synomega

Retrosynthesis with the synomega Python package (pip install synomega) — single-step reactant prediction, multi-step route planning, and a continuous synthesizability score (bb-coverage) for a target molecule given as SMILES. Use this whenever the user asks whether or how a molecule can be synthesized, wants candidate disconnections/reactants, wants a full synthesis route down to purchasable building blocks, or wants to rank/score molecules by makeability.

synomega is a Python package (PyPI) that turns a target molecule (SMILES) into synthesis routes and a continuous synthesizability score. Three layers behind one interface: single-step prediction → AND-OR graph search → synthesizability scoring.

It runs entirely locally. It works out of the box — the default pretrained model and building-block stock are downloaded automatically on first use, so there is nothing to train or configure.

Install

bash

pip install "synomega[gnn]"    # neural D-MPNN backend (torch) — recommended

Requires Python ≥ 3.10.

First run downloads the model + stock (automatic)

The wheel ships only code. The first prediction downloads the default model and stock (a few hundred MB) into ~/.cache/synomega. You can pre-fetch them:

bash

synomega download

Downloads come from the nearest mirror, auto-selected by latency (a USTC GitLab registry, fast in China; and GitHub). Force one with SYNOMEGA_MIRROR=ustc or SYNOMEGA_MIRROR=github; change the cache dir with SYNOMEGA_CACHE.

Fastest path: the bundled helper

scripts/synomega_run.py prints JSON for any of the three operations. No configuration — it downloads the default model/stock on first call:

bash

python scripts/synomega_run.py score        "CC(=O)Nc1ccccc1O" --max-steps 5
python scripts/synomega_run.py score        "CC(=O)Nc1ccccc1O" --exclude-target
python scripts/synomega_run.py plan         "CC(=O)Nc1ccccc1O" --max-depth 5
python scripts/synomega_run.py single-step  "CC(=O)Nc1ccccc1O" --top-k 10

plan and score accept --exclude-target (Python: exclude_target=True), which treats the target as not purchasable even if it is itself in the stock — so a catalogue molecule is not reported as trivially solved in zero steps. Use it when the user asks "how would you actually make X" about a possibly-buyable molecule. Default off.

Python API

python

import synomega

# one ready-to-use planner backed by the default model + stock (downloads once)
planner = synomega.load_default_planner()          # device="cpu" by default

# 1) Single-step — "what reacts to give X?"
for p in planner.model.predict("CC(=O)Nc1ccccc1O", top_k=10):
    print(p.score, p.reactants)

# 2) Multi-step — "how do I make X?"
result = planner.plan("CC(=O)Nc1ccccc1O", max_depth=5)
print(result.solved)
print(result.best_route.describe())

# 3) Synthesizability score — "can X be made / how hard?"
from synomega import SynthesizabilityScorer
r = SynthesizabilityScorer(planner).score("CC(=O)Nc1ccccc1O", max_steps=5)
print(r.bb_coverage, r.min_steps, r.solved)

To use your own checkpoint/stock instead of the defaults, set SYNOMEGA_MODEL and SYNOMEGA_STOCK (the helper reads them), or build the objects directly with TemplateGNN.from_pretrained(...) / InMemoryStock.from_keys_file(...).

How to read the output

• single-step → ranked Predictions; each has .reactants (a tuple of SMILES) and .score (higher = more likely). Present the top few disconnections.
• plan → a route tree. Read it top-down: the target decomposes into the reactants that make it, recursively, until every leaf is a purchasable building block. result.best_route.describe() prints it as numbered steps.
• score → the headline is bb_coverage (0–1): the fraction of the best route's leaves that are purchasable. It is continuous, so 0.8 (a near-miss) is meaningfully better than 0.0 — use it to rank candidates, not just to split solved/unsolved. Also report solved (a fully-purchasable route exists) and min_steps (reactions in the shortest solved route).

Rules for the agent

• Always pass a valid SMILES. If the user gives a name, resolve it first (or ask).
• algorithm ∈ {retrostar (default), mcts, bfs}. Larger max_depth / max_steps finds more but is slower; start at 5.
• The first call downloads a few hundred MB — expect a one-time delay. Loading the model also takes a few seconds; in Python, build planner once and reuse it.

Links

• Package: https://pypi.org/project/synomega/
• Source: https://github.com/zbc0315/synomega
• This skill: https://github.com/zbc0315/synomega-skill