acceptodds
Under review as a conference paper at ICLR 2027

Nonparametric Biologic Design via Discrete Drifting

Abstract

Recent algorithms for sequence-based therapeutic design use diffusion and flow models as statistical priors for optimization in combinatorial spaces. Adapting these priors to new reference datasets can require additional training. We explore whether a frozen representation encoder and reference data can support such adaptation without fitting a generative model. We introduce **NimblDrift**, a *discrete* feature-space drifting algorithm for generating and optimizing populations of biologic candidates. At each step, we select local edits that reduce a discrepancy between the candidate and reference populations in feature space. On three small-molecule targets, the best-scoring drift configurations exceeded greedy-predictor and random-walk baselines in predicted binding. With **GreedyDrift**, we also improved predicted peptide affinity while retaining reference-like distributions of secondary properties. Allele-specific peptide generation further tests the functional information accessible through representation geometry. For practical design, NimblDrift supports modular search with frozen encoders and replaceable reference populations. For representation research, these experiments offer a direct way to compare encoders through their effects on discrete generation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
generative-biology, representation-encoder, peptide, molecules, distribution-matching

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

This paper introduces NimblDrift, a nonparametric method for discrete biological sequence generation that uses frozen pretrained encoders and reference populations to guide generation without needing to to train generative model. The method iteratively applies discrete candidate edits to minimize the distance between generated and reference distributions in the embedding space, and further extends this framework through GreedyDrift to incorporate reward optimization with reference-based regularization. The paper evaluates the approach across a few tasks including sentence sentiment transformation, small-molecule binding, therapeutic peptide optimization and allele-specific MHC peptide generation.

Critical strengths
  1. This paper provides broad empirical evaluation across diverse domains with evaluation that also considers multiple pretrained encoders, reference distributions, and search neighborhood constructions. This demonstrates good flexibility of the method. Additionally the result in biological design and property preservation are encouraging.

2.The method is straightforward and simple. The adaptability for different biological design tasks and objectives without requiring model training is appealing.

Critical weaknesses
  1. A very significant limitation of this work is the limited comparison against natural alternatives, such as nearest reference optimization or pure local search guided by the embedding or other evolutionary search approaches. Without these baselines, the evaluation is incomplete.

  2. The figures are difficult to see and interpret, quite blurry and very dense. Not suitable for the intended purpose (Figure 2 and 3 For Example).

  3. The work is quite incremental with respect to "Generative Modeling via Drifting" [1]. There are some useful extension and interesting applications, but this work is too incremental.

[1] Deng, M., Li, H., Li, T., Du, Y., & He, K. (2026). Generative Modeling via Drifting. arXiv preprint arXiv:2602.04770. https://arxiv.org/abs/2602.04770

Questions for authors
  1. It would be helpful for the authors to include comparisons against simpler search-based baselines, such as nearest-reference embedding search and reference attraction only optimization. These comparisons would help isolate the benefits of population level distribution matching over simpler embedding-guided local search approaches.

  2. Could the authors further clarify the methodological novelty of NimblDrift relative to the original continuous Drifting framework (Deng et al., 2026) and existing greedy search methods? In particular, beyond adapting feature-space distribution matching to discrete spaces through local search and incorporating reward-based regularization, what are the key novel methodological contributions of the proposed approach?

5 comments

We thank the reviewer for the constructive feedback and for recognizing the flexibility and simplicity of NimblDrift. To address the baseline and methodological questions, we undertook new experiments with nearest-reference search, mean-shift attraction, evolutionary search, fine-tuned generative models, and additional encoders. We describe the new comparisons first, then clarify the algorithmic contribution and planned presentation improvements.

Q1. Simpler embedding-guided search

“It would be helpful for the authors to include comparisons against simpler search-based baselines, such as nearest-reference embedding search and reference attraction only optimization.”

We implemented both suggested controls. Fixed nearest-reference search moves each candidate toward the reference closest to its initial embedding. Mean-shift attraction instead moves the candidate toward a kernel-weighted average of the reference embeddings. These baselines update candidates individually. NimblDrift selects replacements by evaluating the full candidate-population discrepancy against the reference distribution.

Table R15. Predicted MHC class I binder percentages.

AlleleRandomFixed nearest referenceMean-shift attractionNimblDrift
A*01:014.3%13.2%14.7%21.3%
A*02:012.7%6.0%12.3%11.2%
A*03:012.7%10.0%10.0%19.2%

NimblDrift exceeds fixed nearest-reference search on all three alleles and mean-shift attraction on A*01:01 and A*03:01. For A*02:01, mean shift reaches 12.3% and NimblDrift reaches 11.2%, both substantially above the 2.7% initialization. The comparisons therefore identify the settings in which population coupling contributes most clearly.

