle whose reliability for out-of-distribution decoded sequences is questionable, and the headline SOTA claim depends on that oracle.

major comments (4)
  1. [Appendix B.3, Table 8; Section 4.1, Table 2] The unconditional sampling result for GFP hard is incompatible with the assumption that the oracle gψ reliably scores generated sequences. The training set for GFP hard has fitness range [0.0, 0.1] (Table 2), and the flow-matching prior is trained only on latents of those low-fitness sequences. With αt = 0 and J = 0, sampling should produce sequences near the training distribution, yet Table 8 reports median oracle fitness 0.42 ± 0.1, more than four times the maximum training fitness. This strongly suggests that the oracle (or the VAE decoder) systematically overestimates fitness for decoded samples that are off the training manifold. Since all headline fitness numbers in Tables 3 and 4 are produced by the same oracle, the state-of-the-art claim may reflect evaluation artifacts rather than real fitness gains. The authors should quantify oracle reliability on out-of-distribution sequences (e.g., retrain on a random split of DMS data and evaluate on held-out mutants, or report oracle calibration versus Hamming distance from the training set) and, if the oracle cannot be trusted, temper the SOTA claim accordingly.
  2. [Equation (5), Algorithm 1, Section 3.4] The guidance update in Line 8 of Algorithm 1 requires ∇_{z'_t} ||gφ(D(ẑ1))−y||², which backpropagates through the decoder D. Section 3.4 says the decoder outputs logits that are mapped to tokens via an argmax operation; argmax has zero gradient almost everywhere. The paper does not specify the differentiable relaxation used (straight-through estimator, Gumbel-softmax, or soft-argmax). Without this detail, the central sampling mechanism in Algorithm 1 is not well-defined and the results are not reproducible. Please state the relaxation explicitly and, if a straight-through estimator is used, discuss its effect on the guidance gradient.
  3. [Appendix B.4, Figure 3] The hyperparameters αt and J are selected per task via grid search on the same benchmarks, and Figure 3's selection criterion is the median fitness (evaluated with the same predictor/oracle pipeline used in the final comparison). The baseline methods (GWG, GGS, GFN-AL, etc.) are not given an equivalent per-task tuning. This introduces a confound in the claimed improvement over baselines: it is unclear whether VLGPO wins because of the method or because of favorable hyperparameters. The authors should either tune the baselines with the same budget or show that the ranking is stable across a range of αt and J values.
  4. [Appendix B.2, Table 7] The VAE reconstruction accuracy for AAV medium is 80.4%, meaning about one in five decoded sequences is not a valid reconstruction of a training-set-like sequence. Yet all generated sequences are scored by the oracle without filtering for validity. Invalid sequences may be scored unreliably, which could bias the reported AAV medium fitness (0.58). The authors should report the reconstruction validity rate of the generated sequences and either exclude invalid decodes or show that they do not affect the conclusions.
minor comments (4)
  1. [Throughout] The term 'AA V' is written with a space in several places (e.g., Abstract, Table 2); standard notation is 'AAV'. Please fix the spacing consistently.
  2. [Section 4.2 and elsewhere] The phrase 'V AE' appears with a space in multiple locations (e.g., 'All V AE and flow matching models'); this should be 'VAE'.
  3. [Equation (5), Algorithm 1] The sign convention in Eq. (5) is inconsistent with Eq. (3) and Algorithm 1: Eq. (5) writes ∇_x log p(y|x1) ∼ 1/2 ∇_{z'_t} ||gφ(D(ẑ1))−y||², while Eq. (3) and Line 8 use a negative sign (with the factor −α_t/2). Please clarify whether the intended log-likelihood is −(1/2)||·||² and correct the sign in Eq. (5).
  4. [Abstract and Section 5] Section 5 appropriately acknowledges the reliance on in-silico evaluation and hyperparameter tuning, but the abstract still asserts state-of-the-art results without these caveats. Consider softening the abstract to match the limitations discussed in Section 5.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: VLGPO's outputs are evaluated by an external oracle, and no equation reduces the headline result to a fitted parameter or self-citation.

