SynOmega

A retrosynthesis toolkit that turns a target molecule into synthesis routes and a continuous synthesizability score. Three decoupled layers:

synthesizability   is this target reachable from purchasable material, in N steps?
     ↑
search             Retro* / MCTS / best-first over an AND-OR graph
     ↑
single-step        product SMILES -> ranked reactant candidates

The layers meet at a deliberately narrow interface — a single-step backend only implements predict(smiles, top_k) -> [Prediction] — so the planner and the scorer do not care whether predictions come from a graph neural network, a transformer, or plain template matching.

Installation

pip install synomega           # core: rdkit + numpy
pip install "synomega[gnn]"    # + the D-MPNN neural single-step backend (torch)

The neural backend is an optional extra on purpose: the template-rule backend runs anywhere, with no GPU and no PyTorch. Requires Python ≥ 3.10.

Zero-config quickstart

pip install synomega ships only code. The first time you ask for the default model or stock, synomega downloads them (a few hundred MB) into ~/.cache/synomega — the same way spaCy and HuggingFace fetch models. So this works out of the box:

import synomega

planner = synomega.load_default_planner()          # downloads model + stock once
print(planner.plan("CC(=O)Nc1ccccc1O").best_route.describe())

from synomega import SynthesizabilityScorer
score = SynthesizabilityScorer(planner).score("CC(=O)Nc1ccccc1O", max_steps=5)
print(score.bb_coverage, score.min_steps)

Or pre-fetch from the command line, then use the CLI with no --model/--stock:

synomega download                                   # cache the default assets
synomega plan --target "CC(=O)Nc1ccccc1O" --max-steps 5

Download mirrors. Assets are hosted on both a USTC GitLab registry (fast in China) and (soon) GitHub. SynOmega auto-selects the faster reachable one by latency; override with SYNOMEGA_MIRROR=ustc|github or point SYNOMEGA_ASSETS_BASE=<url> at your own mirror. Change the cache location with SYNOMEGA_CACHE.

Quick start

from synomega import Planner, SynthesizabilityScorer
from synomega.singlestep import TemplateGNN
from synomega.stock import InMemoryStock

model   = TemplateGNN.from_pretrained("path/to/model_run")   # a trained checkpoint
stock   = InMemoryStock.from_keys_file("building_blocks.keys.gz")
planner = Planner(model, stock, algorithm="retrostar")

result = planner.plan("CC(=O)Nc1ccccc1", max_depth=5, time_limit=60)
print(result.solved)
print(result.best_route.describe())

target: CC(=O)Nc1ccccc1
solved: True  steps: 2  depth: 2  bb_coverage: 1.00
  [1] CC(=O)O.Nc1ccccc1>>CC(=O)Nc1ccccc1  (score=0.4348)
  [2] O=[N+]([O-])c1ccccc1>>Nc1ccccc1     (score=0.2174)

Synthesizability scoring

scorer = SynthesizabilityScorer(planner)

r = scorer.score("CC(=O)Nc1ccccc1", max_steps=5)
r.solved            # True — a complete route to purchasable material exists
r.bb_coverage       # 1.0 — fraction of the best route's leaves that are buyable
r.min_steps         # 2  — reactions in the shortest solved route
r.min_route_depth   # 2  — longest linear sequence of that route

report = scorer.score_batch(targets, max_steps=5)
report.solve_rate         # fraction of targets solved
report.mean_bb_coverage
report.to_dataframe()

Excluding the target from stock. A molecule that is itself a catalogue item is otherwise "solved" in zero steps. Pass exclude_target=True to force a real disconnection — the target is treated as not purchasable, while its intermediates still are. Available on both planning and scoring (default off):

planner.plan("CC(=O)Nc1ccccc1O", exclude_target=True)
scorer.score("CC(=O)Nc1ccccc1O", max_steps=5, exclude_target=True)

Reaction-plausibility filtering

Every single-step prediction is screened by a mapping-free dual-tower reaction- plausibility model: two shared-encoder D-MPNN towers embed the candidate reactants and the target product separately (no atom mapping needed), and score how likely those reactants actually give the product. Candidates below a threshold are dropped — the filter only removes wrong disconnections, it never re-ranks the survivors. Because search and synthesizability both expand through the single-step model, this screens every single-step prediction in the system.

