GsVec

GsVec computes semantic similarity between gene signatures by applying NLP-derived vector representations to enhance biological interpretation of gene expression results from microarray and next-generation sequencing.

Key Features:

• NLP-based semantic representation: Employs a distributed document representation method from natural language processing to capture semantic relationships among genes and signatures.
• Gene-topic vector construction: Creates gene-topic vectors by multiplying a gene’s distributed-representation feature vector with the probability of the gene signature topic and a low-frequency weighting based on occurrence across all gene signatures.
• Signature vector formation: Concatenates gene-topic vectors for genes within each gene signature to produce a comprehensive signature vector.
• Similarity quantification: Calculates cosine distances between signature vectors to quantify relevance and compare specificity and importance among gene signatures.
• Training data: Trained on approximately 5,000 canonical pathway and Gene Ontology (GO) biological process gene signatures from the Molecular Signatures Database (MSigDB).
• Validation: Validated using the BioCarta pathway database and actual differentially expressed gene data, with comparisons to Fisher’s exact test.
• Implementation language: Implemented entirely in R.
• Input types: Operates on gene signatures derived from microarray and next-generation sequencing data.

Scientific Applications:

• Gene signature interpretation: Enables semantic interpretation of gene signatures derived from gene expression studies.
• Disease mechanism analysis: Supports characterization of molecular pathways and mechanisms underlying diseases.
• Target prioritization: Assists in prioritizing candidate genes and pathways for potential therapeutic targeting.
• Pathway and process comparison: Facilitates comparison of canonical pathways and GO biological processes using vector-based similarity metrics.

Methodology:

Generate distributed gene representations; compute gene-topic vectors by multiplying gene feature vectors with topic probabilities and inverse signature-frequency weighting; concatenate gene-topic vectors to form signature vectors; compute cosine distances between signature vectors; train on ~5,000 MSigDB canonical pathway and GO biological process signatures and validate with BioCarta pathways and actual differentially expressed gene data; implemented in R.