Generative Models in Microbiology

One-Sentence Definition

Generative models learn a distribution over sequences, structures, or molecules and can sample new candidates — used for antimicrobial peptide/protein design, data augmentation, and exploratory latent-space biology.

Simple Explanation

Discriminative models answer “what is this?” Generative models answer “invent something like the training set” — a new peptide, backbone, or plausible spectrum.

Detailed Scientific Explanation

FamilyIdeaMicro use
VAE / latent generativeEncode → sample latent → decodeMicrobiome/profile synthesis; exploration
GANs (less dominant now)Generator vs discriminatorHistorical image/spectrum augmentation
Autoregressive LMsNext-token samplingProtein/DNA sequence design
Diffusion / denoisingIterative noise removalBackbone/molecule design (Protein Design for Antimicrobials)
Flow / energy-basedInvertible or energy samplingEmerging molecular design

How to use

  1. Define constraints: length, charge, toxicity filters, synthesizability
  2. Generate many candidates → rank with property predictors (Graph Neural Networks in Microbiology, docking, LM scores)
  3. Synthesize/test a shortlist — generative output is a hypothesis
  4. For images: use augmentation carefully; synthetic Gram stains must not train circular evaluations

Mechanism

Approximate (p(data)) or a conditional (p(data\mid\text{prompt/constraints})); sampling yields novel instances.

Clinical Importance

  • Far upstream of diagnostics; relevant to drug/vaccine research pipelines

Research Importance

Diagnostic Relevance

  • Limited; synthetic data for training must be labeled/validated or it amplifies bias

AMR Relevance

Active Recall Questions

  1. Discriminative vs generative — which predicts MIC from a genome?
  2. Why must generative antibiotic designs still go through wet lab?
  3. Name two generative families used in protein design.

Connections