Prognostic importance of splicing-triggered aberrations of protein complex interfaces in cancer

Abstract

Aberrant alternative splicing (AS) is a prominent hallmark of cancer. AS can perturb protein-protein interactions (PPIs) by adding or removing interface regions encoded by individual exons. Identifying prognostic exon-exon interactions (EEIs) from PPI interfaces can help discover AS-affected cancer-driving PPIs that can serve as potential drug targets. Here, we assessed the prognostic significance of EEIs across 15 cancer types by integrating RNA-seq data with three-dimensional (3D) structures of protein complexes. By analyzing the resulting EEI network we identified patient-specific perturbed EEIs (i.e., EEIs present in healthy samples but absent from the paired cancer samples or vice versa) that were significantly associated with survival. We provide the first evidence that EEIs can be used as prognostic biomarkers for cancer patient survival. Our findings provide mechanistic insights into AS-affected PPI interfaces. Given the ongoing expansion of available RNA-seq data and the number of 3D structurally-resolved (or confidently predicted) protein complexes, our computational framework will help accelerate the discovery of clinically important cancer-promoting AS events.

Figure 1.

Overview of the study. Transcriptomics and clinical data from The Cancer Genome Atlas were integrated with structural protein complex data from the Protein Data Bank to identify cancer-relevant perturbed exon–exon interactions (EEIs), which were then used to derive cancer type-related EEI signatures and prognostic biomarkers.

Figure 2.

Details of the global EEINs at different confidence levels used in this study. The top panel outlines the overlap of the three global EEINs corresponding to the three EEI definition approaches. Given an area of the Venn diagram, the two numbers represent the numbers of edges and the percentage of edges over all edges (26 632) in the union of the three networks, respectively. The bottom panel shows the details of the network sizes of the final three global EEINs with different confidence and coverage. See Supplementary Figure S1 to get an overview of the distribution of NETHIGH EEIs across the considered PPIs.

Figure 3.

Partitioning of edges into edge categories for a given patient or cancer type, illustrated on a dummy set of eight edges. For each patient, three edge categories were defined: (A) gained edges, denoted ‘01’, which are not present in the healthy sample but are present in the paired cancer sample (B) lost edges, denoted ‘10’, which are present in the healthy sample but are absent from the paired cancer sample and (C) non-perturbed edges, denoted either ‘00’ or ‘11’, which are either absent from both healthy or cancer samples or present in both healthy or cancer samples. For each cancer type, we grouped the edges into six categories (last column in the table of this figure).

Figure 4.

Dependence of the number of edges on the choice of expression value threshold (results shown for NETHIGH; results for NETMEDIUM and NETLOW were qualitatively similar to NETHIGH and can be found in Supplementary Figures S2 and S3, respectively). Results were qualitatively similar to an alternative choice of contact-based EEI definition and survival p-value cutoff (Supplementary Figures S4 and S5, respectively). Each point in the figure represents the mean number of edges per patient averaged over all patients across all 15 cancer types, and the vertical bars over the points represent the corresponding standard deviations.

Figure 5.

Cancer-relevant perturbed edges (CRPEs). (A) Counts of CRPEs associated with each cancer type. (B) Percentage of patients in which CRPEs were perturbed, with the median number of patients indicated on top of each box. (C) Correlation between the number of CRPEs and the number of genes significantly associated with patient survival according to Smith and Sheltzer (8). (D) Overlap between the genes from which CRPEs were derived and the genes significantly associated with patient survival according to Smith and Sheltzer (8). Number of genes from which the CRPEs were derived are shown within brackets along the x axis and the q-value of the overlap are shown on top of the bars. Results shown are for NETHIGH; results for NETMEDIUM and NETLOW are qualitatively similar (Supplementary Figures S7 and S8).

Figure 6.