full rationale

The central claim is that VLGPO produces higher-fitness variants than baselines on AAV and GFP. The derivation chain is: (i) train a VAE and a flow-matching prior on the limited task data S; (ii) guide latent sampling with gradients of the re-used predictor g_phi (Eq. 3, Eq. 5, Algorithm 1); (iii) rank generated candidates by g_phi and report oracle g_psi fitness (Section 4.2). The oracle g_psi is a fixed network trained on the full DMS data and sourced from Kirjner et al.; it is not defined in terms of VLGPO's output, nor is any result equation equal to a fitted value. The same predictor is used for guidance and top-k ranking, which can inflate predictor-selected scores, but all reported fitness numbers are oracle-evaluated, so the ranking does not by construction force the oracle result. The acknowledged hyperparameter grid search (Appendix B.4) is tuned on the target tasks through the predictor; this is benchmark overfitting, not circularity. The Table 8 unconditional-sampling anomaly (GFP hard median 0.42 from training fitness range [0,0.1]) is a serious oracle-reliability concern, not a circularity: it does not show that the method's derivation reduces to its inputs. Section 5 explicitly acknowledges the in-silico oracle limitation. I find no self-citation chain, no imported uniqueness theorem, and no ansatz smuggled in by citation; the paper is self-contained against external benchmarks, so the circularity score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim depends on several fitted hyperparameters (guidance strength, gradient steps, VAE capacity) and on domain assumptions about the in-silico oracle, the smoothness of the latent space, and the validity of classifier guidance. None are new entities; the main risk is that per-task tuning plus oracle evaluation make the benchmark results optimistic.

free parameters (6)
  • Guidance strength α_t = AAV medium 0.97, AAV hard 1.2, GFP medium 0.56, GFP hard 0.02
    Selected via grid search per task (Appendix B.4); controls how strongly the fitness predictor steers sampling and directly affects reported fitness.
  • Likelihood gradient steps J = AAV medium 39, AAV hard 19, GFP medium 37, GFP hard 5
    Selected via grid search per task; more steps increase predicted fitness but reduce diversity.
  • VAE KL weight β = 0.01 (AAV), 0.001 (GFP)
    Chosen by hand to balance reconstruction and Gaussian latent prior; affects the latent space used by flow matching.
  • Latent dimension l = 16 (AAV), 32 (GFP)
    Chosen empirically for reconstruction accuracy; central to the continuous optimization space.
  • ODE steps K = 32
    Chosen after observing stability around 10+ steps (Figure 7); discretization of flow integration.
  • Number of generated samples and top-k = 512 samples, top-128
    Heuristic choices affecting diversity and reported fitness after ranking with predictor.
assumptions (5)
  • standard math Conditional flow matching loss is equivalent to the intractable marginal flow matching objective (Lipman et al., 2023).
    Used to train the flow matching prior in latent space (Section 3.2).
  • domain assumption Fitness predictor log-likelihood is modeled as -1/2 ||gφ(x)-y||², so classifier guidance can be applied via gradient of squared error.
    Section 3.3, Eq. (3); this is a heuristic approximation, not derived from the predictor's actual likelihood.
  • domain assumption Oracle gψ trained on full DMS data approximates true fitness of generated sequences.
    Used in all evaluations; authors acknowledge in Section 5 that only in-silico evaluation is performed.
  • domain assumption Manifold-constrained gradient: evaluating likelihood at ẑ1 and backpropagating to z't keeps samples on the learned data manifold.
    Adopted from DPS/D-flow (Chung et al., 2023; Ben-Hamu et al., 2024); Algorithm 1, lines 7-8.
  • domain assumption VAE latent space is smooth enough for flow matching and gradient guidance to produce valid sequences.
    Core to method; AAV reconstruction accuracy 80.4% suggests this assumption is only partially met.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Variational Perspective on Generative Protein Fitness Optimization." pith.science (2026). https://pith.science/paper/66V5OT6E

