Log transformation with log1p()

Step 3: Log transformation

feature_cols <- OmicsProcessing::resolve_target_cols(clean_df, "@")
log_transformed_df <- clean_df %>%
  dplyr::mutate(dplyr::across(
    .cols = tidyselect::all_of(feature_cols),
    .fns = ~ log1p(.x),
    .names = "{.col}"
  ))

Apply a natural log transform with a +1 offset across all feature columns. resolve_target_cols() lets you declare the feature set via names, tidyselect helpers, or regex (e.g., "^met_", "feature[0-9]+", or "@" to use all feature columns). The resulting log1p scaling often reduces skewness, stabilizes variance, and moderates the influence of extreme values without discarding zeros.

Why use log1p?

• Handles zeros gracefully (no need to add a pseudo-count manually).
• Produces smoother distributions that are friendlier for downstream modeling and distance-based methods.
• Keeps column naming intact via .names = "{.col}", so you can swap the transformed data back into your pipeline without additional renaming.

Tips for choosing feature columns

• Prefer explicit targets for clarity: e.g., resolve_target_cols(clean_df, c("met_a", "met_b")).
• Regex or tidyselect helpers are useful for wide matrices: resolve_target_cols(clean_df, tidyselect::starts_with("met_")).
• Ensure feature columns are numeric; convert factors/characters before transformation to avoid unintended NAs.