---
title: AiZynthFinder
description: Break a target SMILES into purchasable starting materials with AiZynthFinder — Monte Carlo tree search over USPTO templates, scored against your stock.
tags:
  - aizynthfinder
---

AiZynthFinder plans retrosynthesis routes, breaking a target molecule down into commercially available starting materials so you can judge synthetic accessibility. Reach for it once you have a candidate worth making — a designed hit or a virtual-screening pick — to scout routes before committing to synthesis or procurement. It is the natural next step after a generator like [GenMol](/docs/genmol) or [REINVENT](/docs/reinvent) proposes new molecules. See [what you can achieve](/docs/capabilities#chemistry) for the full set.

## What AiZynthFinder provides

AiZynthFinder is an open-source AI-powered retrosynthesis planning tool developed by **AstraZeneca's Molecular AI team**. First released in 2020, it has been continuously improved based on 3+ years of industrial application. The latest **AiZynthFinder 4.0** (published May 2024 in Journal of Cheminformatics) includes significant feature enhancements based on real-world medicinal chemistry feedback.

## Citations

> Genheden S, Thakkar A, Chadimová V, Reymond JL, Engkvist O, Bjerrum E. AiZynthFinder: a fast, robust and flexible open-source software for retrosynthetic planning. *Journal of Cheminformatics*. 2020 Nov 17;12:70. doi: [10.1186/s13321-020-00472-1](https://doi.org/10.1186/s13321-020-00472-1)

> Saigiridharan L, Hassen AK, Lai H, Torren-Peraire P, Engkvist O, Genheden S. AiZynthFinder 4.0: developments based on learnings from 3 years of industrial application. *Journal of Cheminformatics*. 2024 May 23;16:57. doi: [10.1186/s13321-024-00860-x](https://doi.org/10.1186/s13321-024-00860-x)

## Performance and industrial validation

- **80-86% success rate** for finding routes where all starting materials are in stock (tested on AstraZeneca designs and REINVENT-generated compounds)
- **71% success rate** on ChEMBL compound targets
- Used routinely by ~30 medicinal chemists at AstraZeneca
- Documented impact on multiple drug discovery programs

## Core capabilities

- **Template-based retrosynthesis**: Applies thousands of reaction templates trained on patent and literature data to break down target molecules into commercially available starting materials.
- **Multi-policy scoring**: Uses different neural network policies (expansion policies, filter policies) to score and rank synthetic routes based on template probability and route feasibility.
- **Stock-aware planning**: Checks route feasibility against chemical inventories (the bundled ZINC stock, or a custom library you upload) to ensure purchasable starting materials.
- **Route clustering and filtering**: Groups similar routes and applies filters to remove duplicates and low-quality pathways.
- **Search algorithm**: The platform runs AiZynthFinder's Monte Carlo Tree Search (MCTS); the library also supports other search strategies, but MCTS is the one configured here.

## New in AiZynthFinder 4.0 (2024)

- **Filter policies**: Apply learned filters to discard unlikely or undesirable reactions
- **Universal one-step support**: Works with any one-step retrosynthesis model
- **Scoring framework**: Flexible scoring of routes with customizable metrics
- **Human-guided synthesis planning**: Prompting system for incorporating chemist knowledge (2025)
- **Bonds-to-freeze constraints**: Specify bonds that should not be modified during retrosynthesis

## Key parameters

### Input molecule
- **SMILES**: Target molecule in SMILES format (required)
- **File input**: Support for SMILES files with one molecule per line

### Search configuration
- **Policies**: The bundled expansion policy is `uspto`, a template-based model trained on USPTO reaction data
- **Stocks**: Chemical inventories to check against, e.g. `zinc`
- **Filters**: Route filters to apply during search — `strict`, `recommender`, or `all`, in increasing order of permissiveness