@misc{pith2026250119200,
  author       = {Pith},
  title        = {Pith review of: A Variational Perspective on Generative Protein Fitness Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/66V5OT6E}},
  note         = {Machine review of arXiv:2501.19200}
}
read the original abstract

The goal of protein fitness optimization is to discover new protein variants with enhanced fitness for a given use. The vast search space and the sparsely populated fitness landscape, along with the discrete nature of protein sequences, pose significant challenges when trying to determine the gradient towards configurations with higher fitness. We introduce Variational Latent Generative Protein Optimization (VLGPO), a variational perspective on fitness optimization. Our method embeds protein sequences in a continuous latent space to enable efficient sampling from the fitness distribution and combines a (learned) flow matching prior over sequence mutations with a fitness predictor to guide optimization towards sequences with high fitness. VLGPO achieves state-of-the-art results on two different protein benchmarks of varying complexity. Moreover, the variational design with explicit prior and likelihood functions offers a flexible plug-and-play framework that can be easily customized to suit various protein design tasks.

Figures

Figures reproduced from arXiv: 2501.19200 by the authors.

Figure 1
Figure 1. Overview of VLGPO sampling. The central section illustrates the VAE framework, showcasing protein sequences, their latent representations z, and the approximate posterior distribution. While the upper section depicts unconditional sampling from the prior p(x) using flow matching in the latent space, the lower section illustrates the modifications introduced by VLGPO during sampling. We additionally incorporate a lik… view at source ↗
Figure 2
Figure 2. Schematic depiction of classifier guidance, with J = 1 and K = 6. Grey lines represent the latent manifolds at different time steps t, the blue line marks the trajectory of the maximum likelihood. Solid arrows indicate how the latent evolves over time. Left: Naive guidance with likelihood gradients ∇zt computed directly at zt pushes the sample off the manifold. This error accumulates, as indicated by the purple regi… view at source ↗
Figure 3
Figure 3. Grid search for median fitness depending on sampling parameters αt and J for the different tasks using the predictor gϕ. In general, higher values of αt and J, corresponding to strong classifier guidance, yield higher predicted fitness values. ble 4 for GFP and AAV respectively, demonstrate the fitness improvement of VLGPO with both predictors (gϕ and gϕ˜) over all other benchmarked methods. In particular, VLGPO sho… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Comparing evaluated fitness ygt from the oracle gψ with required fitness y using the directly learned posterior model (in the same latent space) and our variational approach VLGPO [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: AAV medium B.2. Variational Autoencoder The VAE embeds sequences into a latent space that yields continuous gradients for sampling in Algorithm 1. Empiri￾cally, we choose l = 16 and l = 32 for AAV and GFP, since further compression reduces the decoder D’s reconstructio…
Figure 6
Figure 6. Figure 6: Grid search for diversity depending on sampling parameters αt and J for the different tasks [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Median fitness (left) and diversity (right) for all four tasks depending on employed ODE steps K in sampling. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 20 canonical work pages

  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]

    D-flow: Differentiating through flows for controlled generation

    Ben-Hamu, H., Puny, O., Gat, I., Karrer, B., Singer, U., and Lipman, Y. D-flow: Differentiating through flows for controlled generation. arXiv preprint arXiv:2402.14017, 2024

    arXiv 2024

  3. [3]

    Conditioning by adaptive sampling for robust design

    Brookes, D., Park, H., and Listgarten, J. Conditioning by adaptive sampling for robust design. In International conference on machine learning, pp.\ 773--782. PMLR, 2019

    2019

  4. [4]

