Melanoma 2.0. Skin cancer as a paradigm for emerging diagnostic technologies, computational modelling and artificial intelligence

Abstract

We live in an unprecedented time in oncology. We have accumulated samples and cases in cohorts larger and more complex than ever before. New technologies are available for quantifying solid or liquid samples at the molecular level. At the same time, we are now equipped with the computational power necessary to handle this enormous amount of quantitative data. Computational models are widely used helping us to substantiate and interpret data. Under the label of systems and precision medicine, we are putting all these developments together to improve and personalize the therapy of cancer. In this review, we use melanoma as a paradigm to present the successful application of these technologies but also to discuss possible future developments in patient care linked to them. Melanoma is a paradigmatic case for disruptive improvements in therapies, with a considerable number of metastatic melanoma patients benefiting from novel therapies. Nevertheless, a large proportion of patients does not respond to therapy or suffers from adverse events. Melanoma is an ideal case study to deploy advanced technologies not only due to the medical need but also to some intrinsic features of melanoma as a disease and the skin as an organ. From the perspective of data acquisition, the skin is the ideal organ due to its accessibility and suitability for many kinds of advanced imaging techniques. We put special emphasis on the necessity of computational strategies to integrate multiple sources of quantitative data describing the tumour at different scales and levels.

systems and network biology, machine learning, tumour imaging, tumour sequencing, liquid biopsies, pharmacogenomics

Melanoma. a story of success, but still with great potential for therapeutic improvement

Skin cancer is the most common malignancy in many regions of the world with Caucasian populations, including North America, Europe, and Australia. Melanoma is a potentially lethal tumour that originates from the accumulation of mutations and deregulated pathways in UV-exposed melanocytes, the pigment-producing cells in the skin. Melanoma accounts for approximately 1% of skin cancer but it originates up to 60% of deaths from cutaneous malignancies [1]. When detected and resected in due time, melanoma has an excellent prognosis for the vast majority of the patients with 5-year survival rates of over 95% [2]. However, when melanoma becomes more invasive, the chances decrease rapidly with an overall survival rate of 20–50% after 5 years [3]. For decades, the therapeutic options to treat metastatic melanoma were scarce because no chemotherapy could show an improvement in overall survival.

However, in 2011, this scenario changed drastically after the approval for metastatic melanoma of ipilimumab, an antibody targeting CTLA-4 and the subsequent approval of nivolumab and pembrolizumab, antibodies targeting the PD-1/PD-L1 axis. These therapies are known as immune-checkpoint inhibitors (ICI) because they target central tumour immune evasion mechanisms. Melanoma is susceptible to therapies boosting the immune system due to molecular features such as a high mutational burden and immune cell infiltrates, and thus an ideal tumour entity for testing immunotherapy. ICI have shown robust clinical effectiveness in patients with metastatic melanoma with unprecedented 6-year overall survival rates of over 40% for nivolumab and 50% for the combination of ipilimumab with nivolumab [4]. Melanoma thus serves as a paradigmatic model for the development of new immunotherapies in solid tumours.

However, approximately 40–50% of patients do not respond to ICI due to primary resistance and more follow due to acquired resistance. Furthermore, many patients show severe immune-related adverse events (irAE), including autoimmune colitis, hepatitis or endocrinopathies. Severe irAEs occur in 10–20% of melanoma patients treated with anti-PD-1 antibodies and in 59% of patients treated with the combination of nivolumab and ipilimumab [5, 6].

On the other hand, the pathophysiology of melanoma makes it a perfect case study to explore advanced technologies for diagnostics. The skin is an ideal organ for diagnostic approaches due to its accessibility for obtaining samples from primary tumours and loco-regional metastases. Furthermore, skin tumours are especially suited for any type of imaging technique. An interesting point is that melanoma encompasses a variety of different tumour types, all of them deriving from abnormal melanocytes, but located in different anatomic sites with distinct environments. These subtypes of melanoma have distinct genetic backgrounds, driver oncogenes and deregulated pathways, making them differently susceptible to targeted and immunotherapies [7].

The strong immunogenicity makes melanoma an ideal testbed for new immunological therapies [8]. To discover new melanoma drugs, we need to gain a better understanding of the interplay between melanoma cells, immune cells and their niche both at the molecular and tissue levels. This requires advanced technologies for molecular profiling (e.g. RNA and DNA sequencing), spatial mapping and imaging (e.g. multiplex immunohistochemistry imaging and radiomics, Figure 1).

Figure 1

Emerging technologies for deep profiling of melanoma patients. Some of the technologies are exclusive for melanoma and skin cancer, such as 3D body photography. Others can be employed in other solid tumours, such as mIHC or sequencing of tumour samples. In blue boxes, we listed potential applications for each of the technologies in the context of patient classification and therapy assessment. Here, we include the technologies discussed in the main text, but profiling technologies such as metabolomics, proteomics and epigenomics may be also relevant. MIHC: multiplexed immuno histochemistry; EV: extracellular vesicles; ctDNA or ctRNA: circulating tumor DNA or RNA.

These technologies generate large and complex quantitative data, which require computational algorithms for the analysis and detection of imaging and molecular signatures. A manual analysis of this volume of data in clinical practice is virtually impossible and requires hardware and software tools to efficiently process and aggregate the data. For example, the trained eye of a dermatologist could scan dermoscopic images of a patient for changes and new lesions and compare it with historical images of the patient. However, the time required would exceed the limit the clinician can afford, even more so in the case of total body photography (TBP). Moreover, the data are structurally complex and comes with multi-parameter interconnections that the human brain cannot detect without the help of computational models. This is the case of multiplex immunohistochemistry imaging of skin lesion sections, with dozens of complex imaging data layers, which one would have to scan, aggregate and interpret spatially to get connections between biomarkers, cell types and their co-localization. Genomics and transcriptomics data present similar challenges to identify patterns of co-regulation in datasets composed of thousands of genes quantified from numerous patient samples. Furthermore, tumour progression occurs in multiple, interconnected spatial and time scales, and for each scale, one can obtain different types of clinical, histological and molecular data. Here, the only chance to integrate and establish connections that span time and space is computational modelling.

There is a diversity of computational models. The most common ones are algorithms used to aggregate, compare and visualize datasets, thereby facilitating the human interpretation of biomedical data such as dermoscopic and gene expression data [9]. But we also have more advanced computational models, which we here classify into three categories: machine/deep learning (DL), network analysis and mechanistic models.

Figure 2

DL models in dermoscopic images and melanoma. Large amounts of dermoscopic images (>10.000) are collected, processed and clinically annotated. The images are labelled, which means that each individual image is identified by a panel of dermatologists as representative of a given skin lesion. This way, one can generate a data set composed of images that are grouped in terms of the skin lesion they represent. The data are split in model training and model validation data sets. The data are split into model training and model validation data sets. The model training data set is utilized to train the DL such as CNNs for the classification of the dermoscopic images into skin lesion categories. Specifically, a CNN takes the clinical images as input data in the form of a matrix. Next, it applies dot multiplication to the input data with a matrix filter (i.e. a small matrix) and generates a convolution layer that is a reduced matrix of the input matrix. The convolution layer is further reduced to a pooling layer using the max pooling or average pooling method. Such methods also use matrix filter that moves over the convolution layer to obtain the maximum or the average value of the matrix. The values obtained are the reduced matrix of the convolution layer. Finally, the pooling layer is used as input for a fully connected neural network (consisting of hidden and output layers) to perform classification tasks. In essence, the DLs identify features in the dermoscopic images such as lesion size, shape, zonification or colour and utilize this information to discriminate different types of lesions. The performance of the model is quantified utilizing the validation data set. Once trained and validated, one can take a dermoscopic image from a new patient, process the data and input it into the DL model. The model will make a prediction in terms of the classification of the photographed lesion into one of the groups defined in the training data, for example melanoma, Merkel cell carcinoma or cutaneous lymphoma. Also, this type of imaging data and model can be utilized to make a survey of the lesion evolution or compared them with other patient’s lesions [10]. The main limitation of DL models is that they require very large amounts of annotated clinical image datasets, which are not available for many potential applications. These datasets have to be properly balanced to avoid gender or ethnic-biased predictions. Furthermore, DL are ‘black-box’ models; this means, it remains difficult to extract from them a mechanistic understanding of the combination of clinical and image features utilized to make the predictions.

