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

AlgorithmBest whenMicro exampleWatch out
Logistic / linear (regularized)Need baseline + interpretabilityAMR gene → R/S oddsUnderfits complex interactions
k-NNLocal similarity, small spectral setsMALDI species matchSensitive to scaling; slow at scale
SVM (linear/RBF)High-dim, modest nSpectra; older gene-panel AMRKernel choice; calibration weak
Decision treeTeaching / simple rulesRare as final modelUnstable alone
Random forest / boostingTabular genomics winnerGene matrices → AMRSee Tree Ensembles in Microbiology
Naive BayesFast text-like bags of featuresk-mer bag baselinesStrong independence assumption
Neural nets (MLP)Large n, nonlinear tabularSometimes AMR; often loses to boostingOverfit 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)

  1. Fix labels and Feature Representation for Microbial ML
  2. Train a linear baseline first (Linear and Kernel Models in Microbiology)
  3. Train RF/GBM; compare with honest splits
  4. Inspect errors by lineage and drug (Population Structure Confounding in Microbial ML)
  5. 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

Active Recall Questions

  1. Why start with logistic regression before XGBoost?
  2. Classification vs MIC regression — which aligns better with clinical categories?
  3. Name two supervised uses that are not AMR.

Connections