Information transfer in mammalian glycan-based communication

  1. Felix F Fuchsberger
  2. Dongyoon Kim
  3. Natalia Baranova
  4. Hanka Vrban
  5. Marten Kagelmacher
  6. Robert Wawrzinek
  7. Christoph Rademacher  Is a corresponding author
  1. Department of Pharmaceutical Sciences, University of Vienna, Austria
  2. Department of Biomolecular Systems, Max Planck Institute of Colloids and Interfaces, Germany
  3. Department of Microbiology, Immunology and Genetics University of Vienna, Max F. Perutz Labs, Austria

Abstract

Glycan-binding proteins, so-called lectins, are exposed on mammalian cell surfaces and decipher the information encoded within glycans translating it into biochemical signal transduction pathways in the cell. These glycan-lectin communication pathways are complex and difficult to analyze. However, quantitative data with single-cell resolution provide means to disentangle the associated signaling cascades. We chose C-type lectin receptors (CTLs) expressed on immune cells as a model system to study their capacity to transmit information encoded in glycans of incoming particles. In particular, we used nuclear factor kappa-B-reporter cell lines expressing DC-specific ICAM-3–grabbing nonintegrin (DC-SIGN), macrophage C-type lectin (MCL), dectin-1, dectin-2, and macrophage-inducible C-type lectin (MINCLE), as well as TNFαR and TLR-1&2 in monocytic cell lines and compared their transmission of glycan-encoded information. All receptors transmit information with similar signaling capacity, except dectin-2. This lectin was identified to be less efficient in information transmission compared to the other CTLs, and even when the sensitivity of the dectin-2 pathway was enhanced by overexpression of its co-receptor FcRγ, its transmitted information was not. Next, we expanded our investigation toward the integration of multiple signal transduction pathways including synergistic lectins, which is crucial during pathogen recognition. We show how the signaling capacity of lectin receptors using a similar signal transduction pathway (dectin-1 and dectin-2) is being integrated by compromising between the lectins. In contrast, co-expression of MCL synergistically enhanced the dectin-2 signaling capacity, particularly at low-glycan stimulant concentration. By using dectin-2 and other lectins as examples, we demonstrate how signaling capacity of dectin-2 is modulated in the presence of other lectins, and therefore, the findings provide insight into how immune cells translate glycan information using multivalent interactions.

Editor's evaluation

This manuscript lays out the framework for addressing an important challenge in our understanding of cellular signal transduction: how complex extracellular inputs can be detected and processed using multiple receptors. This problem is addressed in the context of glycan receptors lectins, mediating very common but still not completely understood cell-cell interactions. Using information capacity analysis, the study addresses the importance of glycan input measurement by multiple receptors on the immune cells, showing how the signal detection can benefit from receptor crosstalk.

https://doi.org/10.7554/eLife.69415.sa0

Introduction

Glycans are present in all living cells and play a key role in many essential biological processes including development, differentiation, and immunity. Being surface exposed, glycans often encode for information in cellular communication such as self-/non-self-discrimination, cellular identity, and homing as well as apoptosis markers (Bode et al., 2019; Maverakis et al., 2015; Williams, 2017). Other than linear biopolymers, such as proteins and nucleic acids, glycans are branched structures, where subtle changes in the glycosidic bonds between each monomer can carry essential pieces of information. Adding to this complexity, glycans are products of large cellular machinery and are therefore not directly encoded by the genome (Cummings, 2009). Besides their composition, the recognition of glycans by their receptors is complicated, particularly due to the lack of specificity. Glycans are recognized by lectins, yet no glycan is recognized by a single receptor, and no individual lectin is highly specific for only one glycan. Additionally, affinities are low, and interactions often depend on the multivalency of both the receptor and the ligand. Overall, since alterations of the glycocalyx do not function as a deterministic on/off switch but rather a progressive tuning of the cellular response, glycan lectin communication should be considered as a stochastically behaving system, rather than a deterministic one (Dennis, 2015).

