Heterogeneity of Signaling Complex Nanostructure in T Cells Activated Via the T Cell Antigen Receptor

Abstract

Activation of the T cell antigen receptor (TCR) is a key step in initiating the adaptive immune response. Single-molecule localization techniques have been used to investigate the arrangement of proteins within the signaling complexes formed around activated TCRs, but a clear picture of nanoscale organization in stimulated T cells has not emerged. Here, we have improved the examination of T cell nanostructure by visualizing individual molecules of six different proteins in a single sample of activated Jurkat T cells using the multiplexed antibody-size limited direct stochastic optical reconstruction microscopy (madSTORM) technique. We formally define irregularly shaped regions of interest, compare areas where signaling complexes are concentrated with other areas, and improve the statistical analyses of the locations of molecules. We show that nanoscale organization of proteins is mainly confined to the areas with dense concentrations of TCR-based signaling complexes. However, randomly distributed molecules are also found in some areas containing concentrated signaling complexes. These results are consistent with the view that the proteins within signaling complexes are connected by numerous weak interactions, leading to flexible, dynamic, and mutable structures which produce large variations in the nanostructure found in activated T cells.

Introduction

Formation of protein assemblies comprised of enzymes and adapter molecules, otherwise known as signaling complexes, is a common feature of receptor activation. In the adaptive immune response, activation of the T cell antigen receptor (TCR) begins with the recognition of a specific peptide bound to a molecule encoded by the major histocompatibility complex (pMHC) on an antigen presenting cell (APC). This recognition and binding event initiates the creation of signaling complexes containing the scaffold and effector proteins needed for signal transduction (Rajasekaran et al., 2016; Balagopalan et al., 2020). Many of the participants in TCR-based signaling have been identified, including linker for activation of T cells (LAT), a critical adapter protein that is recruited to signaling complexes and then phosphorylated on several tyrosine residues. Phosphorylated LAT serves as a binding platform for Src homolgy 2 (SH2) domain–containing components, such as growth factor receptor bound protein 2 (Grb2), Grb-2 related adaptor downstream of Shc (Gads), and phospholipase C gamma 1 (PLCγ1) (Zhang et al., 1999a; Balagopalan et al., 2010). Another critical adapter protein, SH2 domain–containing leukocyte protein of 76 kDa (SLP-76), is bound constitutively to Gads and is brought into LAT-based complexes when Gads binds to phosphorylated LAT. SLP-76 is also phosphorylated on multiple tyrosine residues, producing additional docking sites for important effector molecules such as non-catalytic region of tyrosine kinase adaptor protein (Nck), vav guanine nucleotide exchange factor (Vav), and interleukin-2-inducible T-cell kinase (Itk) (Myung et al., 2001; Clements, 2003). Cooperative interactions are important in LAT-based complexes. PLC-γ1 binds to both LAT and SLP-76, and interactions with both sites are needed to stabilize the association of PLC-γ1 with signaling complexes (Braiman et al., 2006; Barda-Saad et al., 2010). Recent studies have shown that these proteins do not cluster simultaneously but arrive at the TCR sequentially, with Zeta chain–associated protein kinase 70 (ZAP-70) arriving the earliest, followed by LAT and then the majority of the other signaling molecules (Yi et al., 2019). However, this is only a partial list of the required components; there are additional proteins, not discussed here, that are also required for proper signal transduction, T cell activation, and subsequent immune responses (Yablonski et al., 1998; Zhang et al., 1999b; Balagopalan et al., 2020).

Imaging studies have shown that TCR engagement leads to dramatic changes in T cells. Even at rest, the surface of the T cell is highly dynamic with undulating membrane protrusions (Cai et al., 2017; Ghosh et al., 2020). Upon contact with an APC displaying an appropriate antigen, the T cell adheres to the APC and initiates the assembly of signaling complexes. The T cell then spreads against the APC, while rearrangement of the membrane, proteins, and the actin cytoskeleton continues, often leading to the formation of the immunological synapse. Visually, activation begins with the rapid formation of clusters of signaling proteins termed microclusters that have been observed by light-level, diffraction-limited microscopy (Bunnell et al., 2002; Yokosuka & Saito, 2010). These microclusters have been studied extensively in T cells activated by pMHC on an APC (Johnson et al., 2000; Krummel et al., 2000; Freiberg et al., 2002; Lee et al., 2002), by activating molecules incorporated into lipid bilayers (Grakoui et al., 1999; Campi et al., 2005; Yokosuka et al., 2005; Kaizuka et al., 2007; Ilani et al., 2009) and by activating antibodies on glass surfaces (Bunnell et al., 2001, 2003; Barda-Saad et al., 2005). Early imaging experiments showed that microclusters form in the regions where lamella on the exterior of the T cell forms tight contacts between it and activating surface (Bunnell et al., 2002). Several studies have provided evidence that these sites of first contact between a T cell and APC or activating surface are the tips of preexisting microvilli where many signaling proteins are concentrated (Jung et al., 2016; Cai et al., 2017; Razvag et al., 2019; Ghosh et al., 2020). The concentration of microclusters along these contact sites persists for many minutes, so the distribution of microclusters is not uniform across the cell surface of fully spread T cells. This uneven distribution is known as first-order clustering, a term that acknowledges that microclusters are mainly confined to areas determined by T cell contacts. Microclusters contain most of the molecules required for TCR signaling, including both phosphorylated LAT and SLP-76, and they appear to be the sites where signal transduction begins (Bunnell et al., 2002; Yokosuka et al., 2005; Varma et al., 2006). Live cell studies have shown that microclusters are dynamic structures, as constituents of the signaling complexes continuously dissociate and reassociate. Other proteins such as CD45 are excluded from the microclusters (Bunnell et al., 2002). Feedback from spreading may influence the composition and fine structure of the microclusters. Additionally, microclusters are now understood to be a kind of membrane-lacking compartment known as condensates formed by liquid–liquid phase separation (Banani et al., 2017; Case et al., 2019). Phase separations at membranes occur when the interactions formed between the proteins are stronger than their interactions with the bulk membrane, resulting in a separate phase. The resulting formation of condensates would explain the apparent decrease in the area occupied by clusters in activated cells (Hu et al., 2016). Some proteins in the microclusters that we are studying contain SH2 and Src homology 3 (SH3) binding domains in intrinsically unstructured regions that have been shown to help produce phase transitions (Cornish et al., 2020). Finally, it has been shown in vitro that phase transitions can drive T cell activation (Su et al., 2016; Zeng et al., 2021).

There has been great interest in understanding how T cell signaling complexes are grouped and if there is organization of TCR-based complexes within the areas of close contact containing concentrated microclusters. While proteins are organized by many binding interactions, such as those delineated above, higher-order interactions such as son of sevenless 1 (Sos1)-dependent oligomerization of LAT are also required for maximal TCR-dependent phosphorylation and activation of phospholipase Cγ1 and Ca²⁺ signaling (Kortum et al., 2013). PLCγ1 oligomerization itself enhances activation (Zeng et al., 2021). In addition, multipoint binding of SLP-76 to adhesion and degranulation-promoting adapter protein (ADAP) facilitates the assembly of SLP-76–containing microclusters (Coussens et al., 2013). These results indicate that clustering of the TCR is not sufficient for T cell activation and further higher-order organization of signaling complexes is needed. The microclusters seen in diffraction-limited microscopy might thus be composed of signaling complexes organized for maximal T cell activation.

Single-molecule localization microscopy (SMLM) should allow a more detailed look at the organization of microclusters. SMLM techniques, which include photo-activation localization microscopy (PALM) and stochastic optical reconstruction microscopy (STORM), use fluorophores that can be switched between on (fluorescent) and off (dark) states, allowing high-resolution localization of single molecules (Sydor et al., 2015). The locations of single molecules should then show the organization of proteins in signaling complexes. A general consensus has developed that most signaling proteins in unactivated T cells are organized into nanoclusters and that this self-clustering increases following T call activation (Balagopalan et al., 2015; Pageon et al., 2016; Feher et al., 2019), although these conclusions have been disputed (Rossboth et al., 2018). Early studies of T cell microclusters using two-color SMLM also showed a variety of interactions between LAT and other proteins. The TCR and LAT form small clusters that tend to be segregated from each other (Lillemeier et al., 2010; Sherman et al., 2011), with some overlap at “hotspots” (Sherman et al., 2011). ZAP-70 kinase mixes uniformly with the TCR but shows only partial mixing with LAT. Grb2 mixes well with LAT throughout the cell, even in the smallest nanoclusters. This study also demonstrated that LAT and SLP-76 do not mix well and are organized with LAT tending to be in the center and SLP-76 distributed on the outside of clusters (Sherman et al., 2016).

