Bringing radiomics into a multi-omics framework for a comprehensive genotype–phenotype characterization of oncological diseases

Diseases are governed by complex biological mechanisms requiring different levels of analyses for a comprehensive interpretation of the underlying pathology. Today, the progress in genomics, transcriptomics, epigenomics and their combination, enables the incorporation of different biological layers of information to predict phenotypic conditions (tumor/normal, early/late stage, survival, etc.). Multi-omics data integration is, therefore, one of the major challenges in the era of precision medicine, particulary in oncology. With the huge increase in genomic data production, the need for specific models and methods for storing and analyzing those data has arisen; an example is MultiAssayExperiment (MAE) [1]. MAE handles multiple and heterogeneous data types for a set of samples of multi-assay genomic experiments (transcript counts, DNA variants or methylation status of genes or regions, etc.). However, these structures consider only data produced by molecular biology experiments and neglect the impact of other ‘omics which also deserve consideration. The progress that has been made in medical imaging techniques and the development of high-throughput algorithms to extract quantitative features from medical images has led to the development of radiomics. In clinical research, radiomics is becoming a meaningful tool and might be considered as an additional and complementary source of ‘omic information, not achievable in a multi-omics biological environment. In this scenario, the growing impact of non-invasive imaging techniques for disease definition, in parallel with the evolution of next-generation sequencing (NGS) tools, provides powerful methods for investigating the phenotype through the combination of imaging characteristics (radiomic features) into a multi-omics biological framework. Indeed, in recent years, correlation of radiomic features with genomic features, rise to a new field of study defined “radiogenomics” [2]. The increasing scale and availability of a high volume of health data requires new and efficient strategies for data management, data linkage and data integration. These types of datasets are defined “multimodal” [3] since multimodal signals are managed together. In this context, there are many challenges to overcome: identifying relationships between data from different modalities, joining multimodal information to execute prediction, learning information to help understand limited data of another modality and, crucial in our case, representing, integrating and summarizing multimodal data. Thus, in order to optimize data management and analysis, it is necessary to reshape the existing information systems into innovative multi-layer data systems by combining statistical and computational methods. So far, no tools integrating genomic and radiomic data have been designed; therefore, consolidating single-omic datasets from different domains in a meaningful manner is an ambitious undertaking. Here, we investigated the role of the MAE structure as a possible bridge for integrating radiomics into a multi-omics framework. To this end, we evaluated the potential of MAE as a structure for storing and managing both imaging and biological ‘omic data derived from different type of experiments, while keeping the coordinated representation of data and ensuring consistency between a single assay and clinical patient data during data subsetting and analysis intact. The extended multi-omics framework proposed here allows researchers to simplify the management of radiogenomic data. In this article, (i) we will first introduce the state of the art of both radiomics and biological ‘omics in the field of cancer research; (ii) we will then summarize the role of radiogenomics in cancer phenotype definition; (iii) we will discuss the current state of radiogenomic public repositories, their limits, challenges and limitations of including radiomics in a multi-omics framework; (iv) finally, we will demonstrate the feasibility of our approach with a case study using The Cancer Genome Atlas (TCGA), for biological data, and The Cancer Imaging Archive (TCIA), for public medical images.

Radiogenomic analyses are generally used for two main purposes: identifying features that might be related to genetic or molecular outcomes and correlating imaging and genomic data to identify suitable markers or predictors of a particular disease. Tipically, a radiogenomic dataset contains genomic (for example, gene expression) and imaging data, without outcomes data. A well-known correlation between specific imaging features and an outcome could enable the discovery of relationship between those features and specific tumor molecular characteristics.  Similarly, investigating a well-known correlation of tumor molecular characteristics related to an outcome may allow the detection of imaging features related with that outcome. Several studies, based on these approaches, have been published. Gevaert et al. [40] tested how well the imaging features, based on specific genomic characteristics, predicted patient survival in non–small-cell lung cancer using sets of imaging and genomic (gene expression) data without outcomes. Other examples are radiogenomic studies in which correlations have been detected between imaging features and tumor subtypes, especially in breast cancer and in glioblastoma multiforme. Mazurowski et al. [41] demonstrated that imaging features describing tumor enhancement dynamics can differentiate breast cancer luminal B molecular subtype from other subtypes. Therefore, an imaging feature might be predictive of outcomes and might not necessitate a genomic analysis. However, in another study, Guo et al. [42] describe that a combination of imaging and genomic features could be useful for better breast tumor characterization. Indeed, they demonstrated that imaging features such as tumor size outperformed genomic features in predicting tumor pathological stage, whereas genomic features outperformed imaging features in predicting breast cancer estrogen receptor (ER) and progesterone receptor (PR) status such as tumor molecular characteristics. Another approach was used in Karlo et al. [43], where correlations between imaging features and mutation of genes (related with stage and diminished survival prognosis) were identified. Through this evidence, imaging features, potentially predictive of outcomes, have been identified. Furthermore, in Glioblastoma Multiforme, 1p/19q co-deletion, a widely used prognostic biomarker for brain tumors, and epidermal growth factor receptor (EGFR) mutations, have been correlated with a wide array of MRI features [44, 45]. Finally, radiogenomics could potentially have an important role in targeted therapies and in improving the performance in cancer outcomes prediction. In order to identify complex phenotypes from a radiogenomics approach, a number of challenges need to be addressed. The introduction of more complex models combining multiple heterogeneous data sources could overcome many of these challenges. For further insights on state of the art of radiogenomics studies see [6, 46, 47].

