Multimodal Foundation Models and Agents for Single-cell Biology
Event details
| Date | 04.08.2025 |
| Hour | 13:00 › 15:00 |
| Speaker | Siba Panigrahi |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Antoine Bosselut
Thesis advisor: Prof. Maria Brbic
Co-examiner: Prof. Amir Zamir
Abstract
Recent advances in the reasoning capabilities of large language models (LLMs) have enabled the development of agentic frameworks capable of autonomous decision-making across diverse scientific domains, including single-cell biology. These frameworks operate through iterative think-act-observe cycles, wherein an LLM performs reasoning to plan actions, executes those actions via a dedicated environment equipped with domain-specific tools, libraries, and databases, and then integrates feedback to refine subsequent steps.
In the context of single-cell biology, such agents function as collaborative co-scientists, formulating novel hypotheses, designing validation strategies, and executing complex computational workflows using the tools embedded within their operational environment. These agents can handle heterogeneous data modalities, including multi-omics datasets, and synthesize insights by integrating prior knowledge with new observations. Notably, they can autonomously generate analysis workflows, process relevant multimodal and unimodal single-cell data, and iteratively refine their workflow to accelerate scientific discovery and enable scalable biological inquiries.
Despite these promising developments, several critical challenges remain. First, there is a lack of rigorous and standardized benchmarks which closely mirror scientific computational workflows using real-world experimental data. Consequently, there is a lack of systematic comparison of different single-cell agentic systems. Additionally, while broadly competent, current LLMs lack single-cell biology domain-specific reasoning, and as a result, they often commit errors that lead to inefficient and computationally expensive workflows that reduce the reliability of generated outputs.
In this talk, I will first present a widely adopted strategy in agent design, generating executable code as actions to improve the flexibility and success rate of agents. I will then discuss constructing a comprehensive, domain-specific environment tailored to single-cell biology, including tool integration, deep learning models, and popular databases. Next, I will introduce an existing benchmark for data analysis agents. Finally, I will briefly highlight my ongoing work on developing multimodal foundation models for single-cell biology as a potential tool within the environment and opportunities to benchmark and improve existing agents.
Selected papers
1. Executable Code Actions Elicit Better LLM Agents (https://arxiv.org/pdf/2402.01030)
2. Biomni: A General-Purpose Biomedical AI Agent (https://www.biorxiv.org/content/10.1101/2025.05.30.656746v1.full.pdf)
3. InsightBench: Evaluating Business Analytics Agents Through Multi-Step Insight Generation (https://arxiv.org/pdf/2407.06423)
Exam president: Prof. Antoine Bosselut
Thesis advisor: Prof. Maria Brbic
Co-examiner: Prof. Amir Zamir
Abstract
Recent advances in the reasoning capabilities of large language models (LLMs) have enabled the development of agentic frameworks capable of autonomous decision-making across diverse scientific domains, including single-cell biology. These frameworks operate through iterative think-act-observe cycles, wherein an LLM performs reasoning to plan actions, executes those actions via a dedicated environment equipped with domain-specific tools, libraries, and databases, and then integrates feedback to refine subsequent steps.
In the context of single-cell biology, such agents function as collaborative co-scientists, formulating novel hypotheses, designing validation strategies, and executing complex computational workflows using the tools embedded within their operational environment. These agents can handle heterogeneous data modalities, including multi-omics datasets, and synthesize insights by integrating prior knowledge with new observations. Notably, they can autonomously generate analysis workflows, process relevant multimodal and unimodal single-cell data, and iteratively refine their workflow to accelerate scientific discovery and enable scalable biological inquiries.
Despite these promising developments, several critical challenges remain. First, there is a lack of rigorous and standardized benchmarks which closely mirror scientific computational workflows using real-world experimental data. Consequently, there is a lack of systematic comparison of different single-cell agentic systems. Additionally, while broadly competent, current LLMs lack single-cell biology domain-specific reasoning, and as a result, they often commit errors that lead to inefficient and computationally expensive workflows that reduce the reliability of generated outputs.
In this talk, I will first present a widely adopted strategy in agent design, generating executable code as actions to improve the flexibility and success rate of agents. I will then discuss constructing a comprehensive, domain-specific environment tailored to single-cell biology, including tool integration, deep learning models, and popular databases. Next, I will introduce an existing benchmark for data analysis agents. Finally, I will briefly highlight my ongoing work on developing multimodal foundation models for single-cell biology as a potential tool within the environment and opportunities to benchmark and improve existing agents.
Selected papers
1. Executable Code Actions Elicit Better LLM Agents (https://arxiv.org/pdf/2402.01030)
2. Biomni: A General-Purpose Biomedical AI Agent (https://www.biorxiv.org/content/10.1101/2025.05.30.656746v1.full.pdf)
3. InsightBench: Evaluating Business Analytics Agents Through Multi-Step Insight Generation (https://arxiv.org/pdf/2407.06423)
Practical information
- General public
- Free
Contact
- edic@epfl.ch