Figure 3

ML models in melanoma. Data generated by new sample profiling technologies require specific processing. High throughput data such as RNA or DNA-Seq, composed of hundreds to thousands of measured features per sample, benefit from computational methods allowing their aggregation, the selection of patient group relevant features and the visualization. For example, PCA can be employed to aggregate the multidimensional data into a 2D figure, in which colour-coding of the samples helps visualize differences between groups. Data from large patient groups, labelled into relevant groups, can be used to train and validate supervised ML models. ML models are utilized for classification of patients in terms of the tumour imaging or molecular makeup. Once trained, one can profile a sample from a new patient, process the data and input it into a trained ML model. The ML will make a prediction of the patient in terms of its classification into one of the groups defined in the training data, for example elevated or low risk of tumour relapse. One should keep in mind that despite their potentially high predictive power, ML models are ‘black-box’ models; this means, it remains difficult to extract from them a mechanistic understanding of the gene and cell phenotype signatures they detect. However, in the last years, there have been efforts to circumvent this issue with computational methods aiming at the interpretability of ML models, including the SHAP (Shapley additive explanations) [11] and LIME approaches [12]. PCA: principal component analysis.

Deep and machine learning models

DL combines cascades of algorithms in layers to generate structures with the shape of artificial neural networks (Figure 2). DL models can learn and deliver intelligent decisions on their own and enable the processing of unstructured data such as documents, images and text. Machine learning (ML) accounts for a large family of computational algorithms used to parse data, learn from that data and perform informed decisions based on what it has learned (Figure 3). The main advantage of ML and DL models is that they can be constructed in a semi-automatic manner, trained with large amounts of imaging or omics datasets and used for classification of patients or assessment of therapies. Such models show high levels of predictability with respect to supervised learning tasks that are quantified by metrics such as sensitivity and specificity.

Gene network analysis

Genes, proteins and cells underlying tumour progression and therapy resistance compose complex networks of interactions, enriched by reoccurring, interlinked gene circuits (Figure 4). The task of interpreting data in terms of the (de)regulation of gene networks is only possible when equipped with algorithms for network-based data visualization, classification and analysis [13]. Gene network analysis, when combined with omics data from patient cohorts, is useful to obtain signatures of interconnected genes and proteins, for example underlying therapy resistance.

Mechanistic modelling

These models generate a mechanistic understanding of the gene circuits and the cell processes underlying progression or therapy efficacy (Figure 5). Once built using existing knowledge and characterized with quantitative data, these models can function as a virtual laboratory, where new hypotheses on biological mechanisms or therapeutic strategies can be tested. Therefore, they support the design of wet-lab experiments and clinical studies, potentially saving time and money through in silico experiments [15].

Figure 4

Molecular network analysis in melanoma. The -omics data (genomics, transcriptomics or proteomics) from patient samples can be investigated utilizing a network approach. Here, one considers not only the differential expression of individual genes but also their mutual molecular interactions and their involvement in cancer pathways. To build comprehensive regulatory networks, one gathers and integrates information on genes and interactions from bioinformatics databases. Next, the -omics data from patient groups are integrated into the network utilizing computational algorithms that detect the most central and differentially regulated fraction of the network (a.k.a. core regulatory network). With the core network, one can visualize multiple levels of information as illustrated in the lower panel, including gene expression in multiple compared conditions (coloured nodes), P-values (yellow background of nodes) or mutual interactions between genes and proteins (lines). The genes in the core network can be used to design mechanistic-based predictive gene signatures. Also, they can be potential targets for new therapies or repurposing of approved drugs. Although network-based models and signatures provide a way of interlinking genes in signatures for tumour progression and therapeutic efficacy, they cannot cope with the temporal–spatial information of tumour progression.

Figure 5

Mechanistic modelling of melanoma. Mechanistic modelling offers a tool to develop computational models that reflect internally the knowledge about how genes, proteins and cells interact with each other. In model derivation, the existing knowledge of melanoma in the literature and databases is used to create regulatory networks depicting the interplay between the genes and cells of interest, and this information is translated into mathematical frameworks such as ordinary differential equations. In model calibration, one iteratively uses quantitative (time series) data of gene expression, protein abundance and cell populations to find the values of the model parameters that can fit the model simulations to the experimental data. A calibrated model generates predictive and mechanistic simulations, which can be used to understand the mechanism of therapy resistance and predict the time course of the disease from an either molecular (which genes and pathways get deregulated and how) or cellular (which clones will prevail in the tumour and spread) perspective. This type of computational simulations has been employed to assess and personalize cancer therapies, but it can also be utilized to detect drug targets compounds [14], design therapy strategies and schedules. A limitation of this type of models is that, in general, they cannot be constructed in an automatic manner: their derivation, calibration and refinement require manual curation. Also, their calibration often requires large amounts of quantitative time-series data, which is not always available.

In this review, we discuss the state-of-the-art of new technologies for imaging and molecular profiling in melanoma, which we think will help us enter into a 2.0 paradigm in the diagnostics and personalized treatment of melanoma, but also the computational modelling approaches necessary to make full advantage of these data. In Exploiting advanced imaging techniques in melanoma diagnostics section, we discuss advanced imaging techniques in melanoma and their paired computational algorithms utilized in dermoscopic-based diagnostics. In Bulk RNA-Seq, scRNA-Seq and molecular profiling of tumor and blood samples in melano- ma diagnostics and therapy section, we review recent publications in which bulk and single-cell sequencing and molecular profiling of melanoma patient tumour and blood samples have been employed in the context of melanoma patient stratification and therapy design. In Sequencing and molecular profiling-based computational models in melanoma diagnostics and therapies section, we review the use of sequencing and molecular profiling-based computational models for clinical utilization in melanoma. In Into the future: multi-level patient data integration and the dream of an in silico melanoma twin patient section, we discuss possible strategies to integrate multi-level patient data in the context of molecular diagnostics. In The keystones for the success of systems and ML-based melanoma diagnostic technologies section, we describe the keystones required for the successful implementation of these technologies. In our Concluding remarks, we discuss open questions in melanoma, which should/can be tackled with computational approaches.