Further research indicated that the organization of SLP-76 around LAT developed when the T cell is fully spread, well after the T cell contacts an activating surface (Barr et al., 2016). This is consistent with photobleaching data that suggests that molecules can freely exchange in newly formed complexes but become more fixed in mature complexes (Bunnell et al., 2002). Additional work examining three components, LAT, SLP-76 and PLCγ1, reaffirmed that LAT is at the center of microclusters, while SLP-76 and PLCγ1 are on the outside (Sherman et al., 2013).

However, despite the high precision reported by localization algorithms, visualization of individual proteins within complex structures has been hampered by several issues. First, the accuracy of individual localizations is severely limited by the difficulty in determining the exact position of a single molecule and problems with the alignment of multicolor images (Yi et al., 2016; Barr et al., 2017). Also, most T cell SMLM studies have analyzed either the entire cell surface or central areas that included contact sites with dense concentrations of microclusters along with the intervening areas containing only a few microclusters. Thus, these studies can conflate the first-order clustering caused by confinement of microclusters in the sites of first contact with the nanoclustering of microclusters that we are interested in studying. In addition, techniques that use genetically based probes, such as PALM, report on the locations of an entire pool of protein, not just the activated, phosphorylated molecules we expect to find organized into signaling complexes. This makes it difficult to see differences between the organization of molecules within signaling complexes and those not engaged in signaling. Finally, due to technical limitations, most SMLM studies have studied only two or three proteins in a single image.

In the current report, we have improved the study of nanostructure organization in T cells in several ways. First, we segmented the T cell surface by formally defining irregularly shaped areas, as necessary, to segment out areas of close contact between the T cell and activating surface where microclusters are concentrated. We contrast those with membrane areas further away from the activating surface that contain rare microclusters as observed by diffraction-limited microscopy. We recently reported development of an SMLM technique, madSTORM, that allows visualization of multiple targets in a single sample (Yi et al., 2016). This technique utilizes the sequential binding and elution of fluorescently labeled antibodies for multiplexed direct stochastic optical reconstruction microscopy (dSTORM) imaging combined with the use of fluorescent nanodiamonds for drift correction and alignment. We used this madSTORM technique to visualize six different proteins in a single sample of activated Jurkat T cells. Furthermore, since this technique uses antibody-based imaging, we could focus on the locations of only activated, phosphorylated signaling proteins. Based on the availability of specific antibodies for proteins involved in T cell activation, particularly antibodies against activated species, we choose to examine phosphorylated LAT (pLAT, base MW 36 kDa), phosphorylated zeta chain of the TCR (pTCRζ, base MW 18 kDa), phosphorylated-ZAP-70 (pZAP, base MW 70 kDa), phosphorylated-SLP-76 (pSLP, base MW 76 kDa), phosphorylated-PLCγ1 (pPLCγ1, base MW 145 kDa), and the nonmicrocluster protein, CD45 (MW 147 kDa).

We also sought to improve the statistical analyses of the images. SMLM data consists of points located in x–y space in contrast to pixel-based diffraction-limited images. These images require the use of point process statistical methods, which are used to study events localized in space or in space and time, to determine relationships between the points. Methods such as exploratory analysis and parametric model-fitting are used to explore, analyze, and model point patterns (Baddeley & Turner, 2006). Exploratory analysis of point process includes investigating the first- and second-order effects of the point process (Baddeley, 2008). First-order methods examine the intensity of the point pattern and the variation of intensity on a large scale due to the change of the underlying area's structure (Baddeley, 2008; Knitter & Nakoinz, 2018). Then, second-order methods describe the distribution of distances between neighboring points to determine if there is any spatial dependence of the point pattern at small scales (Baddeley et al., 2015). Point process parametric model-fitting aims to find a mathematical system that describes the underlying behavior of the point pattern, and additional covariates can be included in the model to help explain the underlying behavior (Baddeley & Turner, 2006). Here, we have applied both exploratory methods and model-fitting to analyze the relationships of the six proteins involved in TCR-based signal transduction that were imaged using madSTORM.

Materials and Methods

Antibodies

Mouse antihuman CD3 (BD Pharmingen San Jose, CA, clone Ucht1, cat# 555330, RRID AB_395737) was used to coat coverslips. The following antibodies from BD Pharmingen (San Jose, CA) were used for madSTORM imaging: mouse antihuman phospho-LAT (pY226) (cat# 558363, RRID AB_647281), mouse anti-CD247 (pY142) (cat# 558402, RRID AB_647307), and mouse antihuman CD45 (cat# 555480, RRID AB_395872). Antibodies from Cell Signaling Tech (Danvers, MA) were also used for madSTORM imaging: rabbit anti-phospho-SLP-76 (pY145) (cat# 14770, RRID AB_2798604), rabbit anti-phospho-ZAP-70 (pY319) (cat# 2717, RRID AB_2218658), and rabbit anti-phospho-PLCγ1 (pY783) (cat# 14008, RRID AB_2728690). Antibodies used for madSTORM were conjugated to Alexa 647 dye using a Molecular Probes kit A20186 (ThermoFisher Scientific, Waltham, MA). Antibodies were tested by Western blotting and immunofluorescent imaging.

Cells and Tissue Culture

Wild-type (E6.1) Jurkat T cells (ATCC, Manassas, VA) were maintained in RPMI 1640 supplemented with 10% fetal bovine serum and antibiotics. Tissue culture reagents were from Gibco-ThermoFisher (Grand Island, NY). The generation of E6.1 Jurkat cells lines stably expressing LAT-YFP has been described previously (Bunnell et al., 2002). Cells were tested three times per year for mycoplasma contamination.

Sample Preparation

The preparation of coverslips follows previously described techniques (Bunnell et al., 2003; Yi et al., 2016), Thermo Scientific Nunc Lab-Tek II 8 well chambered coverslips (ThermoFisher Scientific, Waltham, MA) were cleaned with acidic EtOH, coated with polylysine, and then incubated with nanodiamond fiducial markers. One hundred nm nanodiamonds (cat#ND-NV-100nm_Hi, Adamas Nano, Raleigh, NC) were prepared by diluting a 1 mg/mL stock 1:300 in phosphate buffered saline (PBS), sonicating 10 s, and then incubating 400 mL/chamber for 30 min at room temperature. Following five washes with PBS, the coverslips were subsequently incubated with stimulatory anti-CD3 antibody at a concentration of 10 mg/mL and washed again in PBS. Cells were plated and fixed following established protocols (Bunnell et al., 2003). The fixed cells were permeabilized by incubating 5 min at room temperature in 0.01% Tx-100 (Yi et al., 2016).

Imaging

The madSTORM imaging buffer was prepared following method B from the Nikon N-STORM protocol, except that we added 100 mM 2-mercaptoethanol (M7154; Sigma-Aldrich), 20 mM cysteamine (30070; Sigma-Aldrich), and 2 mM cyclooctatetraene (138924; Sigma-Aldrich). The elution buffer consisted of 3.5 M MgCl₂ (M9272; Sigma-Aldrich) adjusted to pH 6.0, 20 mM PIPES (P6757; Sigma-Aldrich), and 0.1% Tween-2 (BP337-500; Fisher Scientific, Hampton, NH). madSTORM imaging was performed as described (Yi et al., 2017) with the following modifications. Before the madSTORM imaging, a total internal reflection fluorescent (TIRF) image of yellow fluorescent protein (YFP)-conjugated LAT was taken to show that the cell was activated and to show where microclusters were present in the cell. A series of 10,000 STORM images was taken of the unstained sample. Subsequent blocking with 1% fish gelatin in PBS, staining, imaging, antibody elution, and bleaching follow the madSTORM protocol except the treatment with 4% paraformaldehyde (PFA) was reduced to 5 min (see below). A madSTORM image series was taken after each bleaching step for use in finding fiducial markers and to confirm that the previous staining has been removed. madSTORM imaging was done on a TIRF microscope based on an inverted Nikon TI Eclipse microscope (Nikon Instruments Inc., Melville, NY), a 647 nm LUNB laser (125 mW), and a 100 × Apochromat TIRF objective (NA 1.49) with an Andor iXon DU-897 EMCCD 512 × 512 pixel detector (pixel size of 16 nm) plus a 1.5 × intermediate magnification lens.