Table R16. Mean Boltz-2 binding likelihood in the molecular search ablation.

TargetRandomFixed nearest referenceMean-shift attractionNimblDrift
2VT40.1990.2160.2590.466

For 2VT4, population matching raises the score to 0.466, compared with 0.216 for nearest-reference search and 0.259 for mean-shift attraction. This gives a 79.9% increase over mean-shift attraction, directly addressing the reviewer's question about the value of coupling the candidate updates.

New evolutionary and generative comparisons

To extend the evaluation further, we compared against a genetic algorithm and fine-tuned peptide generators across nine allele-seed pairs.

Table R17. Aggregate MHC class I results from the generation and encoder comparisons. Lower NetMHCpan percentile ranks indicate stronger predicted presentation.

MethodMedian percentile rank ↓Predicted binders ↑
Random initialization35.43.2%
Drift, ESM2-35M (encoder comparison)15.618%
Genetic algorithm (GA)8.133%
Drift, ESM C 600M8.031%
Fine-tuned ESM2-MLM-300.9773%
Fine-tuned PepDFM-4000.5478%
Real held-out ligands0.3379%

These results provide a direct comparison with evolutionary search and generative fine-tuning. ESM C 600M drifting achieves a median percentile rank of 8.0 and 31% predicted binders, compared with 8.1 and 33% for the genetic algorithm. Fine-tuned PepDFM reaches 78%, close to the 79% observed for held-out ligands. This expanded comparison quantifies the improvement from direct population updates and motivates the additional experiments on encoder geometry.

For molecular generation, we also fine-tuned ChemGPT, REINVENT4, and GenMol on the same reference datasets used for drifting. NimblDrift achieves a mean Boltz-2 score of 0.816 on 6CM4, compared with 0.532, 0.463, and 0.543, respectively. These new baselines substantially broaden the evidence supporting the approach.

Q2. Contribution relative to continuous drifting and greedy search

“Could the authors further clarify the methodological novelty of NimblDrift relative to the original continuous Drifting framework (Deng et al., 2026) and existing greedy search methods?”

We appreciate the opportunity to clarify this relationship. Deng et al. [1] use a drifting field to train a neural generator with one-step inference. NimblDrift uses frozen biological encoders to directly update populations of discrete molecules or sequences during generation. Each update selects a feasible edit by evaluating its effect on the population discrepancy. This enables immediate use of a new reference population and makes the permitted molecular or sequence edits part of the generation procedure.

1. Direct discrete updates and projected continuous drift. We explicitly compare population-discrepancy minimization with projecting a continuous drift displacement onto feasible edits. Projection selects edits close to the displaced embedding. Direct discrepancy minimization can select a long-range edit when its endpoint improves the population objective. In the synthetic small-world experiment, the mean convergence times are 10.67 steps for NimblDrift and 14.17 for projected drift. The corresponding peptide comparison appears in Appendix A2.2.

2. Molecular connectivity and crossover proposals. We adapt Graph GA crossover operations [2] to propose candidates for population-discrepancy minimization. The analysis using Kleinberg networks [3] motivates the role of long-range connections, and the molecular experiment directly measures the effect of adding those proposals.

Table R18. Molecular reference matching under three neighborhoods, from submission Table 1. MMD² is squared maximum mean discrepancy. The reference self-MMD² level is 0.0065, and each crossover configuration uses 10 partners.

NeighborhoodMinimum MMD² ↓Steps to self-MMD² level ↓
Graph edits0.0128>100
Graph edits + BindingDB crossovers0.006292
Graph edits + within-population crossovers0.004945

Within-population crossovers reach the reference self-MMD² level in 45 steps, compared with 92 for BindingDB crossovers. Graph edits reach a minimum MMD² of 0.0128 after the reported search. This establishes a concrete connection between feasible edit connectivity and the effectiveness of population matching.

3. Reference-based regularization for biological objectives. GreedyDrift combines predicted affinity with a penalty for departure from the reference population. At regularization weight 0.2, mean predicted affinity increases from 6.281 to 7.969 while solubility EMD is 0.280, compared with 0.670 for greedy search. The new attraction-only and nearest-reference controls further isolate the contribution of the population objective.

Together, these contributions connect frozen biological representations, discrete feasible edits, population coupling, and property regularization in a single generation procedure. The new experiments strengthen this account by identifying how simpler search, fine-tuning, and encoder geometry affect the resulting designs.

Improving figure clarity and exposition

“The figures are difficult to see and interpret, quite blurry and very dense.”

We appreciate this feedback. We will enlarge labels, separate the dense panels in Figures 2 and 3, and use high-resolution exports with consistent legends and readable axes. We will also streamline the mathematical exposition around the reference distribution, population discrepancy, and candidate-update rule, and place the new quantitative comparisons alongside the corresponding figures.

