Product Requirements Document: Protein Binder Design MCP Server

Executive Summary

This document outlines the requirements for building an MCP (Model Context Protocol) server that enables LLM agents to run end-to-end protein binder design pipelines. The server wraps RFdiffusion, ProteinMPNN, and ESMFold into high-level, purpose-driven tools that facilitate computational protein design workflows.

Problem Statement

Protein binder design is a complex multi-step process that requires:

1. Generating novel protein backbones that can bind to target proteins (RFdiffusion)
2. Designing amino acid sequences for the generated backbones (ProteinMPNN)
3. Validating that designed sequences fold correctly (ESMFold)

Current approaches require researchers to manually orchestrate these tools, understand their APIs, and handle complex parameter configurations. This creates barriers for:

• Non-expert users trying to leverage AI-driven protein design
• LLM agents attempting to automate protein design workflows
• Rapid prototyping and iteration on design hypotheses

Goals

Primary Goals

1. Enable LLM-driven protein design: Allow AI agents to autonomously execute complete binder design workflows
2. Simplify complex pipelines: Abstract away tool-specific details into intuitive, high-level operations
3. Maintain scientific rigor: Ensure generated designs follow best practices and include quality metrics

Success Metrics

• LLM agents can successfully complete binder design tasks with >90% accuracy in tool selection
• Time to first design reduced by 80% compared to manual workflows
• All designs include validation metrics (pLDDT, pTM, interface metrics)

Key Design Principles

1. Few, High-Level Tools

Research shows LLM tool selection accuracy drops significantly beyond 30 tools. This server MUST maintain fewer than 10 tools total.

Principle	Rationale
Maximum 10 tools	Ensures reliable LLM tool selection
Each tool = complete workflow	Users think in workflows, not API calls
Sensible defaults	Minimize required parameters

2. Workflow-Based Design

Each tool represents a complete user workflow, not a raw API call:

• Bad: run_rfdiffusion(params...) → run_proteinmpnn(params...) → run_esmfold(params...)
• Good: design_binder(target_pdb, hotspot_residues) → returns validated designs

3. Internal Orchestration

Complex multi-step pipelines should be handled internally:

• RFdiffusion → ProteinMPNN → ESMFold validation runs as single tool call
• Retry logic and error handling managed internally
• Progress reporting via MCP notifications

Functional Requirements

Core Tools

1. design_binder

Purpose: Complete end-to-end binder design pipeline

Inputs:

Parameter	Type	Required	Description
target_pdb	string	Yes	Path to target protein PDB file
hotspot_residues	string[]	Yes	Residues on target for binder interface (e.g., ["A45", "A46", "A49"])
num_designs	int	No	Number of designs to generate (default: 10)
binder_length	int	No	Length of binder in residues (default: 80)

Outputs:

{
  "designs": [
    {
      "id": "design_001",
      "sequence": "MKLLVVF...",
      "structure_pdb": "/path/to/structure.pdb",
      "metrics": {
        "plddt": 85.2,
        "ptm": 0.78,
        "interface_plddt": 82.1,
        "binding_energy": -12.5
      }
    }
  ],
  "summary": {
    "total_generated": 10,
    "passed_filters": 7,
    "best_design_id": "design_003"
  }
}

Internal Pipeline:

1. Run RFdiffusion to generate backbone structures
2. Run ProteinMPNN to design sequences (8 sequences per backbone)
3. Run ESMFold to predict structures of designed sequences
4. Calculate metrics and filter designs
5. Return ranked results

2. analyze_interface

Purpose: Analyze protein-protein interface properties

Inputs:

Parameter	Type	Required	Description
complex_pdb	string	Yes	Path to protein complex PDB
chain_a	string	Yes	Chain ID of first protein
chain_b	string	Yes	Chain ID of second protein

Outputs:

{
  "interface_residues": {
    "chain_a": ["45", "46", "49", "52"],
    "chain_b": ["12", "15", "16"]
  },
  "buried_surface_area": 1250.5,
  "hydrogen_bonds": 8,
  "salt_bridges": 2,
  "hydrophobic_contacts": 15,
  "shape_complementarity": 0.72
}

3. validate_design

Purpose: Validate a designed protein sequence

Inputs:

Parameter	Type	Required	Description
sequence	string	Yes	Amino acid sequence
expected_structure	string	No	Path to expected structure PDB for comparison

Outputs:

{
  "predicted_structure_pdb": "/path/to/prediction.pdb",
  "plddt": 87.3,
  "ptm": 0.82,
  "pae_matrix": [[...], ...],
  "rmsd_to_expected": 1.2,  // if expected_structure provided
  "secondary_structure": "HHHHHHHCCCCEEEEEE..."
}