    H., Bashir, A., Sinai, S., Jain, N

    Bryant, D. H., Bashir, A., Sinai, S., Jain, N. K., Ogden, P. J., Riley, P. F., Church, G. M., Colwell, L. J., and Kelsic, E. D. Deep diversification of an aav capsid protein by machine learning. Nature Biotechnology, 39 0 (6): 0 691--696, 2021

    work page 2021

  5. [5]

    Chung, H., Sim, B., Ryu, D., and Ye, J. C. Improving diffusion models for inverse problems using manifold constraints. Advances in Neural Information Processing Systems, 35: 0 25683--25696, 2022

    2022

  6. [6]

    T., Klasky, M

    Chung, H., Kim, J., Mccann, M. T., Klasky, M. L., and Ye, J. C. Diffusion posterior sampling for general noisy inverse problems. In International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=OnD9zGAGT0k

    2023

  7. [7]

    E., Wittmann, B

    Dallago, C., Mou, J., Johnston, K. E., Wittmann, B. J., Bhattacharya, N., Goldman, S., Madani, A., and Yang, K. K. Flip: Benchmark tasks in fitness landscape inference for proteins. bioRxiv, pp.\ 2021--11, 2021

    work page 2021

  8. [8]

    and Nichol, A

    Dhariwal, P. and Nichol, A. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34: 0 8780--8794, 2021

    2021

Show all 44 references
  1. [9]

    Emami, P., Perreault, A., Law, J., Biagioni, D., and John, P. S. Plug & play directed evolution of proteins with gradient-based discrete mcmc. Machine Learning: Science and Technology, 4 0 (2): 0 025014, 2023

    2023

  2. [10]

    Scaling rectified flow transformers for high-resolution image synthesis

    Esser, P., Kulal, S., Blattmann, A., Entezari, R., M \"u ller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., et al. Scaling rectified flow transformers for high-resolution image synthesis. URL https://arxiv. org/abs/2403.03206, 2, 2024

    2024 arXiv

  3. [11]

    C., Berenberg, D., Zadorozhny, K., Kleinhenz, J., Lafrance-Vanasse, J., Hotzel, I., Wu, Y., Ra, S., Bonneau, R., Cho, K., et al

    Frey, N. C., Berenberg, D., Zadorozhny, K., Kleinhenz, J., Lafrance-Vanasse, J., Hotzel, I., Wu, Y., Ra, S., Bonneau, R., Cho, K., et al. Protein discovery with discrete walk-jump sampling. arXiv preprint arXiv:2306.12360, 2023

    2023 arXiv

  4. [12]

    Oops i took a gradient: Scalable sampling for discrete distributions

    Grathwohl, W., Swersky, K., Hashemi, M., Duvenaud, D., and Maddison, C. Oops i took a gradient: Scalable sampling for discrete distributions. In International Conference on Machine Learning, pp.\ 3831--3841. PMLR, 2021

    2021

  5. [13]

    G., Hotzel, I., Lafrance-Vanasse, J., Rajpal, A., Cho, K., and Wilson, A

    Gruver, N., Stanton, S., Frey, N., Rudner, T. G., Hotzel, I., Lafrance-Vanasse, J., Rajpal, A., Cho, K., and Wilson, A. G. Protein design with guided discrete diffusion. Advances in neural information processing systems, 36, 2024

    2024

  6. [14]

    D., Blacklow, S

    Hermes, J. D., Blacklow, S. C., and Knowles, J. R. Searching sequence space by definably random mutagenesis: improving the catalytic potency of an enzyme. Proceedings of the National Academy of Sciences, 87 0 (2): 0 696--700, 1990

    1990

  7. [15]

    P., Glorot, X., Botvinick, M

    Higgins, I., Matthey, L., Pal, A., Burgess, C. P., Glorot, X., Botvinick, M. M., Mohamed, S., and Lerchner, A. beta-vae: Learning basic visual concepts with a constrained variational framework. ICLR (Poster), 3, 2017

    2017

  8. [16]

