Shared Nearest Neighbors Approach and Interactive Browser for Network Analysis of a Comprehensive Non–Small-Cell Lung Cancer Data Set

PURPOSE

Advances in biological measurement technologies are enabling large-scale studies of patient cohorts across multiple omics platforms. Holistic analysis of these data can generate actionable insights for translational research and necessitate new approaches for data integration and mining.

METHODS

We present a novel approach for integrating data across platforms on the basis of the shared nearest neighbors algorithm and use it to create a network of multiplatform data from the immunogenomic profiling of non–small-cell lung cancer project.

RESULTS

Benchmarking demonstrates that the shared nearest neighbors-based network approach outperforms a traditional gene-gene network in capturing established interactions while providing new ones on the basis of the interplay between measurements from different platforms. When used to examine patient characteristics of interest, our approach provided signatures associated with and new leads related to recurrence and TP53 oncogenotype.

CONCLUSION

The network developed offers an unprecedented, holistic view into immunogenomic profiling of non–small-cell lung cancer, which can be explored through the accompanying interactive browser that we built.

INTRODUCTION

Multiomics profiling is becoming increasingly widespread in patient cohort studies as the throughput and resolution of measurement platforms improve.^1-3 Fully harnessing advances in biologic measurement technologies across platforms to realize the potential of systems medicine requires concomitant progress in integrating diverse data types to enable holistic exploration and analysis.^4,5 Insights from integrated analysis across platforms can provide a more complete picture of underlying biology compared with analyzing different platforms in isolation.⁶ Particularly when applied to the study of cancer, holistic approaches to analyze multiplatform data sets can unlock insights for therapeutic targets and biomarkers by providing a fuller view of phenomena driving disease pathogenesis, evolution, and post-therapy recurrence.

CONTEXT

• Key Objective

• How can multiplatform data sets be explored in an integrated and interactive manner?

• Knowledge Generated

• We built the immunogenomic profiling of non–small cell-lung cancer (ICON) data network, which connects measurements on the basis of the interplay between transcriptomic, proteomic, and flow cytometry data; to facilitate hypothesis generation, we created an interactive visualization application for exploration of the ICON data network alongside patient metadata and measurements from multiple platforms. We demonstrate that querying the shared nearest neighbors-based network reveals leads related to relevant tumor characteristics by considering recurrence and TP53 oncogenotype as examples.

• Relevance

• Our method for multiplatform data integration and exploration can be broadly applied to forthcoming large-scale patient profiling studies to generate actionable insights for translational research. Moreover, the interactive ICON data browser made broadly available through this publication is a unique contribution to the field and can serve as a valuable resource for future cancer research.

In this study, we present a novel approach for integrating data across platforms on the basis of the shared nearest neighbors (SNN) algorithm.⁷ We constructed integrated networks of multiplatform biological data by connecting transcriptomic measurements highly correlated with the same measurements from different platforms. Transcripts whose top correlates from orthogonal platforms are similar suggest potential interaction between them and shared participation in overarching biological functions.

We applied the SNN-based network-building approach that we developed to the immunogenomic profiling of non–small-cell lung cancer (NSCLC; ICON) project, an ambitious undertaking to comprehensively characterize immunogenomic diversity in resected NSCLCs across diverse omics platforms. Performing community detection on the resulting ICON data network enabled the identification of modules, collections of closely related transcripts, and focused the exploration of the integrated network, which can be easily explored using the interactive data browser that we developed. The integrated ICON data network offers an unparalleled view into the immune landscape of NSCLC, and the approach for its generation along with its accompanying data browser can serve as templates for the analysis of forthcoming large-scale patient profiling studies to facilitate translational research.

METHODS

Tissue samples were collected from patients enrolled in the ICON project and were processed and analyzed per established protocols (Data Supplement). Given their depth, the platforms of RNA sequencing (RNA-seq),⁸ reverse phase protein array (RPPA), and flow cytometry (FC)⁹ were selected for network building via the SNN-based approach that we developed, which links transcriptomic measurements by their correlation with measurements from selected platforms, in this case RPPA and FC.

