TransMAgent: Cross-Platform Multi-Agent Transcriptional Regulation Analysis System

transmagent_1080p.mp4

English | 中文

TransMAgent is a cross-platform multi-agent transcription regulation analysis system supporting Windows, macOS, and Linux. It integrates enhanced memory mechanisms, multi-agent collaboration architecture, MCP tool services, virtualized secure execution environments, and large-scale transcription regulation databases, providing bioinformatics researchers with an intelligent, all-in-one transcription regulation analysis platform.

Recent Updates

Update Log

• 2025-11-26 — New: Context Compression
  • Overview: Added a "Context Compression" option in the subtask processing flow to automatically streamline historical dialogues and intermediate results, reducing context length and memory usage.
  • Effect: Enhances stability and execution efficiency for long conversations and complex tasks, minimizing redundant information during model calls.
• 2025-11-28 — New: Code IDE Mode
  • Overview: Added code and text editor mode, supporting syntax highlighting, AI completion, AI refactoring, and AI conversational correction features.
  • Effect: Provides a more intelligent code and process writing experience, assisting in the design of complex analysis workflows.

• 2026-2-4 — New: Cross-Session Memory Storage and Retrieval, Important Memory Writing Function
  • Overview: Added cross-session memory storage and retrieval functionality, allowing users to write important memories into persistent storage and retrieve them in subsequent sessions through indexing.
  • Effect: Solves the "memory loss" problem in long-context tasks, enhancing continuity and stability of agents in complex analysis processes.
• 2026-2-6 — New: Heartbeat Mechanism
  • Overview: Added a heartbeat mechanism where agents periodically send heartbeat signals during execution to monitor system status and task progress.
  • Effect: Improves system stability and reliability, allowing timely detection and handling of potential execution issues.
• 2026-3-18 — New: LLM-as-a-Judge for Dynamic Payload Auditing and Runtime Interruption
  • Overview: Integrated an implicit, sandboxed Critic Agent to verify high-sensitivity tool payloads prior to execution, alongside a System Sentinel that asynchronously monitors real-time console outputs.
  • Effect: Accurately intercepts hallucinated data and unauthorized requests, and autonomously halts anomalous processes (e.g., infinite loops, interactive hanging, resource exhaustion), thereby safeguarding host integrity and optimizing compute efficiency.
• 2026-4-6 — New: Multi-Session Parallel Management System
  • Overview: Restructured the underlying session orchestration architecture to support the concurrent execution of multiple independent agent workflows within strictly isolated memory sandboxes.
  • Effect: Achieves absolute contextual and resource isolation across different analytical tasks, granting researchers the ability to seamlessly checkout, switch, and orchestrate multiple time-intensive pipelines simultaneously.

✨ Core Features

📋 Work Execution Flow

TransMAgent interacts with backend large language models through natural language conversations, offering four working modes: Auto Mode, Execution Mode, Planning Mode, and Flash Mode. Users can switch between modes to complete various complex transcription regulation analysis tasks.

Core Workflow:

• Planning Mode: The agent deeply interacts with the user to gather details like data sources and analysis requirements, generating specific subtask plans.
• Execution Mode: The agent automatically thinks and calls appropriate tools based on subtasks, displaying the execution process and results in real-time.
• Auto Mode: No user interaction allowed; all problems are automatically attempted to be solved, suitable for fully automated scenarios.
• Flash Mode: Quick responses to simple questions.

The system adopts a Five-Stage Task Flow: Task Parsing → Task Decomposition → File Validation → Chain Reasoning → Result Summarization, ensuring systematic and complete analysis processes.

🧰 BioTools Integration

We provide a general transcription regulation MCP service BioTools, integrating over 30 bioinformatics analysis tools (e.g., HOMER, DEEptools, ROSE, BETA, TRAPT, etc.).
The BioTools MCP service uses Docker virtualization and streaming HTTP communication mechanisms for unified cloud and local tool calls, allowing users to quickly extend or customize analysis tools.

