Gene2vec

Gene2vec generates 200-dimensional distributed vector representations (embeddings) of human genes from transcriptome-wide gene co-expression data aggregated from 984 Gene Expression Omnibus (GEO) datasets to capture functional relatedness and enable predictive analyses.

Key Features:

• Distributed Representation: Embeds genes into a continuous 200-dimensional vector space that places functionally similar genes closer together.
• Machine-learning-based Embedding: Produces embeddings using machine learning techniques inspired by word embedding methodologies.
• Training Data: Trains on transcriptome-wide gene co-expression patterns compiled from 984 GEO datasets.
• Functional Relatedness Capture: Recovers known biological pathways, with genes within a pathway showing an average inner product 1.52 times greater than random gene pairs.
• Gene Co-expression Mapping: Applies t-SNE (t-distributed Stochastic Neighbor Embedding) to visualize local concentrations of tissue-specific genes.
• Predictive Applications: Uses embedded vectors for predictive tasks such as gene–gene interaction prediction using only gene names.

Scientific Applications:

• Pathway discovery and recovery: Identifies and recovers known biological pathways based on embedding similarity.
• Functional annotation: Supports inference of gene function from vector proximity in embedding space.
• Gene–gene interaction prediction: Predicts interactions between genes using embeddings derived solely from gene co-expression patterns.
• Tissue-specific co-expression analysis: Visualizes and examines tissue-localized gene co-expression patterns via t-SNE maps.

Methodology:

Train embeddings using machine learning methods inspired by word embeddings on transcriptome-wide gene co-expression data from 984 GEO datasets to produce 200-dimensional gene vectors, and apply t-SNE (t-distributed Stochastic Neighbor Embedding) for visualization.