References

[1] Deng et al. (2026). Generative Modeling via Drifting. https://arxiv.org/abs/2602.04770

[2] Jensen (2019). A graph-based genetic algorithm and generative model/Monte Carlo tree search for the exploration of chemical space. https://doi.org/10.1039/C8SC05372C

[3] Kleinberg (2000). The Small-World Phenomenon: An Algorithmic Perspective. https://www.cs.cornell.edu/home/kleinber/swn.pdf

We thank the reviewer again for the helpful review! We hope the additional experiments and clearer methodological explanation support a more favorable assessment, and we would be grateful if the reviewer would consider increasing the score and reassessing confidence in the submission to support acceptance of the work.

I thank the authors for their additional experiments, which address my concerns regarding missing baselines. However, my main concern regarding limited methodological novelty relative to prior work remains unresolved. I appreciate the clarifications and additional results, I will maintain my original score.

Reviewer_bw7BOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes NimblDrift as a method for editing populations of discrete biological candidates using frozen representations and reference data. NimblDrift iteratively selects local edits which reduce a feature-space discrepancy between candidate and reference populations. This avoids fitting a new generative model when incorporating new reference data. The paper also extends this to GreedyDrift, which addresses reward optimization via reference-based regularization. The paper examines neighborhood connectivity through synthetic graph experiments and evaluates predicted molecular binding, peptide property retention, and allele-specific MHC presentation.

Critical strengths
  1. NimblDrift enables incorporating new reference data without retraining a generator. This can be useful under limited resources.
  2. The experiments support the claim that representation-based population matching can preserve useful biological information.
Critical weaknesses
  1. My main concern is positioning and novelty. The methodological contribution is not clearly distinguished from prior distribution-matching and discrete-search methods. Although the updates are defined in Section 3, much of the construction consists of established discrepancy objectives, greedy coordinate updates, and inherited molecular crossover operations. The related-work discussion does not adequately explain the additional contribution relative to kernel herding or greedy MMD minimization. This makes it difficult to assess the significance of the proposed algorithmic contribution.
  2. The method’s practical advantage remains unclear because the baseline comparisons are limited. Molecular binding is evaluated on only three selected targets using predictor scores, with limited search baselines. The peptide experiment includes a generative comparator, but the comparison is restricted to one target and a selected affinity range. Also, the MHC experiments include useful controls but no competing generation method. Therefore, these results demonstrate feasibility but not a clear practical advantage over competing methods.
Questions for authors
  1. Could the authors clarify the precise methodological distinction from greedy discrepancy minimization and kernel-herding approaches? Also, which component of the proposed construction is essential to the observed biological gains?
  2. Do the molecular advantages persist when the proposed method and search baselines use matched information and computational budget?
  3. Could the authors provide complete quantitative comparisons for ESM-2, PeptideCLM-2, greedy search, and MOG-DFM, including affinity and secondary-property statistics? This would help assess how consistently the proposed regularization improves the quality-retention trade-off.
6 comments

We thank the reviewer for the careful assessment and for recognizing the value of incorporating reference data through frozen representations! To address the questions about novelty and practical benefit, we undertook new comparisons with fine-tuned generators, simpler embedding-guided search, and additional encoders. We also clarify the precise objective and the evidence for each component below.

Q1. Relationship to discrepancy minimization and kernel herding

“Could the authors clarify the precise methodological distinction from greedy discrepancy minimization and kernel-herding approaches?”

NimblDrift uses greedy discrepancy minimization to update a fixed-size population of discrete candidates. For neural embeddings, we evaluate an empirical reverse-KL surrogate using Gaussian-smoothed candidate and reference densities. For Morgan fingerprints, we use MMD with a Tanimoto kernel. Each update evaluates a candidate replacement against the current population, so the selected edit depends on both the reference distribution and the other generated candidates.

Classical kernel herding [1] constructs a sample sequence by adding points to approximate a target kernel mean. NimblDrift repeatedly edits existing population members through feasible sequence substitutions or molecular graph operations. Including the unchanged candidate guarantees nonincreasing discrepancy within a sweep at fixed bandwidth. This makes the population size, feasible edits, and neighborhood connectivity explicit parts of generation.

Our contribution combines this discrete population-update procedure with frozen biological representations, reference-based reward regularization, and an analysis of how neighborhood connectivity affects generation. The new experiments isolate the contributions of population matching and representation choice.

New ablations of population matching

“Also, which component of the proposed construction is essential to the observed biological gains?”

