acceptodds
Under review as a conference paper at ICLR 2027

Designing Protein Structures from Natural Language

Abstract

A central challenge in de novo protein design is generating proteins with desired structural and functional properties. Existing design models typically require a dedicated input representation for each property, limiting the design objectives they support. We develop Prompteina, the first all-atom structure generation model that supports new and broader design objectives directly specified in natural language, analogous to text-to-image generation. To train Prompteina, we construct Prompteina-Data, a paired text–protein structure dataset of 1.8M structures and 50M+ captions describing global and residue-specific structural and functional properties. We then adapt Proteina-Complexa, a pretrained structure generation model, through text-conditioned pretraining and finetuning that aligns residues with their corresponding text. Prompteina outperforms evaluated baselines on 15 of 18 structural properties and generates proteins that more closely match descriptions of their intended functions. Prompteina outperforms state-of-the-art binder-design models on at least 13 of 20 targets by up to 14%, while supporting specifications beyond existing conditioning interfaces. With iterative in silico feedback, Prompteina’s gain over non-iterative generation is 57% larger than the next-best baseline on the hardest targets. Finally, we experimentally test Prompteina designs against GM2A, which has no reported de novo binders, and identify one binder whose binding signal increases with concentration and shows both association and dissociation. These results establish natural language as an effective and powerful conditioning interface for protein structure generation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
de novo protein design, protein structure generation, text-conditioned generation, flow matching

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_Xha9OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors propose Prompteina, an all-atom protein design model conditioned on natural-language descriptions. The model incorporates global and token-level textual conditioning, together with residue-level text-alignment fine-tuning, to enable more accessible and controllable generation of all-atom protein structures from natural-language specifications. The authors also construct Prompteina-Data, a large-scale dataset containing approximately 1.8 million protein structure entries paired with 52.4 million textual descriptions. The experimental evaluation covers structural attribute control, computational binder design, an LLM-driven iterative design pipeline, and a preliminary experimental study of GM2A-targeting candidates. The results suggest that the proposed approach is promising for natural-language-guided protein design.

Critical strengths
  1. The paper tackles an interesting and challenging task in text-conditioned protein design. By using natural language as a unified interface, Prompteina aims to support the specification of both global structural properties and local structural constraints.
  2. The proposed combination of global and token-level textual conditioning with residue-level text-alignment fine-tuning is well motivated and provides a promising mechanism for enabling fine-grained structural control.
  3. The authors construct Prompteina-Data, pairing approximately 1.8 million protein structures with 52.4 million textual descriptions. In particular, the residue-level alignment annotations provide a valuable foundation for enabling localized control over generated protein structures.
  4. The paper presents a broad range of experiments, including text-guided protein structure generation, text-guided binder design, agent-driven iterative optimization, and a proof-of-concept experimental study targeting GM2A. This extensive evaluation provides useful insights into the capabilities and potential practical applications of the proposed approach.
Critical weaknesses
  1. How much does the natural-language interface contribute to the model's performance beyond the underlying structural prior? In Table 8, pooled natural-language conditioning performs similarly to conditioning on dedicated CATH labels. It would be helpful to investigate whether natural-language conditioning outperforms structured labels or explicit numerical constraints when the training data and computational budgets are matched. Furthermore, the authors should provide a more systematic analysis of semantic sensitivity (whether changing a meaningful phrase produces the intended structural change) and semantic invariance (whether paraphrasing or stylistic variations preserve the intended outcome). The model's ability to generalize to unseen descriptions also warrants further investigation. These analyses would help establish the advantages of natural language as an expressive and effective interface for protein design in realistic settings. 
  2. The claim that Prompteina is the first all-atom protein design model conditioned on natural-language descriptions appears overly broad. For example, Proteo-R1: Reasoning Foundation Models for De Novo Protein Design (ICML 2026) and Co-Generative De Novo Functional Protein Design (ICML 2026) appear relevant to the paper's claims concerning language-conditioned protein generation and joint sequence–structure design. The authors should discuss more on text-conditioned protein co-design models.
    • Proteo-R1: Reasoning Foundation Models for De Novo Protein Design, ICML 2026
    • Co-Generative De Novo Functional Protein Design, ICML 2026
  3. The evaluation of iterative binder design in an agentic setting is valuable for assessing the practical utility of the proposed approach. The authors use a shared agent and matched generation budgets, which helps ensure a fair comparison. However, the conditioning interfaces of the compared methods differ in expressiveness, potentially confounding the effects of the underlying generation models with those of their input interfaces. It would therefore be valuable to include comparisons with methods that support natural-language-based protein design, like Proteo-R1. Such comparisons would help clarify whether the observed improvements arise primarily from Prompteina's generation capabilities, its more expressive conditioning interface, or their combination.
Questions for authors

Please refer to the weaknesses.

4 comments

We thank the reviewer for their review, and for pointing us to relevant related work which we now cite. We address each weakness below.

