Generalizing predictions to unseen sequencing profiles via deep generative models

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Predictive models trained on sequencing data often fail external validation due to distribution shifts.
• Factors like batch effects and small datasets hinder model generalization across studies.
• Generalizing predictive models without prior knowledge of unseen data is a significant challenge.

Purpose of the Study:

• To propose DeepBioGen, a novel sequencing profile augmentation procedure.
• To enhance the generalizability of predictive models on unseen data.
• To improve cross-study prediction performance in biological and medical applications.

Main Methods:

• DeepBioGen characterizes visual patterns in sequencing profiles.
• It employs a deep generative model to create realistic synthetic profiles.
• The generated profiles are used to train and generalize subsequent classifiers.

Main Results:

• DeepBioGen significantly outperforms existing methods in enhancing model generalizability.
• The generalized classifiers achieved superior performance compared to state-of-the-art methods.
• Successful validation on RNA sequencing tumor expression and WGS human gut microbiome data.

Conclusions:

• DeepBioGen offers an effective strategy for improving the robustness of predictive models.
• The method enhances cross-study prediction accuracy for critical applications.
• DeepBioGen represents a significant advancement in bioinformatics and machine learning for biological data.