Sequential Staining

All the directly conjugated antibodies were diluted and then centrifuged 15 min at 100,000 × g just before use to remove any aggregated material. To image the first antibody, which was always anti-phospho-LAT (pY226) conjugated to Alexa-647, the sample was blocked with 1% fish gelatin in PBS on the microscope stage for 30 min at room temperature. The freshly centrifuged diluted antibody was added to the chamber on microscope for 1 h at room temperature, followed by washing 5× PBS. Then, the buffer was replaced with 1 mL STORM buffer and a 10,000 image series was acquired. The sample was then washed 5× Tris-buffered saline (TBS), and the bound antibody was removed by three 1 min incubations in elution buffer. The sample was washed 3 × TBS followed by 3 × PBS washes and was photobleached for 10–20 s using the 647 nm laser at high power (125 mW) and the 405 nm laser at 2–5 mW to photo-activate any residual A647 dyes in dark state. Then ,another image series of 10,000 frames in STORM buffer was taken of the bleached sample. Following washes of 3 × PBS, 5 min in 4% PFA to prevent reversible cross-linking, and additional washes 5× PBS, the sample was then incubated with next antibody. This procedure was repeated for all six antibodies.

Image Processing

Localization, drift correction, and alignment of localizations were done as described previously (Yi et al., 2016). Thunderstorm plugin on Fiji software (Ovesný et al., 2014) was used to detect localizations, and the previously described MATLAB codes were used for drift correction and alignment (Yi et al., 2016). Localizations from the fiducial markers were identified and removed using a custom knime code where DB.Sc.an (Ester et al., 1996) was used to find groups of localizations found in all comma-separated value (csv) localization files of the last 1,000 images of the images of the bleached samples. Then, these localizations were removed from the csv files of the sample localizations. In addition, localizations were removed if the number of detected photons was below 1,000.

Drawing of Masks

The LAT-YFP TIRF image of the cell of interest was aligned with a rendering of the pTCRζ localizations using bunwarpJ in ImageJ. Regions of interest (ROIs) were drawn on the pTCRζ renderings using the fluorescent image of microclusters as a guide. These ROIs were saved and used as masks in an R algorithm that extracted the localizations within the masks for each protein and compiled them into a csv file. Similarly, ROIs of nonmicrocluster areas were drawn and used to extract localizations that were outside of the microclusters.

Statistical Methods

Exploratory Analyses

The density of the point pattern measures the first-order effects of the underlying point process. Nonparametric kernel density estimation was used to estimate the variation of intensity throughout the study area. Density plots were made using kernel smoothing with an isotropic Gaussian density and the same default bandwidth (one-eighth of the shortest side length of the enclosing rectangle) applied to each type (Baddeley et al., 2015). Cross-type J-functions were used to inspect the dependency of points between different types (van Lieshout & Baddeley, 1996). The J-function for the pair of points of types i and j, J_ij(r), is the ratio of 1 − G_ij(r) to 1 − F_j(r), where G_ij(r) is the distribution function of the nearest distance from a point of type i to a point of type j and F_j(r) is the distribution function of the nearest distance from an arbitrary (empty) location to a point of type j (Tukey, 1977; Baddeley et al., 2015). The benchmark of J_ij(r) is 1 for any pair, i and j. When i = j, the benchmark J_ii(r) = 1 means points of type i follow complete spatial randomness (CSR), and when i ≠ j, the benchmark J_ij(r) = 1 means points of type i are occurring independently of the presence of type j points. A J-function above the benchmark indicates dispersion, while a J-function falling below the benchmark indicates clustering between types. Envelopes were created using Monte Carlo simulations under the null hypothesis of CSR for each type i and complete spatial randomness and independence (CSRI) for each pair of types i and j (Baddeley et al., 2015).

Ripley's K-function is a widely used second-order statistic (Ripley, 1976; Baddeley & Turner, 2005), while the L-function is a scaled version of K-function that stabilizes the variance of K-function (Besag, 1977). The empirical K- and L-functions can be compared with their theoretical values under the assumption of CSR to detect features of a point process such as clustering or inhibition relative to CSR. The K-function and L-function can also be extended to multitype point processes to measure the spatial relationships between two types i and j at each distance r (Baddeley et al., 2015). We used the centered cross-type L-function to compare the sample estimate with the benchmark value of 0. To test if deviations from the theoretical values were statistically significant, envelopes were obtained through Monte Carlo simulations with permutation of labels, in order to construct confidence bounds. The null hypothesis tested is CSRI between different types of points and CSR for individual point types. Each simulation was run 19 times to test the null hypothesis at the significance level of 0.05.

Modeling

Poisson models and Gibbs models were used to model intensity and pairwise interaction patterns among the points. For the Poisson model, linear, quadratic, and cubic trends were considered and each was combined with an additive or interactive protein type term. The intensity function λ(u, m) was modeled as a log linear function of location u = (x, y) and the protein type m. The estimates of the optimal parameters were obtained by maximizing the likelihood using the Berman–Turner method applied to multitype patterns (Baddeley & Turner, 2000).

Gibbs models extend the Poisson process model by incorporating pairwise interaction effects. We considered multitype Strauss and Strauss hard core for the interaction term to model the mark-dependent pairwise interactions among the six types of proteins (Baddeley & Turner, 2006; Illian et al., 2008). Following Baddeley and Turner (Baddeley & Turner, 2006), we assume a multitype point process is represented by:

$$\begin{array}{rl}x& =\{(x_1,m_1),\dots ,(x_{\mathrm{n}},m_{\mathrm{n}})\},x_i\in W,m_i\in M,\\ i& =1,2,\dots ,n(x)\end{array}$$

where x_i and m_i are the location and mark of the ith point, respectively, M is the set of possible types, and W is the observation window in IR². Multitype Strauss models are based on the distances between the points of type i and type j and assume the pair of points will interact if the distance between the points is within the interaction radius r_ij but will have no interaction if the points are further apart than the interaction radius. A multitype Strauss process relies on a set of parameters including the intensity, interaction radius, and interaction parameters to characterize the point process. The intensity parameter β_m(x_i) represents the first-order trend for the type m proteins at location x_i, and the interaction parameter ${\gamma }_{m,m^{\mathrm{\prime }}}$ represents the second-order interaction between points of type m and type $m^{\mathrm{\prime }}$.

The pairwise interaction terms for the model are defined as:

$$c_{m,m^{\mathrm{\prime }}}(x_i,x_j)=\{\begin{array}{c}1if\Vert x_i-x_j\Vert >r_{m,m^{\mathrm{\prime }}}\\ {\gamma }_{m,m^{\mathrm{\prime }}}if\Vert x_i-x_j\Vert \le r_{m,m^{\mathrm{\prime }}}\end{array}$$ (1)

where $r_{m,m^{\mathrm{\prime }}}$ is the interaction distance between the pair of type m and type $m^{\mathrm{\prime }}$ and ${\gamma }_{m,m^{\mathrm{\prime }}}$ are the interaction parameters. All the terms are symmetric, satisfying $r_{m,m^{\mathrm{\prime }}}=r_{m^{\mathrm{\prime }},m}$ and ${\gamma }_{m,m^{\mathrm{\prime }}}={\gamma }_{m^{\mathrm{\prime }},m}$. The interaction radius matrix was estimated using observations based on the sample G-function of the intertype and intratype points.

A Strauss hard-core process combines a hard-core process with a Strauss process. A hard-core process assumes no points are allowed within a radius of some fixed distance around each point. This concept extends to multitype process, where an additional set of hard-core distance parameters is introduced to model the hard-core process between each pair of types. The pairwise interaction terms for the Strauss hard-core process are defined as:

$$c_{m,m^{\mathrm{\prime }}}(x_i,x_j)=\{\begin{array}{c}1if\Vert x_i-x_j\Vert >r_{m,m^{\mathrm{\prime }}}\\ {\gamma }_{m,m^{\mathrm{\prime }}}if{h}_{m,m^{\mathrm{\prime }}}\le \Vert x_i-x_j\Vert \le r_{m,m^{\mathrm{\prime }}}\\ 0if\Vert x_i-x_j\Vert <h_{m,m^{\mathrm{\prime }}}\end{array}$$ (2)

where $h_{m,m^{\mathrm{\prime }}}$ is the hard-core distance satisfying $h_{m,m^{\mathrm{\prime }}}=h_{m^{\mathrm{\prime }},m}$.

The conditional intensity for a multitype Strauss process or a multitype Strauss hard-core process at location u for type m is defined as:

$$\lambda ((u,m((u,|x)={\beta }_m(u)\left[{\prod }_{i,j=1}^{n(x)}{c}_{m,m^{{\mathrm{\prime }}^{\mathrm{\prime }}}}(x_i,x_j)\right]$$

where n(x) is the number of points in the point process.

