SpoMAG is an R-based machine learning tool developed to predict the sporulation potential of Metagenome-Assembled Genomes (MAGs) from uncultivated Firmicutes species, particularly from the Bacilli and Clostridia classes.

SpoMAG leverages the complex combination of presence or absence of sporulation-associated genes to infer whether a genome is capable of undergoing sporulation, even in the absence of cultivation or complete genome assemblies. This strategy allows researchers to assess sporulation potential using only functional annotations from metagenomic data.

SpoMAG predicts the sporulation potential of a given genome through a three-step workflow:

• sporulation_gene_name(), which parses a functional annotation table, such as those generated by eggNOG-mapper, and identifies genes related to sporulation using curated gene names and KEGG orthologs. It requires as input a data frame with the columns Preferred_name, KEGG_ko, and genome_ID (named exactly as such). The function outputs a filtered table of annotations containing sporulation-related genes.

• build_binary_matrix(), which converts the filtered annotations into a binary matrix, with rows representing genomes and columns representing genes (1 = present, 0 = absent). Missing genes are automatically filled with zeros to ensure consistent input for machine learning models.

• predict_sporulation(), which applies a pre-trained ensemble model combining Random Forest and Support Vector Machine predictions into a stacked meta-classifier. It outputs the classification label (Sporulating or Non_sporulating), model probabilities, and final ensemble probability value.

SpoMAG abstracts the complexity of machine learning, allowing users to simply provide an annotation table and receive interpretable predictions. The tool is designed to be accessible even to those without prior expertise in bioinformatics or machine learning.

Its ensemble learning approach combines the predictions from Random Forest and Support Vector Machine classifiers, trained on high-quality labeled datasets of known spore-formers and non-spore-formers. The predictions are then used as features in a meta-classifier using model stacking, enhancing prediction accuracy and allowing SpoMAG to capture complementary decision boundaries from each model.

Model performance was evaluated using cross-validation and standard metrics including AUC-ROC, F1-score, Accuracy, precision, specificity, and recall. As a result, SpoMAG delivers high sensitivity and specificity across MAGs recovered from different hosts' microbiota.

Whether analyzing hundreds of MAGs or a single novel lineage, SpoMAG offers a robust, automated, and cultivation-independent solution to assess sporulation potential. No phenotypic validation or manual annotation is required, making it a practical tool for exploring the ecological and functional roles of spore-forming bacteria.

The repository for SpoMAG is at GitHub on the https://github.com/labinfo-lncc/SpoMAG. In this website, you can report a bug and get help.

Paper under publication.

You can install the SpoMAG package directly from GitHub using:

# Install devtools if not already installed
install.packages("devtools")

# Install SpoMAG from GitHub
devtools::install_github("labinfo-lncc-br/SpoMAG")

SpoMAG depends on the following packages:

• dplyr, version 1.14
• tidyr, version 1.3.1
• tibble, version 3.2.1
• readr, version 2.1.5
• caret, version 7.0.1
• randomForest, version 4.7.1.2

It extracts sporulation-related genes from an annotation dataframe by searching for gene names and KEGG orthologs.

• Input: A dataframe with at least Preferred_name, KEGG_ko and genome_ID columns.

• Output: A filtered dataframe containing sporulation-related hits, each annotated with the standardized columns spo_gene_name and spo_process.

genes <- sporulation_gene_name(df)

It creates a binary matrix indicating the presence (1) or absence (0) of known sporulation genes in each genome.

• Input: A dataframe output from the sporulation_gene_name() function.

• Output: A wide-format dataframe (genome in rows, genes in columns).

matrix <- build_binary_matrix(genes)

Note: The function automatically fills in missing genes with 0 to ensure consistent input for sporulation-capacity prediction.

It applies a pre-trained ensemble machine learning model to predict the sporulation potential of genomes based on the binary matrix of genes.

• Input:

  binary_matrix: Output from build_binary_matrix()

• Output: A dataframe with:

  genome_ID: the genome ID you are using as input

  RF_Prob: Random Forest probability of being a spore-former

  SVM_Prob: Support Vector Machine probability of being a spore-former

  Meta_Prob_Sporulating: Ensemble probability of being a spore-former

  Meta_Prediction: Final prediction (Sporulating or Non_sporulating)

results <- predict_sporulation(binary_matrix = matrix)

To use SpoMAG, your input must be a functional annotation table, such as the output from eggNOG-mapper, containing at least three columns:

• genome_ID: a name you have chosen for the genome you are working on
• Preferred_name: the predicted name of the gene
• KEGG_ko: the KEGG Orthology code (e.g., K07699)

Each row should represent one gene annotation.

Another difference of SpoMAG is its ability to infer gene presence in the annotation file even in cases where annotations are ambiguous. As shown in the example above, some rows can contain a valid KEGG_ko code but a missing or undefined Preferred_name (e.g., “-”), while others have a predicted gene name but no associated KO. SpoMAG integrates both fields to assign a unified spo_gene_name:

• If Preferred_name is missing but KEGG_ko matches a known sporulation-associated KO, the gene is identified based on the KO.