Network parameters were determined using log-likelihood scores (LLSs),^10,11 and Infomap community detection was performed on the finalized network to identify more interpretable subnetworks,¹² termed modules. Modules were annotated using established gene sets¹³ and tested for association with selected metadata features by linear modeling¹⁴ using tabulated single sample gene set enrichment (ssGSEA) scores.¹⁵ An interactive data browser of the network was constructed using Python Dash¹⁶ (see the Data Sharing Statement for more information).

For a detailed methodology, see the Data Supplement.

RESULTS

Overview of ICON and SNN-Based Network Approach

The ICON data set (Fig 1A) represents a uniquely deep and rich profiling of both the genomic and immune landscapes of NSCLC. It is composed of tumor and adjacent uninvolved lung tissue samples collected from nearly 150 patients at the time of surgical resection. Samples underwent RNA-seq,⁸ whole-exome sequencing¹⁷ (WES), T-cell receptor (TCR) sequencing,¹⁸ multiplex immunofluorescence (mIF) for immune cells,¹⁹ FC for immune cells,⁹ and RPPA profiling. Given the richness of the ICON cohort data (Fig 1B), we were motivated to establish connections between measurements from different platforms to create a comprehensive and integrated view of resected NSCLC.

FIG 1.

ICON project. (A) Schematic of data collected at the time of surgery through the ICON project. Red denotes tumor samples, and white denotes adjacent uninvolved tissue samples. (B) Heatmap of measurements available in ICON for each platform (rows) across patients (columns). Colors indicate treatment received, stage, histology, and data availability. FC, flow cytometry; ICON, immunogenomic profiling of non–small-cell lung cancer; mIF, multiplex immunofluorescence; NSCLC, non–small-cell lung cancer; RNA-Seq, RNA sequencing; RPPA, reverse phase protein array; TCR-seq, T-cell receptor sequencing; WES, whole-exome sequencing.

Given their depth, we focused on using RNA-seq, RPPA, and FC data in network building. Measurements from a selected orthogonal platform (eg, RPPA or FC) were ranked by their correlation with a given gene in the RNA-seq data, and for each gene pair, the number of shared top k correlates from the orthogonal platform determined the overlap score of that gene pair (Fig 2A). This process was performed across all gene pairs to compile a SNN matrix for a given orthogonal platform (Fig 2B). From these matrices, gene pairs whose overlap scores met the thresholds selected through LLS benchmarking were included in the network, where each gene is a node and each pair meeting the threshold forms an edge (Fig 2C).

FIG 2.

SNN drives integrated ICON data network. (A) Overlap score is determined by the fraction of shared top k correlates from RPPA, FC. (B) SNN matrix is compiled by performing scoring across all gene pairs for RPPA, FC. (C) Gene pairs meeting score threshold are used to construct the overall network. FC, flow cytometry; ICON, immunogenomic profiling of non–small-cell lung cancer; RPPA, reverse phase protein array; SNN, shared nearest neighbors.

Network Parameter Selection and Benchmarking

We used LLS,^10,11 a metric capturing recovery of validated gene pairs, to perform a parameter sweep to select k and the overlap score threshold (Data Supplement). First, we collected annotated gene pairs from a reported set of GO biologic process terms²⁰ and examined LLS of different k and overlap score threshold combinations to select the finalized SNN-based network configuration. Next, for comparison, we used the same annotated gene collection to calculate the LLS of a similarly sized network based solely on correlations between measurements from RNA-seq (Fig 3A); we found that the SNN-based network had a higher LLS than the RNA correlation network, 7.22 versus 7.11, respectively. This increase in the benchmarking metric is notable as it is on the logarithmic scale; in addition, the validated gene pairs used in its calculation are largely derived from expression-based experiments,²⁰ representing a potential tendency away from relationships observed through other measurement platforms. Taken together, these points underscore the strength of the SNN-based approach that we developed.

FIG 3.