Exploiting advanced imaging techniques in melanoma diagnostics

The skin is perfect for the application of diagnostic methods based on morphology and accessibility. We can distinguish different types of imaging approaches. On the one hand, macroscopic images and TBP, which operates at the scale visible by the naked eye, can detect and document the evolution of pigmented skin lesions over time. On the other hand, dermoscopic images can be taken and digitalized (video dermoscopy). Dermoscopy or epiluminescence microscopy means that skin lesions are analysed with the help of a magnifier (usually 10-fold) connected to a light source (polarized or not polarized) and a transparent plate to allow the visualization of deeper skin structures than with the naked eye. This techniques is useful to distinguish between melanocytic and non-melanocytic lesions and to evaluate whether the lesion is benign or malignant [16]. While dermoscopy has been used in daily practice worldwide for decades, there are more modern imaging techniques such as reflectance confocal microscopy (RCM) which can also be used in melanoma diagnostics [17]. However, RCM is relatively time-consuming, needs an intensive training of the practitioner and is rather cost-intensive. Finally, histopathological analysis of skin lesions is the gold standard in melanoma diagnostics. Microscopic imaging illustrates the composition of the skin lesion at the histological (cell composition of the tissue) and the molecular level (protein markers expressed by the cells).

Macroscopic and dermoscopic images in melanoma

Melanoma can be detected in patients using macroscopic and dermoscopic images of skin lesions that a trained dermatologist can inspect and compare with curated databases of images [18]. An evolution of this technology is TBP [19]. However, the use of conventional 2D TBP for sequential mapping of melanocytic lesions in patients with abundant nevi is time-consuming in terms of data acquisition and analysis. Recently introduced 3D TBP platforms can reduce these hurdles, for example platforms such as VECTRA WB360 3D TBP (Canfield Scientific, Parsippany, New Jersey, USA), which generates high-resolution photography of the (almost) complete skin surface, enables the rapid and standardized documentation of skin lesions and offers methods for visualization and analysis of sequential 3D maps. This is especially advantageous for the diagnostics of patients with hundreds of pigmented skin lesions [20].

Melanoma patients may benefit in several ways from 3D TBP. Firstly, this type of platform can be used for systematic and early detection of new melanomas. There are several groups with an increased melanoma risk including patients with melanoma in the family, patients with a previous melanoma and patients with freckles and multiple pigmented lesions. These patients are followed with clinical check-ups to detect new lesions and changes in pre-existing nevi. Here, semi-automated 3D TBP offers an option to track nevi. Furthermore, in patients with melanoma cutaneous metastases can occur in transit along the lymphatic drainage of the primary melanoma or via the hematologic route as distant skin metastases [21]. Here, sequential 3D TBP can assist to detect and track these metastatic lesions as well as to identify the areas at risk for metastases [22].

Secondly, 3D TBP can help to assess therapy. It is interesting to track systematically the effect of immunotherapy on present metastases, pre-existing nevi or the healthy cutaneous pigmentation system. Erdmann and co-workers demonstrated that 3D TBP could be used to monitor disseminated epidermotropic metastases that correlated with the clinical course of disease during therapy [15]. In addition, it has been reported that melanoma-directed therapies may alter the pigmentation of melanocytic lesions [23]. Furthermore, vitiligo-like depigmentation has been observed in a subgroup of melanoma patients receiving immunotherapy, while an association with clinical response has been discussed controversially [24]. And lastly, targeted therapies such as vemurafenib have been described to induce nevi and/or morphological changes in their existing melanocytic lesions [25]. Thus, 3D TBP can help detect these adverse events. 3D TBP generates abundant imaging data and its manual analysis for diagnosis increases the burden on clinicians. Thus, the implementation of this technology in a general clinical follow-up setting is only manageable when it comes with computational algorithms for imaging data analysis (see Into the future: multi-level patient data integration and the dream of an in silico melanoma twin patient section).

DL algorithms have been applied recently in the context of macroscopic and dermoscopic image segmentation and classification applied to skin lesions (Figure 2; see Kassem et al. [26] for a recent systematic review). Esteva and co-workers have trained a convolutional neural network (CNN), an advanced type of DL model, with more than 100 000 clinical images from 2000 different skin lesions, including melanoma, other skin cancers and autoimmune and non-cancerous lesions [27]. When testing its performance against 21 board-certified dermatologists on biopsy-proven clinical images, the model classified new images of nevi with an accuracy higher than an average dermatologist (AUC > 0.91). These results are not an exception [28]. Tschadl et al. compared 139 algorithms for dermoscopic image-based classification of pigmented skin lesions with the assessment of 511 clinicians having different levels of expertise [29]. Although the latest computational models clearly outperformed clinicians in classification tasks, the algorithms lost accuracy when tested with dermoscopic images obtained from databases not included in the training set, which suggests a bias in the algorithms’ predictions. Furthermore, Soenksen and co-workers [10] generated, trained and validated a DL model for the analysis of wide-field dermoscopic images, which indicates that this technology can be employed with 3D TBP. One key issue in DL models is that they are built as black boxes with limited ability to interpret and understand the model-based diagnostics. To circumvent this, Gareau and co-workers [30] implemented a computational model to classify melanoma lesions using an ensemble approach, which utilizes a set of imaging biomarkers to quantify medically relevant features and therefore increasing the interpretability of the model. The advantage of dermoscopic image-based models is clear: they can process imaging data such as the one from 3D TBP and generate a diagnosis at least as accurate as one trained dermatologists within seconds, thereby reducing the burden of clinicians and making them focus on the fine details of the diagnostics. This model-based diagnostics technology is mature enough to reach clinical and commercial implementation [31, 32].

Molecular imaging in melanoma diagnostics

To date, risk assessment in early-stage melanoma relies on histopathological features of the primary tumour such as Breslow's tumour thickness or ulceration status. However, these parameters provide often inaccurate prognostic estimates and some studies suggest that two out of three early-stage patients eventually die from melanoma despite initially being diagnosed with stage I or II disease [33]. Thus, identifying early-stage melanoma patients with an increased risk of recurrence from histopathological data remains a challenge. Melanoma diagnostics can benefit from microscopic imaging of tissue sections inspecting the cell or protein composition of the lesions. However, standard methods of immunohistochemistry can only quantify the protein expression of a few target biomarkers. This limitation can be circumvented with multiplex immunohistochemistry imaging techniques (mIHC). Here, primary antibodies against specific antigens serve as the basis of all methods and the technology generates one image per detected antibody for the same tissue section. Depending on the method, it is possible to detect up to hundreds of antigens in the same tissue section [34, 35]. The integration of multiple-image layers allows inspecting the abundance of protein network components in a spatial context on the tissue sample, a key feature to detect tumour cell clone- or stroma cell-specific de-regulated pathways and tumour heterogeneity.

In melanoma, mIHC has been used for the evaluation of immune protein and cell phenotypes, as well as the spatial relationship between tumour and immune cells. mIHC has particular potential in immunotherapy prognosis due to its ability to detect and identify tissue infiltrating immune cells by combinatorial staining or to study associations between the therapy response and differential expression of biomarkers such as CTLA-4 and PD-L1 [36]. Halse and colleagues [37] determined the location of different immune cell subtypes in metastatic melanoma and its stroma. They quantified intratumoural immune cell subsets, as well as the distance of these cells to the tumour margin, and found spatial association between intratumoural CD8+ T cells and PDL1-expressing cells, with a prevalence of PDL1 macrophages compared with cancer cells.

