🧬 uncorrupt-smiles

This project implements the approach described in:

Schoenmaker, L., Béquignon, O.J.M., Jespers, W. & Gerard J.P. van Westen
UnCorrupt SMILES: a novel approach to de novo design.
Journal of Cheminformatics 15, 22 (2023). https://doi.org/10.1186/s13321-023-00696-x

uncorrupt-smiles fixes broken SMILES. Give it a trained checkpoint and a SMILES string (or a CSV of them), and it hands back a valid molecule — from the CLI or from Python, one string or a streamed million-row file.

from uncorrupt_smiles import Uncorrupt

uncorrupt = Uncorrupt()  # downloads and loads the bundled pretrained checkpoint, once
uncorrupt.fix_smiles("C1=CC=CC=C1(")
# -> ["c1ccccc1"]

Uncorrupt is the main entrypoint: no checkpoint path, vocab, or device to wire up — it downloads the pretrained checkpoint bundled with this project the first time it's needed (see Fetch example data below) and picks CUDA automatically if available. Pass checkpoint="path/to/your.pkg" to use a self-trained checkpoint instead. Everything else — standardizing datasets, generating synthetic errors, training your own corrector — is secondary tooling for anyone who wants to build a new model, built around the same three composable, argument-driven, streaming-first building blocks: no config files, and no dataset is ever loaded fully into memory (polars streams every CSV in and out).

📦 Installation

Requires Python 3.11+.

pip install uncorrupt-smiles

Optional extras:

pip install "uncorrupt-smiles[analysis]"   # pandas/matplotlib/seaborn for the analysis/ scripts
pip install "uncorrupt-smiles[test]"       # pytest, for running the test suite

To work from a clone instead (e.g. to run the test suite or modify the source):

git clone https://github.com/OlivierBeq/uncorrupt-smiles.git
cd uncorrupt-smiles
pip install -e ".[test]"

Core dependencies — torch >= 2.4, rdkit >= 2024.3.1, polars >= 1.0 — are resolved by pip to whichever wheel (CPU or CUDA) matches your machine.

Note: checkpoints produced by earlier versions of this project (including the archive on Zenodo) use a different, incompatible checkpoint format and cannot be loaded by Seq2Seq.load_checkpoint. Retrain to obtain a checkpoint this version can use.

🚀 Usage

Run uncorrupt-smiles <subcommand> -h for the full flag list of any subcommand. The CLI and the Python API expose exactly the same operations, in the same order of importance: fixing SMILES first, then the supporting tooling for anyone building their own corrector.

CLI

🩹 Fix

uncorrupt-smiles fix \
  --checkpoint data/performance/transformer_multiple_12_PAPYRUS_200_16_3.pkg \
  --input-csv generated/gan_ckpt100.csv --smiles-col SMILES \
  --output-csv generated/gan_ckpt100_fixed.csv

📥 Fetch example data

The example datasets and pretrained checkpoint under data/, generated/, and rawdata/ are not committed to this repository (to keep it small) — they're downloaded on demand from Zenodo and this repo's release assets, streamed straight to disk and checksum-verified.

uncorrupt-smiles fetch-data

Already-present files are skipped by default (pass --force to redownload). Use --only to fetch specific files, e.g. uncorrupt-smiles fetch-data --only rawdata/gbd_8.csv.

🧹 Standardize

Canonicalizes SMILES via RDKit. Anything RDKit can't parse — or an empty string — is dropped, not silently kept.

uncorrupt-smiles standardize \
  --input-csv rawdata/PAPYRUS.csv --smiles-col SMILES --separator ";" \
  --output-csv data/papyrus_std.csv --threshold 200

--smiles-col and --separator are explicit arguments, so this works on any CSV — not just PAPYRUS-shaped exports.

🧪 Generate errors

Introduces synthetic, RDKit-confirmed-invalid corruptions of valid SMILES, for training or evaluating a corrector.

uncorrupt-smiles generate-errors \
  --input-csv data/papyrus_std.csv --smiles-col STD_SMILES \
  --fragment-csv rawdata/gbd_8.csv --fragment-col FRAGMENT \
  --train-csv data/errors/train.csv --dev-csv data/errors/dev.csv \
  --invalid-type all --num-errors 1

--invalid-type selects which corruption(s) to introduce: all, multiple, or one of exists, par, permut, ring, syntax, valence, arom.

🏋️ Train

Training, evaluation, and checkpointing all live as methods on the model itself.

uncorrupt-smiles train \
  --train-csv data/errors/train.csv --dev-csv data/errors/dev.csv \
  --checkpoint-out data/performance/model.pkg \
  --hid-dim 128 --n-layers 2 --n-heads 4 --pf-dim 256 --batch-size 32 --epochs 20

The defaults above are sized for an 8GB GPU. For a full-scale run on more powerful hardware, just pass larger values — e.g. --hid-dim 512 --n-heads 8 --pf-dim 1024 --batch-size 128. --device defaults to CUDA if available, otherwise CPU. Pass --resume <checkpoint> to continue training an existing checkpoint.

