# HAI Seminar with Brian Hie: Genome Modeling & Design Across All Domains of Life Source: https://www.youtube.com/watch?v=ExEoyu14n_g Recap page: https://rapidrecap.app/video/ExEoyu14n_g Generated: 2025-10-13T19:02:23.808+00:00 --- ## Quick Overview Dr. Brian He introduced EVO 2, a genomic foundation model trained on 9.3 trillion DNA base pairs across all domains of life, demonstrating its capability to generate novel, functional biological sequences like CRISPR systems and complete bacterial phage genomes that are significantly divergent from natural counterparts. **Key Points:** - Dr. He introduced EVO 2, a genomic foundation model scaled to 40 billion parameters, trained on 9 trillion DNA base pairs from bacteria, archaea, and eukaryotes, allowing it to process input sequences up to one million base pairs long. - The team successfully used supervised fine-tuning on EVO 1 to generate a functional CRISPR-Cas system whose AI-generated Cas9 protein cleaved double-stranded DNA in vitro as effectively as the state-of-the-art SpCas9, achieving less than 80% sequence identity to anything in nature. - Using prompt engineering with EVO, researchers generated evolutionarily novel anti-CRISPR proteins, experimentally confirming their activity, with some generated sequences showing no significant sequence identity above random null to any gene observed in nature. - Inference time guidance allowed the model to write Morse code patterns into synthetic chromatin accessibility in mouse embryonic stem cells, achieving good agreement between computationally predicted designs and experimental ATAC-seq results. - By applying pre-training, fine-tuning, prompt engineering, and inference guidance, the team designed viable, complete genomes for bacteriophage $\Phi$X174 variants; one generated phage showed less than 95% sequence identity to any phage in nature and qualified as a new species under some taxonomic thresholds. - The generated phage cocktail rapidly overcame bacterial resistance when tested against E. coli C strains resistant to the wild-type $\Phi$X174, demonstrating resilience potential for phage therapy. - Dr. He stressed that computational bottlenecks are currently less severe than experimental bottlenecks, noting that synthesizing and testing complete, large genomes remains the biggest challenge for achieving whole-organism design. **Context:** The seminar, presented by Dr. Brian He, an Assistant Professor at Stanford and investigator at Arc Institute, focused on developing genomic foundation models at the intersection of biology and machine learning. The work aims to move beyond merely reading, writing, and editing DNA sequences, addressing Francis Arnold's observation that current technology cannot yet compose entirely new, functional DNA sequences, which is difficult because biology operates in an unintuitive language. ## Detailed Analysis Dr. He detailed a multi-faceted approach to advancing biological sequence generation, starting with unsupervised pre-training via DNA language models like EVO 1 (prokaryotes) and the expanded EVO 2 (all domains of life), which learn grammar and semantics by predicting the next DNA base pair, mirroring how large language models learn world facts from text. The team then employed supervised fine-tuning to specialize these generalist models, successfully generating novel, working CRISPR-Cas systems and entirely new anti-CRISPR proteins exhibiting high evolutionary novelty. Furthermore, they utilized inference time guidance, coupling EVO 2 with external scoring functions like Informer/Borsa models, to synthetically engineer specific epigenetic patterns, such as writing Morse code into chromatin accessibility in mouse DNA, which was validated experimentally. The culmination of these techniques enabled the design, synthesis, and experimental validation of complete, viable bacteriophage $\Phi$X174 genomes, some of which were highly novel; these AI-generated phages demonstrated practical utility by rapidly overcoming host bacterial resistance when used in cocktails. Dr. He emphasized that while sequence novelty and functionality are now achievable, the major hurdle remaining is scaling experimental synthesis and testing to the scale of whole cells or larger organisms. ### Genomic Foundation Models - EVO 2 trained on 9.3 trillion base pairs spanning all life domains - Model predicts next base pair, embedding biological knowledge for RNA and protein structure - EVO 2 context window extends up to one million base pairs, though human genomes are 3 billion base pairs. ### Supervised Fine-Tuning for Specific Tasks - Specialized EVO 1 on 70,000 CRISPR-Cas sequences to generate a functional Cas9 protein that cleaves DNA in vitro with low sequence identity to natural systems - Fine-tuning enables generation of complex, multi-component biological systems. ### Eliciting Novelty via Prompt Engineering - Prompt length acts as a sweet spot to elicit creativity rather than simple memorization - Prompting with an anti-CRISPR associated gene yielded experimentally confirmed anti-CRISPR proteins with high evolutionary novelty. ### Inference Time Guidance for Epigenetics - Coupled generative EVO 2 with non-generative scoring models (Informer, Borsa) via beam search to control chromatin accessibility - Successfully designed and experimentally validated 20,000 base pair sequences exhibiting desired Morse code patterns in mouse stem cells. ### Whole Genome Design - Applied all generative techniques to design novel $\Phi$X174 bacteriophage genomes, screening 300 sequences to find 16 viable variants - One variant achieved less than 95% sequence identity to any known phage, potentially qualifying as a new species. ### Safety and Open Sourcing - EVO 2 was intentionally not trained on eukaryotic viruses for immediate safety - Open sourcing the model is considered safer as it allows the community to screen for harmful outputs and understand capabilities. ### Future Bottlenecks - Experimental limitations in synthesizing and assembling large DNA constructs (tens of KB currently, needed for whole cells in MB) present the largest bottleneck, surpassing computational challenges.