Many lectin receptors serve as triggers for multiple immunological signaling pathways, often funneling down to NF-κB (nuclear factor kappa-B) as a transcription factor. In this work, we focus on C-type lectin receptors (CTLs). MINCLE (macrophage-inducible C-type lectin), for example, is a CTL involved in the recognition of pathogens as well as self-damage (Miyake et al., 2015; Williams, 2017). MINCLE and its close relative dectin-2 (dendritic cell-associated C-type lectin-2) signal via the FcRγ gamma chain (Miyake et al., 2015; Ostrop et al., 2015; Sato et al., 2006), leading to CARD9-BCL-10-Malt1 activation (Figure 1A). This in turn results in the activation of NF-κB, eventually triggering cytokine release. Importantly, these two receptors share the same signal transduction pathway, while having different functions (Thompson et al., 2021). Therefore, both dectin-2 and MINCLE can be compared of whether these related proteins differently transmit glycan information from the receptor level. In contrast, dectin-1 and dectin-2 have different signal transduction pathways but are both involved in the detection of β-glucans and mannan, respectively (Figure 1A).

Reporter cell system for the observation of glycan-lectin interactions.

(A) Schematic representation of dectin-1, dectin-2, and macrophage-inducible C-type lectin (MINCLE) signaling pathway with GFP under control of nuclear factor kappa-B (NF-κB; left) and the dose dependence of the GFP expression of the cells (right). The input and output distribution of single-cell resolution data are divided by indexed (i.e. i and j) rectangular grids to estimate the joint and marginal probability of the distribution. Ni,j and Nt are the number of cells in the index (i,j) and the total number of cells, respectively. (B) Estimated joint probability distribution between the input glycan concentration and GFP expression using the procedures described in (A). The joint entropy (Hxy) and marginal entropies (Hx and Hy) and therefore mutual information (Ixy) can be calculated from the estimated distribution. The channel capacity of the distribution can be further found by maximizing the mutual information with various trial input distributions (see also Appendix 2—figure 3). Experimental data for FurFurMan stimulation is shown. (C–D) Monoclonal reporter cells expressing dectin-1, dectin-2, MINCLE, or wild type (WT) were stimulated with (C) invertase or (D) FurFurMan (n=3). (E) Dose response of the dectin-2 reporter cells is shown both as geometric mean with SD and boxplot with the whiskers representing the 1 percentile of the cellular population (n=6).

Upon fungal infection, combination of these and other cell surface receptors expressed by antigen presenting cells then leads to a defined immune reaction via signal integration processes (Snarr et al., 2017). Such signal integration can result in synergism between the receptors triggering an effect greater than their individual contributions (Ostrop and Lang, 2017). For example, MCL (macrophage C-type lectin), another CTL present on cells of the innate immune system, is known to synergistically work with dectin-2 (Ostrop et al., 2015; Zhu et al., 2013). Additionally, to this type of synergism, other members of the CTL family, e.g., DC-specific ICAM-3–grabbing nonintegrin (DC-SIGN) and Langerin, rather modulate a response instead of initiating it by themselves (Geijtenbeek and Gringhuis, 2016; Osorio and Reis e Sousa, 2011). Therefore, it is important to quantitatively account for the resulting signaling to describe the complexity of how these cell surface receptors can modulate each other to translate a glycan-encoded information into a biological response.

Accounting for the stochastic behavior of cellular signaling, information theory provides robust and quantitative tools to analyze complex communication channels. A fundamental metric of information theory is entropy, which determines the amount of disorder or uncertainty of variables. In this respect, cellular signaling pathways having high variability of the initiating input signals (e.g. stimulants) and the corresponding highly variable output response (i.e. cellular signaling) can be characterized as a high entropy. Importantly, input and output can have mutual dependence, and therefore, knowing the input distribution can partly provide the information of output distribution. If noise is present in the communication channel, input and output have reduced mutual dependence. This mutual dependence between input and output is called mutual information. Mutual information is, therefore, a function of input distribution, and the upper bound of mutual information is called channel capacity (Appendix 2; Cover and Thomas, 2012).

In this report, a communication channel describes signal transduction pathway of CTL, which ultimately lead to NF-κB translocation and finally GFP expression in the reporter model (Figure 1A). To quantify the signaling information of the communication channels, we used channel capacity. Importantly, the channel capacity is not merely describing the resulting maximum intensity of the reporter cells. The channel capacity takes cellular variation and activation across a whole range of incoming stimulus of single-cell resolved data into account and quantifies all of that data into a single number.

