acceptodds
Under review as a conference paper at ICLR 2027

From Target Molecule to Simulation-Ready Digital Microbial Life

Abstract

Engineering microorganisms to produce target molecules is a central goal of synthetic biology, yet designing a simulation-ready microbial system remains labor-intensive and dependent on expert knowledge. Although large language model (LLM)-based scientific agents can automate parts of this workflow, generating executable biological designs and evaluating their quality remain challenging. We propose digital microbial life design as a new end-to-end problem of transforming a target molecule into a simulation-ready microbial design. To solve this problem, we introduce Origo, a multi-agent system that structures the design workflow using an expert-designed primitive task pool. Each primitive task is bound to specialized MCP tools, while hierarchical multi-agent coordination assigns tasks to specialized agents and coordinates their execution to construct GEMs, and identify enzyme candidates. To address the corresponding evaluation gap, we introduce DLBench, a benchmark for evaluating digital microbial life design through network-level executability and reaction-level biological plausibility, with 40 target products and 22 rate-limiting enzymes. Experiments on DLBench show that Origo outperforms general-purpose LLMs and biomedical agents across both tasks. This work provides a foundation for AI-driven microbial design, downstream simulation, and hypothesis generation for wet-lab validation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
multi-agent system, synthetic biology, genome-scale metabolic model, enzyme mining

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_Kd91OpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Origo, a multi-agent framework that integrates tools based on expert-designed strategies to automate the construction of genome-scale metabolic model (GEM) for producing target molecules and the identification of enzyme candidates. It also introduces DLBench, a benchmark for evaluating these capabilities.

Critical strengths
  1. The motivation for developing the framework is clear. Although the framework consists of numerous components, each component is described in detail.

  2. The work involves substantial engineering effort, and the authors develop a dedicated benchmark to evaluate their proposed framework.

Critical weaknesses
  1. The Related Work section does not clearly articulate the specific differences between this work and prior studies. Furthermore, the Method section lacks sufficient justification for the design choices behind the individual components of Origo and DLBench. As a result, the paper reads more like a technical report describing Origo and DLBench than a research paper.

  2. The paper primarily presents the evaluation results of Origo without providing detailed analyses or ablation studies. Furthermore, to establish DLBench as a meaningful benchmark, the number of benchmark instances should be sufficiently large, and validation by human experts, rather than relying solely on an LLM judge, is necessary.

Questions for authors
  1. Although Origo appears to be primarily an engineering effort, what specific research problems does it aim to address, and how does the proposed framework solve them? Could the authors provide comprehensive ablation studies and analyses to justify the design choices behind each component?

  2. Could the authors substantially increase the number of benchmark instances and provide a human expert baseline? Could they also provide qualitative comparisons with existing models

4 comments

Thank you for asking us to clarify the research question and the reasons behind Origo's design. We study the extent to which an agent system can understand and reconstruct biological metabolic networks. Origo tests this through target-driven de novo construction: starting from pathway and reaction knowledge, agents assemble the supporting metabolism into an executable SBML model and revise it using network diagnostics. Starting without a pre-existing host GEM requires the system to organize the metabolic infrastructure itself. This provides a computational test of metabolic understanding through the connections, biochemical constraints, and functions represented in the generated network.

The task definition and discussion scope is different from related methods, and that also drives us to build DLBench. ChatGEM analyzes and modifies existing GEMs; MetaGEM reconstructs networks from organism-specific metabolites and candidate proteins; D2Cell connects literature-derived engineering knowledge to predictions using existing GEM structures. SubNetX integrates production subnetworks into a host GEM, while EnzRetro predicts enzymatic routes and EC classifications. Origo examines the combination of target conditioning, pathway-based assembly, and system-level SBML construction without a host GEM as input. Under this definition, its scope differs from that of the five systems compared here.

WorkDe novo assemblyRequires an existing host GEMProducing or using whole (not partial) networkConstructs networks under a product constraint
ChatGEM (10.64898/2026.07.20.739662)×✓✓×
MetaGEM (10.48550/arXiv.2605.14812)××✓×
D2Cell (10.1016/j.tibtech.2026.03.026)×✓✓×
SubNetX (10.1038/s41467-025-59827-7)×✓✓✓
EnzRetro (10.1002/exp2.70129)×××✓
This work✓×✓✓

The scientific contribution is a computational investigation of how far agent systems can understand and reconstruct metabolic organization, with de novo SBML generation making that capability open to inspection and testing. It provides an exploratory foundation for the metabolic modeling infrastructure of digital cell factories and virtual cells.