[W1] Does natural language beat label conditioning at a matched budget? Yes. We now include a new experiment suggested by the reviewer: a dedicated CATH-label conditioner with the same pooled and cross-attention pathways as text. We fix the architecture, data and compute budget at 47,275 steps, so the runs differ only in the conditioning signal. Under this matched setting, natural language outperforms label conditioning (92.0% vs. 66.7% on CATH class, 27.0% vs. 25.1% on architecture; Table Xha9.1). We report this new ablation in App. D, Table 9.

We only evaluate CATH for these ablations because the CATH-label conditioning runs are limited to CATH class and architecture, whereas the natural-language runs support many properties beyond CATH (last column of Table Xha9.1; Table 1, App. Table 11). Specifying constraints in natural language lets us scale the number of properties we condition on, while a label interface requires a predefined embedding table for each new property.

Table Xha9.1. CATH top-1 accuracy (%) by conditioning interface and pathway. The last column lists the properties each interface controls; the 18 text tasks are measured on the released model, not on these matched runs.

PathwaysConditioningCATH classCATH architectureProperties shown to be controllable
pooled + cross-attentionCATH label66.725.1CATH class, architecture (2 tasks)
pooled + cross-attentionNatural language92.027.0fold, shape, secondary, tertiary, composite (18 tasks)

[W1] Is Prompteina sensitive to semantic changes? Yes. Our evaluations in Section 3.1 test semantic sensitivity: whether changing the property specified in the prompt changes the generated structure. We measure 18 structural properties, and Prompteina leads on 15 (Table 1; App. Table 11). For helix and strand count we prompt This protein contains {X} helices. with lengths fixed across requested counts, against empty-prompt controls at the same 50 lengths. Changing only X controls the realized count, while the empty-prompt controls show no count control (App. Table 12). We also show a more qualitative example: on CATH architecture, Prompteina is sensitive to a single-word change between two architectures (100% on "Horseshoe", 96% on "Barrel"), while Proteina follows the Horseshoe request and not the Barrel one (84% and 58%, Table Xha9.2). Together, this suggests Prompteina is sensitive to semantic changes in the prompt.

Table Xha9.2. GearNet top-1 architecture accuracy (%). Prompteina's prompts differ by one word in the CATH architecture name. The panel means over all 35 architectures are 51.1% for Prompteina and 27.5% for Proteina.

Modelan Alpha-Beta Horseshoe foldan Alpha-Beta Barrel fold
Prompteina100.096.0
Proteina84.058.0

[W1] Is Prompteina robust to non-semantic changes? Yes. We currently demonstrate semantic invariance by testing Prompteina's robustness to paraphrasing (App. Table 13). To establish comparative robustness, we extend these CATH class and helix count evaluations to Chroma (text), BioMatrix, and Pinal. For each property, we specify the prompts in four ways, with lengths matched across the generations. Prompteina has the highest mean and the lowest standard deviation on both properties (Table Xha9.3). We report this new comparison in App. E.1, Table 14. We also show that Prompteina remains robust on a finer-grained steering task, binder interface residue count, which none of the other models support, at ρ 0.976 to 0.997 (App. Table 19).

Table Xha9.3. Mean and S.d. across four wordings of the same specification. CATH class accuracy is top-1 over three classes (33.3% chance). Helix count is the Spearman correlation between requested and realized counts.

ModelCATH class Mean (%)CATH class S.d.Helix count ρ MeanHelix count ρ S.d.
Prompteina97.92.50.9350.017
Chroma (text)34.29.10.7330.038
BioMatrix35.87.40.5940.107
Pinal35.414.00.7040.038

[W1] Generalization to unseen descriptions. We interpret "unseen descriptions" in two ways: 1) paraphrases of known properties, and 2) new phrases or terms unseen in training.

  1. Paraphrases. As shown above (Table Xha9.3), Prompteina generalizes to paraphrases of known properties in different formats, and when the prompt is written by a different language model than the one used for generating its training data.

  2. Terms absent from training captions. We adopt prompt expansion to support generation from unseen descriptions, as in text-to-image settings [1]: a rewriting step converts the unseen term into an in-distribution description, and Prompteina generates from that. We demonstrate this with one example, the design goal "single-chain transmembrane nanopore", which appears in 0 of the 23.4M captions in our monomer training split. We rewrite the design goal into in-distribution properties represented in our training data, and generate 960 designs per prompt, on the original prompt with the unseen term and on the expanded prompt, matched on length. Designs are scored with three measurements: a constriction diameter of at least 1.0 nm, a lumen that stays open for at least 25 Å across the membrane slab [2], and an inside-out index of at least +0.15 [3]. The unseen term generated no designs that satisfied all three criteria, while the prompt-expanded caption produced 41 (p = 2.9 × 10^-13).

Together, these suggest Prompteina supports both paraphrased and previously unseen descriptions.

