Retrosynthesis Usage Guide

Overview

Plan synthetic routes from target molecules back to commercial building blocks using maintained AiZynthFinder releases and, when useful, a maintained or exactly version-pinned template-free model. Score routes by configured scorer, reaction count, solved state, building-block availability, and independently configured forward-prediction checks.

Prerequisites

pip install aizynthfinder

Download AiZynthFinder USPTO model + building-block stocks.

Chemformer is archived legacy research code, not a supported pip install chemformer dependency. Reproduction requires the archived repository's documented environment, checkpoint, tokenizer, and configuration at a pinned commit.

Quick Start

Tell the AI agent what to do: - "Plan retrosynthesis for this target SMILES using AiZynthFinder" - "Batch retrosynthesize 100 generated molecules; report feasibility" - "Find shorter routes using documented AiZynthFinder rewards or post-search route scorers" - "Validate retrosynthesis with forward prediction via Molecular Transformer" - "Estimate synthesis cost using building-block prices"

Example Prompts

Single target retrosynthesis

"Plan retrosynthesis for SMILES 'CC(=O)Nc1ccc(C(=O)Nc2ccccc2)cc1' using AiZynthFinder. USPTO templates, ZINC stock. Output top 5 routes with depth, score, building-block list."

Batch feasibility screening

"Batch retrosynthesize 100 generated SMILES from candidates.csv. Report whether any route is solved under the configured stock and search budget, plus the shortest solved reaction count. Output to feasibility.csv."

Multi-objective route optimization

"Use AiZynthFinder MCTS and only scorer names available in my installed configuration. Compare the top 10 routes by solved status, route length, and stock availability, reranking after search where necessary."

Forward validation

"For each route from AiZynthFinder, predict the forward reaction with Molecular Transformer; report whether the predicted product matches the target SMILES."

What the Agent Will Do

1. Load AiZynthFinder and explicitly select the configured expansion-policy and stock keys used for the run.
2. Run MCTS with a recorded search configuration and budget.
3. Build routes from solved tree.
4. Compare routes by solved state, reaction count, in-stock leaves, and configured scores.
5. Optionally compare with a maintained or exactly version-pinned template-free/forward model.
6. Output route SMILES, building blocks, scorer values, reaction count, solved state, and stock provenance; report input-parse and execution failures separately from completed searches with no solved route.

Tips

• Template coverage is domain-dependent; report the policy source/training domain and inspect failures rather than assuming a universal coverage percentage.
• Always check in_stock flag for each leaf; non-stock leaves require additional synthesis.
• AiZynthFinder uses MCTS; combine documented search rewards or rerank completed routes instead of selecting a nonexistent mo_mcts algorithm.
• For forward checks, report the observed top-k round-trip match rate for the selected model and reaction representation; do not assume a universal pass rate.
• For batch retrosynthesis on 1000+ compounds, use aizynthcli CLI.

Related Skills

• chemoinformatics/molecular-io - Parse SMILES
• chemoinformatics/molecular-standardization - Standardize before retrosynthesis
• chemoinformatics/generative-design - Add feasibility to generative scoring
• chemoinformatics/reaction-enumeration - Forward direction
• chemoinformatics/admet-prediction - Filter targets before retrosynthesis