ICON data network combines SNN with RNA-based correlation edges. (A) Selection of SNN with RNA-based edges on the basis of improved performance over SNN, RNA alone. (B) Community detection via Infomap identifies 10 top-level modules and 93 mid-level modules with at least 10 genes. The modules can be examined in tandem with orthogonal data modalities and patient characteristics. (C) Overview of the ICON data network composition. For each top-level module, the most significant annotation (with specific database source in parentheses where applicable) of its mid-level modules is shown. Modules in gray are without annotations meeting FDR < 0.05. (D) Nodes balanced across source modalities. Overlaps in gene nodes from each modality shown along with representative summary annotations. (E) SNN modalities provide more edges per node. Overlaps in network edges from each modality shown. ECM, extracellular matrix; FC, flow cytometry; FDR, false discovery rate; ICON, immunogenomic profiling of non–small-cell lung cancer; KEGG, Kyoto Encyclopedia of Genes and Genomes; PID, Pathway Interaction Database; RPPA, reverse phase protein array; SNN, shared nearest neighbors.

Finally, we considered a network combining gene pairs from the SNN-based approach with additional gene pairs from the RNA correlation network and found that this network outperformed both the RNA correlation network and the SNN-based network with a LLS of 7.29 (Fig 3A). We used this blended strategy to build the ICON data network, which integrates orthogonal measurements from FC, RPPA, and RNA-seq.

ICON Data Network Connects More Than 10,000 Measurements

The ICON data network is composed of 10,732 transcriptomic measurements connected by 195,560 edges, offering a comprehensive view of the ICON data. Given its size, we used Infomap to perform community detection and extract more interpretable subnetworks,¹² termed modules. The ICON data network can accordingly be framed as a multilayered network of modules (Fig 3B). We restricted our consideration to modules containing at least 10 genes, leading to 10 top-level modules and 93 mid-level modules within them.

In developing our network-building approach, we connected transcriptomic measurements to each other so that in mining the resulting network, we could leverage pathway databases^13,21 to annotate each module's genes. In this manner, we identified annotations with a false discovery rate (FDR) < 0.05 for seven of 10 top-level modules and 23 of 93 mid-level modules (Fig 3C). Transcriptomic measurements were not restricted to protein-coding genes, which can partially explain modules without annotations. Concomitantly, that the network contains modules with and without annotations underscores the recapitulation of known gene relationships along with the identification of new ones through our approach.

Examining the ICON data network by platforms from which nodes and edges were derived, we found that each source platform, that is, FC (SNN approach), RPPA (SNN approach), or RNA-seq (RNA correlation approach), contributed similar numbers of nodes (Fig 3D). We annotated the collections of nodes shared between at least two source platforms and found that each platform captures different biologic phenomena, underscoring the new, more complete perspective gained by integrating across platforms. We noted that more edges were derived from the SNN-based approach developed (Fig 3E), highlighting the richness of information provided.

Network Mining Facilitated by ICON Data Browser

We explored the ICON data network alongside patient metadata and data from additional orthogonal platforms available from ICON. To do so, for each identified module, we tabulated ssGSEA scores,¹⁵ which provided a metric of upregulation or downregulation of a module's constituent genes. We then used these scores across patients to contrast modules with measurements from orthogonal platforms and to test modules via multivariate analysis for association with features of interest. To facilitate these operations, we built an interactive Python Dash application¹⁶ for data visualization and analysis. This application contains two complementary tools for mining the ICON data network: the first permits browsing of the network, whereas the second enables plotting and modeling of selected measurements, module scores, and patient metadata.

In browsing the ICON data network, the user may examine the relationship between modules and search for specific transcriptomic measurements within the network (Fig 4A). In addition, for a selected module, the user can review its pathway annotation results and determine which underlying measurements from SNN-derived edges drive the module's formation. A selected module's ssGSEA scores or any other measurement across the full ICON data set can be analyzed with the plotting and modeling tool (Fig 4B). In it, the user can plot a selected measurement against individual or combinations of categorical variables from patient metadata. These categorical variables are then used to determine via linear modeling how well they predict the measurement's distribution and to identify other measurements that they predict well. The user can also select a secondary measurement, plot the relationship between the pair of selected measurements, and determine the Spearman correlation between them. With this tool, the user can additionally rank secondary measurements by their correlation with the primary selected measurement. Taken together, the interactive data browser that we developed enables facile exploration of the ICON data network to pave the way for new insights, which we demonstrate through the following vignettes about recurrence and TP53 oncogenotype.

FIG 4.