[W2] Novelty and comparison to Proteo-R1 and CodeFP. The core difference between Prompteina and these two methods is whether the generative process is conditioned directly on text. In Prompteina, the text embeddings of a free-text caption condition the denoising process itself. In Proteo-R1, a reasoning model reads the text and selects key residues in a given antibody-antigen complex, and the generator conditions on those residues. Similarly, CodeFP maps the text to a term from a closed 375-term Gene Ontology vocabulary, retrieves a structural motif for that term, and conditions on the motif (Table Xha9.4). In our paper, we have now made the wording of the claim precise: "the first all-atom protein structure generation model conditioned directly on natural language, supporting diverse design objectives through a single interface". We now cite both works in Related Work and in the conditioning survey in App. A, and we describe what we could and could not run of each in W3.

Table Xha9.4. Comparison of Proteo-R1, CodeFP, and Prompteina.

MethodInputOutputGenerator conditioned onConditioning in natural language
Proteo-R1antigen structure, antibody framework with CDR positions maskedCDR sequences, complex coordinateshidden states at key residue anchorsNo
CodeFPGO terms, from a closed 375-term vocabularysequence, with structure tokens decoded jointly; 3D structure from ESMFoldstructural motifs retrieved per termNo
Prompteinafree-text captionall-atom structure, sequencecaption token embeddingsYes

[W3] Iterative binder design setup. We report the gain from iterative optimization over i.i.d. sampling for the same model, and compare this gain across models. The within-model gain removes each model's base performance from the comparison. The evaluation tests whether a model with a more expressive conditioning interface can gain more from iteration relative to i.i.d. sampling.

[W3] Comparison with Proteo-R1. As released, Proteo-R1 and Prompteina support different design tasks. Proteo-R1 redesigns masked CDR loops inside a supplied antibody-antigen complex, while Prompteina designs a binder from scratch against a target. Nevertheless, Proteo-R1's approach of reasoning over a design request and translating it natively into residue-level constraints is a promising avenue for language-based protein design. Extending Proteo-R1 to de novo binder design would require retraining it, which we leave to future work. We thank the reviewer for the pointer, and now cite and discuss Proteo-R1 in Related Work and in the conditioning survey in App. A.

We thank the reviewer again for their time and feedback, and welcome any further clarifications!

  1. Betker et al. Improving image generation with better captions. Technical report, OpenAI, 2023.
  2. Smart et al. HOLE: A program for the analysis of the pore dimensions of ion channel structural models. Journal of Molecular Graphics, 1996
  3. Dhar et al. Membrane Barrels Are Taller, Fatter, Inside-Out Soluble Barrels. J. Phys. Chem. B, 2021.

The reviewer appreciates the authors' detailed responses and the additional experimental results. The rebuttal has addressed most of the concerns. In light of these clarifications and improvements, the reviewer has raised the recommendation to Weak Accept.

Thank you once again for the constructive feedback and for raising your score!

Reviewer_i2xLOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Prompteina, a general methodology for steering protein generative models with natural language. Prompteina is built upon the publicly available Proteina-Complexa. The paper first curates a large-scale prompt–structure paired dataset, which is used to fine-tune the model. It then experiments with multiple methods to find the empirically optimal strategy for injecting language conditioning. The paper includes extensive in-silico and wet-lab experiments to demonstrate the effectiveness of natural-language steering and of the interface between Prompteina and an LLM agent in protein design. Overall, this paper is creative in its idea and solid in execution.

Critical strengths
  1. Comprehensive experiments and ablation studies.
  2. Plans to open-source the data and code.
  3. Novel idea and careful execution that yield strong results.
Critical weaknesses
  1. It is unclear to me whether the prompt-following capability is preserved under refolding evaluation.
Questions for authors
  1. Prompteina is successful in steering the denoiser-generated protein structures. However, for natural-language-guided binder design, how much of the prompt is followed after refolding the generated binder with a folding model?

  2. During the agentic iterative loop of binder optimization, it seems like the ESMFold2-Fast-computed ipSAE$_{min}$ is the optimization goal. Is there another folding model used for evaluation, so that we know whether the iterative optimization generally improves designs?

  3. For simple conditioning such as CATH class, how is the prompt formatted? Can you show a few prompt–structure examples of accurate generations for the tasks in Table 1b?

  4. The pooled text embedding is injected through AdaLN, but Complexa applies AdaLN to both the sequence and the pair representation. Is the text injected into the pair conditioning too, or does the pair representation still see only the timestep embeddings?

1 comment

We thank the reviewer for their positive assessment of our work, and address each of the raised weaknesses and questions below.

[W1, Q1] Are prompts followed under refolding? Yes. We answer this in two parts. First, we refold each binder from its own co-generated sequence with ESMFold2-Fast, and we compare each property before and after, for all methods (Table i2xL.1). Prompteina preserves every property better than any baseline, at a mean correlation of 0.742 across the five properties against 0.671 for the next best method. Interface geometry (interface residue count, buried surface area) is the hardest to preserve for all four methods. We now include this analysis in the revised paper (App. H.7, Table 27).

