acceptodds
Under review as a conference paper at ICLR 2027

MuseDrift: Navigating Protein Evolutionary Manifolds with Conditional Discrete Diffusion

Abstract

Protein engineering often requires variants that differ from a wild-type (WT) sequence by a specified amount while remaining structurally and functionally plausible. Controlling WT similarity alone does not specify which residues should change or how their substitutions should be coordinated. Here, we introduce MuseDrift, a conditional discrete diffusion model that generates variants from a WT sequence and a target sequence identity, without task-specific property guidance. Trained on WT-homolog pairs spanning different identities, MuseDrift uses pretrained WT representations as context and identity-aware modulation to condition denoising. Compared with editing baselines at the same requested identities, its variants achieve higher mean structural prediction confidence and greater similarity to predicted WT structures. On the CAMEO dataset, it tracks requested identities spanning 30%-95%, with mean absolute errors of 0.52-2.08 percentage points across targets. Its generated variants also receive more favorable scores from external mutation-effect predictors than those from the evaluated baselines. Together, these results support MuseDrift as a framework for controlled exploration of WT-centered protein sequence space.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein design, protein sequence generation, discrete diffusion, controllable generation, protein engineering, protein language models, protein evolution

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_HhPzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes MuseDrift, a conditional discrete diffusion model for generating protein variants with a specified sequence identity to a wild-type (WT) protein. The authors construct a large training dataset of 38.2 million WT–homolog pairs and train a masked-token denoising model that combines WT sequence context, frozen ESM-2 representations, and identity-aware FiLM conditioning. During inference, the model iteratively generates variants while learning which positions to mutate and which amino acids to substitute.

Critical strengths

Conditioning protein variant generation directly on a target sequence identity is an interesting idea. The paper also provides extensive experiments across different identity levels, evaluating structural confidence, structural similarity, sequence diversity, evolutionary patterns, and predicted fitness and energy.

Critical weaknesses
  1. The model is trained using MMseqs2 local-alignment identity but evaluated using full-length Hamming identity. These two definitions can differ substantially, especially for partial alignments and sequences containing insertions or deletions. Surprisingly, the model still achieves highly accurate identity control, but the paper does not explain why this transfer works so well or quantify the discrepancy between the two metrics in the training data. More importantly, I question whether Hamming identity alone is a sufficiently meaningful measure of evolutionary distance for protein engineering. It treats all amino acid substitutions equally, even though mutations at highly conserved or functionally important positions can have much larger effects than mutations at flexible or variable positions. Similarly, conservative and chemically disruptive substitutions contribute equally to the distance. Although the model may learn some of these preferences from evolutionary data, the identity condition itself does not capture them. The authors should better justify this choice and investigate whether more biologically informative measures could provide more meaningful control over protein variation.

  2. The paper mainly evaluates generated variants using structural confidence, similarity to the WT structure, and computational fitness or energy scores. However, as the target identity decreases and more mutations are introduced, pLDDT and predicted fitness scores consistently become worse, while FoldX energy changes become less favorable. Although MuseDrift performs better than the baselines at the same identity levels, the results largely show that its mutations are less damaging, rather than that they provide any actual benefit over the WT. This raises a fundamental question about the practical value of the proposed method. If variants closer to the WT consistently achieve better scores, why not simply retain the original WT sequence? While generating diverse variants is a reasonable objective, the paper does not demonstrate how this diversity translates into useful improvements in protein engineering. For example, there is no evidence that the generated variants achieve better activity, stability, solubility, or other desirable properties than the WT.

