Abstract
Deep Learning and DRUG-seq (Digital RNA with perturbation of genes) have emerged as powerful tools in drug discovery. However, public DRUG-seq datasets are often too small to train deep neural networks from scratch. This study presents a transfer learning approach where a drug efficacy prediction model, DeepCE, is pretrained on the extensive Library of Integrated Network-based Cell-Signature (LINCS) L1000 data and subsequently fine-tuned using human neural cell DRUG-seq data. This pipeline was applied to virtual screening against Alzheimer's disease (AD), identifying 27 potential drug candidates, including Irsogladine, Tasquinimod, and Suprofen.
Introduction: The Challenge of AD Drug Discovery
Alzheimer’s disease (AD) remains one of the most significant challenges in modern medicine. Traditional drug discovery is slow and costly. Drug repositioning—identifying new therapeutic uses for existing, approved drugs—offers a more efficient path forward because the safety profiles of these chemicals are already established.
Computationally, repositioning often involves analyzing how drugs alter gene expression. While the L1000 dataset has been the standard resource for transcriptomics, it is primarily derived from cancer cell lines and can be noisy. DRUG-seq offers a more accurate and cost-effective alternative, but its datasets are currently too limited for standalone deep learning training. This research bridges that gap by applying transfer learning to leverage large-scale "noisy" data for specialized "clean" neural cell applications.
Methodology: Bridging the Data Gap
The researchers employed DeepCE, a neural network architecture utilizing multi-head attention and Graph Neural Networks (GNNs) to predict gene expression profiles from chemical structures (represented as SMILES strings).
The pipeline consists of four key stages:
- Pretraining: The model was trained on a high-quality subset of the L1000 dataset (1,944 treatments) to learn general chemical-gene interactions.
- Fine-tuning: The pretrained weights were transferred to a model fine-tuned on a smaller DRUG-seq dataset of human neural cells (506 treatments).
- Virtual Screening: The trained model predicted gene expression profiles for over 11,000 drugs from the DrugBank database.
- Signature Matching: Predicted drug signatures were compared against AD patient transcriptome data from post-mortem hippocampi. Drugs showing a high negative correlation—effectively "reversing" the AD signature—were prioritized.
Key Results and Insights
The transfer learning approach yielded a substantial performance increase. While training on DRUG-seq data alone resulted in a Spearman correlation coefficient of 0.2556, the transfer learning pipeline improved this to 0.4301, demonstrating the effectiveness of knowledge transfer from cancer cell data to neural cell data.
The screening identified 27 promising candidates for AD treatment. Notable findings include:
- Irsogladine: A PDE4 inhibitor; such inhibitors are known to enhance memory in AD models.
- Tasquinimod: An HDAC4 selective inhibitor that can upregulate memory-related genes.
- Suprofen: A dual COX-1/COX-2 inhibitor, targeting the neuroinflammation central to AD progression.
- Amitriptyline: An FDA-approved antidepressant with previously documented neuroprotective benefits in AD.
Conclusion and Future Directions
This study serves as a successful proof-of-concept for using DRUG-seq data within deep learning frameworks via transfer learning. It demonstrates that knowledge from general cell lines (L1000) can be effectively adapted to specific, high-relevance tissues like neural cells.
Future work will focus on scaling this approach as larger DRUG-seq datasets become available. Refining the model to predict relative gene expression changes directly, rather than absolute values, may further enhance prediction accuracy and the discovery of more nuanced therapeutic candidates.
Summary updated