Herein, we studied dectin-2, dectin-1, MINCLE, DC-SIGN, MCL, TNFαR (TNF alpha receptor), and TLR-1&2 in NF-κB reporter cells using single-cell resolved flow cytometry (Figure 1A, see also Appendix 2). To accurately quantify the information transmission in the receptors’ signaling pathways in response to exogenous glycans, we use the channel capacity as a metric (Figure 1B). By employing channel capacity measurements, we found dectin-2 channel has relatively low signaling capacity, which in turn is synergistically increased in the presence of co-expressed MCL receptor. Furthermore, the channel capacity of dectin-1 and dectin-2 channel for the same glycan ligand was compromised when both receptors are expressed by the cell while increasing the binding sensitivity (EC50) to the ligand. Overall, our findings and approach provide a quantitative description of glycan lectin communication and signal integration of CTLs and other receptors, which may lead to a better understanding of key phenomena such as pathogen recognition and autoimmunity.

Results

Quantifying signal transduction in glycan-based communication

We employed a single-cell resolved reporter system to monitor CTL activity by GFP expression under control of the transcription factor NF-κB in human monocytic U937 cells (Figure 1A). Dectin-2 was expressed in these reporter cells, and stimulation was conducted using various ligands (Figure 1C–E). FurFurMan, an extract of Malassezia furfur, as well as the polysaccharide mannan and invertase, both from Saccharomyces cerevisiae, initiated dectin-2 signaling. In contrast, owing to the lack of multivalency, mannose itself could not initiate signaling but was able to inhibit dectin-2 function (Figure 1C and D, and also Appendix 1—figure 1 A; Ishikawa et al., 2013). In parallel, the invertase treated with α-mannosidase does not activate the NF-κB signaling, indicating the glycosylation-dependent dectin-2 activity (Appendix 1—figure 1 B). The activation of human dectin-2 receptor is in line with previous reports on its murine homolog, which is triggered by Man-α1–2 Man moieties presented on scaffolds like proteins, glycans, or polystyrene beads (Ishikawa et al., 2013; Yonekawa et al., 2014; Zhou et al., 2018). Analogously, introduction of dectin-1 into the reporter cells enabled detection of NF-κB-based GFP expression after stimulation. However, while FurFurMan could also stimulate dectin-1 cells, this was not inhibited by the addition of mannose, which is expected for this β-glucan receptor (Figure 1D).

Next, we studied the dose-response behavior of dectin-2 reporter cells stimulated with FurFurMan over a wide range of input concentrations (Figure 1E). The cellular population revealed an overlap between the unstimulated and the maximally stimulated population, demonstrating the absence of a clear two-state behavior on a population level (Figure 1E, Appendix 1—figure 1 C). To ensure that change in the reporter level is not affected by protein expression rate, we confirmed that GFP expression required at least 16 hr of stimulation to reach its maximum in steady-state protein expression, while short stimulation with for 2–6 hr does not lead the maximum level of GFP production (Appendix 1—figure 1 D). We also ruled out any influence of the selection process for the cellular clones, by sorting dectin-2 expressing cells according to their GFP expression level. When re-stimulated, both populations again showed the same broad GFP expression, confirming the wide range of the response to be independent of genetic differences between individual cells (Appendix 1—figure 1 E). Taken together, observing noisy dectin-2 signaling on a single-cell level in relevant model cell lines reveals a broad population distribution when stimulated.

Dectin-2 transmits less information than other receptors

To investigate whether other receptors with similar signaling pathways follow the same principle, we analyzed the dose response of dectin-1, MINCLE, and the non-CTLs TNFαR and TLR-1 and -2 (Bode et al., 2019; Holbrook et al., 2019; Ishikawa et al., 2009; Ozinsky et al., 2000). To quantify the underlying signal transmission in a cellular population, the channel capacity was used as a metric. Note that we choose the stimulation time, the period of incubation time of the cell with the input ligands, as the time point when GFP response and channel capacity reach the maximum and steady-state value (Appendix 1—figure 2 A and B). And therefore, the stimulation was 13 and 16 hr for TNF-α and the rest of the ligands, respectively. Previous work on TNF-α signaling found the TNF-α channel to have a channel capacity of about 1 bit in particular 1.64 bits when a reporter cell system was used (Cheong et al., 2011). In addition, this channel capacity can be further increased if one can measure the temporal evolution of output dynamics instead of static output dataset (Selimkhanov et al., 2014). Such channel capacity suggests that a cellular population can use a receptor to distinguish between two states: on/off or presence/absence of a stimulant. For U937 cells, we found the TNFαR transmits 1.34 bits of channel capacity for TNF-α stimulant (Figure 2A and B), which was not influenced by the introduction of additional lectins (i.e. MINCLE, dectin-2, and DC-SIGN, see Appendix 1—figure 2 C). In the case of dectin-1 expressing U937 cells, the channel capacities were 1.20 and 1.09 bits for depleted zymosan (DZ) and FurFurMan input, respectively, while both MINCLE and TLR1&2 had a channel capacity of 0.98 and 0.99 bits, respectively. Since these receptors signal via NF-κB, these differences can be explained by receptor expression levels and downstream pathways. In contrast, dectin-2 stimulation resulted in a channel capacity of 0.70 bits using FurFurMan as a ligand. Stimulation using heat inactivated invertase or mannan had 0.80 and 0.49 bits, respectively (Figure 2B). Also, in THP-1 cells, a similar trend of lower GFP expression upon stimulation is observed, further supporting the notion that dectin-2 has a lower signal transmission capacity compared to the other receptors such as TNFαR (Appendix 1—figure 2 D-F).