    Denoising diffusion probabilistic models

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 0 6840--6851, 2020

    2020

  9. [17]

    Ikram, Z., Liu, D., and Rahman, M. S. Antibody sequence optimization with gradient-guided discrete walk-jump sampling. In ICLR 2024 Workshop on Generative and Experimental Perspectives for Biomolecular Design, 2024

    2024

  10. [18]

    F., Ekbote, C

    Jain, M., Bengio, E., Hernandez-Garcia, A., Rector-Brooks, J., Dossou, B. F., Ekbote, C. A., Fu, J., Zhang, T., Kilgour, M., Zhang, D., et al. Biological sequence design with gflownets. In International Conference on Machine Learning, pp.\ 9786--9801. PMLR, 2022

    2022

  11. [19]

    R., Fu, X., Viknander, S., Goldin, C., Monaco, S., Zelezniak, A., and Yang, K

    Johnson, S. R., Fu, X., Viknander, S., Goldin, C., Monaco, S., Zelezniak, A., and Yang, K. K. Computational scoring and experimental evaluation of enzymes generated by neural networks. biorxiv. preprint, 202310 0 (2023.03): 0 04--531015, 2023

    2023

  12. [20]

    E., Fannjiang, C., Wittmann, B

    Johnston, K. E., Fannjiang, C., Wittmann, B. J., Hie, B. L., Yang, K. K., and Wu, Z. Machine learning for protein engineering. In Machine Learning in Molecular Sciences, pp.\ 277--311. Springer, 2023

    2023

  13. [21]

    Kingma, D. P. and Welling, M. Auto-encoding variational bayes, 2022. URL https://arxiv.org/abs/1312.6114

    2022 arXiv

  14. [22]

    S., Barzilay, R., and Fiete, I

    Kirjner, A., Yim, J., Samusevich, R., Bracha, S., Jaakkola, T. S., Barzilay, R., and Fiete, I. R. Improving protein optimization with smoothed fitness landscapes. In The Twelfth International Conference on Learning Representations, 2023

    2023

  15. [23]

    F., Jung, H., Ro, H

    Lee, M., Vecchietti, L. F., Jung, H., Ro, H. J., Cha, M., and Kim, H. M. Robust optimization in protein fitness landscapes using reinforcement learning in latent space. arXiv preprint arXiv:2405.18986, 2024

    2024 arXiv

  16. [24]

    Evolutionary-scale prediction of atomic-level protein structure with a language model

    Lin, Z., Akin, H., Rao, R., Hie, B., Zhu, Z., Lu, W., Smetanin, N., Verkuil, R., Kabeli, O., Shmueli, Y., et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science, 379 0 (6637): 0 1123--1130, 2023

    2023

  17. [25]

    Lipman, Y., Chen, R. T. Q., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling. In International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=PqvMRDCJT9t

    2023

  18. [26]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Liu, X., Gong, C., and Liu, Q. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022

    2022 arXiv

  19. [27]

    N., Marks, D., and Gal, Y

    Notin, P., Dias, M., Frazer, J., Marchena-Hurtado, J., Gomez, A. N., Marks, D., and Gal, Y. Tranception: protein fitness prediction with autoregressive transformers and inference-time retrieval. In International Conference on Machine Learning, pp.\ 16990--17017. PMLR, 2022

    2022

  20. [28]

    Proteingym: Large-scale benchmarks for protein fitness prediction and design

    Notin, P., Kollasch, A., Ritter, D., Van Niekerk, L., Paul, S., Spinner, H., Rollins, N., Shaw, A., Orenbuch, R., Weitzman, R., et al. Proteingym: Large-scale benchmarks for protein fitness prediction and design. Advances in Neural Information Processing Systems, 36: 0 64331--...