• If KEGG_ko is missing but Preferred_name matches a known sporulation gene, the gene is identified based on the name.

• If both are informative and match known references, preference is given to Preferred_name.

This is a quick example using the included files: one_sporulating.csv (a known spore-former) and one_asporogenic.csv (a known non-spore-former). The genome used for the spore-former here is the following:

• GCF_000007625.1 (Clostridium tetani E88, Clostridia class)

The genome used for the non-spore-former here is the following:

• GCF_000006785.2 (Streptococcus pyogenes M1 GAS SF370, Clostridia class)

# Load package
library(SpoMAG)

# Load example annotation tables
file_spor <- system.file("extdata", "one_sporulating.csv", package = "SpoMAG")
file_aspo <- system.file("extdata", "one_asporogenic.csv", package = "SpoMAG")

# Read files
df_spor <- readr::read_csv(file_spor, show_col_types = FALSE)
df_aspo <- readr::read_csv(file_aspo, show_col_types = FALSE)

# Step 1: Extract sporulation-related genes
genes_spor <- sporulation_gene_name(df_spor)
genes_aspo <- sporulation_gene_name(df_aspo)

# Step 2: Convert to binary matrix
bin_spor <- build_binary_matrix(genes_spor)
bin_aspo <- build_binary_matrix(genes_aspo)

# Step 3: Predict using ensemble model (preloaded in package)

result_spor <- predict_sporulation(bin_spor)
result_aspo <- predict_sporulation(bin_aspo)

# View results
print(result_spor)
print(result_aspo)

This is a quick example using the included files: ten_sporulating.csv (ten known spore-formers) and ten_asporogenic.csv (ten known non-spore-formers). The genomes used for the spore-formers here are the following:

• GCF_000011985.1 (Lactobacillus acidophilus NCFM, Bacilli class)
• GCF_000338115.2 (Lactobacillus plantarum ZJ316, Bacilli class)
• GCF_000016825.1 (Lactobacillus reuteri DSM 20016, Bacilli class)
• GCF_000011045.1 (Lactobacillus rhamnosus GG ATCC 53103, Bacilli class)
• GCF_000237995.1 (Pediococcus claussenii ATCC BAA-344, Bacilli class)
• GCF_000145035.1 (Butyrivibrio proteoclasticus B316, Clostridia class)
• GCF_000020605.1 (Eubacterium rectale ATCC 33656, Clostridia class)
• GCF_900070325.1 (Herbinix luporum SD1D, Clostridia class)
• GCF_003589745.1 (Lachnoanaerobaculum umeaense DSM 23576, Clostridia class)
• GCF_000225345.1 (Roseburia hominis A2-183, Clostridia class)

The genomes used for the non-spore-formers here are the following:

• GCF_000011145.1 (Bacillus halodurans C-125, Bacilli class)
• GCF_000008425.1 (Bacillus licheniformis DSM 13 ATCC 14580, Bacilli class)
• GCF_000009045.1 (Bacillus subtilis subsp. subtilis str. 168, Bacilli class)
• GCF_000338755.1 (Bacillus thuringiensis serovar kurstaki HD73, Bacilli class)
• GCF_000010165.1 (Brevibacillus brevis NBRC 100599, Bacilli class)
• GCF_000009205.2 (Clostridioides difficile 630, Clostridia class)
• GCF_000008765.1 (Clostridium acetobutylicum ATCC 824, Clostridia class)
• GCF_000063585.1 (Clostridium botulinum A ATCC 3502, Clostridia class)
• GCF_000013285.1 (Clostridium perfringens ATCC 13124, Clostridia class)
• GCF_000007625.1 (Clostridium tetani E88, Clostridia class)

# Load package
library(SpoMAG)

# Load example annotation tables
file_spor <- system.file("extdata", "ten_sporulating.csv", package = "SpoMAG")
file_aspo <- system.file("extdata", "ten_asporogenic.csv", package = "SpoMAG")

# Read files
df_spor <- readr::read_csv(file_spor, show_col_types = FALSE)
df_aspo <- readr::read_csv(file_aspo, show_col_types = FALSE)

# Step 1: Extract sporulation-related genes
genes_spor <- sporulation_gene_name(df_spor)
genes_aspo <- sporulation_gene_name(df_aspo)

# Step 2: Convert to binary matrix
bin_spor <- build_binary_matrix(genes_spor)
bin_aspo <- build_binary_matrix(genes_aspo)

# Step 3: Predict using ensemble model (preloaded in package)
result_spor <- predict_sporulation(bin_spor)
result_aspo <- predict_sporulation(bin_aspo)

# View results
print(result_spor)
print(result_aspo)