Session 6: Evo, A Foundation Model for Generative Genomics

Session 6: Evo, A Foundation Model for Generative Genomics

by Stanford HAI

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQFGenerative artificial intelligence
🎙 Stanford HAI 👥 34K 📅 October 30, 2025 ⏱ 46 min 👁 658 📄 lecture 🧭 2026-08-06
Available in: English (current) Français

Keywords

EvogenomicsDNA language modelCRISPRfoundation model

Summary

The talk presents Evo, a foundation model for genomics developed by a Stanford research team. Evo is a large language model trained on DNA sequences using next-base-pair prediction. The model, with 7 billion parameters, was trained on 300 billion base pairs from prokaryotic genomes. It demonstrates that DNA sequence alone can encode information about RNA and proteins, and that scaling up the model improves performance. The model can predict the effects of mutations on gene essentiality, generate novel CRISPR-Cas systems with functional activity, and produce large genomic sequences that appear realistic. The team aims to extend Evo to eukaryotic genomes, including humans, to advance understanding of disease and enable therapeutic design. The talk emphasizes the potential of AI to compose DNA sequences, addressing the challenge of controlling biology for human benefit.

131 words

Critical Evaluation

The presentation provides a comprehensive overview of the Evo project, highlighting its innovative approach to applying large language models to genomics. The speaker clearly explains the rationale behind using DNA sequence as a training signal, drawing parallels to natural language processing and emphasizing the evolutionary imprint on genomic data. The technical details, such as the use of the striped hyena architecture and the scaling laws, are presented with sufficient depth for an informed audience. The experimental results, including the generation of functional CRISPR-Cas systems and the prediction of gene essentiality, are compelling and suggest practical applications. However, the talk is primarily a high-level summary and does not delve into potential limitations or failure cases of the model. The speaker mentions that Evo1 generates ‘blurry pictures of genomes,’ but does not elaborate on the specific shortcomings or how they might be addressed. Additionally, while the ethical considerations are briefly touched upon, the discussion is not extensive. The sources cited are limited to the Hoffman-Yee grant program, and no external references are provided to support the claims. Overall, the talk is informative and well-structured, but it would benefit from a more critical examination of the model’s limitations and a broader citation of relevant literature. The title accurately reflects the content, and the presentation is suitable for an audience with some background in AI and biology.

223 words

Title / Content Match

The title accurately reflects the content, which focuses on the Evo foundation model for genomics.

Quality & Reliability

8/10

The talk is presented by a researcher from a Stanford-affiliated project, likely with strong academic credentials. It describes a peer-reviewed research project (Evo) with published results, and includes technical details and experimental validation. However, as a single presentation, it lacks external verification and may omit limitations.

Chapters

Cited Sources

Concurring Sources

Contribution & Novelties

The talk presents Evo, a foundation model for genomics that extends language modeling to DNA sequences, enabling the generation of novel functional biological sequences. The model’s ability to generate CRISPR-Cas systems with high novelty and functional activity is a significant advancement. The approach of training on genomic sequences to capture evolutionary information offers a new paradigm for understanding and engineering biology.

Pour aller plus loin :

  • Evo 1 paper — The preprint describing Evo 1, providing detailed methodology and results.
  • Striped Hyena architecture — The paper introducing the striped hyena architecture used in Evo.
  • AlphaFold — A related AI system for protein structure prediction, illustrating the impact of AI in biology.

111 words

Radar Profile

The radar profile shows high scores across all dimensions, indicating a well-rounded presentation with substantial information, strong technical depth, and high reliability. The model's innovative approach and experimental validation contribute to its high quality.

Reliability 8/10