4. optimize_sequence

Purpose: Optimize an existing binder sequence for improved properties

Inputs:

Parameter	Type	Required	Description
current_sequence	string	Yes	Starting sequence
target_pdb	string	Yes	Target protein PDB
optimization_target	string	No	What to optimize: "stability", "affinity", "both" (default: "both")
fixed_positions	int[]	No	Positions to keep fixed

Outputs:

{
  "optimized_sequence": "MKVVLLF...",
  "mutations": ["A5V", "L12M", "G45A"],
  "predicted_improvement": {
    "stability_delta": "+2.1 kcal/mol",
    "affinity_delta": "-0.8 kcal/mol"
  },
  "metrics": {
    "plddt": 89.1,
    "interface_score": -15.2
  }
}

5. suggest_hotspots

Purpose: Analyze a target protein and suggest potential binding hotspots using multiple data sources

Inputs:

Parameter	Type	Required	Description
target_pdb	string	Yes	Path to target protein PDB
chain_id	string	No	Specific chain to analyze (default: first chain)
criteria	string	No	"druggable", "exposed", "conserved" (default: "exposed")
uniprot_id	string	No	UniProt ID for retrieving known binding sites and functional annotations
include_literature	bool	No	Search PubMed for known binding partners (default: false)

Data Sources:

1. Structural Analysis: SASA-based surface exposure, pocket detection, hydrophobic patches
2. UniProt Integration: Known binding sites, active sites, functional regions from UniProt API
3. Conservation Scoring: Evolutionary conservation from BLAST/MSA analysis
4. Literature Mining: PubMed search for known binding partners and interaction sites

Outputs:

{
  "suggested_hotspots": [
    {
      "residues": ["A45", "A46", "A49"],
      "score": 0.92,
      "rationale": "Large exposed hydrophobic patch",
      "evidence": {
        "structural": {"sasa": 245.3, "pocket_score": 0.85},
        "conservation": 0.78,
        "uniprot_annotation": "Catalytic site",
        "literature_support": ["PMID:12345678"]
      }
    },
    {
      "residues": ["A120", "A123", "A124"],
      "score": 0.85,
      "rationale": "Known functional site from UniProt",
      "evidence": {
        "structural": {"sasa": 180.2, "pocket_score": 0.65},
        "conservation": 0.92,
        "uniprot_annotation": "Binding site for ATP",
        "literature_support": []
      }
    }
  ],
  "surface_analysis": {
    "total_surface_area": 15230.5,
    "hydrophobic_patches": 3,
    "charged_regions": 5,
    "detected_pockets": [
      {"center": [12.3, 45.6, 78.9], "volume": 523.4, "druggability": 0.82}
    ]
  },
  "conservation_profile": {
    "average_conservation": 0.65,
    "highly_conserved_residues": ["A45", "A120", "A121"]
  },
  "uniprot_features": {
    "binding_sites": [{"position": "45-49", "ligand": "ATP"}],
    "active_sites": [{"position": "120", "type": "catalytic"}],
    "known_interactors": ["P12345", "Q67890"]
  },
  "literature_insights": {
    "known_binding_partners": ["ProteinX", "ProteinY"],
    "relevant_publications": [
      {"pmid": "12345678", "title": "Crystal structure of...", "binding_residues": ["A45", "A46"]}
    ]
  }
}

6. get_design_status

Purpose: Check status of running design jobs (for long-running operations)

Inputs:

Parameter	Type	Required	Description
job_id	string	Yes	Job ID from design_binder call

Outputs:

{
  "status": "running",  // "queued", "running", "completed", "failed"
  "progress": {
    "current_step": "proteinmpnn",
    "designs_completed": 5,
    "total_designs": 10
  },
  "estimated_time_remaining": "5 minutes"
}

Resources (Read-Only Data)

1. protein://structures/{pdb_id}

Access to PDB structures by ID

2. protein://designs/{job_id}/{design_id}

Access to generated design files

Non-Functional Requirements

Performance

• Single binder design (10 designs) should complete in <30 minutes on GPU
• ESMFold predictions should complete in <2 minutes per sequence
• Server should handle concurrent requests via job queue

Reliability

• All operations should be idempotent where possible
• Failed jobs should be recoverable/resumable
• Results should be cached to avoid redundant computation

Usability

• Clear error messages with actionable suggestions
• Progress reporting for long-running operations
• Example prompts and workflows in documentation

Security

• Input validation for all PDB files
• Sandboxed execution of computational tools
• Rate limiting for resource-intensive operations

