Oloren ChemEngine

Oloren ChemEngine (oce) is a software package developed and maintained by Oloren AI containing a unified API for the development and use of molecular property predictors enabling

• Direct development of high-performing predictors
• Integration of predictors into model interpretability, uncertainty quantification, and analysis frameworks

Here's an example of what we mean by this. In less than ten lines of code, we'll train, save, load, and predict with a gradient-boosted model with two different molecular vector representations.

import olorenchemengine as oce

df = oce.ExampleDataFrame()

model = oce.BaseBoosting([
            oce.RandomForestModel(oce.OlorenCheckpoint("default"), n_estimators=1000),  # RF w/ our proprietary fingerprint
            oce.SPGNN(model_type="contextpred"),  # fine tune a trained GNN on your data
        ])

model.fit(df["Smiles"], df["pChEMBL Value"])

oce.save(model, "model.oce")

model2 = oce.load("model.oce")

y_pred = model2.predict(["CC(=O)OC1=CC=CC=C1C(=O)O"])

It's that simple! And it's just as simple to train a graph neural network, generate visualizations, and create error models. More information on features and capabilities is available in our documentation at docs.oloren.ai.

If you are wondering what models to build, please visit benchmarks.chemengine.org.

Getting Started with oce

Installation

Prerequisites

• Install Pytorch: https://pytorch.org/TensorRT/tutorials/installation.html
• Install Pytorch Geometric: https://pytorch-geometric.readthedocs.io/en/latest/notes/installation.html

Install Command:

pip install olorenchemengine[full]

Here are some common error messages and solutions: https://oloren-ai.notion.site/Oloren-ChemEngine-Installation-FAQ-f2edec771a7f4350af5fdc361d494604

Install Live Version from GitHub Command

pip install --upgrade "olorenchemengine[full] @ git+https://github.com/Oloren-AI/olorenchemengine.git"

Docker

Alternatively, you can also run OCE from one of our docker images. After cloning the repo, just run:

docker build -t oce:latest -f docker/Dockerfile.gpu . # build the docker image
docker run -it -v ~/.oce:/root/.oce oce:latest python # run the docker image

Replace ".gpu" with ".cpu" in the docker path if you want to run the project in a dockerized environment.

Basic Usage

We have an examples folder, which we'd highly reccomend you checkout--1A and 1B in particular--the rest of the examples can be purused when the topics come up.

Notice

Maintaining and developing Oloren ChemEngine requires a lot of resources. As such, we would like to log for each evaluated model the model hyperparameters, the model performance metrics and a unique, non-identifying hash of the dataset. These logs are used to improve our models. Below is a representative example of such a log:

{dataset_hash: "149eae5c763afcc14f6355007df298b05f4a51c6a334ea933fbe7fc496adb271",

metric_direction: null,

metrics: "{"Average Precision": 0.9479992350277128, "ROC-AUC": 0.7450549450549451}",

name: "BaseBoosting 1zpI0dIb",

params: "{"BC_class_name": "BaseBoosting", "args": [[{"BC_class_name": "RandomForestModel", "args": [{"BC_class_name": "DescriptastorusDescriptor", "args": ["morgan3counts"], "kwargs": {"log": true, "scale": null}}], "kwargs": {"bootstrap": true, "criterion": "entropy", "max_features": "log2", "n_estimators": 2000, "max_depth": null, "class_weight": null}}, {"BC_class_name": "RandomForestModel", "args": [{"BC_class_name": "DescriptastorusDescriptor", "args": ["morganchiral3counts"], "kwargs": {"log": true, "scale": null}}], "kwargs": {"bootstrap": true, "criterion": "entropy", "max_features": "log2", "n_estimators": 2000, "max_depth": null, "class_weight": null}}, {"BC_class_name": "RandomForestModel", "args": [{"BC_class_name": "DescriptastorusDescriptor", "args": ["morganfeature3counts"], "kwargs": {"log": true, "scale": null}}], "kwargs": {"bootstrap": true, "criterion": "entropy", "max_features": "log2", "n_estimators": 2000, "max_depth": null, "class_weight": null}}, {"BC_class_name": "RandomForestModel", "args": [{"BC_class_name": "DescriptastorusDescriptor", "args": ["rdkit2dnormalized"], "kwargs": {"log": true, "scale": null}}], "kwargs": {"bootstrap": true, "criterion": "entropy", "max_features": "log2", "n_estimators": 2000, "max_depth": null, "class_weight": null}}, {"BC_class_name": "RandomForestModel", "args": [{"BC_class_name": "OlorenCheckpoint", "args": ["default"], "kwargs": {"log": true, "num_tasks": 2048}}], "kwargs": {"bootstrap": true, "criterion": "entropy", "max_features": "log2", "n_estimators": 2000, "max_depth": null, "class_weight": null}}]], "kwargs": {"log": true, "n": 1, "oof": false, "nfolds": 5}}"}

The dataset hash is created with the following code:

import joblib

dataset_hash = joblib.hash(X) + joblib.hash(y)

This means that we log no therapeutics-related data whatsoever. We just log hashes of model performance.

If you would still prefer a logging-free version, please fill out the following form to obtain a version with all logging code excised: https://y09gl0qf49q.typeform.com/to/brGMidJ0.

We also require contributor agreements for all versions of Oloren ChemEngine.

oce at a high level

Everything in oce is built around Oloren's BaseClass system, which all classes stem from. Any BaseClass derived objects has its parameters and complete state saved via parmeterize and saves respectively. A blank object (no internal state) can be recreated via create_BC and a complete object (with internal state) can be recreated via loads.

The system includes abstract subclasses of BaseClass are named Base{Class Type} and their interactions, most prominently

• BaseModel, a base class for all any molecular property predictor
• BaseRepresentation, a base class for all molecular representations
• BaseVisualization, a base class for all types of visualizations and analyses

Contributing

First, thank you for contributing to OCE! To install OCE in editable/development mode, simply clone the repository and run:

bash install.sh --dev

This will install the repo in an editable way, so your changes will reflect immediately in your python environment. All tests for OCE are in the tests directory and can be run by running pytest in this directory. Please contact [email protected] if you need any assistance in your development process!

PRs from external collaborators will require a Contributor License Agreement (CLA) to be signed before the code is merged into the repository.

Our Thanks

First, our thanks to the community of developers and scientists, who've built and maintained a repotoire of software libraries and scripts which have been invaluable. We'd like to particularly thank the folks creating RDKit, PyTorch Geometric, and SKLearn who've developed software we strive to emulate and exceed.

Second, we'd like to thank the amazing developers at Oloren who've created Oloren ChemEngine through enoromous effort and dedication. And, we'd like to thank our future collaborators and contributors ahead, who we're excited meet and work with.

Third, huge gratitude goes to our investors, clients, and customers who've been ever patient and ever gracious, who've provided us with the opportunity to bring something we believe to be truly valuable into the world.