Transfer Learning in Multiple Hypothesis Testing.

Area of Science:

• Computational Statistics
• Machine Learning
• Genomics

Background:

• Multiple Hypothesis Testing (MHT) is a complex statistical challenge with traditional methods facing limitations.
• Existing MHT approaches often struggle with large-scale datasets and complex dependencies.
• The integration of advanced machine learning with statistical inference offers potential for novel solutions.

Purpose of the Study:

• To develop a novel approach for Multiple Hypothesis Testing (MHT) by synthesizing Convolutional Neural Networks (CNNs) and Bayesian inference.
• To introduce a sequence-based uncalibrated Bayes factor approach for testing numerous hypotheses within parametric models.
• To demonstrate the utility of this CNN-Bayesian framework in complex data analysis, particularly in genomics.

Main Methods:

• A two-step methodology involving a learning phase with simulated data and a transfer phase with real-world experimental sequences.
• Utilizing Convolutional Neural Networks (CNNs) trained on diverse null and alternative hypotheses.
• Employing a sequence-based uncalibrated Bayes factor for hypothesis evaluation.

Main Results:

• Developed a CNN model that significantly enhances precision in Multiple Hypothesis Testing (MHT) compared to traditional methods.
• Demonstrated robustness of the CNN-based MHT approach under varying conditions, including the number of true null hypotheses and test dependencies.
• Empirical evaluations indicate the methodology's potential usefulness, particularly in genomic applications.

Conclusions:

• The synthesis of CNNs and Bayesian inference presents a powerful and precise new methodology for Multiple Hypothesis Testing (MHT).
• The approach shows promise for complex sequence analysis and has significant potential applications in genomics.
• Further theoretical evaluation is needed, but initial results advocate for continued exploration of this innovative technique.