BioClaw: Human-Bot Research Collaboration Ecosystems in Group Chats

This paper presents BioClaw, a human-bot collaboration ecosystem that transforms group chats into persistent, secure workspaces by converting natural language requests into automated, tool-grounded biomedical analyses executed within isolated Docker containers.

The Gist

Imagine you are in a busy group chat with your research team. You're discussing a new experiment, sharing ideas, and spotting patterns in your data. Suddenly, someone says, "Hey, let's make a volcano plot to see which genes are changing!" or "Can you check if this protein structure matches what we found in the literature?"

In the old days, the conversation would stop. Everyone would have to:

1. Log out of the chat.
2. Open a different program (like Excel or a coding terminal).
3. Upload the files.
4. Run the code.
5. Save the image.
6. Log back into the chat and paste the result.

It's like trying to cook a complex meal while standing in the middle of a crowded party. You have to leave the party, go to the kitchen, chop vegetables, cook, plate the food, and then run back to the party to serve it. It breaks the flow, and you might forget what you were talking about.

BioClaw is the solution to this problem. It turns the group chat itself into a fully equipped, high-tech kitchen.

The Core Concept: The Chat as a "Living Workspace"

Think of BioClaw not just as a chatbot that answers questions, but as a digital research assistant that lives inside your group chat.

• The Kitchen Analogy: Imagine your group chat is a kitchen. In the past, if you needed a blender, you had to go to a different building to get one, use it, and bring it back. With BioClaw, the blender, the oven, the knives, and the recipe books are already built into the walls of the chat room. You just say, "Blend these tomatoes," and the robot arm in the corner does it instantly, handing you the sauce right there in the conversation.
• The Persistent Memory: Unlike a normal chatbot that forgets everything once you close the window, BioClaw remembers everything. It keeps a "workspace" for your group. If you upload a file today, the bot remembers it tomorrow. If you ask for a graph today and then ask to "change the colors" tomorrow, it knows exactly which graph you are talking about because it's been saving the work in the background.

How It Works (The Magic Behind the Scenes)

The paper describes a system that feels like magic but is built on very solid engineering:

1. The Universal Translator (Orchestration):
   BioClaw works on 8 different platforms (like WhatsApp, Slack, WeChat, Discord, etc.). Think of this as a universal translator. Whether you are speaking "Slack-speak" or "WeChat-speak," BioClaw understands you and translates your request into a command the computer can execute.

2. The Secure Sandbox (Docker Containers):
   When you ask the bot to do something dangerous or complex (like analyzing sensitive medical data), it doesn't just run it on the main computer. It creates a temporary, isolated "sandbox" (a Docker container).

   • Analogy: Imagine a chef who, instead of cooking on your kitchen counter, sets up a tiny, self-contained, sterile cooking station right on your table. They cook the meal, serve it, and then immediately pack up the station so nothing messy is left behind. This keeps your data safe and ensures the bot doesn't accidentally break anything.

3. The Super-Toolbox (31 Tools & 95+ Skills):
   Inside that sandbox, the bot has access to a massive library of pre-installed scientific tools. It's like having a Swiss Army knife that never runs out of batteries. It has tools for:

   • Genomics: Reading DNA sequences.
   • Structural Biology: 3D modeling of proteins.
   • Statistics: Making charts and graphs.
   • Literature Search: Finding relevant scientific papers.
   • Image Analysis: Looking at microscope photos.

Real-Life Examples from the Paper

The paper shows how this works in practice:

• The "Volcano Plot" Request: A researcher uploads a spreadsheet of gene data and says, "Make a volcano plot and highlight the important genes." BioClaw instantly runs the code, generates the colorful chart, and posts it back in the chat. No one had to leave the conversation.
• The "Protein Structure" Request: Someone asks, "Show me what this protein looks like in 3D." The bot renders a 3D model right in the chat, which the team can rotate and inspect together.
• The "Paper Hunt": A team needs to know what other scientists have written about "circadian rhythms." The bot searches thousands of papers, summarizes the top findings, and posts a neat list in the chat.

Why This Changes Everything

The paper argues that science is often a collaborative conversation, not just a series of isolated tasks.

• Before BioClaw: The conversation was fragmented. "I have an idea!" -> Leaves chat -> Does work -> Returns with result -> "Okay, what's next?"
• With BioClaw: The conversation is continuous. "I have an idea!" -> "Let's test it!" -> Bot does the work instantly -> "Look at this result!" -> "What if we tweak it?" -> Bot adjusts instantly.

It turns the group chat from a place where you just talk about science into a place where you actually do science. It bridges the gap between having a brilliant idea and verifying it with data, keeping the whole team in the loop the entire time.

The Future: A Team of Specialists

The authors even imagine a future where one group chat doesn't just have one bot, but a team of specialized bots.

• One bot is the "Sequencing Expert."
• One is the "Statistics Guru."
• One is the "Literature Librarian."
• One is the "Visual Artist."

You could talk to all of them at once in the same chat, and they would work together to solve a complex problem, just like a real research team, but without the need for everyone to be a computer expert.

In short: BioClaw is the ultimate research assistant that lives in your group chat, turning casual conversation into powerful, executable scientific discovery without ever making you leave the room.

Technical Summary

1. Problem Statement

Current scientific collaboration often occurs in group messaging platforms (e.g., Slack, WeChat, WhatsApp), where researchers discuss hypotheses, share data, and request analyses. However, a significant workflow fragmentation exists:

• Context Switching: Researchers must leave the conversation to use fragmented local tools, databases, visualization software, and literature search engines to execute analyses.
• Lack of Persistence: Chat threads are treated as transient communication channels rather than persistent computational workspaces. Intermediate results, file states, and execution history are not automatically preserved within the conversation context.
• Execution Gap: While Large Language Model (LLM) agents can plan tasks, existing solutions struggle to embed reliable, tool-grounded bioinformatics execution directly into long-lived, multi-user group chats without compromising security or reproducibility.

