Compared with “non-deep” methods.

We compared our approach with four “non-deep” machine learning methods, including Support Vector Machine (SVM), K-Nearest Neighbors (kNN), Logistic Regression (LR), and Random Forest (RF), which are highly popular in biomedical applications. Upon examining Table 1 and Fig 3A, we observed that among the four conventional machine learning methods, SVM performed the best, boasting a 74.01% accuracy rate and an AUC value of 0.7410. Our method has a slight improvement with the accuracy rate increased by 1.66% and the AUC value increased by 0.0309 over SVM. Moreover, our model exhibited a more balanced specificity and sensitivity. This indicates that our approach is more accurate in predicting DILI compared to conventional machine learning methods.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. ROC curves of BioGL-GCN and other models: (A) Compared with “non-deep” methods; (B) Compared with deep methods; (C) Compared with GCN; (D) Compared with glmGCN.

https://doi.org/10.1371/journal.pcbi.1013423.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Compared with other models.

https://doi.org/10.1371/journal.pcbi.1013423.t001

Compared with deep methods.

We also compared our proposed method with several state-of-the-art deep learning approaches, including DNN, singleDeep [29], and ToxGIN [30]. Unlike these methods that typically take raw gene expression profiles as input or rely on simplified network structures, our proposed BioGL-GCN constructed a gene-gene interaction graph, considering the aggregated effects of neighboring genes. As shown in Table 1 and Fig 3B, our proposed BioGL-GCN method significantly improved performance. Specifically, compared to DNN, we increased ACC by 1.04%, SPEC by 0.91%, SEN by 1.09%, F1 score by 0.78%, and AUC by 0.0086. Compared to singleDeep, BioGL-GCN improved ACC by 0.42% and AUC by 0.0275. In comparison with ToxGIN, our method increased ACC by 0.34% and AUC by 0.0105. Therefore, based on our experimental results, we concluded that incorporating graph-structured data containing gene-gene interaction information was more effective in improving prediction accuracy than relying solely on raw gene expression profiles.

Compared with GCN.

To evaluate the effectiveness of the BioGL layer, we compared our model against a standard GCN. Similar to BioGL-GCN, we fed the PPI network and gene expression data into the GCN network model to extract features, followed by fully connected layers and a softmax function for feature mapping and prediction. The distinction between BioGL-GCN and GCN lies in the depiction of gene-gene interactions, which corresponds to node-node relationships in the gene network graph. While GCN directly utilized the PPI network to encode gene interactions, BioGL-GCN augmented the GCN framework with the BioGL layer. This yielded a new pattern of gene interactions that integrated PPI information with gene similarity, thereby better representing the interactions between genes. From Table 1 and Fig 3C, it was evident that compared to GCN, BioGL-GCN led to improvements in accuracy by 2.29%, sensitivity by 5.52%, F1 score by 2.36%, and AUC by 0.0036. These experimental outcomes suggest that the unique gene-gene interactions derived from the BioGL layer afford deeper insight, thus leading to superior overall performance.

The impact of the improved BioGL layer.

In pursuit of an optimal and adaptable graph representation tailored for graph convolutional layers, our work introduced a BioGL layer that innovated upon the glmGCN framework proposed by Su et al. By incorporating gene frequency information, which indicated the significance of each gene within biological pathways and processes, the BioGL layer specifically highlighted the importance of gene interactions. We adopted the glmGCN model for comparison. Similar to BioGL-GCN, we fed the PPI network and gene expression data into the glmGCN network model to learn gene interaction graphs and extract features, and utilized fully connected layers and softmax for feature mapping and prediction. The distinction is that the glmGCN failed to include gene frequency information in this instance. As evident from Table 1 and Fig 3D, our model achieved an accuracy increase of 1.12%, specificity enhancements of 2.43%, F1 score increments of 0.69%, and an AUC uplift of 0.008. These results showed that incorporating gene frequency into the patterns of gene interactions learned by the graph learning layer could lead to more accurate DILI predictions.

Toxicogenomic profiles for model development.

Ting Li et al. curated a drug-induced transcriptome profiles dataset from the NIH LINCS L1000 dataset. They matched the Level 5 transcriptomic data from LINCS L1000 with the DILIst database using the PubChem Identifier service based on drug names and synonyms, obtaining 23,791 transcription profiles involving 69 cell lines. An improved Kennard-Stone algorithm was then used to extract transcription profiles with maximum explanatory variance. We matched the dataset information table [19] provided by Ting Li et al. into the LINCS L1000 dataset, resulting in a total of 6,000 transcription profiles of 978 landmark genes from 640 drugs (of which 3,568 were DILI positive and 2,432 were DILI negative). In this context, each transcription profile signifies the treatment effect of a unique combination of drug, dosage, duration, and cell line.

