How zou-group/virtual-lab works
The Virtual Lab project is a full-fledged research collaboration platform that orchestrates interactions between human researchers and teams of LLM agents for scientific research. It has a dual nature: a general system framework (located in the src/virtual_lab directory) that manages agent-based “meetings” and a comprehensive demonstration and application (in the nanobody_design directory) that uses a specialized computational pipeline for designing nanobodies against SARS‐CoV‐2.
Out the relationships and interactions
– The data processing scripts feed results into models.
– The models (both “original” and “improved”) output scores and predictions.
– Results are logged in directories such as scores and then visualized (via notebooks and figures).
• A feedback loop from experimental data (and discussion files) back to model adjustments and further agent discussions, emphasizing an iterative design and review workflow.
Architectural patterns and important design principles
• Separation of Concerns: The core Virtual Lab (under src/virtual_lab) is kept separate from the nanobody-specific code. This separation helps maintain a generic research agent framework versus a domain-specific demonstration.
• Modular Pipelines: The nanobody_design folder is organized into separate subdirectories for data, analysis, modeling, discussions, and deployment scripts (including SLURM batch files), reflecting a pipeline that is modular and extendable.
Relevant technologies and frameworks
• Python is the primary language.
• Virtual Lab is published as a pip package with dependencies noted in requirements files.
• External libraries/tools:
– OpenAI’s GPT-4o (via the API).
– Deep learning models like ESM, AlphaFold-Multimer, and Rosetta (the latter with associated XML configuration files).
– Jupyter notebooks for interactive data processing and visualization.
• High-level configuration and build integration (pyproject.toml, conda environment instructions) can be represented at the base or as a global “platform” note.