acceptodds
Under review as a conference paper at ICLR 2027

Cella: Predicting Microbial Cultivation Requirements on a Multiscale Metabolic Graph

Abstract

Microbial growth is determined by interacting processes across a wide range of temporal and spatial scales, from molecular structure to phylogeny. Multiple modalities, including genomic data, genome-scale metabolic models, chemical structures, and media properties, often inform expert predictions. However, previous foundation models have been restricted to data available in large quantities and readily consumable by ML algorithms (PDB, sequences, genomeLMs). In this work, we present Cella, a heterograph model which enables predictions based on different data sources at scale while encoding real mechanistic relationships and handling coverage gaps gracefully. We fit the model on a newly constructed collection of 4,220 prokaryotic organisms and their associated metabolic models sampled exhaustively at the genus level. We evaluate Cella on various industrially relevant tasks, including media and component growth prediction, carbon-source utilization, environmental preference, and component substitution, and strain ranking by shared cultivation requirements. Tasks are scored under held-out organism and component splits, against popularity, phylum-only, and label-blind baselines that the adjacent literature omits. We predict growth media on the nearly half of species in our collection that do not record one, while providing a traceable path to the evidence supporting the predictions. Cella enables multimodal data integration while retaining interpretable interactions and points towards underlying biological patterns learnable with foundation-model-style approaches that generalize across the microbial landscape.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
heterogeneous graphs, relational graph neural networks, multimodal representation learning, metabolism, genome-scale metabolic models, microbiology, GSMMs, GEMs

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 2 reviews, above 27% of the others.
Reviewer_UU8UOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Cella, a multiscale metabolic heterogeneous graph model for predicting microbial cultivation requirements and environmental preferences. It places taxon, organism, metabolite, media component, and medium nodes in one graph, and uses relation-typed message passing to learn representations across genomes, metabolic models, chemical structures, phylogeny, and cultivation conditions. The main contributions are a dataset covering thousands of prokaryotic genera with GEMs reconstructed by three pipelines, a unified compound and reaction namespace, and a heterogeneous graph that handles missing modalities. Experiments compare against popularity, lineage-only, and genome-only baselines under organism-grouped and component held-out splits, evaluating media, component, carbon-source utilization, environmental preference, component substitution, and strain ranking tasks. Results show that Cella achieves the best macro F1 on media and component prediction, and can propose media for many genera without known cultivation records.

Critical strengths
  1. Large-scale, multimodal, cross-scale heterogeneous graph resource. The paper integrates genomes, GEMs from three metabolic reconstruction pipelines, media, components, metabolites, and taxa, and unifies compound and reaction namespaces across many prokaryotic genera. The resource has independent value as a benchmark for microbial phenotype prediction and naturally handles missing modalities, rather than only serving the proposed model.

  2. Rigorous leakage-aware evaluation protocol. The paper uses organism-grouped, component held-out, and out-of-family splits, rebuilds the graph within each split to remove growth relations of test organisms, and adopts degree-matched negatives with popularity and lineage baselines. Cultivation prediction is easily inflated by homology leakage and popular-component bias, and this protocol makes the reported results credible even when gains are modest.

Critical weaknesses
  1. The core performance comparison uses inconsistent evaluation protocols, making the main conclusion unreliable. The footnote in Appendix Table S5 states that Cella was scored on its own protocol and is not directly comparable with the feature baselines above it, yet the main text still places Cella alongside those baselines and uses the table to claim best or comparable performance. If the footnote is correct, all horizontal performance claims lose their basis. If the footnote is wrong, the appendix is misleading. The central empirical claim therefore depends on resolving this inconsistency.

  2. The contribution of the heterogeneous graph structure is not established, and simple baselines are often as good or better. The paper's main selling point is a multiscale metabolic heterogeneous graph, but it lacks key ablations, especially removing the entire GEM layer, metabolite nodes, and their relations while keeping only taxon, organism, component, and medium. Existing results show that simple genome-feature or feature-concatenation baselines are already competitive on multiple tasks, and the authors acknowledge that the genome block supplies most of the learnable signal. The independent gain from graph message passing and the metabolic layer is therefore not established.