Manuscript changes: Section 2 and Appendix E discuss the five related systems and distinguish their task requirements from Origo's settings.

Thank you for requesting ablations to examine the design choices. We added comparisons that change agent organization, knowledge access, MCP availability, and tool exposure. All use the same ten targets sampled from the original set with seed 11062026:

Experimental GroupXML OutputsSBML ValidityGrowth SuccessReference Production SuccessMean Memote ScoreMean LLM Score
Origo (original paper)10/10100%100%100%0.627.20
Single agent7/1070%70%20%0.576.00
No knowledge10/10100%100%100%0.486.80
No MCP2/1010%0%0%0.372.00
All tools/task9/1090%90%80%0.566.22

Success percentages use ten target requests per condition. XML outputs are 10, 7, 10, 2, and 9, respectively. Memote is evaluated only for SBML-valid models. The Origo reference is taken from the original paper for the same sampled targets.

Removing MCP most strongly affects whether a valid, growing model is produced. Removing knowledge leaves the binary functional rates unchanged in this subset but lowers Memote, while the single-agent and all-tools-per-task variants reduce quality. Together, these results show that tool access, knowledge-guided construction, and task organization contribute at different stages of producing an executable and curated network.

We also ran Origo with eight alternative backbones on the same target subset to examine how the results depend on the underlying model:

MethodNSBML validity (%)Growth (%)Product formation (%)Memote score (0-1; mean)LLM overall score (1-10; mean)
Origo-DeepSeek-V4-Flash (Used in This study)10100.00100.00100.000.627.20
Origo-GPT-6-Astra10100.00100.00100.000.616.80
Origo-GPT-6.1-Sol10100.00100.00100.000.616.90
Origo-GPT-6-Luna10100.00100.00100.000.617.20
Origo-GLM-5.310100.00100.00100.000.486.60
Origo-GLM-5.3-Flash9100.0077.80100.000.386.33
Origo-Kimi-K310100.00100.00100.000.526.90
Origo-Qwen3.8-Max10100.00100.00100.000.566.90
Origo-Qwen3.8-Flash10100.00100.00100.000.566.90

Validity and product formation reach 100% among available models for all eight alternatives. Growth is 100% except for GLM-5.3-Flash, at 7/9. Mean Memote ranges from 0.38 to 0.61. The framework therefore supports several backbones, with differences in curation and overall quality.

Manuscript changes: Appendix F adds the ablation study details (Figures S8–S9). Appendix G adds the backbone comparison (Figure S10).

Thank you for recommending human evaluation. We invited a biology PhD student and a rater with industry experience to conduct a double-blind assessment of 90 generated models, scoring all eight dimensions. Five additional LLM judges evaluated the same models. Together with the original judge, this gives the following comparison of six LLM evaluators and two human raters:

Generating methodDeepSeek-V4-Pro(Original)DeepSeek-V4.1-FlashKimi-K3Qwen3.8-MaxGPT-6 LunaGPT-6.1 SolBiology PhD StudentIndustry Expert
Origo7.207.006.806.605.705.208.207.20
Biomni6.206.606.705.804.204.205.405.30
STELLA2.402.302.302.101.501.601.901.60
GPT-5.53.303.504.103.302.602.603.102.80
DeepSeek-V4-Pro5.705.305.004.303.203.304.103.90
GLM-5.24.604.904.804.203.403.504.103.70
Qwen3.7-Max3.302.703.002.602.301.902.101.70
Kimi-K2.64.604.404.103.103.102.705.303.20
Gemini-3.5-Flash2.502.102.001.901.401.501.801.40

Overall quality on a 1–10 scale; means over ten models per method and judge. Bold entries mark column maxima.

To examine agreement beyond method-level means, we also compared Spearman’s rank correlation between paired scores across all 90 models. The complete correlation matrix is below:

JudgeDeepSeek-V4-Pro (Original)DeepSeek-V4.1-FlashKimi-K3Qwen3.8-Maxgpt-6-lunagpt-6.1-solBiology PhD StudentIndustry Expert
DeepSeek-V4-Pro (Original)1.000.920.850.870.840.810.830.85
DeepSeek-V4.1-Flash0.921.000.900.910.880.840.850.90
Kimi-K30.850.901.000.910.840.810.820.87
Qwen3.8-Max0.870.910.911.000.870.830.830.90
gpt-6-luna0.840.880.840.871.000.820.840.87
gpt-6.1-sol0.810.840.810.830.821.000.810.85
Biology PhD Student0.830.850.820.830.840.811.000.88
Industry Expert0.850.900.870.900.870.850.881.00

The backbone LLM use in this study ratings correlate 0.828 and 0.848 with the human scores; the five additional LLM judges have human correlations of 0.813 to 0.896. Pairwise correlations between LLM judges range from 0.809 to 0.920. These results support consistent rankings across evaluators.