Table i2xL.1. Correlation between a property on the generated structure and the same property after refolding, within target and averaged over 20 targets, 4,000 designs per method.

MethodHelix fractionSheet fractionCoil fractionInterface residue countBuried surface area
Prompteina0.9700.9730.9420.4260.397
Complexa0.9630.9510.9220.2900.228
BoltzGen0.8740.8770.8500.3390.318
RFdiffusion30.9250.9500.8920.2600.240

Second, we compute the correlation between the requested value and Prompteina's generated structure: 0.748 for helix fraction, 0.789 for sheet fraction, 0.541 for coil fraction, 0.460 for interface residue count and 0.374 for buried surface area. None of the other baselines support interface size or buried surface area conditioning. Together with Table i2xL.1, these correlations suggest that a requested property is realized in the generated structure and then preserved after refolding.

[Q2] Is iterative optimization robust under a held-out folding model? Yes. During iterative optimization, we use ESMFold2-Fast because it provides similar performance and a 5.6x speedup relative to AlphaFold 3 based on our calibration panel (App. Table 22). Here, we post-hoc score designs from iterative optimization under ESMFold2-Fast with AlphaFold 3 across all methods (Table i2xL.2). All three methods improve under AlphaFold 3, which is not used in iterative optimization, suggesting that optimizing against ESMFold2-Fast generalizes to a second folding model. Prompteina leads both AlphaFold 3 columns.

Table i2xL.2. Expected top-k ipSAE_min, averaged over 20 targets. The i.i.d. columns show matched-budget non-iterative baselines for both folding models, taken from App. Table 25. The iterative optimization columns show scores for both folding models.

MethodESM i.i.d.AF3 i.i.d.ESM iterativeAF3 iterative
Prompteina0.4550.5180.5610.591
BoltzGen0.3590.4200.4790.511
RFdiffusion30.2370.2760.2700.306

[Q3] CATH prompt format and examples. For simple tasks like CATH class, we use prompt formats like This protein has an Alpha Beta structural class. We also test robustness to paraphrasing, testing different formats such as CATH class: Alpha Beta. and The overall fold falls into the Alpha Beta category. We quantify this robustness across four wordings of the same specification, at matched lengths, for CATH class and helix count prompts (Table i2xL.3). Prompteina has the highest mean and the lowest standard deviation on both properties. We report this new comparison in App. E.1, Table 14, alongside the example prompts and paraphrases in Table 13.

Table i2xL.3. Mean and S.d. across four wordings of the same specification. CATH class accuracy is top-1 over three classes (33.3% chance). Helix count is the Spearman correlation between requested and realized counts.

ModelCATH class Mean (%)CATH class S.d.Helix count ρ MeanHelix count ρ S.d.
Prompteina97.92.50.9350.017
Chroma (text)34.29.10.7330.038
BioMatrix35.87.40.5940.107
Pinal35.414.00.7040.038

We show prompt-structure examples for Table 1b tasks in Fig. 3, with additional examples in App. Fig. 7.

[Q4] Pair representation conditioning. Currently, the pooled text embedding is not injected into the pair conditioning; the pair representation receives only the timestep embeddings. Text conditions the pair representation only indirectly, because the pair features include pairwise distances computed from the noisy and self-conditioned coordinates, and earlier text-conditioned steps produced those coordinates. We appreciate the reviewer's suggestion of directly text conditioning the pair representation, and will leave this to future work.

We thank the reviewer again for their time and feedback, and welcome any further clarifications!

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

The paper introduces Prompteina, a natural-language-conditioned model that jointly generates protein sequences and all-atom structures. It extends the pretrained Complexa generator with pooled text conditioning, token-level cross-attention, and an auxiliary text–residue alignment objective. The authors construct Prompteina-Data from PDB, AlphaFold Database, and synthetic Teddymer structures, pairing approximately 1.8 million chain-level entries with 52.4 million captions. Experiments cover structural-property control, generation from protein-family descriptions, interface steering, and agent-guided binder design with iterative feedback. A prospective GM2A campaign identifies one candidate with a concentration-dependent binding signal among 28 expressed designs.

