acceptodds
Under review as a conference paper at ICLR 2027

EnzymeCoT: De Novo Enzyme Design As A Chain Of Reasoning Coupled To A Chain Of Structure

Abstract

Enzyme design aims to produce a sequence and a structure that together catalyse aspecified reaction. Existing conditional generators are given either a compact functional label, from which every decision is then made implicitly inside the sampler, or structural information such as catalytic sites or a motif, which de novo design is meant to produce rather than consume. Neither writes the design down in steps. We propose EnzymeCoT, which runs two coupled chains. From a de- scription of the reaction and a target length, a language model writes a chain of reasoning,the four remaining states of a five-state design record, moving from amechanistic prior through a catalytic core and an active site to a backbone and asequence. Each state names a wider part of the molecule than the last, and a masked discrete diffusion model over joint structure and sequence tokens is re- conditioned on it and asked to commit that part alone, so the written chain drives a chain of structure through the same five states. Where chain-of-thought in a lan- guage model feeds only the language model, each state here conditions the gen- erator directly, and at two depths: residue–position pairs at the input embedding, and the full token sequence by cross attention late in the denoiser. The corpus pairs enzymes with staged design rationales: deterministic code produces every claim and a verifier checks every sentence against it. Conditioned on the reaction and a target length, EnzymeCoT improves by 2.2 points of EC3 agreement on the strongest baseline evaluated under a specification that does not name the target class, and refolds closer to the reference backbone.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
de novo enzyme design, chain-of-thought reasoning, protein language models, conditional generation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_SmwtOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors propose, EnzymeCoT, a framework for de novo enzyme design that couples a LLM-generated chain of reasoning with a protein generator's chain of structural commitments. The LLM produces a staged design record, progressing from a mechanistic prior to a catalytic core, active site, backbone, and full amino acid sequence. The protein generator is reconditioned at each stage to generate the corresponding molecular components. A large dataset is constructed for model training and evaluation. Experimental results demonstrate the effectiveness and performance advantage of the proposed approach.

Critical strengths
  1. The paper introduce a intuitive decomposition of the enzyme-design problem, which provides a natural hierarchy for progressively committing to increasingly large portions of a enzyme protein.
  2. The separation of LLM and protein generator is techniclly sound. The two conditioning mechanisms provide a reasonable way to combine precise molecular commitments with high-level semantic information.
  3. A large dataset is constructed for model training and evaluation, and the experimental results demonstrate the effectiveness and performance advantage of the proposed approach.
Critical weaknesses
  1. The design record and the generated protein update in parallel, but the textual state can be inconsistent with the structural state. The paper should distinguish the staged generation conditioned on a textual plan with a closed-loop reasoning framework. It would be better to quantify the consistency between textual commitments and the generated structures at each stage.
  2. The ablation study should be discussed more in detail. The current evidence do not establish precisely why the staged approach helps. How about a single-pass model receiving the complete design record?
  3. The author should report the computational cost and runtime.
Questions for authors

Please refer to the weaknesses.

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

EnzymeCoT combines a staged language-model design record with progressive sequence–structure generation conditioned on a functional specification. The authors construct a corpus of 56,022 records, check the intermediate text with rules, and train the model with continuous and textual conditioning. The computational evaluation covers functional-class predictions, structural and distributional metrics, component ablations, and results stratified by similarity. The main table reports better EC3 accuracy and micro-F1, but the method does not lead on every metric. The paper also acknowledges the verifier's limited scope and the lack of experimental functional validation.

Critical strengths

The textual records are associated with successive generation states, and the authors build a corpus to supervise those associations. This gives the reader intermediate statements to inspect during generation. Together with the gains in functional-class agreement, this is a promising direction.