### Search limits
- **Time limits**: Maximum search time to prevent runaway computation
- **Route limits**: Maximum number of routes to generate and return
- **Depth limits**: Maximum number of reaction steps in any route

## When to use it

## Primary use cases

- **Lead optimization**: Generate synthetic routes for hit-to-lead compounds to assess synthetic accessibility early in drug discovery
- **Route scouting**: Compare multiple synthetic strategies for the same target molecule
- **CRO preparation**: Generate detailed route proposals with starting material costs for contract research organizations
- **Library design**: Evaluate synthetic feasibility of virtual screening hits before ordering

## Best paired with

- **ADMET-AI**: Score synthetic intermediates for drug-like properties
- **Molecular docking**: Validate binding poses of proposed intermediates
- **Cost analysis**: Estimate total synthesis costs including reagent prices

## Guardrails and limits

## Input constraints

- **Single target only**: Designed for one molecule per job; use batch processing for multiple targets
- **Organic molecules**: Optimized for small organic compounds (<1000 Da); peptides and biologics not supported
- **Valid SMILES**: Requires well-formed, canonical SMILES strings

## Stock requirements

- **At least one stock source**: Must specify at least one chemical inventory (the bundled ZINC stock, or a custom library)
- **Stock availability affects results**: Routes are only marked "purchasable" if all starting materials exist in specified stocks
- **Custom stocks**: Can upload your own building block libraries in SMILES format

## Performance limits

- **Search time**: Typically 5-30 minutes depending on molecule complexity and search breadth
- **Route output**: Limited to top 50-200 routes to prevent overwhelming results
- **Memory usage**: Requires significant RAM for large template databases

## Computational scaling

- **Molecule complexity**: Search time increases exponentially with molecular size and complexity
- **Stock size**: Larger stock databases improve route finding but slow search
- **Filter selection**: A more permissive filter (`all`) surfaces more routes but includes lower-quality ones; `strict` runs faster and cleaner

## Best practices

## Input preparation

- **Standardize SMILES**: Use canonical SMILES and ensure proper atom valence
- **Remove counterions**: Strip salts and counterions that aren't part of the core structure
- **Check connectivity**: Ensure molecules are properly connected (no disconnected fragments)

## Stock selection

- **Match procurement reality**: Choose stocks that reflect your actual purchasing options
- **Combine sources**: Use multiple stock sources for broader coverage
- **Update regularly**: Refresh custom stocks as your inventory changes

## Search strategy

- **Start conservative**: Begin with default settings, then expand search if needed
- **Filter selection**: Start with `recommender` for a balance of speed and route quality
- **Filter judiciously**: Apply filters to reduce noise but don't over-filter

## Result interpretation

- **Route scores**: Higher scores indicate more likely synthetic routes
- **Purchasability**: Focus on routes marked as purchasable in your selected stocks
- **Route diversity**: Review multiple routes to understand synthetic options

## Troubleshooting common issues

- **No routes found**: Check stock selection and molecule validity
- **All routes non-purchasable**: Your target may require custom synthesis of intermediates
- **Search timeout**: Reduce search breadth or use simpler policy
- **Invalid SMILES**: Validate input SMILES using cheminformatics tools

## Chat Usage

### Example Prompts

- "Plan a retrosynthetic route for CC(=O)Oc1ccccc1C(=O)O" (aspirin)
- "Run AiZynthFinder on my top 3 REINVENT hits"
- "Find the most efficient synthesis route for ibuprofen with all reagents in stock"

### Workflow Chaining

- **Feeds into**: ZINC (purchasability and vendor catalog codes for building blocks), Compound Passport (synthesis section)
- **Accepts from**: PubChem (compound SMILES), REINVENT/GenMol (generated molecules), ChEMBL (hit compounds)

## FAQ

### What inputs are supported?

  Single-molecule SMILES targets; stock availability drives route scoring.

### How are results returned?

  Routes, scores, and stock usage arrive as JSON/CSV artifacts with logs.

## Read next