Foundation Models and LLMs in Microbiology
One-Sentence Definition
Foundation models are large models pretrained on broad data (text, sequences, images) and adapted to many downstream microbiology tasks, including large language models applied to literature, records, and laboratory reasoning.
Simple Explanation
One large pretrained model, many uses — from reading the literature to interpreting a genome, instead of a separate small model for each question.
Detailed Scientific Explanation
Model families relevant here:
- Text LLMs — literature synthesis, guideline question answering, structuring free-text microbiology reports, coding assistance for pipelines
- Biological sequence foundation models — Protein Language Models; DNA/genome models (Nucleotide Transformer, Evo-style genome-scale models) for regulatory and mobile-element reasoning
- Multimodal — image + text for microscopy interpretation and report generation
- Agentic use — LLMs orchestrating bioinformatics tools, writing and running workflow code
Serious limitations for clinical/scientific use:
- Hallucination of references, gene names, breakpoints — every factual claim needs verification
- Training cutoff vs rapidly changing breakpoints (EUCAST/CLSI updates)
- Privacy: patient data must not leave approved environments
- Benchmarks poorly reflect real laboratory reasoning
- Retrieval-augmented generation (RAG) grounded in curated sources reduces but does not eliminate error
Mechanism
Self-supervised pretraining at scale → emergent transfer; adaptation via fine-tuning, instruction tuning, or retrieval + prompting.
Clinical Importance
- Realistic near-term wins are documentation, triage of literature, and decision support drafts reviewed by humans — not autonomous diagnosis
Research Importance
- Hypothesis generation, systematic review acceleration, code for Reproducible Bioinformatics Workflows
Diagnostic Relevance
- Report structuring and coding; extraction of phenotype labels for ML datasets
AMR Relevance
- Mining surveillance literature; drafting stewardship summaries; interpreting resistance genotypes with retrieval from AMR Gene Databases
Related MOCs
Active Recall Questions
- Give two failure modes of LLMs in clinical microbiology.
- What is RAG and why does it help?
- Which tasks are realistic near-term uses?