Drug discovery is generally an inefficient process characterized by rising costs1,2, long timelines3 and high rates of attrition4. These inefficiencies are partly rooted in our limited understanding of human biology, in particular, disease-related mechanisms, actionable therapeutic targets and disease response heterogeneity5,6. The lack of sufficiently representative preclinical models, and the limitations of necessarily reductionist disease models, compound the challenges of understanding human systems.

Before single-cell (SC) approaches, cell and tissue characteristics could only be assessed in bulk and from relatively large amounts of starting material. Amplification-based techniques, such as microarrays, bulk RNA sequencing (RNA-seq) and quantitative PCR with reverse transcription (qRT–PCR)7, measured mRNA transcripts in pools of cells and could not distinguish relevant signals from heterogeneous subpopulations or rare cell types. Techniques capable of SC resolution, such as fluorescence-activated cell sorting (FACS), immunohistochemistry and cytometry by time of flight (CyTOF), were limited by the relatively small scale of testable targets and the need for a priori biological insights to enable experimental design8,9,10.

SC technologies that have been developed in the past decade (reviewed in refs. 11,12,13) have made significant inroads towards resolving some of these limitations, while at the same time being complementary to bulk applications that are still commonly used. Among the growing range of technologies, single-cell RNA sequencing (scRNA-seq; Box 1) has advanced substantially14,15 since the demonstration of whole-transcriptome profiling from a single cell in 2009 (ref. 16), and has reached the point where it is being applied in the pharmaceutical industry to investigate key questions in drug discovery and development (Fig. 1). Consequently, scRNA-seq is the focus of this article. SC technologies that extend beyond mRNA to DNA, epigenetic, proteomic and other features17 are also highlighted.

Fig. 1: How single-cell sequencing can inform decisions across the drug discovery and development pipeline.
figure 1

Single-cell technologies are being applied to answer key questions at various stages in the drug discovery and development pipeline. These applications are anticipated to increase the probability of success in the clinic by improving the quality of both the drug candidates emerging from discovery programmes and the clinical development plans for those drug candidates in stratified disease populations.

The rapid and simultaneous development of scalable plate-based and microfluidic-based methods capable of profiling large numbers of single cells has enhanced the utility of SC techniques for industrial-scale applications. Novel computational techniques and other methods (Fig. 2; Supplementary Table 1; Boxes 2 and 3) have also played a key part in leveraging SC data, supported by a growing user community that has helped to improve public data access and generate best practices. The combination of SC profiling platforms and sophisticated computational methods is driving step-change improvements in our knowledge of disease biology and pharmacology. For example, the availability of SC sequencing data for animal model systems is improving our understanding of translatability to humans18. ScRNA-seq has enabled identification of molecular pathways that allow prediction of survival19, response to therapy20, likelihood of resistance21,22 and candidacy for alternative intervention23. Further capabilities provided by SC technologies include the identification of novel cell types24 and subtypes25, the refinement of cell differentiation trajectories and the dissection of heterogeneously manifested human traits26 or constituent cell types that compose multicellular organs or tumours27.

Fig. 2: Computational methods used in single-cell data analysis for drug discovery and development.
figure 2

Representation of the computational tools and/or methods (see Supplementary Table 1 for further details and URLs for the various tools), currently used by pharmaceutical companies for data handling and to probe biological insights through cell-type annotation to reveal genotype and/or phenotype and functional assignment. B cell receptor; CNV, copy number variation; eQTL, expression quantitative trait loci; scATAC-seq, single-cell sequencing assay for transposase-accessible chromatin; scDNA-seq, single-cell DNA sequencing; scRNA-seq, single-cell RNA sequencing; SNV, single-nucleotide variant; ST, spatial transcriptomics; TCR, T cell receptor.

In this Review, we illustrate how SC technologies, primarily scRNA-seq methods, are being applied in the various steps of the drug discovery pipeline, from target identification to clinical decision-making. Ongoing challenges related to study design and data accessibility are also highlighted, as well as potential future directions for the use of SC techniques in drug discovery and development.

Applications in drug discovery and development

SC technologies can be applied throughout drug discovery and development (Fig. 1). Improved disease understanding gained through subtyping based on altered cell compositions and cell states can guide the identification of novel cellular and molecular targets. Target credentialling and validation can benefit from the use of SC sequencing in the identification of relevant preclinical models for a given disease subtype. Highly multiplexed functional genomics screens that merge CRISPR and SC sequencing (scCRISPR screening; Box 2) can enhance target credentialling throughput and augment the perturbation readouts with mechanistic information to improve target prioritization. SC sequencing technologies can provide insights on cell-type-specific compound actions, off-target effects and heterogeneous responses to inform drug candidate selection. In clinical development, these technologies can contribute by helping to identify biomarkers for patient stratification, elucidating drug mechanisms of action or resistance, or monitoring drug responses and disease progression. Opportunities to characterize and improve engineered biologics and cell therapies using SC technologies are also emerging (Box 4).

Below, we discuss representative published studies that demonstrate how SC technologies, particularly scRNA-seq approaches, can be applied in key steps in drug discovery and development, with a focus on those that are widely used in the pharmaceutical industry.

Disease understanding

As most complex diseases involve multiple cell types, SC resolution can significantly advance disease understanding. ScRNA-seq captures differences in cell-type composition and changes in cellular phenotype that are characteristic of a pathological state. Moreover, the unbiased view of scRNA-seq can detect the presence of rare cell types that drive pathobiology.

SC technologies are providing detailed knowledge of underlying disease mechanisms, enabling the investigation of novel therapeutic approaches. Although an exhaustive review is outside the scope of this article, illustrative examples for cancer, neurodegenerative diseases, inflammatory and autoimmune diseases, as well as infectious diseases are presented.


SC molecular phenotyping has been extensively used to understand cancer development. Notable examples include the application of SC technologies to identify the cell of origin or cells associated with prostate carcinogenesis, heterogeneous papillary renal cell carcinoma (pRCC) and Barrett’s oesophagus leading to oesophageal adenocarcinoma28,29,30.

ScRNA-seq has revealed extensive cellular and transcriptional cell-state diversity in cancer and enabled tracking of cancer cell heterogeneity. This has been combined with immunophenotyping techniques to provide a view of stromal–immune niches (ecosystems or ecotypes) with unique cellular composition characterizing different types of tumour. Certain ecotypes are sometimes associated with tumour initiation or progression, sensitivity or resistance to therapeutic agents or clinical outcome as demonstrated by the application of this approach to capture the heterogenicity of diffuse large B cell lymphoma, breast cancer, oesophageal squamous cell carcinoma tumours and papillary thyroid carcinoma31,32,33,34.

