Integrating genetic and non-genetic determinants of cancer evolution by single-cell multi-omics

Abstract

Cancer represents an evolutionary process through which growing malignant populations genetically diversify, leading to tumour progression, relapse and resistance to therapy. In addition to genetic diversity, the cell-to-cell variation that fuels evolutionary selection also manifests in cellular states, epigenetic profiles, spatial distributions and interactions with the microenvironment. Therefore, the study of cancer requires the integration of multiple heritable dimensions at the resolution of the single cell — the atomic unit of somatic evolution. In this Review, we discuss emerging analytic and experimental technologies for single-cell multi-omics that enable the capture and integration of multiple data modalities to inform the study of cancer evolution. These data show that cancer results from a complex interplay between genetic and non-genetic determinants of somatic evolution.

The evolutionary history of cancer includes malignant transformation, followed by progression to more aggressive and resistant forms¹, ultimately leading to its devastating clinical impact². Somatic mutations, including single nucleotide variants (SNV) and structural variations, are critical for cancer initiation and evolution^1,3. Nonetheless, in recent years, pervasive somatic mutations have been identified across healthy tissues^4–8, particularly those exposed to environmental carcinogens, such as skin and the oesophagus, which suggests that cancer often arises from pre-malignant clonal outgrowths. Importantly, somatic mutations in clonal outgrowths that do not progress, or even regress, overlap with recurrent drivers of cancer^4–9. These data suggest that genetic mechanisms alone may be insufficient to drive malignant transformation¹⁰.

Somatic mutations.

Alterations in the DNA acquired post-conception (versus germline mutations) and able to be passed onto progeny of mutated cells. Somatic mutations can be detected by sequencing in otherwise histologically normal appearing tissue, are often associated with age and environmental exposures, and can manifest in cancer driver genes.

Clonal.

A population of cells with the same underlying genetic make-up (that is, with the same somatic mutations), which can beget subclonal populations that have acquired additional genetic aberrancies.

Drivers.

Somatic mutations that increase tumour cell fitness.

As malignant populations grow, cells undergo further genetic diversification that enables tumour progression, relapse and resistance to therapy³. However, clear genetic drivers of progression, metastasis and therapeutic resistance are identified in only a subset of tumours, pointing to non-genetic contributors of cancer progression^11,12. Recent advances, particularly in single-cell technologies, have revealed intratumoural heterogeneity in cell states, epigenetic profiles, spatial dynamics and interactions with the tumour microenvironment. These axes of potentially heritable intratumoural variation may provide additional cues for cancer evolution. Thus, the integration of multiple layers of information for individual cancer cells, via single-cell multi-omics, is critically needed for a comprehensive understanding of the mechanisms of cancer evolution (FIG. 1).

Fig. 1 |. Single-cell multi-omics for deciphering clonal evolution in cancer.

Analytic or experimental integrations of multiple data ‘omics’ modalities in single-cells advance our understanding of mechanisms of clonal evolution. a | Cancer cell representation with heritable traits that can be interrogated via multi-omics platforms. b | Extracting DNA methylation (DNAme) and transcriptomic information from the same cells experimentally has been achieved by modifying plate-based single-cell RNA sequencing (scRNA-seq) protocols (for example, Smart-seq2), in which both RNA and DNA are respectively isolated from the same cells for whole-transcriptome and DNAme data through bisulfite sequencing^121,128,129. Heritable stochastic DNAme changes can then be exploited as native barcodes to directly infer the high-resolution phylogenetic history of tumour cells¹²¹. c | scRNA-seq with integration of protein expression measurements can be performed in parallel for the same cells^122,123. DNA-barcoded antibodies, acting as synthetic transcripts, are used to convert the detection of proteins into a quantitative readout. This allows the immunophenotyping of cells to be integrated with an unbiased transcriptome analysis using existing single-cell sequencing approaches. d | High-sensitivity somatic genotyping in which, for instance, any mutation in mitochondrial DNA may serve as lineage markers. Interrogating these naturally occurring genetic barcodes within scATAC-seq⁹⁷ (or scRNA-seq) provides high-resolution phylogenies coupled with cell state information. e | As an example of spatially aware platforms^{160,162–164,166,169}, spatial transcriptomics¹⁶⁵ utilizes molecular barcodes for the detection of mRNA molecules and maps them to their spatial positioning. gDNA, genomic DNA; indels, insertions or deletions; scRRBS, single-cell reduced-representation bisulfite sequencing; SNVs, single nucleotide variants.

Cell states.

A cell’s phenotype, as inferred by transcriptional or protein markers, that are often transitional (for example, intermediate states within a developmental system such as haematopoiesis or epithelium).

Single-cell multi-omics.

Analytic or experimental integrations of multiple data ‘omics’ modalities in single cells.

Herein, we review recent experimental and analytic innovations in single-cell technologies that integrate multiple dimensions of heritable information in individual tumour cells. This Review addresses both genetic and non-genetic routes of cancer evolution that can be uniquely interrogated through single-cell multi-omics advancements. As comprehensive discussions of genetic sources of cancer heterogeneity³ and the technical aspects of integrative single-cell analyses¹³ have been recently undertaken, we focus herein on nascent but compelling evidence that motivates the integration of multiple strata of information in single cancer cells to decipher tumour heterogeneity and evolution.

Genetic heterogeneity and lineage tracing

Inference of clonal architecture in bulk sequencing.

Genetic heterogeneity through the continuous acquisition of somatic mutations underlies clonal evolution in many cancers. The clonal architecture of these genetically heterogeneous populations has been previously inferred through bulk next-generation sequencing¹⁴. The integration of read depth and variant allele frequencies of somatic mutations in whole-exome or whole-genome sequencing (WGS) data can be used to infer tumour purity, ploidy and local copy number for each mutation and thus to determine the cancer cell fractions (CCFs) harbouring the mutations¹⁴. These data can resolve clonal and subclonal relationships to a limited extent (FIG. 2a). In aneuploid solid tumours, WGS can help time early versus late clonal somatic alterations by determining whether SNVs occurred before or after large amplification processes such as whole-genome duplication¹⁵.

Fig. 2 |. Phylogenetic inference for retrospective lineage tracing.

a | Bulk next-generation sequencing allows inference of clonal architecture phylogenetic trees of genetically heterogeneous populations. However, these data can resolve clonal and subclonal relationships to a limited extent by enabling the assessment of the order of acquisition of mutations (A–D) and are limited in their abilities to resolve the phylogenetic relationships of clones, especially at low cancer cell fractions (CCFs). b | Multi-sampling at different time points (T1, T2) during clonal evolution or at different regions (R1, R2) within a tumour (to assess for intratumoural clonal spatial composition) can provide higher-resolution phylogenetic relationships, even for subclones with low CCFs, owing to coordinated patterns of CCF fluctuations over time. c | Even though additional resolution is gained through multi-sampling, resolving phylogeny at single-cell resolution (by single-cell whole-genome sequencing or targeted sequencing) is required to derive the precise clonal dynamics and evolutionary history of a tumour. High-resolution trees pave the way for critical inferences derived directly from primary patient tumours for defining key parameters of somatic evolution. VAF, variant allele frequency.