mIHC is intensive in quantitative data generation and analysis and has not yet become a routine in clinical research or disease prognostication. One requires computation algorithms to segment the images and systematically detect cell membranes, nuclei and other elements making up the tissue section. It is also necessary to aggregate the information and generate metrics accounting for protein co-localization or cell heterogeneity [38] (See Into the future: multi-level patient data integration and the dream of an in silico melanoma twin patient section). This app-roach was implemented to predict progression-free survival of chemotherapy-treated metastatic melanoma patients using mIHC quantitative data that profile selected protein regulators of cell apoptosis [39]. The authors used an automated imaging pipeline to quantify the abundance of nine protein regulators involved in the intrinsic apoptosis pathway from formalin-fixed paraffin embedded melanoma tissues. Feeding the protein expression profiles and clinicodemographic variables into a supervised ML model generated predictions on disease progression of patients with up to 80% accuracy. Johannet and co-workers [40] utilized DL models on paired imaging from histology specimens and clinical data from a cohort of patients with advanced melanoma to predict the patients’ response to ICI. Johnson et al. [41] analyzed mIHC imaging of pre-treatment tumour biopsies from patients treated with anti-PD-1 with computational algorithms that can quantify and co-localize the stained biomarkers. Their analysis showed that imaging data-derived scores of the PD-1/PD-L1 interaction and IDO-1/HLA-DR co-expression were strongly associated with patients’ response to the anti-PD-1 treatment. Heck et al. reviewed recent progress and update of applications of AI and mIHC in clinical pathology, with a survey of recent studies using these techniques in melanoma [42].

Radiomics

Data produced with radiological imaging technologies such as computational tomography (CT) and magnetic resonance imaging (MRI) can generate features beyond the visible tumour tissue morphology [43]. Specifically, medical images are transformed into high-dimensional data sets utilizing computer algorithms. The algorithms can extract quantitative features such as size, shape, texture and volume of the tumour and its surrounding tissues, with potential applicability for assisting clinical decision-making (see Guiot et al. for a primer and outlook on radiomics [44]). Softy and co-workers utilized preoperative MRI data to perform non-invasive identification of BRAF mutation status in patients with melanoma metastases in brain (54 samples; the performance of the model is quantified by accuracy = 0.79 ± 0.13, precision = 0.77 ± 0.14 and AUC = 0.78) [45]. Radiomics can also be used to identify biomarkers for predicting patients’ survival and therapies’ toxicity. Dercle et al. used baseline and follow-up CT data sets to train a ML model based on a signature of four imaging features. The model predicted overall survival of patients with advanced melanoma and treated with anti-PD-1 (CT data from 575 patients; the prediction power of the model is quantified by AUC = 0.92 ± 0.03) [46]. Radiomics-based risk assessment of side effects can help to design risk-adapted surveillance [47]. However, the implementation of radiomics in clinical routines is still hampered by a lack of computational models and validation standards. In addition, large-scale multi-center trials are needed to demonstrate the feasibility and value of radiomics in clinical applications.

Bulk RNA-seq, scRNA-seq and molecular profiling of tumour and blood samples in melanoma diagnostics and therapy

Melanoma is a tumour with a high mutational load, and this feature fostered large-scale genomic and transcriptomic studies in the early times of cancer genomics, which tried to classify patient samples into tumour subtypes or to detect gene signatures with predictive potential for therapy or survival.

Shain and co-workers [48] selected and tested 293 cancer-relevant genes in laser-microdissected material of 37 primary melanomas and their adjacent precursor lesions. Based on the analysis of the data, the authors hypothesized that melanoma development follows a step-wise process that often starts with BRAF mutations. Subsequently, TERT promoter mutations emerged in intermediate lesions, and inactivation of CDKN2A was exclusively found in invasive melanomas. PTEN and TP53 mutations were characteristic for advanced primary melanomas. An early study utilizing whole-exome, whole-transcriptome and protein array technologies is the Cancer Genome Atlas Network. In this study, more than 300 melanoma samples were profiled and the data were used to propose a genomic classification defined by four major melanoma subtypes linked to the most commonly mutated genes: BRAF-mutant, NRAS-mutant, NF1-mutant and melanomas wild-type for those three oncogenes (triple wild-type) [49]. The latter subtype comprises, among others, KIT-mutant melanomas. Subsequent experimental work supported the involvement of some of these genes in the early malignant transformation of melanoma [50]. More recently, additional tumour genes with a pathogenic role were identified in a meta-study with transcriptomics data coming from more than 1000 melanoma samples, collected from previously published melanoma studies [51]. The authors found that different secondary driver genes were enriched in given mutation subtypes, for example activated TGF-β signalling in BRAF melanomas or inactivated SWI/SNF complex in NRAS melanomas.

Transcriptomics can be used for risk profiling, an idea that is currently under investigation in a clinical study that prescribes adjuvant therapy based on the molecular analysis of the primary melanoma tumour [52]. Also, one can utilize transcriptomics to look for gene signatures predictive of therapy resistance. For example, transcriptomics analysis revealed that resistance to MAPK-targeted therapies induced specific mutational patterns including gene amplification of mutant BRAF, mutations in NRAS and mutations in the PI3K/Akt pathway, suggestive for reactivation of oncogenic pathways [53, 54]. Similar studies investigated the resistance to ICI [55, 56]. Hugo and co-workers [57] detected a transcriptomics signature involving genes linked to epithelial-mesenchymal transition (EMT) such as AXL, WNT5A and TWIST, as well as genes involved in immunosuppression such as IL10, VEGFA and VEGFC. Technologies developed in the last decade give more choices in terms of how samples are characterized, the molecules quantified and the data are aggregated to generate novel insights.

Single-cell transcriptomics of melanoma tumour samples

Single-cell transcriptomics (scRNA-Seq) is an evolution of RNA-Seq, in which the individual cells are labelled and clustered into cell categories, phenotypes or molecular subtypes and get their transcriptome quantified individually. The benefit for understanding cellular heterogeneity and especially tumour heterogeneity is clear: one can detect differential expression of genes in different cell clusters and measure, assuming unbiased sampling, their population size in the tumour. Also, one can profile the infiltration of stroma cells such as fibroblasts or immune cells into the tumour environment. To date, the routine clinical application of scRNA-Seq is not feasible due to its high costs and difficult sample preparation. However, it is a tool to investigate tumour progression and therapy resistance. Tirosh and co-workers [58] analysed 19 primary melanomas and metastases using scRNA-Seq technology. By integration of single-cell and bulk RNA-Seq data, the authors characterized different melanoma subtypes at the cellular and molecular level. They found evidence that cancer-associated fibroblasts have a major impact on T cell infiltration. Since the technique allowed for detecting and profiling tumour clones in a tissue sample, they found samples with small cancer cell subpopulations displaying a signature of MITF-low/AXL-high expression that made them therapy-resistant. More recent studies have exploited scRNA-Seq to investigate melanoma features such as the existence and role of exhausted CD8+ T cells [59], the existence of gene signatures of T cell exclusion linked to the activation of cell cycle regulators like CDK4 [60] or the existence of tumour cell subpopulations with characteristics of neural crest stem cells [61]. Although scRNA-Seq can profile the gene expression of individual cells, one loses the information about the spatial co-localization of the cells in the tumour. In this sense, an emerging technology named spatial transcriptomics could provide this additional information [62]. Also, scRNA-Seq is time-consuming and cost-intensive, and some key parameters influencing the data analysis pipelines require manual tuning and are therefore difficult to be fully automatized.

