# Session 3: Accelerating Drug Discovery and Personalized Treatment Using AI

Source: https://www.youtube.com/watch?v=HlQ95MIbfaY
Recap page: https://rapidrecap.app/video/HlQ95MIbfaY
Generated: 2025-10-30T16:38:35.54+00:00

---
## Quick Overview

The research team developed a human-centered foundation model for cells, leveraging multi-modal data (sequence, structure, image, text) and AI to accelerate drug discovery and personalized treatment, with the RL-trained model (Biomni-R0) outperforming closed-source models like Claude 4 and GPT-5 on key biological benchmarks.

**Key Points:**
- The goal is to build a high-fidelity AI virtual cell model capable of simulating molecular building blocks (DNA, RNA, proteins) and their spatial organization.
- The project relies on four key data modalities: Sequence (DNA, RNA, protein), Structure, Image, and Text (literature), integrated via a unified AI virtual cell foundation model.
- Sequence models like UCE create a universal embedding space for 36 million cells across species and tissues, enabling zero-shot mapping of new data.
- Structure models, like HotPocketNN developed by Russ Altman's lab, identify ligand binding pockets on proteins, outperforming existing methods on challenging targets like KRAS.
- Image models (SubCell) use a multi-task learning framework to learn protein localization from Human Protein Atlas images, providing spatial context that sequence embeddings lack.
- The ultimate vision is an agentic future where AI automates biomedical research, augmenting biologists by providing the impact of an entire specialized team, rather than replacing them.
- The RL-trained Biomni-R0 model surpassed closed-source models like Claude 4 and GPT-5 on overall reward benchmarks, indicating significant progress toward expert-level performance in digital biology tasks.

![Screenshot at 00:44: The vision slide outlines the core concept: combining unprecedented amounts of biological data with Artificial Intelligence to create Virtual Cells capable of simulating biology across scales.](https://ss.rapidrecap.app/screens/HlQ95MIbfaY/00-00-44.png)

**Context:** This presentation details a multi-disciplinary research effort led by PIs including Emma Lundberg, Jure Leskovec, Russ Altman, and Serena Yeung from Stanford University, focused on creating an AI-driven virtual cell model. The project aims to leverage massive, multi-modal biological data—including sequence, structure, image, and text—to accelerate drug discovery and personalized medicine by building models that understand the spatial and functional organization of cells.

## Detailed Analysis

The presentation outlines a comprehensive approach to building an AI-driven virtual cell model to revolutionize drug discovery and personalized treatment. The core philosophy is that the cell is a 3D structured system, necessitating models that integrate data across multiple scales, from molecules to tissues. This integration is achieved by training models on four key modalities: Sequence (DNA, RNA, protein), Structure (protein function, interactions), Image (spatial organization), and Text (literature). Specific projects highlighted include UCE (Universal Cell Embedding Model) for sequence data, which creates a universal embedding space capable of zero-shot mapping across species; HotPocketNN for structure data, which excels at identifying druggable binding pockets on challenging targets like KRAS; and SubCell, a vision foundation model that captures spatial organization missed by sequence-only models. The team emphasizes a collaborative, multi-disciplinary approach, pooling expertise in fields like computer science, bioengineering, genetics, and pathology. The goal is to create an agentic system, Biomni, which automates biomedical research tasks, augmenting scientists rather than replacing them. Initial benchmarks show that the RL-trained Biomni-R0 model outperforms large closed-source models like Claude 4 and GPT-5 on overall reward metrics, demonstrating the power of integrating diverse data modalities. Finally, a case study demonstrated Biomni's ability to autonomously design a successful wet-lab cloning protocol, saving significant time and resources.

### Vision and Approach

- Building a Human-Centered Foundation Model for Cells
- Simulating the cell's molecular building blocks and their spatial organization
- Utilizing sequence, structure, image, and text data modalities via AI.

### Sequence Models (UCE)

- Created a universal embedding space representing 36 million cells (1000 cell types) across species and tissues
- UCE can map new data zero-shot without retraining.

### Structure Models (HotPocketNN)

- Assembled a benchmark of >15 million known and predicted protein pockets
- HotPocketNN identifies ligand binding pockets, outperforming existing methods on targets like KRAS.

### Image Models (SubCell)

- Proteome-aware vision foundation models learn protein localization from Human Protein Atlas images
- Multimodal embeddings integrating sequence (ESM2) and spatial location outperform unimodal embeddings in predicting interactions and distinguishing paralogs.

### Technical Approach Pillars

- Multi-modal multi-scale models
- Joint agentic analysis ecosystem for unified access
- Accessible factual knowledge from biomedical resources
- Application-level evaluations and guardrails.

