A probabilistic model for mining implicit 'chemical compound-gene' relations from literature.

Area of Science:

• Computational Biology
• Bioinformatics
• Cheminformatics

Background:

• Chemical genomics is a growing field in molecular biology, necessitating the identification of associations between chemical compounds (drugs) and genes.
• Co-occurrence analysis of biological entities in literature is a common method for inferring relationships.

Purpose of the Study:

• To develop a probabilistic model for mining implicit relationships between chemical compounds and genes from literature co-occurrence data.
• To evaluate the model's performance using diverse co-occurrence datasets and validate predicted drug-gene pairs.

Main Methods:

• Proposed a Mixture Aspect Model (MAM) and an associated parameter estimation algorithm.
• Utilized MEDLINE records and the ChEBI database for validation.
• Experimented with compound-gene, gene-gene, and compound-compound co-occurrence datasets.

Main Results:

• The MAM trained on all co-occurrence datasets significantly outperformed simpler models.
• Incorporating compound-compound co-occurrences proved most effective in enhancing predictive accuracy.
• The top 20 predicted drug-gene pairs were identified and validated from biological, medical, and pharmaceutical perspectives.

Conclusions:

• The proposed MAM effectively identifies chemical compound-gene relationships from literature.
• The model provides a robust framework for leveraging various co-occurrence data types, with compound-compound interactions being particularly valuable.
• This approach aids in discovering novel drug-gene associations relevant to multiple scientific disciplines.