Variant allele frequencies.

Frequencies of mutated alleles in the sequenced reads from next-generation sequencing. Variant allele frequencies reflect copy number, zygosity, tumour purity and the fraction of cancer cells that harbour the mutation (for example, clonal versus subclonal).

Tumour purity.

Per cent of a tumour mass composed of tumour cells, versus admixed non-neoplastic cells, such as tumour-infiltrating immune cells and stromal cells.

Cancer cell fractions.

(CCFs). Fractions of cells that harbour a given mutation.

When applied to large cohorts, these analyses provide a blueprint of the evolutionary sequence in which driver mutations arise. Thus, driver events can be separated into early, often ancestral mutations (for example, DNMT3A and TET2 mutations in myeloid neoplasms) versus late subclonal drivers that arise as a result of ongoing selection within the malignant population (for example, SF3B1, TP53 and NPM1 mutations in lymphoid and myeloid malignancies)^14,16,17. Similarly, in a large cohort of renal cell carcinomas, tumours could be classified into evolutionary subtypes (for example, tumours with driver events in PBRM1, which subsequently acquire alterations in SETD2 or PI3K) that could be linked with prognostic information (for example, ‘PBRM1 → SETD2’ or ‘PBRM1 → PI3K’ subtypes with an attenuated disease course)¹⁸. Thus, timing analyses have charted different evolutionary trajectories or preferred mutational sequences in cancer, which suggest optimized paths and mutational interdependency^17,19. While the presence of subclonal drivers predicts poor clinical outcome in patients with cancer^14,17,20, early clonal drivers that develop years before diagnosis might present opportunities for early intervention. Nevertheless, methods that use bulk sequencing inference are fundamentally limited in their abilities to resolve the phylogenetic relationships of clones, especially at low CCFs²¹ (FIG. 2a).

Defining clonal dynamics through multi-sampling.

Cancer evolution is a dynamic process, whereby clones of different fitness may reshape the genetic make-up of the tumour in the face of challenges such as treatment with effective therapeutics. Multi-sampling at different time points during clonal evolution can provide higher-resolution phylogenetic relationships even for subclones with low CCFs due to coordinated patterns of CCF fluctuations over time²² (FIG. 2b). With a greater number of sampling time points, individual subclones can be identified at a CCF significantly different from other subclones, especially if they have distinct growth dynamics²¹. Furthermore, serial sequencing not only enhances clonal decomposition but also enables clone-specific fitness measurements^21,23,24. Indeed, dense temporal sampling for either circulating leukaemia cells^{14,17,21,23,24} or circulating tumour DNA in solid malignancies, enables the measurement of clonal growth kinetics in relation to therapy through mathematical modelling^25,26. For example, chronic lymphocytic leukaemia (CLL) relapsed after chemotherapy demonstrated an increased CCF with TP53 mutations compared with the pre-treatment tumour, suggesting a greater fitness advantage of the TP53-mutated subclone in the face of therapy¹⁷. By contrast, targeted CLL therapy provided a fitness advantage to clones harbouring point mutations in the target protein (BTK)²³. In solid tumours, tracking circulating tumour DNA from patients with colorectal carcinomas treated with EGFR blockade identified the clonal dynamics of drug-resistant clone expansion upon treatment and a decline upon drug withdrawal²⁵. Notably, mathematical modelling of clonal growth enables forward and backward extrapolation to predict future clonal fractions and to estimate the number of resistant cells at treatment initiation²³. These data also pave the way for the algorithmic optimization of combination therapies based on continuous measurement of the therapy-specific fitness effects on different clones²⁷.

Another form of multi-sampling is interrogating multiple regions within a tumour to assess for intratumoural clonal spatial composition. Analogous to temporal sampling, multiregional sampling of non-small-cell lung carcinomas helped to refine clonal relationships and, thus, to improve clinical stratification²⁰. In a striking example of evolutionary selection, multiregional sampling revealed the convergent evolution of driver copy number alterations (CNAs), whereby the same driver CNA involved distinct parental alleles across the tumour²⁰. Multiregional tumour sampling also enables comparison of the primary tumour with secondary spread to metastatic foci²⁸, identifying drivers of metastases (for example, CNAs in MYC, YAP1 and MMP13 in brain-metastasized lung adenocarcinomas versus primary tumours²⁹).

Copy number alterations.

(CNAs). Changes in the number of copies of a DNA segment due to deletions or gains in the genome.

While the breadth of genome-wide data enables inference for retrospective lineage tracing, a different, focused approach has leveraged the high mutability of microsatellites in mismatch repair-deficient colorectal carcinomas. Indeed, microsatellite sites have long been appreciated as a molecular clock, providing critical insights into clonal evolution^30,31. For example, the molecular age of pre-malignant colorectal adenomas was comparable to that of adenocarcinomas, arguing against a stepwise progression from adenomas to adenocarcinomas³⁰. In a more recent study, microsatellite sites were used for lineage tracing in multiregional analyses of only 20 poly-G repeat regions, challenging the ‘sequential progression model’, which posits that metastases to distant organs arise from lymph node metastases³². This study demonstrated that, in two-thirds of cases, both lymph node and distant metastases arose directly from the primary colonic adenocarcinomas³², with clear implications for interpreting the staging of biopsy samples. Building on this experience, a similar approach was undertaken targeting microsatellite sites to demonstrate that polyclonal seeding was more frequent in lymph node metastases than in distant metastases³³. Thus, analyses of metastatic disease can further elucidate patterns of clonal spread³⁴ and enable inferences of tumour evolution migration patterns³⁵.

Retrospective lineage tracing.

Clonal architecture and/or phylogenetic reconstruction of primary samples through naturally accumulated heritable marks, such as copy number alterations, single nucleotide variants or DNA methylation, as opposed to prospective lineage tracing, in which lineage barcodes are artificially introduced into a model organism.

Molecular clock.

A method to deduce the temporal history (often in terms of number of divisions) of a cell or group of cells based on genetic or epigenetic changes that reflect time (or number of divisions).

Tracing genetic history in single cells.

Despite the additional resolution that multi-sampling provides to the clonal deconstruction of cancer evolution, resolving phylogeny at single-cell resolution is required to derive the precise clonal dynamics and evolutionary history of a tumour (FIG. 2c). Prospective lineage tracing through optical^36–38 or sequencing barcodes^39–42 has enabled in vitro or in vivo modelling of tumour evolution, including with methods capable of capturing additional cellular features such as single-cell gene expression data^43–48 (see recent reviews^49–51).

