Skip to content

PRIME Flow

The PRIME (Protein Research and Innovation in Molecular Engineering) flow provides comprehensive protein engineering capabilities with 65+ specialized tools across six categories.

Overview

The PRIME flow implements the three-stage architecture described in the PRIME paper: 1. Parse - Query analysis and scientific intent detection 2. Plan - Workflow construction and tool selection 3. Execute - Tool execution with adaptive re-planning

Architecture

graph TD
    A[Research Query] --> B[Parse Stage]
    B --> C[Scientific Intent Detection]
    C --> D[Domain Heuristics]
    D --> E[Plan Stage]
    E --> F[Tool Selection]
    F --> G[Workflow Construction]
    G --> H[Execute Stage]
    H --> I[Tool Execution]
    I --> J[Adaptive Re-planning]
    J --> K[Results & Reports]

Configuration

Basic Configuration

# Enable PRIME flow
flows:
  prime:
    enabled: true
    params:
      adaptive_replanning: true
      manual_confirmation: false
      tool_validation: true

Advanced Configuration

# configs/statemachines/flows/prime.yaml
enabled: true
params:
  adaptive_replanning: true
  manual_confirmation: false
  tool_validation: true
  scientific_intent_detection: true

  domain_heuristics:
    - immunology
    - enzymology
    - cell_biology

  tool_categories:
    - knowledge_query
    - sequence_analysis
    - structure_prediction
    - molecular_docking
    - de_novo_design
    - function_prediction

  execution:
    max_iterations: 10
    convergence_threshold: 0.95
    timeout_per_step: 300

Usage Examples

Basic Protein Design

uv run deepresearch \
  flows.prime.enabled=true \
  question="Design a therapeutic antibody for SARS-CoV-2 spike protein"

Protein Structure Analysis

uv run deepresearch \
  flows.prime.enabled=true \
  question="Analyze the structure of protein P12345 and predict its function"

Multi-Domain Research

uv run deepresearch \
  flows.prime.enabled=true \
  question="Design an enzyme with improved thermostability for industrial applications"

Tool Categories

1. Knowledge Query Tools

Tools for retrieving biological knowledge and literature:

  • UniProt Query: Retrieve protein information and annotations
  • PDB Query: Access protein structure data
  • PubMed Search: Find relevant research literature
  • GO Annotation: Retrieve Gene Ontology terms and annotations

2. Sequence Analysis Tools

Tools for analyzing protein sequences:

  • BLAST Search: Sequence similarity search
  • Multiple Sequence Alignment: Align related sequences
  • Motif Discovery: Identify functional motifs
  • Physicochemical Analysis: Calculate sequence properties

3. Structure Prediction Tools

Tools for predicting protein structures:

  • AlphaFold2: AI-powered structure prediction
  • ESMFold: Evolutionary scale modeling
  • RoseTTAFold: Deep learning structure prediction
  • Homology Modeling: Template-based structure prediction

4. Molecular Docking Tools

Tools for analyzing protein-ligand interactions:

  • AutoDock Vina: Molecular docking simulations
  • GNINA: Deep learning docking
  • Interaction Analysis: Binding site identification
  • Affinity Prediction: Binding energy calculations

5. De Novo Design Tools

Tools for designing novel proteins:

  • ProteinMPNN: Sequence design from structure
  • RFdiffusion: Structure generation
  • Ligand Design: Small molecule design
  • Scaffold Design: Protein scaffold engineering

6. Function Prediction Tools

Tools for predicting protein functions:

  • EC Number Prediction: Enzyme classification
  • GO Term Prediction: Function annotation
  • Binding Site Prediction: Interaction site identification
  • Stability Prediction: Thermal and pH stability analysis

Scientific Intent Detection

PRIME automatically detects the scientific intent of queries:

# Example classifications
intent_detection = {
    "protein_design": "Design new proteins with specific properties",
    "binding_analysis": "Analyze protein-ligand interactions",
    "structure_prediction": "Predict protein tertiary structure",
    "function_annotation": "Annotate protein functions",
    "stability_engineering": "Improve protein stability",
    "catalytic_optimization": "Optimize enzyme catalytic properties"
}

Domain Heuristics

PRIME uses domain-specific heuristics for different biological areas:

Immunology

  • Antibody design and optimization
  • Immune response modeling
  • Epitope prediction and analysis
  • Vaccine development workflows

Enzymology

  • Enzyme kinetics and mechanism analysis
  • Substrate specificity engineering
  • Catalytic efficiency optimization
  • Industrial enzyme design

Cell Biology

  • Protein localization prediction
  • Interaction network analysis
  • Cellular pathway modeling
  • Organelle targeting

Adaptive Re-planning

PRIME implements sophisticated re-planning strategies:

Strategic Re-planning

  • Tool substitution when tools fail or underperform
  • Algorithm switching (BLAST → ProTrek, AlphaFold2 → ESMFold)
  • Resource reallocation based on intermediate results

Tactical Re-planning

  • Parameter adjustment for better results
  • E-value relaxation for broader searches
  • Exhaustiveness tuning for docking simulations

Execution Monitoring

PRIME tracks execution across multiple dimensions:

Quality Metrics

  • pLDDT Scores: Structure prediction confidence
  • E-values: Sequence similarity significance
  • RMSD Values: Structure alignment quality
  • Binding Energies: Interaction strength validation

Performance Metrics

  • Execution Time: Per-step and total workflow timing
  • Resource Usage: CPU, memory, and storage utilization
  • Tool Success Rates: Individual tool performance tracking
  • Convergence Analysis: Workflow convergence patterns

Output Formats

PRIME generates multiple output formats:

Structured Reports

{
  "workflow_id": "prime_20241207_143022",
  "query": "Design therapeutic antibody",
  "scientific_domain": "immunology",
  "intent": "protein_design",
  "results": {
    "structures": [...],
    "sequences": [...],
    "analyses": [...]
  },
  "execution_summary": {
    "total_time": 2847.2,
    "tools_used": 12,
    "success_rate": 0.92
  }
}

Visualization Outputs

  • Protein structure visualizations (PyMOL, NGL View)
  • Sequence alignment diagrams
  • Interaction network graphs
  • Performance metric charts

Publication-Ready Reports

  • LaTeX-formatted academic papers
  • Jupyter notebooks with interactive analysis
  • HTML reports with embedded visualizations

Integration Examples

With Bioinformatics Flow

uv run deepresearch \
  flows.prime.enabled=true \
  flows.bioinformatics.enabled=true \
  question="Analyze TP53 mutations and design targeted therapies"

With DeepSearch Flow

uv run deepresearch \
  flows.prime.enabled=true \
  flows.deepsearch.enabled=true \
  question="Latest advances in protein design combined with structural analysis"

Best Practices

  1. Start Specific: Begin with well-defined protein engineering questions
  2. Use Domain Heuristics: Leverage appropriate domain knowledge
  3. Monitor Quality Metrics: Pay attention to confidence scores and validation metrics
  4. Iterative Refinement: Use intermediate results to guide subsequent steps
  5. Tool Validation: Ensure tool outputs meet quality thresholds before proceeding

Troubleshooting

Common Issues

Low Quality Predictions: 

# Increase tool validation thresholds
flows.prime.params.tool_validation=true
flows.prime.params.quality_threshold=0.8

Slow Execution: 

# Enable faster variants
flows.prime.params.use_fast_variants=true
flows.prime.params.max_parallel_tools=5

Tool Failures: 

# Enable fallback tools
flows.prime.params.enable_tool_fallbacks=true
flows.prime.params.retry_failed_tools=true

For more detailed information, see the Tool Development Guide and Tool Registry Documentation.