Forum Foundation Models Roundtable: Lopez and Hie
Jul 22, 2025
Brian Hie, a Stanford chemical engineering professor, and Marie Lopez, Head of Genomics AI at InstaDeep, dive into the fascinating world of foundation models. They discuss their applications in analyzing DNA and protein structures, and the challenges of training data biases. Brian highlights the importance of evolutionary priors, while Marie shares insights on using genomic 'grammar' for innovative discoveries. The duo also tackles the limitations and risks of deploying these models in clinical settings, emphasizing the need for validation and transparency.
AI Snips
Chapters
Transcript
Episode notes
How Foundation Models Learn Biology
- Foundation models learn broad patterns by reconstructing large, diverse biological datasets rather than training on task-specific labels.
- This enables cheaper specialization to many downstream tasks via transfer learning.
From Early Protein Models To Genome Models
- Brian described early protein language modeling work from his time at MIT and Facebook AI Research, leading to ESM models.
- His lab extended that work toward genome-level models trained directly on DNA sequence.
Evolution As A Training Experiment
- Marie recounted applying language-model ideas to evolutionary genomic data at InstaDeep, producing the Nucleotide Transformer.
- She framed evolution as a long-running experiment the models can learn from across species variation.