    2023

  21. [29]

    Praljak, N., Lian, X., Ranganathan, R., and Ferguson, A. L. Protwave-vae: Integrating autoregressive sampling with latent-based inference for data-driven protein design. ACS synthetic biology, 12 0 (12): 0 3544--3561, 2023

    2023

  22. [30]

    Proximal exploration for model-guided protein sequence design

    Ren, Z., Li, J., Ding, F., Zhou, Y., Ma, J., and Peng, J. Proximal exploration for model-guided protein sequence design. In International Conference on Machine Learning, pp.\ 18520--18536. PMLR, 2022

    2022

  23. [31]

    Rezende, D. J. and Viola, F. Taming vaes. arXiv preprint arXiv:1810.00597, 2018

    2018 arXiv

  24. [32]

    Romero, P. A. and Arnold, F. H. Exploring protein fitness landscapes by directed evolution. Nature reviews Molecular cell biology, 10 0 (12): 0 866--876, 2009

    2009

  25. [33]

    S., Bolotin, D

    Sarkisyan, K. S., Bolotin, D. A., Meer, M. V., Usmanova, D. R., Mishin, A. S., Sharonov, G. V., Ivankov, D. N., Bozhanova, N. G., Baranov, M. S., Soylemez, O., et al. Local fitness landscape of the green fluorescent protein. Nature, 533 0 (7603): 0 397--401, 2016

    2016

  26. [34]

    and Kelsic, E

    Sinai, S. and Kelsic, E. D. A primer on model-guided exploration of fitness landscapes for biological sequence design. arXiv preprint arXiv:2010.10614, 2020

    2010 arXiv

  27. [35]

    Sinai, S., Wang, R., Whatley, A., Slocum, S., Locane, E., and Kelsic, E. D. Adalead: A simple and robust adaptive greedy search algorithm for sequence design. arXiv preprint arXiv:2010.02141, 2020

    2010 arXiv

  28. [36]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., and Ganguli, S. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pp.\ 2256--2265. PMLR, 2015

    2015

  29. [37]

    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

  30. [38]

    Closed-form test functions for biophysical sequence optimization algorithms

    Stanton, S., Alberstein, R., Frey, N., Watkins, A., and Cho, K. Closed-form test functions for biophysical sequence optimization algorithms. arXiv preprint arXiv:2407.00236, 2024

    2024 arXiv

  31. [39]

    Conservative objective models for effective offline model-based optimization

    Trabucco, B., Kumar, A., Geng, X., and Levine, S. Conservative objective models for effective offline model-based optimization. In International Conference on Machine Learning, pp.\ 10358--10368. PMLR, 2021

    2021

  32. [40]

    V., Ngo, N

    Tran, T. V., Ngo, N. K., Nguyen, V. T. D., and Hy, T. S. Latentde: Latent-based directed evolution accelerated by gradient ascent for protein sequence design. In NeurIPS 2024 Workshop on AI for New Drug Modalities, 2024

    2024

  33. [41]

    and Weissman, D

    Van Cleve, J. and Weissman, D. B. Measuring ruggedness in fitness landscapes. Proceedings of the National Academy of Sciences, 112 0 (24): 0 7345--7346, 2015

    2015

  34. [42]

    T., Moriconi, R., Hutter, F., and Deisenroth, M

    Wilson, J. T., Moriconi, R., Hutter, F., and Deisenroth, M. P. The reparameterization trick for acquisition functions. arXiv preprint arXiv:1712.00424, 2017

    2017 arXiv

  35. [43]

    G., Bowden, J

    Yang, J., Lal, R. G., Bowden, J. C., Astudillo, R., Hameedi, M. A., Kaur, S., Hill, M., Yue, Y., and Arnold, F. H. Active learning-assisted directed evolution. Nature Communications, 16 0 (1): 0 714, 2025

    2025

  36. [44]

    Zheng, Q., Le, M., Shaul, N., Lipman, Y., Grover, A., and Chen, R. T. Guided flows for generative modeling and decision making. arXiv preprint arXiv:2311.13443, 2023

    2023 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.