Gene frequency extraction via gene enrichment analysis.

Su et al. proposed that the co-expressed genes in certain biological process (BP) for multiple drugs can be used as indicators of toxicity [38]. Furthermore, genes that played more significant roles in the enriched BP tend to appear more frequently in the BPs for the drugs. Based on this observation, we believed that gene frequency information could be integrated as an important feature into the construction of gene interaction graph. We integrated gene frequency information into the construction of the BioGL. This information was obtained through Gene Ontology (GO) enrichment analysis [39], enabling a better understanding of interactions within complex biological networks. Specifically, we employed rigorous bioinformatics methodologies to identify differential expression genes (DEGs) in cells under drug perturbation conditions. This involved conducting differential expression analysis on the gene expression profiles for each drug, where a gene was classified as a DEG if its normalized expression value had an absolute value meeting or exceeding a predefined threshold of 2. The threshold was directly adopted from the definition of signature strength provided by Subramanian et al. [13], where it is used to quantify the number of landmark genes showing significant expression changes. It has since been widely adopted in large-scale transcriptomic studies such as those using the L1000 platform. We calculated the BPs enriched in each drug transcription profile through the GO enrichment analysis. Specifically, we performed the analysis using the enrichGO function in the clusterProfiler R package, with both the p-value cutoff and q-value cutoff set to 1. This setting ensured that no significance filtering was applied, allowing us to retain all possible GO terms for subsequent gene frequency calculation. One gene might enrich more than once for all the BPs of one drug transcription profile. Assuming we had M drug transcription profiles and N genes, for drug transcription profile j, denoted by D_j, the genes enriched in totally BPs. For the gene, it enriched totally on BPs for D_j.

(1)

Here, p_i represents the frequency of the gene. The gene frequencies obtained for the 978 landmark genes were normalized in preparation for their use as inputs into the network.

Construct the PPI network.

It has been reported that PPI networks play a crucial role in cellular functions and biological processes. Graph theory has demonstrated significant effectiveness when applied to the study of PPI networks [40].

In this study, we utilized the PPI network to model gene relationships, where genetic interaction scores from the STRING database [41] were used to quantify the strength of connections between gene nodes. Specifically, we retrieved high-confidence gene interactions from STRING using a minimum confidence score of 0.7, which corresponded to the “high confidence” threshold recommended by the database. Based on these interactions, we constructed the adjacency matrix , where N was the number of genes (978 landmark genes from the LINCS L1000 dataset). If there was direct interaction between two genes, we recorded the score in matrix A;otherwise, we marked it as 0.

To preserve node self-features during graph convolution, we incorporated an identity matrix I, representing self-loop connections, into A, setting all diagonal elements to 1. The final PPI network derived from the LINCS L1000 dataset contained 4,818 high-confidence gene interactions among the 978 landmark genes.

Active ingredients of Traditional Chinese Medicine.

Compared to small molecules, plant-derived natural products typically feature larger and more complex molecular structures. To validate our module’s ability to effectively predict the hepatotoxicity of natural products based on drug transcriptional profiles, we collected FPKM-normalized gene expression data for 496 active ingredients of TCM from the ITCM platform (http://itcm.biotcm.net) [42]. The expression profiles were obtained via RNA-seq following treatment with each ingredient at a dose of 10 μM for 12 hours in MCF-7 cell line. For each ingredient, the expression profiles included 3 biological replicates. To ensure that the data was aligned with the training dataset (LINCS L1000 level 5 gene expression profiles) of model, we firstly employed the moderated z-score (MODZ) procedure used for LINCS L1000 data processing to derive a consensus replicate signature for each ingredient (including the blank control). Briefly, a pairwise Spearman correlation matrix was computed between the replicate signatures with trivial self-correlations being ignored (set to 0). Weights for each replicate were then computed as the sum of its correlations to the other replicates, normalized such that all weights sum to 1. Finally, the consensus signature was obtained by the linear combination of the replicate signatures with the coefficients set to the weights. This procedure could effectively mitigate the effects of uncorrelated or outlier replicates. Subsequently, to obtain the relative gene expression of each ingredient, we used the following formula:

(2)

Here, CS_ingredient represents the consensus signature for ingredient and CS_control represents the consensus signature for control. The relative expression levels of the 978 landmark genes for each ingredient were extracted and input into the trained BioGL-GCN model to predict their hepatotoxicity.

