Graph-Based Bidirectional Transformer Decision Threshold Adjustment Algorithm for Class-Imbalanced Molecular Data.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Imbalanced data sets, common in biological applications like disease diagnosis and drug discovery, pose challenges for standard classification methods.
• Failure to accurately identify underrepresented classes can lead to significant costs and missed opportunities.

Purpose of the Study:

• To develop an advanced algorithm for robust data classification on highly imbalanced molecular data sets.
• To enhance the identification of elements belonging to underrepresented classes in biological data.

Main Methods:

• The proposed BTDT-MBO algorithm combines Merriman-Bence-Osher (MBO) approaches with a bidirectional transformer architecture.
• It incorporates distance correlation for a similarity graph-based framework and adjusts decision thresholds to manage class imbalance.
• Self-supervised learning via an attention mechanism is utilized within the bidirectional transformer.

Main Results:

• The BTDT-MBO algorithm demonstrated superior performance compared to existing techniques on six diverse molecular data sets.
• The method effectively handled high class imbalance ratios, outperforming competing approaches.
• Validation confirmed the algorithm's efficacy in identifying underrepresented molecular data elements.

Conclusions:

• The BTDT-MBO algorithm offers a significant advancement for classifying imbalanced molecular data.
• Its integrated approach addresses key limitations of existing methods, particularly in biological and medical research.
• This technique holds promise for improving accuracy in critical applications such as disease diagnosis and drug discovery.