TL;DR
This paper introduces DeepCE, a novel deep learning framework that predicts how chemicals affect gene expression, enabling more effective phenotype-based drug screening. By integrating graph convolutional networks and attention mechanisms, DeepCE overcomes limitations of existing methods, particularly in predicting the effects of new, un-tested compounds and handling noisy experimental data.
What Problem Does This Paper Address?
Traditional drug discovery often relies on target-based screening, focusing on how a chemical interacts with a single protein. However, this approach frequently leads to high failure rates because a protein-level interaction doesn't always translate to a beneficial organism-level phenotypic response. Phenotype-based screening, which looks at the overall cellular or organismal effect, is more promising but has historically been slow and difficult to interpret. Chemical-induced gene expression profiles offer a way to bridge this gap, providing a readout of a chemical's systemic effect. However, existing gene expression datasets (like L1000) are plagued by missing data, experimental noise, and low throughput, making it challenging to use them for de novo (new) compound screening. This paper aims to address these limitations by developing a robust computational framework for predicting gene expression profiles of novel chemicals.
Methods
The core of the proposed solution is DeepCE, a deep learning model designed to predict gene expression profiles. Key methodological innovations include:
- Graph Convolutional Network (GCN): Used to learn rich, data-driven chemical substructure representations (neural fingerprints) directly from the chemical's graph structure. This is an advancement over using predefined fingerprints.
- Multi-Head Attention Mechanism: This is crucial for modeling complex interactions. It captures associations between:
- Chemical substructures and genes.
-
Genes and other genes within a cellular context. This allows DeepCE to learn intricate relationships that linear models or simpler neural networks might miss.
-
Data Augmentation: A novel method to extract useful information from unreliable experiments within the L1000 dataset. It leverages the idea that even if an entire experiment is noisy, some of its bio-replicates might still contain valuable, reliable data.
- Multi-Output Prediction: A feed-forward neural network that takes the learned representations and predicts gene expression values for all L1000 genes simultaneously.
Datasets
- L1000 Dataset: This is the primary dataset used for training and evaluating DeepCE. It contains a vast number of gene expression profiles from human cell lines perturbed by various compounds at different doses and time points. The paper specifically uses the "Bayesian-based peak deconvolution L1000 dataset" (level 5 data), which is noted for generating more robust z-score profiles.
- STRING Database: Used to derive vector representations for L1000 genes based on human protein-protein interactions.
- DrugBank Database: Utilized for drug-target interaction and disease prediction tasks, and importantly, for the drug repurposing application. It provided information on drugs, their targets, and ATC labels.
- NCBI Gene Expression Omnibus (GSE147507): Used for patient-specific gene expression data in the context of COVID-19.
Key Insights and Ideas
- Beyond Target-Based Screening: The paper strongly advocates for phenotype-based screening via gene expression profiles, highlighting the limitations of traditional target-centric approaches in drug development.
- The Power of Mechanism-Driven Prediction: By modeling chemical substructure-gene and gene-gene interactions, DeepCE aims to capture the underlying biological mechanisms, leading to more accurate predictions.
- Leveraging Noisy Data: The data augmentation technique is a significant contribution, demonstrating that even "unreliable" experimental data can be mined for valuable insights, thereby expanding the utility of large omics datasets.
- De Novo Screening is Crucial: The ability to predict the effects of entirely new chemicals, not seen during training, is a critical step for high-throughput screening and drug discovery. DeepCE is shown to excel in this area.
- Gene Expression as a Universal Language: The success in downstream tasks like drug-target and disease prediction, and particularly the COVID-19 drug repurposing, underscores the power of gene expression profiles as a common representation for biological effects.
Key Results
- Superior Performance in De Novo Setting: DeepCE significantly outperforms baseline models (vanilla neural network, kNN, linear models, TT-WOPT) in predicting gene expression profiles for novel chemicals, achieving a Pearson correlation of 0.4907 on the testing set.
- Strong Imputation Performance: DeepCE also excels in traditional imputation settings, outperforming TT-WOPT with a Pearson correlation of 0.7010 compared to 0.5113.
- Data Augmentation Boosts Performance: The data augmentation method improved DeepCE's performance, increasing the Pearson correlation from 0.4907 to 0.5014.
- Effective for Downstream Tasks: Gene expression profiles generated by DeepCE led to better performance in drug-target and ATC code classification tasks compared to using original, potentially unreliable, L1000 data.
- COVID-19 Drug Repurposing: DeepCE successfully identified potential COVID-19 treatments by screening compounds against patient gene expression profiles, including several antiviral drugs and Hepatitis C treatments (Elbasvir, Velpatasvir) that show promise.
Limitations and Future Work
- Limitations of L1000 Data: While DeepCE mitigates issues, the inherent sparseness and unreliability of the L1000 dataset remain a challenge. The paper notes that linear models perform poorly, indicating that complex, non-linear relationships are key.
- Computational Resources: The paper mentions that some tensor-based methods (like polyadic regression) are computationally infeasible for high-dimensional data, highlighting the need for efficient deep learning approaches like DeepCE.
- Future Work: The authors suggest that DeepCE provides a powerful framework for robust predictive modeling and screening novel chemicals. The application to COVID-19 drug repurposing is presented as a demonstration of its value, implying further applications in other disease areas. The availability of the source code and generated profiles suggests a path for broader community use and further development.
Summary updated