Package Overview

m6APrediction — An R package for computational prediction of RNA N6-methyladenosine (m6A) modification sites.

Existing m6A prediction tools are either web-based or difficult to integrate into high-throughput workflows.m6APrediction provides a fully local, scriptable solution enabling reproducible m6A site inference inside R pipelines.

This package provides a user-friendly interface for identifying potential m6A methylation sites based on sequence-derived features and transcript-level annotations. It was developed to assist researchers in RNA epitranscriptomics by enabling high-throughput and reproducible prediction of m6A modifications using machine learning models.

Specifically, m6APrediction employs a Random Forest classifier trained on experimentally validated datasets. It allows users to supply 5-mer RNA sequences along with biological and structural features (e.g., GC content, RNA region, exon length, evolutionary conservation) to estimate the probability that a site contains an m6A modification.

The package includes two main functions:

• prediction_multiple() — performs prediction for multiple sites from a feature table (batch mode).

• prediction_single() — performs prediction for one sequence using individual feature inputs (single mode).

Together, these functions make it possible to integrate m6A prediction into downstream RNA analysis pipelines or apply it directly to novel datasets.

Installation

To install m6APrediction from GitHub, you can use either the devtools or remotes package.Both methods ensure that the package is downloaded and compiled from the GitHub repository.

# Install devtools or remotes if not already installed

if(!requireNamespace("devtools")){
  install.packages("devtools") 
}
if(!requireNamespace("remotes")){
  install.packages("remotes")
}

# Install m6APrediction directly from GitHub
devtools::install_github("Hanqi0523/m6APrediction") 
# or equivalently 
remotes::install_github("Hanqi0523/m6APrediction")

After installation, you can load the package with:

library(m6APrediction)

This will make all exported functions (including dna_encoding(), prediction_multiple(), and prediction_single()) available in your current R session.The package automatically includes example data and a pre-trained Random Forest model for demonstration purposes.

Example Usage

The following example demonstrates how to use both core functions — prediction_multiple() for batch predictions and prediction_single() for individual site predictions.

First, we load a pre-trained model (rf_fit.rds) and a small example dataset (m6A_input_example.csv) that come with the package. Then, we perform both batch and single-sequence predictions to illustrate their workflow.

library(m6APrediction)

# ---- Load pre-trained model and example input ----
rf_path <- system.file("extdata", "rf_fit.rds", package = "m6APrediction") 
csv_path <- system.file("extdata", "m6A_input_example.csv", package = "m6APrediction")

ml_fit <- readRDS(rf_path) df <- read.csv(csv_path, check.names = FALSE)

# ---- Batch prediction for multiple sites ----
# This function returns a data frame containing predicted probabilities and binary classifications
res <- prediction_multiple( ml_fit, feature_df = df, positive_threshold = 0.5 )

# Display the first few predictions
head(res[, c("DNA_5mer", "predicted_m6A_prob", "predicted_m6A_status")])

# ---- Single-sample prediction ----
# Here we manually input one RNA site and obtain its m6A probability and label
prediction_single(ml_fit, 
                  gc_content = 0.55, 
                  RNA_type = "mRNA", 
                  RNA_region = "CDS", 
                  exon_length = 1200, 
                  distance_to_junction = 45, 
                  evolutionary_conservation = 0.82, 
                  DNA_5mer = "ATCGA" )

Model Performance

The Random Forest classifier embedded in m6APrediction was trained on experimentally validated m6A sites using balanced positive/negative sampling and 5-mer sequence–derived features.

To evaluate predictive performance, receiver operating characteristic (ROC) and precision–recall (PRC) curves were generated using an independent test set. The model achieved:

Curve Type	AUC
ROC	0.886
PRC	0.701

The ROC AUC of 0.886 indicates strong overall discriminative power between methylated and non-methylated sites. Given the class imbalance inherent to m6A datasets, the PRC AUC (0.701) further demonstrates that the classifier maintains meaningful precision across recall levels rather than relying on majority-class bias.

The corresponding performance curves are shown below:

License

This package is distributed under the MIT License.