A deep learning framework for high-throughput mechanism-driven phenotype compound screening.

Area of Science:

• Computational biology
• Drug discovery
• Bioinformatics

Background:

• Traditional drug discovery relies on target-based screening, which often fails due to poor correlation with organismal phenotypic response.
• Chemical-induced gene expression profiles offer a phenotype-based screening approach but are limited by data sparseness, unreliability, and low throughput.
• Existing imputation methods struggle with *de novo* chemical compound screening.

Approach:

• Developed DeepCE, a mechanism-driven neural network using graph convolutional networks for chemical representation and multi-head attention for chemical substructure-gene and gene-gene associations.
• Introduced a novel data augmentation method to extract information from unreliable experiments in the L1000 dataset.
• Evaluated DeepCE for predicting chemical-induced gene expression in both *de novo* screening and imputation settings.

Key Points:

• DeepCE outperforms state-of-the-art methods in predicting chemical-induced gene expression, demonstrating superior performance in *de novo* screening and imputation.
• Generated gene expression profiles from DeepCE align with L1000 data, proving effective for downstream tasks like drug-target and disease prediction.
• Successfully applied DeepCE to patient-specific drug repurposing for COVID-19, identifying novel lead compounds consistent with clinical evidence.

Conclusions:

• DeepCE offers a powerful framework for robust predictive modeling using noisy omics data.
• The method enables effective screening of novel chemicals for modulating systemic responses to disease.
• DeepCE demonstrates significant potential for advancing drug discovery and repurposing, particularly in complex diseases like COVID-19.