acceptodds
Under review as a conference paper at ICLR 2027

Modeling ecological dynamics via multi-agent systems of microbial metabolism

Abstract

Metabolic interactions govern the assembly and dynamics of microbial communities. These interactions are studied computationally through genome-scale metabolic reconstructions in which an organism's biochemical reaction network is assembled from annotated genomes and curated databases. The capabilities of each organism are then predicted by optimizing flux through the resulting metabolic network using flux balance analysis (FBA) over a particular phenotype (growth, for example). In this work, we asked whether a language model can replicate the predictions of a genome-scale metabolic model, with each organism acting as an agent that chooses its exchange with its environment, which represented by the model context. We first show that prompted single microbial agents were able to recover key metabolic capabilities quantitatively, albeit at a lesser precision than dynamic flux balance analysis (dFBA), over multiple query rounds. Subsequently, by fine-tuning an open-source model on flux envelopes drawn from 897 reconstructions, microbial agents outperformed metabolic models in predicting metabolite concentrations. Our analysis showed that fine-tuning produces valid flux distributions that respect the underlying stoichiometry and effectively turned the language model into a reliable solver of metabolic fluxes under constraints. Then, by scaling to bacterial communities, multi-agent systems correctly captured competition as the main driver of interactions in communities, and correctly modeled smooth network dynamics that transition through building, converging, and stabilizing of the community structure over time, which DyMMM (Dynamic Multispecies Metabolic Modeling), a framework based on dFBA, fails to capture. To do so, we built an observability and evaluation system that was derived from ecological theory and adapted to the multi-agent model setting. Taken together, we present a new framework for modeling microbial communities relevant to bioengineering and biomedicine by using multi-agent large language model systems constructed with domain-specific harnesses. We also propose fine-tuning on metabolic knowledgebases as a strategy to build accurate modeling tools for the prediction of intervention effects in microbial communities and for the design of synthetic communities.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
large language model agents, multi-agent systems, AI for science, computational biology, constrained optimization

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

The authors focus on using multiple language model agents to replicate the predictions of a genome-scale metabolic model. They propose metabolic knowledge fine-tuning with domain-specific harnesses to model microbial communities. First, they explored fine-tuning of language models on FVA/pFBA outputs to give valid flux distributions adhering to the biochemistry. They then studied multi-agent systems scaled to bacterial communities to capture interactions within the community. They claim that this multi-agent approach correctly modelled the network dynamics, whereas dynamic multispecies metabolic modelling failed in comparison. However, my main concerns are that the writing tends to overclaim where experiments do not cover these claims, in particular, that language models can reproduce the state of an organism and can act as linear program solvers.

Critical strengths

The authors target community metabolic modelling, which is a relevant problem, and evaluate against experimental rather than simulated data.

The paper contains very detailed procedures for constructing the benchmark for reproducibility and makes a great effort in describing the setups of the metabolic models.

Critical weaknesses
  1. The framework mainly uses LLM agents with prompting, and LoRA fine-tuned on FBA-based simulation outputs. The methodological novelty is limited to the combination of various components.

  2. The main claim is not tested directly. The authors claim that LLMs are trained on literature, so they carry the information needed to predict the state of microorganisms. However, the results in 4.1.1 and Table 1 show that prompted LLMs do not outperform FBA. This may be used to argue that LLMs cannot predict the metabolic state.

  3. Using LoRA is not training the model to solve a linear program, and current results have not tested this. Finetuning likely imitates the solver outputs, but this is not equivalent to solving the LP. This requires a dedicated experimental section to validate if scoped this way.

Questions for authors
  1. How are the growth fold and metabolites log10 measured in Table 1?

  2. The validator is relevant for many results but is not specified. How is the validator implemented exactly?

3 comments

We thank the reviewer for a careful reading, for their critical feedback to improve our work, and for the opportunity to respond to those critiques. We are glad that the reviewer noted critical strengths in the "high relevance of the problem", the "evaluation against real data", and the "great effort in describing the setups". The reviewer has noted substantial weaknesses that we addressed in the point-by-point response below and in the updated manuscript. We believe these weaknesses relate to supplementary figures and discussion points, and not to core results.