Manuscript changes: Appendix H adds assessment of the sampled ten-target subset by two human raters and multiple LLM judges (Figures S11–S13).

Thank you for the response. While the differences from prior work are now laid out more clearly, I am still not convinced of how meaningful those differences actually are. I also think the number of benchmark instances remains too small (and the additional experiments were conducted on only ten targets). Regarding the human expert baseline, what I meant was that the performance of human experts performing the same task should be included, and a qualitative comparison still does not appear to have been provided. For these reasons, I will maintain my current score.

Reviewer_sLbEOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
3(+2)
Read review
Paper summary

The authors present Origo, a multi-agent system that designs digital microbial life, and DLBench, a benchmark to evaluate digital microbial life systems. Origo defines a workflow of MCP tools, subagents, and coordination to generate a Genome-scale metabolic model (GEM) and enzyme candidates. DLBench pulls together existing metrics like SBML validity, growth validation, product flux validation, and more to evaluate generated GEM designs.

Critical strengths

The paper reports Origo outperforming all evaluated LLMs on its own benchmark by a fairly large margin. If the benchmark and comparisons were sound, this would be a strong performance and contribution.

Critical weaknesses

Origo seems to be packaging a bunch of existing biological tools together into a workflow with multi-agent orchestration. A fair comparison would be to give other LLMs access to these same biological tools, but not have the whole workflow and multi-agent framework. Instead, the baseline for enzyme mining is a comparison to a single prompt asking the LLMs to one shot generate UniProt IDs with no access to tools like Foldseek, MMseqs2, ProstT5, etc. Meanwhile, Origo has access to all of them. The attached repository also doesn’t seem to explain how GEM generations are done with the baseline LLMs. Overall, Origo is mostly an orchestration workflow of subagent definitions and MCP calls without a clear methodological contribution, it is unclear how novel and appropriate DLBench’s chosen evaluation metrics are, and the prompt and available tools used to compare to other LLMs, to me, seems quite unfair, leading to biased results.

Questions for authors
  • How are GEM generations conducted with the baseline LLMs? What tools, if any, are given to them?
  • How would Origo compare to these LLMs if the LLMs also had access to existing biological tools, but without the multi-agent orchestration?
  • DeepSeek V4 Pro is used as an LLM expert evaluator in DLBench, but DeepSeek V4 Pro is also an evaluated model. Does this lead to any benchmark evaluation bias? Qualitatively, it seems like DeepSeek does comparatively much better on the LLM evaluation than on any other metric. Additionally, is DeepSeek V4 Pro known to give reliable evaluation for this task?
  • It would be helpful to show DLBench’s results on GEM and enzyme candidate designs that have been verified by human experts to be good and bad. Currently DLBench is being used to show that Origo is a good framework, but it’s unclear that DLBench itself is accurate.
3 comments
AuthorsOpenReview ↗

W1 and Q1-Q2

W1:Thank you for your comment. Origo addresses a specific reconstruction problem: organizing pathway and reaction building blocks into a target-constrained, system-level SBML network that can be simulated.

DLBench supplies the target and enzyme tasks, output requirements, and complementary assessments of parsing, network behavior, model curation, and biochemical quality. The added blind human scores and multiple LLM judges provide further evidence about the evaluation of generated models. The task pool and scoped tools encode the dependencies needed for this process, while the ablation test what changes when those constraints or capabilities are removed. Related Work detailed in Section 2 and Appendix E distinguish this task from extending an existing host GEM or predicting an isolated synthetic route.

Q1:Thank you for raising this concern. We would like to clarify that our evaluation includes both general-purpose LLMs and two tool-equipped biological agent systems, Biomni and STELLA.

All baseline LLMs and bio-agents receive the same query, “Generate a bacterial SBML model for {target_product} synthesis.” General-purpose LLMs generate the SBML model without access to external tools, while Biomni and STELLA operate with their respective integrated biological toolsets.