Critical strengths
  1. A practical way to guide protein generation with text. Prompteina combines global and residue-level text conditioning, allowing fold, shape, and local structural requirements to be expressed through a single interface. The large paired dataset, supported by caption-quality and text-span audits, is also a valuable contribution. Prompteina leads the reported comparison on 15 of 18 structural-control tasks.
  2. Useful controls make the reported gains more convincing. Matched-length and empty-prompt tests show that structural control is not explained by length alone (Table 11). Removing interface descriptions lowers expected top-k ipSAE_min by 0.091, with a similar drop when epitope and length are held fixed. The ranking also holds under AlphaFold 3 rescoring, and iterative comparisons use matched generation budgets. These checks provide meaningful support for the richer conditioning interface.
Critical weaknesses
  1. Improved scores do not yet establish precise control or stronger binding. The CATH verifier has only 47.2% per-category accuracy on the prompted architecture subset (Appendix F). Count correlations measure trends, not exact satisfaction: disulfide correlation is 0.294 (Table 9), and interface correlations of 0.994–1.000 use means over 50 designs per requested count (Figure 7). Composite tasks do not report joint satisfaction rates.

    ProTrek measures description compatibility, not biochemical activity, and some advantages over Pinal remain statistically unresolved (Appendix G). Likewise, calibration on 5,178 bind/no-bind labels supports ipSAE_min for screening, not affinity ranking. AlphaFold 3 rescoring and the single-replicate GM2A experiment without kinetic fitting do not establish stronger binding or higher experimental hit rates. The wet-lab result is valuable as a proof of concept, but claims such as "strongest predicted binders" need qualification.

  2. The source and generality of the gains remain unclear. In Table 8, pooled text and CATH labels perform similarly; the large gap appears when cross-attention is added only to text. Longer label training does not resolve this architectural mismatch. A label-token control with the same conditioning pathways, information, and budget would better isolate text's contribution. Table 9 similarly lacks a matched 10.2K-step continuation without alignment loss, leaving alignment benefits confounded with additional training.

    Table 10 supports an advantage over the tested systems, not an inherent limitation of inference-time guidance or sequence-first generation in robust, fine-grained sequence–structure control. Data, supervision, and conditioning differ; paraphrase tests do not establish comparative robustness, and ProteinMPNN redesign does not validate the co-generated sequence. Finally, the novelty analysis uses different reference sets and omits Teddymer. It cannot replace an overlap audit covering new and inherited training data or establish generalization to unfamiliar targets and property combinations.

Questions for authors
  1. How separate are training and evaluation? Could the authors clarify the sequence and structural overlap or similarity between the evaluation families/binder targets and the new and inherited training data, including Teddymer?
  2. How often are the requested constraints met? Can the authors report per-design count errors and joint satisfaction rates for composite prompts?
  3. Can the conditioning and alignment effects be separated? For Table 8, can structured label tokens use the same pooled and cross-attention pathways as text, with matched semantic information and training? For Table 9, does alignment retain its advantage over a matched 10.2K-step continuation without the alignment loss?
  4. Do higher binder scores indicate stronger affinity? Is there evidence that ipSAE_min ranks measured affinities within targets, beyond distinguishing binders from non-binders? Otherwise, please qualify the binding-strength claims and report uncertainty across independent agent runs or seeds for the iterative score gains.
  5. What supports the claimed limitation of other generation paradigms? Could the authors clarify what evidence attributes limitations in robust, fine-grained control over both sequence and structure to inference-time guidance or sequence-first generation, rather than differences in data, supervision, and conditioning? If this distinction has not been isolated, the Introduction could present it as motivation rather than an established limitation.
4 comments

We thank the reviewer for their detailed and constructive review. We provide new experiments addressing each weakness and question below.

Precise control and constraint satisfaction

[W1, Q2] How often are constraints met exactly? We report per-design satisfaction for every count-valued property (Table STco.1). Prompteina leads on helix, strand, sheet and disulfide counts, and Pinal leads on salt bridge count.

Table STco.1. Per-design agreement with the requested count: exact match %, within-one match %, and mean absolute error in counts.

MethodHelix countStrand countSheet countDisulfide countSalt bridge count
Prompteina22.6 / 62.6 / 1.4514.6 / 51.1 / 1.7719.1 / 43.7 / 2.1718.6 / 45.1 / 2.256.0 / 22.9 / 4.73
Chroma-text12.0 / 31.7 / 2.8710.3 / 32.0 / 3.2718.0 / 44.6 / 2.001.1 / 15.7 / 3.855.4 / 18.6 / 5.61
Pinal13.1 / 44.0 / 2.4210.0 / 36.0 / 2.6812.6 / 35.4 / 2.5510.6 / 32.0 / 2.9310.6 / 32.9 / 2.89
BioMatrix13.7 / 46.6 / 2.237.1 / 24.0 / 3.2110.0 / 30.6 / 2.817.5 / 26.1 / 3.258.9 / 28.9 / 3.07

Composite prompts request two properties at once, and we report how often a design meets both exactly, over 4,550 designs (Table STco.2). Prompteina leads on all four joint tasks. Its lowest joint satisfaction, 12.1% on CATH class with strand count, is above the highest value any baseline reaches on any task, 8.4%. Each baseline receives the same composite prompt we give Prompteina.

Table STco.2. Joint satisfaction (%): the fraction of composite designs that meet both requested properties exactly.