Sequencing of non-coding RNAs (ncRNAs)

Protein coding regions of the genome only account for about 1.5% of the total sequence in the genome. The vast majority of the genome comprises non-coding regions, with some of these regions accounting for a class of RNA called non-coding RNAs (ncRNAs). A prominent subfamily of ncRNAs are microRNAs (miRNAs) which repress gene expression through the base-pairing to specific target mRNAs [63]. Single miRNAs have many putative mRNA targets and often inhibit the expression of the corresponding genes modestly. When multiple miRNAs target individual mRNAs or common pathways, they can act cooperatively to substantially repress a target or a pathway [64].

miRNAs are important regulators of melanoma pathogenesis and progression [65]. miRNAs are oncogenic when they target tumour suppressor genes [66]. For example, miR-3151 promotes melanoma metastases by directly targeting TP53 and other members of the TP53 pathway [67]. miR-1908, miR-199a-5p and miR-199a-3p convergently target ApoE signalling in melanoma, promoting metastatic invasion, angiogenesis and colonization in melanoma [68]. miRNAs act as tumour suppressors when they target oncogenes. The let-7 family members suppress melanoma proliferation and metastases by targeting cell cycle regulators such as CCND1 and CCND3 as well as ITGB3 [69], a gene strongly associated with the acquisition of invasive features by melanoma cells [70]. Besides, miR-34b, miR-34c and miR-199a can suppress the migratory ability of melanoma cells by downregulating the expression of the oncogene MET [71]. There are other cases of miRNA tumour suppressor in melanoma [72, 73]. We need to be mindful that some miRNAs can play contradictory roles in regulating melanoma. For instance, miR-146a can promote melanoma cell growth by targeting NUMB, a repressor of the NOTCH signalling pathway [71]. However, the same miRNA can suppress metastases formation by downregulating the expression of ITGAV and ROCK1 [74].

Researchers have also found that miRNAs are linked to the effectiveness of melanoma therapies, including targeted therapies [75] and immunotherapy [76]. Having this in mind, the profiling of miRNAs in tumour samples alone or in combination with transcriptomics data can be used to find predictive gene signatures for cancer progression. Segura and co-workers [77] profiled 59 formalin-fixed paraffin embedded melanoma metastases samples and found a signature composed of six miRNAs that can predict post-recurrence survival in metastatic melanoma with an estimated accuracy of 80.2 ± 0.4%. The authors claimed that the signature can be used to stratify stage III patients into different prognostic categories. The same team profiled primary melanoma tumour samples from three patient cohorts (total n = 256) and found a four-miRNA signature (miR-150-5p, miR-15b-5p, miR-16-5p and miR-374b-3p) that has the potential to identify primary melanomas with later metastasis to the brain [78]. The general function of miRNAs as suppressors of gene expression is well-established and has been experimentally validated for many miRNAs. But our understanding of other types of ncRNAs is more limited, which hampers the chances to use them as biomarkers.

Liquid biopsies: quantitative profiling of blood samples

One of the long-time pursued dreams in clinical oncology is to have accurate diagnostic tests based on molecular blood profiling (liquid biopsy), which can be used as a non-invasive, cost-effective technology for repetitive assessment of cancer patients. In melanoma, the serum levels of S100 or MIA were traditionally considered candidates for tumour markers [79]. However, both markers are generally not detectable during early tumour growth and hence their regular profiling does not contribute to early detection of tumour relapse. In recent years, there have been efforts to develop alternative blood-based tests for melanoma.

Tumour DNA-based liquid biopsies

One type of liquid biopsy involves looking for tumour-derived, free-floating DNA fragments loaded into the blood originated from the tumour or circulating tumour cells (circulating tumour DNA, ctDNA) [80]. Diefenbach and co-workers [81] investigated a melanoma liquid biopsy platform based on the profiling of ctDNA. To this end, they implemented a melanoma-specific DNA-Seq panel including 123 amplicons in 30 genes covering driver and targetable mutations, as well as alterations associated with therapy resistance. The panel was tested in a cohort of 74 stage III and IV treatment-naïve melanoma patients, and they found that at least one of their selected cancer-associated mutations was detected in ctDNA in 84% of stage IV patients and 47% of stage III patients. ctDNA is also being used for surveillance of ongoing therapy in order to detect progression of disease before CT/MRI imaging are performed. Although these studies show the potential clinical utilization of circulating DNAs of tumour cells, we still need the investigation of ctDNA profiling in larger and multi-centre cohorts to establish its value as a liquid biopsy in melanoma.

Circulating ncRNAs

ncRNAs can circulate in the bloodstream and therefore can be used in liquid biopsies. Circulating miRNAs (cmiRNAs) show potential for diagnostic due to the ability of miRNA signatures to distinguish cancer subtypes [82, 83] and the chemical stability of miRNAs [84]. Individual cmiRNAs could serve as prognostic biomarkers. For example, miR-221 was identified by different studies as a possible prognosticator in melanoma [85, 86] and increased serum levels of miR-221 were correlated with elevated tumour stage and recurrence. However, predictive models based on single miRNAs may have low specificity and signatures composed of multiple cmiRNAs are a promising option to circumvent this issue. Ledinger and co-workers showed that 16 differentially expressed miRNAs in blood samples can discriminate with high sensitivity (0.95) and specificity (0.98) between patients with metastatic melanoma (stages III and IV) and healthy donors [87]. Although their results are interesting, the analysis would be more conclusive if early melanoma stages samples would be included in the study. Van Laar and colleagues identified 38 cmiRNAs that show high accuracy in discriminating patients from healthy donors (AUC 0.79–0.94). Their miRNA signature is predictive for different sample types including blood, tumour tissue and exosomes, which suggests robustness in its potential use as biomarkers [88].

Extracellular vesicles

Extracellular vesicles are microscopic cell-derived vesicles that are secreted by immune, stem and tumour cells into plasma (pEVs) and other body fluids. The vesicles carry functionally active proteins and RNAs, including miRNAs [89, 90]. The molecular profiling of pEVs has the potential to diagnose and determine prognosis of melanoma patients. Lee and co-authors [91] isolated and profiled pEVs from 14 tumour-bearing melanoma patients, including samples from primary melanomas, skin, lymph node or disseminated organ metastases. They found that the miRNA levels in pEVs are elevated on average 6.6-fold when compared with age-matched healthy controls. Interestingly, they observed that the miRNA levels in pEVs drop back to normal 2 weeks after tumour resection. The authors found 21 up-regulated miRNAs in pEVs of tumour-bearing patients when compared with healthy controls, but also other regulatory factors such as the cancer-associated protease ADAM10 [92]. Pietrowska and co-workers [93] performed mass spectrometry-based proteomic profiling of pEVs from 15 melanoma patients. Since they sorted the vesicles, they could obtain proteomic profiles from melanoma and non-melanoma cell-derived pEVs and found a set of 16 proteins discriminating provenance. They also identified proteins differentially expressed in pEVs from progressive disease patients compared with patients without signs of relapse. Although these results are promising, collection and profiling in larger and multi-centre cohorts is necessary to confirm the value of pEVs as a liquid biopsy in melanoma. To this end, technologies and protocols for standard purification of pEVs have to be established.

While the quantification of single RNAs, mutations or proteins in serum samples could be implemented clinically similar to standard plasma biomarkers, the use of high throughput technologies can generate complex, multiple-marker signatures, which can be established only when combining quantitative profiling and computational modelling.

