E-GuARD

Expert-Guided Augmentation for Robust Interference Compound Detection

THIS NEEDS TO BE CHANGED WITH THIS CODE'S DOI

---

⚠️ Requires Python >= 3.10
(Tested on Linux systems)

Assay interference by small organic compounds poses significant challenges in early drug discovery. Existing computational methods often fall short due to data scarcity and imbalance, limiting their predictive power. E-GuARD introduces a novel framework that combines self-distillation, active learning, and generative models to tackle these challenges effectively and robustly.

Figure: Overview of E-GuARD. Outer loop (black arrows) for molecule generation and inner loop (dashed red arrows) for iterative updates through self-distillation.

---

Key Features

• Molecular Generation: Create novel compounds optimized for specific properties.
• Self-Distillation: Boost model performance through iterative training.
• Active Learning: Enhance efficiency by leveraging simulated or expert feedback.

---

Installation

1. Install REINVENT 4
   E-GuARD builds on REINVENT 4. Follow its installation guide.

2. Create a virtual Python >= 3.10 environment Clone this repository, navigate to its location and install the pip dependencies as follows

   pip install -r requirements.txt

3. Install MolSkill
   Activate the new environment and install MolSkill as follows

   conda install molskill -c msr-ai4science -c conda-forge

---

Usage

1. Train the Teacher Model

Train an initial teacher model using one of the provided datasets. For example, to train a model on the Thiol dataset:

python eGuard/teacher/train.py -d thiol.csv

---

2. Run the E-GuARD Workflow

Generate novel molecules and refine models using active learning:

python eGuard/run_pipeline/run_pipeline.py \
    -d {dataset}.csv \
    -i {number_of_iterations} \
    -s {acquisition_function}

Replace {dataset}.csv with your dataset (e.g., thiol.csv), {number_of_iterations} with the desired number of iterations, and {acquisition_function} with the selected active learning strategy.

---

Data

Datasets used for assay interference are sourced from the paper:
“Lies and Liabilities: Computational Assessment of High-Throughput Screening Hits to Identify Artifact Compounds”
Read the full article here.

Available datasets:

• FI: fluc.csv
• NI: nluc.csv
• TR: thiol.csv
• RR: redox.csv

These datasets are located in the data/ directory.

---

Reproducibility

Recreate figures and analysis results using the Jupyter notebooks provided in the notebooks/ directory.

---

Acknowledgements

We express our gratitude to Roxane Jacob and Vincent-Alexander Scholtz (University of Vienna) for their insightful contributions to the development of machine learning models.

---

📧 For inquiries, contact vincenzo.palmacci@univie.ac.at