Methodclass × helixclass × strandclass × sheetclass × disulfide
Prompteina20.312.123.121.7
Chroma-text5.44.78.41.1
Pinal5.14.45.13.6
BioMatrix5.62.95.32.8

[W1] GearNet verifier accuracy. The reviewer points out that the GearNet CATH verifier has only 47.2% per-category accuracy on CATH architecture. This is the standard fold classifier for this benchmark, introduced by Proteina, and every method in our comparison is scored with it. This reflects how often GearNet labels native structures correctly (App. F), so it measures the verifier quality rather than our designs.

Binding claims, wet-lab validation, and affinity

[W2] Do the wet-lab results establish binder design? Yes. We now report expanded results on our GM2A campaign. In our initial sample, only 1 of 28 assayed designs bound. Our final wet-lab results show nine binders among 89 assayed designs, with a lowest apparent K_D of 107.4 nM. We report the first de novo binders against GM2A, where a prior campaign tested 45 designs with no binders [1, 2]. We describe the expanded campaign in §3.3 and Fig. 6, and report the assay results in App. J.2 and Table 34.

[Q4] ipSAE$_\text{min}$ and binding strength. When referencing our ipSAE$\text{min}$ results, we incorrectly used the term "binding strength". **We now state that ipSAE$\text{min}$ is a proxy for predicted binding confidence rather than for measured binding affinity.** Prior work calibrates interface confidence against binding, not affinity [3, 4]: in [4], a co-folding score separates binders from non-binders on 12 of 13 targets, at 0.52 mean average precision against 0.31 by chance. Estimating binding affinity from a predicted binder-target complex remains an unsolved problem.

[Q4] Uncertainty on the iterative gains. We report the ipSAE$_\text{min}$ gain of iterative optimization over i.i.d. sampling for Prompteina, BoltzGen and RFdiffusion3 (Table STco.3). For each method we give a 95% confidence interval and a sign test across the 20 targets. Prompteina improves on all 20 targets and gains the most, with every interval excluding zero.

Table STco.3. Mean ipSAE$_\text{min}$ gain over i.i.d. sampling, with targets improved, 95% CI of the gain and a sign test across the 20 targets.

MethodGainTargets improved95% CISign test
Prompteina+0.066820/20[+0.037, +0.097]p < 0.001
BoltzGen+0.057420/20[+0.036, +0.079]p < 0.001
RFdiffusion3+0.037116/20[+0.011, +0.063]p = 0.012

Evaluation of prompts specifying function

[W2] Validity of ProTrek as a metric. We agree that ProTrek measures description compatibility. Functional prompts lack a directly verifiable structural property (which we mention in §3.1.2), so we evaluate with the same metric as Pinal and BioMatrix for fairness [5]. ProTrek has shown some utility in tracking biochemical activity: in the Pinal paper, the authors score their candidate pools with ProTrek before synthesis, and designs from four functional classes were active in the wet lab, two of them enzymes with catalytic turnover [6].

Isolating the contribution of natural language

[W3, Q3] Does natural language beat label conditioning at a matched budget? Yes. We now include a new experiment suggested by the reviewer: a dedicated CATH-label conditioner with the same pooled and cross-attention pathways as text. We fix the architecture, data and compute budget at 47,275 steps, so the runs differ only in the conditioning signal. Under this matched setting, natural language outperforms label conditioning (92.0% vs. 66.7% on CATH class, 27.0% vs. 25.1% on architecture; Table STco.4). We report this new ablation in App. D, Table 9.

We only evaluate CATH for these ablations because the CATH-label conditioning runs are limited to CATH class and architecture, whereas the natural-language runs support many properties beyond CATH (last column of Table STco.4; Table 1, App. Table 11). Specifying constraints in natural language lets us scale the number of properties we condition on, while a label interface requires a predefined embedding table for each new property.

Table STco.4. CATH top-1 accuracy (%) by conditioning interface and pathway. The last column lists the properties each interface controls; the 18 text tasks are measured on the released model, not on these matched runs.

PathwaysConditioningCATH classCATH architectureProperties shown to be controllable
pooled + cross-attentionCATH label66.725.1CATH class, architecture (2 tasks)
pooled + cross-attentionNatural language92.027.0fold, shape, secondary, tertiary, composite (18 tasks)

[W3, Q3] Does alignment still help against a matched continuation? Yes, on three of the four metrics in Table 10. We continue training for the same 10,200 steps on the same data, with the alignment loss removed and every other setting matched to the aligned run. The matched control improves on CATH architecture, but degrades on disulfide count, salt bridge count and CATH class (Table STco.5). This suggests alignment is essential for control of fine-grained properties.

Table STco.5. Requested-versus-realized correlation and CATH accuracy (%) before alignment, after alignment, and after a matched 10,200-step continuation without the alignment loss.