Questions for authors
  1. Could the authors clarify why MMseqs2 local-alignment identity is used during training, while full-length Hamming identity is used during inference? How does this definition shift affect the learned conditioning? Moreover, the predicted structural confidence and fitness scores generally deteriorate as target identity decreases. Although MuseDrift performs better than the baselines, what practical benefits do the generated mutations offer over simply retaining the WT sequence? Could the authors provide evidence that the generated variants improve at least some desirable protein properties, even at the expense of others?
  2. Most of the evaluated baselines are relatively simple or based on older models. Could the authors consider comparisons with more recent methods specifically designed for protein evolution or reference-conditioned sequence editing, such as EvoFlows, PEINT, or DPLM-Evo?
3 comments
AuthorsOpenReview ↗

Identity Control and Practical Use (Part 1)

Thank you for the review. We measured the identity mismatch, checked mutation-count control, and clarified the intended use.

1. Why different identity measures? Training uses the MMseqs2 identity labels from our homolog-pair dataset. Generation produces substitutions in a full WT sequence, so we use full-length Hamming identity to count the changes.

We checked 2.91M held-out pairs. We measured identity over the full WT, counting unaligned WT positions as mismatches. For gapped pairs, this is an alignment-based measure, not positional Hamming identity.

Pair typeShareMean training label minus full-WT measureCorrelation
Ungapped, full WT coverage15.1%-0.00051.000
Ungapped, partial coverage25.8%+0.0750.909
Gapped59.1%+0.0460.927
All100%+0.0460.947

The labels are higher by 0.046 on average; the mean absolute difference is 0.053. The measures are strongly correlated, but correlation alone does not guarantee accurate control.

2. Does identity control work? We tested the original sampler on 438 CAMEO WTs, with 64 variants per WT per target.

Target identity0.300.400.550.750.95
Mean absolute error0.0210.0160.0130.0110.005
Within +/-0.0257.9%72.2%79.1%86.1%98.5%
Within +/-0.0594.1%97.6%98.9%99.4%99.4%

This shows approximate control directly on generated sequences. We also tested an exact-count decoder on 354 WTs. It always meets the requested edit count, but lowers pLDDT at the lower identity targets.

TargetExact count met: original / exact-countpLDDT: original minus exact-count [95% CI]
0.9559% / 100%+0.1 [-0.1, +0.4]
0.7534% / 100%+2.5 [+2.0, +3.1]
0.5531% / 100%+4.4 [+3.6, +5.2]
0.3022% / 100%+8.4 [+7.3, +9.4]

CIs use paired WT bootstrap (20,000 draws). Exact-count decoding is optional; the main structural results use the original sampler.

3. What does identity mean biologically? It controls the number of substitutions. For example, identity 0.95 on a 300-residue protein means 15 substitutions. It does not measure how each change affects function. The model learns site and residue choices from homolog pairs.

At target 0.95, a position-matched comparison under the shared exact-edit decoder gives mean BLOSUM62 scores of +0.43 for MuseDrift, -0.41 for ESM3, and -0.48 for ProteinMPNN. MuseDrift's exact-count decoder places 3.8% of edits in conserved columns, versus 9.5% for random sites. These results show nonrandom mutation choices, not proven functional gains. Control by conservation-weighted distance remains untested.

4. Why generate variants instead of keeping WT? MuseDrift builds candidate libraries from a WT sequence and target identity, without needing a structure template. Researchers can choose how far to explore and then screen the variants for a desired property. This serves the library-generation step of directed evolution (Packer & Liu, 2015; Romero & Arnold, 2009).

We have not demonstrated improved experimental activity, stability, solubility, or screening success. At identity 0.95, ProSST-4096 predicts at least one variant above WT among 64 samples for 15.9% of WTs, versus 1.6-7.2% for the original baselines. At targets 0.75 and below, the original additive ProSST/VenusREM evaluation found no variants above WT (0/27,648 per method-target cell).

In a separate common-decoder evaluation at 0.95, the predicted above-WT fractions favor ESM3 (0.88%) and ProteinMPNN (0.81%) over MuseDrift (0.25%). These are computational predictions from different evaluation protocols; they do not establish experimental improvement.

References: Packer & Liu (2015), Nat Rev Genet 16:379-394; Romero & Arnold (2009), Nat Rev Mol Cell Biol 10:866-876.