Core Tool Set:

• add_subtasks: Task initialization and decomposition
• complete_subtasks: Dynamically update subtask status
• mcp_server: Core scheduling interface for calling analysis tools and databases
• ask_followup_question: Proactively ask clarifying questions
• memory_retrieval: Retrieve historical interaction records
• enter_idle_state: Structured task conclusion and summarization

🐳 Virtualized Environment

Utilizes Docker container virtualization technology to effectively isolate execution environments and prevent system-level command errors.
All results are saved in user-specified directories, and support SSH remote command execution for running compute-intensive analyses on high-performance servers.

🔃 Agent Working Modes

We integrate three Agent modes for users to choose from:

• TransAgent: Default mode, serially schedules tasks and tools, follows "subtask-record-reflection" process

• MultAgent: Multi-agent collaboration mode, uses task document-driven division of labor architecture

• BaseAgent: Basic mode for general instruction processing

MultAgent Sub-Agent System:

Sub-Agent	Function
workflow_planner	Analyzes tasks and plans complete workflows
task_executor	Responsible for command execution and result validation
tool_manager	Automatically installs, fixes, updates analysis tools
error_solution_finder	Handles programming and package dependency errors
web_searcher / url_summarizer	Automatic web search and webpage information integration
chart_plotter	Generates high-quality charts (supports ggplot2)

Through this "expert division of labor" structure, the system achieves efficiency and stability in complex analyses.

🔄 Agent Behavior Modes

The system provides four behavior modes that users can dynamically switch between:

• Auto Mode: Fully automated execution without manual intervention
• Execution Mode: Allows interaction and mid-process feedback
• Planning Mode: Generates complete analysis plan before execution
• Quick Response Mode: Prioritizes answering brief questions

This flexible mode switching supports phased processing and debugging of complex research tasks.

⚡ MCP Service Support

TransMAgent supports rapid integration of external MCP (Model Context Protocol) tool services through configuration files, enabling fast plugin-like extensions, including command-line tools, database query services, and third-party API interfaces.

🛠️ Custom Tools

To facilitate users in designing more system-level functionalities, we allow custom tool definitions, including tool calls and tool prompts, to meet personalized analysis needs.

🎛️ System/CLI Tool Prompts

We reserve injection interfaces for system and CLI tool prompts. Researchers can easily fine-tune agent behavior by modifying configuration files, enabling rapid tool customization.

🧩 User Interaction Controls

Each agent output includes interactive controls:

• 👍 Like: Provide feedback on result quality
• ❌ Delete: Remove redundant or incorrect outputs
• 📍 Locate: Jump to corresponding JSON content
• 🪶 Quote: Extract paragraphs for subsequent analysis

🧠 Enhanced Memory Module

To address the "memory loss" problem in long-context tasks, TransMAgent employs a 2D Hybrid Memory Mechanism:

• Precise Memory (PM): Stores complete content of recent interactions
• Fuzzy Memory (FM): Records long-term thought content and indexes
• Memory Retrieval Tool: Can backtrack historical conversation details based on indexes
• Dynamic Memory Management: Users can manually enable or disable memory at any step, achieving "atomic-level control"

The system uses a standardized JSON output format to ensure agent output consistency and parsability:

{
  "thinking": "Current step's thought process and tool call rationale",
  "tool": "Called tool name", 
  "params": "Tool call parameters"
}

🧬 Internal Multi-Omics Annotation Resources

TransMAgent includes a rich transcription regulation resource library:

Data Type	Data Sources	Data Scale
Super Enhancers	SEdb, SEA, dbSuper	2,678,273 entries
Regular Enhancers	EnhancerAtlas, HACER, ENCODE, FANTOM5, etc.	14,797,266 entries
eRNA	eRNAbase	10,399,928 entries
Common SNPs	dbSNP	37,302,978 entries
Risk SNPs	GWAS Catalog, GWASdb	351,728 entries
eQTL Sites	GTEx, PancanQTL, seeQTL, etc.	11,995,221 entries
Chromatin Accessibility	ATACdb, ENCODE DNase-seq	Over 130 million regions
3D Genome Structure	4DGenome, 3D Genome Browser	34,342,926 interaction sites
DNA Methylation	ENCODE, TCGA	Over 190 million sites
Expression Profiles	GTEx, TCGA, CCLE, ENCODE	Multiple tissues and cancers

These data are deeply integrated with MCP services, enabling cross-omics annotation and regulatory network reconstruction.

