Find out more:

Preprint: www.biorxiv.org/content/10.1...
Code: github.com/davidfischerlab/kai

Blog: ai4biomedicine.org/agenticai

kai is an assistant for single-cell biology optimized for human-agent collaboration. Like AI assistants in other domains, kai enhances human efficiency while maintaining accountability – a key advantage over fully autonomous systems in science.

We compared kai with one-shot analysis generation by LLMs by scoring the generated Jupyter notebooks based on various criteria. kai consistently outperforms one-shot analysis generation.

We tested kai on complex scenarios in single-cell biology. kai consistently completed analyses and reasoned (LLM reasoning + analysis execution) for longer than 20 minutes.

The output of kai’s reasoning process is this Jupyter notebook: a transparent documentation of all analyses performed and decisions made. Human scientists can inspect, modify, and give feedback on each step of the analysis.

kai interacts with human scientists through a chat interface in VS Code and directly edits and executes Jupyter notebooks. This design enables kai to autonomously perform analyses while maintaining full accountability.

This motivated us to build kai: an agentic AI that uses Jupyter notebooks – the same interface that humans use to collaborate.

We started by asking: how do humans build trust with each other? In collaborations, they document their reasoning in computational notebooks, e.g. Jupyter notebooks.

For example, how can I verify the product of 20 minutes of autonomous work by an agent without blindly hoping that it didn’t hallucinate at a key decision point at minute 5?

In cell biology, agentic AI systems need to reason over text, code, and analysis results. How do we ensure accountability in this complex multimodal setting to inspire trust in the predictions made by agents?

Our featured article: Adapting systems biology to address the complexity of human disease in the single-cell era go.nature.com/3XBo6Vh #Review by @davidsebfischer.bsky.social, Martin A. Villanueva, Peter S. Winter & ‪@shaleklab.bsky.social‬ @broadinstitute.org @mit.edu @ragoninstitute.bsky.social

This review is a product of a great team effort together with Martin Villanueva, Peter Winter and Alex Shalek! www.nature.com/articles/s41... & rdcu.be/ecTna

In summary, we outline how systems biology is being adapted to the multiscale dynamics of human health and disease in omics-driven as what is effectively a two-loop cycle over discovery and validation.

We review strategies that can manage this distance and dissect how it relates to understanding cellular systems at specific spatiotemporal scales - be it the cellular scale often considered in the context of single-cell-resolved experiments, or the tissue niche scale captured with spatial omics.

Both are needed to build quantitative models are faithful to human biology and validated through perturbation experiments. However, the usage of two distinct systems incurs a "translational distance" that complicates systems biology approaches that utilize information from the two.

In this review, we discuss how one can rationalize what information about a multiscale cellular system is actually captured by on omics study. We leverage that insight to describe how one can translate between discovery efforts in human tissues and validation efforts in experimental model systems.

However, the dynamics of human tissues in disease settings is multiscale - not only does that impact quantitative models, it also reflects in experimental design and the resources of publicly available data that we have access to. This obstructs attempts at building such quantitative models.