Compositional transformations can reasonably introduce phenotype-associated values into sparse features.

Area of Science:

• Microbiome bioinformatics
• Computational biology
• Statistical analysis of high-dimensional data

Background:

• Gihawi et al. questioned the validity of tumor microbiome analyses by Poore et al., citing that sparse features (genera with few reads) became phenotype-associated after batch correction.
• This critique implies that such transformations indicate information leakage and invalidate the analysis, impacting the interpretation of The Cancer Genome Atlas (TCGA) microbiome studies and broader microbiome research.

Purpose of the Study:

• To investigate whether the emergence of phenotype-associated sparse features after data transformation necessarily signifies information leakage or processing errors in microbiome analyses.
• To provide counterexamples demonstrating that such observations can result from valid statistical transformations, challenging the broad invalidation claims.

Main Methods:

• Examined the centered log ratio (CLR) transformation, a common method for compositional microbiome data, noting its sample-wise nature and similarities to batch correction methods.
• Utilized synthetic and vaginal microbiome datasets to demonstrate how CLR transformation, coupled with imputation strategies, can associate sparse features with the geometric mean and, consequently, the phenotype.
• Re-analyzed specific features highlighted by Gihawi et al. to show that the observed phenomenon can occur even after a CLR transformation, serving as a counterexample to information leakage claims.

Main Results:

• The centered log ratio (CLR) transformation, a sample-wise operation, cannot inherently leak information or invalidate downstream analyses.
• Common imputation methods for zero or missing values in CLR-transformed data can lead to associations between transformed features and the sample's geometric mean.
• When the geometric mean is phenotype-associated, sparse and CLR-transformed features also become associated with the phenotype, a phenomenon observed in both synthetic and real microbiome data.
• Re-analysis confirmed that sparse features becoming phenotype-associated can occur after CLR transformation, refuting the claim that this observation alone indicates information leakage.

Conclusions:

• The appearance of phenotype-associated sparse features after data transformation is not sufficient evidence to claim information leakage in machine learning pipelines.
• Sample-wise transformations like CLR can generate such associations without artificially inflating performance, suggesting that the original critique may be overly broad.
• Interpreting individual features in microbiome data requires caution due to the multivariate nature of the data and the impact of transformations and batch correction methods.