Quantification of signal transduction.

(A) Representative histograms of U937 reporter cells dose response, stimulated specifically with invertase for dectin-2, TNFα for the TNFαR, trehalose-6,6-dibehenate (TDB) for macrophage-inducible C-type lectin (MINCLE), Pam3-Cys-Ser-Lys4 (Pam3CSK4) for TLR1&2, and depleted zymosan (DZ) for dectin-1. The number in each histogram is the channel capacity of the corresponding signals. Top right panel shows a schematic representation of the five analyzed receptor channels. (B) Estimated channel capacities between various pairs of ligand and receptor (*p<0.05, Wilcoxon rank-sum test). (C) Monoclonal reporter cells either expressing dectin-2 (n=3), FcRγ (n=2), or dectin-2 and FcRγ (n=4) were stimulated for 16 hr with various concentrations of invertase, labeled with Atto647 dye. The error bars indicate the SDs. (D) Channel capacities from stimulation with invertase and FurFurMan (data also seen in B) and FurFurMan stimulation of U937 reporter cells (**p<0.01, Wilcoxon rank-sum test). (E) Unstimulated reporter cells (mock stimulated) 16 hr after cultivation with 25 mM galactose, or mannose, or under normal conditions for 48 hr (n=3, *p<0.05 and ***p<0.001, Student’s t-test), suggesting that dectin-2 mediated self-recognition leads to a high basal level of cellular activation in FcRγ overexpression cells.

The most striking difference was found between MINCLE and dectin-2, as both lectins use the same signaling pathway via FcRγ (Ishikawa et al., 2013), suggesting that the substantial differences between the channel capacities rely on very early ligand recognition events. We hypothesized overexpression of the signaling protein FcRγ might increase the information transmitted via dectin-2. The overexpression of FcRγ resulted in at least twofold increase of NF-κB controlled GFP expression (Figure 2C). Overexpression of both dectin-2 and FcRγ yielded a high-basal NF-κB activation of the cells while the sensitivity for its ligand (EC50) increased about 50-fold. While the maximal GFP signal of dectin-2 (MFI, mean fluorescence intensity) was increased in the presence of FcRγ overexpression, the channel capacity however decreased simultaneously (0.41 bits; Figure 2 C–E). Since competition with mannose reduced this effect, we speculate that decreased channel capacity might originate from self-recognition of dectin-2 of ligands being present either on the same cell or those in close proximity during the culture conditions (Figure 2E). From this, we concluded the channel capacity of a glycan-based communication channel is not necessarily coupled to its sensitivity. Also, the ability of a communication channel to transmit information is not well described by its maximal signal alone (i.e. MFI), but rather by the channel capacity. Next, we quantified the number of receptors and excluded that the difference in MINCLE and dectin-2 channel capacities is due to differences in receptor expression levels (Appendix 1—figure 2 G). Taken together, dectin-2 has relatively less channel capacity, and while its sensitivity (EC50) can be modulated with FcRγ, the transmitted information does not increase. Additionally, the number of receptors has little influence on the channel capacity or amplitude.

