ORDerly

🧪 Cleaning chemical reaction data 🧪

🎯 Condition Prediction Benchmark 🎯

Quick Install

Requires Python 3.10 (Tested on MacOS and Linux)

pip install orderly

🤔 What is this?

Machine learning has the potential to provide tremendous value to chemistry. However, large amounts of clean high-quality data are needed to train models

ORDerly cleans chemical reaction data from the growing Open Reaction Database (ORD).

Use ORDerly to:

• Extract and clean your own datasets.
• Access the ORDerly condition prediction benchmark dataset for reaction condition prediction.
• Reproduce results from our paper including training a ML model to predict reaction conditions.

📖 Extract and clean a dataset

Download data from ORD

Data in ORD format should be placed in a folder called /data/ord/. You can either use your own data, or the open-source ORD data.

To download the ORD data follow the instructions in the ORD repository (i.e. download Git LFS and clone their repository) and then place it within a folder called /data/ord/.

Extract data from the ORD files

python -m orderly.extract

If you want to run ORDerly on your own data, and want to specify the input and output path:

python -m orderly.extract --input_path="/data/ord/" --output_path="/data/orderly/"

This will generate a parquet file for each ORD file.

Clean the data

This will produce train and test parquet files, along with a .json file showing the arguments used and a .log file showing the operations run.

python -m orderly.clean

🚀 Download the condition prediction benchmark dataset

Reaction condition prediction is the problem of predicting the things "above the arrow" in chemical reactions.

There are three options for donwloading the benchmark.

1. If you have orderly installed you can download the benchmark using this command:

python -m orderly.download.benchmark

2. Or you can either download the ORDerly condition prediction benchmark dataset directly

3. Or use the following code to download it (without installing ORDerly). Make sure to install needed dependencies first (shown below).

import pathlib
import zipfile

import pandas as pd
import requests


def download_benchmark(
    benchmark_zip_file="orderly_benchmark.zip",
    benchmark_directory="orderly_benchmark/",
    version=2,
):
    figshare_url = (
        f"https://figshare.com/ndownloader/articles/23298467/versions/{version}"
    )
    print(f"Downloading benchmark from {figshare_url} to {benchmark_zip_file}")
    r = requests.get(figshare_url, allow_redirects=True)
    with open(benchmark_zip_file, "wb") as f:
        f.write(r.content)

    print("Unzipping benchmark")
    benchmark_directory = pathlib.Path(benchmark_directory)
    benchmark_directory.mkdir(parents=True, exist_ok=True)
    with zipfile.ZipFile(benchmark_zip_file, "r") as zip_ref:
        zip_ref.extractall(benchmark_directory)


download_benchmark()
train_df = pd.read_parquet("orderly_benchmark/orderly_benchmark_train.parquet")
test_df = pd.read_parquet("orderly_benchmark/orderly_benchmark_test.parquet")

📋 Reproducing results from paper

To reproduce the results from the paper, please clone the repository, and use poetry to install the requirements (see above). Towards the bottom of the makefile, you will find a comprehensive 8 step list of steps to generate all the datasets and reproduce all results presented in the paper.

Results

We run the condition prediction model on four different datasets, and find that trusting the labelling of the ORD data leads to overly confident test accuracy. We conclude that applying chemical logic to the reaction string is necessary to get a high-quality dataset, and that the best strategy for dealing with rare molecules is to delete reactions where they appear.

Top-3 exact match combination accuracy (%): frequency informed guess // model prediction // AIB%:

Where AIB% is the Average Improvement of the model over the Baseline (i.e. a frequency informed guess), where $A_m$ is the accuracy of the model, and $A_B$ is the accuracy of the baseline: $AIB = (A_m - A_b) / (1 - A_b)$

Full API documentation

Extraction

There are two different ways to extract data from ORD files, trusting the labelling, or using the reaction string (as specified in the trust_labelling boolean). Below you see all the arguments that can be passed to the extraction script, change as appropriate:

python -m orderly.extract --name_contains_substring="uspto" --trust_labelling=False --output_path="data/orderly/uspto_no_trust" --consider_molecule_names=False

Cleaning

There are also a number of customisable steps for the cleaning:

python -m orderly.clean --output_path="data/orderly/datasets_$(dataset_version)/orderly_no_trust_no_map.parquet" --ord_extraction_path="data/orderly/uspto_no_trust/extracted_ords" --molecules_to_remove_path="data/orderly/uspto_no_trust/all_molecule_names.csv" --min_frequency_of_occurrence=100 --map_rare_molecules_to_other=False --set_unresolved_names_to_none_if_mapped_rxn_str_exists_else_del_rxn=True --remove_rxn_with_unresolved_names=False --set_unresolved_names_to_none=False --num_product=1 --num_reactant=2 --num_solv=2 --num_agent=3 --num_cat=0 --num_reag=0 --consistent_yield=True --scramble=True --train_test_split_fraction=0.9

A list of solvents (names and SMILES) commonly used in pharmaceutical chemistry can be found at orderly/data/solvents.csv

Issues?

Submit an issue or send an email to dsw46@cam.ac.uk.

Citing

If you find this project useful, we encourage you to

• Star this repository :star: