acceptodds
Under review as a conference paper at ICLR 2027

GEM-Reasoner: Protein Function Reasoning for Genome-Scale Metabolic Model Reconstruction

Abstract

Genome-scale metabolic models (GEMs) provide a computational foundation for systems biology and rational chassis design. Existing methods for automatically reconstructing GEMs from genome sequences typically transfer functional annotations from reference proteins based on sequence or protein-embedding similarity. However, these similarity signals do not reliably reflect functional equivalence, particularly under remote homology, functional divergence, or differences in substrate specificity, potentially resulting in erroneous or missing reaction annotations. To address this limitation, we propose GEM-Reasoner, replacing similarity-driven functional transfer with explicit reasoning about functional equivalence. To support this reasoning when retrieved annotations are incomplete or conflicting, we develop CAPA-RAG, a conflict-aware retrieval-augmented method that uses conflicts across functional dimensions to guide iterative retrieval and evidence integration. GEM-Reasoner uses CAPA-RAG to construct a structured functional profile for each protein and then applies a large language model to assess functional equivalence from paired profiles. We further construct the Protein Function Similarity (PFS) Benchmark, comprising PFS-Core and PFS-Challenge; PFS-Challenge pairs low-similarity positives with high-similarity negatives. On PFS-Challenge, GEM-Reasoner improves MCC by 0.168 over CLEAN, the strongest baseline. In GEM reconstruction for three bacterial species, GEM-Reasoner achieves the highest macro-average F1 for both substrate utilization and gene essentiality prediction. These results suggest that explicit functional reasoning can improve automated GEM reconstruction. Our code is available at https://anonymous.4open.science/r/GEM-Reasoner-825F.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
large language model reasoning, retrieval-augmented generation, protein function prediction, genome-scale metabolic reconstruction

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_uk5MOpenReview ↗
Rating
1(−1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GEM-Reasoner, a framework for genome-scale metabolic model reconstruction that combines retrieval-augmented generation (RAG) with LLM-based protein function reasoning. The method retrieves multidimensional functional evidence and uses conflict detection to guide iterative retrieval. An LLM then assesses functional equivalence between proteins to support metabolic network reconstruction.

Critical strengths

The paper addresses the limitations of similarity-based methods in identifying protein functional relationships under remote homology, functional divergence, and differences in substrate specificity. By combining RAG with LLM-based reasoning, the proposed framework introduces a conflict-aware retrieval mechanism to improve protein function annotation for genome-scale metabolic model reconstruction.

The framework is clearly presented, and the experiments cover both protein functional similarity prediction and downstream metabolic model evaluation. The results show substantial improvements over existing methods, particularly on challenging protein pairs.

Critical weaknesses

The main weakness of this paper is its limited methodological novelty. The proposed framework mainly combines established techniques, including protein function retrieval, conflict detection, iterative retrieval-augmented generation, and LLM-based reasoning. Although CAPA-RAG introduces conflict-guided retrieval, its technical contribution beyond existing retrieval and reasoning frameworks remains limited. Moreover, the GEM reconstruction stage directly follows the CarveMe pipeline, replacing sequence similarity scores with functional similarity scores.

Another concern is whether the performance improvements can be attributed to the proposed conflict-aware retrieval mechanism. Although the ablation experiments demonstrate the effectiveness of CAPA-RAG, the paper does not compare it with alternative retrieval strategies under comparable evidence and computational budgets. Therefore, it remains unclear whether the improvements mainly come from conflict-aware retrieval or from the additional functional evidence and LLM processing.

Finally, the proposed framework may suffer from high computational costs. GEM-Reasoner involves multiple LLM agents, iterative retrieval, and pairwise functional reasoning, which could introduce substantial inference overhead, particularly for genome-scale reconstruction. The computational efficiency and practical scalability of the framework require further investigation.

Questions for authors
  1. What are the main technical innovations of CAPA-RAG beyond existing retrieval and LLM-based reasoning frameworks?

  2. How can the authors distinguish the contribution of conflict-aware retrieval from the benefits of additional functional evidence and LLM processing?

  3. What are the computational costs and practical scalability of GEM-Reasoner compared with existing GEM reconstruction methods?

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

The authors proposed GEM-Reasoner, a framework for automated GEM reconstruction where an LLM reasons about the functional equivalence between a pair of proteins instead of relying on sequence similarity as other tools/frameworks.

GEM-Reasoner starts by picking the top five proteins, similar in sequence for a query protein. Then, these go to CAPA-RAG a conflict aware retrieval augmented method built by them. It uses a model called ProTrek to retrieve different type of information about the query protein across 14 dimensions (EC number, catalytic activity, etc.). Then it uses LLM agents, to detect cross-dimensional conflicts, and suggest an order while retaining all information. A final LLM compares the two profiles and outputs a similarity level, a score and a rationale. At the end, the GEM is reconstructed using CarveMe pipeline (a top-down tool for GEM reconstruction) using GEM-Reasoner scores instead of sequence similarity, the rest of the pipeline is unchanged (GPR based score aggregation and the MILP).

They also introduced the PFS benchmark, built with Swiss-Prot, where pairs of proteins are labeled as functionally equivalent if they share the full EC number and the opposite if they differ in at least two EC fields. PFS-Core uses randomly sampled pairs and PFS-Challenge selects low-similarity positives and high-similarity negatives. This benchmark was used to validate GEM-Reasoner against strong baselines. On the other hand, GEMs reconstructed for three bacteria were compared with CarveMe, ModelSEED and AlphaGEM on substrate utilization and gene essentiality.