CheckpointDisulfide count ρSalt bridge count ρCATH class accuracyCATH architecture accuracy
Before alignment0.0540.56694.055.5
After alignment0.2940.66396.051.1
Matched control, no alignment-0.1070.53990.755.8

Other generation paradigms and robustness

[W4, Q5] Limitations of other generation paradigms. Our goal is to show that text-conditioning an all-atom structure generation model gives zero-shot control over properties of both sequence and structure. Search is expensive and needs a reward tuned per property and per target: Complexa runs 16 GPU hours per easy target and over 100 on TNF-α, H1, and IL17A, where it also adds a pLDDT term to its reward [7]. DiSCO needs a custom reward, distance thresholds and hyperparameters for each property [8]. Sequence-first models such as Pinal cannot condition on structural constraints, which many protein design tasks require, including protein and small-molecule binding [6]. We have revised Related Work to make this motivation clearer.

[W4, Q5] Is Prompteina more robust to paraphrasing than the baselines? Yes. To establish comparative robustness, we extend our robustness evaluation on the CATH class and helix count properties (App. Table 13) to Chroma (text), BioMatrix, and Pinal. For each property, we specify the prompts in four different ways, with lengths matched across the generations. Prompteina has the highest mean and the lowest standard deviation on both properties (Table STco.6). We report this new comparison in App. E.1, Table 14. We also show that Prompteina remains robust on a finer-grained steering task, binder interface residue count, which none of the other models support, at ρ 0.976 to 0.997 (App. Table 19).

Table STco.6. Mean and S.d. across four wordings of the same specification. CATH class accuracy is top-1 over three classes (33.3% chance). Helix count is the Spearman correlation between requested and realized counts.

ModelCATH class Mean (%)CATH class S.d.Helix count ρ MeanHelix count ρ S.d.
Prompteina97.92.50.9350.017
Chroma (text)34.29.10.7330.038
BioMatrix35.87.40.5940.107
Pinal35.414.00.7040.038

[W4, Q5] ProteinMPNN redesign and validation of the co-generated sequence. Inverse-folding redesign is the standard self-consistency protocol for generated backbones, used by RFdiffusion [9], BoltzGen [1] and Complexa [7], which reports redesign, redesign with a fixed interface, and its own generated sequence. For validation of the co-generated sequence, we show in Table 3 that Prompteina reaches 0.455 expected top-k ipSAE$_\text{min}$ under the self condition, which scores each model’s own sequence, against 0.400 for the next best method.

Training/evaluation overlap, novelty, and generalization

[Q1] How separate are the training and evaluation data? For functional evaluation (§3.1.2), we evaluate models on whether they can recover a family’s fold. 121 of the 122 evaluation families have a near-identical sequence in our training data, which we define as a match over at least 80% of the family sequence’s length with at least 90% of those residues identical. All 122 are present in Swiss-Prot, and both Pinal and BioMatrix train on Swiss-Prot, so the families are in every compared model’s corpus. For binder design we use the 20-target panel as in Complexa, and the same Teddymer and PDB dimer dataset. One target, insulin, has a close match in that data, at 99.3% identity to a chain of one Teddymer dimer. Prompteina’s designed insulin binders are less similar to that dimer’s partner chain than Complexa’s are, at a highest TM-score of 0.306 against 0.488. Both are below the 0.5 threshold for being considered a shared fold [10], so neither model reproduces the training structure.

[W4, Q1] Are the generations novel against the training data, including Teddymer? Yes. We now include the Teddymer dataset as part of our novelty analysis. For both Complexa and Prompteina, we use the model’s own training data as the reference set. Prompteina’s reference set is the larger of the two (1,688,289 chains against 1,416,251), which biases the comparison against us. Our generations are more novel than Complexa’s (Table STco.7). Teddymer supplies the nearest match for 22.2% of our designs and 38.2% of Complexa’s, so leaving it out never favoured us.

Table STco.7. Mean novelty of generated monomers, where novelty is one minus the TM-score to the nearest chain in the model’s own training data. The second column adds the 908,094 Teddymer chains to each reference set.

ModelWithout TeddymerWith Teddymer
Prompteina0.2300.227
Complexa0.2170.215

[W4, Q1] Does Prompteina generalize to an unfamiliar target? Yes. Our wet-lab campaign on GM2A demonstrates generalization to unseen targets, given that GM2A appears in none of the 454,105 Teddymer dimers or 45,856 PDB complexes in our binder training data, which we searched by sequence and by accession. The nearest match among the Teddymer chains covers 64.2% of GM2A at 28.5% identity and an E-value of 224.9, which indicates no detectable homology. We obtain nine binders among 89 assayed designs, with a lowest apparent K_D of 107.4 nM. Prompteina yields binders against a target where no other design model has, despite prior campaigns.

We thank the reviewer again for their time and feedback, and welcome any further clarifications!

