Agent-driven protein engineering platform integrating 40+ protein language models and 11+ biological databases, supporting end-to-end mutation design, function prediction, and model fine-tuning.
VenusFactory2 is an agent-driven protein engineering platform developed by Hong Liang's Lab at Shanghai Jiao Tong University. It features a layered architecture: a resource layer comprising 40+ protein language models, 30+ datasets, and 11+ biological databases; an application layer with Train/Eval/Predict/Tools modules; a LangGraph + LangChain multi-agent orchestration layer; and interface layers including Web UI, REST API, and CLI.
Core capabilities include zero-shot mutation effect prediction via ESM-2/ProSST/ProtSSN (<1 min), protein function prediction with 30+ fine-tuned models (<30 sec), custom model training supporting 7 PEFT methods (LoRA, QLoRA, DoRA, AdaLoRA, SES-Adapter, IA3, Freeze), one-stop multi-source data retrieval from AlphaFold/UniProt/RCSB/KEGG, and AI-assisted literature search.
The Agent-0.1 module (currently Beta) enables automatic planning and execution of multi-step protein engineering workflows from natural language instructions (e.g., "Design thermostable mutations for PDB:1ABC" → PDB download → ESM-2 scanning → stability scoring → ranked report). The platform offers dual Web UIs (Gradio v1 and FastAPI+React v2) with local and online deployment modes, full REST API documentation, and CLI scripts. Related work published at ACL 2025 (v1), arXiv 2026 (v2 tech report), with sub-models VenusREM (ISMB/ECCB 2025) and VenusX (ICLR 2026) integrated.
Unconfirmed: license type (LICENSE file exists but specific terms not stated), model weight distribution (no HuggingFace link provided), online Playground feature completeness.