Overview
Supervised learning uses a dataset of input-output pairs, where the correct answer is known, to train a model that can predict the output for new, unseen inputs. In bioinformatics, this paradigm maps naturally onto problems such as classifying a tissue sample as cancerous or healthy, predicting whether a genetic variant is pathogenic, or assigning a function to an uncharacterized protein. The model learns decision boundaries that separate classes in feature space, guided by a loss function that penalizes prediction errors.
Methods
Logistic regression models the probability of class membership and remains a baseline for binary classification, especially when interpretability is valued. Support vector machines construct hyperplane separators with maximum margin and perform well in high-dimensional settings typical of omics data. Random forests aggregate many decision trees trained on bootstrapped samples and provide built-in feature importance measures. Gradient-boosted trees (XGBoost, LightGBM) iteratively correct predecessor errors and often win structured-data competitions. Neural networks with hidden layers can learn complex non-linear boundaries but require more data and careful regularization. All methods benefit from class-weight adjustment or resampling when classes are imbalanced.
Practical Protocol
A practical supervised learning workflow for gene expression classification begins with loading the expression matrix, where rows represent genes and columns represent patient samples. The data is first normalized using log-transformed counts per million (log-CPM) or quantile normalization. A typical dataset might contain 20,000 gene features across 100 samples, requiring careful validation. The researcher splits the data into training (70%), validation (15%), and test (15%) sets while preserving class proportions through stratification. A random forest classifier with 500 trees is trained on the training set. Five-fold cross-validation on the training set assesses stability: the model is fitted on four folds and evaluated on the held-out fold, repeated five times, yielding an average cross-validated accuracy. After confirming that cross-validation performance is acceptable, the model is retrained on the full training set and evaluated once on the held-out test set. Feature importance scores from the trained forest reveal which genes most strongly discriminate the classes. The top 20 ranked genes are inspected against known biological pathways via KEGG or Reactome enrichment analysis to verify mechanistic plausibility. This workflow has been used to develop multi-gene prognostic signatures for breast cancer subtypes, where a panel of 50 genes stratified patients into risk groups with significantly different survival outcomes. The same approach applies to classifying drug response phenotypes from pharmacogenomic datasets and predicting pathogenic variants from genomic features.
Applications
Supervised learning classifies bacterial pathogens using genomic markers in bacterial genetics, predicts patient prognosis in cancer biochemistry, and identifies differentially expressed genes from DNA microarrays and gene expression data. It also powers clinical decision support by stratifying patients into risk groups and annotating regulatory elements across the genome.
