📄 bioengineering

eBiota: Designing microbial communities from large seed pools with desired function using rapid optimization and deep learning

The paper introduces eBiota, an integrated platform combining graph-based search, extended flux balance analysis, and deep learning to rapidly design and simulate functional microbial communities from large seed pools for target product generation and pathogen inhibition.

Original authors: Jiang, X., Hou, J., Zhang, H., Guo, J., Gu, S., Vandeputte, D., Liao, Y., Guo, Q., Yang, X., Zhou, Y., Geng, P. X., Wang, C., Li, M., Jousset, A., Shen, X., Wei, Z., Zhu, H.

Published 2026-03-31

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jiang, X., Hou, J., Zhang, H., Guo, J., Gu, S., Vandeputte, D., Liao, Y., Guo, Q., Yang, X., Zhou, Y., Geng, P. X., Wang, C., Li, M., Jousset, A., Shen, X., Wei, Z., Zhu, H.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are a master chef trying to create the perfect dish. You have a pantry filled with 21,514 different ingredients (microbes), and you want to cook a meal that produces a specific flavor (a useful chemical like hydrogen or a medicine).

The problem? Trying to mix and match every possible combination of these ingredients to find the perfect recipe would take longer than the age of the universe. It's like trying to find a specific needle in a haystack the size of a mountain, where the haystack is constantly growing.

Enter eBiota. Think of eBiota as a super-smart, AI-powered sous-chef that can instantly scan your massive pantry, figure out which ingredients work well together, and design the perfect recipe for you.

Here is how this "digital kitchen" works, broken down into three simple steps:

1. The Fast Filter (CoreBFS): "The Ingredient Scanner"

First, the chef needs to know which ingredients can actually make the dish.

The Analogy: Imagine you want to bake a cake. You don't need to check every single spice in the world; you just need to find the ones that have flour, eggs, and sugar.
How eBiota does it: It uses a super-fast "searchlight" called CoreBFS. Instead of reading the entire recipe book of every microbe, it quickly scans the "table of contents" (the metabolic pathways) to see if a microbe has the necessary tools to turn raw materials into your target product. It filters out the millions of useless ingredients in seconds, leaving you with a shortlist of candidates.

2. The Taste Test (ProdFBA): "The Efficiency Chef"

Now you have a shortlist, but you need to know which ones actually make the best cake without burning the kitchen down.

The Analogy: Some ingredients might make a cake, but they might be lazy (grow slowly) or make a mess (waste energy). You want a team that works hard, grows fast, and produces the maximum amount of flavor.
How eBiota does it: It uses a tool called ProdFBA. This is like a simulation engine that runs thousands of "what-if" scenarios. It asks: "If we mix Microbe A and Microbe B, will they grow well together? Will they produce more hydrogen than Microbe A alone?" It finds the perfect balance where the microbes are happy, healthy, and super productive.

3. The Social Network (DeepCooc): "The Team Builder"

Even if two ingredients are great individually, they might hate each other when mixed. One might eat the other!

The Analogy: Think of a workplace. You can have two brilliant employees, but if they fight constantly, the project fails. You need a team that gets along.
How eBiota does it: It uses a deep learning model called DeepCooc. This model was trained on 23,000 real-world "social media" posts from nature (microbiome samples from soil, guts, oceans, etc.). It learned the "social rules" of microbes: "Oh, I see that Bacteria X and Bacteria Y always hang out together in nature, so they probably get along." It predicts which microbes will coexist peacefully and which will fight, ensuring your designed community is stable.

What Did They Prove?

The team didn't just build the tool; they tested it in the real world:

The Hydrogen Challenge: They asked eBiota to design a team of microbes to produce hydrogen fuel. It successfully recreated known natural processes and even found new, more efficient combinations that humans hadn't thought of.
The Plant Doctor: They used it to find microbes that could fight a nasty plant disease (R. solanacearum). They tested 94 different bacteria in a lab, and eBiota correctly predicted which ones would act as "bodyguards" for the plants.
The Big Crowd: They mixed all 94 bacteria together in a jar. eBiota predicted exactly how much of each bacteria would survive, matching the real-life results almost perfectly.

Why Does This Matter?

For a long time, designing microbial communities was like trying to build a house by randomly throwing bricks together and hoping a wall forms. It was slow, expensive, and limited to small piles of bricks.

eBiota changes the game. It allows scientists to:

Design from scratch: Create custom microbial teams for making medicines, cleaning up pollution, or creating green energy.
Go big: Use the entire library of known bacteria (21,000+ species) instead of just a handful.
Save time and money: Run simulations on a computer before ever touching a test tube.

In short, eBiota is a "Digital Twin" of the microbial world. It lets us simulate, design, and perfect the invisible workforce of bacteria that could solve some of our biggest challenges in energy, health, and agriculture.

1. Problem Statement

The rational design of microbial communities (consortia) to perform specific functions (e.g., metabolite production, pathogen inhibition) is a critical goal in synthetic biology and biotechnology. However, current approaches face three major bottlenecks when scaling from small, predefined pools to the vast diversity of available microbial genomes:

The "Curse of Dimensionality": The search space for optimal strain combinations grows exponentially with the size of the seed pool. Existing tools are typically limited to pools of a few hundred strains, missing the vast potential of the >20,000 available bacterial genomes.
Suboptimal Production Efficiency: Traditional Flux Balance Analysis (FBA) often maximizes growth rate rather than target metabolite production. This leads to solutions where growth is high but production is low, or where multiple flux states yield the same outcome (solution redundancy), failing to identify the most efficient consortia.
Ecological Instability: Designed communities often fail in practice because they lack ecological stability. Existing methods struggle to predict co-occurrence patterns and higher-order interactions (beyond pairwise) necessary for stable coexistence, especially for novel taxa not present in training datasets.

