Predicting DNA damage response using synthetic cell painting profiles and experimental analysis.

Area of Science:

• Computational biology
• Machine learning
• Drug discovery

Background:

• Detecting DNA damage response (DDR) via cell painting is hindered by small sample sizes and imbalanced datasets.
• Developing robust predictive models for DDR is crucial for efficient drug discovery pipelines.

Purpose of the Study:

• To establish a machine learning framework for enhanced DDR prediction using synthetic cell painting data.
• To evaluate different synthetic data generation algorithms and classifiers for DDR detection.

Main Methods:

• Generated synthetic cell painting profiles using Gaussian copula, CTGAN, VAE, and CopulaGAN on the idr-0080 dataset.
• Assessed synthetic data quality using fidelity metrics.
• Trained and evaluated Support Vector Machine (SVM) classifiers using real and synthetic data, addressing class imbalance.
• Utilized SHAP analysis to identify key morphological features driving predictions.

Main Results:

• An SVM model trained on real data augmented with Gaussian copula synthetic data achieved the highest performance (F1-score = 0.87, AUROC = 0.94).
• SHAP analysis identified critical morphological features contributing to DDR prediction.
• The model successfully identified known and novel DDR inducers in an external dataset (cpg-0012), validated by γH2AX marker and cell viability assays.

Conclusions:

• The machine learning framework integrating synthetic cell painting profiles offers a robust and scalable approach for DDR prediction.
• This method provides an effective virtual prescreening strategy for identifying potential drug candidates in early-stage drug discovery.