Towards the genome-scale discovery of bivariate monotonic classifiers

Area of Science:

• Computational biology
• Machine learning
• Genomics

Background:

• Bivariate Monotonic Classifiers (BMCs) are interpretable machine learning models capable of capturing nonlinear patterns in high-dimensional data.
• Previous BMC applications were limited by high computational complexity in feature pair selection.
• The genome-scale application of BMCs was hindered by the computational cost of leave-one-out performance estimation.

Purpose of the Study:

• To introduce a computationally efficient algorithm for identifying BMCs.
• To enable large-scale BMC analysis for improved classification performance and biomarker discovery.
• To provide an open-source implementation for reproducible research.

Main Methods:

• Developed the fastBMC algorithm, leveraging a mathematical bound for BMC performance estimation.
• Empirically evaluated fastBMC's speedup compared to traditional methods.
• Applied fastBMC to biomedical datasets, including glioblastoma and breast cancer, for performance assessment.

Main Results:

• fastBMC achieves a speedup factor of at least 15 for BMC identification compared to traditional approaches.
• Improved classification performance was observed on smaller biomedical datasets by enabling analysis of larger feature sets.
• Demonstrated the interpretability of BMCs through a glioblastoma survival predictor, leading to novel hypotheses.

Conclusions:

• fastBMC facilitates rapid construction of robust and interpretable ensemble models using BMCs.
• The algorithm accelerates the discovery of gene pairs predictive of phenotypes and their interactions.
• Enabled biomarker identification and biomedical hypothesis generation in cancer datasets.