However, these methods are not applicable for retrospective lineage tracing in primary human tissues, where reliance on ‘native barcodes’ is key (FIGS 2,3). High-throughput, single-cell-targeted DNA sequencing, using a droplet-based microfluidics platform, for a panel of recurrent driver mutations has allowed the highly sensitive capture of somatic genotypes of thousands of cells to reveal subclonal shifts from diagnosis to relapse in acute myeloid leukaemia (AML)⁵². Another leading method of retrospective lineage tracing inference, applied directly to human cancers, is single-cell WGS (scWGS). scWGS enables the inference of CNAs that serve as natural barcodes through which the phylogenetic relationships between individual cells can be drawn^53–55. Early single-cell sequencing studies of breast cancers resolved distinct clones within a tumour and demonstrated that metastatic foci were seeded from a single expanded clone⁵⁴. Individual cells with non-recurrent complex CNAs provided a glimpse of the underlying genetic diversity that fuels clonal evolution⁵⁴. Recently, high-throughput scWGS approaches have been introduced, enabling the sequencing of thousands of breast cancer cells and identifying early ancestral (for example, MCL1 and MYC) versus subclonal (for example, RAD18) amplifications⁵⁶. scWGS is characterized by sparse coverage of SNVs, limiting the resolution of lineage reconstruction based on a relatively small number of large somatic CNAs. Furthermore, examples of convergent evolution in which the same CNAs are seen in both parental alleles within a single tumour, as noted above^19,57, caution against classic phylogenetic reconstructions with the infinite sites assumption (that is, that mutations used for lineage tracing do not recur and are not lost). Thus, phylogenetic methods with ‘finite sites’ models may be better suited to address aneuploid tumours⁵⁸ (the challenges of tumour phylogenetics have been reviewed elsewhere⁵⁹). Other limitations include the introduction of technical errors from whole-genome amplifications such as allelic dropouts, non-uniform coverage, and PCR errors and recombinations⁴⁹. Novel computational tools have been developed that adapt the classic phylogenetic methods for species evolution to address the technical noise of single-cell data^60,61.

Fig. 3 |. Interrogating native barcodes for retrospective lineage tracing.

Clonal architecture and/or phylogenies can be reconstructed from primary samples through naturally accumulated heritable marks, that is, ‘native barcodes’, such as copy number alterations (CNAs), single nucleotide variants (SNVs), small insertions or deletions (indels) in microsatellite repeat regions, DNA methylation changes, and mutations in mitochondrial DNA (mtDNA). Emerging and potential multi-omics technologies for lineage inference display a trade-off between throughput and lineage inference resolution (black bars). For instance, single-cell retrospective lineage tracing inference methods using CNAs (from single-cell whole-genome sequencing or single-cell RNA sequencing datasets) provide a low resolution of the underlying genetic diversity that fuels clonal evolution but can be applied to high number of cells. High-sensitivity somatic genotyping of a large number of loci may enable clonal reconstruction and a high-resolution retrospective lineage tracing with methods that interrogate mtDNA or microsatellite sites. Finally, heritable stochastic DNA methylation changes can serve as a molecular clock and therefore be exploited as native barcodes to infer phylogenetic history.

Increasing DNA input through the use of single-cell-derived colonies (that is, populations of cells derived from a single clone) obviates the technical artefacts from whole-genome amplifications. The high-resolution lineage reconstruction of single-cell colony or organoid data has been applied to normal haematopoietic development⁵⁴ and cancerous tissues⁶². The high-resolution trees paved the way for critical inferences derived directly from primary patient material of key parameters of somatic evolution such as determining the size and rate of growth of the self-renewing haematopoietic stem cell pool in a normal adult⁶³. This study also began to link clonal dynamics with lineage fates, demonstrating that the human adult haematopoietic stem cells contributed to myeloid and B cells but not to T cells in the 140 sequenced single-cell clones. While this experiment was limited in throughput, it provided a model for retracing the genetic lineage history and fate decisions of normal and cancer stem cells in primary human samples, anticipating the integration of genetic data with other known axes of variation at transcriptional and epigenetic levels.

Heterogeneity of cell states in tumour evolution

Cell state plasticity as a mediator of cancer evolution.

Intratumoural genetic variability underlies tumour heterogeneity and provides a heritable diversity that drives clonal evolution across cancer. Nevertheless, genetic mechanisms alone may not capture the full spectrum of intratumoural heterogeneity. One such variability that has long been appreciated is the plasticity of cell states within a single tumour. A classic example is the epithelial–mesenchymal transition (EMT) in epithelial tissue-derived cancers, which is reminiscent of transitions between epithelial and mesenchymal cells during embryonic development⁶⁴. EMT, a heterogeneous and dynamic disposition with intermediary or partial EMT meta-states, is governed by a complex network of transcription factors such as SNAI1, SNAI2, ZEB1, TWIST1 and other regulators⁶⁴. Clinically, EMT is associated with varying invasive and metastatic potential⁶⁵.

More broadly, cancers often recapitulate physiological developmental programmes and may thus be composed of stem-like and differentiated cell types, as has been observed in heamatopoietic^66,67, brain⁶⁸ and epithelial tumours^69,70. These observations support a hierarchical model of cancer⁷¹. While still debated in the cancer field, in this model, cancer stem-like cells are at the apex of the tumour differentiation hierarchy and reflect traits of normal stem cells such as an extensive proliferative and self-renewal capacity. For example, AML cells with the most immature phenotype demonstrate robust repopulating capacities across leukaemia subtypes regardless of distinct somatic drivers⁶⁷. Consistent with the suggestion that cancer initiates in cells with stem-like properties, a primitive stem-like programme has been shown to precede malignant transformation¹⁰. Specifically, activating an embryonic neural crest progenitor state in BRAF-mutated and TP53-mutated melanocytes resulted in the induction of melanoma¹⁰. These results indicate that somatic mutations cooperate with stem-like states for cancer initiation¹⁰. While these data provide compelling evidence for therapeutically targeting the most stem-like cells, tumour cell plasticity may complicate this strategy. Differentiated cancer cells demonstrated the capacity to de-differentiate into stem-like states^66,69,70, providing evidence for a plastic bidirectional interchange between stem and differentiated cell states in malignant populations.

Stochastic cell state transitions can also serve as mediators of cancer resistance. In vitro and mouse models have identified transcriptional persister states in rare cells that enable survival through drug insult and facilitate the acquisition of full-blown genetically driven resistance to therapy^11,12,72. Furthermore, lineage plasticity, such as the transformation of prostatic⁷³ or lung adenocarcinoma⁷⁴ into small cell carcinoma or of CLL into diffuse large B cell lymphomas in Richter transformation⁷⁵, has long been appreciated as a non-genetic mechanism of therapy resistance. This form of lineage plasticity provides a path to therapy resistance via cell-identity reprogramming that eliminates dependencies on therapeutically targeted pathways. Collectively, these data show that cancer evolution is a result of a complex interplay between genetic diversity and cancer’s ability to toggle between cell states that jointly allow cancer populations to scan the fitness landscape for superior evolutionary trajectories and roadmaps to circumvent therapeutic barriers. We thus require methodologies that integrate cell state heterogeneity with a purely genetic cancer evolution model.

Heterogeneous transcriptional states at single-cell resolution.

