Home

Welcome to the MRS wiki

MRS (Metabolism-Related Score) is an R package designed for gene prioritization in positive-only scenarios using Spy-based positive-unlabeled (PU) learning. It integrates reliable negative identification, model comparison, ablation-based feature selection, and final model tuning to identify candidate genes from unlabeled gene sets.

Some jargons

• PU learning: Positive-unlabeled learning is a machine learning framework for binary classification using only positive and unlabeled data. MRS adopts a Spy-based mehtod, which injects a small fraction of positive samples into the unlabeled set ("spies") to estimate the class prior and identify reliable negatives.
• Ablation-based feature selection: A technique to evaluate feature importance by systematically removing features.

General introduction

The MRS package provides a robust and modular machine learning framework for gene classification in datasets containing only positive samples. It is built upon the well-established caret infrastructure, enabling flexible modeling, cross-validation, and performance evaluation.

Originally developed to prioritize metabolism-related genes in Plasmodium falciparum, MRS implements a semi-supervised learning pipeline centered around the Spy PU-learning strategy. This approach identifies reliable negative samples from unlabeled data, enabling accurate classifier training without known negatives.

The full MRS pipeline consists of:

1. Tune the parameters for Spy PU-learning to control the selection of reliable negatives.
2. Identify reliable negative samples from unlabeled genes using the Spy strategy.
3. Compare ten classification models to select the best-performing one based on cross-validation.
4. Optionally perform feature selection using an ablation study to retain informative features.
5. Train the final classification model on the selected feature set and evaluate its performance.

While designed for gene prioritization tasks, MRS is broadly applicable to other binary classification problems where only positive labels are available. It is particularly suitable for biological, clinical, and high-throughput screening scenarios involving semi-supervised data constraints.

Examples

• MRS example on Plasmodium falciparum data

Questions

Visit issues or contact @yang-qingxin regarding issues.