⚗️ Model Training and Reinforcement Learning

TransMAgent's underlying large language model is fine-tuned based on Qwen3-8B in two stages:

Stage 1: Supervised Fine-Tuning

• Dataset Generation: Generate high-quality Q/A pairs via API calls to DeepSeek-V3
• Data Content: Tool usage instructions + transcription regulation expertise
• Training Method: Uses LoRA low-rank adaptation technology, trained in 8*A800 GPU distributed environment
• Optimization Objective: Maximum likelihood estimation with sliding window for long contexts

Stage 2: Reinforcement Learning Fine-Tuning

• Dataset Construction: Filter high-quality trajectory data, use 2D sliding window to simulate real environment
• Training Method: Group Relative Policy Optimization (GRPO) method to eliminate length bias
• Reward Functions:
  • MCP Tool Validation Reward: Checks parameter completeness and validity
  • Text Similarity Reward: Combines Jaccard, cosine, and edit distance similarities
  • Format Compliance Reward: Ensures standard JSON output

Reward Function Formula:

Comprehensive Similarity = w1×Jaccard Similarity + w2×Cosine Similarity + w3×Edit Distance Similarity

Optimization Objective:

L_GRPO = -E[Â_{i,t} × log(π_ϕ/π_ref)] + β×D_KL(π_ϕ||π_ref)

This method significantly improves model stability and accuracy in complex tool calling and transcription regulation reasoning tasks.

🧱 Core Architecture Principles

TransMAgent adopts a Standardized Task Processing Architecture:

1. Task Parsing: Clarify user requirements, handle ambiguous or vague questions
2. Task Decomposition: Use add_subtasks to break down complex problems into manageable subtasks
3. File Validation: Automatically verify file authenticity and structural integrity
4. Chain Reasoning: Call biological toolkits via mcp_server, execute system commands via cli_execute
5. Result Summarization: Generate analysis reports, recommend follow-up questions

Each subtask completion automatically triggers a reflection mechanism to continuously optimize execution logic.

💡 Ollama Support

To ensure data security and privacy, the system supports connecting to Ollama local large models, enabling completely offline agent reasoning and tool control.

📊 Other Features

To meet researchers' one-stop needs, we also integrate:

• Basic conversation functionality
• Chain calling
• Session saving and loading
• Automated dataset review website

🎯 Application Examples

• TransAgent Mode Demo:

• MultAgent Mode Demo:

📋 System Requirements

• Windows: Windows 10 or later
• Ubuntu: Ubuntu 18.04 or later
• macOS: macOS 10.14 or later

🐳 Docker Environment Building

Before using TransMAgent, you need to build the BioTools MCP service environment. BioTools integrates over 30 bioinformatics analysis tools, using Docker virtualization container technology for secure, isolated cloud and local unified tool calling.

⚠️ Important: Please complete Docker environment building before software installation or build from source.

For detailed build instructions, see: BioTools MCP Server Environment Setup Guide

🚀 Quick Start

📦 Software Installation

Directly download and use pre-built packages, no additional configuration required:

📥 Go to GitHub Releases to download the latest version

💡 Download the compressed package for your platform, extract and run the executable directly.

🔨 Build from Source

💡 Due to rapid version iterations, we recommend compiling yourself to experience the latest features.