As transcriptional state heterogeneity is a key ingredient in tumour initiation and progression^11,12,72, single-cell RNA sequencing (scRNA-seq) of individual cells from primary tumours has emerged as a transformative technology. Studies across cancer types have revealed intratumoural cell state heterogeneity as the rule rather the exception. Heterogeneity in cell states is attributed not only to basic cellular processes, such as cell cycle, metabolic and hypoxia-induced stress states, but also to developmental programmes and clinically relevant resistance programmes^68,76,77.

scRNA-seq provided high-resolution mapping of stem-like states⁷⁸ and developmental hierarchies⁶⁸ directly in patient samples. For example, in IDH1/2-mutated oligodendrogliomas, scRNA-seq identified stem-like cancer cells with activation of neural stem cell programmes at the top of the hierarchy that branches into two distinct cellular states resembling astrocytic and oligodendrocytic lineages⁶⁸. Consistent with the cancer stem cell model⁷⁹, the expression of cell cycle-related genes was highly enriched in stem-like cells. In single-cell analyses of melanoma, a stemness signature coincided with an AXL-high drug-resistance programme to RAF and MEK inhibitors⁷⁶, directly co-mapping critical enabling properties, such as growth and resistance to therapy, to the same cells. scRNA-seq further enabled the detection of rare cells in patients who represent residual disease after therapy⁸⁰. Comparison of scRNA-seq of pre-treatment and post-treatment multiple myeloma identified residual tumour cells (often comprising a small minority of the plasma cell population) based on the known malignant signatures of pre-treated myeloma⁸⁰. Thus, scRNA-seq can distinguish rare tumour cells amid non-neoplastic cells for early disease or minimal residual disease detection and for the identification of targetable resistance programmes.

Furthermore, not only can scRNA-seq measure cell states at a given time but emerging analytic technologies can also leverage scRNA-seq data to inform the likely dynamics of cell states. Various methods have emerged that reconstruct differentiation trajectories, collectively termed pseudotime projection analyses, based on the inference of similarity gradients in transcriptional data^50,51. For example, one of the first of these techniques, Monocle, constructs minimum spanning trees to build backbones onto which single cells are mapped⁸¹. This perspective capitalizes on the fact that scRNA-seq is not dependent on sorting based on known cell surface markers and thus captures transitional states, which may be informative for cell state dynamics inference. Another distinct framework predicts future cell states by quantifying the ratio of unspliced to spliced mRNA within standard scRNA-seq to determine ‘RNA velocity’, that is, the rate of change in mRNA abundance⁸². As genes that are ‘turned on’ show a higher ratio than ‘turned off’ genes, RNA velocity can predict the transcriptional profile that the cell will assume, enabling prediction of the future cell state together with measurement of the current cell state. Both these approaches provide powerful measurements of cell state dynamics in both normal and cancerous tissues. Notably, these approaches are independent of genetic lineage tracing and thus offer the tantalizing prospect of linking cell state dynamics with genetic identity through single-cell multi-omics integration to elucidate the genetic underpinnings or decoupling of cell state dynamics in somatic evolution.

Coupling genetic lineage history with single-cell transcriptional states.

Retracing the lineage history of cancer cells with distinct phenotypic variations (for example, along developmental hierarchies or in treatment tolerance) requires the direct integration of genetic information with scRNA-seq data. For example, prospective barcoding through lentiviral transfection of patient-derived xenografts demonstrated that tumour cells with the same barcodes (that is, originating from the same cell) occupied distinct lineage-specific states, supporting the plasticity of glioblastoma cells⁷⁷. However, methods that rely on artificial lineage barcodes cannot be applied to primary human samples, which necessitate the use of naturally occurring genetic barcodes of somatic mutations (FIGS 2,3).

Experimental methods that integrate DNA and RNA expression in single cells have been developed^83–85. For example, G&T-seq can capture single-cell genomic and whole-transcriptome data by extracting and separating genomic DNA and mRNA for amplification and sequencing⁸⁵. Another method, sci-L3, improves the throughput of the dual assay using combinatorial indexing⁸⁴. These technologies share the same limitations of unimodal single-cell WGS, as discussed above, such as scalability, sparsity, PCR errors and allelic dropouts. The application of scWGS has limited value for largely euploid malignancies, for which a targeted sequencing approach may be more informative. Still, these methods provide the powerful ability to link genetic lineage history with the phenotypic states within the same cells for aneuploid tumours.

Furthermore, the rich genetic information available in the transcriptome itself has been leveraged to infer CNAs in large gains and losses by assessing imbalances in gene expression across chromosomes. This approach enabled the interrogation of the link between clonal identities and developmental states in these early landmark single-cell studies of aneuploid cancers^68,76,78. In the setting of oligodendrogliomas, clonal divergence did not align with the predominant cell states, although skewed enrichment of subclones could be identified in distinct cell states⁶⁸. In glioblastomas, specific copy number gains, such as those in EGFR, CDK4 and PDGFRA, respectively, underpin the biases in the transcriptional identities of tumour cells into astrocytic, neural-progenitor and oligodendrocyte-progenitor states⁷⁷. Advanced analytic methods have since integrated SNPs for copy-neutral loss-of-heterozygosity events in conjunction with global gene expression levels for CNAs^86,87. Application of these methods to multiple myeloma linked the subclonal deletion del(16) in the initial cancer cells with a transcriptional signature resembling relapsed extramedullary myeloma cells, demonstrating a potential link between aggressive behaviour and subclonal drivers in the original tumour⁸⁶.

scRNA-seq data have also been interrogated for SNV drivers via the Smart-seq2 protocol, which captures full-length transcripts⁸⁷. Mutation status in the oncohistone H3K27M was determined in 34% of tumour cells, with greater efficiency in the polyadenylated transcript H3F3A versus the non-polyadenylated HIST1H3B, reflecting the poly-A capture method of the Smart-seq2 platform⁸⁷. Detection was markedly improved with the addition of H3F3A mutation locus-specific primers within the Smart-seq2 procedure (97% of 44 analysed cells). Similar methods using Smart-seq2 have been developed to link genetic alterations and transcriptional states in BCR–ABL-positive chronic myeloid leukaemia stem cells to uncover signatures of persistence in the setting of treatment with tyrosine kinase inhibitors⁸⁸. Altogether, these studies demonstrate that clonal divergence maintains a complex relationship with cell-state heterogeneity, varying between tumour types. These results further highlight the need to integrate genetic identities with transcriptional states at the single-cell level at higher throughput.

Somatic genotyping in high-throughput single-cell RNA-seq.

The efficiency of somatic genotyping in the studies described above was low due to sparse coverage and the low throughput of scRNA-seq methods that profile the entire transcript (for example, Smart-seq2). Highly sensitive somatic genotyping through targeted amplification of both genomic DNA and cDNA enabled the integration of clonal structure with transcriptomic states in chronic myeloid neoplasms⁸⁹. The added advantage of genomic DNA amplification is the genotyping of lowly or non-expressed mutations. While this method was built within the Smart-seq2 protocol, its throughput was improved to thousands of cells with the development of a 384-well format that generates 3´-biased libraries⁸⁹. To measure genotype to phenotype relationships across tens of thousands of cells, the detection of SNVs has also been pursued in high-throughput droplet-based scRNA-seq (10x Genomics) by deeply sequencing the final library (~200,000 reads per cell versus the recommended minimum of ~20,000 reads per cell)⁹⁰. This approach provided a framework to dissect subclonal transcriptional identities of AML such as that of GATA2-mutated subclones⁹⁰.

