Moving Genomics into Tissues

Abstract

The formation and maintenance of tissue integrity requires complex, coordinated activities by thousands of genes. Until recently, transcript levels could only be quantified for a few genes in tissues, but advances in DNA sequencing, oligonucleotide synthesis, and fluorescence microscopy have enabled the invention of a suite of spatial transcriptomics technologies capable of measuring the expression of many, or all, genes in situ. These technologies will help biologists address three classes of scientific problems: (1) determining the cell-type architecture of tissues; (2) querying cell-cell interactions; and (3) monitoring molecular interactions between tissue components. Here, we review the technological progress in two main classes of spatial transcriptomics methods: those that utilize next-generation sequencing for gene detection, and those that employ imaging-based detection approaches. In particular, for both approaches, we outline standard experiments and analyses that will help compare across individual technologies and accelerate further development in this exciting field. Spatial transcriptomics represents the beginning of a larger scale technological effort to move genomic measurements directly into tissues, which will catalyze discoveries in a wide range of scientific disciplines.

Introduction

New technologies to measure biomolecules have been critical drivers of biological progress. When measuring biomolecules, researchers have historically faced a key trade-off in selecting experimental methodologies. On the one hand, “-omics” tools generate broad, often comprehensive measurements of many biomolecules from a purified specimen. On the other hand, a suite of targeted tools–like immunostaining or in situ hybridization–can localize a much smaller number of specific molecules within intact cells and tissues. Research projects therefore often have two phases: a researcher first formulates a hypothesis using an -omics technology, then performs targeted, hypothesis-driven work to characterize the role of specific genes or proteins within intact tissues of interest.

This historical bifurcation in methodologies is being upended by the recent, rapid development of spatial transcriptomics (ST) technologies. These tools enable the transcriptome-wide quantification of RNAs within intact tissue sections. Broadly speaking, we see ST as best suited to answering three kinds of biological questions (Fig. 1A). Most fundamentally, ST technologies are capable of elucidating the cell-type composition of tissues. Cell type definitions are frequently imported from large-scale single cell RNA-seq or epigenetics datasets, and then computationally projected onto ST datasets to learn their spatial distributions, but definitions can also be generated from the ST data themselves. To date, this has been the most commonly used application of ST in published studies. Using ST, compositional atlases have been generated for a wide variety of tissues, including nervous system tissues^1–5human kidney^6–8, heart9, testes¹⁰ and lung¹¹.

Figure 1. Usage of spatial transcriptomics in biological experimentation.

A) Three primary classes of biological questions addressed by spatial transcriptomics. Colored hexagons represent different types of cells indicated by ɑ, ꞵ, and ɣ in the corresponding model equations. B) Summary of the two classes of ST technology: sequencing-based and imaging-based methods. The input to ST technologies are generally tissue sections, and the output is generally a digital gene expression matrix and a table of spatial coordinates for each spot. Detailed descriptions of individual methods within these two classes are shown in Figure 2 and Figure 3.

A second kind of question relates to cellular interactions: which rules and patterns define how individual cell types spatially covary with each other? For example, an ST study of mouse visual cortex⁴ found a predilection for inhibitory neuron subtypes to be more spatially proximal to each other than would be expected if they were distributed randomly. Similar kinds of proximity analyses have been used to define gene expression patterns upregulated by amyloid plaques found in Alzheimer’s disease¹², and to characterize histopathological responses to traumatic brain injury¹³.

Finally, because ST often delivers transcriptome-wide data on gene expression in situ, it can help to elucidate molecular interactions between tissue components. By defining ligand-receptor pairs amongst cell types at different spatial proximities, we can determine if, and how, individual cell types are communicating with each other. Such analyses should be immensely clarifying for many cell non-autonomous phenomena, including interactions between tumors and their surrounding environment^14,15, immune infiltrates in tissues, or the establishment of developmental gradients¹⁶.

To understand how current and future technologies can improve our capacity for addressing these problems, we describe in detail the existing ST technologies in the sections below. In particular, we emphasize the importance to the field of adopting experiments that validate technological claims, and enable comparison across platforms. The design of quality control experiments with well-defined conditions that can be easily replicated across labs and methodologies can be extremely catalytic for technology development. For example, the mixed-species experiment used in single cell technology development^17,18–in which cells from well-established cell lines, derived from different organisms, are mixed together prior to the assay–has become a widely accepted benchmark for validating the sensitivity and specificity of new methodologies. The field of ST could similarly benefit from defining a standard set of reference experiments, by which to gauge the performance of the key experimental parameters.

