OMGenomics Podcast 011: AI in biology: distinguishing hype from reality
Aug 13, 2025
Valerie de Crissy-Lagard, a seasoned scientist specializing in enzyme function prediction, and Rachel Thomas, co-founder of Fast AI and an AI expert pursuing a PhD in immunology, discuss the promising yet precarious role of AI in biological research. They delve into a case where AI's enzyme predictions fell short and emphasize the dire need for collaboration between machine learning experts and biologists. With insights on automation bias and the intricacies of genomic annotation, they highlight the importance of a multidisciplinary approach to enhance research quality.
AI Snips
Chapters
Transcript
Episode notes
Data And Competitions Power Breakthroughs
- Algorithmic breakthroughs like AlphaFold depended on decades of curated data and competitions, not just model innovation.
- Foundational datasets and benchmarks are prerequisite catalysts for major AI wins.
Biologically Impossible AI Predictions
- AI annotated some E. coli unknowns with biologically impossible functions like mycothiol synthase.
- Valerie spotted these errors instantly using basic metabolic knowledge and pathway context.
Validate Annotations With Biological Context
- Cross-check predicted functions against pathway context and presence of pathway genes to detect biological impossibilities.
- Use metabolic models and gene-neighborhood information to validate annotations before accepting them.