W1, novelty. To our knowledge, no prior work has evaluated LLMs as models of microbial communities nor modeled these communities using agents given bacterial personas. The closest work, Eubiota (Lu et al., 2026), is a biomedical AI agent that acts as a co-scientist with knowledge about gut bacteria, whereas we model the bacteria directly as agents. No prior work has shown that fine-tuning an LLM on a metabolic graph is feasible and provides accuracy higher than either the metabolic graph or the LLM (Table 1), which indicates learned emergent knowledge. We do agree, however, that the basic recipe (LLM + LoRA fine-tuning) is standard, although a novel recipe is not a requirement for the applications to physical sciences track. Our work has generated new knowledge and novel insights into a pertinent basic ecological problem using LLMs.

W2, LLMs are less accurate than metabolic networks. We first tested the ability of prompted LLMs to recover known capabilities quantitatively, and they do so at lower precision than metabolic networks (Table 1). We also show in a new figure (Figure S1) that this is directly proportional to LLM model size, where larger models do better in predicting network fluxes for more common and better-studied organisms (E. coli, yeast, P. putida), reflecting a knowledge bias. All this preliminary data was what drove our work to the next phase, which was fine-tuning LLMs for microbial species. We used this fact as a ramp to come up with a fine-tuning strategy, which then showed that LLMs fine-tuned on metabolic networks now have higher performance than metabolic networks and than prompted LLMs (Table 1 and Table S2).

W3, Fine-tuning does not solve a linear program. We agree and withdraw the claim. This was a discussion point and not a core result. It was merely a hypothesis we came up with to explain the results and model behavior, given that the admissible fraction of LLM-proposed fluxes rose from 0.868 to 0.928 with fine-tuning (Table S5), and given prior work by Nair et al. (2020) solving MILPs using deep neural networks. The revised manuscript now states only that the fine-tuned fluxes are highly consistent with the linear program results.

Q1, growth fold and metabolite errors. These are measurements used as ground truth data for testing, from Clark et al. (2021). Growth fold is each species' final biomass after 48 hours of culture, and the metabolite error is the absolute log10 of predicted over measured concentration in mM for butyrate, acetate, lactate and succinate, averaged over the four. We added the formulas for computing them (Equations 12 and 13) to Appendix C.2, where these results are reported in full (Table S2).

Q2, the validator. A validator is a set of functions that verify that the proposed LLM flux distribution respects a pre-specified output format (a JSON object with a specified format) and that the proposed fluxes respect mass balance (no uptake fluxes are larger than what is available in the medium). It is a basic check step that is required so that downstream analyses, including dynamic simulation, can run consistently. We included a description of the validator in the manuscript (Section 3.4), and Table 1 reports its outcome for each model (rows "mass balanced fluxes" and "unparsable LLM outputs").

Ethics flag. The footnote that the reviewer flagged was added automatically by the ICLR submission system to all submissions. Any other submission has it as well.

I appreciate the effort the authors have put into their response and revision. My main concern is not addressed. I feel the paper's idea is good but needs to justify the relevance of LLMs.

The main claim is not tested directly. The authors claim that LLMs are trained on literature, so they carry the information needed to predict the state of microorganisms. However, the results in 4.1.1 and Table 1 show that prompted LLMs do not outperform FBA. This may be used to argue that LLMs cannot predict the metabolic state.

Prompted LLMs lose to FBA and outperform it only after fine-tuning. This shows the useful signal actually comes from the metabolic models. Are LLMs just acting as surrogates for FBA? What properties of LLMs make them better surrogates than the metabolic models themselves? The language modeling, more parameters, or the architecture? For example, a non-language surrogate trained on the same corpus, as a control, could show what LLMs bring.

Thank you @RbGd for continuing the conversation. Your initial comment was "LLMs cannot predict the metabolic state." which we clarified by stressing that fine-tuning increased performance beyond that of metabolic networks. Then your follow-up comment was "the useful signal actually comes from the metabolic models", which we agree with; our hypothesis is that fine-tuning allows a form of synergy between modalities, that's why it outperforms both baselines (LLMs+Metabolic models). For example, on single-agent models, fine-tuning allowed LLMs to improve prediction on lactate and butyrate (fermentation pathways).