AuthorsOpenReview ↗

Comparisons with Newer Models (Part 2)

We added ESM3, ProteinMPNN, and PEINT. Official checkpoints for EvoFlows and DPLM-Evo are not publicly available.

5. ESM3 and ProteinMPNN. MuseDrift generates variants from a WT sequence and target identity, without a structure template. Including its frozen WT encoder, MuseDrift has 128.5M parameters (95M trainable), compared with ESM3's 1.4B.

We evaluated the first sample per WT and target on 429 WTs at T=1.0. MuseDrift uses its original sampler without enforcing exact counts. ESM3 and ProteinMPNN choose sites using their own 1 - p(WT residue), then exclude WT residues at those sites to meet an exact edit count. ESM3 receives only the WT sequence; ProteinMPNN also receives the ESMFold-predicted WT backbone.

Differences are MuseDrift minus baseline; brackets give 95% paired WT-bootstrap CIs (20,000 draws).

TargetpLDDT vs ESM3TM-score vs ESM3pLDDT vs ProteinMPNN
0.95+0.3 [+0.05, +0.6]+0.013 [+0.006, +0.019]0.0 [-0.3, +0.2]
0.75+1.2 [+0.5, +1.9]+0.026 [+0.017, +0.036]-0.3 [-0.8, +0.2]
0.550.0 [-1.0, +1.0]+0.032 [+0.020, +0.044]-1.5 [-2.2, -0.8]
0.30-3.7 [-4.9, -2.5]+0.008 [-0.005, +0.021]-1.6 [-2.8, -0.4]

MuseDrift has higher TM-score than ESM3 at 0.95, 0.75, and 0.55, but lower pLDDT at 0.30. ProteinMPNN has higher pLDDT at 0.55 and 0.30; its mean TM-score is also higher there (differences -0.026/-0.035).

Sequence quality. With all three models using the same exact-edit decoder at T=1.0, MuseDrift has higher mean homolog-column support at all four targets (354 WTs). Mean differences across these targets are 0.0072-0.0216 versus ESM3 and 0.0135-0.0424 versus ProteinMPNN. Support counts substitutions observed in natural homologs and is related to our training objective.

ProSST-2048 (348 WTs) favors ESM3 at every target. Compared with ProteinMPNN, MuseDrift has higher mean scores at 0.95/0.75 and lower scores at 0.30; the difference at 0.55 is not significant. These are predicted mutation effects, not measured fitness.

6. PEINT. We ran the official code/checkpoint with default sampling on 429 WTs and the same ESMFold settings. We calibrated its evolutionary-time input to identity on 30 of the 438 CAMEO WTs, using identity only.

TargetMuseDrift identityPEINT identityPEINT length differspLDDT vs PEINTTM-score vs PEINT
0.950.95000.90648%+0.67 [+0.27, +1.09]+0.031 [+0.023, +0.040]
0.750.75180.74276%+0.15 [-0.48, +0.80]+0.029 [+0.018, +0.040]
0.550.55150.53789%+0.17 [-0.65, +0.98]+0.033 [+0.022, +0.044]
0.300.30450.26598%+9.45 [+8.08, +10.87]+0.131 [+0.115, +0.147]

Identities are means; length changes are relative to WT. TM-score is WT-length normalized. At 0.95/0.75, one three-residue PEINT output per target could not be TM-scored (n=428); assigning zero leaves the conclusions unchanged. MuseDrift has higher TM-score throughout, but achieved identities and lengths differ, so this is not a strictly matched comparison.

Sampling affects the ranking. Greedy ESM3 gives pLDDT 84-85 across targets on 429 WTs, above MuseDrift. PEINT with nucleus sampling (p=0.90; 150 WTs) reaches identity 0.964 at target 0.95 and exceeds MuseDrift in pLDDT (-0.53 [-0.90, -0.16]). On 354 WTs, MuseDrift's exact-count decoder scores below default PEINT at 0.75 (78.7 vs 81.1) and 0.55 (72.1 vs 76.5).