Q2:Thank you for asking whether the gains persist when a single agent has access to the biological tools. To examine this, we added single-agent, no-MCP, no-knowledge, and all-tools-per-task comparisons on the same ten targets:

Experimental GroupXML OutputsSBML ValidityGrowth SuccessReference Production SuccessMean Memote ScoreMean LLM Score
Origo (original paper)10/10100%100%100%0.627.20
Single agent7/1070%70%20%0.576.00
No knowledge10/10100%100%100%0.486.80
No MCP2/1010%0%0%0.372.00
All tools/task9/1090%90%80%0.566.22

Success percentages use ten target requests per condition. XML outputs are 10, 7, 10, 2, and 9, respectively. Memote is evaluated only for SBML-valid models. The Origo reference is taken from the original paper for the same sampled targets.

Removing MCP leaves only one SBML-valid model among two XML outputs and yields no growth- or production-competent model. Removing knowledge preserves the 10/10 validity, growth, and production rates but lowers mean Memote from 0.62 to 0.48, indicating a loss in curation quality. The single-agent and all-tools-per-task variants also reduce output coverage and aggregate quality. Together, these results show that tool access, knowledge-guided construction, and task organization contribute at different stages of producing an executable and curated network.

Manuscript changes: Appendix F adds the single-agent, no-MCP, no-knowledge, and all-tools-per-task comparisons on the shared target subset. Figures S8–S9 report their output coverage, aggregate metrics, and dimension-level assessments.

Thank you for raising the possibility of evaluator bias. We conducted an additional agreement study and invited two human raters to assess the same 90 generated models through a blind review:

Generating methodDeepSeek-V4-ProDeepSeek-V4.1-FlashKimi-K3Qwen3.8-MaxGPT-6 LunaGPT-6.1 SolBiology PhD StudentIndustry Expert
Origo7.207.006.806.605.705.208.207.20
Biomni6.206.606.705.804.204.205.405.30
STELLA2.402.302.302.101.501.601.901.60
GPT-5.53.303.504.103.302.602.603.102.80
DeepSeek-V4-Pro5.705.305.004.303.203.304.103.90
GLM-5.24.604.904.804.203.403.504.103.70
Qwen3.7-Max3.302.703.002.602.301.902.101.70
Kimi-K2.64.604.404.103.103.102.705.303.20
Gemini-3.5-Flash2.502.102.001.901.401.501.801.40

Overall quality on a 1–10 scale; means over ten models per method and judge. Bold entries mark column maxima.

The human raters and all additional LLM judges give Origo the highest mean overall-quality score. To examine agreement beyond method-level means, we also compared Spearman’s rank correlation between paired scores across all 90 models. The complete correlation matrix is below:

JudgeDeepSeek-V4-Pro (original)DeepSeek-V4.1-FlashKimi-K3Qwen3.8-Maxgpt-6-lunagpt-6.1-solBiology PhD StudentIndustry Expert
DeepSeek-V4-Pro (original)1.000.920.850.870.840.810.830.85
DeepSeek-V4.1-Flash0.921.000.900.910.880.840.850.90
Kimi-K30.850.901.000.910.840.810.820.87
Qwen3.8-Max0.870.910.911.000.870.830.830.90
gpt-6-luna0.840.880.840.871.000.820.840.87
gpt-6.1-sol0.810.840.810.830.821.000.810.85
Biology PhD Student0.830.850.820.830.840.811.000.88
Industry Expert0.850.900.870.900.870.850.881.00

The backbone LLM use in this study ratings correlate 0.828 and 0.848 with the human scores; the five additional LLM judges have human correlations of 0.813 to 0.896. Pairwise correlations between LLM judges range from 0.809 to 0.920. These results support consistent rankings across evaluators.

Manuscript changes: Appendix H specifies the human assessment and the additional LLM judges. Figures S11–S12 report their scores; Figure S13 and the accompanying text report agreement with the original evaluator and between human raters.

Thank you for your response. Some of my questions have been answered, but some still remain unclear:

(Q1/Q2) Thank you for the clarification and new comparisons. The single-agent, no-MCP, no-knowledge, and all-tools/task comparison is helpful. These comparisons help isolate what each part of the framework contributes, but I’m still not convinced of the contribution. Although it increases performance, I’m not very surprised that adding tools specialized for the task and defining when the agent should use the tool increases performance. Additionally, many of the reported columns are measures of technical correctness rather than biological validation, and the MCPs seem to be doing most of the heavy lifting regardless.