Signal integration compromises between dectin-1 and dectin-2 receptors when both are engaged

To expand our insight from isolated cell surface receptors to the interplay between multiple lectins, we prepared reporter cells expressing dectin-2 and dectin-1 simultaneously. FurFurMan served as a stimulant since it interacts with both dectin-1 and dectin-2. First of all, we found that the level of receptor expression did not change upon expression of an additional lectin (Figure 3A). Dectin-1 expressing cells gave a higher maximal signal (i.e. maximal MFI) and channel capacity than dectin-2 expressing cells; however, the latter channel showed higher sensitivity (EC50) to FurFurMan. We found that the double positive cells did compromise between the two receptors, displaying the values corresponding to the intermediate values of the EC50 and channel capacity of dectin-1 and dectin-2 (Figure 3B and C). Additionally, mannose could be used to interfere with dectin-2 signaling, thus U937 dectin-1 dectin-2 expressing cells showed the same dose-response curve as dectin-1 expressing cells (Figure 3D). When DZ, a dectin-1 specific stimulants, was used, dectin-2 expression did not significantly influence the response of the double positive cells. Hence, dectin-2 specific signaling was not influenced by dectin-1 expression (Appendix 1—figure 3 A-C). Moreover, inhibition of dectin-2 signaling initiated by FurFurMan by the addition of 25 mM mannose resulted a response that was not a compromise. Taken together, we see that the simultaneous stimulation of dectin-1 and dectin-2 resulted in a compromise between their channels, which demonstrates how these two channels integrate glycan signal into response.

Signal integration of dectin-1 and dectin-2.

(A) Quantitation of surface expression of U937 dectin-1, dectin-2, and dectin-1 dectin-2 U937 reporter cells. Fluorescence intensity (FI) values were transformed into the number of proteins expressed using a PE-quantitation. Graph shows geometric mean ± robust SD of the cellular population. (B) Monoclonal reporter cells either expressing macrophage-inducible C-type lectin (MINCLE), dectin-2, dectin-1, or both dectin-2 and dectin-1 were stimulated for 16 hr with various concentrations of FurFurMan (n≥3). The error bars indicate the SDs. (C) Channel capacities of U937 reporter cells expressing either dectin-1, dectin-2, or both stimulated with FurFurMan (*p<0.05 and **p<0.01, Wilcoxon rank-sum test). (D) Monoclonal reporter cells either expressing dectin-2, dectin-1, or both dectin-2 and dectin-1, stimulated with various concentrations of FurFurMan (n=3). Dectin-1 expressing cells were stimulated either with or without 25 mM of mannose. The error bars indicate the SDs.

Macrophage C-type lectin (MCL) increases the channel capacity of dectin-2

To further expand our insights into signal transmission through multiple lectins, we wondered whether co-expression of other lectins would synergistically increase the channel capacity of dectin-2 signaling. For this, we included DC-SIGN and MCL (Figure 4A). Although DC-SIGN does not elicit NF-κB signaling by itself in U937 cells, it is known to recognize high-mannose structures present on invertase (Gringhuis et al., 2009). As expected, U937 dectin-2 DC-SIGN cells experience significantly increased ligand binding (Appendix 1—figure 4 A and B). We then speculated that this would either (a) promote the ligand recognition by pre-concentration of the stimulants on the cell surface or (b) sequester the input signal from dectin-2, reducing the cellular response. In fact, DC-SIGN-mediated ligand binding did not alter the dectin-2 channel capacity for FurFurMan or invertase stimulation or did DC-SIGN expression itself modulate TLR4 signaling (Figure 4B, Appendix 1—figure 4 C). However, the sensitivity, as assessed by EC50, increased for dectin-2 DC-SIGN expressing cells (Figure 4A). The increased sensitivity due to DC-SIGN co-expression might increase the channel capacity if the allowed dose range spans low-concentration region. Therefore, we calculated the channel capacity by increasing the maximum input concentration. However, this was not the case (Figure 4C). Contrary to DC-SIGN, overexpression of MCL significantly increased the channel capacity of dectin-2 expressing cells, particularly when limiting our dataset to low-maximum invertase concentrations (Figure 4B and D, Appendix 1—figure 4 D and E). This indicates that MCL enhances the fidelity of invertase information transmission of dectin-2 channel, providing quantitative measurement of synergistic effect of MCL (Ostrop et al., 2015; Zhu et al., 2013).

Signal response of dectin-2 in the presence DC-specific ICAM-3–grabbing nonintegrin (DC-SIGN) or macrophage C-type lectin (MCL).

(A) Representative dose response of invertase stimulation of U937 cells expressing dectin-2, MCL, or dectin-2 co-expressed with either DC-SIGN or MCL. (B) Channel capacities of dectin-2 in combination with DC-SIGN and MCL after stimulation with either FurFurMan or invertase (*p<0.05, Wilcoxon rank-sum test). (C and D) Channel capacities calculated from different maximum invertase concentrations of dectin-2 expressing cells compared with either DC-SIGN (C) or MCL (D) co-expression. The shaded regions represent the 95% CI of the channel capacity. The right side of the dashed line in (D) is the region that shows statistical significance between dectin-2 and MCL co-expressed dectin-2 (*p<0.05, Wilcoxon rank-sum test).