2. Methodology: The eBiota Platform

The authors introduce eBiota, an integrated platform designed to address these challenges through a hierarchical workflow combining graph theory, metabolic modeling, and deep learning. The platform operates on a seed pool of 21,514 bacterial genomes (eBiota-GEM).

The workflow consists of three core modules:

A. CoreBFS (Rapid Pathway Screening)

Function: A graph-based Breadth-First Search (BFS) algorithm tailored for metabolic networks.
Innovation: Unlike classical BFS, CoreBFS accounts for the scale-free and small-world properties of metabolic networks. It pre-stores "currency metabolites" (high-degree nodes like ATP, metal ions) to avoid search explosion and limits reaction steps to 30 (covering ~99% of pathways).
Goal: Rapidly screens the 21,514 strains to identify those possessing complete metabolic pathways to convert specific substrates into target products.

B. ProdFBA (Production-Oriented Flux Balance Analysis)

Function: An extended FBA algorithm that optimizes for both production efficiency and growth stability.
Innovation: It integrates Parsimonious FBA (pFBA) and Flux Variability Analysis (FVA) into a multi-step optimization:
1. Maximize community growth rate.
2. Maximize target production flux while constraining growth to $\ge$ 99% of the maximum.
3. Minimize total flux to reduce redundancy.
Goal: Identifies strain combinations that maximize target metabolite yield without sacrificing viability, overcoming the limitations of classical FBA which often decouples growth from production.

C. DeepCooc (Deep Learning for Co-occurrence)

Function: A ResNet-based deep learning model trained to predict microbial co-occurrence and mutual exclusion.
Training Data: Trained on 23,323 microbiome samples from the Earth Microbiome Project (EMP), covering >1,500 genera. Labels were derived using the HiOrCo method.
Input Features: Uses 335 transport reactions from GEMs (including metabolite charge and flux bounds) as input.
Goal: Predicts pairwise and higher-order (triplet) co-occurrence patterns to ensure the designed communities are ecologically stable and likely to coexist in nature.

3. Key Contributions

Scale: Successfully designed microbial communities from a seed pool of 21,514 strains, a scale previously computationally prohibitive for existing tools.
Integration: First platform to unify pathway topology screening (CoreBFS), production-focused metabolic optimization (ProdFBA), and data-driven ecological inference (DeepCooc) in a single workflow.
Algorithmic Improvements:
- Developed CoreBFS to reduce the design space from $10^{16}$ combinations to a manageable set with high precision (Accuracy: 0.938 vs. 0.678 for MultiPus).
- Developed ProdFBA to resolve the trade-off between growth and production, identifying 11.38% of strains with high efficiency (vs. 1.5% with classical FBA).
- Developed DeepCooc achieving balanced accuracy of 0.966 for pairs and 0.911 for triplets, outperforming traditional ML models.

4. Results and Validation

The authors validated eBiota using public datasets and extensive in vitro experiments:

Hydrogen Production:
- Single Strains: eBiota simulations matched experimental data for 24 genera with F1-scores of 0.839 (Hydrogen), 0.757 (Formate), and 0.909 (Acetate), outperforming the state-of-the-art tool MICOM.
- Consortia: Successfully simulated 29/34 two-member and 9/9 three-member hydrogen-producing consortia found in literature. It correctly ranked hydrogen yields for 11 mixed microbial cultures (MMC) with a Spearman correlation of $R=0.861$ ( $p<0.001$ ), whereas MICOM showed no correlation.
- Efficiency: Co-cultures designed by eBiota showed a median 1.5-fold increase in hydrogen production efficiency compared to mono-cultures.
Pathogen Inhibition (Plant Pathogen R. solanacearum):
- Tested against 94 bacterial strains isolated from plant roots.
- eBiota correctly identified 88/94 strains as competitors (growth inhibitors).
- Mechanism: Identified that competitors had significantly lower metabolic distance scores and higher competition indices.
- Pareto Analysis: Used Pareto front analysis to delineate trade-offs between pathogen and competitor growth, accurately distinguishing between strong competitors and independent strains.
Community Composition Prediction:
- Simulated a 94-strain synthetic community.
- Predicted relative abundances showed a strong correlation with in vitro 16S rRNA sequencing data ( $R=0.827$ , $p<0.001$ ) and a low Bray-Curtis distance (0.386).
Multi-Metabolite Production:
- Designed consortia for 28 metabolites using mouse gut strains. eBiota achieved a recall of 0.770, significantly outperforming MICOM (0.361).

5. Significance

Digital Twin of the Microbiome: eBiota serves as a "digital twin," capable of simulating diverse microbial behaviors (production, interactions, abundance) from first principles combined with ecological data.
Accelerated Discovery: By enabling ab initio design from massive seed pools, it uncovers novel functional combinations and highly efficient consortia that would be missed by traditional, small-pool approaches.
Broad Applicability: The platform offers a versatile framework for applications in green energy (hydrogen), sustainable agriculture (biocontrol), microbiome therapeutics, and industrial fermentation.
Open Source: The code, GEM database (eBiota-GEM), and trained models are publicly available, fostering reproducibility and further development in synthetic ecology.

Limitations noted by authors: Currently relies on automatically reconstructed GEMs (curated models would improve accuracy), focuses only on bacteria (excluding fungi/archaea), uses static FBA (lacks dynamic temporal modeling), and predicts co-occurrence rather than explicitly modeling stable coexistence mechanisms. Future work aims to address these through dynamic FBA and expanded taxonomic coverage.