SC technologies such as Perturb-seq hold promise in the mapping of genotype to phenotype changes — not only for oncology but also in other diseases — by assessing the impact of rare and common human disease genetic variants. This has been applied to assess the phenotypic consequences of somatic coding variants in the oncogene KRAS and the tumour suppressor gene TP53 in an unbiased and high-throughput fashion35.

As the extensive transcriptional cell-state diversity found in cancer is often observed independently of genetic heterogeneity, many studies have investigated the epigenetic coding of malignant cell states. Understanding epigenetic mechanisms is vital as they may enable adaptation to challenging microenvironments and may contribute to therapeutic resistance. Multi-omics SC profiling (Box 2) has provided insights into intratumoural heterogeneity in glioma and identified epigenetic mechanisms that underlie gliomagenesis36,37.

Longitudinal studies provide insights into the biological mechanisms associated with tumour progression and fitness of polyclonal tumours. Most studies have been carried out using mouse models or patient-derived xenografts (PDXs). Examples of this approach include a longitudinal SC analysis of samples from a myeloma mouse model that led to the identification of the GCN2 stress response as a potential therapeutic target38, and multi-year time-series SC whole-genome sequencing (scWGS; Box 2) of breast epithelium and primary triple-negative breast cancer (TNBC) PDX, which revealed how clonal fitness dynamics was induced by TP53 mutations and cisplatin chemotherapy39.

SC studies have also improved understanding of metastasis. A Cas9-based, SC lineage tracer has been applied to study the rates, routes and drivers of metastasis in a lung cancer xenograft mouse model, revealing that metastatic capacity was heterogeneous, arising from pre-existing and heritable differences in gene expression, and uncovering a previously unknown suppressive role for KRT17 (ref. 40). This study demonstrated the power of tracing cancer progression at subclonal resolution and vast scale. Further, SC immune mapping of melanoma sentinel lymph nodes (SLNs) identified immunological changes that compromise anti-melanoma immunity and contribute to a high relapse rate41. The progressive immune dysfunction found to be associated with micro-metastasis in patients with stage I–III cutaneous melanoma may motivate new hypotheses for neoadjuvant therapy with potential to reinvigorate endogenous antitumour immunity42. A similar suppressed immune environment was observed in acral melanoma compared with that of cutaneous melanoma from non-acral skin43. Expression of multiple, therapeutically tractable immune checkpoints was observed, offering new options for clinical translation that may have been missed without SC approaches. Metastasis studies based on SC analysis of circulating tumour cells (CTCs) have also been carried out44,45. The spatial heterogeneity and the immune-evasion mechanism of CTCs in hepatocellular carcinoma (HCC) have been dissected using scRNA-seq44, identifying chemokine CCL5 as an important mediator of CTC immune evasion, and highlighting a potential anti-metastatic therapeutic strategy in HCC. Further, it was recently shown that the spread of breast cancer cells occurs predominantly during sleep. ScRNA-seq analysis of blood CTCs, which increase during rest in both patients and mouse models, revealed a marked upregulation of mitotic genes, exclusively during the resting phase, thus enabling metastasis proficiency45.

A step change in our understanding of cancer is anticipated from initiatives such as the Human Tumour Atlas Network (HTAN)46 established by the National Cancer Institute, the primary focus of which is to elucidate the evolution of cancer from its pre-malignant forms to the state of metastasis at SC and spatial resolution. HTAN will generate SC, multiparametric, longitudinal atlases and integrate them with clinical outcomes. This initiative has already resulted in studies that capture in detail tumour initiation and progression as demonstrated by the creation of a SC tumour atlas covering the transition of polyps to malignant adenocarcinoma in colorectal cancer (CRC)47.

Neurodegenerative diseases

Parkinson disease is caused by the degeneration of dopaminergic neurons in the substantia nigra48, but not all dopamine-producing neurons degenerate. SC genomic profiling of human dopamine neurons found that although there are ten transcriptionally defined dopaminergic subpopulations in the human substantia nigra, only one population selectively degenerates in Parkinson disease, and the transcriptional signature of this population is highly enriched for the expression of genes associated with Parkinson disease risk49. The vulnerability of this population of dopaminergic neurons may provide insights for potential therapeutic interventions.

A different approach was used to study somatic DNA changes in single Alzheimer disease neurons. By comparing more than 300 individual neurons from the hippocampus and the prefrontal cortex of patients with Alzheimer disease with matched controls using scWGS, genomic alterations implicating nucleotide oxidation in the impairment of neural function were identified50. This work provided a different perspective on disease evolution, suggesting that the known pathogenic mechanisms in Alzheimer disease may lead to genomic damage in neurons that can progressively impair their function.

The role of immune cells in neurodegenerative diseases is posited in many recent studies. ScRNA-seq studies of brain tissues from both healthy mice and Alzheimer disease mouse models highlight disease-associated microglia, suggesting that a cell-state-targeting strategy may benefit patients with Alzheimer disease51 (Fig. 3). In addition, SC transcriptome and T cell receptor (TCR) profiling (Box 2) has revealed T cell compartments that are activated and expanded in Parkinson disease52.

Fig. 3: Single-cell RNA sequencing in disease understanding.
figure 3

Single-cell RNA sequencing (scRNA-seq) reveals a novel microglia type in an Alzheimer disease (AD) mouse model. Unbiased clustering of single immune cells (CD45+) sorted from wild-type (WT) and AD mouse brains classified the cells into ten subpopulations, according to the expression patterns of the 500 most variable genes. The analysis thus allowed for de novo identification of rare subpopulations and revealed three microglia types: 1 (yellow), 2 (orange) and 3 (red). As the distinct microglia states of the orange and red clusters are found only in the AD model mice, they are called ‘disease-associated microglia’ (DAM). Microglia 1 cluster corresponds to homeostatic monocyte states found in both WT and AD. Differential expression analysis between DAM (microglia 3) and homeostatic microglia (microglia 1) from the AD mouse brain shows that DAMs are characterized by a significant downregulation of homeostatic markers and upregulation of several known AD risk factors. Microglia 2 is an intermediate Trem2-independent state between microglia 1 and microglia 3. t-Distributed stochastic neighbour embedding (t-SNE) map adapted with permission from ref. 51, Elsevier.

Novel SC technologies have been developed to study the brain. Examples include Patch-seq53,54 — a robust platform that combines scRNA-seq with patch clamp recording — and VINE-seq55, which is based on single-nucleus RNA sequencing (snRNA-seq). These approaches have been used to identify cell types in the neocortex that were selectively depleted in Alzheimer disease and to chart vascular and perivascular cell types at SC resolution in the human Alzheimer disease brain, respectively55,56.