To address this directly, we added two simpler baselines. Fixed nearest-reference search moves each candidate toward the reference closest to its initial embedding. Mean-shift attraction moves each candidate toward a kernel-weighted average of reference embeddings. Both update candidates individually. NimblDrift evaluates each replacement through the full population discrepancy.

Table R8. Predicted MHC class I binder percentages.

AlleleRandomFixed nearest referenceMean-shift attractionNimblDrift
A*01:014.3%13.2%14.7%21.3%
A*02:012.7%6.0%12.3%11.2%
A*03:012.7%10.0%10.0%19.2%

NimblDrift exceeds fixed nearest-reference search on all three alleles. It also exceeds mean-shift attraction on A*01:01 and A*03:01, while mean shift achieves 12.3% on A*02:01 compared with 11.2% for NimblDrift. These results identify where population coupling provides the clearest gains.

Table R9. Mean Boltz-2 binding likelihood in the 2VT4 search ablation.

TargetRandomFixed nearest referenceMean-shift attractionNimblDrift
2VT40.1990.2160.2590.466

For 2VT4, NimblDrift achieves 0.466, compared with 0.216 for fixed nearest-reference search and 0.259 for mean-shift attraction. The 79.9% increase over mean-shift attraction directly supports the benefit of population-level matching in this molecular experiment.

Reference

[1] Chen, Welling, and Smola (2010). Super-Samples from Kernel Herding. https://proceedings.mlr.press/r8/chen10a.html

Q2. Reference information and computational budget

“Do the molecular advantages persist when the proposed method and search baselines use matched information and computational budget?”

To strengthen this comparison, we fine-tuned ChemGPT, REINVENT4, and GenMol using the same task-specific reference datasets supplied to NimblDrift. This directly compares how the methods incorporate the available target-specific examples.

Table R10. Best reported mean Boltz-2 binding likelihood in the generator comparisons.

Method2VT45SDV6CM4
ChemGPT, fine-tuned0.5930.3050.532
REINVENT4, fine-tuned0.450.3310.463
GenMol, fine-tuned0.3850.4100.543
NimblDrift, SMI-TED289M0.4230.3180.816

NimblDrift gives the highest score on 6CM4, exceeding GenMol by 50.3%. On 2VT4, it exceeds GenMol, while ChemGPT gives the highest score, followed by REINVENT4. GenMol leads on 5SDV. The initial ChemGPT checkpoint also reproduced reference molecules, so we considered nearest-reference Morgan-Tanimoto distance alongside binding. The resulting Pareto comparisons support competitive novelty for NimblDrift across the evaluated molecular tasks.

We further separated reference information from predictor feedback by running uniform-weight ablations for both molecular encoders.

Table R11. Uniform and Boltz-2-weighted reference comparisons.

Method2VT45SDV6CM4
SMI-TED289M, uniform weights0.4170.3180.816
SMI-TED289M, Boltz-2 weights0.5440.4190.812
Molformer, uniform weights0.3100.4020.495
Molformer, Boltz-2 weights0.3360.4230.721
Random walk (submission Table 4)0.1970.1910.156

Both encoders improve over random walk on all three targets with uniform reference weights. In these runs, Boltz-2 evaluates the generated molecules after the encoder-guided search. This directly establishes gains from the reference molecules and encoder representations, with predictor-weighted reference selection reported as a separate configuration.

Table R12 gives the complete measured runtime at each experiment's stated population size and step count.

Table R12. NimblDrift runtime on one H100 GPU, from Appendix Table A1.

Task / encoderPopulationSeconds/stepStepsTotal seconds
Molecules / SMI-TED289M1009.840392
Molecules / Molformer1007.940316
Molecules / Morgan fingerprints1006.240248
Peptides / ESM-2 35M4076.41596
Peptides / PeptideCLM-240772.0151080
MHC class I / ESM-2 35M2002.12041
MHC class II / ESM-2 35M2003.92078

These experiments match the task-specific reference datasets and document the computational cost of the NimblDrift configurations. For the revised comparison, we will additionally specify pretraining sources, proposal counts, evaluation calls, and fine-tuning and generation times for each baseline so that reference information and computational budgets are explicit.

Together, the new generator comparisons and uniform-weight controls show that direct population updates can produce competitive molecular candidates from the same reference examples used to adapt generative models. The simpler search ablations in our preceding comment further isolate the contribution of population matching.

Q3. Affinity and secondary-property comparisons

“Could the authors provide complete quantitative comparisons for ESM-2, PeptideCLM-2, greedy search, and MOG-DFM, including affinity and secondary-property statistics?”

We appreciate the request to make the property-retention comparison easier to assess. For target 7JVS, we compared ESM-2 and PeptideCLM-2 regularization, unregularized greedy search, and MOG-DFM [1]. We optimized the PeptiVerse affinity score and then evaluated solubility, non-fouling, and hemolysis with the corresponding secondary-property predictors. This tests whether reference-based regularization preserves useful properties during affinity optimization.

