PISTON: Predicting drug indications and side effects using topic modeling and natural language processing.

Area of Science:

• Computational biology
• Bioinformatics
• Pharmacogenomics

Background:

• Drug discovery is time-consuming and costly, with significant risks from unforeseen side effects.
• Computational methods, particularly text mining, offer a cost-effective alternative to experimental approaches for analyzing biological literature.
• Systems biology leverages text mining to uncover hidden relationships between drugs, genes, and diseases.

Purpose of the Study:

• To develop an algorithm for predicting novel drug-phenotype and drug-side effect associations.
• To utilize topic modeling and natural language processing (NLP) for efficient analysis of scientific literature.
• To provide a computational tool for identifying potential drug indications and mitigating risks from adverse drug reactions.

Main Methods:

• Extracting drug-gene co-occurrence sentences from literature abstracts.
• Employing NLP to identify gene-regulatory relationships (up-regulation/down-regulation) described in text.
• Applying topic modeling to group genes and regulatory relationships into topics, creating a drug-topic probability matrix.
• Constructing a classifier using the drug-topic matrix to predict drug-phenotype and drug-side effect associations.

Main Results:

• The algorithm successfully predicts novel drug indications and side effects using a unified approach.
• It enables the exclusion of candidate drugs with undesirable side effects, enhancing patient safety.
• The system facilitates continuous updates of candidate drug lists as new literature becomes available.

Conclusions:

• The proposed text mining algorithm offers an efficient, low-cost method for drug repositioning and adverse effect prediction.
• This approach accelerates the identification of novel therapeutic applications and potential safety concerns.
• The PISTON resource provides a valuable tool for researchers in drug discovery and systems biology.