Critical weaknesses
  1. Table 1 gives an EC3 accuracy of 47.2%, but Figure 3b labels 1673 designs as correct and 1128 as incorrect. Those counts give 59.73% over 2801 cases. The figure describes the full test set, and I could not find a different checkpoint or definition of correctness that reconciles the two results. It is difficult to judge the size of the main empirical gain.
  2. The EC3 improvement over CodeFP in the main table is 2.2 percentage points, and macro-F1 is lower than CFP-Gen. The method adds four language-generation stages and four denoising rounds. Yet the baselines receive different conditioning information, and the ablations change that information together with the staged computation. The effect of reasoning is therefore entangled with extra information and compute. A simpler baseline with comparable conditioning and cost is needed to judge the value of the extra stages.
  3. As the paper acknowledges, the verifier checks entities, quantities, and expression rules against source material, it does not verify mechanistic relations. A statement can pass these checks without accurately describing the generated design. The functional-class and structural metrics do not close that gap, either. The paper needs evidence that the intermediate records describe the outputs accurately or help identify failures. At present, the claimed interpretability benefit is largely assumed.
Questions for authors
  1. I understood Table 1 and Figure 3b to refer to the same model and full test set, but they give 47.2% and 59.73%(1673/(1673+1128) ), respectively. Is there a difference in checkpoint or the definition of correctness? Which result is used in the subsequent analyses?
  2. Which comparison isolates the reasoning content while keeping conditioning information and total computation comparable? How do the extra stages affect training cost and inference time?
  3. What evidence shows that the intermediate claims accurately describe the generated outputs? How often do the records contain statements that pass the source-based verifier but conflict with the final output?
  4. When a design matches the target at EC3 but has a different full EC label, what functional differences can remain? How often does this occur in the reported results, and which classes account for the lower macro-F1?
  5. Figure 3c reports EnzymeCoT's results by similarity to the training data. How do the baselines perform in those same groups, particularly the 803 test records least similar to the training set? Does the advantage reported in Table 1 remain in that group?
Reviewer_pkyuOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EnzymeCoT, a framework for de novo enzyme design that couples LLM-generated reasoning with protein sequence and structure generation. The method uses a language model to progressively describe catalytic residues, active sites, backbone structures, and sequences, which condition a discrete diffusion model at different stages. The authors construct approximately 56,000 enzyme design records with automatically generated and verified rationales. Experiments on known enzyme classes show modest improvements in computationally predicted enzyme-function agreement and structural similarity over the evaluated baselines.

Although the integration of structured reasoning into protein generation is interesting, I have significant concerns regarding the practical relevance, evaluation, and novelty of the generated designs.

Critical strengths
  • The integration of step-by-step textual reasoning with protein sequence and structure generation is an interesting direction.
  • The construction of a large enzyme design dataset with automatically verified rationales is potentially useful for future research.
  • The method demonstrates improvements over the evaluated conditional protein generation baselines, and the ablation studies support the contribution of staged conditioning.
Critical weaknesses
  1. Limited relevance to the central challenges of de novo enzyme design

My primary concern is that the paper does not adequately address the central challenges of modern de novo enzyme design. The evaluation primarily measures whether generated proteins reproduce known enzyme classes and resemble their reference structures.

However, two important directions in enzyme engineering are (1) generating novel protein scaffolds that preserve functional catalytic sites and (2) designing enzymes with new catalytic activities or altered substrate selectivity. The paper does not demonstrate either capability.

While reproducing known enzyme functions is a meaningful task, it remains unclear how the proposed model would be useful for practical enzyme engineering. Furthermore, the comparison is limited to a few conditional protein generators and does not sufficiently consider established enzyme-design approaches, particularly motif-scaffolding methods such as RFdiffusion3 and recent structure-guided design pipelines. Although these methods operate under different input conditions, comparisons under appropriately defined tasks would help establish the practical significance of EnzymeCoT.

  1. Lack of evidence for novel catalytic-site generation

If the authors aim to position EnzymeCoT as a framework for genuinely novel enzyme design, they should demonstrate whether it can propose new catalytic motifs or active-site configurations.

The training rationales are derived from annotations of known enzymes and predicted enzyme–ligand complexes. However, the framework does not explicitly model reaction chemistry, transition-state stabilization, or catalytic energetics, and the paper provides no evaluation of whether the generated active sites are chemically plausible or catalytically functional.

The reported EC3 agreement is based on a function prediction model rather than actual enzymatic activity. Without evaluating active-site geometry, substrate binding, or catalytic activity, it is difficult to establish that the generated proteins could perform the intended reactions.

  1. Insufficient evaluation of sequence and structural novelty

The paper uses maximum 8-mer containment to characterize similarity between test enzymes and training sequences. However, this is a relatively limited measure of sequence similarity and does not adequately establish evolutionary or structural novelty.

More importantly, this analysis concerns reference test enzymes rather than the novelty of the generated designs themselves.

I would expect a systematic evaluation of generated-sequence identity against known enzymes, as well as structural similarity to existing protein structures using established tools such as MMseqs2 and Foldseek. Such analyses would help distinguish genuinely novel enzyme designs from the reconstruction of known enzyme families.

Questions for authors
  1. Can EnzymeCoT generate novel protein scaffolds while preserving functional catalytic-site geometry? How does it compare with established motif-scaffolding methods under a comparable design task?
  2. Can the model propose novel catalytic motifs or support new substrate specificities beyond those represented in the training data?
  3. Why was 8-mer containment chosen to characterize training–test similarity? Can the authors provide systematic sequence-identity and structural-novelty analyses of the generated proteins?
  4. Is there evidence that the generated catalytic sites are geometrically and chemically plausible, beyond agreement with a computational enzyme-function classifier?
  5. How does the computational cost of the multi-stage reasoning framework compare with simpler conditional generators, and does the improvement justify this additional complexity?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.