Supervised Learning Algorithms in Microbiology
One-Sentence Definition
Supervised learning trains a model on input–label pairs so that new microbial samples can be assigned phenotypes, taxa, or clinical outcomes — the workhorse paradigm for AMR prediction, species ID, and diagnostic triage.
Simple Explanation
You show labeled examples (“this genome was ceftriaxone-resistant”) until the model can guess the label for new samples.
Detailed Scientific Explanation
| Algorithm | Best when | Micro example | Watch out |
|---|---|---|---|
| Logistic / linear (regularized) | Need baseline + interpretability | AMR gene → R/S odds | Underfits complex interactions |
| k-NN | Local similarity, small spectral sets | MALDI species match | Sensitive to scaling; slow at scale |
| SVM (linear/RBF) | High-dim, modest n | Spectra; older gene-panel AMR | Kernel choice; calibration weak |
| Decision tree | Teaching / simple rules | Rare as final model | Unstable alone |
| Random forest / boosting | Tabular genomics winner | Gene matrices → AMR | See Tree Ensembles in Microbiology |
| Naive Bayes | Fast text-like bags of features | k-mer bag baselines | Strong independence assumption |
| Neural nets (MLP) | Large n, nonlinear tabular | Sometimes AMR; often loses to boosting | Overfit on small cohorts |
Task types:
- Classification — R/S, species, outbreak yes/no
- Ordinal / multiclass — S/I/R; serotype
- Regression — log₂MIC (harder; breakpoint edges matter)
How to use (recipe)
- Fix labels and Feature Representation for Microbial ML
- Train a linear baseline first (Linear and Kernel Models in Microbiology)
- Train RF/GBM; compare with honest splits
- Inspect errors by lineage and drug (Population Structure Confounding in Microbial ML)
- Calibrate probabilities before clinical thresholds (Model Evaluation in Clinical Microbiology)
Clinical Importance
- Supervised models are what labs actually evaluate against AST — unsupervised tools do not give R/S alone
Research Importance
- Benchmarks (e.g., CRyPTIC TB) standardize supervised AMR challenges
Diagnostic Relevance
- Closed supervised systems power automated plate readers and spectral libraries
AMR Relevance
- Default path for Machine Learning for AMR Prediction and Genotype to Phenotype Prediction
Related MOCs
Active Recall Questions
- Why start with logistic regression before XGBoost?
- Classification vs MIC regression — which aligns better with clinical categories?
- Name two supervised uses that are not AMR.