Inflammatory and autoimmune diseases

ScRNA-seq was used to characterize a particular regulatory T cell present in spondyloarthritis57 and helped the discovery of cytotoxic T cells in the synovium in psoriatic arthritis. Clonal expansion of these synovial immune cells was demonstrated via complementary TCR-seq58. Differentiation of peripheral blood mononuclear cell (PBMC) samples of patients with anti-citrullinated peptide antibody-positive (ACPA+) and negative (ACPA) rheumatoid arthritis at the SC level mapped immune correlates to each of these two different rheumatoid arthritis subtypes59, while profiling of the immune compartment of skin biopsies revealed that common dermatological inflammatory diseases each have distinct T cell resident memory, innate lymphoid cell and CD8+ T cell gene signatures59,60.

In multiple sclerosis, comparing PBMC samples at SC resolution from sets of twins discordant in multiple sclerosis revealed an inflammatory shift in a monocyte cluster, together with a subset of naive helper T cells that are IL-2-hyper-responsive in the multiple sclerosis cohort61. SC techniques have also helped to explain epidemiological evidence implicating Epstein–Barr virus (EBV) as a necessary aetiological factor in multiple sclerosis62. Using single-cell B cell receptor sequencing (scBCR-seq; Box 2) of both cerebrospinal fluid and blood from patients with multiple sclerosis revealed expansion of B cell clones in multiple sclerosis that bind a similar antigen in glia (GlialCAM) and EBV (EBNA1)63.

Further studies in rheumatoid arthritis, modelling expression quantitative trait loci (eQTLs) at SC resolution in memory T cells found several autoimmune variants enriched in cell-state-dependent eQTLs64, identifying risk variants for rheumatoid arthritis enriched near the ORMDL3 and CTLA4 genes. It is important to note that eQTLs depend on the functional cell state, thus their identification is complicated in studies that aggregate cells.

Technological advancements building on SC protocols can further enhance disease understanding. For example, tetramer-associated T cell antigen receptor sequencing (TetTCR-SeqHD) helped to unravel the role of cytotoxic T cells in type 1 diabetes by combining TCR-seq readouts with cognate antigen specificity, gene expression and surface marker presence65.

Infectious diseases

A prominent example of the use of SC approaches to advance understanding of infectious diseases is in the recent study of coronavirus disease 2019 (COVID-19) to identify immune correlates of disease severity in human tissue. Comparing bronchoalveolar lavages of patients with COVID-19 of different disease severity found local immune profiles associated with disease status66. Analyses of SC transcriptome, surface proteome and T and B lymphocyte antigen receptors of PBMC samples from patients with COVID-19 found a monocytic role in platelet aggregation, circulating follicular helper T cells in mild disease and clonal expansion of cytotoxic CD8+ T cells and an increased ratio of CD8+ effector T cells to effector memory T cells in the more severe cases67. These findings indicate cellular components that might be targeted therapeutically. Similarly, scRNA-seq of circulating immune cells and readouts of metabolites in plasma of patients with COVID-19 revealed an intricate interplay between immunophenotypes and metabolic reprogramming. Emerging rare, but metabolically dominant, T cell subpopulations were found, along with a bifurcation of monocytes into two metabolically distinct subsets that correlated with disease severity68. Further, combining SC transcriptomics and SC proteomics (Box 2) with mechanistic studies found that generation of the C3a complement protein fragment by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infection drives differentiation of a CD16-expressing T cell population associated with severe COVID-19 disease outcomes69.

SC analysis of lung tissue samples collected post-mortem from patients with COVID-19 identified molecular fingerprints of hyperinflammation, alveolar epithelial cell exhaustion, vascular changes and fibrosis70. Data suggested FOXO3A suppression as a potential mechanism underlying the fibroblast-to-myofibroblast transition associated with COVID-19 pulmonary fibrosis, providing insights into potential symptomatic treatments for SARS-CoV-2. A complementary study compiling lethal COVID-19 multi-tissue SC data sets from scRNA-seq and snRNA-seq analyses identified potential disease-relevant mechanisms, such as defective alveolar type 2 differentiation, expansion of fibroblasts and putative TP63+ intrapulmonary basal-like progenitor cells in the lungs of dead patients71. A review of the SC immunology of SARS-CoV-2 infection has provided interactive and downloadable curated SC data sets72.

Other notable applications of SC technologies in infectious diseases include the study of bacterial heterogeneous clonal evolution during infection and the characterization of granulomas in tuberculosis.

Parallel sequential fluorescence in situ hybridization (Par-seqFISH) was developed to capture gene expression profiles of individual prokaryotic cells while preserving spatial context73. This technology showed heterogeneity in growing Pseudomonas aeruginosa populations and demonstrated that individual multicellular biofilms can contain coexisting but separated subpopulations with distinct physiological activities73.

Coupling sophisticated SC analyses with detailed in vivo measurements of Mycobacterium tuberculosis-associated granulomas was used to define the cellular and transcriptional properties of a successful host immune response during tuberculosis74. Lack of clearance of granulomas and persistence of M. tuberculosis was characterized by type 2 immunity and a wound-healing involvement, whereas granulomas that drove bacterial control were dominated by the presence of pro-inflammatory type 1, type 17 and cytotoxic T cells74.

Target discovery

The precision and granularity that SC technologies bring to disease understanding can not only accelerate the discovery of new drug targets, but also potentially reduce attrition by providing insights into issues that affect the likelihood that drug candidates modulating these targets will progress successfully. Below, we discuss examples that illustrate the general impact of SC technologies in target discovery, while being mindful that the terms associated with target progression, such as identification, validation, credentialling and qualification have different but overlapping meanings.

Target identification

Oncology is at the forefront of the application of SC approaches to target identification. A clear example of the use of SC analysis in the discovery of novel cell-type-specific targets is the identification of S100A4 as a novel immunotherapy target in glioblastoma, following an integrated analysis of >200,000 glioma, immune and other stromal cells from human glioma samples at the SC level. Deleting this target in non-cancer cells reprogrammed the immune landscape and significantly improved survival75. Developing strategies to directly target cancer cells remains a primary focus, and SC technologies can also provide significant benefits here. As an example, SC genomics has recently provided a map charting potential new tumour antigens76. These are ideal targets for cell-depleting therapeutic monoclonal antibodies, as has been demonstrated for haematological cancers (for example, rituximab or alemtuzumab).