We will add these comparisons and present MuseDrift as a sequence-based library generator, with quality depending on the target, metric, and sampler.

Thank you for the detailed response and additional experiments. I appreciate the authors' efforts to quantify the identity mismatch, validate identity control, and include stronger baselines. Some important issues remain:

  1. Positioning MuseDrift as a library generator clarifies its intended use, but does not establish its practical advantage. The new analyses show nonrandom mutation preferences, but not improved functional properties or screening outcomes. More importantly, the predicted benefits over WT remain limited: no above-WT variants were found at identities of 0.75 or below in the original evaluation.

  2. The new comparisons reveal mixed results across metrics and identity levels. Furthermore, MuseDrift uses its original sampler while ESM3 and ProteinMPNN enforce exact mutation counts, even though the authors' own ablations show that exact-count decoding can substantially reduce pLDDT. This raises concerns about whether the reported improvements reflect the model itself or differences in decoding constraints.

I am therefore inclined to maintain my rating.

Reviewer_LqrVOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces MuseDrift, a conditional discrete diffusion model to generate variants of wild-type (WT) sequences. The core idea is to condition the denoising process with prefix WT tokens, ESM2 features of the WT sequence and a soft target sequence identity that measures how close the generated samples are to the WT. Additionally, the paper also proposes a dataset of 38.2M aligned WT-homolog pairs.

Critical strengths
  1. The paper is clearly written and easy to follow.
  2. The description of data construction is detailed.
  3. The modeling idea looks plausible to me. Using closeness information to control the generation of WT variants makes sense.
Critical weaknesses
  1. My main concern is the selection of baselines. In the related work section, the paper mentions several methods for reference-conditioned editing and evolution, including EvoFlows, PEINT, DPLM-Evo, and ESM-3. However, the paper doesn't compare the proposed method with any of them. While tables 1 and 2 show large gains of MuseDrift compared to other baselines, I would have expected more comprehensive comparisons with those newer approaches.
Questions for authors
  1. How sensitive are the metrics in tables 1 and 2?
1 comment
AuthorsOpenReview ↗

Additional Baselines and Evaluation Sensitivity

Thank you for the review. We added ESM3, ProteinMPNN, and PEINT. Official checkpoints for EvoFlows and DPLM-Evo are not publicly available.

1. ESM3 and ProteinMPNN. MuseDrift generates variants from a WT sequence and target identity, without a structure template. Including its frozen WT encoder, MuseDrift has 128.5M parameters (95M trainable), compared with ESM3's 1.4B.

We evaluated the first sample per WT and target on 429 WTs at T=1.0. MuseDrift uses its original sampler without enforcing exact counts. ESM3 and ProteinMPNN choose sites using their own 1 - p(WT residue), then exclude WT residues at those sites to meet an exact edit count. ESM3 receives only the WT sequence; ProteinMPNN also receives the ESMFold-predicted WT backbone.

Differences are MuseDrift minus baseline; brackets give 95% paired WT-bootstrap CIs (20,000 draws).

TargetpLDDT vs ESM3TM-score vs ESM3pLDDT vs ProteinMPNN
0.95+0.3 [+0.05, +0.6]+0.013 [+0.006, +0.019]0.0 [-0.3, +0.2]
0.75+1.2 [+0.5, +1.9]+0.026 [+0.017, +0.036]-0.3 [-0.8, +0.2]
0.550.0 [-1.0, +1.0]+0.032 [+0.020, +0.044]-1.5 [-2.2, -0.8]
0.30-3.7 [-4.9, -2.5]+0.008 [-0.005, +0.021]-1.6 [-2.8, -0.4]

MuseDrift has higher TM-score than ESM3 at 0.95, 0.75, and 0.55, but lower pLDDT at 0.30. ProteinMPNN has higher pLDDT at 0.55 and 0.30; its mean TM-score is also higher there (differences -0.026/-0.035).

