scAgent

Agentic AI framework that probes our local single-cell data. Enables non-computational experts to seamlessly investigate transcriptional signatures.

Methodology

• Converted our seurat object to anndata, enabling versatile scanpy analysis features
• Downloaded 20 billion parameter openAI LLM (gpt-oss:20b)
• Agentic framework designed with LangChain in Python. LLM provided with tools and context to execute analysis on our data

Architecture

Phase 1: Planner

Expert bioinformatician takes prompt and generates a detailed step-by-step plan

Phase 2: Executor

Extracts plan and runs code on sc data using defined libraries

Phase 3: Self Corrector

Feedback loop that catches an error in the script. Debugs and updates

Phase 4: Summarizer

Summarizes results from executor