synrxnval

synrxnval is a Python package to validate synthetic chemical reactions (that can be generated with the synrxnforge package) by exploiting chemical knowledge of 2 seq2seq Transformer models.

For a given synthetic reaction (A>>C), the reagents B are first predicted. Then, the combination of tagged reactants and reagents (A!>B) are submitted to a forward reaction prediction model*, that predicts the outcome C' of the given mixture.

If C'== C and the confidence score of the prediction is >0.95, the synthetic reaction is considered valid and is kept.

---

Installation

You can install synrxnval directly from source:

conda create -n synrxnval python=3.10 -y
conda activate synrxnval
conda install -c conda-forge rdkit -y

git clone https://github.com/yvsgrndjn/synrxnval.git
cd synrxnval
pip install -e .

---

Setup

The input file can have several formats, the easiest is the output parquet file from synrxnforge.

you need first to download the reagents prediction (T2) model and the forward tag prediction model (T3*). Models trained on USPTO data are available [https://zenodo.org/records/14017743?token=eyJhbGciOiJIUzUxMiJ9.eyJpZCI6ImY2NmZmNTkxLTE2YzYtNDU0OS04NjAzLTJiNzg1YzFhMGQ5NSIsImRhdGEiOnt9LCJyYW5kb20iOiJmMjBhNjlkNDRkYTE1YWMwNDVjODQ2YjkwOTQ1ZjgyNCJ9.DRQXQBjRcv6MTW1hEYDSYZ6j11dmKBAQI-nytHBTHKu66KYTS3TgriJW_pOTfayHcditLS4MNKa9okI4FLSD2Q|here].

Run validation

$ synrxnval python -m --input path/to/parquet
                      --t2_model path/to/model
                      --t3_model path/to/model
                      --out_dir path/to/out