Skip to content

AiZynthFinder feature parity

This page maps common AiZynthFinder options to their RENKIN equivalents. Each equivalent is scoped to what RENKIN actually computes. A matching name does not mean a matching algorithm.

AiZynthFinder RENKIN Notes
time_limit --time-limit-secs N / time_limit_seconds=N Standard search only. The deadline is cooperative, and routes found before it are kept. JSON gains time_limit_secs and termination.
exclude_target_from_stock --exclude-target-from-stock / exclude_target_from_stock=True The target is never treated as a stock terminal, so no depth-0 route is returned.
AiZynthExpander renkin expand / renkin.expand() One-step disconnections from RENKIN's rules, merged by precursor set, with stock membership for each precursor.
route_distances + RouteCollection.cluster --cluster, --n-clusters K, --max-clusters N Unit-cost tree edit distance and average-linkage clustering. See the limits below.
trees.json / ReactionTree.to_dict() --format aizynth A route array with the same structure as AiZynthFinder's output. Reactions are not atom-mapped.

Time limit

renkin --target "CC(=O)Oc1ccccc1C(=O)O" --depth 8 --max-routes 100 --time-limit-secs 10

termination is deadline_exceeded when the budget stopped the search. The deadline is checked at cooperative checkpoints, so it is not a hard real-time bound. Coverage and recovery modes keep --coverage-timeout-secs and --recovery-timeout-secs.

Excluding the target from stock

renkin --target "CC(=O)O" --depth 2 --exclude-target-from-stock

Stock identity for precursors is unchanged: exact standardized canonical SMILES.

Single-step expansion

renkin expand --target "CC(=O)Oc1ccccc1C(=O)O" --max-candidates 10 --output human
renkin expand --target "..." --templates data/templates_extracted.smi --top-templates 5000

import json, renkin
result = json.loads(renkin.expand("CC(=O)Oc1ccccc1C(=O)O", max_candidates=10))

Candidates are ordered by RENKIN's heuristic step cost, then by the number of in-stock precursors. This order is not a neural-policy probability, and it says nothing about feasibility or yield.

Route distance and clustering

renkin --target "..." --max-routes 20 --cluster            # k chosen by silhouette (<= 5)
renkin --target "..." --max-routes 20 --n-clusters 3       # fixed k

route_clusters contains distance_matrix and labels (one label per route, in route order; the first route is always in cluster 0), along with n_clusters, silhouette, and selection.

Limits:

  • Routes become molecule/reaction trees. Children are placed in a canonical order, and distance is the Zhang–Shasha ordered tree edit distance with unit costs. Two nodes of the same kind with identical labels cost 0 to match; different labels cost 1.
  • AiZynthFinder weights molecule substitutions by fingerprint Tanimoto distance. RENKIN does not, so the two tools' distances are not numerically comparable. RENKIN's value is an upper bound on the unordered unit-cost distance.
  • The distance measures structural differences between routes. It is not a chemical-similarity score.

AiZynthFinder-format export

renkin --target "CC(=O)Oc1ccccc1C(=O)O" --format aizynth > trees.json
renkin audit-route trees.json --format aizynthfinder --stock data/building_blocks.smi

Reaction smiles follows AiZynthFinder's retro direction (product>>reactants). mapped_reaction_smiles, template_hash, and policy_probability are left out on purpose. RENKIN's own template identity is recorded as metadata.renkin_template_id. Because the reactions have no mapping, the audit reports forward validation as not_evaluable for those steps.

Not covered

  • MCTS, DFPN, and Retro with neural expansion or filter policies. RENKIN searches with A over rules and templates.
  • Bond constraints (freeze_bonds, break_bonds). These require atom-map tracking through rule application.
  • Route images (PNG/SVG). --format mermaid and --format tree provide text renderings.