(Q3/4) The human evaluation is helpful, and the human evaluation tracks the LLM evaluation with reasonable correlation. However, it is unclear how the human evaluation was conducted. What was provided to the humans? Additionally, although the human evaluation was good on DLBench, there is not convincing evidence that high scores on DLBench equate to biologically correct. Specifically, the humans evaluated the systems based on DLBench’s criteria, but it is not shown that DLBench’s criteria corresponds to biologically accurate systems. The overall score D_8 is approximately equal to the average of central metabolism completeness (D_2), model systematicity (D_3), and biological plausibility (D_4), which may not cover everything necessary for a biologically accurate system. With just these criterion, it seems likely that some biologically inaccurate systems could score well. As asked for in my previous review, it would be helpful to see how systems (both good and bad) that have been separately verified by experts to be biologically correct perform on DLBench.

I choose to maintain my rating, but increase soundness. Overall, it is unclear if the main contribution is as an applications paper aiming to generate biologically meaningful systems, or as a methodological paper with the multi-agent framework. I’m unconvinced that DLBench is biologically accurate and that the multi-agent framework is a methodological contribution, so neither of these directions are fully supported.

Reviewer_MLaXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2(+1)
Read review
Paper summary

This paper introduces Origo, a multi-agent system for automating the design of microorganisms capable of producing target molecules. It combines GEM generation with enzyme candidate identification using specialized computational tools. The authors also introduce DLBench to systematically evaluate the generated results. Experiments show that Origo outperforms general-purpose LLMs and biomedical agents on most benchmark metrics.

Critical strengths
  1. The paper presents a domain-specific multi-agent architecture that combines expert-defined metabolic design steps, specialized tools, and iterative validation. This approach offers a promising way to automate complex scientific workflows and demonstrates improvements over general-purpose LLMs and the selected biomedical agents.
  2. DLBench evaluates metabolic design results at both the network and enzyme levels using specialized computational checks. A consistent evaluation protocol allows for more objective and reproducible comparisons across systems.
Critical weaknesses
  1. The proposed concept of digital microbial life seems broader than what the system actually models, which is primarily metabolic functionality. The practical applicability of Origo also remains insufficiently demonstrated. GEM generation and enzyme mining are evaluated separately, and no independent assessment of complete microbial designs by domain experts is reported. As a result, it remains unclear whether the proposed metabolic pathways, enzymes, and host organisms are compatible and suitable for experimental implementation.
  2. The evaluation lacks comparisons with specialized metabolic design tools and ablation studies. This makes it difficult to determine where Origo's improvements come from and how much each component contributes, including the multi-agent architecture, expert-defined templates, and specialized tools.

Minor Comments:

  1. Figure S1 appears to have an incorrect caption.
  2. Table S7 has formatting issues that should be corrected.
  3. The label "Figure 2" is repeated in the caption of Figure 2.
Questions for authors
  1. Could the authors provide an end-to-end evaluation of at least a few generated designs, ideally involving independent experts, to assess their potential feasibility for experimental implementation?
  2. How does Origo select a host organism and verify that the generated metabolic network is consistent with its biological capabilities?
  3. The practical motivation for the target-driven design approach could be clarified. What are the advantages of generating GEMs de novo rather than integrating target-specific biosynthetic pathways into existing host models, as demonstrated in SubNetX (10.1038/s41467-025-59827-7)?
  4. How is the enzyme mining query (q) derived from the reactions in a generated GEM? What happens if no reference enzyme is known for a required reaction?
  5. The Related Work section would benefit from a discussion of more closely related studies, including ChatGEM (10.64898/2026.07.20.739662), MetaGEM (10.48550/arXiv.2605.14812), D2Cell (10.1016/j.tibtech.2026.03.026), SubNetX (10.1038/s41467-025-59827-7), EnzRetro (10.1002/exp2.70129). It would also be useful to include comparisons with relevant specialized methods where feasible, or explain why they were not selected as baselines.
  6. How necessary is the multi-agent architecture when the workflow is largely predefined? An ablation study would help clarify the contributions of agent coordination, expert-defined templates, and task-specific tool restrictions.
  7. When multiple gap-filling solutions are possible, how does GEM Repair decide which reactions to add? How is the biological plausibility of these additions ensured?
  8. Have the authors assessed the agreement between the LLM judge's scores and evaluations by independent experts or other LLMs? Such a comparison would help establish the reliability of the LLM-based evaluation in DLBench.
  9. Why was DeepSeek V4 Flash selected as the backbone model? It would be useful to understand how sensitive Origo's performance is to the choice of the underlying LLM.
12 comments