2. Sequence quality. With all three models using the same exact-edit decoder at T=1.0, MuseDrift has higher mean homolog-column support at all four targets (354 WTs). Mean differences across these targets are 0.0072-0.0216 versus ESM3 and 0.0135-0.0424 versus ProteinMPNN. Support counts substitutions observed in natural homologs and is related to our training objective.

ProSST-2048 (348 WTs) favors ESM3 at every target. Compared with ProteinMPNN, MuseDrift has higher mean scores at 0.95/0.75 and lower scores at 0.30; the difference at 0.55 is not significant. These are predicted mutation effects, not measured fitness.

3. PEINT. We ran the official code/checkpoint with default sampling on 429 WTs and the same ESMFold settings. We calibrated its evolutionary-time input to identity on 30 of the 438 CAMEO WTs, using identity only.

TargetMuseDrift identityPEINT identityPEINT length differspLDDT vs PEINTTM-score vs PEINT
0.950.95000.90648%+0.67 [+0.27, +1.09]+0.031 [+0.023, +0.040]
0.750.75180.74276%+0.15 [-0.48, +0.80]+0.029 [+0.018, +0.040]
0.550.55150.53789%+0.17 [-0.65, +0.98]+0.033 [+0.022, +0.044]
0.300.30450.26598%+9.45 [+8.08, +10.87]+0.131 [+0.115, +0.147]

Identities are means; length changes are relative to WT. TM-score is WT-length normalized. At 0.95/0.75, one three-residue PEINT output per target could not be TM-scored (n=428); assigning zero leaves the conclusions unchanged. MuseDrift has higher TM-score throughout, but achieved identities and lengths differ, so this is not a strictly matched comparison.

4. Sensitivity. Sampling affects the ranking. Greedy ESM3 gives pLDDT 84-85 across targets on 429 WTs, above MuseDrift. PEINT with nucleus sampling (p=0.90; 150 WTs) reaches identity 0.964 at target 0.95 and exceeds MuseDrift in pLDDT (-0.53 [-0.90, -0.16]). On 354 WTs, MuseDrift's exact-count decoder scores below default PEINT at 0.75 (78.7 vs 81.1) and 0.55 (72.1 vs 76.5).

For Tables 1 and 2, we also resampled homology groups rather than individual WTs. MMseqs2 gives 348/373 groups at 30%/50% identity and >=80% coverage among 429 WTs. Of 135 comparisons with the original baselines and ablation, 133 remain significant under all three bootstrap schemes. No direction or significance decision changes; median CI widths increase 1.44x/1.34x. This group analysis does not include the new baselines.

These comparisons show that quality depends on the target, metric, and sampler. We will add the results and clarify this scope.

Reviewer_FDxfOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The authors present MuseDrift, a discrete diffusion model that generates mutated sequences given a fixed desired sequence identity to a wild-type sequence. The model is trained on homologous sequence pairs. The experiments show that, compared to baselines, MuseDrift can sample sequences with the requested sequence identity, while retaining good pLDDT and TM-Score to the wild-type structure even at low sequence identity. The generated sequences also retain better scores under general mutation effect predictors.

Critical strengths

The technical approach and experiments seem sound and demonstrate improvement over baselines in terms of the considered metrics. For example, MuseDrift retains a pLDDT of 66 at 40% sequence identity to the wild-type, compared to 40 for the best baseline.

Critical weaknesses

The main weakness of this paper is a lack of motivation for the problem. While the approach and experiments demonstrate that MuseDrift can generate reasonable sequences (in terms of the pLDDTs of their predicted structures) with a targeted sequence identity to a wild-type sequence, it is unclear why this is useful or desirable. Usually the goal is to optimize a protein for a particular function, and natural proteins usually evolve under the selective pressure of optimizing a particular function, yet MuseDrift is not conditioned on any such function to optimize. And how does sequence identity control help when exploring variants? It seems like an ancillary quantity to control when optimizing a sequence, as opposed to an uncertainty about the protein's properties for example. The paper would be significantly strengthened if the authors could demonstrate how MuseDrift actually enables better protein design. For example, can MuseDrift be used to optimize a particular property? Or can it be used to optimize a binder, something for which there exist well-established in-silico evaluation metrics?