Given the conditional intensity, the log pseudolikelihood of the model is defined as:

$$\log PL={\sum }_{i-1}^{n(x)}log\lambda ((((x_i,m_i)|x)-{\sum }_{m\in M}{\int }_w\lambda ((u,m|x)du$$

Then, the parameters for a multitype Gibbs model can be obtained by maximizing the pseudolikelihood using the Berman–Turner device (Baddeley & Turner, 2000).

To reduce the number of parameters to estimate, we used the minimum interpoint distances as the estimated hard-core distances. The Strauss interaction parameter, ${\gamma }_{m,m^{\mathrm{\prime }}}$, quantifies between-type interactions and within-type interactions. ${\gamma }_{m,m^{\mathrm{\prime }}}=1$ corresponds to a homogeneous Poisson process, and ${\gamma }_{m,m^{\mathrm{\prime }}}=0$ corresponds to a hard-core process. ${\gamma }_{m,m^{\mathrm{\prime }}}>1$ implies positive associations between points of types m, and $m^{\mathrm{\prime }};{\gamma }_{m,m^{\mathrm{\prime }}}<1$ indicates inhibition between the types m and $m^{\mathrm{\prime }}$.

Models were compared via their corresponding Akaike Information Criterion (AIC) scores (Akaike, 1974).

$$AIC=-2\log L(\hat{\theta })+2p$$

where p is the number of parameters in the model and $L(\hat{\theta })$ is the model's maximum likelihood. Models with higher maximum likelihood and fewer parameters receive lower overall AIC scores, and models with lower AIC scores are preferred. Throughout this analysis, R software (version 4.0.4) was used, and the “spatstat” package was used to the exploratory analysis and modeling of point pattern data (Baddeley & Turner, 2005).

Figure Production

Renderings were produced with Thunderstorm (Ovesný et al., 2014), while representations of point patterns and graphs were produced with spatstat (Baddeley et al., 2015). Images were made into figures using Photoshop and Illustrator (Adobe Systems Inc., San Jose, CA). Scale bars stamped on the images were removed and replaced with bars drawn in Photoshop for improved clarity.

Results

We have produced stable versions of the transformed T cell line, Jurkat, transfected with constructs to express various T cell molecules modified with the addition of fluorescent protein tags to enable their detection by confocal microscopy. In this study, Jurkat T cells stably transfected with LAT-YFP were activated by 3 min of contact with a coverslip coated with an activating anti-TCR antibody followed by fixation, permeabilization, and blocking as previously described (Yi et al., 2016). We chose this cell line to be certain that activated cells containing LAT microclusters were imaged. To find activated cells, YFP fluorescence was observed using TIRF microscopy to visualize LAT-YFP microclusters in areas of close contact with the activating surface. Three activated cells with LAT-YFP microclusters were identified in a TIRF image of the selected field of view; the edges of the cells are outlined in red, orange, and magenta (Fig. 1a). After capturing the TIRF image, a dSTORM series of the unstained field was obtained to determine the level of background localizations due to noise and to visualize fiducial markers.

Fig. 1.

Production of point patterns. (a) Diffraction-limited fluorescence image of LAT-YFP taken before the madSTROM imaging. Scale bar is 5 μm. (b) Renderings of localizations obtained from madSTORM images. Localizations from each antibody staining are shown as average histograms. The cells used for analyses are outlined and identified. Scale bar is 5 μm. (c) Masks used to define microcluster and nonmicrocluster areas. Masks were drawn on cell 1, cell 2, and cell 3 using the LAT-YFP image and the rendering of pTRCζ as guides; darker (green) areas mark microcluster areas, while lighter (yellow) areas mark nonmicrocluster areas.

MadSTORM imaging was then performed sequentially for pLAT, pZAP, pTCRζ, pSLP-76, pPLCγ1, and CD45 (Yi et al., 2016). A dSTORM series of 20,000 images was obtained for each staining. Localizations arising from bursts containing fewer than 1,000 photons were removed as the unstained series showed that most of these were due to noise (Supplementary Fig. 1). Renderings of the remaining localizations from the dSTORM images are shown in Figure 1b; each panel is labeled with the protein visualized in that dSTORM series. Fiducial markers have been removed from these images. The same three cells identified in the TIRF image are outlined in all the panels showing madSTORM localizations. The TIRF image was combined with the rendered localizations of pTCRζ (the first panel in 1B) to produce masks of areas of close contact containing concentrated signaling complexes (dark-colored microcluster areas, shown in green; Fig. 1c) and those areas further from the activating surface with more dispersed signaling complexes (lighter-colored nonmicrocluster areas, shown in yellow; Fig. 1c), while the edges of the cells are again outlined in red, orange, or magenta (Fig. 1c). Generally, each individual microcluster mask follows the contour of one lamellar ridge in close contact with the antibody-coated activating surface. Each outlined cell contains several microcluster masks delineating different contact sites. This segmentation allows us to analyze microcluster and nonmicrocluster areas separately. Areas close to the edges of lamellar ridges are not included in either kind of mask, as it is difficult to know whether they contain concentrated microclusters. Localizations within the masks formed the point patterns used for analysis of microcluster (Fig. 2a) and nonmicrocluster areas (Fig. 2b). Only areas within masks appear in this visualization of the point patterns; the cell outlines are no longer shown. The point patterns have marks attached to the points in addition to the location information; in this case, marks identify the protein in the point pattern. Each visualized protein is shown in a different color, although the large number of localizations makes it difficult to clearly see the distribution of the various proteins.

Fig. 2.

Point patterns in the selected areas. Point patterns produced from the localizations of the proteins extracted by the masks. The top row shows the microcluster point patterns, while the bottom row shows the nonmicrocluster point patterns. Localizations from each protein are shown in a different color as indicated in the scale.

Point process methods can be used to study the relationships between events occurring in space such as our data depicting localizations in 2D space. In Figure 3, we begin the analysis of our point patterns with an examination of the intensity distribution of each protein (Fig. 3, Table 1, and Supplementary Fig. 2), which is a measure of the number of points per unit area of the study region. This is a first-order property defining the spatial characteristics of the localizations. The intensity distributions are shown as heat maps with variable scales to make it easier to see differences in the distribution of each protein. The maps show regions with higher density as warm colors, areas with the most localizations are yellow, while less dense regions are cooler colors, and those with the fewest localizations are dark blue. The density distributions of cell 1 are shown in Figure 3a, those of cell 2 are shown in Figure 3b, and those of cell 3 are shown in Figure 3c. In each set (a–c), the panels in the top row show the distributions within the microcluster masks, while the bottom row shows the densities for the nonmicrocluster areas. These intensity distributions are also shown on a uniform scale in Supplementary Figure 2 to allow direct comparison of the localization densities.

Fig. 3.

Intensity densities of each protein. (a) Intensity densities for cell 1. Top panels show the microcluster areas, and bottom panels show the nonmicrocluster areas. (b) Intensity densities for cell 2. Top panels show the microcluster areas, and bottom panels show the nonmicrocluster areas. (c) Intensity densities for cell 3. Top panels show the microcluster areas, and bottom panels show the nonmicrocluster areas. Scale bars show the density as #/area; warm colors show areas of higher density, while cool colors show areas of lower density.

Table 1.

Intensities of Each Type of Protein in Each Cell's Microcluster and Nonmicrocluster Segment (Intensity = #/nm ×10⁻⁴).

Protein Type	Cell 1			Cell 2			Cell 3			Average Ratio
	Micro	Non	Ratio	Micro	Non	Ratio	Micro	Non	Ratio
TCRζ	5.18	0.41	13.8	2.50	0.44	5.7	3.55	0.38	9.4	9.6
pZAP	3.14	0.55	5.8	1.83	0.65	2.8	2.40	0.59	4.1	4.2
pSLP	5.57	0.41	13.5	1.61	0.20	7.8	2.04	0.24	8.5	10.0
pLAT	1.55	0.48	3.2	0.78	0.44	1.8	0.80	0.54	1.5	2.2
pPLCγ1	1.58	0.62	2.6	0.76	0.34	2.2	1.08	0.58	1.9	2.2
CD45	0.80	0.51	1.6	0.54	0.42	1.3	0.62	0.46	1.3	1.4

The number of detected localizations is dependent on both the affinity of the antibodies used for staining (primarily determined by the antigenicity of the protein or peptide antigen) and the relative abundance of the protein, making it difficult to compare the number of localizations of different proteins. However, comparison of the same protein in different areas is possible. Looking first at cell 1, which showed the best overall staining, the five proteins found in areas of close contact in diffraction-limited studies, pTCRζ, pZAP, pSLP, pLAT, and pPLCγ1 (Bunnell et al., 2002), show increased numbers of localizations in the microcluster areas. This difference is so large that the heat map scales generally differ 5–10-fold for the microcluster and noncluster areas. In contrast, CD45, which is excluded from microclusters in diffraction-limited imaging (Bunnell et al., 2002), shows more equal numbers of localizations in microcluster and nonmicrocluster areas. These differences are quantified in Table 1. Another striking feature of the density plots is heterogeneity, both between different masks and within a given mask.

