OmicVerse Bulk Metabolomics — Multivariate & Biomarker

Goal

Given a preprocessed metabolomics AnnData (output of the omicverse-bulk-metabol-preprocessing skill), discriminate two groups with PLS-DA / OPLS-DA, rank features by VIP, inspect the OPLS-DA S-plot, then build and validate a biomarker panel via per-feature ROC AUC and nested-CV multi-metabolite classification with a permutation null. The output is a ranked feature list, model-quality metrics (R²X / R²Y / Q² / mean AUC / permutation p-value), and an actionable biomarker panel.

This skill is discrimination + biomarker selection only. Imputation / normalization / transformation belong in the preprocessing skill; pathway interpretation of the resulting hits belongs in the pathway-multifactor skill.

Quick Workflow

1. Confirm the input AnnData has been PQN-normalized + log-transformed + Pareto-scaled. Multivariate models on un-Pareto-scaled data are dominated by a few high-intensity features.
2. Fit ov.metabol.plsda(...) first as a baseline; inspect r2x, r2y, q2. A positive Q² is the minimum bar for a model that beats mean prediction.
3. Fit ov.metabol.opls_da(n_ortho=1) to separate group-discriminating signal (predictive component) from within-group variation (orthogonal components). Compare r2y / q2 against the PLS-DA baseline.
4. Pull the VIP table with result.to_vip_table(adata.var_names) and visualize the top features with ov.metabol.vip_bar(...).
5. Run ov.metabol.s_plot(...) to see covariance vs. correlation against the predictive component — features in the corners are robust biomarker candidates.
6. For biomarker selection: get per-feature univariate AUC + bootstrap CI with ov.metabol.roc_feature(...).
7. Build a multi-metabolite panel with ov.metabol.biomarker_panel(...) using nested CV; include a permutation null to test whether the panel beats random feature selection.
8. Plot the held-out ROC curve from panel.outer_predictions / panel.outer_labels for the report.