ICON data browser enables exploration of ICON data and network. (A) The network panel offers a view into the multilayered structures of constituent modules, pathway annotation results, and top correlates from the shared nearest neighbors-based network-building approach tied to module formation. (B) The plotting and modeling panel allows for a selected module's ssGSEA scores or any other measurement across the full ICON data set to be compared against categorical metadata or against a secondary continuous measurement. FDR, false discovery rate; ICON, immunogenomic profiling of non–small-cell lung cancer; KEGG, Kyoto Encyclopedia of Genes and Genomes; N/A, not available; PID, Pathway Interaction Database; RPPA, reverse phase protein array; ssGSEA, single sample gene set enrichment; TID, tumor identifer.

Recurrence-Focused Exploration Highlights Metabolic Module

Despite improvements in treatments, approximately half of patients with resectable NSCLC will relapse after curative-intent treatment,²² motivating our inquiry in this area. We leveraged the ICON data network to delve into disease recurrence, focusing on patients who had not received neoadjuvant therapy (Data Supplement). In considering associations between patient characteristics and disease relapse, we found no significant differences in recurrence status of patients who had undergone upfront surgical resection except for the variable of pathologic stage (P = 2.2e^–02; Data Supplement).

We tested module signatures in this cohort and controlled for stage and histology. This led to the identification of several modules linked to disease recurrence, and we considered in greater detail the most closely associated one (P = 2.14e^–06) with pathway annotations meeting FDR < 0.05. The signature score of this module tends to be lower in patients who ultimately relapse (Fig 5A), particularly in patients with stage II-III nonsquamous NSCLC (Fig 5B), suggesting its potential impact in prognosis. We used pathway annotation to further examine this module and found that the genes within it strongly associate with several metabolic pathways (Fig 5C), underscoring the mechanistic insights provided by our approach and highlighting areas for future exploration.

FIG 5.

Network mining identifies module strongly tied to patient recurrence. (A) Distribution of ssGSEA scores by patient response for the neoadjuvant-free cohort. (B) Distribution of ssGSEA scores by patient response for a subset of cohort with nonsquamous, stage II+ tumors. (C) Top pathways for genes driving annotation of the selected module with the false discovery rate < 0.25. (D) Correlation of module scores with the Morisita overlap index from T-cell receptor sequencing (ϱ = 0.53, P = 1.98e^–05). ssGSEA, single sample gene set enrichment.

We also considered which measurements from orthogonal platforms not included in the construction of the ICON data network correlated strongly with the distribution of ssGSEA scores from the highlighted module. We found that the correlation of the Morisita Overlap Index (MOI) from TCR sequencing was pronounced (ρ = 0.53, P = 1.98e^–05; Fig 5D). As MOI describes the homology between TCRs in the tumor and uninvolved lung,²³ we observe that patients who recur tend to have lower MOI, which indicates less similarity between TCRs in the tumor and uninvolved lung. This could suggest that the immune system of patients who eventually relapse do not have widespread presence of a TCR that recognizes the tumor, thus impeding an effective immune response.²⁴

TP53 Oncogenotype in Squamous NSCLC Tied to Transfer RNA

The ICON data network can also be mined from the perspective of patient oncogenotype by incorporating WES data relevant to a query of interest. Given that it is often mutated in NSCLC, we turned our attention to TP53.^25,26 TP53 mutations have been reported to associate with therapeutic resistance,²⁷ but further study can help better elucidate the potential mechanisms of action responsible for this association. We focused our inquiry on patients with lung squamous cell carcinoma (LUSC), which although comprising 30% of all new lung cancer diagnoses,^28,29 has historically been less studied than lung adenocarcinoma.³⁰

We used WES data to annotate patients with LUSC in ICON by whether they had a functionally relevant mutation in TP53 as determined by the annotation of those mutations in the International Agency for Research on Cancer TP53 database.³¹ We then performed linear modeling to identify modules associated with the presence or absence of a functionally relevant TP53 mutation in patients with LUSC; we included pathologic stage, whether the patient had received neoadjuvant chemotherapy, and smoking status as covariates. Furthermore, we confirmed that no comutations in KRAS or EGFR were observed in the population of interest, as expected.³⁰

Through this approach, we identified a module linked to TP53 mutation status in patients with LUSC (P = 4.72e^–02), and its ssGSEA scores tend to be higher in patients without functionally relevant TP53 mutations (Fig 6A). We performed pathway annotation on the transcriptomic measurements contained within the module; however, we did not find any associated pathways meeting FDR < 0.05. Next, we compared which measurements from orthogonal platforms not included in constructing the ICON data network correlated strongly with the distribution of ssGSEA scores from the highlighted module. We found that the correlation of malignant cell presence in the tumor and stroma (count/mm²) as measured by mIF was robust (ρ = 0.55, P = 6.39e^–03; Fig 6B).

