Single-step retrosynthesis

Answer one scientific question: given one product, which precursor sets could produce it in one reaction? Do not recurse, check stock, invent conditions, or call the output a synthesis route. Hand accepted candidates to retrosynthesis_planning for multi-step search.

Use RetroChimera 1 as the default. Its ensemble combines edit-based and de-novo components, exposes a direct Syntheseus-compatible Python API, and publishes Pistachio, USPTO-FULL, and USPTO-50K checkpoints. The OpenAI4S adapter already runs it in an isolated process so PyTorch and model dependencies never enter the stdlib core.

Run through the checked adapter

Create a separate environment and install the model:

conda create -n retrochimera python=3.10 -y
conda run -n retrochimera python -m pip install "retrochimera==1.2.0"

The USPTO-50K checkpoint uses RetroChimera's Graphium architecture and requires "retrochimera[graphium]==1.2.0" instead. Install that extra before using the smaller checkpoint as a smoke test.

Acquire and verify a reviewed checkpoint with retrosynthesis_planning/model_deployment.py; keep weights outside git. Then:

The checked adapter and deployment notes are deliberately owned by the retrosynthesis_planning Skill. A delegated specialist that runs this recipe must therefore be allowlisted for both single-step-retrosynthesis and retrosynthesis_planning; loading a Skill never widens that allowlist. Load and read the dependency through the Skill APIs before importing it:

host.load_skill("retrosynthesis_planning")
backend_notes = host.skills.read("retrosynthesis_planning", "MODEL_BACKENDS.md")

If either call is refused, stop and ask the caller to add the dependency to the specialist profile. Do not bypass the gate with workspace file reads or a ../ resource path. Once access is confirmed, the USPTO-50K smoke-test checkpoint created by those notes lives under the same workspace root. Run the adapter:

from pathlib import Path

from retrosynthesis_planning.external_backends import SyntheseusBackend

workspace = Path.cwd().resolve()
model_dir = workspace / "models" / "retrochimera" / "uspto50k"
manifest = model_dir / "model-manifest.json"

backend = SyntheseusBackend(
    model="RetroChimera",
    model_dir=model_dir,
    manifest=manifest,
    python_command=(
        "conda",
        "run",
        "--no-capture-output",
        "-n",
        "retrochimera",
        "python",
    ),
)
result = backend.single_step("Oc1ccc(OCc2ccccc2)c(Br)c1", num_results=5)
for proposal in result["predictions"]:
    print(proposal["rank"], proposal["reactants_smiles"], proposal["score"])

Require a path-free manifest containing model version, checkpoint ID and hash, training dataset, and code/weight licenses. Leave automatic model download off. The adapter caps requests at ten candidates because low-ranked beams become increasingly hallucination-prone.

Compare candidates correctly

• Canonicalize each molecule, sort dot-separated components, and collapse exact duplicate precursor sets before comparing models.
• Preserve raw rank and raw model score. Do not calibrate a probability without a held-out set matching the deployment domain.
• Reject unparsable outputs and obvious atom/charge pathologies, but label this as structural screening rather than feasibility validation.
• Use reaction-forward-prediction for round-trip product recovery and reaction-atom-mapping only after both sides of a proposed reaction are known.
• Keep disagreements between edit-based and sequence-based models as review diversity; do not average scores from unlike models.

For a class-unknown benchmark, run the deterministic protocol after model inference. The protocol fails closed without RDKit because identity/string fallbacks would corrupt exact-match science; install the repository's optional chemistry environment first:

uv sync --extra chemistry

Then normalize the frozen public output:

uv run python skills/retrosynthesis_planning/single_step_benchmark.py normalize \
  --targets input/targets.csv \
  --predictions results/predictions.jsonl \
  --model-manifest input/model_manifest.json \
  --top-k 10 \
  --output results/intermediate_results.json

The public target CSV is intentionally strict: it accepts only target_id and product_smiles, so a reaction class, reference precursor, patent identifier, or accidental extra column fails closed. Run evaluate only in the separate evaluator process after predictions are frozen:

uv run python skills/retrosynthesis_planning/single_step_benchmark.py evaluate \
  --targets input/targets.csv \
  --predictions results/predictions.jsonl \
  --references private_evaluator/reference_precursor_sets.jsonl \
  --top-k 10 \
  --output private_evaluator/metrics.json

The evaluator compares dot-separated precursor molecules as unordered multisets, preserves invalid and duplicate beams, scores each target before aggregation, and supports multiple recorded precursor sets per product. It does not turn patent-record recovery into a feasibility label.

Optional diversity model

Use sagawa/ReactionT5v2-retrosynthesis when a second sequence model is useful. It is MIT, 0.2B parameters, and loads directly through Transformers. Record whether the checkpoint is the ORD-pretrained model or the USPTO-50K fine-tune: their benchmark meanings are very different. It is not the default proposal model.

Output contract

Return product SMILES, ordered precursor sets, model/checkpoint provenance, raw scores, parse status, duplicate group, and explicit caveats. A precursor set is a hypothesis for chemist review, not evidence of literature precedent, selectivity, available conditions, yield, safety, or experimental success.

Failure modes

Symptom	Action
model_dir is required	Install a reviewed checkpoint and pass its directory; do not enable an implicit download.
backend timeout or OOM	Lower num_results, use the smaller USPTO-50K checkpoint for a smoke test, or move the isolated worker to a GPU environment.
many invalid or repeated beams	Stop expanding the beam; report low candidate diversity and try an independent model.
high score but failed forward recovery	Keep it as a disagreement requiring chemistry review; never overwrite either raw result.

Primary model source: https://github.com/microsoft/retrochimera. Read deployment details and reviewed checkpoint metadata with host.skills.read("retrosynthesis_planning", "MODEL_BACKENDS.md") after the dependency has been allowed and loaded.