Sequencing and molecular profiling-based computational models in melanoma diagnostics and therapies

ML/DL-based exploitation of genomics data in melanoma

Using genomics data to build predictive ML models is a powerful tool that is being widely utilized nowadays (Figure 3). Hou and co-workers integrated transcriptomics and methylation profiles from uveal melanoma patients to implement and trained an ML model that can classify the cohort into two risk subtypes with significantly different overall survival (HR = 29, 95% confidence interval: 6.7–126, P < 0.001) [94]. Lai and co-workers [95] integrated genomics data and DL to find gene signatures for classifying melanoma patients for prognosis. In their approach, they combined network analysis and an autoencoder DL model to reduce the dimensionality of genomics data into a patient score profile. They utilized the score and identified three subtypes in TCGA melanoma patients that show different survival times. Finally, they computed SHAP coefficients to quantify and rank the impact of genomic features, thereby rending the interpretability of their DL model. Xie et al. [96] developed an autoencoder DL model trained with whole exome sequencing data and identified a group of melanoma patients showing resistance to ICI treatment such as anti-CTLA-4. The patient group had somatic copy number alterations with low tumour mutation burden and high microsatellite instability. Atak et al. [97] presented a CNN model called DeepMEL2 to identify and characterize functional enhancer mutations in different melanoma cell lines. The model was trained on integrative genomics data that represent melanoma chromatin accessibility and provided explanation to the consequence of enhancer mutations on gene regulation, such as how non-coding variants affect functional cis-regulatory programs of the melanocytic and mesenchymal-like melanoma cell states.

One can also utilize ML models based on transcriptomics data to assess tumour immunity. Immunoscores are scoring systems based on the quantification of the tumour infiltration levels of lymphocytes. The quantification of the tumour infiltration of well-defined populations such as CD8+ T cells, macrophages, B cells and DCs has predictive value for disease progression or response to immunotherapy [98–100]. The standard way of assessing immunoscores relies on the staining and quantification of the T cell infiltration ratio by pathologists in tumour sections [101], but this procedure is time- and personnel-intensive and has some level of subjectivity. Image segmentation algorithms can reduce the resources required for imaging data analysis, but the hurdles linked to preparation and quantification of the samples persist. Transcriptomics data offer a quantitative way of accessing tumour composition: computational algorithms can be used to aggregate the expression levels of genes preferentially expressed by given types of immune cells and use this information to infer corresponding levels of tumour infiltration. The literature contains several algorithms performing RNA-Seq-based computational quantification of tumour infiltration [102]. To mention one, CIBERSORT is an algorithm that infers cell composition of fresh, frozen and fixed tissues from their gene expression profiles [103]. Nie and colleagues [104] found using RNA-Seq data that a high immunoscore implies a better response to anti-PD1 therapy in metastatic melanoma. Vokes and co-workers [105] recently performed transcriptome sequencing from pre-treatment tumours in melanoma patients treated with ICI and used CIBERSORT to infer relative proportions of 22 types of immune cells in the tumour. Interestingly, they identified higher total immune infiltrate in stable patients compared with patients with complete/partial response or progressive disease.

Sequencing-based personalization of immunotherapy

Although ICI therapy is the current standard of care for metastatic melanoma, other immunotherapies are under investigation for unresponsive patients. In the case of immunization with tumour epitopes, one key element is to select a set of tumour neoepitopes or non-mutated epitopes able to elicit a robust antitumour T cell response [106]. Algorithms such as NetMHCpan and MHCflurry combine molecular modelling and ML to predict peptides’ binding affinity to specific MHC-I alleles [107]. Transcriptomics data in combination with bioinformatics algorithms offer the chance to make an optimal and personalized selection of tumour epitopes. Lischer and co-workers [108] created a database of non-mutated tumour epitopes by mining transcriptomics datasets for cutaneous melanoma. Their algorithm selected MHC-I restricted epitopes from genes overexpressed in tumour and minimize the risk of off-site target effects. To enhance the predictability of these methods, much work has to be done to improve the computational algorithms predicting epitope binding affinity and to integrate data concerning patient’s T cell repertoire.

Network analysis in melanoma

Network-derived gene signatures

With network modelling, one can detect signatures of interacting genes underlying progression or therapy responsiveness [109] (Figure 4). Dreyer and co-workers [110] combined a manually curated network of melanoma gene regulatory pathways with transcriptomics data from tumour samples of patients with different responses to the anti-PD1 immunotherapy. They detected a core regulatory network of 41 interacting factors discriminating responders from non-responders. Interestingly, the network contained many proteins involved in EMT, as well as members of the E2F TF family such as E2F1, known for its pivotal role in melanoma progression [111], EMT [112], angiogenesis [113] and drug resistance [114, 115]. Song and co-workers used network analysis to integrate bulk and scRNA-Seq data with clinical, epigenetic and proteomic data from primary melanoma. They detected core regulatory subnetworks and cell types underlying melanoma progression and tested in vitro and in vivo 17 key network genes associated with poor prognosis [116]. The analysis suggested a prominent role in tumour progression for immune genes such as MYO1F and TFs such as ZNF180.

Network-based drug discovery and repurposing

Kohsravi and collaborators combined genome-wide association studies, transcriptomic datasets and metabolomic datasets with network analysis to detect melanoma-linked factors known to be the target for drugs included in DrugBank [117]. Using this approach, they found 35 drugs interacting with 20 unique melanoma targets. In more advanced computational setups, network analysis can be combined with other computational models [118]. For example, one can employ molecular-level simulations such as molecular docking [119] or pharmacophore modelling [120] to investigate the binding between the targeted proteins and small molecules proposed as drugs. Dreyer and co-workers [110] identified E2F1 as a central player in a network underlying melanoma therapy resistance, but the precise role of E2F1 in orchestrating therapy resistance gene regulatory networks relies on the interaction with its transcriptional cofactors. Hence, a way of targeting E2F1 is to search for small molecule inhibitors interfering in the TF-cofactor interaction. Goody and co-workers [121] found that the MTA1 cofactor triggers E2F1 targets involved in tumour metastasis. They used pharmacophore modelling to prioritize amino acid residues that participate in the formation of the E2F1-MTA1 binding. Next, they screened computationally a large FDA-approved library of molecules and found 16 compounds with the potential to fit to their E2F1-MTA1 pharmacophore model. Further in vitro/in vivo studies identified argatroban as a small molecule effectively interfering in the formation of the MTA1-E2F1 complex.

Network analysis in other melanoma therapies

Using transcriptomic data from melanoma patients with different responses to anti-PD1 treatment and a comprehensive network of melanoma pathways, Lai and co-workers [122] identified and ranked cooperative miRNAs that could target dysregulated genes in melanoma, which could be deliver into the tumour as therapy using molecular carriers such as nanoparticles [123]. miRNAs can also be used to engineer immune cells for immunotherapy. Lai et al. [124] utilized a comprehensive network on DC activation and transcriptomics analysis to detect miRNA-gene interactions with the potential to enhance DC immunotherapy for metastatic melanoma. When applied to DCs transfected with mRNA encoding constitutively active IKKβ and able to expand cytotoxic T cells [125], the targeting of IKK-β by miR-15a-5p and miR-16-5p was identified as a candidate interaction to further activate NF-κB signalling and boost the immunogenic potency of therapeutic DCs.