Our group and others have further modified scRNA-seq platforms to enable the highly sensitive capture of somatic mutations in conjunction with whole transcriptomes in tens of thousands of single cells via targeted amplification of barcoded cDNA of the mutated locus of interest^91,92 (FIG. 4a). Leveraging a nanowell-based scRNA-seq technology, van Galen et al. performed targeted somatic genotyping in scRNA-seq in the context of AML⁹¹. These data have demonstrated that AML transcriptional identities reflecting haematopoietic differentiation states closely correlated with the underlying genetic drivers (for example, FLT3 internal tandem duplications in the primitive progenitor cell state versus FLT3 Asn841Lys in differentiated cells within the same tumour)⁹¹, analogous to the close alignment of genotype and cell state in glioblastoma described above⁷⁷. While the nanowell throughput enabled the transcriptomic profiling of thousands of cells, the efficiency of mutation locus capture was low for most targets owing to the low expression level of the targeted genes and the distance between the targeted locus and the transcript ends. However, the concurrent capture of whole-transcriptome information enabled the inference of mutation status based on transcriptional similarities between genotyped and non-genotyped cells through a Random forest classifier.

Fig. 4 |. Somatic mutations reshape differentiation topologies.

a | A schematic workflow of genotyping in single-cell RNA sequencing (scRNA-seq) by which cell state and somatic genotyping can be simultaneously captured for single cells. High-throughput digital scRNA-seq platforms (shown on the left) employ tagmentation or fragmentation for transcript-end biased cDNA short-read sequencing. Thus, loci harbouring somatic mutations are often lost. To overcome this limitation, these multi-omics techniques split the full-length cDNA for the targeted amplification of loci of interest on the one hand (shown on bottom panel with blue background) and for standard digital scRNA-seq on the other (shown on top panel with orange background). The two libraries are then intersected via shared cell barcodes analytically (not shown) to co-map somatic mutations and whole transcriptome data at single-cell resolution. b | The direct linking of somatic genotypes with whole transcriptome enables researchers to superimpose and chart two differentiation topologies within the same sample, namely the native wild type and the mutated one, thus turning the co-mingling of mutated and wild-type cells from a limitation to an advantage. The differentiation topologies (graphs) are built from scoring of stemness, pseudotime and differentiation states (that is, cell fate #1 and #2). c | By superimposing two differentiation topologies, we can identify the fitness impact of a somatic mutation within each cell state. Differential gene expression between mutant and wild-type cells can be identified (as shown by the volcano plots of differentially expressed genes). Pathway enrichment analysis of differentially regulated genes reveals activated or downregulated pathways (as shown by the annotated points on the volcano plots and heatmap, showing degree of enrichment). The differentially expressed genes and regulated pathways may vary as a function of cell state (that is, stem versus cell fate #1 versus cell fate #2). WT, wild type; MUT, mutant.

As an alternative approach, we developed Genotyping of Transcriptomes (GoT), which modifies droplet-based scRNA-seq (10x Genomics)⁹². GoT enabled highly sensitive genotyping in thousands of cells, resulting in a genotyping efficiency of ~90% for the main somatic mutation targets in the CALR driver gene. Nonetheless, the distance between the targeted locus and the transcript ends limited the sensitivity of genotyping for some targets, likely due to the inefficiency of large amplicons to cluster in standard short-read sequencing. To overcome this limitation, we introduced two complementary approaches, including long-read sequencing and circularization-GoT, in which sequential rounds of circularization and inverse PCR remove the intervening sequence between target locus and cell barcodes, producing a fragment that can be sequenced with standard short-read sequencing. We also noted that cells heterozygous for mutations may be erroneously assigned as wild type owing to partial sampling of the mRNA pool in scRNA-seq or as a result of transcriptional bursts. To overcome these potential biases, we undertook down-sampling of genotyping unique molecular identifier (UMI) counts for mutation calling and multivariate modelling with genotyping UMI. We also systematically examined the relationship of genotype-based findings to varying the minimum UMI count thresholds to assess the robustness of the findings. While GoT remains a challenge for lowly expressed genes, continued optimization of digital scRNA-seq platforms for enhanced overall RNA capture efficiencies will improve cDNA-based genotyping rates.

The ability of GoT to jointly capture cell state and genotype information at high-throughput showed that, in CALR-mutated myeloproliferative neoplasms, the predominant transcriptional identities of haematopoietic progenitor cells were uncoupled with the somatic mutation status; that is, mutant and wild-type cells were mingled throughout the haematopoietic differentiation tree, confirming that precision genotyping is necessary to distinguish mutant from wild-type cells⁹². Thus, in this context, GoT enables the overlay of the mutant differentiation trajectories onto the normal map of haematopoietic development within the same individual, eliminating potential biological confounders and technical batch effects (FIG. 4b). The direct comparison of mutant and wild-type differentiation landscapes identified that the mutant cell frequency increased with myeloid differentiation, suggesting a differential fitness advantage as a function of progenitor identity. To corroborate this result, we integrated pseudotime analysis (to provide cell state dynamics) with genetic mapping of the cells and demonstrated that mutated cells were biased towards a differentiated state compared with wild-type cells. Consistently, the mutation-related cell cycle gene expression increase was more pronounced in differentiated progenitors, such as megakaryocytic progenitors, compared with haematopoietic stem cells. Mutation-induced alterations in the unfolded protein response, NF-κB activity and JAK–STAT signalling all varied as a function of progenitor type and stage of maturation, demonstrating that the underlying cell identity constrains the impact of the driver mutation and that the resulting cancer cell phenotype is a function of the interaction between cell identity and the somatic mutation (FIG. 4c).

Such methods may be particularly helpful in studying early clonal expansions in human tissues⁹³. Recently, clonal expansions within morphologically normal tissues, resulting in somatically acquired mosaicisms, were identified throughout the body^4–6. These clonal expansions often harbour somatic mutations in known cancer drivers (for example, TP53 and NOTCH1), especially in tissues under environmental stress such as the skin, oesophagus and lung^4,5,7,8. Similarly, within the haematopoietic system, recurrent drivers of myeloid malignancies, for example, DNMT3A and TET2 mutations, have been demonstrated often at low variant allele frequencies in individuals without haematological abnormalities⁹⁴. This state, termed ‘clonal haematopoiesis’, nonetheless predisposes these individuals to an increased risk of developing blood cancers^94,95 and cardiovascular disease⁹⁶. However, a critical question remains as to the phenotypic consequences of these mutations that enable their clonal outgrowth. Thus, the application of multi-omics technologies to these otherwise normal-appearing tissues may help identify deviant transcriptional programmes that enable the earliest clonal growths in human tissue.