Cancer-relevant perturbed edges and alternative splicing (AS). (A) Correlation between the number of CRPEs and the number of AS events significantly associated with the overall patient survival. Data on AS events significantly associated with overall patient survival across all cancers except PRAD and THCA were obtained from the Supplementary Table S1 of Zhang et al. (34). (B) Number of CRPEs belonging to partially perturbed protein complexes (Supplementary Table S2). (C) Percentage of patients in which the CRPEs belonging to partially perturbed protein complexes were found, with the median number of patients indicated on top of each box. Results shown are for NETHIGH; results for NETMEDIUM and NETLOW are qualitatively similar (Supplementary Figures S9 and S10).

Figure 7.

Unique and shared CRPEs across cancer types. (A) Counts of unique CRPEs in each cancer type. (B) Counts of CRPEs shared between cancer types. Given two cancer types, p-values of the overlap between their CRPEs with respect to the background set of edges in a global EEIN (here, NETHIGH) were calculated using the hypergeometric test and corrected using the Benjamini-Hochberg procedure to obtain the corresponding q-values. q-values ≤0.05 were considered significant. Numbers in each cell denote the counts of overlapping CRPEs between the corresponding cancer types, while the color indicates whether the overlap was significantly greater (red color) or smaller (green color) than expected by chance. For example, in the bottom-left green color cell, 32 CRPEs were shared between 495 CRPEs of BLCA and 365 CRPEs of BRCA, which was significantly less than would be expected by chance. Results shown are for NETHIGH; results for NETMEDIUM and NETLOW are qualitatively similar (Supplementary Figures S11 and S12).

Figure 8.

Unique and shared significantly enriched KEGG pathways across cancer types. (A) Number of uniquely enriched KEGG pathways in each cancer type. (B) Numbers of enriched KEGG pathways shared between cancer types. Given two cancer types, q-values of the overlap between their enriched KEGG pathways with respect to the total number of annotated KEGG pathways in a global EEIN (here, 285 KEGG pathways in NETHIGH) were computed, as for the overlapping CRPEs in Figure 7. Numbers in each cell denote the number of overlapping enriched KEGG pathways, while the color indicates whether the overlap was significantly higher (red color) than expected by chance. Results shown are for NETHIGH; results for NETMEDIUM and NETLOW are shown in Supplementary Figures S14 and S15.

Figure 9.

Distribution of cancer-relevant perturbed edge categories across cancer types. Results shown are for NETHIGH; results for NETMEDIUM and NETLOW are qualitatively similar (Supplementary Figures S17 and S18).

Figure 10.

BioCRPEs across cancer types. (A) Number of BioCRPEs in each cancer type. (B) Pairwise overlaps between cancer types in terms of BioCRPEs. Given two cancer types, q-values of the overlap between their BioCRPEs were computed as for the overlaps between CRPEs in Figure 7. Results shown are for NETHIGH; results for NETMEDIUM and NETLOW are qualitatively similar (Supplementary Figures S19 and S20). Results were qualitatively similar to an alternative choice of contact-based EEI definition and survival p-value cutoff (Supplementary Figures S21 and S22, respectively). (C) Kaplan–Meier curve for a KIRC BioCRPE with the lowest log-rank p-value (displayed in the top-right area). Ensembl IDs of the participating exons are shown above the figure. (D) Kaplan–Meier curve for a KIRP BioCRPE with the lowest log-rank p-value.

Figure 11.

Number of novel BioCRPEs. For each cancer type, given the number of BioCRPEs (written on top of the corresponding bar), the percentage of known or novel BioCRPEs are outlined in different shades of orange.

Figure 12.

An example of two novel BioCRPEs related to BRCA. (A) Kaplan–Meier curves corresponding to the two BioCRPEs. (B) The left portion of the figure shows the 3D structure of the part of the GINS complex containing subunits 2 (Q9Y248) and 4 (Q9BRT9). The protein Q9Y248 is shown in blue color with the associated exon ‘ENSE00001303150’ participating in the two BioCRPEs in cyan color. The protein Q9BRT9 is shown in gray color with the exons ‘ENSE00003554215’ and ‘ENSE00003585396’ participating in the two BioCRPEs in yellow and pink colors, respectively. The right portion of the figure highlights inter-protein interactions (within 4Å, shown as yellow dashed lines) between the exons participating in the two BioCRPEs.