FIG 6.

Network mining identifies the transfer RNA–focused module strongly associated with functionally significant TP53 mutations in LUSC. (A) Distribution of ssGSEA scores by TP53 status for LUSC tumors. (B) Correlation of module scores with the presence of malignant cells (count/mm²) in the tumor and stroma via multiplex immunofluorescence (ϱ = 0.55, P = 6.39e^–03). (C) Components of the identified module highlight strength of interaction with transfer RNA transcripts. LUSC, lung squamous cell carcinoma; ssGSEA, single sample gene set enrichment.

To delve into this module, we examined its constituents (Fig 6C). We found that the module was primarily composed of transfer RNA transcripts, and most nodes were connected by edges derived from RNA-RNA correlations. Taken together, the generally lower module ssGSEA scores exhibited by TP53-mutant tumors indicate diminished transcription of transfer RNA, which suggests translational dysregulation compared with TP53 wild-type tumors. Moreover, the correlation between the module's ssGSEA scores and the presence of malignant cells attributes generally lower malignant cell presence in the tumors of patients with functionally relevant TP53 mutations. This may be explained as reflecting the dysregulation in tumors with TP53 mutations driving a more diverse tumor microenvironment in which additional cell types have been recruited.^32-35

DISCUSSION

We report a SNN-based network approach for integrating multiplatform data and its application to the ICON data set to create a unique resource for the study of resected NSCLC. By harnessing similarities in top correlates between measurements, our approach offers a flexible and generalizable method for holistically analyzing sample data from diverse platforms, thus providing a more complete perspective on underlying biology than possible by considering data from different platforms separately. We show that the network we developed contains established connections between genes in addition to new ones, underscoring the power of the integrative approach presented. Our approach can be used to comprehensively examine multiplatform cohort data sets to better understand diverse patient attributes and explore the interaction between biological phenomena at play. Furthermore, the ICON data browser enables future researchers to explore unique questions with relative ease by making the ICON network accessible in an interactive manner.

We demonstrate that our approach enables the facile identification of collections of measurements that distinguish recurrence and TP53 mutation status in ICON; moreover, these measurements can be further contextualized within the overall ICON data set to learn more about mechanisms linking them. Through our network browser, we show that networks like ours can be explored by pathways associated with collections of measurements, enabling the examination of interactions between pathways. Thus, SNN-based networks can be probed in both a top-down approach, that is, by querying modules, and a bottom-up approach, that is, by querying individual measurements.

Our approach builds upon rich work to date for constructing biologic networks. We drew inspiration from the study by Price et al,³⁶ who directly linked highly correlated measurements from different platforms to create personal, dense, and dynamic data clouds. However, we noted that RNA-seq coexpression networks³⁷ lend themselves well to annotation using established expression-focused resources and approaches like gene set enrichment analysis.¹³ To achieve multiplatform integration while being able to use existing annotation tools, we developed our SNN-based approach to create a network of transcriptomic measurements informed by measurements from complementary orthogonal platforms (FC, RPPA, and RNA-seq), thus providing a more complete perspective into mechanisms associated with tumor biology.

Our work also fits into a growing collection of interactive interfaces for multiplatform data sets; we highlight one such complementary project, the PanCancer Immunogenomics (PCIG) tool.³⁸ PCIG provides a web platform for mining the Pan-Cancer Analysis of Whole-Genomes study's clinical, genomic, and transcriptomic data from 2,658 samples across 40 indications.³⁹ The PCIG tool presents immune signatures calculated from transcriptomic data to offer perspective on tumor immune phenotype, which can be viewed alongside tabulated copy number, structural variant, and mutation load profiles for a selected indication. Although our work focuses on a smaller cohort of patients from a single broad indication, that is, NSCLC, we provide direct immunophenotype measurements via FC and mIF along with measurements from other platforms for interactive visualization within the ICON data browser; moreover, these data can be viewed in the context of the integrated SNN-based network that we developed. Both the PCIG tool and our project examine cancer from an immune perspective but differ with respect to cohort sizes and data types for exploration. Together, the growing availability of integrated and interactive projects such as these will spur future advances in cancer research.