Questions for authors
  1. What exactly does the Table S5 footnote mean by "its own protocol"? How do Cella and the feature baselines differ in label sets, candidate space, or threshold fitting? If such differences exist, how should the horizontal comparison in Table 2 be interpreted?

  2. Have you run an ablation that removes the entire GEM layer, i.e., dropping metabolite nodes and requires/asserts relations while keeping only taxon, organism, component, and medium? If not, how do you establish that the multiscale metabolic heterogeneous graph itself, rather than additional input features, provides the performance gain?

  3. Are the medium proposals for genera without cultivation records independently validated, for example through temporal holdouts, external data, or wet-lab experiments? If not, can you at least report sample sizes, performance metrics, and confidence intervals by phylogenetic distance bin to support the extrapolation claim?

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

Cella is a heterogeneous graph model for predicting microbial cultivation requirements. The authors take one quality-controlled isolate genome per named GTDB genus, giving 4,220 prokaryotes, reconstruct three genome-scale metabolic models (GEMs) per genome with CarveMe, ModelSEED and gapseq, and reconcile them with BacDive and metaTraits traits and MediaDive recipes. Taxa, organisms, metabolites, components and media form a 7,377-node graph in which GEM-derived relations are typed by how many reconstructions assert them, and a two-layer relational GNN with 407k parameters encodes it. The evaluation spans three nutrition tasks cast as link prediction, nine environment-tolerance tasks and three applications, against popularity, lineage-only and random-forest baselines on GenomeSPOT and GEM features. Cella reports the best macro F1 on media components (0.502) and cultivation media (0.351) with AUROC level with or below the best concatenation arm (Table 2), and the paper further analyses cross-pipeline GEM agreement and proposes media for 1,973 genera without records.

Critical strengths
  • (S1) The resource is substantial and useful in its own right. Building three GEMs for each of 4,220 genera enables cross-pipeline comparison at scale, and Sec. 3.3 turns this into a practical finding, since only 130 of 808 compounds are covered by all three tools while 82.8% of organism-compound pairs agree on that shared vocabulary. Provenance is tracked per record in Table S2, predicted-only traits are deliberately excluded as targets to avoid circularity, and ion-aware curation resolves 409 of 627 components where name matching alone yields 6 edges.

  • (S2) The evaluation is built to resist leakage, and the reporting is candid. Rebuilding the graph per split without grows_in edges for held-out organisms, out-of-family splits for environment tasks, degree-matched negatives and a value-only control for envelope tasks all address real pitfalls (Sec. 2.3–2.4). The authors also acknowledge that the genome block supplies most of the learnable signal and that the substitution evaluation is confounded (Table S7).

  • (S3) Typing GEM relations by tool agreement is a clean way to encode reconstruction uncertainty. Instead of averaging or voting across pipelines, the model keeps agreement as a relation type and learns its weight through per-relation transforms, while structurally undefined mixtures such as peptone enter only through graph position with a null chemistry vector. The encoder is small and trains full-batch, so the approach is cheap to reproduce.

Critical weaknesses
  • (W1) The headline table compares arms scored under different protocols, so the central claim is not yet supported. Table S5 states that the Cella row is "scored on its own protocol, so its row is not directly comparable with the feature arms above it", yet Table 2 shows it beside those arms without the caveat and Sec. 3.1 draws the best-F1 claim from it. The protocol is also described inconsistently, with an 80/20 split over three seeds in Sec. 2.3, 5% held out over five folds in Sec. 2.4 and five seeds in Table 2, and the thresholding rule used for Cella is never stated.

