NEGradient_GenePriority

The repository NEGradient_GenePriority (short for "Non-Euclidean Gradient Methods for Matrix Completion in Gene Prioritization") provides the code and data to reproduce the results presented in the paper "Gene prioritization using Bayesian matrix factorization with genomic and phenotypic side information." This study introduces a novel method for gene prioritization by combining Bayesian matrix factorization (BMF) with genomic and phenotypic side information, enabling robust predictions and improved identification of disease-associated genes.

---

Overview

1. Problem Formulation:

   • Input: Matrix $O$ of gene-disease associations ($O_{ij} = 1$ for known associations, $O_{ij} = 0$ for unknown or missing associations).
   • Objective: Predict potential associations $\hat{O}$, where $\hat{O}_{ij}$ are floating-point scores indicating the likelihood of gene-disease associations.

2. Data Preparation:

   • Positive Class: Known gene-disease associations ($O_{ij} = 1$).
   • Negative Class: Randomly sampled negatives ($O_{ij} = 0$), ensuring no overlap with positives.
   • Unknown Class: Represented as missing values in the sparse encoding.
   • Procedure:
     • Generate Negative Samples: Randomly sample negatives and merge with the positive class to form $O_1$.
     • Create Train-Test Splits: Split $O_1$ into train and test sets across six random iterations for cross-validation.

3. Incorporating Side Information:

   • Load side information (e.g., gene or disease features) and normalize using the Frobenius norm.
   • Integrate side information into the training process to enhance predictions.

4. Model Training:

   • Train matrix completion models on the train split of $O_1$, using side information.
   • Compute the Root Mean Squared Error (RMSE) during training as a measure of model performance.
   • Repeat training across six random splits to yield six independent models, $M_i \quad \forall i \in \set{1,2,3,4,5,6}$.

5. Matrix Completion:

   • Use the trained models to generate the completed matrix, generating six predicted matrices ($\hat{O_i} \quad \forall i \in \set{1,2,3,4,5,6}$) with floating-point scores.
   • Aggregate the six matrices using a mean operation to produce $\bar{O}$ (averaged matrix).

6. Model Testing and Ranking:

   • Rank genes for each disease (columns of $\bar{O}$) based on the predicted scores.
   • Assume missing values are zeros for ranking purposes.
   • Use only positive associations from the test set for evaluation. For each disease, highly ranked genes in $\bar{O}$ should correspond to true positive associations.

7. Evaluation and Metrics:

   • Compute ranking metrics to assess the quality of predictions, focusing on the ability of the model to rank true positive associations higher.
   • Metrics are aggregated across the six splits to ensure robustness.

---

Dataset Generation

The above procedure describes the general approach. However, the reference paper uses two datasets, OMIM1 and OMIM2, respectively named $O_1$ and $O_2$. The procedure for $O_2$ differs in the following way:

• Subset Relationship: $O_2$ is a subset of $O_1$.
• Model Generation: For $O_2$, models are generated using an N-Fold cross-validation approach, rather than the N random splits used for $O_1$.

OMIM1 Dataset Construction

Input:

• Gene-disease DataFrame: $d$
• Number of splits: $N$
• Sparsity factor $\alpha$, the ratio of 0s to 1s.

Output:

• Unified sparse matrices: $O_{1_{i}} \quad \forall i \in {1, N}$
• Combined splits: $S_{1_{i}} \quad \forall i \in {1, N}$

Procedure:

1. Convert $d$ into a sparse matrix $O_{1_{1s}}$.
2. Sample $N$ zero matrices $O_{1_{0s, i}}$ from $O_{1_{1s}}$ using $\alpha \quad \forall i \in {1, N}$.
3. Combine $O_{1_{1s}}$ and $O_{1_{0s, i}}$ into unified matrices $O_{1_{i}} \quad \forall i \in {1, N}$.
4. Split $O_{1_{1s}}$ indices randomly into subsets $S_{1_{1s}}$.
5. Split $O_{1_{0s, i}}$ indices randomly into subsets $S_{1_{0s, i}} \quad \forall i \in {1, N}$.
6. Merge splits from positive ($S_{1_{1s}}$) and zero ($S_{1_{0s, i}}$) samples into $S_{1_{i}} \quad \forall i \in {1, N}$.

---

OMIM2 Dataset Construction

Input:

• Gene-disease DataFrame: $d$
• Number of folds: $N$
• Association threshold: $T$
• Sparsity factor $\alpha$, the ratio of 0s to 1s.

Output:

• Unified sparse matrix: $O_2$
• Combined folds: $S_2$

Procedure:

1. Filter diseases with fewer than $T$ associations.
2. Convert $d$ into a sparse matrix $O_{2_{1s}}$.
3. Sample zeros from $O_{2_{1s}}$ using $\alpha$.
4. Combine $O_{2_{1s}}$ and $O_{2_{0s}}$ into a unified matrix $O_2$.
5. Split $O_{2_{1s}}$ indices randomly into subsets $S_{2_{1s}}$.
6. Split $O_{2_{0s}}$ indices randomly into subsets $S_{2_{0s}}$.
7. Merge splits from positive ($S_{2_{1s}}$) and zero ($S_{2_{0s}}$) samples into $S_2$.

---

Installation

Steps:

1. Clone the repository:

   git clone https://github.com/azimgivron/NEGradient_GenePriority.git
   cd NEGradient_GenePriority

2. Create and activate a virtual environment:

   python -m venv venv
   source venv/bin/activate

3. Upgrade pip and install dependencies:

   pip install --upgrade pip
   pip install -r requirements.txt

---

Requirements

• Python 3.11
• pip (Python package installer)
• Virtual environment (venv) module

---

Usage

Run the main script to reproduce the results:

python main.py

This script performs gene prioritization using Bayesian matrix factorization with genomic and phenotypic side information.

---

Repository Structure

• NEGradient_GenePriority/: Main modules for preprocessing, matrix operations, and evaluation.
• main.py: Main script for running the gene prioritization pipeline.
• requirements.txt: List of dependencies for the project.
• Dockerfile: Configuration for containerized deployment.
• pyproject.toml: Project and dependency configuration for building and distribution.
• .gitignore: Specifies files to ignore in the repository.
• LICENSE: Project license.

---

Key Features

1. Matrix Completion with Bayesian Matrix Factorization (BMF):

   • Predicts gene-disease associations based on sparse data.

2. Support for Genomic and Phenotypic Side Information:

   • Combines genomic and phenotypic data to improve prioritization accuracy.

3. Flexible Preprocessing and Evaluation:

   • Tools for data preprocessing, creating folds/splits, and computing metrics like ROC-AUC and BEDROC.

---

License

This project is licensed under the MIT License. See the LICENSE file for details.

---

Contact

For questions, suggestions, or issues, please open an issue on the repository.