Yusuf Roohani at ARDD2025: Building a Virtual Cell: an AI platform for engineering cell state

Yusuf Roohani at ARDD2025: Building a Virtual Cell: an AI platform for engineering cell state

🎙 Yusuf Roohani 👥 9K 📅 January 19, 2026 ⏱ 22 min 👁 299 📄 expert opinion 🧭 2026-08-16
Available in: English (current) Français

Keywords

virtual cellAIsingle-cellperturbation predictionfoundation models

Summary

Yusuf Roohani, leading a machine learning group at the Arc Institute, presents his vision for a ‘virtual cell’—an AI platform to predict cellular responses to perturbations and ultimately engineer cell state. He frames AI as a tool to accelerate scientific discovery, citing AlphaFold as a paradigm. The talk outlines four research fronts: model development, evaluation, data curation, and data generation. He introduces ‘STATE’, a transformer-based model that predicts perturbation outcomes by learning over sets of cells, outperforming baselines in capturing differentially expressed genes and recapitulating perturb-seq experiments. To evaluate such models, the Arc Institute launched the ‘Virtual Cell Challenge’, a community competition with over 3,000 participants. For data, they built ‘SCBase’, an agent-curated repository of over 600 million cells, five times larger and more diverse than existing ones. Finally, they use AI agents to propose and iteratively optimize genetic perturbation experiments, achieving a 21% average improvement in hit rate. Roohani emphasizes a closed-loop platform integrating experimental and computational efforts, and concludes with the concept of AI as a ‘virtual biologist’.

170 words

Critical Evaluation

Value of the Information & Strength of the Argument

The presentation provides valuable insights into the state-of-the-art in AI-driven cell modeling. The argumentation is solid, grounded in published research and concrete results. Roohani clearly explains the motivation, methodology, and outcomes of their work, making a compelling case for the virtual cell approach. He acknowledges limitations and open questions, such as the need for better evaluation frameworks and the potential for errors in agent-curated data. The talk is well-structured and persuasive, though it primarily presents the speaker’s own work without deep critical comparison to alternative approaches.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is high, as the work is published in reputable journals like Cell. The speaker cites specific papers and provides a link to his website for further details. The title accurately reflects the content. The talk is a conference presentation, so it does not provide full methodological details, but it references the relevant literature. The sources are credible and directly related to the presented work.

168 words

Title / Content Match

The title accurately reflects the content, which focuses on building a virtual cell platform using AI for engineering cell state.

Quality & Reliability

8/10

The presentation is by a leading researcher in the field, affiliated with the Arc Institute and Stanford, and describes peer-reviewed work published in Cell. The claims are supported by specific results and references to published papers. However, the talk is a conference presentation and does not provide full methodological details, and some claims about the superiority of their model are based on their own evaluations.

Key Moments

Cited Sources

  • Perspective on virtual cells published in Cell — Roohani mentions a perspective paper published in Cell last year describing the vision for a virtual cell.
  • STATE paper — The model STATE is described, and the paper is referenced as published.
  • Virtual Cell Challenge paper — The competition was published in Cell in June of this year.
  • SCBase paper — The agent-curated repository SCBase is described, with a preprint mentioned.

Concurring Sources

  • AlphaFold — Referenced as a successful AI model for protein structure prediction.
  • Foundation Models paper — The paper that introduced the term 'foundation models'.

Contribution & Novelties

The talk presents novel contributions in AI-driven cell modeling, including the STATE model that learns over sets of cells, the Virtual Cell Challenge for community evaluation, and the use of AI agents for data curation and generation. The emphasis on a closed-loop platform integrating experimental and computational efforts is a distinctive approach.

Pour aller plus loin :

  • Foundation Models — The paper that introduced the term ‘foundation models’.
  • AlphaFold — The AI system for protein structure prediction.
  • Perturb-seq — A technique for high-throughput perturbation screens with single-cell readouts.
  • Single-cell RNA sequencing — Overview of the technology used in the talk.

100 words

Radar Profile

The radar profile shows high scores across all dimensions, indicating a well-rounded presentation with strong information content, technical depth, and reliability. The talk is particularly strong in terms of novelty and practical impact, with slightly lower scores in breadth of sources due to the focus on the speaker's own work.

Reliability 8/10

💬 Sur les 0 commentaires analysés, aucune tendance n'a pu être dégagée.