Oloren ChemEngine

Made by Oloren AI, the Python library Oloren ChemEngine enables the creation of state-of-the-art, complex molecular property predictors in just a few lines of code. Models defined and trained with olorenchemengine achieve super-leaderboard performance in less than 10 lines.

Installation

In a Python 3.8 environment, you can install the package with the following command:

bash <(curl -s https://raw.githubusercontent.com/Oloren-AI/olorenchemengine/master/install.sh)

Feel free to inspect the install script to see what is going on under the hood. This will work fine in both a conda environment and a pip environment.

Quick Start

 import olorenchemengine as oce

 ## Loading in a dataset

 # df is a Pandas Dataframe with the following columns:
 # "Smiles" (structure)
 # "pChEMBL Value" (property to be predicted)
 df = oce.ExampleDataFrame()

 ## Defining a model

 # The model is a gradient boosted model with the learners being:
 # 1. Random Forest model learning from a set of molecular descriptors
 # 2. GIN model pretrained using contrastive learning on PubChem
 # 3. a Random Forest model trained using a representation
 #    learned using contrastive learning on PubChem
model = oce.BaseBoosting([
         oce.RandomForestModel(oce.DescriptastorusDescriptor("rdkit2dnormalized"), n_estimators=1000),
         oce.RandomForestModel(oce.OlorenCheckpoint("default"), n_estimators=1000)])

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

 ## Saving the model
 oce.save(model, "model.oce")

 ## Loading the model
 model2 = oce.load("model.oce")

 ## Predicting property of new compounds
 y_pred = model2.predict(["CC(=O)OC1=CC=CC=C1C(=O)O"])

Philosophy

With the rapidly changing and diverse landscape of molecular property predictors, Oloren ChemEngine provides a common framework for the development, testing, and usage of AI model. Previously an exercise in chaos, with Oloren ChemEngine many different types of models–including ensembles of different modelling strategies and features–can be defined, trained and saved, imposing structure on the development of molecular property predictors while maintaining flexibility.

The following are the guiding principles for the development of Oloren ChemEngine models:

• Simplicity: models can be defined, trained, saved, and used with minimal effort.
• Flexibility: differing molecular representations, experimental datapoints, model architectures, ensembling strategies, and other innovative methodologies can be implemented in a consistent framework
• Accuracy: the capabilities of the library match or supercede top-of-the-leaderboard molecular property predictors, with a concerted focus on improving the utility of molecular property predictors in real-world settings, leveraging available experimental data.

Defined as subclasses of BaseModel, models including graph neural networks, descriptor- and fingerprint-based machine learning models, and Oloren AI proprietary models are all supported.

Contents

• Module Reference
  • olorenchemengine package
    • Subpackages
    • Submodules
    • olorenchemengine.base_class module
    • olorenchemengine.basics module
    • olorenchemengine.dataset module
    • olorenchemengine.ensemble module
    • olorenchemengine.gnn module
    • olorenchemengine.hyperparameters module
    • olorenchemengine.internal module
    • olorenchemengine.interpret module
    • olorenchemengine.manager module
    • olorenchemengine.reduction module
    • olorenchemengine.representations module
    • olorenchemengine.splitters module
    • olorenchemengine.uncertainty module
    • Module contents

Getting Started

• Representing Molecules
  • Transforming Molecules with Representations
  • Defining Models with Representations
  • Concatenating Representations
  • Defining New Representations
• Visualizing Models and Data

Examples

• 0 - Minimal Example
• 1A First Model
  • Getting the data
  • Making the model
• Model Searching
• Hyperparameter Optimization
• 1D - BasicVis
  • Visualizing the chemical space of a dataset
  • Visualizing model errors by cluster
  • VisualizeFeatures - exploratory data analysis (EDA) tool
• Predicting Confidence Intervals
  • SDC
  • ADAN
  • Aggergate Error Models: Random Forest
  • Analysis
• Plotting counterfactuals
  • Background
  • Methods
  • Example plots
• Integrated Error Models
• Production Level Models
• Visualizing Errors
• Putting it all together
  • Creating the dataset
  • Training the production-level model
  • Visualizing results
    • Ground truth output
    • Predicted output and error margin
• PyG ChemProp Implementation
  • Background
  • ChemProp Model Training
• Building your own molecular representation class
  • Parent Class Specs
  • Initialization
  • _convert function
  • Convert!
  • Using the representation in a model
• Creating your own visualizations
  • The BaseVisualization class
    • Basic flow
    • Notes on use
    • Notes on attributes class
  • Creating by example
    • First run-through
    • Second run-through

Indices and tables

• Index

• Module Index

• Search Page