Technical Architecture

Technology Stack

• MCP Framework: Python MCP SDK
• Core Tools:
  • RFdiffusion (structure generation)
  • ProteinMPNN (sequence design)
  • ESMFold (structure prediction)
• Supporting Libraries:
  • BioPython (PDB parsing)
  • PyRosetta (optional, for interface analysis)
  • NumPy/SciPy (numerical operations)

System Components

┌─────────────────────────────────────────────────────────────┐
│                      MCP Server                              │
├─────────────────────────────────────────────────────────────┤
│  Tools Layer                                                 │
│  ┌──────────────┐ ┌──────────────┐ ┌──────────────┐        │
│  │design_binder │ │analyze_iface │ │validate_seq  │  ...    │
│  └──────────────┘ └──────────────┘ └──────────────┘        │
├─────────────────────────────────────────────────────────────┤
│  Orchestration Layer                                         │
│  ┌──────────────────────────────────────────────────────┐  │
│  │  Pipeline Manager (RFdiffusion→MPNN→ESMFold)         │  │
│  │  Job Queue & Status Tracking                          │  │
│  │  Result Caching                                       │  │
│  └──────────────────────────────────────────────────────┘  │
├─────────────────────────────────────────────────────────────┤
│  Computation Layer                                           │
│  ┌────────────┐ ┌────────────┐ ┌────────────┐              │
│  │RFdiffusion │ │ProteinMPNN │ │  ESMFold   │              │
│  └────────────┘ └────────────┘ └────────────┘              │
└─────────────────────────────────────────────────────────────┘

Directory Structure

protein-design-mcp/
├── src/
│   └── protein_design_mcp/
│       ├── __init__.py
│       ├── server.py              # MCP server entry point
│       ├── tools/
│       │   ├── __init__.py
│       │   ├── design_binder.py   # Main binder design tool
│       │   ├── analyze.py         # Interface analysis tools
│       │   ├── validate.py        # Validation tools
│       │   └── optimize.py        # Sequence optimization
│       ├── pipelines/
│       │   ├── __init__.py
│       │   ├── rfdiffusion.py     # RFdiffusion wrapper
│       │   ├── proteinmpnn.py     # ProteinMPNN wrapper
│       │   └── esmfold.py         # ESMFold wrapper
│       ├── utils/
│       │   ├── __init__.py
│       │   ├── pdb.py             # PDB file utilities
│       │   ├── metrics.py         # Quality metrics
│       │   └── cache.py           # Result caching
│       └── resources/
│           ├── __init__.py
│           └── structures.py      # PDB structure access
├── tests/
│   ├── __init__.py
│   ├── test_tools.py
│   ├── test_pipelines.py
│   └── fixtures/
│       └── test_pdbs/
├── examples/
│   ├── basic_binder_design.py
│   └── batch_design.py
├── docs/
│   └── usage.md
├── environment.yml
├── pyproject.toml
├── CLAUDE.md
├── README.md
└── LICENSE

Implementation Phases

Phase 1: Foundation

• Set up project structure and MCP server skeleton
• Implement validate_design tool with ESMFold integration
• Basic PDB parsing utilities
• Unit tests for core functionality

Phase 2: Core Pipeline

• Implement RFdiffusion wrapper
• Implement ProteinMPNN wrapper
• Implement design_binder tool with full pipeline
• Job queue and status tracking

Phase 3: Analysis Tools

• Implement analyze_interface tool
• Implement suggest_hotspots tool
• Add quality metrics calculations

Phase 4: Optimization

• Implement optimize_sequence tool
• Add result caching
• Performance optimization

Phase 5: Polish

• Comprehensive documentation
• Example workflows
• Integration tests
• Error handling improvements

Risks and Mitigations

Risk	Impact	Mitigation
GPU memory limitations	Can't process large targets	Implement chunking, provide clear size limits
Long computation times	Poor user experience	Job queue, progress reporting, caching
Model availability	External dependencies	Containerized deployment, fallback options
LLM tool confusion	Incorrect tool selection	Clear tool descriptions, examples, <10 tools

Appendix

Example LLM Interactions

Simple Binder Design:

User: Design a binder for the protein in target.pdb, targeting residues 45-50 on chain A

LLM: I'll use the design_binder tool to create binders targeting those residues.
[Calls design_binder with target_pdb="target.pdb", hotspot_residues=["A45","A46","A47","A48","A49","A50"]]

Interface Analysis:

User: Analyze the interface between chains A and B in my complex

LLM: I'll analyze the protein-protein interface for you.
[Calls analyze_interface with complex_pdb="complex.pdb", chain_a="A", chain_b="B"]

References

• MCP Specification
• RFdiffusion Paper
• ProteinMPNN Paper
• ESMFold Paper