VenusFactory2

VenusFactory2

Recursive Self-improvement AI Agents for Protein Engineering

Web server · Tutorials · Installation · 中文

VenusFactory2 brings protein research into a conversational workspace. Ask a question, upload experimental data, or start with a protein of interest to explore the literature, run predictions and develop models for your research.

• Ask and analyse. Find protein information, use scientific tools and interpret results.
• Develop from your data. Turn a research goal and measurements into model code, tests and a research report.
• Investigate a protein. Bring literature, database records and predictions together in a report with sources.
• Reuse the results. Run models directly, combine tools in a workflow, or request publication of a reviewed Scientist model.

Start with a question

Click New Chat, choose a mode and describe what you want to learn. Agent is selected by default. You can paste a sequence or attach files, including structures, measurements and papers.

Mode	Example request	What you get
Agent · Ask and analyse	“Find homologues of this protein and explain its annotated domains.”	An answer supported by searches, tool results and downloadable files
Scientist · Develop and improve	“Use my solubility measurements to develop and evaluate a predictor.”	A project with code, test and evaluation results, a report, and model artifacts when trained
Researcher · Investigate a protein	“Investigate this enzyme, focusing on substrate specificity and engineering opportunities.”	A multi-section report combining literature, protein information and predictions

To explore first, use Agent’s suggested questions or open a completed example in Scientist or Researcher. Completed examples include results you can read and download without starting a new run. You can also ask Agent how to use the platform; it reads the same tutorials as the Manual.

How Scientist works

Scientist uses recursive self-improvement (RSI): it develops a solution, tests it, uses the feedback to revise the approach, and repeats within the research budget. A PI (research lead) coordinates the work:

Goal and data → PI and plan → Develop and experiment ↔ Critique and revise → Review → Deliver

Follow the plan, intermediate files and progress in the conversation. The final project and report can be downloaded as a ZIP. A useful result may be a new analysis tool, a fitted machine-learning model or a method for your experimental setting. Scientist outputs stay private. To share a model, email its project and test report to [email protected]. The platform operator selects reviewed models for publication; each published entry displays the author email. See operator publication for local registration.

Scientist delivers a selected model or method with a README, runnable interface, examples and a test report. Experiments and earlier versions stay in the research topic; they do not each become another default tool. See the Scientist guide for using and downloading results.

Choose models for your experiment

Model Plaza groups models by biological function. Open a model to run a prediction, or use Train, Evaluate and Predict with your own dataset. Workflow lets you connect tools and follow the data through each step.

Research task	Example tools and models
Find related proteins and annotate function	NCBI BLAST, Foldseek, InterProScan
Rank mutations	VenusREM, ProSST, ProtSSN
Design sequences	ProteinMPNN, LigandMPNN
Predict protein properties	ProtSolM, VenusX, VenusVaccine and Factory-trained models
Study small molecules	ADMET prediction and Vina docking
Build a model for your experiment	Custom model training and reviewed Scientist models

The catalog brings together established tools, Factory-trained models and reviewed Scientist contributions. Available tools depend on the installation and administrator settings; the model page shows what can run.

Run locally

Use Linux or WSL2 with Git and Python 3.10+ to run the installer; it prepares a Python 3.12 runtime for the platform. This setup installs the web platform, built-in search and protein services, and one agent backend:

git clone --depth 1 https://github.com/ai4protein/VenusFactory2.git
cd VenusFactory2
python3 install_full.py --components harnesses --harnesses deepseek-harness
.venv/bin/python scripts/serve.py --mode local --port 7861

Open http://localhost:7861. In Settings → Administration, configure a language-model API and enable the tools you need. Use DMXAPI or OpenRouter free models, or configure your own OpenAI, Claude, Grok, GLM, MiniMax and other provider APIs in local mode. See API configuration. Web search can use Tavily or SerpAPI; public protein and literature databases are built in.

The installer detects the environment, prepares dependencies and shows download progress. Model packages and weights can be added later. For example:

python3 initialize.py install --tools proteinmpnn,ligandmpnn

Each runtime gets its own Python environment and model directory. Optional structure and design runtimes include AlphaFold2, Protenix, Mol-1, RFD4-Proteina, RFantibody, ESMFold and ESMFold2-Fast, alongside BindCraft and BoltzGen. See the runtime guide for versions, installation and validation.

For a full installation, use python3 install_full.py. Some components require a GPU or a separate license. See the installation guide for hardware requirements, component selection and health checks, and the service guide for online or multi-process deployment.

For PBS clusters, see online/local deployment. The job supervises both services, keeps their data separate and cleans up its processes on exit.

Online or local?

Use the web server to get started without installation. Run locally when you need your own datasets, a larger training model or additional compute-intensive tools. The two modes share the same interface.

	Online	Local
Conversations	Shared daily allowance across all three modes	Your own API and resource settings
File uploads	Up to 10 MB per file	No platform file-size limit
Custom model training	ESM2 8M	Installed models and available hardware
Library	Unavailable	Browse uploaded and generated files

Learn more

• Tutorials and examples: Chat, Model Plaza, Workflow and working with your data.
• Installation: add models, weights and agent backends as needed.
• Developer guide: repository layout, model integration, services and checks.

License and citation

VenusFactory2 is available for non-commercial use under the project license. Individual tools and models retain their own licenses. Please use the published citations when referring to the platform and its methods. For questions, contact [email protected].