AiZynthFinder outputs & interpretation

AiZynthFinder returns ranked retrosynthetic routes as JSON and CSV, each with a score, step count, and a purchasability flag showing whether every starting material is in your selected stocks. Filter for purchasable, high-scoring routes first — those tell you which candidates are worth advancing to procurement (a ZINC purchasability check) or a Compound Passport synthesis section.

Output artifacts

AiZynthFinder generates multiple output files containing route information, scoring data, and metadata. All outputs are available for download from the job page or via API.

Primary artifacts

Routes JSON (routes.json)

Format: Hierarchical JSON structure containing complete retrosynthetic trees Contents:

• Full route trees with intermediates and reactions
• Template IDs and probabilities for each step
• Stock availability information for starting materials
• Route metadata (scores, lengths, purchasability)

Structure:

{
  "routes": [
    {
      "route_id": "route_001",
      "score": 0.85,
      "length": 3,
      "is_purchasable": true,
      "tree": {
        "smiles": "CCc1ccc(cc1)C(C)C(=O)O",
        "children": [
          {
            "smiles": "CCc1ccc(cc1)C(C)C(=O)Cl",
            "template_id": "ester_hydrolysis_rev",
            "template_prob": 0.92,
            "children": [...]
          }
        ]
      }
    }
  ]
}

Routes CSV (routes.csv)

Format: Tabular format with one row per route Columns:

• route_id: Unique identifier for the route
• score: Overall route score (0.0-1.0)
• length: Number of reaction steps
• is_purchasable: Boolean indicating if all starting materials are available
• template_ids: Comma-separated list of template IDs used
• target_smiles: Target molecule SMILES
• starting_materials: SMILES of required starting materials

Example:

route_id,score,length,is_purchasable,template_ids,target_smiles,starting_materials
route_001,0.85,3,true,ester_hydrolysis_rev,coupling,Cc1ccc(cc1)C(C)C(=O)O,Cc1ccc(cc1)C(C)C(=O)Cl.CO
route_002,0.72,4,false,suzuki_coupling_rev,ester_hydrolysis_rev,Cc1ccc(cc1)C(C)C(=O)O,Cc1ccc(cc1)B.O=C(Cl)C(C)C,CCO

Route analysis JSON (analysis.json)

Format: Detailed analysis of route characteristics Contents:

• Route complexity metrics
• Starting material availability breakdown
• Template usage statistics
• Stock coverage analysis

Metadata and logs

Job metadata (metadata.json)

Format: Complete configuration and runtime information Contents:

• Input parameters (SMILES, stocks, policies, filters)
• Model versions and policy information
• Search statistics (nodes explored, time taken)
• Stock database versions
• Template database information

Example:

{
  "input": {
    "smiles": "CCc1ccc(cc1)C(C)C(=O)O",
    "stocks": ["zinc", "emolecules"],
    "policy": ["uspto_50k"],
    "filters": ["price"]
  },
  "versions": {
    "aizynthfinder": "2.3.1",
    "policy": "uspto_50k_v1.2",
    "templates": "uspto_2023_q4"
  },
  "runtime": {
    "duration_seconds": 245.3,
    "nodes_explored": 15420,
    "routes_found": 67
  }
}

Search logs (search.log)

Format: Detailed execution log with timestamps Contents:

• Search progress and milestones
• Template application attempts
• Stock availability checks
• Performance metrics and warnings

Key metrics and fields

Route scoring

• Overall score: Composite metric (0.0-1.0) balancing:
  • Template probabilities (how likely each reaction is)
  • Route length (shorter routes preferred)
  • Stock availability (purchasable materials boost score)
• Score interpretation:

  • 0.8: High-confidence routes

  • 0.6-0.8: Moderate confidence
  • < 0.6: Low confidence, may need manual validation

State score vs. route score

• State score: Measures search tree node quality during exploration (used internally by MCTS/search algorithms)
• Route score: Final composite metric for completed routes, combining template probabilities across all steps
• Tree search metrics: The underlying Monte Carlo Tree Search (or alternative algorithms like Nested MCTS or Greedy Best First) use state scores to guide exploration

Success rate context

In published benchmarks, AiZynthFinder achieves 80-86% success rates for finding at least one route with all starting materials available in commercial stock databases. Routes with higher scores and shorter step counts generally correlate with experimental success.

