Adjusting for covariates and assessing modeling fitness in machine learning using MUVR2

Area of Science:

• Omics research
• Bioinformatics
• Computational biology

Background:

• Machine learning (ML) is integral to Omics research for analyzing molecular data and identifying associations with exposures and health.
• Existing ML methods often lack robust frameworks for assessing model fitness and adjusting for covariates, hindering biological interpretation.
• The previous MUVR algorithm demonstrated state-of-the-art performance in prediction and variable selection.

Purpose of the Study:

• To introduce the MUVR2 framework, an advancement of the MUVR algorithm.
• To address the limitations of existing ML approaches by incorporating model fitness assessment and covariate adjustment capabilities.
• To enhance the utility of ML in Omics research for deeper biological insights.

Main Methods:

• MUVR2 integrates the elastic net regularized regression framework alongside partial least squares and random forest models.
• The framework employs advanced cross-validation strategies to ensure state-of-the-art performance and minimize overfitting.
• Covariate adjustment is enabled within the elastic net modeling component of MUVR2.

Main Results:

• MUVR2 consistently achieved state-of-the-art performance in both prediction and variable selection across various cross-validation strategies.
• The framework effectively minimizes overfitting, leading to more reliable model assessments.
• MUVR2 successfully demonstrated the capability for covariate adjustment using elastic net, a feature not available with partial least squares or random forest in this context.

Conclusions:

• MUVR2 provides a significant advancement for ML applications in Omics research, offering improved model evaluation and covariate adjustment.
• The open-source availability of MUVR2, including algorithms, data, and tutorials, promotes wider adoption and reproducibility in the scientific community.
• This framework facilitates more robust biological interpretation by enabling ML models to account for confounding factors.