Spatial transcriptomics: born from two distinct domains of biotechnology

Spatial transcriptomics emerged from parallel, synchronous efforts by two distinct groups of technologists (Fig. 1B). In genomics, advances in massively parallel DNA sequencing, molecular biology, DNA-based molecular barcoding, and computational analysis made possible the measurement of gene expression–as well as more recently, epigenetic regulation–within many individual cells. These strategies and concepts were creatively adapted to capture RNA locally from intact tissue sections on a pixelated, DNA-barcoded surface, and read out their gene identities using next-generation sequencing¹⁹. We term this family of technologies “sequencing-based spatial transcriptomics” (sST). In parallel, technologists working on microscopy techniques developed several strategies for simultaneously detecting the presence of many mRNA transcripts within tissue using fluorescence in situ hybridization (FISH)^20,21 or direct in situ sequencing^22,23. We call such technologies “imaging-based spatial transcriptomics” (iST). Both of these classes of technology deliver similar, but still complementary, measurements of gene expression in situ. We will describe the major instantiations of sST and iST emphasizing key metrics and experimental frameworks for their characterization and validation.

Sequencing-based spatial transcriptomics

Sequencing-based ST technologies typically begin with the construction of a spatially indexed surface (Fig. 2A), in which each pixel contains a barcoded DNA primer that uniquely marks a pixel’s location in two-dimensional space. Tissue is then placed atop the surface, and the resident mRNA is brought into contact with the primer, either by diffusion of RNA from the tissue to the surface, or diffusion of the barcoded primers into the tissue. Typically, they utilize primers with polyT sequences on their 3’ termini to capture, transcriptome-wide, mRNAs across the transcriptome.

Figure 2. Sequencing-based spatial transcriptomics methodology and characterization.

A) Workflow of sequencing-based spatial transcriptomics methods. Strategies for the fabrication of indexed pixel surfaces wherein DNA barcoded primers are associated with spatial localization. Microarray-based strategies utilize deterministic DNA barcodes printed on glass slides. Bead-based strategies utilize DNA-conjugated beads with diverse, clonal barcodes whose spatial locations are ascertained. Nanoball- or polony-based strategies utilize local clonal amplification to generate clusters of clonally barcoded primers. Microfluidic barcoding utilizes channels to deterministically deliver row and column barcodes to a tissue, forming a 2D grid. B) Steps of sequencing library generation downstream of surface indexing. Basic computational processing of the data results in a digital gene expression matrix, with a paired table of coordinates for each pixel. C) Top, sensitivity of selected sST technologies ^{16,19,24–26,89,90}, represented by the average number of UMI counts per normalized 10μm spot, and colored by the tissue type; bottom, the reported spot size for each technology is shown in gray, and the binned resolution used for analysis shown in purple. D) Proposed approaches for benchmarking sequencing-based ST technologies. Representative regions of interest (outlined areas in images at left) are selected from tissues and different quality control metrics are applied to the selected region to quantify the sensitivity of RNA capture, and the spatial resolution. A marker gene specifically expressed in selected regions was chosen. For capture sensitivity (bar plot, top right), the total counts of the marker gene are summed within the selected region, and compared to the counts ascertained by an smFISH reference assay. For resolution (density plot, bottom right), the intensity of the marker gene expression is quantified across a dimension of the feature, and the feature thickness is compared to the full width half maximum of the profile. Methods A, B, and C represent theoretical ST technologies.