We then wondered whether the difference in channel capacity between dectin-2 and TNFαR could simply be a result of affinity. Since TNFαR has a nanomolar affinity for its ligand (Grell et al., 1998), we applied an anti-dectin-2 antibody to stimulate dectin-2 cells. Even under these conditions, we did not monitor an increase in channel capacity (Appendix 1—figure 4 F). Therefore, we found that MCL but not DC-SIGN significantly increase the dectin-2 channel capacity, while both MCL and DC-SIGN enhance cellular binding of the stimulants and the resulting cellular sensitivity to invertase.

Dectin-2 channel has a low signal-to-noise ratio

The relatively low channel capacity of dectin-2 could be a result of its limited maximum GFP expression even at high-stimulant concentrations compared to the other channels (Figure 2B). For this, we define the signal power as the variation of the mean GFP expression under individual stimulant dose (Figure 5A). In addition, the level of background noise (i.e. noise power) of the channel can be defined as the average of the variance of GFP expression at a given stimulant dose. These definitions allow to decompose signal and noise power (Appendix 3) and analyze them separately to infer how those two parameters shape the channel capacity.

Decomposition of the signaling channels into signal power and noise power.

(A) Schematic description of signal and noise power. meani and vari are the average and variance of the output at ith dose, respectively. (B) Decomposed signal power (red) and noise power (blue) of the individual signaling channel. (C) The ratio between signal power and noise power (i.e. signal-to-noise ratio) of individual channel given in (A) (*p<0.05, Wilcoxon rank-sum test).

TNFaR, MINCLE, and dectin-1 have a similar level of noise power. Amongst the three receptors, TNFaR shows the highest signal power and consequently the highest signal-to-noise ratio (Figure 5B and C). All three channels have a signal-to-noise ratio higher than one. For dectin-2, both signal and noise power are low compared to the other receptors; however, the noise power exceeds the signal power, resulting in a significantly lower signal-to-noise ratio. Since the signal power is independent of the noise power, our data indicate that the lower variation of the mean GFP expression (i.e. signal power) of dectin-2 dictates the reduced channel capacity compared to the other receptors. A similar conclusion cannot be drawn for the noise power since it is inherently coupled to the signal power (see Appendix 3).

We further employed the decomposition method to dectin-2 signaling in the presence of either dectin-1 co-expression or FcRγ overexpression (Appendix 1—figure 5). Analogous to the compromised channel capacity when dectin-1 and dectin-2 were co-expressed (Figure 3C), the analysis revealed that both the signal and the noise power were compromised as well (Appendix 1—figure 5 A and B). In case of FcRγ overexpression, dectin-2 signaling after invertase stimulation is characterized by increased noise power, resulting in decreased signal-to-noise ratio (Appendix 1—figure 5 C-E). Therefore, despite the high-GFP expression at high-stimulant concentrations (Figure 2C), the overexpression of FcRγ, as additional signaling hubs, involved in the dectin-2/NF-κB cascade did not increase the signal power but instead elevated the noise power, leading to reduced channel capacity. Taken together, the relatively low-channel capacity of dectin-2 is directly related to its low-signal power, and the overexpression of FcRγ further decreases the channel capacity through increasing the noise power.

Discussion

We set out to better understand how glycan-encoded information is read in cellular communication. We established a glycan-responsive in vitro model and exploited the channel capacity as a quantitative metric. For the receptors other than dectin-2, the channel capacities were around 1 bit or higher, similar values that have been reported for other systems previously (Suderman et al., 2017). In particular, TNF-α receptor has a channel capacity of 1.64±0.36 bits, which was found in a comparable reporter cell system (Cheong et al., 2011). Interestingly, the number of receptors expressed on the cell surface did not determine the channel capacity of a signaling channel (Appendix 1—figure 2 G). Our results exemplify that lectin signaling pathways and especially the dectin-2 pathway should not be viewed as a deterministic on/off-switch, but rather as difference in the probability of cells to be active at a certain dose. This is in line with previous reports strengthening a quantitative view of cellular signaling and taking the cellular microheterogeneity into account (Levchenko and Nemenman, 2014; Zhang et al., 2017). We found that the mannose binding CTL dectin-2 to transmits less information compared to other receptors of the same family (Figure 2B).