Purchasability assessment

• is_purchasable: Boolean flag indicating complete stock coverage
• Stock checking: Each starting material verified against specified stock sources
• Partial availability: Routes may be partially purchasable if some materials are missing

Route complexity

• num_steps (length): Number of reaction steps required
• Template diversity: Number of different reaction types used
• Intermediate complexity: Molecular weight and synthetic complexity of intermediates

Template information

• template_ids: Machine-readable identifiers for applied reaction templates
• Template metadata: Reaction class, substrate scope, literature references
• Probability scores: Confidence in template applicability

Interpreting results

Route prioritization

1. Filter for purchasability: Focus on is_purchasable: true routes first
2. Sort by score: Higher scores indicate more reliable routes
3. Consider length: Shorter routes are generally preferred for simplicity
4. Review intermediates: Check for synthetic feasibility of key intermediates

Quality assessment

High-quality routes

• Score > 0.8 and purchasable
• 2-4 reaction steps
• Common reaction types (esterification, coupling, etc.)
• Commercially available starting materials

Moderate-quality routes

• Score 0.6-0.8 or partially purchasable
• May require custom synthesis of one intermediate
• Uncommon reactions but literature precedent exists

Low-quality routes

• Score < 0.6 or non-purchasable
• Very long sequences (>6 steps)
• Novel reactions without clear precedent
• Require rare or expensive starting materials

Common patterns to watch for

Realistic disconnections

• Good: Breaking at functional groups with clear synthetic precedent
• Questionable: Breaking stable rings or introducing unusual functionality
• Poor: Routes requiring protection/deprotection of every functional group

Starting material quality

• Good: Common building blocks, catalog compounds, natural products
• Questionable: Complex intermediates requiring multi-step synthesis
• Poor: Hypothetical compounds not available commercially

Export and sharing

Download options

• Individual files: Download specific artifacts from job page
• Bulk download: ZIP archive containing all outputs
• API access: The finished job record from GET /api/v1/jobs/{jobId} lists each artifact

Sharing for collaboration

• Job links: Include complete job URL in ELN entries
• Configuration export: Share exact parameters used for reproducibility
• Route visualization: Export routes as images or interactive viewers

Integration with other tools

• CRO packages: Include purchasable routes with cost estimates
• ELN integration: Embed route information in electronic lab notebooks
• Literature search: Use intermediate SMILES for patent/ literature checks

Troubleshooting outputs

Empty results

• Cause: No valid routes found within search constraints
• Solutions:
  • Expand stock sources
  • Use less restrictive policies
  • Increase search depth/limits
  • Check input SMILES validity

All routes non-purchasable

• Cause: Target requires intermediates not in selected stocks
• Solutions:
  • Add more comprehensive stock sources
  • Consider custom synthesis of key intermediates
  • Review if target molecule is synthetically accessible

Unexpectedly low scores

• Cause: Molecule outside training data scope or unusual chemistry
• Solutions:
  • Try different policies
  • Manually validate high-scoring routes
  • Consider alternative retrosynthesis tools

Performance issues

• Cause: Large search spaces or complex molecules
• Solutions:
  • Reduce search breadth
  • Use faster policies
  • Increase computational resources

Canvas Visualization

When an AiZynthFinder job completes, a result card appears on the canvas.

• Visualization: Interactive retrosynthetic tree diagram

Tier 2 Metrics

• Number of routes found
• Best route score (0-1, closer to 1.0 = more feasible)
• Step count
• Complexity score
• Building block availability

Interpreting the Tree

• Nodes: Molecules (target -> intermediates -> building blocks)
• Green nodes: Building blocks that are commercially available (in stock)
• Yellow nodes: Available to order from vendors
• Red nodes: Not commercially available -- may need custom synthesis
• Edges: Reaction steps connecting molecules

Read next