Expanding highly sensitive somatic genotyping to a larger number of loci may also enable clonal reconstruction and high-resolution retrospective lineage tracing coupled with cell identity and cell state information. For this purpose, any mutation, including those in mitochondrial DNA^97–99 or microsatellite sites¹⁰⁰, may serve as a lineage marker (FIGS 1d,2,3). The incorporation of these naturally occurring genetic barcodes into multi-omics single-cell sequencing may provide high-resolution phylogenies coupled with nuanced cell state data, ultimately allowing us to decipher the cell fate decisions of cancer cells.

Epigenetic plasticity in cancer evolution

Epigenetic profiles underlie cell states.

The cancer-enabling phenotypes described above, such as persistence, self-renewal and lineage plasticity, need to be inherited to contribute to the evolution of cancer to progression and resistance. However, these cell states have often been shown to be uncoupled from genetic identity. Accumulating data suggest that cell states may be encoded and propagated epigenetically^12,101–103, consistent with epigenetic encoding of cell identity in normal developmental biology. Epigenetics encompasses additional layers of heritable changes¹⁰⁴, including DNA methylation (DNAme), chromatin accessibility states and histone modifications, mediated largely through the propagation of key transcription factors¹⁰⁵. As a whole, however, epigenetic modifications result in the highly coordinated regulation of transcriptional activity, which is important for normal development and tissue specification.

In cancer, the high prevalence of mutations in epigenetic modifiers, such as TET2, DNMT3A and ASXL1 in haematopoietic neoplasms^16,106, and SWI/SNF (BAF) chromatin remodelling complexes in solid tumours¹⁰⁷, points to the significance of epigenetics in mediating tumorigenesis. In addition, recent mechanistic explorations have pointed to epigenetics as underlying transcriptional signatures of persistence and resistance^11,102. In glioblastomas, persistent tumour stem cells displayed a global reconfiguration of the repressive histone methylation (H3K27me3), with a focal increase in the active enhancer mark (H3K27ac) associated with Notch signalling and quiescence¹⁰⁸. In oestrogen receptor-positive breast cancer, KDM5 histone demethylase expression underpinned transcriptional heterogeneity and drug resistance¹⁰². Thus, a growing body of evidence points to the significance of epigenetic encoding of cell states in cancer.

Corrupted epigenetic identities enable tumour plasticity.

Epigenetic identities are faithfully propagated from the first transformed cell, allowing inference of cell of origin^109–111 akin to the faithful propagation of the genetic information of the ancestral cancer cell. Nonetheless, the epigenome also parallels the genome in undergoing diversification in the growing malignant population. Thus, just as stochastic errors in the genome result in genetic intratumoural heterogeneity, stochastic aberrations in the epigenome generate epigenetic intratumoural diversity. This understanding has emerged from measurement of DNAme at the level of the single DNA fragment. These studies, based on bulk bisulfite sequencing, have shown thousands of loci that exhibit a ‘noisy’ stochastic pattern of DNAme changes, representing large heritable cell-to-cell variation in the epigenome of somatic tissues, both normal and cancerous¹¹².

Building on these foundational studies¹¹², the examination of primary cancer tissues revealed that tumour cells exhibit DNAme diversification in parallel to genetic evolution. Sampling of multiple regions within the primary tumour, metastases and pre-malignant outgrowths in prostatic adenocarcinomas revealed that lineage tracing via DNAme heterogeneity closely mirrored the phylogenetic relationships built based on copy number genetic diversity¹¹³. Similarly, in colorectal cancer models, DNAme analysis of single-cell-derived colon cancer organoids revealed a marked DNAme heterogeneity that was propagated stably, in parallel with genetic diversification⁶². Large-scale analysis of DNAme in primary CLL samples confirmed greater intratumoural DNAme heterogeneity than normal B cells and was predictive of adverse outcome¹¹⁴.

Multiple, coordinated epigenetic layers regulate gene expression and cell identity, and early data suggest that epigenetic diversification extends beyond the level of DNAme. For example, corrupted coordination between layers of the epigenome resulted in co-mapping to the same genomic loci of typically mutually exclusive activating and repressive histone modifications in CLL, likely reflecting cell-to-cell diversity in histone modification and associated with greater transcriptional diversity¹¹⁵.

Importantly, recent evidence supports the notion that selection may operate on epigenetic diversity, akin to the well-established genetic diversity model of clonal evolution^116,117. In IDH-mutant gliomas, stochastic hypermethylation of the CTCF insulator protein-binding motifs resulted in the loss of insulation between enhancers and genes, which could then be hijacked as a mechanism for oncogenic gene activation¹¹⁶. Conversely, aberrantly restrictive states, through promoter hypermethylation or polycomb-mediated repression, can inhibit the induction of differentiation programmes, arresting the cancer cells at a proliferative state, as was shown in EZH2-mutated B cell lymphomas^117–119. Finally, corrupted epigenetic encoding through the process of epimutation (for example, stochastic DNAme changes) can lower the barriers for transition between cell states¹¹⁷ (FIG. 5). This phenomenon may underlie the increased plasticity of malignancies, undermining differentiation hierarchies and enabling processes such as de-differentiation into stem-like states^114,117. Such plastic differentiation topologies have been shown through mathematical modelling to amplify positive selection and therefore the evolutionary capacity of cancer¹²⁰. Taken together, epigenetic information emerges as a central heritable encoding of critical cell states and may thus provide an additional field of operation for the evolutionary process of diversification and selection¹¹⁵ (TABLE 1).

Fig. 5 |. An integrative model of cancer progression.

Single-cell multi-omics profiling of malignancies and clonal expansions in normal tissues may help unravel the underlying model through which rogue somatic evolutionary processes are suppressed in the multi cellular human host. Emerging data suggest that genetic constraints (that is, the time needed for the accumulation of multiple driver events)¹⁷³ may need to act in concert with other mechanisms to suppress somatic evolution. One such mechanism may involve spatial constraints, as in the case of the colonic crypts that reduce the effective population size by splitting the overall population of colonic stem cells to isolated habitats, thereby favouring drift over selection^36,174 (top panel). Greater spatial mixing with malignant progression can thus serve to amplify the selection and development of resistance to therapy¹⁴². Another such mechanism may be the complex differentiation hierarchies that suppress positive selection¹²⁰. Thus, the patterns of de-differentiation related to the relaxation of epigenetic identity barriers in cancer may have the opposite effect, serving to amplify positive selection^117,120 (bottom panel). Of note, these differentiation hierarchies may be encoded extrinsically in highly organized tissues, such as epithelial organs via cytokine gradients¹⁷⁵, or intrinsically in less spatially defined tissues, such as the bone marrow, through complex, deep epigenetic hierarchies¹⁷⁶.

Table 1 |.

