Retrosynthetic Planning

The Make half of design-make-test-analyse, and the check that a proposed molecule is more than a picture. AiZynthFinder runs Monte Carlo tree search over reaction templates until every leaf of a route is something you can buy.

Tool: AiZynthFinder 4.4.1, MIT. pip install aizynthfinder — Python 3.10–3.12 only, it will not install on 3.13. Then download_public_data <dir> for models and a ZINC stock (several GB). CPU is sufficient. Checked against: 4.4.1, December 2025.

Read references/aizynthfinder-setup.md before your first run, references/synthesizability-scores.md when triaging more molecules than you can search, and references/route-quality.md before acting on a route — that one is judgement, not syntax.

The two scripts

The stock file is the answer

This is the thing to get right. A target is solved when every leaf of a route is in your stock file — so "solved" is a statement about the stock at least as much as about the molecule.

A solved fraction quoted without naming the stock is meaningless — the same molecule is solved against eMolecules and unsolved against a cupboard. route_report.py says so on every run.

Configuration changed at version 4

Version 3 took bare lists of file paths; version 4 takes typed blocks. Every tutorial older than 2024 shows the incompatible form, and the resulting error is unhelpful.

python skills/retrosynthesis/scripts/aizynth_config.py config \
    --model uspto_model.onnx --templates uspto_templates.csv.gz \
    --filter-model uspto_filter_model.onnx --stock zinc:zinc_stock.hdf5 > config.yml
python skills/retrosynthesis/scripts/aizynth_config.py check --config config.yml

expansion:
  uspto:
    type: template-based
    model: uspto_model.onnx
    template: uspto_templates.csv.gz

check verifies every referenced file exists, because AiZynthFinder discovers a missing model after the run starts. It also warns when a config looks like the version-3 form.

Always set a filter policy. Without one the search proposes reactions the expansion model likes but that do not work, and the solved fraction stops measuring anything.

Budget before you start

time_limit is per target. At the 120 s default, ten molecules is twenty minutes and ten thousand is nearly two weeks. Use aizynthcli --nproc 8 to parallelise across targets.

Reading the result

python skills/retrosynthesis/scripts/route_report.py summary --output out.json.gz
python skills/retrosynthesis/scripts/route_report.py routes --output out.json.gz

# 1/1 solved (100.0%)
targets       1
solved        1
median_steps  2

target  route  steps  starting_materials  leaves_in_stock  score
TARGET  0      2      3                   3                0.95

Step count matters more than existence — yields multiply, so five steps at 70% is 17% overall, and past about six steps a route is rarely run as written. route_report.py flags those.

blocks counts how many routes share each starting material. If twenty targets converge on three intermediates, the campaign is cheap; that is a different and more useful fact than the solved fraction.

Unsolved does not mean unmakeable

It means no route was found within the time and depth limits, using these templates, terminating in this stock. Four distinct fixes, and working out which applies is the useful step: raise the time limit, raise max_transforms, broaden the stock — or accept that the chemistry is not in USPTO templates.

That last case is systematic. Template models only know reactions in their training corpus, so novel methodology, photoredox, electrochemistry, and enzymatic steps are largely invisible.

Four ways this misleads

1. A solved route is a proposal, not a validated synthesis. The templates come from reactions that worked on other substrates; nothing here knows your chemoselectivity or protecting-group needs.
2. Convergent beats linear at equal step count. Overall yield depends on the longest linear sequence, so read the tree shape, not just its depth.
3. Where the disconnections sit matters more than step count for a series. A route that decorates late gives analogues from a common intermediate; one that installs the variable group first needs a full resynthesis each time.
4. Pre-filtering with RAscore inflates the solved fraction, because RAscore is trained to predict AiZynthFinder's own verdict. Fine as a pipeline, misleading as a statistic — report the pre-filter.

Triage at scale

Route search is seconds to minutes per molecule; scores are microseconds. For a generated library: SAscore or RAscore across everything, full search on the survivors, and a chemist reading the routes for the handful you will actually order. The honest hierarchy is SAscore < RAscore < route search < a chemist's opinion < the compound in a vial, and each step right is more expensive and more real.

Composing with the rest of the bundle

• generative-design → here: the essential check, since nothing in a REINVENT objective knows what can be made. Better still, add SAscore as a scoring component during the run.
• chemical-space → alongside: if it is already purchasable, you do not need a route.
• medchem → before: no point routing molecules that fail structural alerts.
• admet-prediction → alongside: makeable and developable are different filters.

Reporting results honestly

Name the stock, always. Give solved fraction and median step count — 90% solved at nine steps is worse than 60% at three. State the time and depth limits, since unsolved is partly a statement about them. And say plainly that a proposed route is a hypothesis no chemist has yet reviewed.