Build from source to get the latest features and performance optimizations:

# Use Node.js 23
nvm use 23

# Install dependencies
pnpm install

# Compile TypeScript code
cd src/backend/src
pnpm run watch
cd src/frontend/js/main
pnpm run watch

# Start application
pnpm run start

# Package application
pnpm run dist

⚙️ Installation Configuration

TransMAgent uses a two-level configuration file mechanism:

Type	Path	Description
System Default Config	src/backend/configs/config_*.json	Factory default configuration
User Config	~/.transmagent/story.json	User custom configuration (auto-created)

Configuration Loading Mechanism

System Default Config + User Config → Merged Enhanced Config

• User config overrides system default config
• If user config doesn't exist, system will auto-create it

Agent Mode Configuration Files

Mode	System Config File
TransAgent (default)	configs/config_transagent.json
BaseAgent	configs/config_baseagent.json
MultiAgent	configs/config_multagent.json

🔧 Agent Parameter Configuration

See mcp_server

📦 Tool Dependency Installation

See plugins

🤖 Large Model and Software Detailed Configuration

Refer to Configuration Example for advanced configurations like Ollama Support

📚 Full Documentation

For comprehensive usage guides, please refer to the complete documentation:

Documentation Type	Path	Description
User Guide	docs/users/	Detailed installation, interface operation, Agent usage tutorials
Developer Documentation	docs/developer/	Project architecture, API documentation, development guide

🌐 API Interface

Print Session List

curl -X POST http://localhost:3005/chat/list \
  -H "Content-Type: application/json"

Start New Session

curl -X POST http://localhost:3005/chat/checkout \
  -H "Content-Type: application/json"

Switch Session

curl -X POST http://localhost:3005/chat/checkout \
  -H "Content-Type: application/json" \
  -d '{"chat_id": "chat-ba0091e2-a942-425d-8783-115d82011781"}'

Switch Mode

curl -X POST http://localhost:3005/chat/mode \
  -H "Content-Type: application/json" \
  -d '{"mode": "auto/act/plan/flash"}'

Switch Agent Mode (transagent/baseagent/multagent)

curl -X POST http://localhost:3005/chat/agent_mode \
  -H "Content-Type: application/json" \
  -d '{"agent_mode": "transagent/baseagent/multagent"}'

Switch Model

curl -X POST http://localhost:3005/chat/model \
  -H "Content-Type: application/json" \
  -d '{"model": "runapi"}'

Switch Tool Format (toolcalls/prompt)

curl -X POST http://localhost:3005/chat/tool_format \
  -H "Content-Type: application/json" \
  -d '{"tool_format": "toolcalls"}'

Submit Query

curl -X POST http://localhost:3005/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [
        {"role": "user", "content": "Hello"}
    ],
    "max_step": 2
  }'

🧠 AI Agent Evaluation Website

TransMAgent includes Eval Agent Web — a web-based multi-model AI Agent benchmarking platform with 10-dimension automated evaluation and visualization reports.

Quick Start

# 1. Enter the evaluation website directory
cd eval_agent_web

# 2. Install dependencies
pip install -r requirements.txt

# 3. Start the server (default: 0.0.0.0:5000)
python app.py

Open http://localhost:5000 in your browser to access the evaluation interface.

Core Features

Feature	Description
Multi-format Upload	Supports model result JSON (transagent_result_test_set format) and generic CSV/JSON
10-Dimension Evaluation	Data Authenticity, Response Relevance, Tool Use Accuracy, Hallucination Control, Code Execution, Domain Knowledge, Result Interpretation, Autonomous Planning, Error Recovery, Output Standardization
Real-time Progress	SSE-based real-time evaluation progress streaming
Visualization Reports	Auto-generated model comparison bar charts, dimension heatmaps, radar charts (PDF)
Checkpoint Resume	Supports resuming from interruption checkpoints
One-click Export	Download all evaluation results as ZIP

Workflow

1. Configure API (supports DeepSeek / OpenAI-compatible) → 2. Upload model output files → 3. Start evaluation → 4. View reports & charts

📞 Contact Us

For any questions, please contact us via:
📧 Email: mp798378522@gmail.com

TransMAgent - Making Transcription Regulation Analysis Simple and Efficient