Critical strengths

1- Transparent and reproducible because the full prompts, the code and its settings are provided. They also did ablations and backbone tests, and mentioned the limits of the method.

2- Explainable model: the reasoner LLM provides justification of the predictions.

Critical weaknesses

1- Hard to read: heavy notation for simple operations (Eq. 4) (View_m, P_m, not defined). Agents are used before they are introduced and key settings only found in the appendix.

2- Limited evaluation: substrate specificity is excluded from the evaluation (pairs differing in exactly one EC field), which is exactly the “differences in substrate specificity” problem the authors uses to motivate the work. In addition, baseline thresholds are tuned on an unknown validation set (I don't know how it was created) and transferred to PFS-Benchmark, where I don't know if the similarity distributions are different, while GEM-Reasoner’s fixed s > 0.5 without justification.

Questions for authors

1- By picking the top five similar proteins, dissimilar proteins (remote homologs) in sequence may not reach the reasoning step in the GEM. Why did you choose five proteins? Why did you choose DIAMOND over BLAST or FoldSeek?

2- ProTrek is trained on Swiss-Prot/TrEMBL. Did you check that the proteins on PFS-Challenge were in TrEMBL?

3- What is the validation set used to find the thresholds that are used to make predictions and evaluate the tools?

4- Why the thresholds for the baselines are not provided?

5- Why the substrate specificity (ΔEC = 1) is excluded from the evaluation?

6- Which set of proteins were used as base in the GEM reconstruction? How was AlphaGEM ran for E. coli?

7- If the PFS benchmark evaluates the similarity among two proteins based on the EC number why you did not compare CLEAN performance using their own predictions? For example, two proteins with at least one same predicted EC number could be a positive.

Reviewer_JZx3OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

GEM-Reasoner combines DIAMOND candidate retrieval, ProTrek functional-text retrieval, and language-model comparison of protein profiles to assess whether a query and reference protein share an enzyme activity. These scores replace sequence-similarity scores in CarveMe and guide reaction selection. The paper evaluates protein-pair classification on PFS-Core and PFS-Challenge, followed by substrate-utilization and gene-essentiality prediction in three bacterial strains. The authors report the highest macro-average F1 on both downstream tasks. A proline-utilization case study further examines pathway content and model-derived fluxes.

Critical strengths

The evaluation goes beyond protein-pair classification. Comparing reconstructed models against experimental phenotype labels provides useful evidence that changing functional-transfer scores can improve a downstream metabolic-modeling task.

Critical weaknesses

1.The functional-matching criterion does not fully justify the subsequent reaction transfer. Section 3.1 considers two proteins a match if they share at least one activity at full four-field Enzyme Commission (EC) specificity. Yet Equation D.6 in Appendix D.2 propagates that score to every reaction associated with the reference protein. A correct protein-pair match can therefore support reactions for which no shared activity has been established. The authors acknowledge this distinction but do not quantify its effect on the reconstructed networks. The proline example in Appendix F.4.5 illustrates why this matters: PRO1x and the PutB annotation differ in cofactor specificity. Improved phenotype prediction does not establish that the transferred reactions are biochemically correct.

2.The ablation does not clearly separate conflict handling from access to additional annotations. Although the “w/o CAPA-RAG” variant retains functional retrieval, Appendix E.1 does not specify whether it receives the same expanded candidates as the full system. If it only sees the initial candidates, some of the reported gain could come from finding a correct annotation rather than resolving conflicts more effectively. The CLEAN and ProTrek comparisons do not settle this question either: they use representation similarity, whereas GEM-Reasoner receives retrieved functional descriptions. Using the same candidates and final language model, while removing only the conflict assessments and ranking suggestions, would give a more direct test of the claimed contribution.

3.For GEM reconstruction, the paper does not clearly state whether the DIAMOND reference collection excludes query proteins or full-length sequence-identical records. Separately, the Swiss-Prot temporal filter does not exclude earlier exposure through ProTrek’s TrEMBL training data, as Appendix A.4 acknowledges. These uncertainties limit conclusions about previously unseen proteins; neither establishes leakage, and the presence of correct EC descriptions in a shared functional vocabulary is not itself leakage.

4.Negative records require EC numbers differing in at least two fields, while records differing in exactly one field are excluded. This leaves important distinctions between closely related activities, including categories differing only in the fourth field, untested by that construction. The results therefore do not directly establish the fine-grained substrate-specificity discrimination emphasized in the motivation. (Appendix A.2.)

Questions for authors

1.How many retained reactions inherit scores from multi-activity reference proteins, and can the existing records show whether the shared-activity evidence supports each specific transferred reaction?

2.Does “w/o CAPA-RAG” receive the initial candidates or the same expanded candidates as the full method? Are results available with identical candidates and final-model inputs, differing only in conflict-processing information?

3.Does the DIAMOND reference collection exclude query proteins and full-length sequence-identical records? Please report the overlap and distinguish performance on already-covered proteins from queries without an exact reference match.

4.Is overlap between PFS-Challenge and ProTrek’s TrEMBL training data checkable? If so, what overlap is observed, and which cases remain of unknown training membership?

5.Given the exclusion of one-field EC differences, which claims about substrate-specificity discrimination are directly supported by the current benchmark, and which should be restricted?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.