We can also look at whether different proteins show similar patterns, that is, whether they tend to show higher and lower densities in the same places. The microclusters masks of pTCRζ, pZAP, and pSLP, which also show the best staining (see Fig. 1b), have very similar density patterns. For example, in the microcluster masks of cell 1 (Fig. 3a), pTCRζ, pZAP, and pSLP generally show high and low densities in the same places. In contrast, the density distributions of pLAT and pPLCγ1 are similar to the others in many places but differ in several of the microcluster masks. For example, pLAT is dense in the center of the large central mask of cell 1 (indicated with an *) where pTCRζ, pZAP, and pSLP show lower relative densities. pPLCγ1 displays higher relative density in parts of three smaller masks (indicated with triangles) where the other proteins show lower relative densities. Previous studies have shown a separation between clusters of pTCRζ and LAT (Lillemeier et al., 2010; Sherman, 2011), so it is not surprising that these two show different density distributions. However, SLP binds to LAT via Gads so our expectation was that SLP should follow the LAT distribution rather than pTCRζ. Previously, it was shown that there are LAT pools separate from SLP (Barr et al., 2006), so areas of dense LAT with low SLP are explicable (such as the area marked *), but we also see a few areas with higher relative SLP (marked with a circle on the pSLP panel) or PLCγ1 (marked with triangles) density and lower relative LAT density, which is surprising. CD45 also shows different density patterns from the other proteins within most of the microcluster masks; that is, the pattern of high and low intensities for CD45 does not match the pattern of any other protein. CD45 also shows substantially lower average intensities in the microcluster areas of all three cells (Table 1). In the nonmicrocluster areas, no two proteins show the same pattern across all the masks, suggesting more random distributions. Cells 2 and 3 display dimmer staining and fewer localizations for all the proteins, but there is an increased concentration of signaling complex proteins in the microcluster areas of both cell 2 and cell 3, with the quantification shown in Table 1. Once again, the density distributions of pTCRζ, pZAP, and pSLP are correlated, but in these two cells, there are one or two masks that do not match. The density distributions in the nonmicrocluster areas of cell 2 and cell 3 are generally different for each protein. Overall, this analysis showed a striking concentration of the five proteins known to be in signaling complexes within the microcluster areas, confirming the expected first-order distributions. There are also indications that the intensities of pTCRζ, pZAP, and pSLP are correlated in the microcluster areas. However, there is a surprising inconsistency in the association of the density distributions of pSLP, pLAT, and pPLCγ1 within microclusters, despite our expectation that these proteins would all be highly correlated as they bind to each other. Therefore, we decided to apply second-order methods to further explore the spatial dependencies within the various point patterns, particularly within the microcluster areas. Second-order properties measure the interactions between the observed points. We used two second-order methods to examine the spatial dependence within the various point patterns, particularly within the microcluster areas.

To begin studying the dependency of points between different types, that is, the relationship of one protein to itself and others, we plotted cross-type J-functions (Fig. 4), which compare nearest-neighbor and empty-space distances (van Lieshout & Baddeley, 1996) (see Materials and Methods). The J-function has several advantages including that it is easy to evaluate and is insensitive to edge effects. In a random distribution of points, nearest-neighbor and empty-space distances have the same probability distribution, resulting in equal distances to either another point or an empty space. Thus, CSR or no relationship appears as a red dotted horizontal line with a benchmark of 1 in each graph. Deviation above this line shows more dispersion of the two proteins than expected, while deviation below the line shows clustering between the two proteins being analyzed. The gray shading area surrounding the red dotted line identifies the area containing 95% of randomly generated curves for each graph. All the individual J-function graphs are completely gray indicating that the observed deviations from CSR overlap with some randomly generated curves. Although the observed deviations are small, they do show consistent trends that are worth examining in more detail.

Fig. 4.

Cross-type J-functions for all protein–protein combinations. (a) Cross-type J-functions for cell 1. Left panels show the microcluster areas, and right panels show the nonmicrocluster areas. (b) Cross-type J-functions for cell 2. Left panels show the microcluster areas, and right panels show the nonmicrocluster areas. (c) Cross-type J-functions for cell 3. Left panels show the microcluster areas, and right panels show the nonmicrocluster areas. J(r) is on the y axis, and r is on the x axis. Rows show J_{i, j}(r) where i is the row label and j is the column label. Columns show J_j.i(r) where i is the row label and j is the column label.

The J-functions of cell 1 are shown in Figure 4a, those of cell 2 are shown in Figure 4b, and those of cell 3 are shown in Figure 4c. The left panels display analyses of the microcluster regions, while the right panels show those of the nonmicrocluster regions. The graphs arranged in a horizontal row show how each type of protein is arranged around the protein identified to the left of the row; that is, the first row of graphs shows distances to other proteins from central pTCRζ molecules. The vertical columns of graphs show distances from the protein identified at the top to other proteins; that is, the first column consists of graphs showing distances to pTCRζ from central molecules of the other proteins. The graph panels forming a diagonal line from the upper left to lower right display distances from a central molecule to other molecules of the same protein. Supplementary Table 1 contains a chart showing how the individual graphs of each J-function are arranged.

Looking first at the J-functions of molecules in the microcluster areas of cell 1, the first panel of Figure 4a (upper left corner) reveals that pTCRζ is more clustered with itself than would be expected from CSR as the graph falls below the red dotted line. An examination of the remaining graphs along the diagonal shows that pZAP, pSLP, and pLAT all group with themselves, while pPLCγ1 shows less clustering with itself, and CD45 shows only a small interaction with itself. An examination of the first row shows that pZAP, pSLP, and pLAT are more clustered around pTCRζ than expected, while pPLCγ1 and CD45 show more nearly random distributions around pTCRζ molecules. The next row shows clustering of pTCRζ, pSLP, and pLAT around pZAP with decreased clustering of pPLC γ1 and CD45. Similar patterns are seen in the remaining rows. The last row shows little clustering of the other proteins around CD45.

The graphs in the columns are complimentary to the rows, showing how the protein labeled at the top is arranged around different central molecules. The first column shows that pTCRζ clusters around pZAP, pSLP, and pLAT with less clustering around pPLC γ1 and very little around CD45. Similar clustering of pZAP appears in the second column. These clustering patterns are seen in the microcluster areas of cells 2 and 3 (Figs. 4b and 3c, left panels), except that pLAT shows less clustering in these two cells, and pSLP, pPLCγ1, and CD45 show little clustering around pLAT. In contrast, in the nonmicrocluster areas (right panels of Figs. 4a–4c), the interactions between all combinations of molecules show little deviation from CSR in all cells.

Thus, these J-function plots indicate that there are small, but detectable deviations from CSR in the arrangements of most of the proteins found in signaling complexes in the microcluster regions. However, CD45, which is not found in signaling complexes, has J-functions that are all very close to CSR, including the arrangement of the other proteins around CD45 (bottom row in all panels), the arrangement of CD45 around other proteins (last column in each panel), and the arrangement of CD45 with itself (last graph in the bottom row and last column). Thus, there is little organization of this protein in activated Jurkat T cells even in the microcluster regions. In the nonmicrocluster regions, all the J-functions of all of the proteins are close to CSR. So, nonrandom distributions appear to be restricted to proteins found in signaling complexes in the microcluster areas. Having seen these interesting developments, we sought to confirm these observations with another commonly used statistic, the L-function.

Figure 5 shows graphs of the centered cross-type L-function, which is a scaled version of Ripley's K-function (Besag, 1977) (see Materials and Methods). Ripley's K-function is a widely used second-order statistic that calculates the expected number of points in a distance r around an arbitrary, central point (Ripley, 1976; Baddeley & Turner, 2005), but it can be difficult to interpret as the benchmark of a completely random distribution is πr², which is not linear. The L-function is a normalized version of Ripley's K-function that produces a linear benchmark (Besag, 1977), and a centered L-function is further modified to produce a linear benchmark value of 0. When the observed L-function value L_ij(r) − r is higher than this theoretical benchmark, the point process shows that more type j points are around the type i point at the given distance r, indicating clustering. When the empirical values are below the theoretical line, the type j points are fewer than expected around type i points at distance r, indicating dispersion. To test if the deviation from the theoretical values is statistically significant, envelopes obtained through Monte Carlo simulations with permutation of labels are used to indicate the confidence bounds. The gray shading surrounding the red dotted line identifies the area containing 95% of randomly generated curves obtained by Monte Carlo simulations for each graph. Plots that fall within this area overlap with some random distributions. Deviations that lie outside of this envelope are generally considered significant. The null hypothesis tested is CSR for interactions of a protein with itself and CSRI for interactions between different types of protein.

