MOC - AI in Microbiology

Artificial intelligence and machine learning applied to microbes — structure prediction, diagnostics, resistance, discovery, and surveillance.

Parent: Home · Map: Encyclopedia Map
Companion: MOC - Bioinformatics in Microbiology (the data layer underneath) · Media: Learning Media Hub

Overview

AI in microbiology learns patterns from large biological datasets — sequences, images, spectra, electronic records — to predict structures, identify organisms, infer resistance, design molecules, and detect outbreaks. It sits on top of bioinformatics outputs, and never replaces wet-lab confirmation (Antimicrobial Susceptibility Testing, culture).

flowchart LR
  Data[Sequences / images / spectra / AST labels] --> BI[[MOC - Bioinformatics in Microbiology]]
  BI --> Feat[Features / embeddings]
  Feat --> AI[ML / DL models]
  AI --> Out[Predictions]
  Out --> Eval[[Model Evaluation in Clinical Microbiology]]
  Eval --> Val[Wet-lab / clinical validation]
  Val --> Data

1. Methodological Foundations

1b. Algorithm Atlas (types + how to use)

2. Structure, Proteins, Design

3. Diagnostics

4. Resistance and Therapy (bacteria-focused)

5. Prevention and Population Level

6. Frontier

Core Concepts to Internalize

  • Data quality dominates model choice. Label noise from imperfect AST ceilings performance.
  • Population structure is the microbiology-specific confounder — models learn lineages, not mechanisms (Population Structure Confounding in Microbial ML).
  • Calibration and error types matter more than AUC — very major errors are the currency of clinical acceptance.
  • Explainability — clinicians act on reasons, not scores.
  • Human-in-the-loop — AI proposes, laboratory confirms, clinician decides.
  • Drift — pathogens and breakpoints change; models decay silently.
  • MIC ≠ persistence — tolerance phenotypes need different labels and models.

Data & Resources

ResourceUse
AlphaFold DB, PDBStructures for design and mechanism
UniRef / UniProtPretraining protein language models
CARD / ResFinder / AMRFinderPlusAMR labels and features (AMR Gene Databases)
NCBI Pathogen Detection, EnteroBaseGenomes + metadata at scale
PATRIC/BV-BRCGenome–phenotype pairs
MIMIC / local EHR (governance!)Clinical outcome models — privacy critical
Public AST collections (e.g., CRyPTIC for TB)Benchmarks for genotype→phenotype

Important Methods & Inputs

Important Papers

  • Paper - AMR Database M.Centner 2026 — database quality limits AI labels
  • Add: AlphaFold2 (Jumper 2021); RoseTTAFold/RFdiffusion; halicin (Stokes 2020) and abaucin; ESM-2/ESMFold; TRIPOD+AI

Research Questions

  1. When is genomic ML accurate enough to replace phenotypic AST, and for which drug–bug pairs?
  2. How do we detect model failure on novel plasmids or unseen species?
  3. What governance is required before AI reads clinical Gram stains autonomously?
  4. Can protein language models prioritize truly novel resistance determinants prospectively?
  5. Do AI stewardship tools change patient outcomes, not just prescribing metrics?

Review Article Opportunities

  • Structure prediction → antimicrobial discovery: what actually reached the bench
  • Clinical validation checklist for AMR prediction models
  • AI for hard phenotypes: Biofilm, persistence, tolerance
  • LLMs in the clinical microbiology laboratory: realistic scope

Learning Aids

Build Status

ClusterStatus
ML/DL foundations + evaluation + ethicsdone
Algorithm atlas (supervised/unsupervised/CNN/transformer/GNN/generative/transfer + features)✅ 2026-08-02
Structure, protein LMs, designdone
Diagnostics and image AIdone
AMR, stewardship, discoverydone
Bacteria-specific confounders + plasmid-host ML✅ 2026-08-02
Biofilm / persistence AI
DNA genome LMs + agentic workflows
Vaccines, outbreak detectiondone
Foundation models / LLMsdone
Hands-on notebooks with real databacklog