Python

🩹 Fix

from uncorrupt_smiles import Uncorrupt

uncorrupt = Uncorrupt()  # or Uncorrupt(checkpoint="data/performance/model.pkg", device="cpu")

# a single SMILES, or a list
fixed = uncorrupt.fix_smiles(["C1=CC=CC=C1(", "CC(=O)Oc1ccccc1C(=O)O"])

# or stream an entire CSV in and out
uncorrupt.fix_smiles_csv(
    "generated/gan_ckpt100.csv", "SMILES", "generated/gan_ckpt100_fixed.csv",
)

Uncorrupt is a thin wrapper around Seq2Seq, the model class itself — reach for it directly if you're already holding a model (e.g. right after fit(), or loaded via Seq2Seq.load_checkpoint(path, device)), since a checkpoint's src_vocab/trg_vocab are attached to the model, not passed around separately:

from uncorrupt_smiles.transformer import Seq2Seq

model = Seq2Seq.load_checkpoint("data/performance/model.pkg", device="cpu")
model.fix_smiles(["C1=CC=CC=C1("])

fix_smiles/fix_smiles_csv work with a checkpoint trained by this model architecture, or any future architecture exposing the same generate(...) method — nothing here is Transformer-specific beyond Seq2Seq itself.

📥 Fetch example data

from uncorrupt_smiles.fetch_data import fetch_all

fetch_all()

Already-present files are skipped by default (pass force=True to redownload). Use only=[...] to fetch specific files, e.g. fetch_all(only=["rawdata/gbd_8.csv"]).

🧹 Standardize

from uncorrupt_smiles.preprocess import standardize_smiles, standardize_stream

standardize_smiles("C(C)O")  # -> "CCO"

standardize_stream(
    "rawdata/PAPYRUS.csv", smiles_col="SMILES", output_csv="data/papyrus_std.csv",
    length_threshold=200, separator=";",
)

🧪 Generate errors

from uncorrupt_smiles.invalid_smiles import generate_errors, reservoir_sample_fragments, write_errors_split

# stream straight to train/dev CSVs
write_errors_split(
    "data/papyrus_std.csv", "STD_SMILES",
    "rawdata/gbd_8.csv", "FRAGMENT",
    "data/errors/train.csv", "data/errors/dev.csv",
    seed=42, invalid_type="all",
)

# or generate pairs in memory, e.g. for a handful of molecules
fragments = reservoir_sample_fragments("rawdata/gbd_8.csv", "FRAGMENT", k=20_000, seed=42)
pairs = list(generate_errors(["CCO", "c1ccccc1"], fragments, seed=42))

🏋️ Train

from uncorrupt_smiles.data import iter_csv_column, make_loader
from uncorrupt_smiles.transformer import Seq2Seq
from uncorrupt_smiles.vocab import Vocab

src_vocab = Vocab.build_from_lines(iter_csv_column("data/errors/train.csv", "ERROR"))
trg_vocab = Vocab.build_from_lines(iter_csv_column("data/errors/train.csv", "STD_SMILES"))

model = Seq2Seq.build(
    len(src_vocab), len(trg_vocab), max_length=202, device="cuda",
    src_pad_idx=src_vocab.pad_idx, trg_pad_idx=trg_vocab.pad_idx,
    hid_dim=128, n_layers=2, n_heads=4, pf_dim=256,
)

train_loader = make_loader("data/errors/train.csv", "ERROR", "STD_SMILES", src_vocab, trg_vocab, batch_size=32)
dev_loader = make_loader("data/errors/dev.csv", "ERROR", "STD_SMILES", src_vocab, trg_vocab, batch_size=32)

model.fit(train_loader, dev_loader, src_vocab, trg_vocab, epochs=20, checkpoint_path="data/performance/model.pkg")

The defaults above are sized for an 8GB GPU; pass larger hid_dim/n_heads/pf_dim/batch size keyword arguments for a full-scale run on more powerful hardware. Use Seq2Seq.load_checkpoint(...) to continue training an existing checkpoint.

🗂️ Package layout

Module	Purpose
src/uncorrupt_smiles/uncorrupt.py	Uncorrupt — the main entrypoint. Wraps Seq2Seq.load_checkpoint, defaulting to the bundled pretrained checkpoint (downloaded on demand).
src/uncorrupt_smiles/transformer.py	The Seq2Seq model. Fixing SMILES (fix_smiles/fix_smiles_csv), training, evaluation, and checkpoint save/load are all methods on the model itself; src_vocab/trg_vocab are attached to the model, not passed around separately.
src/uncorrupt_smiles/fetch_data.py	Downloads example datasets / the pretrained checkpoint on demand.
src/uncorrupt_smiles/invalid_smiles.py	Per-SMILES error functions (exists_error, par_error, permutation, ring_error, syntax_error, valence_error, arom_error, introduce_error) plus streaming orchestration (generate_errors, write_errors_split, reservoir_sample_fragments).
src/uncorrupt_smiles/preprocess.py	standardize_smiles / standardize_stream.
src/uncorrupt_smiles/vocab.py	Vocab — character-level vocabulary, built by streaming.
src/uncorrupt_smiles/data.py	Streaming CSV helpers and the training IterableDataset.
src/uncorrupt_smiles/cli.py	The uncorrupt-smiles entrypoint.

✅ Testing

pip install -e ".[test]"
pytest

The suite uses small synthetic fixtures and tiny model dimensions — it runs in a few seconds and comfortably fits an 8GB-VRAM/modest-CPU machine.

🙏 Acknowledgments

uncorrupt-smiles builds on the original SMILES-corrector research and codebase by Linde Schoenmaker.

📄 License

MIT — see LICENSE.