Bio Retrosynthesis

Version Compatibility

Reference examples tested with: AiZynthFinder 4.4+, RDKit 2024.09+, RDChiral 1.1+, and ASKCOS Lite 0.5+. Chemformer is archived legacy research code; if reproducing it, pin an exact repository commit, checkpoint, and configuration rather than assuming a current pip package or stable Python API.

Before using code patterns, verify installed versions match. If versions differ:

• Python: pip show <package> then help(module.function) to check signatures
• CLI: aizynthcli --version

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

Retrosynthesis

Plan synthetic routes from a target molecule back to commercially available building blocks. AiZynthFinder combines Monte Carlo Tree Search (MCTS), template-based expansion, configurable search rewards, and route scorers (Saigiridharan et al. 2024). Chemformer is a published template-free transformer baseline whose official repository is now archived; use its exact historical environment for reproduction or select a maintained model with a documented interface. ASKCOS is another open-source synthesis-planning platform. A useful workflow combines retrosynthesis with building-block availability and an independently configured forward-prediction check, while recognizing that a round-trip model match is not experimental validation.

For generative design pipelines that need synthetic feasibility, see chemoinformatics/generative-design. For reaction enumeration (forward direction), see chemoinformatics/reaction-enumeration.

Retrosynthesis Method Taxonomy

Decision: For most users, AiZynthFinder with its documented public USPTO expansion policy and a current stock is a practical open-source starting point. For high-stakes routes, apply expert review and an independently configured forward-prediction check; neither model agreement nor a solved search route is experimental validation.

Decision Tree by Scenario

AiZynthFinder Setup

Goal: Configure AiZynthFinder with USPTO templates + a building-block stock and run MCTS retrosynthesis planning on a target SMILES.

Approach: Create a version-appropriate YAML configuration with the expansion policy and stock, instantiate AiZynthFinder from that file, set the target SMILES, then call tree_search() followed by build_routes(). Use the schema documented for the installed release rather than copying legacy policy / finder dictionaries.

from aizynthfinder.aizynthfinder import AiZynthFinder

finder = AiZynthFinder(configfile='config.yml')
finder.expansion_policy.select('uspto')  # key defined in config.yml
finder.stock.select('zinc')              # key defined in config.yml
# finder.filter_policy.select('uspto')   # optional configured filter
finder.target_smiles = 'CC(=O)Nc1ccc(C(=O)Nc2cccc(C(F)(F)F)c2)cc1'
finder.tree_search()
finder.build_routes()

Output: finder.routes, a RouteCollection containing ranked reaction_trees, initial scores, serialized route dictionaries, and route metadata.

Route Output Analysis

for tree, score in zip(finder.routes.reaction_trees, finder.routes.scores):
    leaves = list(tree.leafs())
    n_steps = len(list(tree.reactions()))
    print(f'Steps: {n_steps}, Score: {score}, Solved: {tree.is_solved}')
    print(f'In-stock: {sum(tree.in_stock(node) for node in leaves)} / {len(leaves)}')
    print(f'Building blocks: {[node.smiles for node in leaves]}')

Critical metrics:

• Number of reactions: count tree.reactions(); do not assume graph depth and reaction count are interchangeable for branched routes
• Score: interpret according to the configured scorer; scale and direction are scorer-specific
• In-stock: how many leaf nodes are commercially available
• Solved state: tree.is_solved is true only when all leaf nodes satisfy the configured stock criterion
• Stock origin: record the selected stock name, source snapshot, and access date

Search Rewards and Route Scoring

AiZynthFinder retains the mcts search algorithm. It can combine configured search rewards through search.algorithm_config.search_rewards and corresponding weights, and it can rank completed routes with loaded scorers. There is no built-in mo_mcts algorithm or finder.mo_mcts configuration block. Available scorer names depend on the installed version, configuration, and plugins, so inspect the loaded scorers and use only documented names. If the desired objective is not available during search, export routes and rerank them explicitly after search.

Building Block Stocks

AiZynthFinder accepts HDF5 stocks built from plain-text SMILES with its documented smiles2stock command:

smiles2stock --files zinc_building_blocks.smi --output zinc.hdf5

Forward Validation with Molecular Transformer

AiZynthFinder predicts retrosynthesis (target -> precursors), while a forward model predicts products from reactants. For each proposed reaction step, serialize reactants and reagents in the format expected by the installed forward model, request its ranked product predictions, and compare standardized product structures with the planned product. The Molecular Transformer literature does not define a universal molecular_transformer.predict_forward Python function, so use the documented interface of the chosen implementation. Report the observed top-k round-trip match rate for the model, reaction representation, and dataset; no universal 30–50% pass rate is established by the AiZynthFinder 4.0 paper.

Template-Free with Chemformer

Chemformer uses a BART-style transformer trained on USPTO reactions for SMILES-to-SMILES prediction. Its official repository is archived and does not expose the Chemformer.load_pretrained(...).predict(...) convenience API sometimes shown in informal examples. To reproduce the published model, use the inference entry point, Hydra configuration, tokenizer, and checkpoint bundled with one pinned archived commit, and record that environment. For new work, prefer a maintained template-free implementation with a documented inference interface and benchmark it on the intended reaction domain.

Trade-off: A template-free model can propose disconnections outside a fixed template library, but its outputs require syntax checks, atom/reaction consistency checks, route-level review, and prospective validation. Treat it as a comparison or complementary hypothesis generator rather than assuming that merging its routes with AiZynthFinder is always superior.

Disconnection-Aware Design (DAD)

Modify generative design to also score retrosynthetic feasibility with AiZynthFinder batch mode.

Goal: Add retrosynthetic feasibility scoring to generative design pipelines for hundreds-to-thousands of candidate molecules.

Approach: Batch-process generated SMILES through aizynthcli, use the reported solved state and number of reactions for each route, and feed a documented feasibility definition back into the generative scoring function.

aizynthcli --smiles compounds.smi --output routes.json \
           --config config.yaml --policy uspto --stocks zinc

For each compound, returns top-K routes. Score-feasibility for generative design:

• "Solved" = at least one extracted route has all leaves in the configured stock
• "Short solved route" = a solved route whose reaction count is below a project-defined threshold
• "Unsolved" = no extracted route is solved under the search budget and stock; this does not prove that the target is unsynthesizable