It is off by default: benchmarks (scripts/BENCHMARKS.md) show it does not improve top-k retrieval of the recorded reaction (it is marginally negative, −0.2…−0.9 pp) and adds latency (×1.7 GPU / ×4.6 CPU). Enable it when you want the candidate list pruned of implausible disconnections rather than maximal recall.

# off by default:
planner = synomega.load_default_planner()

# enable / tune:
planner = synomega.load_default_planner(plausibility=True)
planner = synomega.load_default_planner(plausibility=True, plausibility_threshold=0.5)

# bring your own single-step model + explicit scorer:
from synomega.plausibility import PlausibilityScorer
scorer = PlausibilityScorer.default(device="cpu")
planner = Planner(model, stock, plausibility=scorer, plausibility_threshold=0.4)

When enabled, each surviving prediction carries its raw plausibility in prediction.meta["plausibility"].

Simplification-constrained model

An alternative single-step model is restricted, at the reaction-template level, to simplifying (fragmentation) disconnections — those that split the target into two or more precursors. In multi-step search it reaches purchasable material with fewer node expansions at matched solvability (about 30% fewer expansions and ~2x faster median search on a drug-like benchmark; see benchmark/ and the accompanying paper). It downloads on first use like the default model:

planner = synomega.load_default_planner(simplify=True)   # fragmentation-only model

from synomega.singlestep import TemplateGNN
model = TemplateGNN.simplify(device="cpu")               # or load it directly

Override the download with SYNOMEGA_SIMPLIFY_MODEL=/path/to/run_dir.

Two synthesizability metrics

These are conflated in the literature; SynOmega keeps them apart because they answer different questions.

bb_coverage matters because most targets are unsolved at realistic step limits. A 5-step route with 4 of 5 leaves buyable scores 0.8, not 0 — so a near-miss is distinguishable from a total failure, and a set of molecules can be ranked rather than merely split into solved/unsolved. The headline score = 1/(U+1)**U, where U is the number of non-purchasable starting materials in the best route, falls off sharply with each missing building block (U=0 → 1.0, 1 → 0.5, 2 → 0.11, 3 → 0.016), so it cleanly separates a solved target, one missing a few materials, and one missing many.

Search algorithms

All three share the AND-OR graph, the budget, and the route extractor, so their results are directly comparable.

Command line

# one-time: precompute building-block InChIKeys so later loads take seconds
synomega build-stock --catalogue catalogue.smi.gz --out building_blocks.keys.gz

synomega plan  --target "CC(=O)Nc1ccccc1" --model path/to/model_run \
               --stock building_blocks.keys.gz --stock-is-keys --max-steps 5

synomega score --targets targets.smi --model path/to/model_run \
               --stock building_blocks.keys.gz --stock-is-keys \
               --max-steps 5 --out report.json

Bring your own single-step model

Any object implementing the SingleStepModel interface plugs into the planner:

from synomega.singlestep import SingleStepModel, Prediction

class MyModel(SingleStepModel):
    name = "my-model"
    def predict(self, smiles: str, top_k: int = 50) -> list[Prediction]:
        # return candidate disconnections, best first
        return [Prediction(reactants=("CCO", "CC(=O)O"), score=0.9)]

planner = Planner(MyModel(), stock, algorithm="retrostar")

Built-in backends: TemplateGNN (D-MPNN template classifier, needs [gnn]) and TemplateRuleModel (pure template matching, no PyTorch).

Design notes

• AND-OR graph, not a tree. A molecule is solved if it is in stock or any of its reactions is solved; a reaction is solved if all its reactants are. Molecules are interned by InChIKey, so an intermediate reached down two branches is one node, expanded once. Cycles are rejected at edge creation.
• Batched expansion. The search pulls a batch of frontier molecules and issues one predict_batch, so a GPU-backed model is not left idle.
• Caching. Planner(cache=True) (default) memoizes expansions; cache_path= persists them to SQLite across runs.
• Stock membership is by InChIKey, matching a vendor catalogue written by a different toolkit. There is deliberately no Bloom-filter backend — false positives would inflate solve-rate and break comparability with published numbers.

Development

git clone https://github.com/zbc0315/synomega
cd synomega
pip install -e ".[gnn,dev]"
pytest

License

MIT — see LICENSE.