• Quickstart
• Colab
• Web
• Docker
• Reproduction / Custom Models

Please cite: https://www.biorxiv.org/content/10.1101/2023.01.28.526009v1

@article{10.1101/2023.01.28.526009, 
  year = {2023}, 
  title = {{Explainable prediction of catalysing enzymes from reactions using multilayer perceptrons}}, 
  author = {Probst, Daniel}, 
  doi = {10.1101/2023.01.28.526009},
  note = {bioRxiv 2023.01.28.526009},
}

As you need at least Python 3.9 to get started, I suggest you use conda to create an environment with an up-to-date Python versions (3.11 is really, really fast, so I suggest using this as soon as rdkit supports it). For now, let's go with Python 3.10: conda create -n theia python==3.10 && conda activate theia is all you need (ha). Then you can go ahead and install theia using pip (theia was taken, so make sure to install theia-pypi, except if you want to parse log files):

pip install theia-pypi

Next, download the models using the CLI command:

theia-download

Thats pretty much it, now you can start theia by simply typing:

theia

and open the url https://127.0.0.1:5000/ in your web browser.

In case you don't want or need an UI, you can also use the cli to simply predict an EC number from an arbitrary reaction:

theia-cli "rheadb.ec123" "S=C=NCC1=CC=CC=C1>>N#CSCC1=CC=CC=C1"

If you want a bit more information than just the predicted EC class, you can also get the top-5 probabilities:

theia-cli "rheadb.ec123" "S=C=NCC1=CC=CC=C1>>N#CSCC1=CC=CC=C1" --probs

Or, if you want human-readable output, you can make it pretty:

theia-cli "rheadb.ec123" "S=C=NCC1=CC=CC=C1>>N#CSCC1=CC=CC=C1" --probs --pretty

and you'll get a neat table...

┏━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━┓
┃ Prediction ┃ Probability [%] ┃
┡━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━┩
│ 2.7.4      │           14.22 │
│ 2.3.2      │           11.03 │
│ 2.3.1      │            7.15 │
│ 2.7.8      │            4.62 │
│ 2.6.1      │            4.05 │
└────────────┴─────────────────┘

Of course, there are more models than rhea.ec123, which we used in the previous examples. Here's a complete list of all the included models:

You can also give the API a spin in this Google colab notebook. Keep in mind that Colab has old and slow CPUs with outdated instruction sets, so you might want to turn the GPU on. On a modern CPU both training and inference is fairly fast.

A demo of the web application can be found here. Keep in mind that this service has limited resources and that a locally installed version (even on your laptop) will be much, much faster.

A docker image is available on the docker hub here. After running the docker image, the app will be available at https://localhost:8000/theia.

To get started, install the reproduction requirements with:

pip install .
pip install -r reproduction_requirements.txt

The training, validation, and test sets used in the manuscript can be recreated using the following two commands (of course, you can plug in your own data sets here to get a custom model):

mkdir experiments/data
python scripts/encode_split_data.py data/rheadb.csv.gz experiments/data/rheadb
python scripts/encode_split_data.py data/ecreact-nofilter-1.0.csv.gz experiments/data/ecreact

The training of the models can be started with:

mkdir experiments/models
chmod +x train_all.sh
./train_all.sh

If you want to train the 6 additional models for cross-validation, you can run the following:

chmod +x train_all_cross.sh
./train_all_cross.sh

Finally, to reproduce the figures, you first have to run some additional data crunching scripts:

python scripts/class_counts.py data/ecreact-nofilter-1.0.csv.gz experiments/data/ecreact_counts.csv
python scripts/class_counts.py data/rheadb.csv.gz experiments/data/rheadb_counts.csv

Then it's time to draw:

cd figures
chmod +x generate_figures.sh
./generate_figures.sh

As a bonus, the ablation studies, where a fraction of the training set labels are randomised can be run using an additional script:

chmod +x train_rhea_ablation.sh
./train_rhea_ablation.sh

fin.