Spatially indexed pixels have been generated in a variety of ways. The first strategy, employed by Stahl and colleagues and used in the Visium technology sold by 10X Genomics, utilizes a microarrayer spotting robot to deliver a unique barcode to a fixed, known location on the surface of a slide¹⁹. These spots are 50–200 microns in size (though upcoming Visium products will reportedly have smaller pixels), and are separated by a similar amount of white space to prevent mixing during liquid handling. A second strategy utilizes solid microparticles for spatial barcoding. In Slide-seq¹³, 10-micron diameter beads are used as the solid support for oligonucleotide synthesis, in which the bead barcode is created by split-pool cycles, a process first developed for single-cell barcoding¹⁷. The beads are fabricated into a tightly packed uniform monolayer on a slide, and the locations of each barcode ascertained by in situ sequencing. Alternative strategies–either for the method of affixing barcodes to beads or for DNA synthesis–enable pixel size to be reduced to the 1–5 micron range²⁴. A third strategy is to locally amplify unique barcode sequences by rolling circle amplification (RCA) or bridge amplification. In Stereo-seq, DNA nanoballs are generated by rolling circle amplification that span ~200 nm in diameter, with 500–715 nm center-to-center spacing, and each rolony barcode sequenced. A polyT capture sequence is then ligated onto the barcoded nanoballs, to enable capture of released RNA. In Seq-scope²⁵, barcoding is accomplished by local bridge amplification of DNA randomers directly onto an Illumina sequencing flow cell to create polonies that are 0.5–1 micron in diameter. Pixels can also be formed combinatorially: in DBiT-seq²⁶, microfluidic channels deliver barcoded reverse transcription primers to the RNA within tissue; the channel apparatus is then rotated 90 degrees to deliver a second set of primers in the orthogonal direction that are ligated in situ, creating a paired barcoding scheme for recovering two-dimensional coordinates.

Once surface barcoding is complete, the downstream workflows of these technologies are remarkably convergent. In most technologies, tissue is placed in contact with the barcoded surface, and mRNA diffuses to the barcoded primers (with the exception of DBiT-seq, in which primers are diffused into the tissue) (Fig. 2B). Reverse transcription, cDNA amplification, and short-read sequencing generate reads that contain mRNA fragments for transcriptome alignment, paired with barcode sequences that are matched back to the pixel white lists. The result is the spatial localization of each detected transcript to each pixel.

Quality control.

In sST, the two most important quality parameters are: (i) the mRNA capture sensitivity per unit area; and (ii) the spatial accuracy of mRNA detection, which can be reduced by simple lateral diffusion or more complex technical artifacts. In addition, there are several additional parameters for characterizing and comparing sequencing-based ST technologies. One is the sequencing efficiency: the amount of DNA sequencing required to ascertain a particular number of mRNA molecules. Another is the spatial area covered by the technology–and its flexibility to accommodate tissues of different sizes, compositions, and shapes–which can be quite germane to a technology’s applicability to specific biological problems.

The capture sensitivity of a technology is highly influenced by the specific molecular and cytoarchitectural features of the tissue being assayed. Tissues can vary by several orders of magnitude in how much RNA can be captured per unit area, because of differences in RNase content, cell density, extracellular matrix composition and other features. To date, the performance of new technologies has not been assayed with a set of consistent, standardized tissues (Fig. 2C shows select sensitivity and resolution metrics across a variety of technologies and tissue types). To properly characterize, benchmark, and compare new technologies and improvements, it will be highly beneficial for the field to establish a set of reference tissues with well-defined histological structures that test different technological challenges (Fig. 2D). Stahl et al validated their technology with the adult mouse olfactory bulb, which has been subsequently utilized by several other sST technologies for validation as well. The olfactory bulb is composed of five discrete layers, each with well-known molecular markers, making it a useful model for technology validation. In addition, the use of mouse embryos is reasonably common(Liu et al. 2020; Stickels et al. 2021; Chen et al. 2022)^16,26,27–though with some variation in the exact age of the specimens used–because they are compact, harbor extensive cytoarchitectural variation, and have known spatial axes of transcriptional variation. We propose the additional inclusion of an E12-E14 mouse embryo, sectioned sagittally at a position along the medial-lateral axis that includes the eye . The eye is clearly identifiable, even by non-experts, and is a symmetric structure with clear histological layers, making direct comparisons across experimental replicates and regimes more straightforward. The consistent usage of these two standardized tissue types would be enormously helpful to overall technology development goals in the field.

To quantify mRNA capture, a histological structure can be segmented from an ST-derived pseudoimage, generated from shading individual pixels by the number of UMIs, or by plotting the intensity of a metagene that correlates with the chosen histological structure¹³ (Fig. 2D). Once segmented, counts of individual genes are summed within the area; smFISH performed on the same genes, in the same histological structure, provides a direct, rigorous comparison of a technology’s sensitivity. If the histological structure is composed of a uniform cell type, it is also possible to compute an average expression per cell (by counting cells in the feature), for direct comparison to single cell data.