SC techniques have been applied in target identification in other therapeutic areas besides oncology. Of particular interest are studies in diseases with a fibrotic component, as there are few therapeutic options currently available. For example, scRNA-seq in mice comparing healthy and ischaemic hearts identified CKAP4 as a potential target for preventing fibroblast activation and thereby reducing the risk of cardiac fibrosis77. In cardiac samples from patients with ischaemic heart disease, expression of CKAP4 positively correlated with genes known to be induced in activated cardiac fibroblasts. In human chronic kidney disease, the creation of a multi-model SC atlas facilitated the discovery of myofibroblast-specific naked cuticle homologue 2 (NKD2) as a candidate therapeutic target in kidney fibrosis78. In addition, in a mouse model of kidney fibrosis, the transcription factor RUNX1 was identified as a potential target to block myofibroblast differentiation, after further analysis of sparse single-cell sequencing assay for transposase-accessible chromatin (scATAC-seq; Box 2) data79.

Human genetic data are a key resource for target identification4. Integrating information on cell-type-specific expression with disease-associated genetic variants from genome-wide association studies (GWAS) — so-called sc-eQTL — can identify the cell types and effector genes that have a causal role in disease, providing insight into potential therapeutic approaches80. Other strategies that combine GWAS summary statistics with SC transcriptomics quantify the heritability of a gene expression signature derived from scRNA-seq data sets (capturing either a cell type or a biological process)81. Via a method called SC Linker (Box 3), novel relationships between GABAergic neurons in major depressive disorder, disease progression programmes in M cells in ulcerative colitis and a disease-specific complement cascade process in multiple sclerosis have been identified81.

Computational frameworks integrating complementary molecular information have been used extensively to prioritize potential drug targets. For example, GuiltyTargets annotates on protein–protein interaction networks with differentially expressed genes linked to a disease, learns an embedded representation and uses this to predict new targets82. The incorporation of SC data sets into these computational approaches enables the prediction of cell-specific targets. For example, a network-based approach based on SC data sets has been used to prioritize drug targets in arthritis83.

Target credentialling and validation

In target credentialling and validation, confidence in a gene target is established by acquiring and combining evidence from various sources (disease biology, target biology and tractability, genetic studies, etc.). The translational validity of study models may also be examined to better understand potential gaps between the models and the disease biology or therapeutic aim. ScRNA-seq data can inform each of these facets.

Routes to improving confidence in a target include validating functional linkages between the target and the disease biology. Gene targets, gene signatures and cell states affected by individual perturbations and their genetic interactions may all be assessed at once through a scCRISPR screen, allowing target categorization and prioritization. Traditionally, significant resources are involved in target credentialling, and so compromises are often made between the number of targets examined and the complexity and number of readouts. ScCRISPR screening alone or after a genome-wide pooled screen (Box 2) can mitigate this trade-off by allowing tens to hundreds of perturbations to be pooled and profiled at once84,85,86.

An application of this scCRISPR screening approach first involved the identification of regulators of T cell stimulation and immunosuppression using a genome-wide pooled CRISPR screen, with candidate hits followed up with functional assays and Perturb-seq to reveal affected gene programmes, leading to at least four potential antitumour targets87. More recently, the platform has been expanded to allow paired CRISPR activation (CRISPRa) and CRISPR interference (CRISPRi) screening and pooled scRNA-seq profiling, advancing the range and depth of target validation. Perturb-seq could also be performed in vivo88, allowing investigation of gene functions in multiple cell types in a physiological context.

Targets may be further credentialled and validated for their impact on disease-relevant mechanisms by using functional genomics or pharmacology studies in vitro or in vivo. Currently, readouts of these studies are usually low-dimensional, focusing on only dozens of predefined proteins or specific disease-related phenotypes89,90,91. However, coupling these studies with unbiased omics readouts can provide more granularity, allow exploration of drug mode of action (MoA) (see also next section) and even reveal any unexpected toxicity profiles. Transcriptomic readouts are often the most cost-effective and relatively straightforward to interpret, and SC transcriptomics has the additional advantage of high resolution, especially for complex models. For example, dual specificity phosphatase 6 (DUSP6) has been proposed as a potential target for inflammatory bowel disease (IBD)92 and the roles of Dusp6, which had remained unclear previously from a study using bulk RNA sequencing93, have been dissected in mice in a cell-type-specific manner using scRNA-seq94.

De-orphaning studies are typically needed if the target of the drug candidate is unknown. These studies are particularly interesting for drug combinations or bispecific treatments, because biological mechanisms that are different from those of the individual drugs may be involved. For example, scRNA-seq profiling of CD45+-enriched cells from livers of mice treated with an anti-CTLA4 immune checkpoint inhibitor (ICI), and/or the IDO1 inhibitor epacadostat showed that the combination promotes CD8+ T cell proliferation and activation, and the enrichment of an interferon-γ (IFNγ) gene signature95. Similarly, flow cytometry and CyTOF were applied to demonstrate that anti-CD47–PDL1 bispecific treatment reduced binding on red blood cells and enhanced selectivity to the tumour microenvironment (TME), compared with anti-CD47 and anti-PDL1 monotherapies or combination therapies96. ScRNA-seq enabled further exploration of the mechanism, including myeloid population reprogramming, activation of the innate immune system and T cell differentiation, which cannot be directly measured using traditional methods.

ScRNA-seq can be conveniently combined with scATAC-seq for chromatin information, DNA-barcoded antibody staining for surface and/or intracellular protein expression (such as CITE-seq/ECCITE-seq97 and INs-seq98) and is therefore useful when target modulation results in pre- and/or post-transcriptional changes (Box 2). For instance, to study ICI resistance (ICR), Perturb-seq was extended and coupled with antibody staining and TCR profiling99. This work targeted 248 genes of the ICR signature identified in a previous study22 and revealed novel ICR mechanisms including downregulation of CD58 along with known resistance mechanisms.

Preclinical studies

Selecting the appropriate models for target credentialling maximizes clinical translatability. In vitro models include cell lines, primary cells and patient-derived organoids (PDOs), the latter incorporating some elements of higher-order tissue organizational complexity. In vivo models include syngeneic models, in which murine cancer cells are isografted into genotypically similar mice, PDX in immunodeficient mice, and genetically engineered mouse models (GEMMs), which recapitulate genetic alterations crucial to human carcinogenesis. Before the advent of SC omics technologies, the relative translatability of derived research models could be assessed using bulk and/or antibody-targeted SC methods (for example, flow cytometry) capable of demonstrating that characteristics of patients or donors were, in fact, recapitulated by the research models100. SC sequencing methods expand the granularity with which model or patient fidelity can be examined by shifting assessments from wholesale pools or averages to measurements of cell-type composition, intra-tissue heterogeneity and detection of rare cell phenotypes.

It has long been suggested that therapeutic strategies that account for the cellular pathogenic diversity present in complex diseases such as cancer are more likely to be successful in patients. ScRNA-seq profiling of the Cancer Cell Line Encyclopedia (CCLE) revealed patterns of heterogeneity shared between tumour lineages and specific cell model lines, suggesting that derivative cell models are promising tools for the discovery of therapeutic strategies that are not compromised by cellular heterogeneity101.