To understand how these insights could be expanded on the interplay between multiple receptors like the CTLs occur on innate immune cells rather than isolated lectins, we employed combinations of CTLs on our model cells. Dectin-2 and dectin-1 recognize different epitopes on FurFurMan, and we found that the effects were not additive, but a compromise between the two receptors, showing intermediate sensitivity (EC50) and channel capacity between dectin-2 and dectin-1 (Figure 3A and B). This effect implies at high concentrations of FurFurMan the dectin-2 channel is actively inhibiting dectin-1 signaling, resulting in a lower cellular NF-κB activation. It is well known that lectins are able to modulate the signals of other receptors (Geijtenbeek and Gringhuis, 2009; Gringhuis et al., 2009; Miyake et al., 2015). Yet this compromise is an exciting discovery since to the best of our knowledge previous studies have not quantified lectin signal integration. Hence, it is likely that during a fungal infection, exposing multiple epitopes of pathogens are recognized by the precise arsenal of immune receptors, and their underlying signaling pathways are integrating the information contained within the epitopes. This in turn leads to a compromise of all activated receptors and results in a specifically tailored biochemical response of the given immune cell (Ostrop and Lang, 2017).

Dectin-2 itself we found to have relatively less channel capacity when compared to the closely related MINCLE that uses the same pathway with more signal power (Figure 2A and B). It is therefore likely the receptor itself determines very early on the information flow into the cell. This could be a result of MINCLE being stimulated with crystalline insoluble ligands which could result in larger signaling clusters at the cellular surface. Alternatively, dectin-2 signaling could be influenced by mannose structures that are present on the cellular surface by giving rise to background signaling and selection for reducing signaling power in an in vitro setting of high-cellular density. Additionally, since dectin-2 binds high-mannose structures of eukaryotic origin (McGreal et al., 2006), a too sensitive reaction might lead to permanent self-recognition of human Man9 structures for example and hence potential autoimmune reactions. This hypothesis is supported by the dectin-2-dependent high-basal activity of FcRγ overexpressing dectin-2 cells, which in turn is responsible for a lower channel capacity in dectin-2 FcRγ cells (Figure 2C–E). Hence, dectin-2 could have evolved to use the CARD9-BCL-10-Malt1 pathway to NF-κB less effective. Along the same lines, recent reports show that CTLs are in general becoming more important in autoimmunity, dectin-2 in particular is known to be responsible for the development of allergic reactions (Dambuza and Brown, 2015; Parsons et al., 2014).

We first thought a combination of multiple lectins might synergistically enhance signaling capacity of dectin-2. But while DC-SIGN greatly enhanced ligand binding to the cells, meaning the increased sensitivity (EC50), it did not significantly increase the channel capacity (Figure 4A–C, Appendix 1—figure 4A and B). In contrast to DC-SIGN, the closely related MCL to dectin-2 has a significant synergetic effect on dectin-2 channel capacity at particularly low-stimulant concentrations, potentially making double positive cells more discriminative, at earlier timepoints of infection compared to dectin-2 expressing cells, substantiating the importance of signal integration to understand an cellular innate immune response (Ostrop and Lang, 2017).

Finally, it is important to take into consideration that our conclusions came from model cell lines, which were used as a surrogate for cell-type-specific lectin expression patterns of primary immune cells. Human monocytes and dectin-2 positive U937 cells have comparable receptor densities and respond similar to stimulation with zymosan particles (Appendix 1—figure 6A and B). Importantly, since our channel capacity calculations are applicable regardless of the nature of signal and medium, one could use it to quantify cellular responses in similar assays in the future. Work is ongoing to address central questions of cellular communication based on glycan lectin interactions.

Materials and methods

All reagents were bought from Sigma Aldrich, if not stated otherwise.

Reporter cell generation and reporter cell assay

Request a detailed protocol

U937 cells were transduced with an NF-κB-GFP Cignal lentivirus (Qiagen) according to the manufacturer’s instructions to generate NF-κB reporter cells. 0.5 mL of 2e5 cells were mixed with the lentivirus at an MOI (multiplicity of infection) of 15 and spin transduced for 1.5 hr at 33°C and 900 g. After 48 hr of rest, cells were selected with puromycin (gibco) for three passages. Eight cultures from a single cell each were subsequently made and evaluated according to their GFP expression, clone #5 only monoclonal cells were used for all experiments of this paper.