From this same analysis, the spatial resolution of the technology can also be assessed. A dimension of the segmented histological structure can be extracted by measuring the distance between full-width half maximum of intensity. That distance is then compared to the same dimension measured with an image of the same feature, generated by histological or smFISH staining. This provides a quantitative assessment of lateral diffusion of transcripts from their source within the tissue. A similar approach can be taken–with the same data–to quantify false-positive noise (although this is less commonly a problem in sequencing-based ST technologies): individual genes known to be excluded from specific histological structures are quantified within that excluded feature in the ST dataset, and compared to counts of that gene from smFISH data. These simple analyses, if performed on widely available, agreed-upon tissue samples, provide a common language for characterizing and comparing ST technologies.

Imaging-based spatial transcriptomics

In parallel to the development of sequencing-based ST technologies, there has been an explosion of imaging-based ST approaches (iST). These approaches have primarily been driven by advances in three fields – oligonucleotide synthesis²⁸, fluorescence microscopy, and single-cell transcriptomics. Recent advances in oligonucleotide synthesis now enables specific synthesis of 10⁵–10⁶ individual oligonucleotide sequences in a pooled fashion, critical for generation of barcoded hybridization probes. Recent developments in sensitive sCMOS detectors^29,30 and organic fluorophores now enable sensitive, high-throughput detection of labeled RNAs in cells and tissues. Lastly, comprehensive single-cell atlases allow selection of informative RNA subsets for labeling.

In Fig. 3, we outline the broad concepts behind iST. In iST, RNA molecules are specifically tagged with fluorescent probes by complementary hybridization. These probes are then imaged via fluorescence microscopy. While fluorescence-based RNA detection in situ has been widely used for more than two decades, spectral limitations prevent simultaneous imaging of significantly more than ~5 distinct organic fluorescent molecules. To overcome this limitation, recent imaging based ST approaches generally utilize multiple sequential imaging rounds and combinatorial strategies for detection of transcripts. Thus, a specific iST approach is largely defined by: i) the detection modality (how are RNAs molecules labeled), and ii) the multiplexing approach (how multiple RNA transcripts are detected across sequential imaging rounds).

Figure 3. Imaging-based spatial transcriptomics methodology and characterization.

A) Depiction of the three fundamental steps in imaging-based spatial transcriptomics. Targeting chemistry summarizes how the target mRNA is labeled; black lines represent mRNA molecules and blue lines indicate oligonucleotide probes. Encoding summarizes two strategies for gene encoding to enable multiplexing. Linear encoding labels different mRNAs in each imaging round. Exponential encoding labels each mRNA in multiple imaging rounds. Image processing highlights major steps in downstream image processing after data collection. First, samples are registered between imaging rounds to the same coordinate space. Then, spots corresponding to single RNA molecules are identified, and assigned to imaging rounds. Lastly, gene identity is decoded for each spot based on the imaging rounds. B) Summary of different encoding and targeting chemistry and key methods in each category. C) Summary of the detection efficiency of select iST technologies at different numbers of genes simultaneously measured ^{20,34,37,38,40,42,43,91}. D) Experiments to measure and compare the performance of methods, and corresponding quality control metrics. Sensitivity and specificity experiments assess the rate of false-positives and false-negatives through comparisons with smFISH (External Validation), and internal positive and negative control transcripts (Internal validation). As the number of molecular features increase in imaging ST, quality control in dynamic range, accuracy, and scalability need to be examined. Dynamic range represents the maximum number of molecules that can be measured, accounting for molecular crowding. This may differ between methods as a function of signal to noise, and the sparsity of codebooks, and is dependent on the accuracy of the measurement. Scalability considers how the experimental time and cost scales with the number of features.

Detection modality.

There are three main strategies for labeling RNA molecules in situ: i) direct probe-based detection, ii) enzymatically assisted probe-based detection, and iii) direct enzymatic sequencing of RNA molecules in situ (Fig 3A).All detection modalities start with fixed cells and tissues, wherein the RNA molecules are crosslinked to cellular matrix, thereby fixing their positions throughout processing. Direct probe-based detection is based on fluorescence single-molecule FISH (smFISH) protocols pioneered by Robert Singer, and subsequently Arjun Raj and colleagues^31–33, in which RNA transcripts are tiled with many (>20) short (20–50 nts) fluorescently labeled complementary oligonucleotide probes recruited to a single, diffraction-limited spot, generating a punctate, high-specificity signal.