Although cell lines are easy to manipulate and have limited associated costs, more complex biological model systems better recapitulate the cell–cell interplay and emergent functions of human physiology. Using scRNA-seq to expand and quantify the extent of this recapitulation helps to guide efforts towards the most translatable systems for preclinical development, and recent areas of focus include mouse102 and human organoids103. Human liver organoids have been shown to be highly predictive for drug-induced liver injury (DILI)104, and human PDOs derived from pancreatic duct adenocarcinoma malignant ductal cells have been assessed as a good model for the human counterpart105.

Taking model complexity a step further, SC sequencing studies of hepatoblastoma and lung adenocarcinoma have demonstrated that tumour state and heterogeneity are preserved in PDX models despite differences in TME106 and that they can help to identify heterogeneity in drug responses and likely associations with anti-drug resistance107.

Characterization of well-established GEMMs at SC resolution108 and compendiums of mouse SC transcriptomic data have facilitated the identification of genes with similar murine and human expression profiles109, ligand–receptor interactions across all cell types in a microenvironment of syngeneic mouse models110, and similarities across murine–human cell populations or subpopulations in lung cancer18 (Supplementary Fig. 2). Similarly, recent SC studies revealed mechanisms underlying chemotherapy-induced ototoxicity after comparing healthy and cisplatin-exposed mice111, as well as mechanisms of ICI-induced liver injury following comparisons of treated versus untreated mice95.

A growing number of public SC data sets, representing models of interest, healthy and diseased human donors, are enabling researchers to better assess translatability18,109,112 (Table 1).

Table 1 Examples of publicly available single-cell data sets and their applications in different phases of drug discovery

Drug screening and MoA analysis

High-throughput screening (HTS) in drug discovery is traditionally performed using coarse (cell viability or proliferation) or highly specific (marker expression) readouts. If a more unbiased phenotypic assessment is chosen, using bulk assessments such as RNA-seq assumes that all cells in the assay behave similarly. In comparison with bulk RNA-seq, SC transcriptomics offers more detailed views of the responding cell types, and the corresponding cell-type-specific changes (pathway, off-target effects, dose–response profiles), allowing for separation of confounding factors such as cell cycles. Therefore, HTS approaches have recently been combined with scRNA-seq readouts. Standard HTS tests a much larger number of compounds but typically at a single dose and under very limited biological conditions, whereas the novel HTS approaches that use SC gene expression readouts test several doses and conditions at the same time and are well adapted for drug MoA studies (Fig. 4).

Fig. 4: Single-cell high-throughput screening.
figure 4

a, Standard high-throughput screening (HTS) tests a much larger number of compounds than HTS using single cells, but typically at a single dose and a single biological condition. The most active compounds obtained by standard HTS must be further studied (for example, dose–response analysis) but finally provide hits that are the starting point for drug discovery of active and safe drugs. b, HTS using single-cell approaches allows for testing of several doses and conditions at the same time and it is mainly used for drug mode of action (MoA) studies. In the uniform manifold approximation and projection (UMAP) embeddings shown, each cell is coloured either by the type of perturbation or the perturbation dose. k, thousand; M, million; t-SNE, t-distributed stochastic neighbour embedding. Elements of part b adapted from: ref. 200, CC BY 4.0; ref. 115. © The Authors, some rights reserved; exclusive licensee AAAS.

To mitigate the costs of scRNA-seq as a readout for chemical perturbation studies and to increase its throughput, multiplexing techniques have been developed. Hundreds of compounds can now be simultaneously profiled, considering multiple doses, time points and cell types, leading to a comprehensive understanding of compound function at scale and SC resolution. Using pre-existing genetic diversity and barcode-labelled antibodies or lipids, samples originating from different experimental conditions (time points, compounds, dose) can be pooled together; techniques that are collectively called hashing. For example, MIX-seq increases throughput using single-nucleotide polymorphism (SNP)-based demultiplexing of scRNA-seq readouts of cell lines and has been used to identify treatment-induced transcriptional changes for 13 drugs on up to 99 cell lines113. Another application of this approach relied on transient transfection of cells with short oligo barcodes114. The technology was validated by first multiplexing cell samples from various species (human or mouse) and, in a subsequent experiment, by multiplexing different time exposures of a human chronic myelogenous leukaemia cell line to a drug perturbation (imatinib, a BCR–ABL-targeting drug). Multiplexing the response of this cell line to 45 drugs (mostly kinase inhibitors) revealed drug-induced differential gene expression. A recent extension of single-cell combinatorial indexing sequencing (sci-RNA-seq), called sci-Plex, introduces a precursory step for sample multiplexing by single-stranded DNA (ssDNA) oligo uptake in single nuclei. This technique has been applied to screen exposure of 188 compounds in three cancer cell lines and profiled up to 650,000 cells115. Common and dose-dependent pathways associated with HDAC inhibitors, interfering with epigenetic cellular mechanisms, across these three diverse cancer cell lines were discovered. A metabolic consequence to depletion of cellular acetyl-CoA reserves in HDAC-inhibited cells was found, providing insight into the MoA of histone deacetylase (HDAC) inhibitors.

The field of deep learning has embraced the rich and high-dimensional data sets generated by SC multiplexed perturbation experiments (see review116). These methods enable the prediction of the cellular changes induced by a drug117 or exploration of the prohibitively large combinatorial space when combining chemical perturbations (for example, compositional perturbation autoencoder (CPA)118). The latter can identify potential combination treatments from the large multiplex SC data sets generated by techniques such as sci-Plex.

SC approaches using human samples can also help to explore the MoA of drugs or vaccines. As an example, elucidating the nature of the induced immunological memory after SARS-CoV-2 vaccination from real-world evidence has complemented the preclinical and clinical studies of these vaccines. SC technologies were used to compare the immunological changes induced by natural infection, vaccine-based antigen exposure or a combination of the two. The immunological B cell response to BNT162b2 vaccination was charted using scRNA-seq and scBCR-seq (Box 2), and the effectiveness of this mRNA vaccine against emerging variants of concern was analysed119. On the basis of SC data, it was discovered that the antibody response resulting from hybrid exposure (previously infected people vaccinated with the BNT162b2 mRNA vaccine) has an increased potency for neutralization120. These findings were later proved to be clinically relevant in a much larger cohort of patients121. Regarding therapies, the RECOVERY trial established dexamethasone as an effective treatment for hospitalized patients with COVID-19 receiving oxygen or mechanical ventilation122. Subsequent SC studies unravelled the immunological components that underlie the effectiveness of dexamethasone. A prominent role for neutrophils in response to this potent corticosteroid in patients with severe COVID-19 was discovered123. These insights may thus help the development of more targeted treatment options for severe COVID-19.

