Using word evolution to predict drug repurposing.

Area of Science:

• Computational biology
• Natural language processing
• Pharmacology

Background:

• Traditional drug discovery relies on linking knowledge from separate publications.
• This can lead to over-generation of potential drug pairs.
• An alternative approach using word evolution is explored to improve efficiency.

Purpose of the Study:

• To investigate the use of word evolution for detecting drugs suitable for repurposing.
• To leverage changing word contexts to identify novel therapeutic applications for existing drugs.

Main Methods:

• Word embeddings were generated from chronologically ordered MEDLINE publications at bi-monthly intervals.
• A time series of word embeddings was created for each word.
• Clinical drugs were the focus, with repurposing status determined using the Unified Medical Language System (UMLS) or SemRep extracted semantic triples.

Main Results:

• Deep learning classification models were trained and evaluated using 5-fold cross-validation.
• Performance reached 65% accuracy with UMLS labels and 81% accuracy with SemRep labels.
• These results indicate the effectiveness of the word evolution technique for drug repurposing detection.

Conclusions:

• The word evolution method shows significant promise for identifying candidate drugs for repurposing.
• The performance is dependent on the annotation approach (UMLS vs. SemRep).
• Different deep learning architectures are optimal depending on the available training data quantity and annotation method.