Fig. 5.

Centered cross-type L-functions for all protein–protein combinations. (a) Cross-type L-functions for cell 1. Left panels show the microcluster areas, and right panels show the nonmicrocluster areas. (b) Cross-type L-functions for cell 2. Left panels show the microcluster areas, and right panels show the nonmicrocluster areas. (c) Cross-type L-functions for cell 3. Left panels show the microcluster areas, and right panels show the nonmicrocluster areas. L(r) − r is on the y axis, and r is on the x axis. Rows show L_{i, j}(r) − r where i is the row label and j is the column label. Columns show L_j.i(r) − r where i is the row label and j is the column label.

The L-functions of cell 1 are shown in Figure 5a, those of cell 2 are shown in Figure 5b, and those of cell 3 are shown in Figure 5c. The left panels display analyses of the microcluster regions, while the right panels show the analyses of the nonmicrocluster regions. The graphs arranged in a horizontal row show how the other kinds of proteins are arranged around the protein identified to the left of the row; that is, the first row of graphs shows how various proteins are arranged around central pTCRζ molecules (these graphs show L_ij(r) − r, where i is pTCRζ and j represents the other protein). The vertical columns of graphs then show how the protein identified at the top is arranged around the other proteins; that is, the first column consists of graphs showing how pTCRζ is arranged around the other proteins (these graphs show L_ji(r) − r, where i is pTCRζ and j represents the other protein). The graph panels forming a diagonal line from upper left to lower right display the L-function for each protein arranged around itself. Supplementary Table 2 contains a chart showing how the individual graphs of each L-function are arranged.

Beginning with the L-functions of points within the microcluster regions of cell 1, the first graph of Figure 5a (upper left corner) demonstrates that there are more pTCRζ molecules found around central pTCRζ molecules than would be expected from CSR, confirming that pTCRζ clusters with itself. An examination of the remaining graphs along the diagonal shows that pZAP, pSLP, pLAT, and pPLC γ1 all cluster with themselves. These effects are above the 95% confidence levels, that is, the graphs rise above the gray shaded areas surrounding the red dotted, benchmark line and are considered significant deviations from CSR. CD45 shows an interaction with itself only on the left side of the graph which shows the results for CD45 molecules found at short distances from central CD45 molecules. The right side of the graph shows results for CD45 molecules at longer distances from central CD45 molecules; here, there is no significant deviation from CSR. So, all proteins do tend to cluster with themselves, with CD45 demonstrating less self-clustering.

Moving across the pTCRζ row, there are also more molecules of pZAP, pSLP, pLAT, and pPLCγ1 around pTCRζ than expected from CSRI, indicating that all of these proteins cluster around pTCRζ. Again, these effects are above the 95% confidence levels; the graphs rise above the gray shaded areas surrounding the red dotted, benchmark line and are significant deviations from CSRI. The distribution of CD45 in the last graph of this row is different. On the left side of the graph, the results for CD45 molecules found at short distances from the central pTCRζ molecule show that the number of CD45 molecules is close to the expected value for CSRI and the calculated L-function falls within the gray shaded area. Thus, there is little clustering of CD45 around TCRζ molecules at short distances. The right side of the graph shows results for CD45 molecules at longer distances from the central pTCRζ molecule. Now, CD45 shows significant dispersion from pTCRζ as the L-function falls below the benchmark line and the gray shading representing the 95% confidence level. The next row shows the L-functions of molecules surrounding pZAP. Again, there are more molecules of pTCRζ, pSLP, pLAT, and pPLCγ1 around pZAP than expected from CSRI and the deviations appear to be significant at all distances. Again, CD45 is close to the expected value for CSRI on the left side of its graph and the calculated L-function falls within the gray shaded area, while on the right side of the graph, CD45 shows significant dispersion from pZAP. Similar results are seen in the remaining rows; the molecules found in signaling complexes show significant clustering around pSLP, pLAT, and pPLCγ1, while there is little clustering of CD45. Interestingly, the molecules found in signaling complexes show a tendency to cluster around central molecules of CD45 (bottom row).

The graphs in the columns are complimentary to the rows, showing how the protein identified at the top is arranged around different central molecules. The arrangement of pTCRζ around the other proteins is seen in the graphs in the first column that show significant clustering of pTCRζ around all the other molecules, including CD45. The remaining columns confirm that all the signaling complex molecules show clustering around each other and around CD45. CD45 shows no significant clustering around the other proteins at short distances and no clustering or dispersion at longer distances.

Similar results are seen in the microcluster areas of cell 2 (Fig. 5b, left panels), with less significant clustering around pLAT and pPLCγ1 and a bit more clustering of CD45 around the other proteins. In the microcluster area of cell 3 (Fig. 5c, left panels), there is very little clustering of pLAT and pPLCγ1 around other proteins, along with no clustering of CD45. Overall, the L-functions demonstrate significant reciprocal clustering of the proteins visualized in signaling complexes, while CD45 clusters only with itself at short distances. The extent of clustering is correlated with the strength of the staining which could indicate that pLAT and pPLCγ1 may appear less clustered because we did not detect enough molecules to visualize the pattern. The nonmicrocluster areas of all cells (Figs. 5a–5c, right panels) show only occasional variable clustering of proteins, with little significant clustering of any protein. Within these nonmicrocluster areas, there is no detectable pattern; proteins that show clustering in one cell generally do not show clustering in the other cells. Thus, the L-function verifies and strengthens the conclusions from the J-function. In the microcluster areas, pTCRζ, pZAP, and pSLP are all clearly clustered around each other, while pLAT and pPLCγ1 show some clustering. CD45 shows little clustering except with itself. In the nonmicrocluster areas, the proteins show little to no clustering, indicating a random dispersion in these regions.

These exploratory analyses produced evidence of nanostructure within the microcluster areas, but we were intrigued by the variability of intensity distributions of different masks within the same cell and the differences in J- and L-functions between the different cells. So, we examined the second-order properties of each mask in cell 1 to study this heterogeneity. These masks are numbered in Supplementary Figure 3.

Analysis of the cross-type J-functions for each mask showed some heterogeneity in the molecular interactions detected in the microcluster areas (Supplementary Fig. 4). The graphs in each panel are arranged as shown in Supplementary Table 1  Supplementary Table 2. Noticeable clustering was apparent mainly in microcluster mask 4, with little clustering seen in the other microcluster masks, even in interactions of proteins with themselves. Mask 4 contains the most localizations, reinforcing our sense that the appearance of order is correlated with the density of the localizations. The molecules still show random interactions in the nonmicrocluster areas (Supplementary Fig. 5).

Centered cross-type L-functions (Fig. 6) show significant clustering in all seven masks in contrast to the J-function which showed clustering mostly confined to one mask. The graphs in each panel are arranged as shown in Supplementary Table 2. The graphs along the diagonals of the panels show that all proteins cluster with themselves in all the masks and recapitulate the results of the combined microcluster mask shown in Figure 5 with only minor variations. The first three rows of each set of graphs show the results for clustering around pTCRζ, pZAP, and pSLP, respectively. They all show significant reciprocal clustering as they did the combined analyses. However, the clustering of other molecules changes in the different masks. The clustering of pLAT around pTCRζ is variable, with good clustering seen in masks 4, 5, and 7 and some clustering in masks 1, 2, and 3, but no clustering in mask 6. pLAT shows the same pattern of clustering around pZAP, but clusters well around pSLP in all masks except for masks 1 and 6, which show less clustering. So, we now see a hint of our expected pattern where LAT clusters more with SLP than with pTCRζ or ZAP. The clustering of other molecules around pLAT varies in the different masks: good in masks 1, 5, and 7 but inconsistent in the other four masks. The organization of pPLCγ1 is quite variable with clustering most often detected around pTCRζ, pZAP, and pSLP along with clustering of these proteins around pPLCγ1. The last row shows that pTCRζ and pZAP usually cluster around CD45 except in mask 5. pSLP, pLAT, or pPLCγ1 do not cluster around CD45 in most, but not all masks. The last column demonstrates that CD45 does not generally cluster around the other molecules, except in masks 1 and 7 which show significant but erratic clustering of CD45. The correlation between dense localization distributions and clustering of proteins is less pronounced in these analyses, although the bright masks 3, 4, 5, and 7 do have more pairs that show clustering. Interestingly, clustering was detected in the individual nonmicrocluster masks (Fig. 7): results that differ from the analyses of the combined nonmicrocluster masks (Fig. 5). Masks 9 and 10, the brightest of the nonmicrocluster masks, show significant clustering of most molecules. pTCRζ, pZAP, and pSLP show significant clustering in all of the nonmicrocluster masks. pLAT and pPLCγ1 have similar patterns where the most prominent clustering is seen in mask 10, which also shows clustering around CD45. Of the nonmicrocluster masks, only the brightest, mask 10, shows consistent clustering of most proteins. Overall, more structure is found in the microcluster masks, and of all the proteins, CD45 still shows the least organization, but there is substantial variation in the nanostructure found in the different masks.

