AI for Science: Building scientific foundation models - genomics edition

Abstract

This talk will examine how foundation models for genomic sequences are built, using Carbon as a case study. It will cover the key choices involved in curating genomic data, representing DNA as tokens, designing biologically appropriate training objectives, and evaluating models across generation, variant-effect prediction, perturbation, and long-context tasks. The talk will also consider why methods developed for natural language cannot be applied directly to genomes, and how open models, datasets, benchmarks, and cross-disciplinary collaboration can support progress in AI for genomics.

Bio

Georgia Channing leads the AI for Science team at Hugging Face, working at the intersection of machine learning and the natural sciences. She read for her Master’s and PhD in machine learning at the University of Oxford, with a focus on applying AI to scientific discovery. Her work has spanned a wide range of AI-for-science areas, including remote sensing, biophysics, and materials design. She now focuses on building open tools and models for genomics and electron microscopy.