The second approach, enzymatically assisted probe detection, solves the specificity and sensitivity problem through enzymatic detection and polymerase-based amplification with RCA^23,34. There are several advantages to enzymatic assisted probe detection. First, RCA amplicons are bright and detectable with high signal-to-noise ratio(SNR) in fluorescence microscopy, even with lower magnification and exposure times. Second, enzymatic gapfill of padlocks allows interrogation of genetic variation (e.g. SNPs) and barcodes^23,35,36. Finally, increased SNR allows for more diverse iterative barcoding approaches, such as commercial fluorescence sequencing chemistries. The initial methodology circularized probes hybridized to in situ generated cDNA, but more recent methods have amplified directly from RNA to increase detection efficiency^4,37.

The third detection approach, direct enzymatic sequencing, utilizes in situ enzymatic reactions to perform RNA sequencing library construction within cells and tissues. This approach, pioneered by Church and colleagues, and subsequently expanded by Boyden and colleagues, utilizes in situ reverse transcription with random hexamer primers to generate cDNA, which is subsequently fragmented, and intra-molecularly circularized^22,37. These circularized molecules are subsequently amplified via RCA. Here, specificity for individual RNA molecules is conferred through alignment of in situ sequencing reactions (see Combinatorial Multiplexing) performed on the RCA amplicons. Due to its untargeted nature, this approach is most similar to sST and offers the possibility of hypothesis generation from transcriptome-wide data. However, the low conversion rate of RNA molecules to sequenced RCA amplicons limits its application in many ST experimental contexts that require sensitive gene expression quantification.

Multiplexing approach.

There are two main classes of imaging-based multiplexing which combine with the detection modalities in the previous section – sequential readout and combinatorial multiplexing (Fig. 3A–B). Both classes leverage multiple imaging rounds to overcome the limitations in spectral bandwidth. In sequential readout approaches, in each imaging round, a unique set of mRNA molecules is labeled, imaged, and the fluorescent probes are removed following imaging. Subsequently, in the next imaging round, a new set of mRNA molecules are labeled, and the process repeated. In this way, the number of unique mRNA targets imaged scales as (number of rounds) * (number of fluorescent channels). In one example, in osmFISH³⁸, smFISH probes are stripped via denaturants during each imaging step, followed by a new round of smFISH staining. The main advantage of sequential readout methods is that they are simple to implement: they do not need sophisticated image processing and alignment over many imaging rounds, and are robust to encoding errors and the density of labeled RNAs. However, this ease of implementation comes with a significant tradeoff in multiplexing scalability (Fig. 3C).

In combinatorial multiplexing approaches, each mRNA molecule is repeatedly interrogated over multiple imaging rounds, and its identity decoded by the combination and order of images it is detected in^{20,21,39–41}. For example, in MERFISH, Zhuang and colleagues assign a binary code to each mRNA molecule interrogated, wherein a 1 corresponds to an imaging round where the mRNA molecule is labeled, and a 0 corresponds to an imaging round where it is dark. Here, the codebook size in theory scales as 2^N, where N is the number of imaging rounds. Critically, Zhuang and colleagues implement a hamming-distance-corrected error robust codebook, which utilizes bits in the diversity of the codebook to enable detection and correction of errors in single-imaging rounds. This allowed for 140 genes to be imaged in 16 rounds, with the ability to correct for errors in one round, and the ability to detect if an error has happened in two imaging rounds. Under the umbrella of combinatorial multiplexing, there exists a host of implementations, which mainly differ in cycling chemistry (Fig. 3A, cycling chemistry). The simplest cycling chemistry is reversible hybridization of a fluorescently labeled probe, utilizing heat/denaturants and/or DNAse to remove DNA probes after imaging^38,41. To increase the speed of cycling, reducing-agent cleavable dyes conjugated to detection oligonucleotides have been used⁴⁰. Lastly, cyclic fluorescent sequencing chemistries such as sequencing by ligation (SBL) and sequencing by synthesis (SBS) have been directly used in situ for combinatorial readout, primarily for RCA amplicons^4,35,37. Given that the encoding space of combinatorial multiplexing scales exponentially, the main barrier to increasing the number of genes detected is density of molecules labeled. Recent advancements have enabled detection of 1000–10,000+ genes with high efficiency through either sparsification of the codebook or expansion microscopy, with a decrease in throughput due to increased imaging rounds (Fig. 3C)^3,4,20,42,43. We summarize the combinations of read and detection methods used, highlighting the degree of multiplexing and detection efficiency of non-exhaustive technology examples in Fig 3B–C, with more technologies being developed recently^44–48.

