Transfer learning of classification rules for biomarker discovery and verification from molecular profiling studies

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning in Biology

Background:

• Biomarker discovery often involves analyzing multiple related datasets.
• Integrating knowledge from existing datasets can enhance the learning process for new data.
• Current methods may not fully leverage prior information from source datasets.

Purpose of the Study:

• To present a novel framework for integrative biomarker discovery.
• To incorporate prior knowledge, in the form of interpretable rules, into the learning process for new datasets.
• To evaluate two methods of knowledge transfer: whole-rule transfer and rule-structure transfer.

Main Methods:

• Developed a framework that integrates prior knowledge (interpretable, modular rules) into the learning process on new datasets.
• Implemented two knowledge transfer methods: transfer of whole rules and transfer of rule structures.
• Evaluated the framework on three dataset pairs (one genomic, two proteomic) using standard classification performance measures and novel transfer amount measures.

Main Results:

• Whole-rule transfer demonstrated improved classification performance compared to using the target data alone.
• Performance gains were particularly notable when the source dataset was larger than the target dataset.
• The proposed transfer method also outperformed using the union of source and target datasets.

Conclusions:

• The novel framework effectively integrates prior knowledge for enhanced biomarker discovery.
• Whole-rule transfer is a promising method for improving classification performance in integrative studies.
• This approach offers advantages over traditional methods, especially in scenarios with varying data sizes.