Fig. 6.

Cross-type L-functions of individual microcluster masks labeled by mask number. L(r) − r is on the y axis, and r is on the x axis. Rows show L_{i, j}(r) − r where i is the row label and j is the column label. Columns show L_j.i(r) − r where i is the row label and j is the column label.

Fig. 7.

Cross-type L-functions of individual nonmicrocluster masks labeled by mask number. L(r) − r is on the y axis, and r is on the x axis. Rows show L_{i, j}(r) − r where i is the row label and j is the column label. Columns show L_j.i(r) − r where i is the row label and j is the column label.

Poisson models and Gibbs models were applied to examine the trends in intensity along with pairwise interaction patterns among the points (Baddeley and Turner, 2000, 2006; Illian et al., 2008). In Poisson models, logical regression is used to predict the probability of the presence of a localization at each point in the area being studied; by definition, each point is independent. Strauss models are extensions of the Gibbs hard-core process that use the “potential energy” determined by attractive and repulsive interactions between points to calculate the likelihood of a particular spatial arrangement of the points. Hard-core processes account for the space occupied by each point. Due to the large amount of computational power needed for this modeling, we were unable to model the entire data set or even a data from a single mask containing many points. Therefore, we modeled the data from one small, low-density microcluster mask (mask 1) and one nonmicrocluster mask (mask 8) (Supplementary Figs. 6a and 6b). AIC was used to compare the models produced for these two areas (Akaike, 1974). AIC rewards models for fitting the data and penalizes models for complexity; the model with the lowest AIC is considered the best. The AIC scores showed that Gibbs models with Strauss hard-core interaction terms outperformed Poisson models without interaction terms, indicating the strong presence of interactions within the pattern (Supplementary Table 3). The Strauss hard-core model with quadratic trend terms required fewer parameters with only a slightly higher AIC, so it was used for analysis. In this model, the microcluster areas had more statistically significant interaction parameters, indicating more pairwise interactions were found in microcluster areas compared with the nonmicrocluster areas (Supplementary Tables 4 and 5 and Supplementary Fig. 6c). Thus, the best model shows more structure in the microcluster areas.

The model was then used to predict the intensity distribution for each type of protein. The fitted trends, which show how each type of protein is distributed according to the model, are presented in Figure 8. The predicted intensity trend of each protein is displayed as a heat map superimposed on the points arranged according to the Strauss hard-core model with quadratic trend terms. These fitted trends captured some of the intensity distribution but did not match the entire pattern. Using a fitted conditional intensity to account for dependencies between the different points did not improve the fit (Supplementary Fig. 6d). The discrepancies between the fitted trend and the actual intensity could be due to the edge effects, which could be significant in these small masks, or the simplicity of the model. The model may fail if different interactions take place in different parts of a single mask, as we think likely. For example, changes in the density of one species may affect the interactions with other kinds of proteins. In the future, it might be possible to incorporate more covariates into the model to help explain the characteristics of the point pattern produced directly from the data. These problems highlight the difficulties in producing adequate mathematical models for the complex process of T cell signaling.

Fig. 8.

Fitted trends according to the model for selected masks. (a) Fitted trend for each type of protein for the microcluster mask. (b) Fitted trend for each type of protein for the nonmicrocluster mask. Scale bars show the density as #/area; warm colors show areas of higher density while cool colors show areas of lower density.

Discussion and Concluding Remarks

Analysis of the molecular organization of TCR signaling proteins should lead to a better understanding of the interactions between the proteins and the relationship of nanostructure to T cell activation. Diffraction-limited imaging studies have already established the importance of the micron-scale organization of signaling proteins following engagement of the TCR. Yet, the molecular architecture or nanostructure within these large microclusters is not well understood despite numerous studies using imaging methods that can detect individual molecules (Dietz & Heilemann, 2019; Brameshuber et al., 2022). Thus, many questions remain regarding the spatial arrangement of signaling complexes, the role of physical forces in organizing these complexes, and how the configuration of signaling complexes influences the final activation of T cells. The development of a more precise picture has been hindered by variability in the results of different studies, possibly due to differences in SMLM techniques, sample preparation, and analysis methods.

Early SMLM studies of T cells used PALM, where switchable fluorescent proteins are used to produce localizations (Lillemeier et al., 2010; Sherman et al., 2011). This technique requires overexpression of the proteins under study, which increases the number of molecules that may induce interactions, thus possibly creating nanostructure that does not normally exist. Interactions can also be induced by some of the fluorescent proteins used in these studies (Wang et al., 2014; Barr et al., 2016). Additionally, PALM does not discriminate between active and inactive molecules, making it more difficult to determine how nanostructure is related to activation. Other studies employed dSTORM imaging, where a labeled antibody is used to produce localizations (Hu et al., 2016; Pageon et al., 2016). Often, two antibodies are applied, one that recognizes the protein of interest along with a labeled second antibody that recognizes the first one. The use of two antibodies increases the distance of the fluorescent probe from the target protein and decreases resolution. Frequently, the labeled antibody contains multiple fluorophores per molecule, which can lead to apparent, but artifactual clustering. Both PALM and dSTORM techniques are also limited by problems with aligning images, especially if fluorophores of different colors are used to produce images containing localizations of more than one protein in the same sample. Finally, SMLM data is subject to errors in undercounting the number of localizations if the detection efficiency is low (Patterson et al., 2010) as well as overcounting if the same fluorophore or label is counted repeatedly (Annibale et al., 2011; Sengupta et al., 2011). By employing the madSTORM technique based on sequential binding of multiple antibodies, we were able to use the most efficient dye, Alexa 647, for labeling all the antibodies to produce easily detected probes and avoid alignment problems (Yi et al., 2016). We also used directly conjugated antibodies to decrease the distance from the probe to the target. We chose antibodies specific for phosphorylated, active forms of five of our proteins of interest. This allowed us to visualize all these activated proteins in the same sample, thereby avoiding the problem of trying to build a comprehensive picture from images of different samples. As our study shows significant variability in the nanostructure found in different cells and different areas of the same cell, directly observing the locations of many proteins simultaneously is crucial to obtaining an accurate picture of activated protein organization in T cells. We did not attempt to group localizations, preferring to error on the side of overcounting rather than undercounting localizations (Pageon et al., 2016). Although overcounting can produce spurious clusters that exaggerate nanostructure, we do not think that was a significant problem here, as our study is notable for the number of areas showing random or near random distributions of signaling proteins rather than extensive clustering.

SMLM studies, including those of T cells, require the use of statistical or segmentation analyses to make sense of the patterns in the localizations (Khater et al., 2020; Wu et al., 2020). Early T cell studies used spatial descriptive statistics such as Ripley's K (Ripley, 1976) or the related O ring statistic (Wiegand and Moloney, 2004) to explain how localizations are arranged (Lillemeier et al., 2010; Sherman et al., 2011; Hu et al., 2016; Pageon et al., 2016). Ripley's K, a cumulative distribution function, is more robust with respect to noise (Khater et al., 2020), while the O ring statistic, a probability density function, is better at distinguishing clusters at specific distances (Wiegand & Moloney, 2004). Here, we used the J-function, another cumulative statistic (van Lieshout & Baddeley, 1996), and the L-function (Besag, 1977), a common variant of Ripley's K, to assess clustering behavior by producing pairwise comparisons of all the proteins being studied. The J-function has not been used in previous T cell nanocluster studies and has the advantage of being insensitive to edge effects (Khater et al., 2020), an important quality given the small ROIs used in our study. A disadvantage of the J-function is that it tends to report less order than other functions (Baddeley et al., 2015). Ripley's K and related methods, such as the L-function, are regarded as some of the most effective analytical methods for exploring patterns at various scales. The L-function is a linearized version of Ripley's K with a benchmark of 0 that makes the graphs easy to understand, as deviations from CSR are readily apparent. A disadvantage of all functions based on Ripley's K is their sensitivity to large-scale heterogeneities (Khater et al., 2020). Future research could try applying weighted versions of the K- and L-functions to SMLM data (Veen & Schoenberg, 2006; Adelfio & Schoenberg, 2008).