Image processing.

Following the collection of multi-round image data, there are three primary steps in image processing to generate primary spatial transcriptomic data: i) spot detection, ii) image alignment and registration, and iii) decoding into spatial mRNA localizations. During spot detection, local maxima detection is used to localize the centroid of fluorescent spots corresponding to individual mRNA molecules. In both sequential readout and combinatorial multiplexing approaches, the same cells and tissues are imaged over many imaging rounds. As such, a critical aspect of the multiplexing approach is to align each imaging round to the same coordinate framework. This is generally performed using image features such as fluorescent nuclear staining (DAPI), fiduciary markers such as fluorescent beads, or the mRNA molecule localizations themselves. For combinatorial multiplexing, the alignment needs to be accurate to the resolution of individual mRNA localization (~ single diffraction limited spot) for decoding. Lastly, following image alignment and registration, for combinatorial multiplexing, the order of fluorescent signals for each localized spot is decoded – either through matching to a codebook in multiplexed FISH and targeted in situ sequencing approaches, or through matching to the transcriptome for untargeted in situ sequencing. For transcriptome mapping, while in situ sequencing is limited in read length, there have been pioneering approaches to paired sequencing of the same molecules with ex situ high throughput sequencing to enable longer read lengths.

Quality control.

In iST, just as in sequencing-based ST, the two key data quality parameters are sensitivity and specificity. These parameters need to be assessed given the specific conditions for detection, multiplexing, and image processing for a given technology. There are two primary approaches to assess sensitivity and specificity: i) direct external validation via single-molecule FISH, and ii) the use of internal positive and negative controls. Several studies (MERFISH, ExSEQ), have directly benchmarked multiplexing performance against smFISH for the same images. The near quantitative detection rate and low false positivity rate of smFISH allows for direct measurements of detection efficiency (false-negatives), as well as false-positives. In addition to direct external validation, the use of built-in positive controls can provide internal validation for multiplexed imaging experiments. These include built-in codebook controls with no probes assigned to measure decoding accuracy, scrambled probes to measure false-positive detection rate and standard control mRNA targets to allow for comparison of detection efficiency across experiments (Fig. 3D).

An underexplored aspect of imaging-based approaches is how performance varies with the degree of multiplexing. As the number of interrogated mRNAs increases, molecular crowding can prevent efficient detection in the case of enzymatic approaches and make image processing and decoding difficult for FISH based detection. In MERFISH³⁹, and in situ sequencing³⁷, such molecular crowding has been addressed through expansion microscopy^49,50 at the expense of volumetric throughput. In addition, as more genes are encoded, either longer codebooks are needed, or fewer bits can be devoted to error correction. As such, a key benchmark needed is how sensitivity and specificity scale as a function of number of genes and spatial molecular density (Fig. 3D).

Lastly, there is a need to systematically assess how image processing pipelines affect data outputs. This is hampered by the fact that current analysis pipelines for imaging transcriptomics are boutique and technology specific. In addition, there is a current lack of standardized data formats across the field for images, making it difficult to develop and benchmark generalized image processing pipelines (from registration, spot-calling, decoding to cellular segmentation) for iST. As such, there is a critical need in the field for generalizable, open source image processing tools for iST⁵¹ and standardized primary datasets to benchmark such tools.

Critically, all of these benchmarks and QC parameters need to be explored in the context of a set of standardized reference tissues, preferably shared with sST approaches outlined in Fig. 2.

Beyond ST: genomics technological innovation in cells and tissues

The recent, rapid progress in the development of ST methodologies is a harbinger of a broader technological transformation of genomics to the study of intact cells and tissues. The same tools that enabled ST–high-throughput DNA sequencing, novel barcoding strategies, new microscopy tools, and innovations in molecular enzymatics–will increasingly allow genomics to be used to answer questions in cell and tissue biology. We anticipate technology development proceeding in three main domains (Fig. 4).