Finally, SC expression profiling has also been applied to study the biological mechanisms of drug resistance at cellular resolution. Analysing SC data from pre- and multiple post-treatment time points from a lung adenocarcinoma cell line demonstrated the mechanism of acquired resistance to epidermal growth factor receptor (EGFR) tyrosine kinase inhibitors such as erlotinib in non-small-cell lung carcinoma and the existence of intracellular heterogeneity in treatment sensitivity, highlighting the importance of unbiased SC readouts124.

Biomarkers and patient stratification

In some settings, patients can be stratified into refined populations on the basis of disease prognosis or therapeutically relevant markers that predict drug response. These prognostic or predictive biomarkers are often used as eligibility criteria in clinical trials to identify patients who are more likely to have disease progression or respond to a drug, respectively (Fig. 5a).

Fig. 5: Biomarker discovery and patient stratification.
figure 5

a, Single-cell RNA sequencing or single-cell multi-omics technologies enable the identification of a predictive biomarker from a cohort of patients enrolled in an early-phase clinical study. Such a predictive biomarker can be used to identify patients who can benefit from a given treatment as a biomarker enrichment strategy. b, Single-cell analysis of immune cells from samples from patients with metastatic melanoma treated with immune checkpoint inhibitor (ICI) therapies uncovers a TCF7+ memory-like state in the cytotoxic T cell population associated with a positive outcome. t-SNE, t-distributed stochastic neighbour embedding. Elements of part b reprinted with permission from ref. 19, Elsevier.

Bulk transcriptomic signatures have been typically used to determine prognostic biomarkers in cancer, as in the case of the four consensus molecular subtypes (CMS1–4) defined by an international consortium for CRC125. However, the CMS classification has not yet proved convincingly useful in the clinic126. Bulk sequencing inherently lacks the resolution to capture crucial cell populations of CRC tumours and their complex microenvironment; and the underlying epithelial cell diversity remains unclear in the CMSs. Recently, scRNA-seq has helped to define more precise prognostic biomarkers in CRC127,128. Analysis of the transcriptomes of single cells from tumour and adjacent normal samples led to the definition of two epithelial cell groups with different intrinsic CMSs (named iCMS2 and iCMS3). Combining them with microsatellite instability and fibrosis status, a new classification called IMF has been proposed128. IMF includes five subtype classes, having distinct signalling pathways, mutational profiles and transcriptional programmes. Although promising, the value of this new classification is yet to be proved in the clinic.

ICI therapy has been successful in achieving durable responses in a subset of patients in a wide range of malignancies. However, there are still many unanswered questions around why not all patients respond to ICI therapy, and identification of predictive biomarkers for the response of ICI remains a key goal. Through these efforts, several predictive biomarkers, including tumour mutation burden (TMB), have been discovered129,130. Unfortunately, these predictive biomarkers fail to explain response to ICI for all patients. Recent SC sequencing studies have demonstrated the ability to identify new predictive biomarkers for the response or resistance to ICI. A study of CD8+ T cellular states at baseline19 revealed that responders to checkpoint inhibitors are enriched in the TCF7+CD8+ T cell state, which is also present in other indications responsive to checkpoint blockade (Fig. 5b). Beyond the conventional CD8+ T cell mediated mechanisms associated with ICI response, SC sequencing is also highlighting other cell types that shape response, such as TREM2hi macrophages, γδ T cells, CXCL9+ tumour-associated macrophages, T cell exclusion signatures and lung cancer activation module (LCAMhi) characterized by PDCD1+CXCL13+ activated T cells, IgG+ plasma cells and SPP1+ macrophages131,132,133,134,135,136. Promisingly, some of these cell types and states have been recurrent in multiple independent studies across tumour types137 and have outperformed currently used predictors such as TMB, tumour infiltrating lymphocyte (TIL) levels and PDL1 expression. In addition to scRNA-seq, there are examples of SC spatial analysis being applied to identification of potential predictive biomarkers of response. The proximity of exhausted CD8+ T cells to PDL1+ cells has been reported to predict the clinical response of combined PARP and PD1 inhibition in ovarian cancer138, while the proximity of antigen-presenting cells to stem-like CD8 T cells in intratumoural tertiary lymphoid structures has been reported to predict ICI efficacy139,140.

ScRNA-seq has also been applied to characterize chemotherapy resistance processes in cancer, as exemplified by a study in high-grade serous ovarian cancer (HGSOC). SC analysis of tissue samples collected before and after chemotherapy showed that stress-associated cancer cell populations pre-exist and are subclonally enriched during chemotherapy. The stress-associated gene signature also predicted poor prognosis in HGSOC141. In addition, scRNA-seq may be applied to predict future relapse, as seen in MLL-rearranged acute lymphoblastic leukaemia (ALL) by quantifying the proportion of cells that are identified as resistant or sensitive to treatment142. In this study, the relapse prediction outperformed the current risk stratification scheme143.

Outside oncology, SC studies are, for the first time, providing an opportunity to stratify disease into actionable subtypes. In IBD, scRNA-seq identified a cellular module called GIMATS in inflamed tissues from patients with Crohn’s disease144, consisting of IgG plasma cells, inflammatory mononuclear phagocytes, activated T cells and stromal cells. A high GIMATS score in patients was associated with failure to achieve durable remission after antitumour necrosis factor (TNF) therapy. In addition, profiling patients with ulcerative colitis and healthy individuals identified immune and stromal cells (including inflammation-associated fibroblasts) associated with resistance to anti-TNF treatment145. Furthermore, scRNA-seq analysis of PBMCs from patients with acute Kawasaki disease revealed the decreased abundance of CD16+ monocytes and downregulation of pro-inflammatory cytokines such as TNF and IL-1β in response to high-dose intravenous immunoglobulin (IVIG) therapy146. There have now also been several studies that have applied scRNA-seq approaches to diseased tissues and reported on biomarkers predictive of drug response or resistance124,131,147; however, there is still a gap in terms of understanding how well these findings translate into the clinic.

Although these SC studies are limited in terms of patient numbers, conditions and samples, methods such as cell-type deconvolution allow them to be used to complement existing bulk RNA-seq studies that typically have more mature response and outcome data22.

Monitoring of drug response and disease progression

Clinical monitoring of both disease progression and response to therapy with SC sequencing approaches is starting to influence clinical decision-making. The field of oncology has taken the lead in this area. The concept of minimal residual disease (MRD) as a metric to indicate remaining cancer cells during or after completing therapy has been a central tenet in measuring drug response. For example, patients with acute myeloid leukaemia (AML) often harbour multiple subclones, each with complex molecular abnormalities148. Clinical practice today defines complete remission as <5% blasts detected by morphological evaluation in the bone marrow without an assessment of subclonal molecular abnormalities or their evolution during therapy. Evidence is mounting that MRD assessments below this 5% threshold are a relapse risk factor and could therefore guide treatment decisions149. MRD assessment with SC mutational profiling (in contrast to more traditional MRD methods) allows for subclonal assessment at lower detection limits and for analysis of subclonal evolution throughout treatment150. SC mutational profiling improved sensitivity and specificity of MRD detection and was also able to identify relapse-causing resistant clones.