The following table reproduces all numerical results from submission Table 6 for the PeptideCLM-2 regularization sweep, including the greedy baseline. Earth-mover distance (EMD) measures the difference between each generated property distribution and the natural-peptide reference distribution. Lower EMD indicates closer agreement with that reference.

Table R13. Full-population property EMD and mean predicted affinity.

Population / regularization weightNon-fouling EMD ↓Solubility EMD ↓Non-hemolytic EMD ↓Mean affinity ↑
Initial random 12-mers0.1010.0100.0086.281
0.0 (greedy search)0.6020.6700.1228.651
0.10.4420.4110.1028.294
0.20.3870.2800.0897.969
0.50.3270.1390.0337.431
1.00.2600.0220.0197.052

At a regularization weight of 0.2, mean predicted affinity increases from 6.281 to 7.969. Relative to greedy search, this setting reduces solubility EMD by 58.2%, non-fouling EMD by 35.7%, and non-hemolytic EMD by 27.0%. The complete sweep shows how the regularization strength controls the balance between predicted affinity and retention of reference-like properties.

To compare secondary properties at similar binding scores, Figure 2 evaluates all four methods within the shared affinity interval of 8.0-8.5. PeptideCLM-2 regularization places candidates in favorable regions of all three property distributions. ESM-2 shows a similar pattern, with PeptideCLM-2 providing stronger solubility and non-fouling profiles. This comparison supports property retention at comparable predicted affinity. We will accompany the distributions with per-method affinity and secondary-property summaries for all four methods in the revision.

Additional generative comparisons for MHC peptides

To address the broader request for peptide-generation baselines, we also fine-tuned the ESM-2 masked-language-model head and PepDFM on the same IEDB reference datasets used for drifting.

Table R14. Predicted binder percentages by allele.

MethodA*01:01A*02:01A*03:01
Random initialization4.32.72.7
NimblDrift, ESM2-35M21.311.219.2
ESM2-35M MLM, fine-tuned81.745.590.5
PepDFM, fine-tuned84.053.894.8
Reference ligands84.057.795.3

These comparisons quantify improvements from population matching and the higher binder fractions achieved through task-specific generative fine-tuning. We then studied larger encoders and contrastive representation learning. In the representation experiment, ESM C 600M drifting reaches 37.5%, 22.5%, and 33.7% predicted binders while maintaining a median nearest-reference distance of five edits for each allele. A separately contrastively trained ESM2-35M encoder reaches 96.3%, 50.0%, and 92.0%, at median distances of three, four, and four edits. Its contrastive training excludes the generation reference set and the three test alleles.

Reference

[1] Chen et al. (2025). Multi-Objective-Guided Discrete Flow Matching for Controllable Biological Sequence Design. https://openreview.net/forum?id=8YIMLoHP9J

The expanded evaluation provides concrete evidence for the roles of population coupling, reference regularization, and encoder geometry in biologic design. We thank the reviewer for prompting these additions and apologize for the length of the response. We respectfully ask the reviewer to consider raising the score and revisiting confidence in the submission in light of the new comparisons and clarifications.

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

This paper introduces NimblDrift, a non-parametric method for discrete biological sequence and molecule design. Instead of training generative models (such as diffusion or flow models), NimblDrift directly guides local search over candidate populations using frozen representation encoders and reference datasets.

The authors also propose GreedyDrift to incorporate scalar reward objectives alongside reference-based regularization. The approach is evaluated across multiple benchmarks.

Critical strengths
  1. The motivation for this work is very practical: training discrete diffusion or flow models from scratch can be computationally expensive and difficult to adapt to small, ad-hoc reference sets. Making direct population updates based on frozen representation is an appealing framework for biologic design.

  2. The methodological comparison with projected continuous drift and the analysis using Kleinberg small-world networks are interesting. They clearly explain why naive single-edit graph neighborhoods suffer from slow mixing and justify the empirical benefit of incorporating crossover proposals to introduce effective shortcuts.

Critical weaknesses
  1. The empirical validation relies exclusively on computational predictors (e.g., Boltz-2 for small molecules, NetMHCpan, and PeptiVerse surrogates). More critically, in Table 4, the highest-performing variant incorporates reference weights derived directly from Boltz-2 scores, which is the same model used as the final evaluation metric. Could this be an issue of oracle overfitting? It's definitely making things obscure to evaluate the method's true generalizability.

  2. The paper positions NimblDrift as a competitive alternative to discrete diffusion and flow models, yet the experimental baselines are mostly limited to random walks, FlashBind greedy search, etc (Table 4, Figure 2). Why? Lacking direct comparisons against state-of-the-art discrete generative models weakens the claim.

