Classification Models for Omics Data Usage Guide

Overview

Build diagnostic and prognostic classifiers on omics feature matrices with regularized logistic regression, random forest, and gradient-boosted trees. The discipline this skill enforces is that in the p>>n regime a regularized linear model is often the strongest baseline, the probability (not the label) is usually the product so calibration matters more than accuracy, and a suspiciously perfect AUC is most often a batch artifact rather than biology.

Prerequisites

pip install scikit-learn xgboost imbalanced-learn pandas

Conceptual prerequisites: know whether the deliverable is a probability or a hard label, whether batch is confounded with the outcome, and whether feature selection and scaling are inside the cross-validation loop (machine-learning/model-validation).

Quick Start

Tell your AI agent what you want to do: - "Build an elastic-net logistic classifier as a baseline on my expression data" - "Train XGBoost with early stopping and check whether it beats the linear model" - "Is my high AUC real, or is it learning the batch?" - "My classes are imbalanced -- should I use SMOTE?"

Example Prompts

Algorithm choice

"Build a classifier from my RNA-seq matrix. Start with regularized logistic regression and only escalate to trees if there is interaction signal it misses."

"Train an XGBoost classifier with a low learning rate, shallow depth, and early stopping; the dataset is small."

Batch shortcuts

"My cross-validated AUC is 0.97. Check whether the model is learning batch instead of biology by predicting the batch from the features and using a batch-aware split."

Imbalance and calibration

"My positive class is 8%. Handle the imbalance without destroying calibration, and tell me whether SMOTE is appropriate."

"Check whether my random forest probabilities are calibrated and recalibrate them if not."

Mixed features and missing data

"My features mix continuous expression and categorical genotype with missing values. Set up the preprocessing and a model that handles NaN natively."

What the Agent Will Do

Start with a regularized linear baseline (often the ceiling in p>>n)
Escalate to RF/XGBoost only when nonlinear or interaction signal is present, with early stopping
Run the batch-shortcut checks (predict the batch; leave-one-batch-out) before trusting any AUC
Handle imbalance by class weighting or threshold tuning, not resampling, for risk models
Check and recalibrate probabilities; report AUPRC and MCC under imbalance
Keep all preprocessing inside the CV fold and defer unbiased evaluation to model-validation

Tips

"Random forest is the obvious choice for expression data" is a myth; regularized logistic/linear SVM frequently win in p>>n (Statnikov 2008)
GBDTs beat deep nets on tabular/omics matrices (Grinsztajn 2022); reserve deep nets for very large n or raw/multimodal inputs
RF compresses probabilities toward 0.5, boosting pushes them to the extremes -- both need recalibration if the probability is used
SMOTE/oversampling destroys calibration for no AUC gain on risk models; class-weight or tune the threshold instead, and if you must resample use an imblearn Pipeline
XGBoost 2.x: set early_stopping_rounds and eval_metric in the constructor, not fit()
XGBoost/LightGBM handle missing values natively; in omics, missingness is often informative (below detection limit)
Trees are scale-invariant; standardize for logistic/SVM/kNN/PCA

machine-learning/model-validation - Nested CV, calibration, and net benefit for the trained classifier
machine-learning/biomarker-discovery - Select features before modeling (inside the CV fold)
machine-learning/prediction-explanation - Interpret the classifier and detect shortcuts with SHAP
machine-learning/survival-analysis - Time-to-event outcomes that classifiers cannot handle
differential-expression/batch-correction - Batch correction done design-aware, not across the split
expression-matrix/normalization - Per-sample normalization that is safe outside the CV fold