Block selection in multiblock partial least squares for modeling genotype-phenotype relations in Saccharomyces

Area of Science:

• Genomics
• Systems Biology
• Biostatistics

Background:

• Correlations among explanatory variables in data-based modeling can form distinct gene blocks.
• Identifying influential gene blocks is crucial for understanding complex biological systems and genotype-phenotype relationships.
• Limited sample sizes pose challenges in traditional statistical modeling for gene analysis.

Purpose of the Study:

• To propose and evaluate a novel method, weighted block importance on projection in partial least squares (BwIP-mbPLS), for identifying influential gene blocks.
• To apply this method to genotype-phenotype mapping in Saccharomyces cerevisiae, focusing on modeling copper chloride and melibiose.
• To compare the performance of BwIP-mbPLS against traditional methods for predicting phenotypic variations.

Main Methods:

• Utilized partial least squares (PLS) with multiple blocks (mbPLS) to handle correlated variables and identify gene blocks.
• Developed BwIP-mbPLS to pinpoint influential gene blocks and variable importance on projection (VIP) for selecting key genes within blocks.
• Employed k-means clustering, silhouette index, and total distance metrics for gene classification into 18 blocks.

Main Results:

• Classified 5629 genes into 18 distinct gene blocks using k-means clustering.
• BwIP-mbPLS identified an average of 4 influential gene blocks, significantly improving efficiency-based phenotype prediction.
• The proposed methods consistently outperformed conventional PLS and mbPLS in phenotype prediction, with most identified blocks containing fewer than 10 genes.

Conclusions:

• BwIP-mbPLS is an effective method for identifying key gene blocks and variables in genotype-phenotype mapping, particularly for efficiency traits.
• The developed approach offers a robust solution for data-based modeling with high-dimensional, correlated data, even with limited sample sizes.
• These findings advance the potential of data-driven methods for dissecting complex genetic architectures and predicting phenotypes in model organisms.