Questions for authors
  1. Could the authors provide a clearer ablation separating the contribution of the drift mechanics from the Boltz-2 reference weighting? How well does SMI-TED289M Drift perform when reference weights are uniform, or when evaluated against an orthogonal scoring function (e.g., docking scores via AutoDock Vina / GNINA)?

  2. Direct comparison with generative models is missing. It greatly weakens the claim made in the abstract and introduction that NimblDrift serves as an alternative to fine-tuning diffusion/flow priors? How does NimblDrift compare in performance, compute efficiency, and sample novelty against a standard fine-tuned masked LM or discrete diffusion baseline?

4 comments

We thank the reviewer for the thoughtful assessment and for recognizing the value of adapting frozen representations to small reference sets. The questions motivated substantial new experiments, including uniform-weight controls, five fine-tuned generative baselines across molecules and peptides, simpler search baselines, and an expanded encoder study. These experiments clarify both the contribution of population matching and the design settings in which NimblDrift is most effective.

Q1. Separating reference weighting from generation

“Could the authors provide a clearer ablation separating the contribution of the drift mechanics from the Boltz-2 reference weighting?”

To address this concern directly, we repeated molecular generation with uniform reference weights for both SMI-TED289M and Molformer. In these runs, the encoder and reference molecules guide generation, and Boltz-2 scores the resulting candidates after generation.

Table R1. Mean Boltz-2 binding likelihood, higher is better.

Method2VT45SDV6CM4
SMI-TED289M, uniform weights0.4170.3180.816
SMI-TED289M, Boltz-2 weights0.5440.4190.812
Molformer, uniform weights0.3100.4020.495
Molformer, Boltz-2 weights0.3360.4230.721
Random walk (submission Table 4)0.1970.1910.156

With uniform weights, SMI-TED289M achieves 2.12x, 1.66x, and 5.23x the random-walk scores on 2VT4, 5SDV, and 6CM4, respectively. Molformer also improves over random walk on all three targets. Moreover, uniform SMI-TED289M weighting achieves 0.816 on 6CM4, compared with 0.812 under Boltz-2 weighting. Thus, the reference-weighting ablation demonstrates substantial improvements from encoder-guided population updates with Boltz-2 reserved for evaluation. The weighted runs separately quantify the benefit of incorporating predictor information into the reference distribution.

Q2. Direct comparisons with fine-tuned generative models

“How does NimblDrift compare in performance, compute efficiency, and sample novelty against a standard fine-tuned masked LM or discrete diffusion baseline?”

To broaden the molecular comparison, we fine-tuned ChemGPT, REINVENT4, and GenMol on the same task-specific reference datasets used by NimblDrift.

Table R2. Best reported mean Boltz-2 binding likelihood in the new generator comparisons.

Method2VT45SDV6CM4
ChemGPT, fine-tuned0.5930.3050.532
REINVENT4, fine-tuned0.450.3310.463
GenMol, fine-tuned0.3850.4100.543
NimblDrift, SMI-TED289M0.4230.3180.816

NimblDrift reaches 0.816 on 6CM4, 50.3% above the strongest fine-tuned comparator, GenMol at 0.543. On 2VT4, NimblDrift also exceeds GenMol. ChemGPT gives the highest 2VT4 score, followed by REINVENT4, and GenMol leads on 5SDV. These comparisons establish competitive molecular design through direct reference-guided generation with a frozen encoder.

The additional 2VT4 search ablation further demonstrates the contribution of population matching. NimblDrift reaches 0.466, compared with 0.216 for nearest-reference search and 0.259 for mean-shift attraction, a 79.9% increase over mean-shift attraction.

To assess novelty alongside binding, we additionally compared the Pareto frontiers using median Morgan-Tanimoto distance to the nearest reference molecule. Larger distances indicate greater structural novelty. The comparisons support competitive novelty for NimblDrift, with method rankings varying across targets. We also observed reference-set memorization in the initial ChemGPT checkpoint, which reinforces the importance of reporting binding and novelty together.

Q2 continued. Fine-tuned peptide generators

To extend the generative comparison to allele-specific peptide design, we fine-tuned PepDFM and the ESM-2 masked-language-model head on the same IEDB reference ligands used for drifting. We then evaluated the generated peptides with NetMHCpan.

Table R3. Predicted binder percentages by MHC class I allele.

MethodA*01:01A*02:01A*03:01
Random initialization4.32.72.7
NimblDrift, ESM2-35M21.311.219.2
ESM2-35M MLM, fine-tuned81.745.590.5
PepDFM, fine-tuned84.053.894.8
Reference ligands84.057.795.3

NimblDrift increases the predicted binder fraction from 4.3% to 21.3% for A*01:01, from 2.7% to 11.2% for A*02:01, and from 2.7% to 19.2% for A*03:01. Fine-tuned peptide generators achieve higher fractions, providing a useful benchmark for investigating how the representation affects direct population matching.