  1. Stark et al. BoltzGen: Toward Universal Binder Design. bioRxiv, 2026.
  2. Uçar et al. BoltzProt-1: Towards Efficient De Novo Binder Design with Good Developability. bioRxiv, 2026.
  3. Overath et al. Predicting Experimental Success in De Novo Binder Design: A Meta-Analysis of 3,766 Experimentally Characterised Binders. bioRxiv, 2025.
  4. Shanehsazzadeh. Autonomous De Novo Protein Binder Design with Claude. arXiv, 2026.
  5. Su et al. A trimodal protein language model enables advanced protein searches. Nat Biotechnol, 2025.
  6. Dai et al. Toward De Novo Protein Design from Natural Language. bioRxiv, 2025.
  7. Didi et al. Scaling Atomistic Protein Binder Design with Generative Pretraining and Test-Time Compute. arXiv, 2026.
  8. Rector-Brooks et al. General Multimodal Protein Design Enables DNA-Encoding of Chemistry. arXiv, 2026.
  9. Watson et al. De novo design of protein structure and function with RFdiffusion. Nature, 2023.
  10. Xu and Zhang. How significant is a protein structure similarity with TM-score = 0.5? Bioinformatics, 2010.

Thank you for the response. I will maintain my current positive score.

Comments on the paper

1 comment

We thank the reviewers for their feedback. We summarize and address their main concerns below, with further details in our replies to each reviewer.

1. Do the GM2A experiments establish successful binder design and generalization to unseen targets? (Reviewer STco). No de novo binders have previously been reported for GM2A, including in a prior campaign that tested 45 designs without success [1, 2]. In our expanded experiments, we identify nine binders out of 89 designs tested, compared to one out of 28 in our initial submission, with the strongest binder at an apparent K$_\text{D}$ of 107.4 nM. Since GM2A is absent from our binder training data, these results suggest that Prompteina generalizes to unseen targets and yields binders where other state-of-the-art models do not. We describe the expanded campaign in §3.3 and Fig. 6, and report the assay results in App. J.2 and Table 34.

2. Does natural language conditioning outperform label conditioning with the same architecture, training data, and compute? (Reviewers STco and Xha9). To isolate the contribution of natural language, we train a label-conditioned model with the same architecture, data, and number of training steps. The text-conditioned model outperforms the label-conditioned one (92.0% vs. 66.7% on CATH class, 27.0% vs. 25.1% on CATH architecture), suggesting that the gains come from natural language conditioning. Specifying constraints in natural language lets us scale the number of properties we condition on, while a label interface requires a predefined embedding table for each new property. We report this new ablation in App. D, Table 9.

3. Do designs satisfy the requested constraints exactly? (Reviewer STco). As requested by the reviewer, we now report per-design exact match rates for count-based properties, and joint satisfaction rates for composite prompts. Prompteina has the highest exact match rate on helix, strand, sheet and disulfide counts, and leads on all four composite tasks: on CATH class with helix count it reaches 20.3% against 5.6% for the best baseline. We report the per-design results in our response to Reviewer STco.

4. Is prompt-following preserved after refolding, and do gains from iterative optimization generalize to a held-out folding model? (Reviewer i2xL). We now add a comparison of each property before and after refolding, across all methods (App. H.7, Table 27). We measure five properties: helix, sheet and coil fraction, interface residue count and buried surface area. Prompteina preserves properties better than any baseline, at a mean correlation of 0.742 against 0.671 for the next best method. We also score the iteratively optimized designs with AlphaFold 3, which is not used during optimization. Prompteina leads on expected top-k ipSAE$_\text{min}$ at 0.591 against 0.511, so optimization against ESMFold2-Fast generalizes to a held-out folding model. We report the AlphaFold 3 scores in our response to Reviewer i2xL.

5. Novelty in relation to Proteo-R1 and CodeFP (Reviewer Xha9). The core difference between Prompteina and these two methods is whether the generative process is conditioned directly on text. In Prompteina, the text embeddings of a free-text caption condition the denoising process itself. In Proteo-R1, a reasoning model reads the text and selects key residues in a given antibody-antigen complex, and the generator conditions on those residues. Similarly, CodeFP maps the text to a term from a closed 375-term Gene Ontology vocabulary, and conditions on a structural motif retrieved for that term. We have made the wording of the claim precise, and we now cite both works in Related Work and in the conditioning survey in App. A [3, 4].

We have updated the manuscript to reflect these results, with the revised text highlighted in blue. We thank the reviewers again for the suggestions that improved both our results and the clarity of our claims!

  1. Stark et al. BoltzGen: Toward Universal Binder Design. bioRxiv, 2026.
  2. Uçar et al. BoltzProt-1: Towards Efficient De Novo Binder Design with Good Developability. bioRxiv, 2026.
  3. Wu et al. Proteo-R1: Reasoning Foundation Models for De Novo Protein Design. ICML, 2026.
  4. Chen et al. Co-Generative De Novo Functional Protein Design. ICML, 2026.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.