Predicting drug-gene relations via analogy tasks with word embeddings

Area of Science:

• Computational biology
• Bioinformatics
• Natural Language Processing

Background:

• Word embeddings transform text into numerical vectors, enabling analogy tasks via vector arithmetic.
• BioConceptVec is a biological embedding model trained on PubMed abstracts.
• Drug-gene relations are crucial for understanding pharmacology and developing new therapeutics.

Purpose of the Study:

• To investigate if BioConceptVec and similar embeddings capture drug-gene relationships.
• To assess the predictive power of these embeddings for identifying drug targets.
• To evaluate the impact of biological pathway information on predictive performance.

Main Methods:

• Trained custom embeddings on PubMed abstracts.
• Utilized analogy computations (vector arithmetic) on embeddings to predict drug-gene relations.
• Incorporated biological pathway categorization for drugs and genes.
• Evaluated performance on datasets split by year to assess temporal prediction capabilities.

Main Results:

• Demonstrated that BioConceptVec and custom embeddings contain drug-gene relation information.
• Showed that analogy computations can predict target genes from drugs.
• Found that categorizing entities by biological pathways significantly improves prediction performance.
• Confirmed that embeddings trained on past data can predict future drug-gene relations.

Conclusions:

• Vector embeddings, like BioConceptVec, offer a computationally efficient method for predicting drug-gene interactions.
• Integrating biological pathway information enhances the accuracy of these predictions.
• This analogy-based approach shows promise as a viable alternative to complex models for drug-gene relation discovery.