Given the limitations of SMLM methods and analyses along with the differences in study design, it is not surprising that each study has come to somewhat different conclusions. By carefully choosing ROIs that separate areas likely to have different patterns of organization, we sought to minimize the problems of heterogeneity in the study regions. The combination of two analysis methods allowed us to minimize edge effects with the J-function and then confirm the trends we found with the more effective clustering algorithm of the L-function. Previous studies were limited in the number of proteins that could be analyzed in a single sample. Although we still had to perform pairwise comparisons, we were able to produce a complete set of interactions between six different proteins, allowing us to directly compare their interactions. These improvements should produce a more meaningful analysis of clustering of multiple molecular pairs within the same sample. The dSTORM data of multiple proteins combined with point pattern analysis of carefully delineated masks that separate microcluster areas from unactivated portions of the membrane clearly show that nanoscale organization is only found in the microcluster areas, which are the areas that are in close contact with the activating surface. By comparing multiple microcluster masks, we were able to document heterogeneity in the molecular interactions in different microcluster areas.

Although previous SMLM studies showed that numerous lymphocyte proteins form small clusters with themselves (Balagopalan et al., 2020; Lamerton et al., 2021), including pTCRζ and LAT (Lillemeier et al., 2010; Sherman et al., 2011; Jensen et al., 2022), our results show that self-association is not uniform across the cell surface. All six proteins investigated in this study exhibit self-clustering most prominently in the parts of the membrane that contain visible microclusters. Self-clustering is greatly decreased in nonmicrocluster areas. These results indicate that the nanoscale order of signaling proteins is modulated by membrane location and that areas of low protein density are less likely to promote self-clustering. It also argues that nanoclustering is not an artifact because an imaging artifact would generate the same results in microcluster and nonmicrocluster areas. Overall, we observe consistent support for the clustering of proteins with themselves within the microcluster areas in all three cells that were studied. We found near random distributions of signaling complex proteins in nonmicrocluster areas although these proteins were all phosphorylated and therefore capable of signaling. This result is consistent with the binding of Grb2 to all LAT clusters, even very small ones (Sherman et al., 2011). The origin of these dispersed, activated proteins is unclear, although we know endocytosis and recycling of signaling complex components are prominent features of TCR signaling (Balagopalan et al., 2009). Likewise, the function of these activated proteins in nonmicrocluster areas is unknown; they could produce signals that are different from those generated in microcluster areas.

The sites of first contact between a T cell and APC or activating surface appear to be the tips of preexisting microvilli (Jung et al., 2016; Cai et al., 2017; Razvag et al., 2019; Ghosh et al., 2020), but whether there is substantial nanoclustering of proteins within the tips of microvilli is not known. It is possible that much of the clustering reported in quiescent cells (Lillemeier et al., 2010; Sherman et al., 2011) reflects the confinement of proteins in the small area of microvilli tips. Our results show a large increase in the density of localizations of pTCRζ, pZAP, pSLP, pLAT, and pPLC γ1 in microcluster areas, providing further evidence for the concentration of these proteins in microvilli. However, the intensity distributions in the areas of close contact show neither a series of dense circles as expected from images of microvilli dispersed along the contact site nor a uniform dense distribution as expected from concatenated microvilli filling the contact surface. Instead, we see distinct areas of higher concentration along an individual mask that outlines a membrane ruffle containing microvilli in contact with the activating surface. This suggests that the distribution of proteins found in signaling complexes changes after the first contact of microvilli. The distribution of CD45 has been controversial. Some studies have shown distinct separation of CD45 from microclusters (Bunnell et al., 2002; Chang et al., 2016; Pageon et al., 2016), while others showed a more uniform distribution (Cai et al., 2017; Ghosh et al., 2020). Here, CD45 did not show a large increase in density within the microcluster areas but was not excluded from them. Thus, our results support a uniform distribution of CD45.

Early SMLM studies also found interactions between different T cell signaling proteins and found clusters containing pTCRζ were separate from those containing LAT (Lillemeier et al., 2010; Sherman et al., 2011). All of our data, including the cross-type J- and L-functions and our preliminary modeling, showed that the microcluster areas are clearly different from the nonmicrocluster areas, with little detectable cross-clustering, that is, clustering of one protein around another species, in the nonmicrocluster areas. In the microcluster areas, we found cross-clustering of most proteins found in TCR signaling complexes. While these results are consistent with the expected clustering of signaling complex proteins with each other in microclusters, the details of these interactions are not what we anticipated. Generally, there were more random distributions than previous studies showed, particularly in the analyses using the J-function. In addition, we expected that clusters of pTCRζ which binds pZAP would be separate from clusters of pLAT which binds pSLP and pPLCγ1. Instead, our analyses show clustering of pTCRζ with both pZAP and pSLP and no indication that pLAT and pPLCγ1 cluster more with pSLP than pTCRζ. In particular, we did not find evidence of previously reported patterns of interactions of SLP, LAT, and PLCγ1 (Sherman et al., 2011; Barr et al., 2016; Sherman et al., 2016; Yi et al., 2019). Although we cannot explain all of the discrepancies, the techniques used here differ from those studies in many ways including the use of dSTORM instead of PALM, the imaging temperature, and the use of phospho-specific antibodies. Thus, we are examining different pools of proteins under different conditions from previous experiments. If clustering is weak with only small deviations from random distributions, as shown here in T cells, the final result will depend greatly on the statistical methods used. Our use of J- and L-functions is different from the previous studies which will influence the reported pattern of interactions. In particular, the J-function is more likely to describe random distributions. Our results from this algorithm suggest that previous studies may have overstated the extent of nanostructure in T cells. We also noted a tendency to find more order in denser, brighter cells and microcluster areas. Consequently, the selection of samples may also be responsible for some inconsistencies, as the selection of bright samples may lead to detection of more ordered structures.

A comparison of the cross-type J-function or cross-type L-function between the various microcluster masks and the whole data set showed the individual masks do not share all the characteristics of the whole data set. In addition, different microcluster masks often display different clustering patterns. This heterogeneity in nanostructure contributes to the evidence that structural rearrangements, presumably caused by liquid–liquid phase separations (Banani et al., 2017; Case et al., 2019), occur in different locations after the first contact of the microvilli. High concentrations of ligand, as used here, could contribute to the production of areas with greater clustering of signal complexes and make phase transitions more likely (Pageon et al., 2016). Several of the proteins required for T cell signaling that are studied here contain disordered domains which may also help drive phase transitions (Cornish et al., 2020). Accordingly, small variations in the distribution of these crucial proteins could produce a diverse set of phase domains. Once denser areas are induced by phase transition, the proximity of molecules could promote further binding. In the case of T cells, this could drive oligomerization of LAT (Houtman et al., 2006) and SLP-76 (Coussens et al., 2013). Perhaps cross-linking induced by proximity at higher densities can help stabilize nanostructure.

Our results provide evidence for the restriction of nanostructure of proteins found in TCR-based signaling complexes to the microcluster areas of activated T cells. Furthermore, even within microcluster areas, we observed extensive heterogeneity in the arrangement of these proteins, including zones of randomly distributed molecules and areas containing organized clusters. Many commonly used clustering algorithms may fail to capture the true organization of a study area when different parts of the region of interest are organized differently. This may explain the difficulties in producing a consensus picture of nanostructure in T cells, as often the study areas have been large and may contain more than one kind of nanostructure. We also found significant cell-to-cell variability in nanostructure, so the selection of which cells are analyzed will affect study outcomes. Because the organization of signaling complexes depends on numerous weak interactions that produce flexible and variable structures, exactly which kinds of nanostructures are reported may change in different experiments. The organizational variations seen here may be related to functional outcomes, emphasizing the importance of understanding the true range of T cell nanostructure. The search for an accurate representation of signaling protein organization clearly requires more high-quality SMLM data, produced with an understanding of pitfalls of sample selection.

Availability of Data and Materials

Computer codes used to process the localizations can be accessed by contacting Dr. Lawrence Samelson, while computer code for point pattern analyses and modeling can be accessed by contacting Dr. Frederic Schoenberg.

Supplementary Material

To view supplementary material for this article, please visit https://doi.org/10.1093/micmic/ozad072.

Financial Support

This research was supported by the National Cancer Institute, USA (Intramural funding of The Center for Cancer Research) Intramural Research Programs of the National Cancer Institute (The Center for Cancer Research).