### Biomni Success Metrics (First 3 Months)

- 9,531 users signed up, 2,328 monthly active users, 63,893 tasks automated (millions of research hours saved)
- Biomni-R0 (RL-trained) surpassed closed-source models (Claude 4, GPT-5) in overall reward.

### Case Study

- Wet-Lab Protocol Design
- Biomni designed a successful molecular cloning protocol in 10 minutes, which traditionally takes experts 3 years of training and effort.

![Screenshot at 0:44: The vision slide outlines the core concept: combining unprecedented amounts of biological data with Artificial Intelligence to create Virtual Cells capable of simulating biology across scales.](https://ss.rapidrecap.app/screens/HlQ95MIbfaY/00-00-44.png)
![Screenshot at 1:36: The central dogma slide illustrates the flow of genetic information \(DNA -\> RNA -\> Protein\) and states that the AI Virtual Cell model must learn this molecular language.](https://ss.rapidrecap.app/screens/HlQ95MIbfaY/00-01-36.png)
![Screenshot at 2:54: The self-organization slide emphasizes that the AI Virtual Cell model needs to be spatially and structurally aware to model the hierarchical organization of biology.](https://ss.rapidrecap.app/screens/HlQ95MIbfaY/00-02-54.png)
![Screenshot at 4:00: The slide illustrating the four key data modalities \(Sequence, Structure, Image, Text\) that feed into the AI Virtual Cell Foundation Model.](https://ss.rapidrecap.app/screens/HlQ95MIbfaY/00-04-00.png)
![Screenshot at 5:25: A collaborative approach diagram showing the five principal investigators and their respective areas of expertise feeding into the central cell model.](https://ss.rapidrecap.app/screens/HlQ95MIbfaY/00-05-25.png)
![Screenshot at 7:00: A slide detailing 'How to Build the Virtual Cell with AI,' emphasizing the need to simulate both building blocks and their spatial organization.](https://ss.rapidrecap.app/screens/HlQ95MIbfaY/00-07-00.png)
![Screenshot at 8:00: A summary of Year 1 artifacts, showing 17 models, 3 datasets, 8 benchmarks, 3 AI agents, and 21 papers produced.](https://ss.rapidrecap.app/screens/HlQ95MIbfaY/00-08-00.png)
![Screenshot at 10:30: The structure models section explaining that learning protein structure advances understanding of interactions and cellular pathway activity.](https://ss.rapidrecap.app/screens/HlQ95MIbfaY/00-10-30.png)
![Screenshot at 11:39: The structure models slide shows the HotPocketNN model identifying binding pockets on KRAS, outperforming existing methods.](https://ss.rapidrecap.app/screens/HlQ95MIbfaY/00-11-39.png)
![Screenshot at 12:30: The image models slide introduces SubCell, a vision foundation model that learns protein localization from Human Protein Atlas images.](https://ss.rapidrecap.app/screens/HlQ95MIbfaY/00-12-30.png)
![Screenshot at 13:55: A comparison chart showing that multimodal embedding \(SubCell + ESM2\) outperforms unimodal embeddings in predicting protein-protein interactions and distinguishing paralogs.](https://ss.rapidrecap.app/screens/HlQ95MIbfaY/00-13-55.png)
![Screenshot at 20:00: The multimodal text models slide emphasizes the need for AI to use scientific literature and introduces multimodal benchmarks for vision-language understanding.](https://ss.rapidrecap.app/screens/HlQ95MIbfaY/00-20-00.png)
![Screenshot at 21:25: A slide introducing the accessible virtual cell interface using AI Agents that allow biologists to interact with complex data via natural language.](https://ss.rapidrecap.app/screens/HlQ95MIbfaY/00-21-25.png)
![Screenshot at 23:05: A slide summarizing the agentic approach, where the AI agent automates research tasks in a closed-loop system.](https://ss.rapidrecap.app/screens/HlQ95MIbfaY/00-23-05.png)
![Screenshot at 28:07: A slide summarizing Biomni's success in the first three months as an open-source initiative, including 9,531 users signed up and 63,893 automated tasks.](https://ss.rapidrecap.app/screens/HlQ95MIbfaY/00-28-07.png)
![Screenshot at 30:14: The Popper framework slide illustrates an iterative hypothesis testing process where scientific progress is driven by constant falsification.](https://ss.rapidrecap.app/screens/HlQ95MIbfaY/00-30-14.png)
![Screenshot at 33:20: A concluding slide summarizing the vision: AI biologists will speed up research by augmenting scientists, making research universally accessible, and amplifying potential. AI does not replace scientists.](https://ss.rapidrecap.app/screens/HlQ95MIbfaY/00-33-20.png)
![Screenshot at 37:36: A slide illustrating the Popper agent's iterative process for hypothesis testing and validation using Bioomni for execution.](https://ss.rapidrecap.app/screens/HlQ95MIbfaY/00-37-36.png)