AI Algorithms in Microbiology

One-Sentence Definition

An algorithm atlas for microbiology: which machine-learning and deep-learning families to choose for a given data type, clinical question, and validation constraint — and which traps make the choice irrelevant.

Simple Explanation

Pick the tool that matches your data shape (table, image, sequence, graph), not the newest name. In microbiology, how you split and label beats which algorithm almost every time.

Decision Guide (start here)

Your data	Typical question	First algorithms to try	Deep notes
Gene presence/absence, k-mers, SNPs (tabular)	Species, AMR, virulence class	Tree Ensembles in Microbiology → regularized linear baseline	Supervised Learning Algorithms in Microbiology
MALDI / FTIR peak tables	Species ID, resistance flag	k-NN, SVM, RF	Linear and Kernel Models in Microbiology
Colony / Gram / smear images	Organism class, parasite detect	Convolutional Neural Networks in Microbiology	Digital Microscopy and Image AI
Protein or DNA sequences	Function, AMR motif, variant effect	Transformers and Attention in Microbiology · LMs	Protein Language Models · DNA and Genome Language Models
Molecules / assembly graphs / networks	Antibiotic hit, plasmid links	Graph Neural Networks in Microbiology	AI for Antibiotic Discovery
Unlabeled genomes / microbiome	Structure, clusters, outliers	Unsupervised Learning in Microbiology	Population Structure and Clustering
Small labeled set + large unlabeled	Rare phenotypes	Transfer Learning in Microbiology · self-supervised pretrain	Foundation Models and LLMs in Microbiology

Always run the workflow in Figure - Machine Learning Workflow in Microbiology and judge with Model Evaluation in Clinical Microbiology.

Algorithm Families (catalog)

A. Learning paradigms

Paradigm	Idea	Microbiology use
Supervised	Learn from labeled examples	AMR R/S, species from spectra, sepsis risk
Unsupervised	Find structure without labels	Strain clusters, microbiome ordination
Semi- / self-supervised	Pretrain cheap, fine-tune scarce labels	Protein/DNA LMs → phenotype heads
Transfer learning	Reuse a model trained elsewhere	ImageNet→plates; ESM→AMR; foundation→local lab
Reinforcement	Learn by sequential reward	Rare: adaptive experimental design, robotic labs
Generative	Sample new sequences/structures	De novo peptides, backbone design

B. Classical supervised (tabular)

See Supervised Learning Algorithms in Microbiology · Linear and Kernel Models in Microbiology · Tree Ensembles in Microbiology.

C. Unsupervised

See Unsupervised Learning in Microbiology.

D. Deep architectures

See Deep Learning in Microbiology · Convolutional Neural Networks in Microbiology · Transformers and Attention in Microbiology · Graph Neural Networks in Microbiology · Generative Models in Microbiology.

Feature Representation (the real “how to use”)

Before choosing an algorithm, choose features — see Feature Representation for Microbial ML:

Modality	Common features
Bacterial WGS	k-mers; unitig/DBG; gene/allele presence; SNPs vs reference; embeddings from DNA LMs
AMR task	Detected genes (AMR Gene Databases) + mutations + plasmid markers
Microbiome	ASV/OTU counts, CLR-transformed abundances, pathway scores
Images	Raw pixels / patches (CNN); rarely hand-crafted morphology scores
Spectra	Peak lists, binned intensities, learned embeddings

Practical Playbooks

Playbook 1 — Genomic AMR classifier

1. Labels from phenotypic Antimicrobial Susceptibility Testing (document method/breakpoint)
2. Features: gene presence ± key mutations (Machine Learning for AMR Prediction)
3. Baseline: L2 logistic regression → Tree Ensembles in Microbiology (XGBoost/LightGBM)
4. Splits: hold out lineages / hospitals / years (Population Structure Confounding in Microbial ML)
5. Report very major errors, not only AUC

Playbook 2 — Digital microscopy

1. Define label source (expert read / culture confirmation)
2. Train Convolutional Neural Networks in Microbiology with heavy augmentation
3. External validate across microscopes/stains
4. Human-in-the-loop triage first (AI Diagnostics in Microbiology)

Playbook 3 — Outbreak / population structure

1. Prefer epidemiological methods (MLST and cgMLST, SNP trees) as ground truth
2. Use unsupervised embeddings/clustering as hypothesis tools (Unsupervised Learning in Microbiology)
3. Do not treat a UMAP cluster as a legal outbreak definition

Playbook 4 — Molecule discovery

1. Represent compounds as graphs/SMILES
2. Property models + generative design (Graph Neural Networks in Microbiology · Generative Models in Microbiology)
3. Wet-lab confirmation is the product (AI for Antibiotic Discovery)

What Not to Do

• Random train/test split on clonal bacteria → inflated accuracy
• Deep learning on 200 tabular genomes when RF suffices
• Deploy without calibration and drift monitoring
• Confuse MIC prediction with persistence/biofilm phenotypes (AI for Biofilm and Persistence Phenotypes)

Related MOCs

• MOC - AI in Microbiology · MOC - Bioinformatics in Microbiology

Active Recall Questions

1. For gene presence/absence AMR tables, which family is usually the first strong model?
2. When do CNNs clearly beat tree ensembles in the lab?
3. Why can unsupervised clustering mislead outbreak response?

Connections

• Machine Learning Basics for Microbiology · Deep Learning in Microbiology · Computational Microbiology Study Path