Explainable multi-task learning improves the parallel estimation of polygenic risk scores for many diseases through...

Area of Science:

• Genetics
• Computational Biology
• Bioinformatics

Background:

• Complex diseases often share genetic underpinnings and occur together in individuals.
• Improving the accuracy of polygenic risk scores (PRS) for complex diseases is crucial for personalized medicine.

Purpose of the Study:

• To investigate if multi-task learning (MTL) can enhance the accuracy of PRS by exploiting shared genetic determinants across multiple diseases.
• To test the hypothesis that simultaneous PRS estimation improves predictive power compared to single-task learning (STL).

Main Methods:

• A multi-task learning (MTL) approach using an explainable neural network architecture was employed.
• PRS were estimated in parallel for 17 prevalent cancers (pan-cancer MTL model) and 60 prevalent non-cancer diseases (pan-disease MTL model).
• Performance was compared against independent single-task learning (STL) models for individual diseases.

Main Results:

• The pan-cancer MTL model demonstrated generally more accurate PRS estimations for individual cancers compared to STL models.
• Similar performance improvements due to positive transfer learning were observed in the pan-disease MTL model.
• Interpretation revealed significant genetic correlations between key single nucleotide polymorphisms (SNPs) used in MTL models.

Conclusions:

• Simultaneous PRS estimation using MTL effectively leverages shared genetic etiology to improve predictive accuracy for multiple diseases.
• The findings suggest a well-connected network of diseases with overlapping genetic bases.
• MTL offers a promising strategy for enhancing polygenic risk prediction in complex diseases.