Mechanistic modelling in melanoma

Mechanistic computational models to understand melanoma genes and cell networks

One can utilize these models to understand the connection between genes circuits and cell phenotypes in the context of progression and therapy resistance (Figure 5). Lai and Friedman [126] developed a continuum model of the immune response to melanoma and used it to simulate the effect of combining BRAF/MEK and PD-1 inhibitors. The model accounts for avascular tumour growth, dynamics of DCs, myeloid-derived suppressor cells and various T cell subpopulations, as well as the effect of immunosuppressive cytokines on these cells. Interestingly, their model suggests that for low doses, the combination therapy is effective, while it becomes antagonistic for high doses. Vera and co-workers [114] built a computational model linking the deregulation in melanoma of a gene network around E2F1, p73 and miR-205 with cell population dynamics and genotoxic/cytostatic therapy efficacy. The model simulations predicted gene signatures inducing therapy resistance and suggested that tumour heterogeneity in terms of gene network regulation enhances resistance. These predictions were validated by experiments. A recent review by Albrecht and co-workers [127] gives an overview of published mechanistic computational models for melanoma.

Mechanistic model-based therapy discovery and assessment

Mechanistic modelling is routinely used in the pharmacologic industry to establish safe dosage (pharmacokinetics) and mechanism of actions (pharmacodynamics) for anti-cancer drugs [128, 129]. But mechanistic models can also be used to identify combinatorial therapeutics, assess the effectiveness of therapies or stratify patients in terms of the foreseen response to therapy [130]. For instance, Shin et al. [131] modelled the EGFR-PYK2-c-Met interaction network to identify effective drug combinations for cancer and predict patients’ response to the drug combinations based on their gene expression profiles, thereby demonstrating the potential of mechanistic modelling to develop personalized treatments. Using the similar data-driven modelling approach, researchers from the same lab investigated combinatorial therapeutic strategies for reducing the resistance of tumour cells to FGFR4-targeted therapy. They found that co-targeting FGFR4 and ErbB kinases or AKT showed synergistic effects in tumour suppression [132]. Based on a generic model, Schmucker et al. [133] developed a computational framework to search for novel combination therapies for cancer cell lines and optimize sequential treatment plans for preventing inherent and acquired drug resistance of tumour cells. This work showcases the advantage of in silico experiments in facilitating and accelerating discovery of efficient therapy for cancer.

On the other hand, Brady and Enderling [134] differentiated between academic and translational mechanistic models, the latter able to predict the outcome of novel therapies or treatment protocols and therefore directly motivating clinical trials. Passante and co-workers [135] utilized protein expression data to train a computational model predicting the sensitivity of melanoma cells to apoptosis. With model simulations, the baseline expression of the selected apoptosis regulatory proteins was sufficient to predict the response of melanoma cell lines to the apoptosis-inducing agents TRAIL (tumour necrosis factor-related apoptosis-inducing ligand) and dacarbazine, a standard of care chemotherapeutic option in melanoma. In an in silico clinical trial with metastatic melanoma patient data, the model successfully estimated response rates to the pro-apoptotic combined treatment of a therapeutically relevant TRAIL variant and Birinapant [136]. Under the assumption that the drugs reach their target sites, the model predicts that 30% of otherwise unstratified metastatic melanoma patients might respond to the combined treatment.

Mechanistic modelling has also been used in the context of immunotherapy. De Pillis and co-workers [137] developed a mathematical model describing the dynamics of dendritic cell vaccination in melanoma. Through simulations, they designed a therapy schedule with the potential to improve efficacy and detected patient-specific parameters with an impact on treatment efficacy. Santos and co-workers [138] combined transcriptomics data and mechanistic modelling of melanoma DC vaccination to predict gene signatures discriminating immune sensitivity and therapy resistance. One of several mechanisms for therapy resistance predicted requires intermediate expression for antigen presentation genes in cancer cells that could minimize the combined antitumour activity of cytotoxic T and natural killer cells. Interestingly, the identified signatures matched with the transcriptomics data of therapy-resistant patients [139]. Finally, mechanistic modelling can be used to link the dynamics of DC biodistribution during vaccination with the activation of intracellular pathways mediating the DC-T cell interaction [140].

Into the future: multi-level patient data integration and the dream of an in silico melanoma twin patient

An important issue that remains is the necessity to integrate all these different types of clinical, imaging and molecular patient data and to reduce their complexity for a clinical application. Rather than thinking about technologies competing for space in clinics and markets, we foresee a future, in which data coming from all these platforms are integrated to enhance diagnostics and therapy. Here again, computational modelling is a must to make this possible. We see two possibilities (Figure 6).

Figure 6

The future of melanoma care: integration of multi-layered quantitative patient data into computational models. Orange panel. In future, we will generate a multiplicity of clinical, imaging and molecular data for each patient, taken from clinical records, as well as blood and tumour samples. To enhance diagnostics and therapy, we will have to integrate these multi-layered data. We foresee two possibilities for this integration. Grey panel. Medical computational dashboards are tools for aggregating and visualizing quantitative patient data and facilitating clinical decisions, as well as discussion in multidisciplinary tumour boards. Blue panel. Ensemble computational models build on top and integrate ML, network and mechanistic models specific for different patient data. They could speed up and improve clinical decisions by integrating multi-layered data in models for patient classification.

In the last decades, there is intense work in developing computational platforms allowing systematic and privacy protection complying with integration of clinical, imaging and molecular data. A paradigmatic case is TranSmart [141]. In an evolution of these types of tools, computational algorithms could be used to aggregate and visualize the data. Different types of figures and metrics generated this way could be integrated into the health care analogous of a car dashboard (Figure 6, grey panel). This may facilitate the overview and handling of the data, but the interpretation of the data in a clinical sense would still fall primarily into the expertise of well-trained physicians. These medical computational dashboards are an optimal tool to inspire discussions in multidisciplinary tumour boards [142].

One can go one step further. Since we talk about many types of multi-parametric data, training a clinician to scrutinize all these data in a time-effective manner is difficult. Furthermore, some important data features can be hidden in the complexity of the data. To circumvent this, one can analyse different types of data with specific, patient data-driven computational models (Figure 5, blue panel). In order to guide their decisions, physicians can obtain an augmented dashboard with both data and computer model predictions on patient classification and therapy assessment. Moreover, some algorithms allow the integration of the individual predictions into an aggregated model that could speed up clinical decisions (we name them ensemble computational models). For example, Lai and co-workers [95] integrated multiparametric genomics data, disease network modelling and DL to classify melanoma patients for prognosis. These types of computational data structures and models could be used to feed and aggregate continuously the patient’s clinical and profiling data, thereby obtaining a digital replica for the patient with predictive capabilities that some call ‘digital patient twin’ [143]. This idea is still in its infancy and faces several problems such as the lack of sufficient curated medical data to build them [144] or the risk of undetected biases in the computational models due to factors such as gender or ethnicity [145]. But certainly, this concept underlies several ongoing research programs funded publicly or by large information technology companies.

The keystones for the success of systems and ml-based melanoma diagnostic technologies

Each of the technologies discussed here has its repertoire of potential use and requires specific types of computational modelling to analyse the data produced. To evaluate the potential of the described technologies for actual implementation in clinical practice, we assess them in terms of clinical application (capability) and economic impacts as well as required behavioural changes (Table 1).

Table 1

Technologies for profiling melanoma patients

Notes: mIHC: multiplexed immunohistochemistry; sc: single-cell; nc: non-coding; c-:circulating; pEV: plasma extracellular vesicles. Computational models. Black: published results, grey: feasible; white: not applicable. Applications. Dark blue: published results; light blue: feasible. Economics. Investment: red: <25 k Euro; pink: <250 K Euro; white: >250 K Euro. Consumables (cost per sample): red: <50 Euro; pink: <250 Euro; white: >250 Euro. Processing time: red: <30 min; pink: <5 h. Operator skill level: red: basic care nurse; pink: technical assistant; white: clinician. Acceptance by clinicians: black: current high acceptance; grey: foreseen acceptance; white: foreseen difficulties for acceptance. Fits clinicians behaviour: black: fits current behaviour; grey: slight changes required; white: large changes required. Fits industry structure: black: fits structure; grey: affordable changes required; white: large changes required. To evaluate the potential of the described technologies for actual implementation, we employed the model of the innovation cube proposed by Witt and Olschewski (2017; https://www.theron.com/wp-content/uploads/2020/08/2017_the_sight_2_17.pdf).