Questions for authors

Apart from addressing the concerns raised above, I have the following questions about the baselines

  • Does ESM2 handle the WT prefix natively given how it was trained? And how is the sequence identity controlled in that case, or are only the edit sites filled in?
  • Why does the ESM-2 selected-site infilling baseline decode in a single forward pass instead of several steps like the other ESM2 baseline?
  • The fixed edit sites seems very restrictive. Instead of fixing random edit sites, what about using steering methods that would just bias towards/away from WT residues to accomplish a particular sequence identity?
  • Have the authors tried simple structured-based baselines? For example take the predicted target structure and redesign the sequence with ProteinMPNN? This is already well-established for optimizing binding (see ProteinHunter [1])
3 comments
AuthorsOpenReview ↗

Motivation and Identity-Controlled Libraries (1/2)

Thank you for the careful review. Your main concern is motivation, so we start there.

Why identity control? MuseDrift is a generator for directed-evolution libraries, not a function optimizer.

In directed evolution, researchers start from a wild type (WT), introduce mutations to build a variant library, and screen it for desired properties (Packer & Liu, 2015; Romero & Arnold, 2009). The assay selects for function. The library supplies the candidates, so useful sequence diversity and the ability to retain a fold matter (Bloom et al., 2006; Wittmann et al., 2021).

Requested identity sets the mutation load: 0.95 on a 300-residue protein corresponds to 15 substitutions. It acts as an evolutionary dial for sequence divergence. Adjusting it lets the same model explore nearby or more distant WT-centered neighborhoods. The model decides where to mutate and what residues to place, using preferences learned from natural WT-homolog pairs.

The original sampler tracks this requested load in the submitted CAMEO evaluation (438 unique WTs; 64 variants per reference record and target):

Target identity0.300.400.550.750.95
Mean absolute identity error0.0210.0160.0130.0110.005
Variants within +/-0.02 of target57.9%72.2%79.1%86.1%98.5%
Variants within +/-0.05 of target94.1%97.6%98.9%99.4%99.4%

The intended use is to supply evolution-informed candidate libraries for subsequent screening. MuseDrift does not prescribe a functional direction or require campaign-specific assay labels. Researchers can generate libraries at different distances and screen them for activity, stability, binding, or other goals. We therefore evaluate identity control, library diversity, and computational plausibility, rather than claiming a specific functional optimum.

We performed no wet-lab screening and do not claim that the variants outperform the WT, that MuseDrift designs binders, or that it has demonstrated improved experimental screening efficiency.

References

Packer MS, Liu DR (2015). Nat Rev Genet 16:379-394.

Romero PA, Arnold FH (2009). Nat Rev Mol Cell Biol 10:866-876.

Bloom JD, Labthavikul ST, Otey CR, Arnold FH (2006). PNAS 103:5869-5874.

Wittmann BJ, Yue Y, Arnold FH (2021). Cell Syst 12:1026-1045.e7.

AuthorsOpenReview ↗

ESM-2 Decoding, Model-Selected Edit Sites, and ProteinMPNN (2/2)

Q1. ESM-2 baselines. ESM-2 was not trained specifically for WT-prefix-conditioned generation. Our full-mask protocol provides a visible WT prefix and reveals the variant over up to 50 random-order steps. Selected-site infilling instead masks only the chosen edit sites within the WT. Both enforce an exact edit count: unedited positions retain the WT residue, and edited positions exclude it.