The relapse risk associated with MRD is partially explained by the presence of persister cells that are induced in response to treatment. This type of drug resistance is often driven by non-genetic adaptive mechanisms, although these are poorly understood. To study the rare and transiently resistant persister cells, a high-complexity lentiviral barcode library called Watermelon was developed to simultaneously trace the clonal lineage, proliferation status and transcriptional profile of individual cells during drug treatment151 (Supplementary Fig. 3). This approach identified rare cancerous persister lineages that are preferentially poised to proliferate under drug pressure and found that upregulation of antioxidant gene programmes and a metabolic shift to fatty acid oxidation are associated with persister proliferative capacity. Obstructing oxidative stress or rewiring of the metabolic programme of these cells alters their proportion. In human tumours, programmes associated with cycling persisters are induced in response to multiple targeted therapies. Persister cell states should thus be targeted to delay or even prevent cancer recurrence. In addition, the PERSIST-SEQ consortium ( was initiated to create a SC atlas of persister cells to improve the understanding of therapeutic resistance in cancer. Similarly, initiatives like HTAN46 could potentially contribute to consistent mapping of persister cell states among the set of clinical transitions of adult and paediatric malignancies when exploring therapeutic resistance. A study in TNBC showed that treatment-resistant clones originated from pre-existing cancer cells. By combining bulk whole-exome sequencing (WES) with SC transcriptomics, it was demonstrated that some of these adaptive changes were not induced by somatic mutations but were characterized by transcriptional reprogramming of these cells152.

As discussed previously, ICI therapy is a promising new therapeutic modality for some cancer patients, and understanding which subpopulation benefits from this treatment option is important. In addition, monitoring of pharmacodynamic changes and closely following response to ICI treatment from a molecular level are required for better patient selection and overall treatment outcome improvement. Mechanisms by which PD1/PDL1 blockade either revives pre-existing TILs or recruits novel T cells have been examined recently with the application of paired scRNA-seq and scTCR-seq on site-matched tumours from patients with basal or squamous cell carcinoma before and after anti-PD1 therapy153. Analysis of TCR clones and their transcriptional phenotypes revealed that drug response is driven by the expansion of novel T cell clones not previously observed in the same tumour, probably derived from a distinct repertoire of T cell clones that recently migrated into the tumour. Another SC study154 showed that CXCL13+CD8+ T cells were expanded in response to PDL1 treatment and identified a circulating T cell subtype that shared higher levels of TCR clones with tumour CXCL13+CD8+ T cells. The number of T cell clonotypes induced during early treatment provides a good proxy for future treatment success. This metric was used to identify SC changes induced by successful ICI treatment during a window of opportunity study155. These findings have also been recently confirmed in a multiple tumour type study155,156, thereby not only providing insight into the PD1/PDL1 blockade MoA, but also suggesting that liquid biopsies that sample TCR repertoire and identify clonal changes upon treatment may provide an actionable pharmacodynamic response.

Current challenges

Several challenges remain for industry to harness the transformational capabilities of scRNA-seq technologies, which will require changes to infrastructure and ways of working. Moreover, as the generation of scRNA-seq data in the public domain has outpaced that of internal efforts from any single pharmaceutical company, effective integration of all relevant scRNA-seq data is particularly challenging. In addition, owing in part to sample requirements and cost of scRNA-seq data generation, it is not likely to quickly replace bulk molecular profiling of early discovery or clinical samples, and so effective integration of scRNA-seq and bulk molecular profiling data is also needed.

Study design and implementation

Standardized design and implementation of SC experiments is still in its infancy. Although SC resolution has the potential to improve understanding of cell states and subsets of rare populations, discerning a cell type precisely and consistently across different experiments for rare cell populations is difficult, especially when fine distinctions guide cell-type identification. A uniform analysis pipeline, together with consistent methodology and vocabulary, are prerequisites to addressing this. Multi-omics approaches, by providing orthogonal indicators including cell surface and intracellular proteins or epigenetic markers, can further refine cell-state delineation but also imply new analysis challenges157,158,159,160,161.

SC sequencing throughput is primarily limited by the cost, but also by sample processing and computation capacity. For scRNA-seq, tissue samples need to be dissociated and processed immediately after collection to preserve high RNA quality145,162. SC library preparation poses a challenge to clinical sites where personnel may not necessarily be trained to handle sample preparation and specialized equipment. Sample quality and consistency are also hard to control, especially in large-scale multi-site clinical studies. Technology development of single-nucleus sequencing on cryopreserved or even formalin-fixed paraffin embedded (FFPE) samples provides a potential solution to this issue, allowing clinical sites to bank biopsies for later processing163,164,165. This technology also makes it possible to take advantage of banked samples from previous studies. However, care should be taken when selecting technologies as each has its own limitations166,167.

An online calculator ( can help to determine the number of cells to be interrogated in a sample given prior assumptions on the diversity and relative composition of cells in the biology under investigation. Guidance in deciding which protocol to use or how deeply to sequence the collected cells has been provided168. In addition, design considerations for setting up longitudinal SC experiments have been reported169.

Design of SC experiments presents unique opportunities and challenges compared with bulk transcriptomics assays. On one hand, the availability of many SC samples within the experiment allows application of machine learning approaches that may be inappropriate for the typically powered bulk experiment. However, the results may have limited generalizability, owing to the low number of biological samples used to generate the SC data. On the other hand, compared with bulk RNA-seq, scRNA-seq is more expensive, and samples are more difficult to access and process. Bulk techniques have been optimized to deal with poor-quality RNA, frozen samples and even FFPE samples, whereas SC technology is only recently expanding beyond the use of fresh tissue. Enabling technologies, such as cryopreservation170 or snRNA-seq165, are still undergoing considerable optimization. A balance in complexity and budget can be achieved by combining bulk and scRNA-seq in a single experiment. SC samples can be used to computationally deconvolute cell-type abundance from bulk samples collected using an experimental set-up that favours fewer SC and more bulk sequenced samples. In addition, leveraging publicly available SC data sets can mitigate budget constraints.