As described in the previous sections, we propose the use of MultiAssayExperiment (MAE) object as data structure to integrate genomic, radiomic and clinical data, providing coordinated representation, operations on multiple and heterogeneous data and focusing on two fundamental aspects of data at stake: multisampling and data longitudinality. We tested this solution at first studying existing MAE objects of TCGA unrestricted data of different cancer tissue obtained through curatedTCGA R package [77] and then creating a new MAE based on the TCGA breast cancer data and the respective radiomic features, extracted from T1 weighted Dynamic Contrast Enhanced (DCE) MRI images of TCIA [71].

Objects from curated TCGA contain data from different ‘omic experiments carried out on the same patient. Each experiment may contain different sample types for the same patient (for example, primary solid tumor and metastatic samples). To manage these data, in the TCGA project, a barcode is used as primary identifier. The TCGA barcode consists of a sequence of values associated to labels, each of which specifically identifies a TCGA data element. For example, the “Sample” label describes the sample type of a particular collection of data related to a patient and may take different value corresponding to a sample type according to Sample Type Codes table [78]. The barcode TCGA-A1-A0SB-01A example indicates Primary Solid Tumor data (Sample Type Code: 01) of the patient A0SB belonging to TCGA project on breast invasive carcinoma (Tissue Source Site Codes: A1) (Fig. 2). This nomenclature was also used by curated TCGA in the construction of the TCGA data MAE object. In more detail, the barcode is used as value of colname column of MAE sampleMap (a DataFrame that relates the “primary” data - that describes the biological unit, which can refer to specimens, patients, etc.—to the experimental assays—for example, RNAseqGene). This DataFrame allows an unambiguous map from every experimental observation to one and only one biological unit, such as a patient, and allows different technical and biological replication for each assay. Moreover, identifiers allow consistency between data during subsetting and re-ordering. We propose to use both MAE data structure and a TCGA-barcodes-like structure to manage radiomic experiment data, together with biological omic data, in a single data structure. Typically, in a radiomic workflow, each single ROI, and the respective features, may represent a different lesion, or a region of a lesion, and, therefore, may exist in multiple samples for each patient of a radiomic experiment. According to our proposal, this multi-sampling feature can be managed by using MAE data structure and a specific barcode, in the same way as genomic data. In this scenario, a key role is played by the ExperimentList component of MAE, which contains all experimental data. This component can contain different type of elements, two of which are Matrix (a base element used for ID-Based dataset) and SummarizedExperiment [79]. The latter may contain one or more assays (a matrix-like element that store the data). For each assay, rows represent features of interest and columns represent samples. For example in a genomic experiment, the rows of an assay represent genes or transcript, the columns represent the patients and each different assay in a Copy Number Alterations experiment may represent the values of segment mean and number of probes (Fig. 3). A radiomic experiment differs substantially from a genomic experiment in that it consists of one level of data complexity less than a genomic experiment. The latter is performed on a set of samples (the columns of the assay) and the analysis is performed on a set of genes, transcripts or protein (the rows of the assay). Finally, the experiment produces different results from the various assays. A radiomic experiment, instead, is performed on a set of samples without analysing of the molecular sub-level and therefore, from data structure point of view, we have two options:

In both cases, the rows of each assay stored radiomic features.