Figure 4. The future of contextual genomics.

The in situ measurement of transcriptomes marks the beginning of a larger technological effort to import genomic measurements into intact biological systems. Innovation along three major axes–resolution, modality, and dynamics–will provide powerful new tools for interrogating tissue structure and function.

Modality.

Modern genomics has assembled an impressive array of measurement technologies of biomolecules through creative combinations of enzymatic and biochemical manipulations with DNA sequencing readouts. However, most of these technologies can only be used in vitro, on substrates extracted and purified from tissues or cells. Adapting genomics technologies to function within intact tissue sections represents an enormous opportunity for biological discovery. We anticipate that the progress ST technologies have made in digitally counting transcripts will be increasingly applied towards other modalities of genomic measurements. Most immediately, iST has been used to quantify nascent mRNA by targeting introns⁵², which could be similarly adapted to quantifying splice variants by targeting exon-exon junctions.⁵² The discovery of isoform variants currently require long-read technology and/or plate-based methods to assess. Future adaptation of long-read pipelines, or targeted isoform sequencing in situ, should enable scalable adaptation to the spatial domain. While the immense diversity of isoforms across cell-types is well known, there remain many questions about how this variability is functionally manifested across cell types and tissue localizations.

Beyond the transcriptome, spatial profiling of genomic variation has begun to be explored, but tools are still in relative nascency. Available technologies have primarily measured genome structure through imaging^53–56, but more recently, sequencing-based strategies have also been reported⁵⁷. Many of the same techniques used by these technologies could be applied to study epigenomic modification and regulation as well⁵⁸. Beyond obvious applications in cancer mutational profiling, spatial profiling of genomic variation may help elucidate the functional relevance of somatic mutations in aging and disease.

Spatial methods may offer the opportunity to unify the world of genomics and proteomics. While genomics has traditionally been focused on dissociated measurements, interrogation of protein localization with affinity reagents has long been performed with low multiplexity in situ. Recent developments in DNA-barcoding of affinity reagents such as antibodies have enabled highly multiplexed protein readouts via sequencing^26,59. These approaches are readily adaptable to the spatial domain, especially in the context of spatial-capture ST measurements. In the near future, whole proteome affinity libraries (including antibodies, nanobodies, and aptamers) may be utilized and interrogated in situ. Furthermore, proximity-based enzymatic reactions–such as ligation⁶⁰, and polymerase extension^61–63–may enable high throughput protein-protein, DNA-protein, and RNA-protein interaction measurements within tissues. Lastly, there has been rapid development in numerous approaches for novel protein sequencing methods which leverage single-molecule imaging⁶⁴ which, one day, may allow for direct protein sequencing in tissues.

Resolution.

Current spatial genomic approaches span a wide range of spatial resolution from broad tissue regions to subcellular localization, with an overall tradeoff between volumetric throughput of tissue and spatial resolution of data collected. Different spatial resolutions are suited to different classes of biological problems. The ability to perform cellular segmentation on the measured transcripts is crucial for many downstream applications, such as quantifying the cell type composition and organization of tissues. In both sST and iST, segmentation is a computational problem. For sST, it involves the deconvolution of cell type mixtures from a pixel^65–67; the problem is made much simpler when pixel size is reduced to the size of individual cells. For iST, the computational task is to assemble individually detected transcripts into cells from microscopy images^68–70. Progress on both of these efforts has been made, but additional computational innovation will greatly accelerate the ability of ST tools to be applied to problems in tissue biology⁷¹.

After tissue composition, increasing resolution will enable models of tissue organization which take into account cellular interactions. Fundamental questions include: how does one extract the relevant cellular networks which compose tissues? And, how does one discover the functional receptor-ligand interaction networks between cells? Such questions are most suited to technologies with cellular resolution or higher, to enable accurate assignment of molecular signatures to networks of interacting cells.

Third, technologies that enable the precise in situ colocalization of biomolecules – potentially at resolutions exceeding the diffraction limit – remain almost entirely unexplored. One intriguing exception was the proposal by several groups to use in situ PCR amplification with local concatenation to detect spatial proximity between two nucleic acids^61,62. A method for in situ molecular colocalization would be enormously biologically enabling, from the quantification of transcription factor binding sites, to the colocalization of biomolecules to specific organellar compartments, or the quantification of gap junctions or other cell-cell interactions. While genomic technology for quantifying such proximity or interaction events does exist – such as ChIP-seq or ribosomal profiling – they are most suited to bulk-level analyses of lysed tissue, rather than being deployable in situ.