Data accessibility

The current organization of public SC data generally falls short of the FAIR principles for data stewardship in several aspects171, in particular with respect to data accessibility. Ongoing cataloguing efforts (for example, the BROAD Single Cell Portal —, spreadsheet of data set metadata172) and international collaborations to generate healthy reference databases (for example, Human Cell Landscape (HCL)173, Tabula Sapiens174 provide an initial entry point for discovery of data sets. However, none of these initiatives is comprehensive, resulting in the need to manually search the publication databases (for example, PubMed) and omics repositories (for example, GEO). Without uniform metadata across these databases, the search strategy must also be varied between various resources to ensure completeness.

Within a given organization, some data are likely to be accessible only to a subset of analysts. Tracking designations flagging permissible data use in the metadata versus in an external system each present different barriers related to internal risk management and compliance, as well as to scientists and analysts seeking to use those data or to build on previously completed analyses. For public data sets, similar issues exist — data access might be restricted behind security portals, as in the case of dbGaP and EGA, because of privacy laws, contractual considerations or the sensitivity of human data. This is especially true for raw reads from full transcript protocols such as Smart-Seq2 and is equally likely to be applicable to internally generated data.

Data interoperability and reusability

Most SC transcriptomics data sets of published work are made available publicly. Unfortunately, there is considerable variability in the format and layout of data. Digital formats for expression or count matrices (scRNA-seq) and experimental metadata are not standardized175. In addition, lack of comprehensive sample metadata is a common problem. Therefore, the interoperability of these data sets is limited.

Moreover, the non-uniformity of data processing, including the quality control (QC), cell-type annotation and the lack of a well-defined cell-type nomenclature (that is, either ‘flat’ or ‘shallow’ nomenclatures are used, with different levels of detail across studies), necessitates reprocessing of the data sets to interrogate them for new research questions.

Currently, the pharmaceutical industry either resorts to in-house curation efforts to augment their internal library of SC data sets with uniformly processed public entries and/or engages with external vendors for this service (see Box 5 for an example from a company and Box 6 for general use of SC public data sets by industry). The maturity, range and type of services provided by vendors varies greatly, from project-based and ad hoc curation of a small set of data sets, to platforms that house an industrialized pipeline, SC web viewers and exploratory research environments. The extent of the curation is also highly variable: some vendors start from raw sequence reads, whereas others reuse published gene expression matrices and cell-type annotations. Another big challenge to overcome is technical variations in SC data introduced by multiple factors such as laboratories and conditions. It is crucial to properly handle technical variations in the data integration and curation step (see Box 3 for computational tools for batch-effect correction and data integration). However, these approaches are expensive and time-consuming. To avoid duplication of work across companies and academic institutions, the community could benefit from collaboratively adopting and developing common standards. The academic sector has clearly paved the way by showing the value generated by creating repositories of uniformly processed and/or integrated data sets (Table 1).

Direct exploration of published data sets is being facilitated by both online viewers hosted by some researchers and general purpose scRNA-seq platforms that provide more elaborate exploratory analysis capabilities. Researcher-hosted viewers are useful to quickly check the expression of a gene but do not support maximal reuse of published data sets. Even the most advanced viewers, such as Cellxgene176 limit the scope of interrogation to selected use cases. These viewers are not a durable resource and often rely on temporary web hosting and are therefore more appropriate for accessing the data immediately after publication. By contrast, general purpose platforms such as Cumulus/Pegasus, which runs on Terra.Bio177, provide a cloud infrastructure tailored to run scRNA-seq bioinformatics pipelines and a notebook system for exploratory analysis. The EMBL-EBI Single Cell Expression Atlas (SCEA)178 has built a uniform pipeline for transcript quantification, quality control and cell-type annotation, and it runs on the browser-based Galaxy platform179. A final example, the HCA Data Coordination Platform (DCP), is a public, cloud-based platform on which scientists can share, organize and interrogate SC data.

Conclusions and future perspectives

Most complex diseases for which treatment remains elusive have a multicellular aetiology, and a SC perspective could be crucial in advancing our understanding and ability to select the most therapeutically impactful cellular or molecular targets. SC protocols combined with sophisticated multiplex strategies have increased the scale and resolution at which assays can be performed. In addition, SC profiling of commonly used preclinical models enables researchers to select the model that best recapitulates essential human pathobiology. Interrogating human samples at cellular resolution can help to advance personalized medicine, by expediting the discovery of new biomarkers to help stratify patients on the basis of prognosis or prediction of treatment effect. A longitudinal SC view on diseased tissues during treatment can also provide physicians with a more direct and mechanistic view on response to treatment.

Having established the more mature scRNA-seq-based methods for routine use in industry, effort is increasingly focused on adopting other methods such as SC proteomics and spatial omics technologies, as industrial SC capabilities are expanded. As the core technologies become standardized, the requisite skills become more widely available and the costs fall, the rate of SC data generation is likely to continue to accelerate180,181.

As the technical challenges involved in SC data generation, curation and access are addressed, new opportunities are emerging. For example, upstream of target discovery, the focus is already shifting from the discovery of novel cell types and cellular marker genes towards hypothesis generation rooted in deeper understanding of cellular mechanisms. The integration of additional data types supports this shift as omics and other multiparametric data enhance the granularity of insight into the cellular environment. For example, mapping genetic cues on disease provided by GWAS on SC profiles from scRNA-seq experiments can help to elucidate cellular phenotypes linked to complex diseases81,182.

With the increasing maturity of spatial profiling technologies, we are beginning to better understand human tissue organization and microenvironment niches. Spatial profiling enables cell types to be accurately counted and localized within the broader tissue architecture. In addition, it facilitates the mapping of intricate auto- and paracrine interactions between cell types within a tissue. However, the resolution of the most unbiased and comprehensive approaches (for example, 10X Visium) remains supracellular. We expect that such approaches will evolve to provide SC resolution, and thus complement and extend the pipeline of methods applicable to intercellular interaction discovery from scRNA-seq (for example, CellPhoneDB183). Moreover, advances in spatial profiling are lining up with the recent progress made in digital pathology. Combined with automated feature extraction and molecular classification of digitized pathology images via deep learning techniques184, orthogonal informational cues assayed via sequencing or multiplex imaging technologies will enable researchers to develop a deeper knowledge of the complex biology involved in some diseases.

Given the enormous technical, computational and scientific complexities involved in SC data generation and translating those data into benefits to patients, collaboration has a key role. This is clearly demonstrated by the Accelerating Medicines Partnership and LifeTime initiatives, and the rapid growth of SC research around SARS-CoV-2 (ref. 185). LifeTime established a special task force to study COVID-19 and to identify SC-based biomarkers and novel modalities. In this case, HCA and LifeTime created a common framework for sharing knowledge, data, tools and other resources. As the scale and complexity of SC data and our understanding of human biology continue to deepen, collaborative efforts between academia and industry will be increasingly vital to realize the transformational potential of SC technologies.