Q2. Why one forward pass? The single-pass baseline predicts all masked sites simultaneously from the visible WT. To isolate the effect of decoding rounds, we now refill the same selected edit sites in 5 or 10 rounds (150 WTs, one variant per WT and setting):

IdentityMetricSingle pass5 rounds10 roundsMuseDrift
0.30pLDDT34.031.732.666.2
0.30TM-score0.2610.2740.2790.711
0.55pLDDT68.168.168.476.3
0.55TM-score0.7320.7340.7310.821
0.95pLDDT84.284.184.384.5
0.95TM-score0.9310.9360.9310.943

Only 12-35% of edited residues match the single-pass outputs. Nevertheless, at identities 0.55 and 0.95, pLDDT differences from single-pass filling are within 0.3 and their confidence intervals include zero. At 0.30, multi-step filling is 1.4-2.3 points lower. MuseDrift retains higher mean scores at 0.30 and 0.55; pLDDT means are close at 0.95.

Q3 and Q4. Model-selected edit sites and ProteinMPNN. The original baselines use random edit sites. In the new comparison, ESM3 and ProteinMPNN select sites using their own 1 - p(WT residue), then exclude the WT residue at those sites. This tests model-based site selection rather than uniform random placement; it is not logit-bias steering.

The common-decoder experiment uses T=1.0 and 354 WTs. ProteinMPNN receives an ESMFold-predicted WT backbone; MuseDrift receives no explicit structure. The pLDDT comparison is separate: it uses 150 WTs and MuseDrift's original sampler, while ESM3 and ProteinMPNN use the protocol above. Thus, the structural comparison does not use a shared decoder across all three models.

All entries below are MuseDrift minus the named comparator. Brackets give pLDDT confidence intervals.

IdentityComparatorSupportpLDDTProSST-2048
0.95ESM3+0.0216+0.2 [-0.1, +0.6]-2.0
0.95ProteinMPNN+0.0424+0.2 [-0.1, +0.6]+4.0
0.75ESM3+0.0190+1.9 [+0.7, +3.1]-19.7
0.75ProteinMPNN+0.0347+0.2 [-0.7, +1.0]+5.6
0.55ESM3+0.0150+0.4 [-1.3, +2.2]-39.8
0.55ProteinMPNN+0.0284-1.3 [-2.5, -0.1]-1.7 (n.s.)
0.30ESM3+0.0072-2.8 [-4.5, -0.9]-74.3
0.30ProteinMPNN+0.0135+0.7 [-1.0, +2.4]-43.5

Support is the fraction of substituted residues observed in the corresponding homolog-alignment columns. This measure aligns with MuseDrift's homolog-based training objective and is not independent functional validation. MuseDrift has higher mean support in all comparisons. ProSST-2048 measures predicted substitution effects, not experimental fitness; n.s. denotes a nonsignificant difference.

The stronger baselines give mixed results. MuseDrift has higher pLDDT than ESM3 at 0.75, lower pLDDT than ESM3 at 0.30, and lower pLDDT than ProteinMPNN at 0.55. The remaining pLDDT intervals include zero. ESM3 has higher mean ProSST scores throughout; ProteinMPNN has higher means at 0.30 and 0.55, but the latter difference is not significant. We do not claim overall superiority. Logit-bias steering of a diffusion model remains untested.

The original approximately 66-versus-40 pLDDT comparison is at identity 0.30 against EvoDiff, not these stronger baselines. We will add the new comparisons, distinguish their evaluation protocols, and revise the motivation and claims accordingly.

I thank the authors for their response. I am still not convinced by the motivation of the work, as I don't see the advantage of producing a fixed mutational burden if there is no relation to better functional outcomes. I understand that the approach is agnostic to a specific function to optimize, but demonstrating that the libraries produced by MuseDrift have some kind of utility as a directed evolution library, even if just based on in silico measures, would be important to convince readers of the utility of the method. Otherwise, why not just use methods like ESM3 to propose mutations? Given this and the mixed results compared to stronger baselines I will maintain my original score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.