Dynamics.

While almost all current genomic technologies are end-point measurements which provide a snapshot of biology, there is an exciting future for approaches which provide temporal context to spatial genomic measurements. Looking forward, there are three important areas which are ripe for future innovation, i) integration of molecular recording technologies, ii) computational inference of dynamics, and iii) in situ perturbations.

First, an especially promising intersection exists between spatial-genomic technologies and synthetic genetic manipulation approaches for molecular recording. There has been a recent explosion of gene-editing based molecular recording technologies which encode lineage and signaling dynamics into genomic sequence^72–77. These approaches have leveraged single-cell sequencing for subsequent readout of the genomic information. Given the inherent similarities in chemistry between single-cell and sequencing-based spatial transcriptomics, genomic recordings of lineage may be easily translated to the tissue context. Such measurements may represent a holy grail for developmental biology, wherein information regarding cellular lineage, cell-state, and adult tissue organization can be simultaneously combined to form the developmental picture. In tumor evolution, such approaches can answer important questions regarding the spatial heterogeneity of tumor clones, and the relationship between tumor clone fitness^57,78 and the cellular microenvironment. Beyond lineage, novel recording technologies are beginning to encode cellular histories and transcriptomic states⁷⁴, which may offer the promise of fully 4D measurements of cell states with genomics.

Second, there has been an explosion of computational tools to enable the inference of dynamics in single-cells such as pseudotime^79,80 and RNA-velocity^81,82. Approaches developed for single-cell have been successfully applied to ST data, such as the application of RNA velocity to ST of the developing mouse cortex⁸³. However, future development of spatial-oriented toolboxes for computational dynamics will more directly leverage the unique aspects of ST data. One may leverage the subcellular localization of RNA to infer additional temporal information regarding dynamics (e.g. RNA is made in the nucleus, but translated in the cytoplasm⁴²). More importantly, spatial information may provide contextual ground truth in pseudotime analyses, such as utilizing cellular localization to assign nodes in cellular trajectories.

Third, perturbations allow causality to be inferred from end-point measurements. Recent approaches have enabled nucleic-acid barcoding of both genetic and chemical perturbations⁸⁴ for single-cell readout. Similar to molecular recording, perturbation barcoding approaches^84–86 may be inherently compatible with spatial capture methods. Additionally, barcode-readout via in situ sequencing has been demonstrated for imaging methods^36,87,88. Analyzing the effects of genetic perturbation in situ using ST will enable profiling of phenotypes which cannot be accessed in the absence of tissue context. These phenotypes are numerous, including cellular localization and cell-cell interactions.

Conclusion

Application of genomics to tissues represents an exciting, multifaceted domain of technology development. Faced with many exciting opportunities for new measurements in the future, what can the field do to further accelerate the pace and quality of these developments? The early and easy sharing of data and protocols are key catalysts of technological progress. This lesson is perhaps best exemplified by the recent rapid progress in the field of single-cell genomics, where manuscripts are usually preprinted, processed and raw data have been shared with publication (and often before), and protocols made easily accessible through online portals. In ST, the heterogeneity of methodologies breeds a large variety of file formats and data structures. This makes data and protocol sharing more challenging, but also all the more important in order to contextualize the utility of new technologies as they become available. Relatedly, we see the adoption of standardized tissues and QC measurements–which we detailed extensively in the sections above–as essential to developing a common language for technological progress in ST. Without the proposed standards, a clear comparison of even basic metrics–like the capture efficiency of RNA–is not possible.

The advent of spatial genomics also poses an additional important challenge of how to facilitate hypothesis generation and testing by biological experts, using these powerful new tools. Undoubtedly, a key to this will be invention and adoption of computational tools for analysis of these new data types. Progress in development of these tools has been exciting, and is discussed extensively in the accompanying review. Computational tools will be important not only for the extraction of biological insights from the data, but also for informing the design of spatial genomics experiments. We anticipate the maturation and dissemination of spatial genomics technologies will become a critical driver of biological discovery across many fields of tissue and disease biology.