2. Methodology: The BioClaw System

BioClaw is a human-bot research collaboration ecosystem designed to transform group chat threads into persistent, executable computational workspaces.

A. System Architecture

The system operates on a multi-channel orchestration model with isolated containerized execution:

1. Orchestration Layer:
   • Multi-Channel Support: Connects to 8 distinct platforms (WhatsApp, Feishu, WeCom, Discord, Slack, WeChat, QQ, and a Local Web UI) via channel adapters that normalize platform-specific events into a unified message representation.
   • State Management: Maintains per-group state ($C_L$) including conversational history ($M_L$), a dedicated group workspace ($N_L$), and an execution queue ($O_L$).
   • Concurrency: Uses SQLite for state persistence and implements "cursor advance-then-rollback" semantics to handle concurrent multi-group operations and ensure failure recovery.
2. Execution Runtime (Containerized):
   • Isolation: Every analysis runs inside an isolated Docker (or Apptainer for HPC) container, ensuring filesystem isolation and security.
   • Dual Backend Support: Supports both Anthropic's Claude Agent SDK (with native MCP integration) and OpenAI-compatible APIs (with explicit tool-calling loops).
   • Communication: Uses a dual-channel IPC mechanism: stdin/stdout for prompts and streaming, and shared filesystem directories for artifacts and scheduling.
3. Bioinformatics Runtime:
   • Pre-bundled Environment: Each container image includes a reproducible environment with 17 command-line tools (e.g., BLAST, BWA, SAMTools, PyMOL, FASTQC) and 14 Python libraries (e.g., for single-cell analysis, cheminformatics).
   • Skill Architecture:
     • Built-in Skills (25+): Pre-packaged workflows for common tasks (e.g., differential expression, literature retrieval).
     • Skills Hub (70+): A community-maintained repository allowing dynamic discovery and execution of domain-specific skills across 10 domains.

B. Key Design Features

• Skill Design Philosophy: Skills are designed as workflow-level abstractions rather than simple tool wrappers. They are:
  • Modular & Composable: Breaking complex tasks into checkpointed stages.
  • Artifact-Persistent: Intermediate outputs are written to the group workspace, allowing inspection and partial reruns.
  • Human-in-the-Loop: Users can review intermediate results and redirect the workflow without discarding prior progress.
• Reproducible Notebook Generation: Upon completion of an agent run, the system automatically reconstructs a Jupyter Notebook (.ipynb) from the execution trace. This includes extracted Python code, file contents, and agent reasoning (Markdown), ensuring every chat-initiated analysis is a self-contained, re-runnable artifact.

3. Key Contributions

1. Novel Paradigm: Introduces the concept of the "group chat thread" as a persistent computational workspace rather than just a front-end for one-shot Q&A. It integrates conversational context, file uploads, and execution results into a single continuous loop.
2. Robust System Architecture: Implements a scalable, multi-platform orchestration layer with per-group isolation, dual-backend model support, and secure host-container communication mechanisms.
3. Comprehensive Bioinformatics Runtime: Delivers a production-ready environment with 31+ pre-installed tools and 95+ skills covering genomics, structural biology, and clinical data, accessible via natural language.
4. Real-World Validation: Successfully deployed across 8 major messaging platforms and validated through diverse case studies, proving the feasibility of chat-native scientific discovery.

4. Results and Representative Applications

The authors validated BioClaw through representative case studies spanning various data modalities and domains:

• Sequencing & QC: Processing raw FASTQ files with FastQC and returning quality control summaries directly in chat.
• Sequence Analysis: Performing BLAST searches with structured top-hit reporting.
• Data Visualization: Generating publication-quality Volcano plots from differential expression data (CSV) without external software.
• Structural Biology: Rendering protein structures via PyMOL, predicting structures using ESMFold, and visualizing ligand binding sites.
• Literature & Wet-Lab: Retrieving and summarizing PubMed papers and analyzing experimental SDS-PAGE gel images uploaded by users.
• Long-Horizon Workflows: Demonstrated a "Manuscript-planning" skill that decomposes research ideas into structured stages (novelty assessment, task definition, figure planning), allowing iterative human feedback at checkpoints.

Deployment: The system is live on Discord, WhatsApp, Feishu, QQ, WeCom, WeChat, Slack, and a local web UI, demonstrating cross-platform interoperability.

5. Significance and Future Outlook

• Transformative Collaboration: BioClaw bridges the gap between scientific discussion and executable analysis, shortening the path from hypothesis to verification. It preserves the continuity of collaborative reasoning by keeping all artifacts within the team's native communication channel.
• Reproducibility: The automatic generation of Jupyter notebooks from chat interactions addresses a critical need in computational biology for reproducible, auditable workflows.
• Scalability: The architecture supports a "multi-agent" future where specialized "Claws" (e.g., a sequencing agent, a structural biology agent) can operate simultaneously within the same group chat, mimicking the division of labor in real scientific teams.
• Accessibility: By lowering the barrier to entry for complex bioinformatics tools, BioClaw democratizes access to advanced computational analysis for researchers who may not be proficient in coding or command-line interfaces.

In summary, BioClaw represents a significant step toward conversational and reproducible computational biology, establishing group chats as a viable, secure, and powerful environment for scientific discovery.