As shown in our case study, this data organization enables the use of MAE to collect, manage and then analyze radiomic data together with genomic and clinical data. In our case study the TCIA data consist of 36 quantitative radiomic features extracted from primary tumor images of 91 patients of the BRCA study, each acquired at a single time point. TCGA data consist of a MAE object composed of several experiments. We selected RNA-seq (expression quantification of 20,502 genes from 878 samples) and miRNA-seq (expression quantification of 1046 miRNA from 849 samples) experiments and integrated it together with TCIA data in a single MAE object. To do this, we first downloaded TCGA experiments, using curated TCGA, and TCIA features released. Subsequently, we created two R objects: one for TCGA and one for TCIA data. In the case of TCGA data we have directly obtained a MAE object through curated TCGA while in the case of TCIA, we first converted an xls file in an R dataframe and then in a summarizedExperiment object. This experiment contains two assays that represent two time point data, one of which is simulated. The radiomic experiment was subsequently integrated with the two pre-existing summarizedExperiment (RNA-seq and miRNA-seq experiment extracted from the downloaded MAE) using the workflow described in MAE vignettes on Bioconductor [80]. Figure 5 shows a generalized Venn diagram for sample membership in multiple assays. The visualization of set intersections was performed using the UpSet matrix design using UpSetR package [81]. The script code to reproduce above described procedure is available at https://​gitlab.​com/​Zanfardino/​radiogenomics-mae-case-study. We also propose an architecture, shown in Fig. 6, for a modular integration platform. Through a graphic interface, the users are able to (i) create or upload a MAE object, (ii) summarize MAE data through basic statistics and plots, (iii) manage and subset the uploaded data and (iv) execute different type of analyses (through independent modules). All the functions allow to work with highly complex data in an intuitive and simplified way. One way to manage and understand the meaning of large dimensional data is to place it in a visual context such as we have done here (Fig. 7).

The high-throughput production of ‘omics data has led to an increase of data sets of different types that need to be integrated in order to better understand disease mechanisms and how these multiple molecular data generate the observed phenotypes in complex diseases. Merging imaging phenotypes with multi-omic biological data may lead to new prognostic cancer  models, new support for patient treatment strategy and development of improved survival predictors. Accordingly, increased attention is paid to statistical methods and algorithms to analyze and correlate multivariate imaging, clinical and molecular data for disease diagnosis and prognosis. Bringing these datasets together in a meaningful manner is the main goal of this study. Here, we identified three main challenges to overcome: the management of missing data caused by data filtering or non-execution of a specific analysis on a subset of samples, different spatial and temporal scales of imaging data and the need to manage radiomic features related to multiple lesions or sub-regions of a lesion. Our proposal to use MAE as data structure to combine radiogenomic data aims to integrate and facilitate the use and the exploration of heterogeneous and complex data derived from these deeply distant domains. Our integrated design enables regular operations of MAE on all experiments of a radiogenomic dataset, including radiomic data. With our MAE design, operations like: (i) selecting complete cases or subsettings, (ii) selecting samples with information in all dataset and/or in all ‘omics of a set of experiments (crucial to set a multi-omics analysis when more samples with missing data exist) and (iii) selecting subjects with specific phenotypes and clinical outcomes, may be carried out in order to ensure correct alignment of assays and patients, making radiogenomic integrative analysis more attainable. Moreover, the MAE structure, and the ability to organize data of different experiments in different interlinked data objects, has facilitated the handling of the additional spatial and temporal scales added by radiomic data. All of these operations and, therefore, our selection of the MultiAssayExperiment as the radiogenomic data container have been successfully tested with the integration of TCGA-BRCA data of 91 patients with radiomic features available on TCIA for the same patients. Despite the existence of specific software that make some of these functions available, such as TCGAbiolinks, RTCGA Toolbox and mixOmics (Table 1), none of these tools was designed to work on radiogenomic data. Conversely, our proposal facilitates radiogenomic studies since it allows user exploration across genomic as well as imaging datasets in data type independent manner. For example, the use of the TCGA barcode is clearly suitable for biological ‘omics data description but not for radiomic data, except for “Patient ID”, “Sample” and “Center” labels. The latter represents an open challenge and, therefore, a possible future direction might be the development of a unique nomenclature for a new type of barcode to specifically describe radiogenomic data. Moreover, for data different from TCGA, also a custom id can be used as colName in SampleMap. In conclusion, understanding the relationships among genomic profiles, imaging phenotypes and outcomes has great potential to improve cancer treatment and management. In this context, genomic features are closely related to genetic and molecular profile of a cancer and, consequently, to outcomes like receptor status, while, radiomic features characterize tumor phenotypes and, consequently, outcomes like tumor stage. Bringing radiomics and genomic data together into a single data structure is the first step to achieve effective radiogenomic analysis, integrating information arising from different aspects of the tumor. The use of the current version of MAE by an interdisciplinary research community can pave the way to further development of extended MAE object for implementing new functionalities specific to radiogenomic domain in order to correlate phenotype and genotype features.