Mapping drug physico-chemical features to pathway activity reveals molecular networks linked to toxicity outcome

Abstract

The identification of predictive biomarkers is at the core of modern toxicology. So far, a number of approaches have been proposed. These rely on statistical inference of toxicity response from either compound features (i.e., QSAR), in vitro cell based assays or molecular profiling of target tissues (i.e., expression profiling). Although these approaches have already shown the potential of predictive toxicology, we still do not have a systematic approach to model the interaction between chemical features, molecular networks and toxicity outcome. Here, we describe a computational strategy designed to address this important need. Its application to a model of renal tubular degeneration has revealed a link between physico-chemical features and signalling components controlling cell communication pathways, which in turn are differentially modulated in response to toxic chemicals. Overall, our findings are consistent with the existence of a general toxicity mechanism operating in synergy with more specific single-target based mode of actions (MOAs) and provide a general framework for the development of an integrative approach to predictive toxicology.

Figure 1. Analysis strategy to compute indices of pathway activity.

To compute the indices of pathway activity the first step is to summarize the gene expression profiles using PCA according to KEGG pathways. This results in 148 pathway indices summarized using two PCs. These PC can then be used as an input to a Hotelling's statistics to compute the perturbation index for a specific drug as compared to a matched control group. The third step is to visualize the relationship between the drugs with the use of a hierarchical clustering. We can then show that the dimensionality reduction in step 1 is biologically relevant to use in the subsequent analysis.

Figure 2. Integrating pathways associated to PCFs and toxicity.

Alongside using a regression based model to identify pathways associated to PCFs (Step 1) we also identified pathways associated to toxicity by the use of the statistics (Step 2). The resulting pathways were then mapped onto a KEGG pathway map to identify clusters of pathways associated to both PCFs and toxicity (Step 3). Finally we asked the question if the PCFs we have found to be associated to pathways are a better predictor of toxicity.

Figure 3. Hierarchical clustering of chemicals based on pathway modulation profiles.

The figure shows the clustering of chemicals on the basis of the extent of change of transcriptional activity in molecular pathways after exposure. Panel A represent the relationship between the samples when the change in pathway activity is represented simultaneously by the first and second PCs (multi-variate Hotelling test). Panels B and C represent respectively the results of cluster analysis when the change in pathway activity is estimated by the PC1 or the PC2 (univariate t-test). Notice that toxic chemicals cluster (highlighted areas on Panels A and C) on the basis of the multivariate test and that the information associated to toxicity is primarily represented by the PC2. Toxic chemicals have been highlighted using a red square on the left of each clustering.

Figure 4. Visual summary of descriptor connections.

The network represents the number of interactions between PCFs descriptor groups computed from pooled models. The thickness of the line is proportional to the number of pathways in which PCFs of a given descriptor group are selected in an interaction component of a predictive models. The highest value edge is found between ET-State and Geometrical descriptors in which 11 out of 19 pathways were found to contain models based on features from these 2 descriptor groups.

Figure 5. Example models linking PCFs with molecular pathway activity.

The figure shows the relationships between the observed (x axis) and predicted (y axis) indices of pathway activity for a number of exemplar KEGG pathways. Nephrotoxic samples are represented by red dots whereas non-nephrotoxic samples are represented by black dots. Gap Junction and ErbB Signaling Pathway contain features belonging to ET-State indices, Geometrical descriptors and RDF descriptors. The values are 0.55 and 0.57 respectively. Wnt Signaling Pathway and Adipocyte Signaling Pathway contain features belonging to GSFRAG, Information indices, Edge adjacency indices and 3D-MoRSE descriptors. The values are 0.52 and 0.51 respectively. Note that models containing a feature from E-State indices and RDF descriptors better separate nephrotoxic and non-nephrotoxic samples.

Figure 6. KEGG Pathway topology map.

The Figure shows a dendrogram representing the degree of similarity between different KEGG pathways. Pathways marked in red are pathways that were found to be associated to chemical features (19), and pathways marked in blue have been found to be predictive of toxicity (21). Pathways whose activity is predicted by PCFs group in a tight cluster. Note that the majority of toxicity annotated pathways cluster towards the lower half of the dendrogram, close to pathways linked to PCFs.

Figure 7. PCFs linked to molecular response are better predictors of toxicity.

Panel A shows the comparison between the classification accuracy of models predictive of toxicity and developed selecting from PCFs which are predictive of molecular response and those developed using uncorrelated PCFs. Note that PCFs linked to molecular response have a higher predictivity (). Panels B and C show the PCA representation of the samples using the 3 most represented features in the model populations. The information for the best separation in both instances is present in PC2 and PC3. The equations for panel B show that a high increase in symmetry and high polarizability and low electronegativity is predictive of toxicity. In the case of the unselected features panel C toxic chemicals do cluster together but are specific to containing a nitrogen with a triple single bond and a low autocorrelation.

Figure 8. Association between PCFs and toxicity associated pathways.

The figure represents the detailed relationship between pathways associated to chemical hits and pathways associated to toxicity. Pathways with membrane component were mostly associated to chemical hits whereas pathways with downstream signalling components were mostly associated to toxicity. This figure represents three possible links between pathways associated to chemical hits (Wnt Signaling Pathway, Long-Term Depression and ErbB Signaling Pathway) and toxicity (Focal Adhesion, ALS and Pancreatic Cancer) through shared genes between the pathways. Although each link presents a mechanism of action these were only implied by the pathway associated to toxicity. Genes found to be up or down regulated have been marked with a red or a blue arrow respectively.