We sincerely appreciate the care with which you examined the biological scope of Origo and the relationship between its two workflows. Your questions about host assumptions, the construction of enzyme queries, and the distinction between model generation and biological implementation concern details that readers need in order to judge the work. We recognize that these connections deserved a clearer explanation in our submission.We have revisited those explanations, added a comparison with the related systems you recommended, and included multiple LLM & blind human assessment of the generated models.

Origo addresses the metabolic modeling layer of digital microbial life. Its central contribution is the self-organized construction of a system-level SBML network from pathways and reaction building blocks within an expert-defined task framework. And evaluate, through de novo generation, the extent to which artificial agent systems can understand and reconstruct metabolic networks, thereby providing a theoretical exploratory attempt for digital cell factories and virtual cells.In this process, starting without a host GEM makes the reconstruction of supporting metabolism part of the problem. Appendix E explains how this setting differs from existing metabolic engineering workflows.

We added 5 different LLM models and 2 blind human assessment of 90 GEM generated: nine generating methods on the same ten targets. A biology PhD student and a rater with industry experience each scored all eight dimensions, yielding 630 more model reviews.

Generating methodDeepSeek-V4-ProDeepSeek-V4.1-FlashKimi-K3Qwen3.8-MaxGPT-6 LunaGPT-6.1 SolBiology PhD StudentIndustry Expert
Origo7.207.006.806.605.705.208.207.20
Biomni6.206.606.705.804.204.205.405.30
STELLA2.402.302.302.101.501.601.901.60
GPT-5.53.303.504.103.302.602.603.102.80
DeepSeek-V4-Pro5.705.305.004.303.203.304.103.90
GLM-5.24.604.904.804.203.403.504.103.70
Qwen3.7-Max3.302.703.002.602.301.902.101.70
Kimi-K2.64.604.404.103.103.102.705.303.20
Gemini-3.5-Flash2.502.102.001.901.401.501.801.40

Overall quality on a 1–10 scale; means over ten models per method and judge. Bold entries mark column maxima. DeepSeek-V4-Pro contains the archived ratings.

The agreement analysis supports the reliability of the rubric-based evaluation. Across the 90 models evaluated, the two human raters have Spearman ρ=0.875\rho=0.875. The LLM judges also agree closely with one another, with pairwise Spearman correlations of 0.8090.809--0.9200.920, indicating that the relative ranking of generated models is stable across judge models. Appendix H reports the complete correlation matrix, dimension-level results.

Section 3.1 now clarifies that GEM generation and enzyme mining are connected through the reaction steps in the target-product design. For each GEM example, domain experts manually select key enzyme-catalyzed reactions in the target pathway and identify a known reference enzyme; these selected enzymes constitute the objects evaluated in the enzyme-mining track. We evaluate the two components separately in the present study so that network construction and enzyme candidate mining can each be characterized clearly with task-specific metrics. Larger-scale integrated validation of host–pathway–enzyme compatibility is planned as future work.

Manuscript changes: Section 3.1 now explains the expert-mediated connection between GEM generation and enzyme mining. Appendix E defines the de novo construction setting and compares it with related methods (Table S8). Appendix H adds the blind human assessment, dimension-level scores, and judge agreement (Figures S11–S13).

Thank you for raising this important point. Origo is primarily a system for de novo generation of a digital metabolic network from a target molecule and biochemical pathway building blocks; its core construction procedure is therefore not restricted to a particular chassis. This setting provides a computational test of how far an agent system can understand, organize, and reconstruct biological metabolic networks, and offers exploratory infrastructure validation for future digital cell factories and virtual cells.

We establish this digital usability through SBML validation, flux balance analysis (FBA), growth testing, product-flux analysis, and model-quality assessment. Thermodynamic consistency, kinetic feasibility, and strain-specific compatibility require additional constraints and organism-specific evidence; these are goals for the next stage of the work and are already identified in the Limitations and Future Work section.

Comments on the paper

1 comment

Hi reviewers,

Thank you for your reviewing service. The authors have provided fairly extensive rebuttals. Please confirm that you have read them, state which of your concerns are or are not addressed, and whether that changes your recommendations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.