Automated annotation of disease subtypes

Area of Science:

• Computational biology
• Genomics
• Medical informatics

Background:

• Accurate disease classification into subtypes is essential for effective research and treatment.
• The Open Targets Platform (OT) aids disease ontology but often has incomplete annotations, especially for rare diseases.
• Manual expert input for disease annotation is resource-intensive and time-consuming.

Purpose of the Study:

• To develop and evaluate a machine learning approach for identifying diseases with potential subtypes.
• To leverage existing biomedical data for automated disease classification and subtype discovery.
• To improve the accuracy and scalability of disease annotation within the Open Targets Platform.

Main Methods:

• Utilized a machine learning model applied to approximately 23,000 diseases in the Open Targets Platform.
• Derived novel features from direct evidence to predict disease subtypes.
• Integrated pre-trained deep-learning language models to enhance predictive performance.

Main Results:

• Achieved an 89.4% ROC AUC in identifying known disease subtypes.
• Successfully identified 515 candidate diseases predicted to have unannotated subtypes.
• Demonstrated the benefit of integrating deep-learning language models for subtype prediction.

Conclusions:

• The developed models effectively partition diseases into distinct subtypes, enhancing disease ontology.
• This scalable methodology improves knowledge-based annotations and aids in personalized medicine.
• Identified candidate diseases represent valuable targets for further investigation and potential new therapeutic indications.