Reaction Condition Recommendation

Reaction-condition recommendation

Answer one scientific question: for a fixed reaction, which condition labels does a trained model rank highest? Conditions are hypotheses used to focus literature/ELN retrieval. They are not an experimental procedure and must not be generated before reactants and products are specified.

Parrot is the implementation. The original repository code is MIT, but its Google Drive archives do not carry separate machine-readable terms in the official downloader and remain blocked. The approved deployment instead uses the first author's separately published Hugging Face repository, whose repository card declares MIT. Admission is limited to revision b9ef6049d341bfc62d835f09ad6ce33b6f86b047, USPTO_condition.mar (SHA256 4418693a91a7a3b5f2aa101a39d58702b154e58901ddbf1ac94edc4c28de8e7d) and condition_predictor_metadata.zip (SHA256 dfdf7fff11fe2d52af49146b1080dd6304ddd2b51665907fa759ffd4c5fca820).

Install and run

Keep Parrot in its own environment because it pins an older Transformers stack. This recipe is Linux-only: upstream states that Parrot was tested on Linux, and envs_cpu.yaml contains Linux-specific packages such as ld_impl_linux-64 and libgcc-ng. On macOS or another non-Linux platform, stop and route the task to a reviewed Linux container or remote host rather than trying to solve that lock file locally.

From an operator terminal whose current directory is the writable session workspace, clone the code under a workspace-owned model root and detach at the reviewed commit. Do not run a moving branch:

set -eu

PARROT_ROOT="$PWD/models/parrot"
PARROT_COMMIT="0fb2325567e21011589641544e32427c8244e2a9"

mkdir -p "$PARROT_ROOT"
if [ ! -d "$PARROT_ROOT/source/.git" ]; then
  git clone https://github.com/wangxr0526/Parrot.git "$PARROT_ROOT/source"
fi
git -C "$PARROT_ROOT/source" cat-file -e "${PARROT_COMMIT}^{commit}"
git -C "$PARROT_ROOT/source" checkout --detach "$PARROT_COMMIT"
test "$(git -C "$PARROT_ROOT/source" rev-parse HEAD)" = "$PARROT_COMMIT"
SOURCE_STATUS="$(git -C "$PARROT_ROOT/source" status \
  --porcelain --untracked-files=all)"
test -z "$SOURCE_STATUS"
conda env create -n parrot -f "$PARROT_ROOT/source/envs_cpu.yaml"

The final assertion must remain empty; if a reused checkout has modified or untracked files, stop instead of executing it as reviewed source.

That source revision and the approved Hugging Face snapshot have been verified in the external deployment root. The repository-native ../retrosynthesis_planning/parrot_mar_inference.py adapter consumes a safely expanded MAR through model_location; the OpenAI4S worker has completed a real GPU canary and returned 15 joint condition beams. This is an engineering inference check, not benchmark accuracy or experimental validation. Invoke the snapshot through ReactionModelBackend("parrot", ...); temporary files must use an explicit external workspace_dir.

Do not execute the official download_data.py directly. At the reviewed revision it constructs an unquoted shell=True extraction command and does not propagate extraction failure, so a workspace path containing spaces can fail silently and shell metacharacters are unsafe. Review the downloader URLs and checkpoint terms and record an explicit allow decision in the model manifest before acquiring anything; a missing or deny decision must stop. Only after that decision, use an approved operator workflow that streams each archive to private staging, verifies its recorded size and digest, and extracts it without a shell while rejecting traversal and links. Place only the verified dataset, label dictionaries, and checkpoint at the repository-relative paths named by the reviewed configuration, and add the acquisition receipts to the manifest before inference.

The Google Drive files remain unapproved. Do not substitute the MIT source-code license for those artifacts or silently replace the admitted Hugging Face revision. A missing/deny admission decision, an unexpected filename, size, or digest, or a path/link-unsafe archive must stop before extraction or inference.

Write one complete reaction SMILES per line. For the reviewed MAR deployment, pass the expanded model directory in the backend manifest. The legacy upstream CLI example below applies only to a separately reviewed legacy snapshot:

SESSION_WORKSPACE="$PWD"
PARROT_ROOT="$SESSION_WORKSPACE/models/parrot"
conda run -n parrot --cwd "$PARROT_ROOT/source" python inference.py \
  --config_path configs/config_inference_use_uspto.yaml \
  --input_path "$SESSION_WORKSPACE/reactions.txt" \
  --output_path "$SESSION_WORKSPACE/predicted_conditions.csv" \
  --num_workers 2 --inference_batch_size 8 --gpu -1

Run these blocks from the session workspace root. Shell expansion makes both input and output absolute before conda run changes to the repository working directory; do not rely on Parrot's process directory for session I/O.

The USPTO checkpoint recommends categorical condition components. Use the Reaxys configuration only when its separately obtained data/checkpoint terms have been reviewed and temperature prediction is required. Never imply that all Parrot checkpoints predict temperature.

Scenario 5 benchmark contract

Use ../retrosynthesis_planning/condition_benchmark.py with the checkpoint's frozen label dictionaries. Submit ranked complete five-slot tuples (catalyst1, two solvents, and two reagents), preserving explicit empty slots. Do not form an unscored Cartesian product from independent marginal labels. The evaluator scores multi-reference exact tuples, Top-1 slot recall, OOV tuples, duplicates, and unused Top-K budget.

Interpret the result

• Preserve the label dictionary and model configuration used to decode each categorical ID.
• Return top-k condition sets rather than combining marginal top-1 labels into a condition set the model never emitted.
• Keep catalyst, reagent, solvent, and temperature fields separate.
• Use predictions to construct targeted literature/ELN searches for the exact transformation and close substrate analogues.
• Mark missing condition classes as unknown. Do not let an LLM fill them and relabel the result as model output.

Output contract

Return canonical reaction SMILES, ordered condition sets, raw component labels and decoded names, checkpoint/config provenance, temperature support status, and validation state (model_only, literature_analog, exact_precedent, or eln_verified). Model-only is the default.

Failure modes

Primary sources: https://github.com/wangxr0526/Parrot and the first-author checkpoint distribution https://huggingface.co/xiaoruiwang/ChemEnzyRetroPlanner_metadata. The source paper is Wang et al., Research (2023), DOI 10.34133/research.0231.