ProjAtlas

An AI-assisted multi-agent platform for interactive exploration of single-neuron projectomes across mouse and macaque brains.

ProjAtlas integrates atlas-registered single-neuron projectomes from mouse and macaque brains and supports literature interpretation, data querying, interactive visualization, and quantitative analysis through natural-language interaction.

Platform: https://digital-brain.cn/ProjAtlas

Repository Structure

ProjAtlas/
├── platform/                  # ProjAtlas web platform
│   ├── frontend/              # Web client (interactive 3D visualization, data querying)
│   └── backend/               # Backend services (user/session/message management, LLM assistant)
├── multi-agent-framework/     # Multi-agent framework (paper-interpretation, neuron-selection,
│                               # brain-visualization, and textual-summarization agents)
├── rag-evaluation/            # RAG benchmark and evaluation (paper interpretation agent)
└── function-calling/          # Function-calling model training and evaluation
                                # (neuron-selection and brain-visualization agents)

Each subdirectory has its own README with setup and usage instructions.

How the Pieces Fit Together

Two different relationships exist between these directories — a runtime dependency and an experiment-produces-asset relationship. They are not the same thing:

• platform/backend depends on multi-agent-framework at runtime. The backend imports it as the atlas_assistant Python package (installed in editable mode from multi-agent-framework/, see platform/backend/README.md). multi-agent-framework is what actually runs the four conversational agents (paper interpretation, neuron selection, brain visualization, textual summarization) that platform/backend exposes over HTTP to platform/frontend.
• rag-evaluation and function-calling are standalone experiment code, not runtime dependencies. They are used once, offline, to produce the assets that multi-agent-framework consumes at runtime:
  • rag-evaluation benchmarks the RAG pipeline (retrieval + generation) used by the paper interpretation agent, and documents how to build the same HippoRAG2 knowledge base that multi-agent-framework loads at startup (see multi-agent-framework/README.md).
  • function-calling trains and evaluates the fine-tuned model used by the neuron selection and brain visualization agents; the resulting LoRA adapter weights, published at zhang-manyi/ProjAtlas-xLAM2-8B-FunctionCalling-lora (see function-calling/README.md), are what multi-agent-framework calls out to via FUNCTION_CALLING_MODEL in its .env.

In short: to run the ProjAtlas platform, you need platform/ + multi-agent-framework/ (plus a RAG index and a deployed fine-tuned model, built using the other two directories). To reproduce the paper's reported numbers, rag-evaluation and function-calling are self-contained and don't require the platform to be running.

License

Code in this repository is licensed under Apache License 2.0. See NOTICE for third-party acknowledgements and licensing terms that apply to model weights distributed alongside this repository (which are not covered by the Apache-2.0 license).

Citation

Citation will be added upon publication.

Developed By

Brain Science Data Center, Center for Excellence in Brain Science and Intelligence Technology (CEBSIT), Chinese Academy of Sciences.