Bio-graph learning layer.

The input to the GCN is denoted as , where X_gcn signifies the features of individual nodes and A_gcn represents the adjacency matrix that encodes the relational information between nodes. Su et al. proposed a novel approach, named glmGCN, which embeds a graph learning (GL) module within the GCN framework. This glmGCN integrated GL with GCN to attain enhanced graph representations and facilitated semi-supervised learning. First, the PPI network was employed to obtain the adjacency matrix A_gcn, which represented the graphical relationships between genes. In the GL layer, given node information , a non-negative function S_glmGCN was employed to represent the relationship between data points x_i and x_j based on their nodal distance connections. By combining the adjacency matrix A_gcn with the function S_glmGCN, a new graph representation was constructed. Inspired by Su et al., we also incorporated a BioGL module in our architecture to learn the optimal graph structure, as depicted in Fig 1C. In contrast to the approach of Su et al., our model was based on gene expression profiles and incorporated additional biological information—the frequency information of genes obtained through enrichment analysis—within the BioGL layer.

To optimally construct the neighborhood structure of the data, we devised a nonlinear function S based on the gene expression matrix , adjacency matrix , and gene frequency information . We defined g_i and p_i as the gene expression level and gene frequency of the gene, respectively. As mentioned, gene frequency information was regarded as an important feature of genes. Su et al. used to represent the distance between genes. Similarly, we considered that the same approach was applicable to gene frequencies. Since gene frequency is a one-dimensional measure, we used the product to represent the distance between the gene and the gene. Let represent the relationship between the and genes. Here, we reduced the dimensionality of S by performing calculations in a low-dimensional space parameterized by the projection matrix , where d<N , to enhance computational efficiency. The graph S was learned through the BioGL layer as follows:

(3)

(4)

Here, A_ij represents the connection between the gene and the gene in the adjacency matrix A derived from the constructed PPI network. ϕ is a tunable constant that, by raising A to the power of ϕ, emphasizes the importance of the initial graph and magnifies the distinction between strong and weak interactions. σ is an activation function, and α is a learnable parameter vector. Ultimately, a final softmax operation ensured that the learned graph S adhered to the following properties:

(5)

The weight vector α and W_g were optimized through the following loss function:

(6)

Here, γ and β are two tunable constants, F denotes the Frobenius norm. If the distance between the gene and the gene is large, S should be relatively small. The second term serves as regularization to control the sparsity of the learned gene relationship matrix S, with γ being the regularization coefficient. We also incorporated the initial relationship matrix A of genes into the process.

Graph convolutional network.

Our proposed network, built upon GCN and incorporating the BioGL layer, utilized the BioGL layer to learn the graph representation S, which was subsequently used in the graph convolutional layers. The output of the graph convolutional layers can be calculated as follows:

(7)

Here, ;is the output activation of the layer. is a trainable weight matrix for each convolutional layer. is an input-to-hidden weight matrix for a hidden layer with h(0) feature maps. is a hidden-to-output weight matrix for a hidden layer with h(L–1) feature maps (C is the class number. Here, C = 2). And denotes an activation function.

Assuming y_i denotes the true label of the sample, and z_i indicates the probability predicted by the model that the sample belongs to the positive class (designated as class 1). Cross entropy loss function was used here:

(8)

All parameters of the entire architecture were optimized through the following approach:

(9)

The entire model encompassed an input layer, multiple hidden layers, and an output layer. The hidden layers consisted of a BioGL layer, two graph convolutional layers, and three fully connected layers. Batch normalization was employed after the BioGL layer and the two graph convolutional layers to optimize parameters, stabilizing the learning process. A Flatten layer was adopted to transform pooled features into a one-dimensional vector. Following this, the fully connected layers were used to map the distributed features, with softmax being employed for the final prediction.

The pseudo-code for the proposed method is outlined in Algorithm 1:

Algorithm 1 Constructing the BioGL-GCN Model for Predicting DILI.

Input:

  Gene expression matrix G, hepatotoxicity labels Y

Output: The BioGL-GCN model for predicting DILI.

1: Construct the PPI network based on the 978 landmark genes,

  obtaining the adjacency matrix A;

2: Include self-linkage relations: Form , where I

  is the identity matrix;

3: Perform gene enrichment analysis to obtain the gene frequency

  matrix P;

4: Split the data into training set Gtrain and testing set

  Gtest;

5: Calculate the new pattern of gene interactions S;

6: Feed Gtrain, Ytrain, and S into the GCN and train the model;

7: Five-fold cross-validation;