GTC 2025 Foundation Models in Biology

Category: Biology Special Address · Year: 2025 · ▶ Watch

Switch language → 中文

Segments (9)

  • 00:00 · Introduction
    • Rory Kelleher introduces the session and Anthony Costa introduces the panelists.
  • 04:28 · Defining Foundation Models in Biology
    • The panel discusses what constitutes a foundation model in the context of biology compared to natural language.
  • 11:25 · The Language of DNA and Data Modalities
    • Discussion on the challenges of interpreting biological data and the need for diverse data modalities.
  • 16:38 · Multimodality in Biological Models
    • Exploring how models can integrate DNA, RNA, proteins, and cellular data to understand complex biological systems.
  • 20:14 · From Virtual Cells to Scientific Agents
    • The conversation shifts towards building AI agents capable of reasoning and executing scientific workflows.
  • 27:28 · Open Source vs. Closed Source and Business Models
    • Panelists debate the merits of open-sourcing models and the evolving business landscape in biotech.
  • 31:24 · Success Stories and Concrete Examples
    • Speakers share specific examples of how their models have solved real-world biological problems.
  • 35:52 · The Future of Foundation Models in Biology
    • Predictions on the killer apps for biological AI and its impact beyond drug discovery.
  • 37:08 · Q&A: Controlling Harmful Uses
    • The panel addresses a question on how to prevent the misuse of powerful biological models.

Specific Numbers (4)

Timestamp Metric Value Context
11:32 Percentage 1% Analogy used to describe how poorly humans currently understand the ‘language’ of DNA.
25:57 Cells 400 million The number of single cells aggregated in the Virtual Cell Atlas dataset.
32:54 Probability 10% The current therapeutic probability of success from preclinical to approved drug.
34:53 Years 500 million The estimated evolutionary time simulated by the Evo model to design a novel fluorescent protein.

Customer Stories (2)

  • [22:20] Scientists using Chai-1
    • Used the model to figure out the 3D structure of a molecule and protein complex.
    • outcome: Overcame resistance in a specific biological pathway, demonstrating the model’s utility in complex problem-solving.
  • [39:04] Generate Biomedicines
    • Designed a novel drug for severe asthma using generative models.
    • outcome: Created a molecule that requires dosing only twice a year instead of every month, significantly improving patient experience.

Key Technologies (4)

  • ESM (Evolutionary Scale Modeling): A family of protein language models that learn the evolutionary patterns of protein sequences.
  • Evo: A genomic foundation model capable of designing DNA, RNA, and proteins across multiple scales.
  • Chroma: A generative model for programmable protein design.
  • AI Agents: Systems that can reason, plan, and execute complex workflows autonomously, moving beyond simple input-output models.

Predictions / Commitments (3)

  • [26:18, Near future] Models are graduating from tools into products that can own end-to-end workflows, acting as scientific agents.
  • [34:08, 5 to 10 years] The killer app will be connecting these models together and how scientists use them to accelerate discovery.
  • [35:38, Long term] Biology will be used for computation, terraforming planets, and creating biomaterials, expanding far beyond therapeutics.

Companies Mentioned (4)

Cursor, Codium, Cognition · DeepMind · OpenAI, Anthropic, xAI, Mistral · NVIDIA

Notable Quotes (3)

We speak DNA with a really heavy accent. — Patrick Hsu @ 11:23

These models are graduating from tools into products where they can actually own more of these end-to-end workflows. — Joshua Meier @ 26:18

Drugs are just a small part of biology. Biology can terraform planets, it can make biomaterials, it can be used for computation. — Patrick Hsu @ 35:38

Key Topics

Foundation Models · Digital Biology · Generative AI · Protein Design · Genomics · Multimodality · AI Agents · Drug Discovery · Open Source · AI Safety · Virtual Cells · Scientific Workflows

Takeaways

  • Foundation models in biology are evolving from single-modality sequence models to multimodal systems that understand DNA, RNA, proteins, and cellular contexts.
  • The next major leap is the development of AI agents that act as ‘virtual scientists,’ capable of reasoning, planning, and executing complex experimental workflows.
  • A significant bottleneck in the field is the lack of high-quality, diverse, and paired datasets needed to train more capable and generalizable models.
  • Open-source models play a critical role in advancing academic research and building trust, though commercialization strategies for these technologies are still being figured out.
  • The ultimate impact of biological foundation models will extend far beyond therapeutics, potentially revolutionizing fields like biomaterials, agriculture, and environmental engineering.