AI Models Could Represent DNA as Embeddings, Experts Suggest
TL;DR. Hacker News discussion explores how AI models might represent DNA as embeddings, enabling biological simulations. - The concept involves modeling DNA strands or entire genomes similar to how language models process text. - This approach could create AI 'sandboxes' for genetics, allowing for new protein or gene design. - Researchers confirm foundation models like Evo and Genomics-FM already treat DNA as language.
- The idea proposes representing DNA sequences as numerical embeddings for AI processing.
- Two main approaches are discussed: modeling individual DNA strands/genes or entire genomes.
- Potential applications include simulating genetic outcomes, inventing new genes, and testing drugs.
- Experts confirm that foundation models currently exist that treat DNA like language for protein design.
Sources
- Ask HN: Could DNA be represented as "an embedding" in an AI model? — news.ycombinator.com