New experiments on encoder size, pooling, and contrastive training

To investigate this result, we conducted further experiments with larger encoders, concatenation pooling, and a contrastively trained ESM2-35M encoder. This keeps the population-update mechanism fixed while examining different representation choices.

Table R4. Predicted binder percentage / median edit distance to the nearest reference in the representation study. Each cell reports both quantities. Larger edit distances indicate greater sequence novelty.

MethodA*01:01 % / distanceA*02:01 % / distanceA*03:01 % / distance
Random initialization4.3 / 62.7 / 62.7 / 6
Drift, ESM2-35M23.0 / 511.7 / 519.2 / 5
Drift, ESM2-650M35.8 / 519.2 / 530.0 / 5
Drift, ESM C 600M37.5 / 522.5 / 533.7 / 5
Fine-tuned ESM2-MLM-3081.7 / 445.5 / 590.5 / 4
Fine-tuned PepDFM-40085.2 / 453.7 / 594.7 / 4
Drift, contrastively trained ESM2-35M96.3 / 350.0 / 492.0 / 4
Drift, ESM2-35M with concatenation pooling96.8 / 071.2 / 090.3 / 0
Reference ligands82.0 / 060.8 / 093.7 / 0

Changing the encoder from ESM2-35M to ESM C 600M raises predicted binder fractions from 23.0% to 37.5%, 11.7% to 22.5%, and 19.2% to 33.7%, while preserving a median distance of five edits for all three alleles. This demonstrates that stronger representations can improve the same generation algorithm while retaining sequence novelty.

We then examined contrastive representation learning [1]. The contrastive training excluded the generation reference set and all three test alleles. With this separately trained encoder frozen during generation, NimblDrift reaches 96.3%, 50.0%, and 92.0% predicted binders, with median distances of three, four, and four edits. These results support representation geometry as an important determinant of generation quality and show substantial improvements through encoder selection and training.

Concatenation pooling reaches 96.8%, 71.2%, and 90.3%, with median nearest-reference distances of zero. This setting emphasizes recovery of reference sequences. Reporting binding and edit distance together makes the distinction between reference recovery and novel sequence generation explicit.

Taken together, these new experiments substantially expand the original evaluation. They establish competitive molecular generation, quantify the gains achievable through direct peptide population updates, and identify representation choices that improve those gains.

Reference

[1] Radford et al. (2021). Learning Transferable Visual Models From Natural Language Supervision. https://arxiv.org/abs/2103.00020

Q2 continued. Aggregate comparisons and runtime

For completeness, the following tables report the aggregate MHC class I results across nine allele-seed pairs, including the evolutionary baseline and every encoder in the expanded comparison.

Table R5. Aggregate results from the generation and encoder comparisons. Lower NetMHCpan percentile ranks indicate stronger predicted presentation.

MethodMedian percentile rank ↓Predicted binders ↑
Random initialization35.43.2%
Drift, ESM2-35M (encoder comparison)15.618%
Genetic algorithm (GA)8.133%
Drift, ESM C 600M8.031%
Fine-tuned ESM2-MLM-300.9773%
Fine-tuned PepDFM-4000.5478%
Real held-out ligands0.3379%

This comparison quantifies the benefit of each approach. ESM C 600M drifting improves the median percentile rank from 35.4 at initialization to 8.0 and increases the predicted binder fraction to 31%. The genetic algorithm reaches 33%, and fine-tuned PepDFM reaches 78%, close to the 79% observed for held-out ligands.

Table R6. Aggregate encoder comparison, with each encoder frozen during drifting.

EncoderParametersMedian percentile rank ↓Predicted binders ↑
Random initializationN/A35.43.2%
PeptideCLM23M19.514%
ESM2-35M35M15.618%
PeptideCLM-286M13.817%
ESM C 300M332M9.030%
ESM C 600M574M8.031%
ESM2-650M650M9.728%
ESM C 6B6.3B18.317%

ESM C 600M gives the strongest results among the encoders in this comparison. The variation across model families and sizes further motivates the representation-geometry experiments in our preceding comment and provides practical guidance for selecting an encoder.

Measured computational cost

NimblDrift incorporates new reference data through population updates using a frozen encoder. To make the cost concrete, we reproduce the full runtime measurements from Appendix Table A1, collected on a single H100 GPU at each experiment's stated population size and number of steps.

Table R7. Runtime of the reported NimblDrift configurations.

