Python API

Load a released model

import torch

from prexsyn.shortcuts import AllInOneLoader, MoleculeProjector

config_path = "./data/trained_models/enamine2310_rxn115_202511.yml"
device = "cuda" if torch.cuda.is_available() else "cpu"

loader = AllInOneLoader(config_path)
projector = MoleculeProjector(
    model=loader.model().to(device).eval(),
    detokenizer=loader.detokenizer(),
    descriptor="ecfp4",
    num_samples=16,
)

loader.model(), loader.chemical_space(), and loader.detokenizer() are cached after their first call. Missing released assets are downloaded from the URLs in the YAML.

Project one molecule

one() accepts a SMILES string, an RDKit Mol, or a prexsyn_engine.chemistry.Molecule:

result = projector.one("COc1ccc(-c2ccnc(Nc3ccccc3)n2)cc1")

for i, item in enumerate(result.items[:3]):
    print(item.molecule.smiles(), item.similarity)
    print(item.get_tree())

    # Graphviz is required for image rendering.
    img = item.get_image()
    img.save(f"output_{i}.png")
    img.close()

Products are sorted by descending Tanimoto similarity. result.best() returns the highest-ranked item or None; result.best_similarity() returns 0.0 when no product was generated. Timing is available as result.time.model, result.time.detok, and result.time.total.

Project a batch

batch = projector.many(["CCO", "c1ccccc1"])

for target_result in batch.results:
    best = target_result.best()
    if best is not None:
        print(best.molecule.smiles(), best.similarity)

many() returns one result per input molecule. Sampling is internally chunked according to batch_size_limit, which defaults to 64.

Generate from fingerprints

desc() accepts a two-dimensional NumPy array or PyTorch tensor with shape (batch_size, descriptor_size). Its descriptor must match the name passed to MoleculeProjector. For a CUDA model, pass a tensor on the model device.

import torch

from prexsyn_engine.chemistry import Molecule

fingerprint_array = projector.descriptor_function(Molecule.from_smiles("CCO"))[None, :]
fingerprint = torch.from_numpy(fingerprint_array).to(projector.model.device)
result = projector.desc(fingerprint).results[0]

The released model supports ecfp4 and fcfp4.