Highly Scalable Task Grouping for Deep Multi-Task Learning in Prediction of Epigenetic Events.

Area of Science:

• Genomics
• Computational Biology
• Machine Learning

Background:

• Deep neural networks (DNNs) predict cellular events from DNA sequences, aiding genome-wide association studies (GWAS).
• Multi-task learning (MTL) enhances DNN training but often causes negative transfer, where performance degrades for some tasks.
• Existing MTL frameworks share a single feature extraction network, limiting scalability and effectiveness.

Purpose of the Study:

• To develop a scalable framework for task grouping in MTL to mitigate negative transfer.
• To improve the performance of DNNs trained for predicting cellular events from DNA sequences.
• To enhance the elucidation of biological mechanisms underlying GWAS findings.

Main Methods:

• Proposed a highly scalable task grouping framework for MTL.
• Jointly trained only tasks that are potentially beneficial to each other.
• Exploited network weights from task-specific classification heads obtained via one-time joint training.

Main Results:

• Demonstrated the effectiveness of the proposed task grouping framework.
• Showcased superiority over baseline methods on a dataset of 367 epigenetic profiles.
• Successfully reduced negative transfer in multi-task DNN training.

Conclusions:

• The proposed task grouping framework effectively addresses negative transfer in MTL for biological sequence analysis.
• This approach offers a scalable and superior alternative to existing MTL strategies.
• Enhanced DNN models can accelerate the discovery of biological mechanisms from genomic data.