Task / encoderPopulationSeconds/stepStepsTotal seconds
Molecules / SMI-TED289M1009.840392
Molecules / Molformer1007.940316
Molecules / Morgan fingerprints1006.240248
Peptides / ESM-2 35M4076.41596
Peptides / PeptideCLM-240772.0151080
MHC class I / ESM-2 35M2002.12041
MHC class II / ESM-2 35M2003.92078

The molecular runs take 4.1–6.5 minutes per population. Peptide regularization takes 1.6 minutes with ESM-2 and 18.0 minutes with PeptideCLM-2, and MHC generation takes 0.7–1.3 minutes. These measurements describe population-update runtime for the listed configurations. The practical advantage is immediate use of a new reference set through the encoder and search procedure.

We will bring these comparisons into the revised manuscript and clarify the contribution around direct biologic design from representations and reference data, supported by explicit comparisons with generative training, evolutionary search, and simpler embedding-guided updates. The additional experiments give a much clearer account of molecular performance, peptide novelty, and the role of representation geometry.

We thank the reviewer again for the constructive suggestions! Given these new experimental results and clarifications, we would be very appreciative for a increased score and reassessed confidence in the submission, if possible.

We would like to thank reviewer neGa again for the important feedback that encouraged the authors to thoroughly baseline NimblDrift. Seeing that the reviewer has not yet left updated feedback or score and that the discussion period is ending, we encourage the reviewer to look through the wholistic rebuttal comment "Rebuttal Follow-up, Paper Revision and Discussion Takeaway" which summarizes the positioning of the paper and revisions to the manuscript.

Importantly, we show that NimblDrift is competitive in all tasks (though pMHC requires a contrastive fine-tune to improve geometry of the representation space) and thoroughly baseline with simplified versions of the method.

We hope that the reviewer neGa has some time before the discussion period closes to look through the new manuscript and provide feedback to the authors.

Thanks, The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Follow-up, Paper Revision and Discussion Takeaway

We thank the reviewers for the time to look through our responses. The authors have spent time reviewing experiments conducted during the discussion period, and want to make a final argument for the value of our paper before the discussion period closes.

Closing Argument for Publication of NimblDrift

NimblDrift is a simple method for sequence design given oracle access to foundation encoders. We thoroughly study NimblDrift in a variety of molecular and peptide design tasks to show that it is competitive with fine-tuned generative models. Importantly, the NimblDrift update can push arbitrary candidate populations toward a target distribution, an algorithmic primitive that only diffusion / flow-models allow.

In one regime (pMHC ligand generation), we find that NimblDrift is quite suboptimal relative to generative baselines; in this setting, we attribute sub-optimality to poor representations rising from the mean-pooling in the embedding, and we close the gap with a short and simple contrastive fine-tuning objective for the ESM encoder. This setting informs us of the importance of representation quality for NimblDrift's performance. In particular, we show how encoding explicit geometry through contrastive objectives may benefit generation.

We believe our study of NimblDrift is comprehensive, spanning many biologic tasks, both simple and state of the art baselines, as well as analysis tools, and encoder choices. Because of how simple the algorithm is, we believe the others in the generative biology research community will be curious about our work and benefit from the takeaways of our experiments. For these reasons, we graciously request the reviewers to take a final look through Nonparametric Biologic Design via Discrete Drifting in the updated manuscript and reflect upon their recommendation for this paper

Manuscript Change Log (ordered by importance):

We use [reviewer code] to prefix changes that correspond to concerns of a specific reviewer.

  1. [Zdte] We drop analysis of "Projected Continuous Drift" in favor of comparing NimblDrift to nearest-reference similarity maximization between pairs of candidates and references ("SingletonDrift"). We believe this helps the reader understand what sets NimblDrift apart from the simpler approach which was in part the concern of reviewer Zdte.
  2. [neGa] Bootstrapped confidence bounds show that NimblDrift with the SMI-TED_289M is the only method competitive with all methods and all targets for molecular generation. This clarifies the result from Table R2 which did not incorporate confidence bounds over re-seeds of the initial candidate population.
  3. [Zdte] SingletonDrift (nearest reference similarity maximization) and Density ascent baselines added for molecular generation
  4. [bw7B] GreedyDrift section shows violin plots of exact secondary-property distributions at various choices of λ\lambda. New metric, "# of therapeutically viable & stronger binder peptides" show that NimblDrift strongly outperforms baselines.
  5. [All] pMHC ligand generation section shows fine-tune baselines outperform NimblDrift and introduces contrastive NEW fine-tuned encoder which can achieve competitive binder generation results when applied to NimblDrift.
  6. [All] Simplified the main paper and appendix to focus on core results (removed SentenceDrift & some appendix experiments)
  7. [All] General revision made through out to increase clarity of exposition

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.