What properties of LLMs make them better surrogates than the metabolic models themselves? Our initial hypothesis was that LLMs had learned new capacities by solving mixed-integer linear programs for metabolic models which was shown by the increase in feasible solutions proposed by LLMs, but we removed that claim at the request of the reviewer. A plausible hypothesis hinted at by the reviewer is language modeling; the LLMs we tested have reasoning capacity which might be tied to increase in performance.

As requested, we ran a non-LLM baseline on the same corpus, by training an artificial neural network, which receives the same information as the prompt (media composition, species names) as numerical input and was run in the same simulation as the fine-tuned LLM. This setting showed that there is signal that comes from metabolic models for the species in the training set (Table1: accuracy of 0.72/F1 score of 0.62). However, for species and media outside the corpus, the MLP fails (error on metabolites is 1.9 vs 0.6 for LLMs), while the LLM are able to take any medium as input even if not part of training, which brings a notion of transfer learning outside the training set.

The reviewer have also alluded on more parameters being more helpful to learning and our figure S1, shows that larger models are able to recapitulate metabolic knowledge better with an accuracy proportional to popularity of the microorganism (well-studied organisms are better predicted). The fine-tuned model used Llama-3.1-8B, which was less accurate than FBA without fine-tuning. This provides a natural control on the important components that led to the improvements in accuracy.

To sum up, on the reviewers’ question about the increase in performance of fine-tuned LLMs over metabolic models, we showed in the first section of the paper that metabolic models can be incomplete in a new medium where they haven’t been curated (they don’t have a notion of statistical learning). In fact, only 1/25 species grew on our test media DM38, and we had to manually debug those draft models to make them compliant with experiments, which then all 25/25 ended up growing on it, as expected. We mitigated these gaps during fine-tuning by generating a corpus that included 897 species grown on different media with ablated nutrients so the fine-tuned model does not depend on one single gappy reconstruction.

Finally, on the general question on why LLM perform well in recapitulating metabolic knowledge, there have been recent reports that state of the art models are being trained on biomedical literature and genomic sequences, which may explain why Haiku was the only model that recapitulated such a complex behavior as diauxic growth.

We’re glad to answer any other results the reviewer might have and we will upload the amended manuscript with new results shortly.

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

This paper proposes an LLM-based multi-agent framework for simulating microbial metabolism and ecological dynamics. The authors combine prompting, rule-based validation, and LoRA fine-tuning on FBA-generated data, evaluating the resulting models against experimental measurements.

Critical strengths

The paper explores an interesting biological application and includes comparisons with experimental measurements and conventional metabolic models.

Critical weaknesses
  1. Limited novelty. The framework primarily assembles existing prompting, validation, multi-agent simulation, and fine-tuning techniques. No meaningful methodological innovation or comparison against simpler learned predictors is demonstrated.

  2. Unsupported and inconsistent claims. Only 9.3% of the fine-tuned model's complete flux proposals are jointly feasible (Table 7), undermining the claim that it learns to solve constrained optimization problems. Furthermore, the main text claims Haiku 4.5 reproduces diauxic growth, whereas Appendix E reports "no switch reached."

  3. Weak empirical evidence. Predictive improvements are modest and inconsistent. The claimed ecological dynamics lack ground-truth temporal validation, making the conclusions regarding biological fidelity unconvincing.

Questions for authors
  1. How can the model be considered a reliable metabolic solver when only 9.3% of complete flux proposals are jointly feasible?

  2. How do the authors reconcile the contradictory diauxic growth results between the main text and Appendix E?

  3. What evidence establishes the advantage of LLM agents over simpler non-LLM surrogate models?

1 comment

We appreciate the reviewer's feedback for improving our work, and we are glad that the reviewer noted critical strengths in "an interesting biological application" and the "comparisons with experimental measurements and conventional metabolic models".

W1, novelty. Our contribution is a multi-agent harness that represents a graph of bacteria taking turns to share and compete for resources. To our knowledge, no prior work has attempted this exercise. Beyond the technical curiosity, this work has direct implications for the bioengineering of synthetic microbial communities for bioreactors, and for modeling the gut microbiome and diseases with disrupted bacterial communities (inflammatory bowel disease, colon cancer). It is highly relevant for reverse ecology, which models the community formation that leads to an observed ecological behavior. Since we started this work, we have laid out plans to expand it to spatial settings and to gradually increase complexity, in order to model bacterial communities relevant to health and industry.

