ReactionClassifier

Hierarchical reaction classification. Given a reaction SMILES, it predicts a class in an LLM-derived reaction taxonomy and confirms it symbolically: a Morgan difference–product (MDP) fingerprint MLP gate proposes a class, the exact retrosynthetic templates in that class's tier-3 subtree are applied to the reaction, and a label is returned only if a template reproduces the recorded product

Install

pip install reactionclassifier          # requires rdkit, torch, numpy

Quickstart

from reactionclassifier import ReactionClassifier

clf = ReactionClassifier()              # loads the bundled gate + templates + taxonomy
r = clf.classify("CC(=O)O.NCc1ccccc1>>CC(=O)NCc1ccccc1")
r.reaction_code     # '2.1.2.1'   (deterministically confirmed)
r.reaction_name     # 'Amidation using Carboxylic Acids | Primary Amine + Carboxylic Acid to Secondary Amide'
r.tier_path         # ['2.1', '2.1.2', '2.1.2.1']
r.confidence        # Top-1 probability of the neural layer

classify() returns a ClassificationResult:

So: if reaction_code is populated you have a verified label; otherwise neural_code/neural_name give the model's best (unverified) guess.

Taxonomy and granularity examples

from reactionclassifier import load_taxonomy, name_for, full_class_name, tier_path, load_granularity
load_taxonomy()["1.3.6"]      # -> single class name
name_for("2.1.2.1")           # -> single-level name
full_class_name("2.1.2.1")    # -> pipe-joined L3|L4|L5 names, e.g.
                              #    'Amidation using Carboxylic Acids | Primary Amine + Carboxylic Acid to Secondary Amide'
tier_path("1.3.6.2")          # -> ['1.3', '1.3.6', '1.3.6.2']
load_granularity()            # the two granularity-comparison tables

What's included

Full reaction database

The full labelled reaction database (≈666k reactions) is hosted externally (Zenodo) rather than shipped in the wheel:

from reactionclassifier.database import download_database
path = download_database()     # downloads + caches the parquet, returns its path

The released database excludes NameRXN-derived columns (NAME, CLASS); NameRXN is proprietary and its labels are not redistributed.

Detail

• The MDP fingerprint is a Morgan difference (reactant⊕product bit-unions) concatenated with the product fingerprint (RDKit, radius 2, 2048 bits each; 4096-dim).
• The full generalised-SMIRKS library is not released; only the small subset embedded in the granularity examples is included.

License

MIT (code and bundled data). See LICENSE.