NVIDIA GTC 2025: Foundation Models in Biology

NVIDIA GTC 2025: Foundation Models in Biology — Summary

Key technological concepts and definitions

Biological foundation models are distributional learners trained on large corpora of biological sequences and data (DNA, RNA, proteins, etc.). They learn statistical/evolutionary patterns that can be reused across downstream tasks rather than being task-specific.

• Evolution as a unifying training signal: pretraining on evolutionary sequence data provides strong priors from natural selection (sequence → structure → function) that transfer broadly across protein and genomics tasks.
• Domain differences: chemistry and small-molecule spaces lack the same evolutionary structure, making foundation-model approaches less straightforward there; these domains often require different data, architectures, or post-training steps.
• Multimodality and length scales: effective biological foundation models must span modalities and abstraction levels — nucleotide sequences, proteins/structures, cellular images, transcriptomics, spatial data, and patient/clinical records — and connect them for end-to-end reasoning.
• Agents and automation: combining models with orchestration enables persistent agents that can scrape/standardize public data or drive lab workflows, providing scale and capabilities beyond manual processes.

Data, evaluation, and model training practice

• Empirical evaluation: lab validation is the gold standard; robust in-silico proxies that predict lab outcomes accelerate progress.
• Iterative closed loop (active learning): targeted experimental data and short feedback loops (generate → test → retrain) are critical — internet-scale pretraining data alone is finite.
• Multimodal/pairing requirement: paired datasets (e.g., sequence ↔ phenotype, image ↔ gene expression, structure ↔ function) are essential to connect layers and enable transfer.
• Tooling and UX: visualizers, designer GUIs, and interpretable outputs (reasoning traces, feature activations) increase adoption and allow both experts and non-experts to use models effectively.

Concrete success stories and demos

• Model-assisted structural insight: an AI model helped resolve a difficult protein–ligand complex and accelerated follow-up experiments (Joshua / Chai example).
• EVO / ESM series (Evolutionary Scale / ESM3):
  • Designed a novel fluorescent protein (sequence not seen in nature) that folded and fluoresced green in the lab — demonstrating generation of new functional proteins.
  • Co-design of proteins and non-coding RNA components (e.g., guide RNAs for CRISPR) to improve editing activity beyond mining natural sequences.
  • EVO2: state-of-the-art pathogenicity prediction for disease-causing mutations using evolutionary learning without supervised labels.
• Drug design (Generate Bio / Molly): generative modeling produced a therapeutic candidate with improved pharmacology (reduced dosing frequency), now in Phase 2 clinical trials — an example of in-silico design translating toward clinical validation.
• Virtual Cell Atlas (ARC Institute): an agent-driven pipeline reanalyzed hundreds of millions of single-cell RNA-seq datasets (SRA) into a standardized atlas — an example of automated large-scale data aggregation and curation.

Products, features, and tooling mentioned

• EVO Designer / EVO Visualizer: tooling to prompt/model DNA/protein outputs, fold designs, and inspect model internals and nucleotide-resolution activations.
• Virtual Cell Atlas: a large, standardized single-cell corpus produced by agentic pipelines.
• Open model releases and community tooling: open-source models, model hubs, and interfaces to accelerate adoption and gather community feedback.
• Multi-agent frameworks and ELN integration: chains of specialized models (agents) calling each other and instrument APIs to run end-to-end lab or analysis tasks; emphasis on preserving “reasoning traces” (chain-of-thought) for reproducibility.

Translation, business, and economics

• Compute vs. clinical costs: training large models is expensive, but failed clinical trials are far costlier; improved model guidance could reduce downstream failure and overall program costs.
• Cost trends: training and lab experimentation are becoming cheaper (cloud GPUs, commoditized experimental services), enabling more shots on goal.
• Commercial vs. open science balance: open-source releases accelerate validation, talent recruitment, and community-driven use cases; many companies follow a portfolio strategy (open releases + commercial services/tooling).
• Monetization: early revenue models focus on partnerships, platform/API access, and enterprise tooling layered above core models.

Safety, interpretability, workforce, and governance

• Interpretability & mechanistic insight: inspecting model internals to discover biology is a research goal and a means to increase trust.
• Human-in-the-loop & training scientists: experts currently benefit more from models than junior users; training the next generation to work with model-enabled pipelines is a challenge.
• Risks and dual-use: unintentional and intentional misuse are concerns. Suggested mitigations include curated training datasets (excluding certain sequences), access controls, operational safety layers, and research into alignment/safety for lab agents.
• Skill atrophy concern: panelists discussed the risk that human hands-on skills may decline as models automate more of the workflow.

Predictions and “killer apps” (panel views)

• No single killer app: value will arise from connecting models across levels (sequence → cell → patient) and providing reusable “app store” layers that enable many high-value applications.
• High-impact outcomes (5–10 years): programmable biology (engineerable biological systems), virtual cells/patients (digital twins), and new materials/biomanufacturing.
• Two axes of impact:
  1. Ubiquitous AI accelerating routine lab work (faster, cheaper, better).
  2. Enabling entirely new classes of biology-enabled applications not feasible today.

Guides, tutorials, and resources cited

• Virtual Cell Atlas — standardized, large single-cell dataset produced by an agent pipeline.
• EVO model series and EVO Designer/Visualizer — model weights, papers, and interactive tooling for generative protein/genome outputs.
• Open model releases (examples: Chai’s initial open model, ESM/EVO releases) — used to gather feedback, broaden evaluation, and build trust.
• Best-practice guidance implied: prefer iterative lab-in-the-loop evaluation, build paired multimodal datasets, produce interpretable outputs, and expose tooling to the community.

Speakers / main sources (as introduced)

• Rory Keller — NVIDIA (intro)
• Moderator: Anthony — leads digital biology / developer relations at NVIDIA
• Molly Gibson — Flagship Pioneering; founding president of Lio Sciences; former Generate Biomedicines
• Patrick (surname in transcript: “Sue”) — co-founder ARC Institute; assistant professor of bioengineering
• Nicholas (surname in transcript: “Sophia”) — leads frontier intelligence team at Evolutionary Scale; previously at Chan Zuckerberg Initiative (CZI)
• Joshua (surname in transcript: “Omire”) — co-founder & CEO of Chai Discovery; prior work on ESM/protein transformer models

(Note: the transcript contained several auto-generated word errors; names and organization spellings are reported as given in the subtitles.)