  • (W2) Against the strongest baseline, plain genome features, the graph adds little or loses, while the main text relies on weaker controls. In Table S6, Cella GNN trails the GenomeSPOT random forest on the pH and salinity optima, oxygen tolerance and 12-class habitat AUROC, and its pH-envelope AUROC of 0.586 is the lowest of all arms, below even the value-only control at 0.614. Temperature optimum is the only clear win, at 0.819 versus 0.787. On carbon sources Cella falls below the lineage-only control in both F1 and AUROC, and the statement in Sec. 3.2 that multimodal prediction "beats all baselines on every task" holds only for AUROC, since the all-modalities arm is below genome-only in F1 on all three nutrition tasks. Sec. 3.1 and the abstract nonetheless report comparisons only against popularity and lineage-only.

  • (W3) The baselines and ablations needed to attribute any gain are missing. There is no graph-learning baseline, such as an R-GCN without provenance typing, HGT or a DistMult embedding of the same graph, and no comparison with Canopy, so the effect of graph construction cannot be separated from that of the encoder. Domain and mechanistic baselines are also absent, including MediaMatch and phylogeny-based collaborative filtering (Oberhardt et al., 2015). This matters because all three GEMs are gap-filled on the gapseq-predicted medium, which is itself the output of a medium predictor, so the GEM edges partly encode an existing prediction and the reconstructions share one nutritional prior that may also inflate the 82.8% agreement. Finally, no ablation is reported for the agreement relations, the virtual node, the metabolite layer or the MolFormer features, the abstract's claim of "handling coverage gaps gracefully" is untested, and the held-out component splits described in Sec. 2.4 have no results.

  • (W4) Several application claims go beyond the evidence. The substitution scorer is validated by the share of near-identical proposals, but Sec. 2.2 maps hydrates of a salt to the same metabolite exchanges, which makes such proposals top-ranked by construction, and Table S7 concedes that the evaluation "cannot presently separate a good substitution score from a bad one", with Cella at 0.617 AUROC against 0.608 for a simple shared-component count. The claim in Sec. 3.5 that the embedding is "tracking function rather than relatedness" conflicts with Table S7, where Cella's nearest-neighbour transfer of 0.362 matches patristic distance at 0.364 and, on carbon sources, falls below the best-fixed-neighbour control at 0.160 versus 0.195.

References

  • Matthew A. Oberhardt, et al. Harnessing the landscape of microbial culture media to predict new organism–media pairings. Nature Communications, 6:8493, 2015.

  • Petr Máša, Tomáš Kliegr, and Marcin P. Joachimiak. Explainable rule-based prediction of cultivation media for microbes. Computational and Structural Biotechnology Journal, 27:5194–5206, 2025.

  • Aaron Weimann, et al. From genomes to phenotypes: Traitar, the microbial trait analyzer. mSystems, 1(6):e00101-16, 2016.

  • Zeqian Li, Ahmed Selim, and Seppe Kuehn. Statistical prediction of microbial metabolic traits from genomes. PLoS Computational Biology, 19(12):e1011705, 2023.

  • Ziniu Hu, et al. Heterogeneous graph transformer. In Proceedings of The Web Conference 2020 (WWW), pp. 2704–2710, 2020.

  • Bishan Yang, et al. Embedding entities and relations for learning and inference in knowledge bases. In International Conference on Learning Representations (ICLR), 2015.

Questions for authors
  • (Q1) Are the supervised models initialised from the DistMult-pretrained encoder, and how do held-out results compare with and without pretraining? A clear gain would support the foundation-model framing and raise my assessment of significance.

  • (Q2) Per-label macro metrics fix the popularity control at 0.5 AUROC by construction, whereas under per-organism recall@k a popularity ranking of common salts would be a strong baseline. How does Cella compare with popularity and lineage-only under such metrics? A clear margin would raise my view of the practical significance.

  • (Q3) BacDive metabolite-utilization records include explicit negative outcomes, so why are carbon-source labels treated as positive-unlabeled? Do the conclusions hold on measured negatives or with a PU-aware objective (Elkan & Noto, 2008), and why does Fig. S2, where AUROC still rises from 0.535 to 0.641, indicate noise rather than data limitation? Robust results would increase my confidence in this task.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.