W2 and Q1. To reiterate the setup of this experiment, we compared the flux values proposed by an LLM to those of a metabolic network model and found that the fluxes of the fine-tuned LLM respect the constraints of the metabolic network individually (92.8%) but not jointly (9.3%, Table S5). Our attempt to interpret these results was that LLMs learned how to solve a metabolic network, which is a linear program. We have removed this argument and clarified the result as a comparison between the LLM and the metabolic network. This does not challenge our core result, which is accuracy against ground truth data, on which our fine-tuned model was the most accurate (Table 1). The ground truth is not represented by a metabolic network model but rather by measurements of metabolite concentrations.

W2 and Q2. We agree and thank the reviewer for catching the discrepancy in the supplementary table. What matters for this result is the main figure (Figure 3), which clearly shows diauxic growth, with a switch visible in panel (c), where acetate accumulates while glucose is consumed and drops to zero after glucose is depleted. The Appendix E table was a stale version of an earlier experiment without calibrated parameters, and we updated it to match the main figure and text (now Table S3). Haiku 4.5 satisfies all criteria for diauxic growth, which importantly hints that model complexity is needed for more complex behaviors.

W3. We agree with the reviewer that the improvements are modest, but our proposed framework does much better than the state of the art in the field, which is represented by dynamic metabolic models (Table 2). We recognize that the problem is generally hard, which is what drew our attention to using LLMs to solve it. Higher prediction accuracy can be reached using statistical or black-box models, but optimizing for mechanistic interpretability costs accuracy. Interpretability is essential for our setting because it will allow the design of interventions.

On temporal validation, we agree with the reviewer, but temporal data is hard to obtain and measure, and the whole field of metabolic modeling is exactly an attempt to reverse engineer community formation using first principles and snapshot measurements. We show that community formation using our system is smooth (Figure 5), that it respects ecological constraints such as the dissimilarity-overlap relation (Figure 6), and that negative interactions shape the structure of the agent graphs (Figure 4), which are known phenomena from ecology that our model recovers.

Q3. We agree with the reviewer that simpler linear models or baselines can outperform our predictions of final community composition. But our object of study, and that of the whole field of metabolic modeling, is to provide mechanistic models that can nominate direct targets for interventions, which a statistical model alone cannot provide. So on general accuracy, we recognize that our approach is not the most accurate, but for the specific class of mechanistic models, we reiterate that our approach is novel and more accurate than state-of-the-art community models (Table 2). In addition, a statistical model depends on its training data and cannot model unseen species. By searching the literature, we found a model trained on the same communities as ours, which predicted final metabolite concentrations with correlations of 0.80 to 0.92 when the test species were in its training data, but 0.01 to 0.13 when all species in the test communities were unseen (Qian et al., 2026), which is exactly the setting our agents are evaluated in, since they do not see any measurement from these communities. We also added a new baseline called MICOM (Diener et al., 2020) as a second mechanistic community model. Here too, the fine-tuned agents were more accurate, with a rank correlation of 0.195 against 0.059 and a median log10 error of 0.631 against 1.482 (Table 2).

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

This paper proposes a multi-agent LLM framework for simulating microbial metabolism. Each species is represented by an agent that predicts growth and metabolite exchange from the current medium, while community agents interact through a shared metabolite pool. The authors further fine-tune Llama-3.1-8B on pFBA/FVA flux examples from 897 gut microbial models. Validation covers 25 gut species grown in DM38, two E. coli phenotypes, and multispecies communities. Predicted biomass and metabolite concentrations are compared with experimental measurements; dFBA and DyMMM serve as single-species and community baselines. Community simulations are also assessed through species abundances, metabolite levels, inferred interaction networks, and the dissimilarity-overlap relationship. Fine-tuning improves several endpoint metrics, with a mean Spearman correlation of 0.195 for community composition.

Critical strengths
  1. The idea of using one LLM agent per microbial species and coupling the agents through a shared medium is original. It gives the paper a clear modeling contribution and makes the simulated exchanges easier to inspect than a black-box endpoint predictor.