The cost to implement a new diagnostics technology depends on not only the initial investment of the diagnostic device but also the price of consumable kits (often expensive) and the time spent by clinicians and practitioners (Table 1). From the technologies analysed, we see a variety of situations. For example, bulk RNA-Seq of tumour samples requires a large initial investment but relatively low consumable costs per sample. The technology also has established standards for data processing and analysis. Besides, other technologies, such as mIHC imaging and dermoscopic images or 3D TBP, come with more affordable devices but require expensive consumables and are costly in terms of handling time of skilled operators or clinicians in the analysis of the data.

The other important aspect for the success of these technologies relies on the behavioural changes in hospitals and health care industries they require (Table 1). Some of the technologies such as 3D TBP adapt themselves quite well to the current practice in clinics. Such technologies are based on knowledge and principles that are widely accepted by practitioners and also fit well into the productive structures of industries. Others such as mIHC imaging may require significant changes in clinical routines and clinicians’ behaviours. More challenging technologies such as profiling of pEVs require further investigation to generate the evidence for clinical utilization.

Since all of these novel technologies will need the aid of data analysis and computational modelling to maximize their power in clinical practices, they face new issues that were not encountered by conventional diagnostics technologies. Firstly, the integration of computational models and AI into health products and technologies needs permission in line with new regulation and policies, especially for AI-driven tools with high risk of sacrificing data privacy and making biased predictions. Thus, regulatory organizations could draw a boundary between different technologies, with some of them facing significant hurdles in clinical use due to their intensive utilization of data with high privacy issues such as patients' genomic information. Furthermore, patent protection is not effectively enforceable in products that rely heavily on data and algorithms. Instead, what is often relevant in these data-intensive diagnostics technologies is the exclusive access to relevant biomedical datasets and the know-how on both algorithms and the patient data used to train them. This attributes a unique role to hospitals in the development and improvement of these technologies since they are the only stakeholders that have access to the patient data necessary to train and test computational models.

Concluding remarks

In this paper, we have introduced and discussed imaging and molecular profiling technologies that can improve melanoma diagnostics and therapy in the next decade. Beside the mentioned technologies, there are other technologies that are still in an early development phase for melanoma. For instance, tumour or blood sample proteomics, lipidomics and/or epigenetics analyses of tumour or blood samples are under development for diagnosing melanoma and can also benefit from the computational modelling methods discussed by us [146]. The metagenomics analysis of the gut microbiome has the potential to assess the efficacy of cancer immunotherapy [147] and bioinformatics plays a key role in dissecting the metagenomics data [148]. There are open questions for melanoma and skin lesions that could be addressed with computational approaches. For example, it remains to be investigated whether the combination of dermoscopic imaging and molecular profiling of lesions can improve the accuracy of the current diagnostics methods.

Whether all these technologies will be implemented in clinical practice in the coming years will depend on a trade-off between the aforementioned points including the costs of relevant devices, the computational resources/ability required for the deployment of the technologies in clinical routine, legal issues linked to patient privacy and the intensity of the changes in hospitals and health care industries required by the technologies.

Key Points

• Melanoma is a paradigm for deploying advanced tumour profiling and computational modelling due to its tissue accessibility and suitability for imaging techniques.

• Macroscopic body photography can be combined with artificial intelligence algorithms to detect and classify pigmented skin lesions and document their evolution.

• Multiplex immunohistochemistry imaging techniques provide the composition of the skin lesion at the histological and molecular level, but the complexity of the spatial data generated demands algorithms for image segmentation.

• The high mutational load of melanoma favors large-scale bulk and single-cell genomic and transcriptomic studies, but interpreting these data to the full extent is only possible when equipped with network-based algorithms for classification.

• Mechanistic computational models of the gene circuits and the cell processes underlying melanoma progression or therapy resistance can be used to delineate biological mechanisms or design therapeutic strategies.

• In the future, computational models will be used to aggregate the patient’s clinical and molecular data and create a digital replica of the patient with predictive capabilities.

Acknowledgements

The content of this paper is based on a workshop on system medicine held in the Deutscher Hautkrebskongress 2020, Nürnberg (ADO 2020). We thank the organizers of the ADO 2020 for the opportunity they offered to us. The idea of the paper was originally conceived by J.V. and C.B. X.L. and M.H. revised and edited the draft. All the authors contributed text, revised and approved the final version. We apologize to colleagues whose work could not be discussed and cited due to space restrictions.

Funding

The European Union's Horizon Research and Innovation Programme Grant Agreement No. 101057250 [CANCERNA to J.V.]; German Federal Ministry of Education and Research (BMBF) [KI-VesD 161L0244A to J.V., A.B., S.W.; e:Med-MelAutim 01ZX1905A to X.L., J.V.; e:Med-MelAutim 01ZX1905B to O.W., S.G.; e:Med-MelAutim 01ZX1905D to B.P.; e:Med-MelAutim 01ZX1905E to L.H.; NADIM 13GW0406E to C.O.]. Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy (EXC 2075 – 390740016 M.R. and C.G.); Stuttgart Center for Simulation Science (SimTech).

Author Biographies

Julio Vera is a physicist and a professor of systems tumour immunology.

Xin Lai is a computer scientist and has expertise in gene network-driven patient classification and therapy assessment.

Andreas Baur is a dermatologist and a translational researcher who investigates the role of extracellular vesicles in cancer and immunity.

Michael Erdmann is a dermatologists with expertise in skin cancer, dermatohistopathology and total body photography as a diagnostics tool in skin cancer.

Shailendra Gupta is a computer scientist and has expertise in computer-driven drug repurposing.

Cristiano Guttà is an experimentalist who works in computer model-based cancer patient classification.

Lucie Heinzerling is a professor of onco-dermatology and translational research.

Markus Heppt is a dermatologist with focus on skin cancer and his research interest is drug discovery in melanoma.

Philipp Maximilian Kazmierczak is a clinician and his expertise is radiomics.

Manfred Kunz is a professor of onco-dermatology and translational research.

Christopher Lischer is a bioinformatician and his expertise is in cancer transcriptomics and next generation sequencing.

Brigitte Pützer is a professor of molecular oncology and the director of the Institute for Experimental Gene Therapy and Tumor Research at the University Hospital Rostock.

Markus Rehm is a professor of molecular oncology and the director of the Institute of Cell Biology and Immunology at the University of Stuttgart.

Christian Ostalecki is a chemist and his expertise is multiplex immunohistochemistry.

Jimmy Retzlaff is a computer scientist and his expertise is in computational modelling in cancer.

Stephan Witt is an economist and has expertise in the evaluation of emergent technologies.

Olaf Wolkenhauer is a professor of bioinformatics and the director of the Department of Systems Biology and Bioinformatics at the University of Rostock.

Carola Berking is a professor of dermatology, a skin cancer expert and the director of the Department of Dermatology at the University Hospital Erlangen.

Author notes

Julio Vera and Xin Lai have contributed equally to this work.