TGSA: protein–protein association-based twin graph neural networks for drug response prediction with similarity augmentation

Abstract

Motivation Drug response prediction (DRP) plays an important role in precision medicine (e.g., for cancer analysis and treatment). Recent advances in deep learning algorithms make it possible to predict drug responses accurately based on genetic profiles. However, existing methods ignore the potential relationships among genes. In addition, similarity among cell lines/drugs was rarely considered explicitly. Results We propose a novel DRP framework, called TGSA, to make better use of prior domain knowledge. TGSA consists of Twin Graph neural networks for Drug Response Prediction (TGDRP) and a Similarity Augmentation (SA) module to fuse fine-grained and coarse-grained information. Specifically, TGDRP abstracts cell lines as graphs based on STRING protein-protein association networks and employs Graph Neural Networks (GNNs) for representation learning. SA views DRP as an edge regression problem on a heterogeneous graph and utilizes GNNs to smooth the representations of similar cell lines/drugs. Besides, we introduce an auxiliary pre-training strategy to remedy the identified limitations of scarce data and poor out-of-distribution generalization. Extensive experiments on the GDSC2 dataset demonstrate that our TGSA consistently outperforms all the state-of-the-art baselines under various experimental settings. We further evaluate the effectiveness and contributions of each component of TGSA via ablation experiments. The promising performance of TGSA shows enormous potential for clinical applications in precision medicine. Availability The source code is available at https://github.com/violet-sto/TGSA. Supplementary information Supplementary data are available at Bioinformatics online.

Deep learning methods for drug response prediction in cancer: Predominant and emerging trends

“…Integrating multiomics into the learning process further exacerbates the already problematic feature size of single omics data. A few papers indeed demonstrate significant performance boost with multiomics ( 45 , 49 , 108 ) but the majority report only marginal improvement ( 104 , 108 , 118 , 126 , 141 , 152 ). As opposed to DREAM participators which utilized agreed-upon datasets and scoring metrics, the DRP models in Supplementary Table 1 substantially differ among them as discussed earlier, largely contributing to discrepancies and mixed conclusions.…”

“…There is slightly more diversity of approaches in this space as compared to drugs partially due to the use multiomic data which allow constructing graphs with heterogeneous node attributes. The multiomics can be utilized to encode gene relationships and attributes using one or more data modalities, including correlations between genes ( 109 , 111 , 115 , 116 ), known protein interactions (i.e., PPI) ( 109 , 111 , 117 , 118 ) using STRING database ( 119 ), and relationships based on known gene pathways ( 109 ) using GSEA dataset ( 120 ). Recently, novel approaches have been explored such as heterogeneous graphs where both cell lines and drugs are encoded as graph nodes ( 108 , 116 , 118 , 121 ), and a model that utilizes diverse data types for building graphs, including differential gene expressions, disease-gene association scores and kinase inhibitor profiling ( 111 ).…”

“…Considering N drugs in the dataset, the model was trained in an MTL fashion with N prediction tasks, one task for each drug, outperforming the single weight layer configuration. Zhu et al ( 118 ) used MTL in a pre-training step to accumulate chemical knowledge by training a GNN to predict more than a thousand biochemical properties of molecules, a strategy adopted from another work ( 132 ). The pre-trained model was integrated into the complete DRP model called TGSA.…”

“…However, the actual baselines and ablation experiments that are chosen to benchmark models are a critical aspect that affects credibility. Some DL models explore alternative data representations, such as images ( 76 , 77 ) or graphs ( 105 , 118 ). In these cases, ablation studies with common vector-based representations (e.g., FPS, descriptors) should be considered.…”

Data augmentation and multimodal learning for predicting drug response in patient-derived xenografts from gene expressions and histology images

“…However, the high dimensionality of gene expressions and the relatively small sample size can lead to overfitting (2). To address this issue, several gene selection methods have been utilized, including filtering genes based on variability across samples (29)(30)(31) or using gene subsets such as LINCS (10,(32)(33)(34)(35)(36) or COSMIC (35, [37][38][39][40] that are known to be associated with cancer and/or treatment response. We are not aware of any systematic analysis that studied which filtering method better addresses overfitting and improves prediction generalization.…”