The role of epigenetics in tumour evolution

Epigenetic defects in cancer	Mechanism of tumour evolution	Examples
Disrupted epigenetic modifiers	Somatic mutations in epigenetic modifiers	DNMT3A, TET2 and AXL1 mutations in haematological malignancies105; SWI/SNF (BAF) mutations in solid tumours107
	Somatic mutations in epigenetic substrates	Histone H3F3A and HIST1H3B K27M mutations in paediatric high-grade gliomas177,178
	Dysregulated expression of epigenetic modifiers	KDM5 overexpression linked with drug resistance in breast cancer102
Epigenetic diversification and/or plasticity	Changes in epimutation rate	Increased epimutation in CLL versus normal B cells121; AML subsets defined by epimutation rates179
	Suppression of differentiation	EZH2-activating mutations in DLBCL inhibit differentiation programmes118,119
	Discoordination of epigenetic layers	Discoherence of DNA methylation and histone modifications in CLL115
	Global hypomethylation linked with genomic instability	MMR-positive CRC displays de novo methylating defect whereas MMR-deficient CRC does not, suggesting two modes of chromosomal instability180
	Positive selection of epigenetic disruption	Hypermethylation of CTCF insulator protein binding motifs results in loss of insulation between enhancers and genes116

AML, acute myeloid leukaemia; CLL, chronic lymphocytic leukaemia; CRC, colorectal cancer; DLBCL, diffuse large B cell lymphoma; MMR, mismatch repair.

Epimutation.

Heritable stochastic errors in epigenetic marks (best described in DNA methylation).

Multi-omics technology links genetic, epigenetic and transcriptional information in single cells.

Novel analytical and experimental methods are under development to capture the single-cell epigenome alongside cell state readouts such as transcriptional and protein expression phenotypes¹²¹. scRNA-seq with integration of protein expression^122,123 and single-nuclei chromatin accessibility assays (ATAC-seq) have been performed in parallel on the same cancer samples (FIG. 1c), with analytic linking of transcriptional states with dysregulation of the regulatory networks as inferred from chromatin accessibility¹²⁴. In mixed lineage biphenotypic leukaemias (with markers of both myeloid and lymphoid differentiation in the same leukaemia), mapping tumour cells to the map of normal haematopoietic differentiation with both mRNA and ATAC-seq data enabled causal linking of transcription factor activation with its downstream targets (for example, RUNX1 upregulating CD69)¹²⁴. Further application of such analytically or experimentally joined dual high-throughput scRNA-seq and ATAC-seq protocols (for example, sci-CAR)¹²⁵ to cancer specimens will likely help discern precise cis-regulatory sites and target genes and identify the key regulatory networks that govern clonal evolution.

The experimental extraction of DNAme and transcriptomic information from the same cells has been achieved by modifying plate-based scRNA-seq protocols, for example, Smart-seq2, in which both RNA and DNA are respectively isolated from the same cells for whole-transcriptome and DNAme data (through bisulfite sequencing)^{85,121,126–129}. One method, scTrio-seq was applied to multiregional sampling of hepatocellular and colorectal carcinomas to retrace the genetic lineage histories of cancer cells in the context of transcriptional and methylation states^126,127. In colorectal carcinomas, methylation profiles were closely linked with genetic lineages as defined by CNAs¹²⁷. While transcriptional programmes were negatively correlated with the methylation of promoters, as expected, transcriptional states were not consistently linked with subclonal genetic identities.

Linking DNAme and whole-transcriptome data in single CLL cells revealed that the cell-to-cell methylation variation was closely associated with gene expression changes¹²¹. While the limitation of sparsity in single-cell genomics data also extends to single-cell bisulfite sequencing data (resulting in a capture rate of ~10% of the targeted methylome per cell), global changes in DNAme provide valuable insights into clonal evolution such as the ability to measure epimutation rates. Notably, heritable DNAme epimutations also serve as a molecular clock^130–134 and can therefore be exploited as native barcodes to directly infer the high-resolution phylogenetic history of tumour cells in primary CLL samples (FIGS 1b,3). Further integration of somatic genotyping in this multi-omics procedure was applied to the subclonal SF3B1 mutation, which revealed that mutated cells congregated within one clade in the DNAme-based phylogenetic tree with distinct transcriptional output, providing orthogonal validation to tree inferences and enabling the estimation of nodal age at SF3B1 mutation acquisition¹²¹. Finally, the multi-omics projection of transcriptional information directly onto the lineage tree revealed the activation of distinct pathways in different clades with therapeutic exposure¹²¹. While the application of joint single-cell epigenetic multi-omics platforms^135,136 to cancer is still nascent, these data highlight the promise of single-cell integrations of various heritable, yet plastic, dimensions of cancer cells for tracking — and ultimately predicting — clonal evolution.

Spatial dynamics and the microenvironment

Spatial dynamics as a heritable source of tumour heterogeneity.

Spatial localization represents another heritable dimension in tumour cell evolution associated with fitness, as tumour cells tend to co-localize with their parent cells in spatially constrained growth^19,20,54,137. In certain cases, for example, cells at the leading edge of a solid tumour mass are in the active cell cycle (‘boundary-bound’ growth), pointing to spatial constraints as an important heritable trait of fitness related to metastatic potential and survival^138,139. Recent data suggested that the spatial distribution of tumour cells is associated not only with fitness but also with clonal evolution¹⁴⁰. The multiregional sampling of single glands (composed of <10,000 cells) within the same tumour further revealed a high degree of clonal mixing in colorectal carcinomas in striking contrast to pre-malignant adenomas of comparable size, which demonstrated segregation of clonal identities¹⁴⁰ (FIG. 5). Consistent with these data, mathematical simulations identified the clonal mixing of tumour cells to be an early event of colorectal carcinomas¹⁴¹. Thus, analogous to the ongoing corruption of epigenetic profiles of cancer cells, the break-down of tissue architecture may be a critical feature of clonal evolution, as even minute cell dispersals increase tumour fitness and its ability to overcome therapeutic challenges¹⁴². Notably, while leukaemias, or more broadly blood cancers, are less strictly bound by spatial constraints, the disorganization of the bone marrow is a long-appreciated feature of myeloid neoplasms¹⁴³ and, thus, the distortion of tissue architecture is a shared feature of progression between solid cancers and blood malignancies.

Spatial embedding of the evolutionary process also results in varying microenvironmental interactions for the growing malignant population. Across all cancer types, the tumour ecosystem, including the immune cells, endothelial cells and stromal cells, has been demonstrated to be a critical determinant of transformation, progression and response to therapy^144–146. In a recent fascinating example, non-small-cell lung carcinomas were found to modulate neoantigen expression for immune escape through either promoter methylation or CNAs in a manner highly correlated with specific immune microenvironments¹⁴⁷. Similarly, tumour-to-immune cell interactions were observed in multiregional sampling of high-grade ovarian carcinoma¹⁴⁸, demonstrating that tumour-associated T cells negatively correlated with clonal heterogeneity, selecting for clones with immune escape mechanisms¹⁴⁸. These investigations have shown that heritable evolutionary changes (at either the genome or epigenome level) vary within the same malignant population as a function of the local interactions with the non-malignant cellular neighbours. Specifically, these studies suggest that the microenvironment is exerting selective pressures on subclones and actively moulding cancer evolution.