Our approach is not without limitations. As with other network-building approaches, parameter selection can be somewhat subjective. We attempt to mitigate this by performing benchmarking using LLS and acknowledge that the resulting network can vary on the basis of the combination of k and overlap score cutoffs selected. In mining our network, we recognize that false discoveries can persist despite stringent multiple hypothesis correction. We anticipate that leads from SNN-based networks will require validation through analysis of complementary data sets and experiments in preclinical models.

Taken together, the SNN-based network approach that we developed enables integration of multiplatform data sets, which can help realize the promise of large-scale patient cohort studies across cancer types. Our application of this method to ICON allowed us to a create a rich and deep integrated network that can easily be mined around patient characteristics of interest. We demonstrate that mining the network identifies modules associated with phenomena such as recurrence and TP53 oncogenotype, which provide insights toward a prognostic biomarker and better understanding of disease biology. Ultimately, our SNN network–building approach and accompanying ICON data browser enable integration and exploration of patient data from diverse omics platforms, providing a more complete view of phenomena driving disease and thus unlocking insights for therapeutic targets, biomarkers, and treatment plans.

DATA SHARING STATEMENT

A data sharing statement provided by the authors is available with this article at DOI https://doi.org/10.1200/CCI.22.00040.

AUTHOR CONTRIBUTIONS

Conception and design: Stephanie T. Schmidt, Neal Akhave, Natalie Vokes, Andrew Futreal, Ara Vaporciyan, Tina Cascone, Jianjun Zhang, Christopher A. Bristow, Timothy P. Heffernan, Marcelo V. Negrao, Don L. Gibbons

Financial support: Timothy P. Heffernan

Administrative support: Beatriz Sanchez-Espiridion, Ignacio I. Wistuba, Timothy P. Heffernan

Provision of study materials or patients: Jun Li, Lauren A. Byers, Beatriz Sanchez-Espiridion, Ignacio I. Wistuba, Boris Sepesi, John V. Heymach, Marcelo V. Negrao

Collection and assembly of data: Stephanie T. Schmidt, Neal Akhave, Ryan E. Knightly, Alexandre Reuben, Natalie Vokes, Jianhua Zhang, Junya Fujimoto, Lauren A. Byers, Beatriz Sanchez-Espiridion, Lorenzo Federico, Annikka Weissferdt, Younghee Lee, Carmen Behrens, Ignacio I. Wistuba, Andrew Futreal, Boris Sepesi, John V. Heymach, Chantale Bernatchez, Cara Haymaker, Jianjun Zhang, Christopher A. Bristow, Marcelo V. Negrao, Don L. Gibbons

Data analysis and interpretation: Stephanie T. Schmidt, Neal Akhave, Alexandre Reuben, Natalie Vokes, Jun Li, Lauren A. Byers, Lixia Diao, Jing Wang, Marie-Andree Forget, Daniel J. McGrail, Shiaw-Yih Lin, Erika Suzuki, Jeffrey J. Kovacs, Ignacio I. Wistuba, Andrew Futreal, Boris Sepesi, Cara Haymaker, Tina Cascone, Jianjun Zhang, Christopher A. Bristow, Marcelo V. Negrao, Don L. Gibbons

Manuscript writing: All authors

Final approval of manuscript: All authors

Accountable for all aspects of the work: All authors

AUTHORS' DISCLOSURES OF POTENTIAL CONFLICTS OF INTEREST

The following represents disclosure information provided by authors of this manuscript. All relationships are considered compensated unless otherwise noted. Relationships are self-held unless noted. I = Immediate Family Member, Inst = My Institution. Relationships may not relate to the subject matter of this manuscript. For more information about ASCO's conflict of interest policy, please refer to www.asco.org/rwc or ascopubs.org/cci/author-center.

Open Payments is a public database containing information reported by companies about payments made to US-licensed physicians (Open Payments).