Research map / Graph learning / GNN architectures
Modelling Microbial Communities with Graph Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Uses graph neural networks (GraphSAGE and a message-passing GNN) on bacterial genome presence/absence features to directly predict steady-state relative abundances of microbial communities, with a Lotka-Volterra-based simulator.
Problem
Predicting bacterial community composition at steady state from bacterial genomes, generalizing to unseen bacteria and larger communities, where existing gLV models cannot generalize to new species and need time-series data.
Approach
Treat each community as a fully connected graph with one node per bacterium whose attributes are binary genome (KEGG KO) vectors, and regress equilibrium relative abundances directly with permutation-equivariant GNNs instead of fitting growth dynamics; also introduces a simulator that generates genomes encoding Lotka-Volterra parameters.
Key results
- GNNs predict unseen bacterial mixes with R2 = 0.81 (Friedman2017) and R2 = 0.77 (BaranwalClark2022), outperforming permutation-sensitive MLPs whose accuracy drops significantly when bacteria are shuffled.
- Generalization to larger communities is only marginal: best R2 = 0.55 (Friedman2017, trained on 2-3 bacteria, tested on 7-8) and 0.45 on BaranwalClark2022 (trained up to 22, tested on 23-26); training on larger to predict smaller gives R2 < 0.
- Generalization to unseen bacteria is partial: e.g., R2 = 0.55 and 0.58 for Ea and Sm on Friedman2017, but negative R2 for Pch.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use a permutation-equivariant GNN rather than an MLP | community abundance prediction on Friedman2017 · Friedman2017 | R2 | Helped |
| shuffle bacteria order at test time (ablation) | MLP community abundance predictor · Friedman2017 | prediction accuracy | Hurt |
| train on communities closer in size to the test communities | GNN ensemble · BaranwalClark2022 communities with 23–26 bacteria | R2: 0.26 -> 0.45 | Helped |
| use SoftMax instead of sigmoid output (ablation) | GNN models · real-data validation sets | MSE: 0.007532 (sigmoid) vs 0.007538 (SoftMax); 0.811864 with SoftMax | No clear effect |
| increase simulated interaction edge density (ablation) | GNN models · simulated communities | R2: GraphSAGE: 0.4749, 0.4034, 0.4184; MPGNN: 0.4324, 0.6016, 0.4173 | Mixed |
| exclude an unseen keystone bacterium from training rather than a random bacterium (ablation) | MPGNN · simulated communities with unseen bacteria | R2: 0.5594 -> 0.6438 | Helped |
| exclude an unseen keystone bacterium from training rather than a random bacterium (ablation) | GraphSAGE · simulated communities with unseen bacteria | R2: 0.4465 -> 0.4900 | Helped |
| increase the diversity of bacteria seen during training (ablation) | MPGNN · simulated communities with unseen bacteria | R2 | Helped |
| train on fewer samples (ablation) | GNNs on simulated communities · simulated communities | R2 | Hurt |
| train only on smaller communities than the test communities (ablation) | GNNs on simulated communities · larger simulated communities | R2 | Hurt |
| use MLP with shuffled input bacteria rather than fixed input order (ablation) | MLP on simulated communities · simulated communities | R2: 0.8364 -> -0.1911 | Hurt |
| use 2-layer GCNII with 100 hidden features rather than GATv2 with 2 layers and 100 hidden features (ablation) | graph architecture · BaranwalClark2022 | R2: -0.1305 -> 0.7090 | Helped |
Experimental setup
Datasets: Friedman2017 BaranwalClark2022 simulated Lotka-Volterra communities
Benchmarks: Friedman2017 BaranwalClark2022 simulated communities
Metrics: R2 MSE
Baselines: MLP GATv2 GCNII generalized Lotka-Volterra models
Builds on: GraphSAGE MPNN (Gilmer et al.) generalized Lotka-Volterra model
Closest papers by meaning
- The Map Equation Goes Neural: Mapping Network Flows with Graph Neural Networks (NeurIPS 2024 · similarity 0.83)
- Whole Genome Transformer for Gene Interaction Effects in Microbiome Habitat Specificity (AAAI 2025 · similarity 0.83)
- OgBench: A Framework for Evaluating Graph Neural Networks on Omics Data (NeurIPS 2026 · similarity 0.83)
- On the Scalability of GNNs for Molecular Graphs (NeurIPS 2024 · similarity 0.82)
- Learning Persistent Community Structures in Dynamic Networks via Topological Data Analysis (AAAI 2024 · similarity 0.82)
- Homomorphism Counts for Graph Neural Networks: All About That Basis (ICML 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.