Reaction Forward Prediction

Forward reaction prediction

Answer one scientific question: given reactants and a separately declared reagent/condition string, which product structures does the model rank highest? For retrosynthesis review, test whether the intended product appears in the forward model's top-k outputs. Call this round-trip recovery, not proof that the reaction works.

Use sagawa/ReactionT5v2-forward by default. It is a 2025 peer-reviewed, MIT-licensed 0.2B model distributed as safetensors and runs through ordinary Transformers.

Install and run

Install in a separate environment; do not add these packages to OpenAI4S core:

conda create -n reactiont5 python=3.11 -y
conda run -n reactiont5 python -m pip install \
  "torch" "transformers==4.40.2" "tokenizers==0.19.1" \
  "huggingface_hub[cli]==0.35.0" \
  sentencepiece rdkit datasets accelerate pandas

Acquire an immutable local model snapshot and a reviewed source checkout from an operator terminal whose current directory is the writable session workspace. The revisions below are the reviewed revisions for this recipe; do not replace either with main. A future revision requires a new review and provenance record before use.

set -eu

REACTIONT5_ROOT="$PWD/models/reactiont5"
SOURCE_COMMIT="76eb08068e10fe255cae5d563a91e1c1e9abac54"
FORWARD_REVISION="933114058cb2604dc1bf536dbebdfcefbe83d4fc"

mkdir -p "$REACTIONT5_ROOT"
if [ ! -d "$REACTIONT5_ROOT/source/.git" ]; then
  git clone https://github.com/sagawatatsuya/ReactionT5v2.git \
    "$REACTIONT5_ROOT/source"
fi
git -C "$REACTIONT5_ROOT/source" cat-file -e "${SOURCE_COMMIT}^{commit}"
git -C "$REACTIONT5_ROOT/source" checkout --detach "$SOURCE_COMMIT"
test "$(git -C "$REACTIONT5_ROOT/source" rev-parse HEAD)" = "$SOURCE_COMMIT"
SOURCE_STATUS="$(git -C "$REACTIONT5_ROOT/source" status \
  --porcelain --untracked-files=all)"
test -z "$SOURCE_STATUS"

conda run -n reactiont5 hf download sagawa/ReactionT5v2-forward \
  --revision "$FORWARD_REVISION" \
  --local-dir "$REACTIONT5_ROOT/forward-$FORWARD_REVISION"

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

Record the two revisions and hashes of the downloaded regular files. Keep the snapshot outside version control. The batch CLI imports repository-local modules, so run prediction.py with task_forward as its working directory and pass only the reviewed local snapshot:

REACTIONT5_ROOT="$PWD/models/reactiont5"
FORWARD_REVISION="933114058cb2604dc1bf536dbebdfcefbe83d4fc"

HF_HUB_OFFLINE=1 conda run -n reactiont5 \
  --cwd "$REACTIONT5_ROOT/source/task_forward" \
  python prediction.py \
  --input_data "$PWD/reactions.csv" \
  --model_name_or_path "$REACTIONT5_ROOT/forward-$FORWARD_REVISION" \
  --input_max_length 150 --num_beams 5 --num_return_sequences 5 \
  --batch_size 16 --output_dir "$PWD/forward-output"

Run that block from the session workspace root so $PWD expands to absolute workspace input/output paths. For a single record, select the environment in its own OpenAI4S Python Cell:

host.env.use("reactiont5")

After the switch succeeds, load only the reviewed local snapshot in a new Cell:

import os
from pathlib import Path

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

reviewed_revision = "933114058cb2604dc1bf536dbebdfcefbe83d4fc"
snapshot = Path.cwd() / "models" / "reactiont5" / f"forward-{reviewed_revision}"
if not snapshot.is_dir():
    raise FileNotFoundError(f"reviewed snapshot is missing: {snapshot}")
os.environ["HF_HUB_OFFLINE"] = "1"
tokenizer = AutoTokenizer.from_pretrained(snapshot, local_files_only=True)
model = AutoModelForSeq2SeqLM.from_pretrained(snapshot, local_files_only=True)
model.eval()
text = "REACTANT:CCBr.OCCREAGENT:"
inputs = tokenizer(text, return_tensors="pt")
generated = model.generate(
    **inputs,
    num_beams=5,
    num_return_sequences=5,
    return_dict_in_generate=True,
    output_scores=True,
)
products = [
    tokenizer.decode(row, skip_special_tokens=True).replace(" ", "").rstrip(".")
    for row in generated.sequences
]

Record the model ID, reviewed revision, local file hashes, source commit, package versions, device, beam settings, and input string. Never fall back from a missing local snapshot to a moving Hub model ID.

For a reproducible OpenAI4S deployment, use the pinned reactiont5v2 plan in ../retrosynthesis_planning/reaction_model_deployment.py, download sagawa/ReactionT5v2-forward at revision 933114058cb2604dc1bf536dbebdfcefbe83d4fc, snapshot every downloaded file, and pass the local snapshot to ReactionModelBackend("reactiont5_forward", ...). The worker forces local_files_only=True; implicit Hugging Face downloads are not allowed during inference. top_k is limited to 1--10 and max_new_tokens to 1--256; record both values with each run.

The pinned snapshot has passed a real CPU model-card canary in the external model root: the declared reactant/reagent example returned CN1CCC=C(CO)C1, exactly matching the published expected product. This proves that the pinned files load and the input protocol is reproduced; it is not a chemistry-wide accuracy claim.

Scenario 4 benchmark contract

Use ../retrosynthesis_planning/forward_benchmark.py with the frozen separated reactant/reagent inputs. Preserve every submitted beam, including empty, invalid, and duplicate products. The private evaluator compares against all recorded products and reports both isomeric and connectivity Top-K accuracy so stereochemistry-only failures remain visible. A connectivity hit is not silently promoted to an exact stereochemical hit.

Pin the Hugging Face revision for reproducible work and record resolved commit, model ID, package versions, device, beam settings, and input string.

Round-trip check

1. Keep precursors and reagents in different fields; missing reagents are an explicit unknown, not an empty condition claim.
2. Generate no more top-k products than the review can inspect.
3. Parse and canonicalize each predicted product with RDKit.
4. Compare canonical intended product against the top-k set and record its rank.
5. Preserve nonmatching top products as possible model disagreements or byproduct hypotheses.

Do not multiply a backward-model score by a forward-model score unless both were calibrated together on a deployment-matched held-out set. If the backward and forward checkpoints share training data, round-trip agreement is correlated evidence rather than an independent experiment.

Output contract

Return reactants, reagents, ranked canonical products, invalid outputs, intended product rank or null, top-k recovery, raw sequence scores when available, and model provenance. Do not emit a boolean feasible field.

Failure modes

Primary sources: https://github.com/sagawatatsuya/ReactionT5v2 and https://huggingface.co/sagawa/ReactionT5v2-forward.