The tumour microenvironment at single-cell resolution.

While the above-described bulk multi-omics analyses of primary cancer samples have provided strong support for the impact of the microenvironment on the evolutionary process, the cellular complexity of the microenvironment requires single-cell analysis to provide a high-resolution map of its interactions with the tumour. In head and neck squamous cell carcinomas, through scRNA-seq, interactions between tumour cells with a partial EMT programme and the stromal compartment were identified¹⁴⁹. The complexity of the microenvironment has also been recently demonstrated with scRNA-seq, showing biases in the immune microenvironment that support immune evasion such as through expansion of T cells with exhaustion programmes⁷⁶. Indeed, single-cell omics profiling has been transformative for the fields of immunology and immuno-oncology^150,151; however, an in-depth discussion of immune cell biology is beyond the scope of this review and has been reviewed recently elsewhere^150,151.

Nonetheless, as scRNA-seq is typically performed on fully dissociated tissue, it does not preserve the architecture of the tumour cells and microenvironment. Cell-to-cell interactions may be nonetheless inferred from scRNA-seq by correlating the expression of known ligands and receptors^152,153. While analytic tools have also been developed to predict the spatial arrangement of highly structured tissue architectures^154,155, highly disordered cancer tissues may prove more challenging for inferences based on scRNA-seq of dissociated samples.

Examining tumour microenvironments with spatially aware single-cell technologies.

Spatially aware single-cell platforms may thus be particularly transformative in the study of the cancer microenvironment. For protein detection, multiplexed labelling with metal isotopes (for example, imaging mass cytometry (IMC)¹⁵⁶ and multiplexed ion beam imaging (MIBI)¹⁵⁷) or fluorophores (for example, multiplex immunofluorescence (MxIF)¹⁵⁸ and cyclic immunofluorescence (CycIF)¹⁵⁹) enable the detection of dozens of markers at cellular or subcellular resolution. The application of these methods to cancer samples revealed significant intratumoural heterogeneity, including spatially dependent diverging signalling pathways¹⁵⁸. The detection of mRNA with immunofluorescence (for example, multiplexed error robust-fluorescence in situ hybridization (MERFISH)¹⁶⁰ and sequential FISH (seqFISH)^161,162) and in situ RNA sequencing of amplified cDNA (FISSEQ and STARmap)^163,164 have increased the targets to hundreds of genes. Spatial transcriptomic platforms that utilize molecular barcodes for the detection of mRNA molecules have drastically increased the dimensionality^165,166 (FIG. 1e). These data can be analytically integrated with high-throughput scRNA-seq platforms to inform each dataset, whereby the dissociated cells from the scRNA-seq data may be remapped to their spatial positioning based on spatially patterned gene expressions^167,168. The multiplexing of protein and gene expression was achieved through extension of the IMC platform by integration of in situ hybridization of targeted mRNA molecules such that RNA to protein expression correlations could be determined¹⁶⁹. For example, in breast cancer, HER2 gene and protein expressions were highly correlated, whereas CK19 showed poor correlation between protein and corresponding gene levels. The same method identified the T cell-recruiting cytokine CXCL10 as expressed in tumour cells associated with T cells, providing insights into the tumour–immune cell interface. Computational tools have been developed in parallel to quantify contributions of gene expression or protein expression from cell-intrinsic versus environmental factors¹⁷⁰, with respect to both spatial positioning¹⁷¹ and cell-to-cell interactions¹⁷². Application of these spatially aware platforms, integrated with high-throughput single-cell multi-omics, promises to enhance our understanding of the cell-to-cell interactions and spatial constraints that enable the fitness optimization and evolution of cancer cells.

Conclusions and perspectives

Cancer evolution encompasses a complex interplay of genetic, cell state, epigenetic, spatial and microenvironmental factors. Recently, novel multi-omics technologies have begun to integrate across these genetic and non-genetic determinants of tumour evolution at the critical resolution of the single cell — the fundamental evolutionary unit.

These methods pave the way to address central questions on cancer evolution through the study of human tissue. For example, to define the factors that drive the malignant transformation versus involution of clonally expanded cells requires the ability to genotype cells for driver mutations as well as to gather information about their transcriptional and epigenetic states. Another open field of investigation in human malignancies is the assessment of lineage fate decisions of cancer stem cells that culminate from the interactions of somatic mutations with cell states. These cell states may in turn be determined by intrinsic epigenetic underpinnings and/or extrinsic cues from the environment, emphasizing the need for multi-omics single-cell data integration across modalities.

As a final perspective, single-cell multi-omics profiling of malignancies and clonal expansions in normal tissues may help unravel the underlying model through which rogue somatic evolutionary processes are suppressed in the multicellular human host (FIG. 5). Emerging data that show the pervasiveness of somatic driver mutations in normal tissue^4–6,94 suggest that genetic constraints (that is, the time needed for the accumulation of multiple driver events)¹⁷³ may need to act in concert with other mechanisms to suppress somatic evolution. One such mechanism may involve spatial constraints, as in the case of the colonic crypts that reduce the effective population size by splitting the overall population of colonic stem cells to isolated habitats, thereby favouring drift over selection^36,174. Greater spatial mixing with malignant progression can thus serve to amplify selection and the development of resistance to therapy¹⁴². Another such mechanism may be the complex differentiation hierarchies; this notion is supported by mathematical modelling of evolutionary dynamics, showing that evolutionary graphs that reflect differentiation hierarchies have an organization that suppresses positive selection¹²⁰. By contrast, the patterns of de-differentiation related to the relaxation of epigenetic identity barriers in cancer may have the opposite effect, leading to evolutionary graphs that serve to amplify positive selection^117,120. Of note, these differentiation hierarchies may be encoded extrinsically in highly organized tissues, such as epithelial organs via cytokine gradients¹⁷⁵, or intrinsically in less spatially defined tissues, such as the bone marrow, through complex, deep epigenetic hierarchies¹⁷⁶. As clonal outgrowths represent a first critical step towards circumventing this barrier to somatic evolution, they tend to affect cytokine-related mechanisms in epithelial tissues (for example, mutations related to Notch and WNT signalling) and epigenetic mechanisms in haematopoietic tissue (for example, DNAme modifier mutations). Growing clonal populations can then serve as a superior substrate for an effective evolutionary process, ultimately leading to the selection of malignant clones.

In other words, we envision that the application of single-cell multi-omics to somatically evolving human tissue will provide critical clues as to the basic system properties that discourage the trillions of somatic cells from rescinding the multi-cellular contract and optimizing their fitness at the expense of the host. Such integrative analysis, empowered by the simultaneous interrogation of the multi-faceted axes of diversity that fuel somatic evolution, may thus advance this exciting new frontier in evolutionary biology at the intersection between multi-cellular species evolution and asexual reproduction of somatic cells more akin to unicellular lifeforms.