  2. The paper evaluates both monocultures and communities, including growth, metabolite endpoints, E. coli phenotypes, and ecological interaction patterns. Fine-tuning improves several results over prompted models, providing reasonable empirical support for the approach.

Critical weaknesses
  1. The comparison study is limited. The paper compares mainly with dFBA, DyMMM, and prompted LLMs, but does not include matched learned baselines or key ablations. It is therefore unclear whether the gains come from the LLM, the training data, or the iterative simulation setup.

  2. The evaluation protocol needs stronger justification. The paper combines standard measures with several task-specific criteria, including the butyrate call, clean mass balance, usable-answer count, and phenotype scorecards. The biological basis for these scoring rules is not fully established, and the paper does not show whether the model rankings remain stable under other reasonable definitions.

  3. The paper lacks an overview figure of the framework. A diagram showing the training pipeline, species agents, shared-medium update, and inference loop would make the method easier to follow and reproduce.

Questions for authors

Please see the weaknesses.

1 comment

We thank the reviewer for the constructive feedback and for the opportunity to improve our work. We are glad that the reviewer found that coupling the agents "through a shared medium is original", that it "gives the paper a clear modeling contribution", and that fine-tuning provides "reasonable empirical support for the approach". The reviewer has noted weaknesses that we address in the point-by-point response below and in the updated manuscript.

W1, comparison study and ablations. The reviewer requested a comparison that separates the contributions of the LLM, the training data, and the iterative simulation setup. On the training data, the fine-tuned agent and the prompted Llama-3.1-8B share the same base weights, so their difference isolates the effect of fine-tuning. Fine-tuning single strains allowed the growth rank correlation to increase from -0.030 to 0.325 and the metabolite error to drop from 1.021 to 0.748 (Table 1), and for communities the rank correlation to rise from 0.058 to 0.195 (Table 2). On baselines, we added MICOM (Diener et al., 2020) as a second mechanistic community model baseline. It had a rank correlation of 0.059 and a median log10 error of 1.482, against 0.195 and 0.631 for the fine-tuned agents (Table 2). In the literature, we found a statistical model trained on the same communities we evaluated on (Qian et al., 2026), which predicted final metabolite concentrations with correlations of 0.80 to 0.92 when the test species were in its training data, and 0.01 to 0.13 for three of the four metabolites when they were not, the latter being the setting our agents attempt to model by never seeing any measurement from these communities. More generally, higher accuracy can be reached with statistical models, but our object of study is mechanistic models that can nominate targets for interventions.

W2, evaluation protocol. We agree and we conducted additional analyses for the stability of the evaluation protocol. First, the growth fold error, the metabolite error and the growth rank compared predictions directly with measured optical density and metabolite concentrations (Equations 12 and 13). A higher importance was given to butyrate production which was included as well (Clark et al., 2021), to evaluate our model for identifying the butyrate producers among members of the community. Next, we evaluated mass balance in predicted fluxes and validated LLM outputs through validity checks on the output of the LLM (Section 3.4).

As an orthogonal evaluation, we also included phenotypic and ecological criteria which include known signatures of diauxic growth and overflow metabolism (Mahadevan et al., 2002; Basan et al., 2015) (Tables S3 and S4). These are emergent behaviors that were not encoded in the community and reflect a quality control on the design of communities.

To test stability, we recomputed the model rankings under alternative definitions of our performance metrics (Table S6). The ranking on metabolites did not change for any floor of the log error between 0.01 and 1 mM, the butyrate call ranking did not change for thresholds of 0.5 and 1 mM, and the fine-tuned agent ranked first on growth under every definition. Metabolite rankings were largely recapitulated under alternative metrics.

W3, overview figure. We added an overview figure (Figure 1) with three panels including the fine-tuning pipeline and the inference loop.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Second round of revisions

Dear all,

We addressed the remaining reviewer scomments, in a second round of revisions. The updated manuscript has additional experiments that compared fine-tuned LLM to a baseline Artificial Neural Networks (ANN), which clearly segregated the "language" and reasoning components in the observed performance. We summarized these results in table 1.

Recent reports in the field have indicated that frontier labs may be training their models directly on biological data, which may suggest why for example we see that Haiku is the only model that recapitulates diauxic growth which is a complex behavior, and further reinforces the timeliness of this report.

Let us know if there are any additional concerns!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.