Reporter cell assay

Request a detailed protocol

U937 reporter cells were used in its log phase, and 100 µL were plated in a 96-well plate with 3e4 cells per well. Cells were challenged in complete media (RPMI with 10% FBS (fetal bovine serum), 1% Glutamax, 1% Pen/Strep, and all by gibco) with TNF-α and various other ligands and at various concentrations for 13 hr and 16 hr, respectively. After incubation, cells were re-suspended once in DPBS (Dulbecco's phosphate-buffered saline) and the expressed GFPs fluorescent intensity was measured by flow cytometry (Attune Nxt, Thermo Fisher).

Cell culturing and passage

Request a detailed protocol

U937 cells were kept between 1e5 and 1.5e6 cells/mL in complete media with passage 2–4 times a week. 293 F cells were adherently cultured in DMEM (Dulbecco's modified Eagle’s medium) with 10% FBS, 1% Glutamax, 1% Pen/Strep (Gibco), and split 2–3 times per week. All cells were tested for mycoplasma contamination using Minerva biolabs VenorGeM Classic.

Antibody staining and quantitation

Request a detailed protocol

For the surface staining, cells were incubated in with the respective antibodies and isotype controls for 30 min at 4°C in DPBS, then washed once in DPBS +0.5% BSA and measured via flow cytometry. For perforated stains cells were first fixed in 4% PFA (Carl Roth) at 4°C for 20 min, then perforated in perforation solution (DPBS +0.5% BSA+0.1% Saponine) for 20 min at 4°C. The cells were then re-suspended in perforation solution containing the respective antibodies, incubated for 20 min at 4°C and measured via flow cytometry after being washed once. To quantify the fluorescent intensities, we used the BD PE quantitation kit, which allowed us to calibrate FI to the number of PE molecules present in a sample. A list of all used antibodies can be found in the Supplementary file 1.

Generation of lectin overexpressing cells

Request a detailed protocol

cDNA of MINCLE, dectin-2, MCL, FcRγ, dectin-1, and DC-SIGN were cloned into vector BIC-PGK-Zeo-T2a-mAmetrine:EF1a as previously reported (Wamhoff et al., 2019). This bicistronic vector expresses mAmetrine under the PGK promoter. To combine multiple GOI (gene of interest), we also used the lentiviral vector EF1a-Hygro/Neo a gift from Tobias Meyer (Addgene plasmid # 85134). Briefly, 293F cells were transfected with vectors coding for the lentivirus and GOI. Lentivirions were generated for 72 hr, and the supernatant was frozen to kill any remaining 293 F cells. This supernatant was used to transduce the GOI into U937 cells via spin infection at 900 g and 33°C in the presence of 0.8 µg/mL polybrene (van de Weijer et al., 2014). After 48 hr of rest, the U937 cells were selected with appropriate antibiotics (Zeocin 200 µg/mL, G418 500 µg/mL, or Hygromycin B 200 µg/mL; Thermo Fisher, Carl Roth, Thermo Fisher, respectively). A list of used primers can be found in the Supplementary file 2.

Labeling of proteins

Request a detailed protocol

Invertase (5 mg in 1 mL) was heat inactivated for 40 min at 80°C and mixed with 3×molar excess of Atto647N-NHS dye (AttoTech) according to the manufacturer’s protocol. The labeled protein was purified using Sephadex G-25 column, and aliquots were frozen at –80°C. Since we found the labeled invertase to contain less impurities, we used Atto647 labeled invertase for all experiments shown in this study. Human TNF-α (Peprotech) was labeled with the same procedure, yet without heat inactivation. The degree of labeling was determined to be around 1 as determined with a labeled protein concentration measurement of a NanoPhotometer NP80 (Implen).

Channel capacity calculation

Request a detailed protocol

Calculations of channel capacity were based on Cheong et al., 2011 and Suderman et al., 2017. See Appendix 2 on channel capacity calculation for details.

Data representation, software, and statistical analysis

Request a detailed protocol

Data is shown as mean ± SD. Statistical analysis of data was performed by unpaired two-tailed t-test, with significant different defined as (p<0.05). EC50 values were calculated in graph pad prism version 8.4.2 using four parametric dose vs. response function. When necessary statistical differences between EC50 values were compared using an extra-sum-of-squares F test. Detail of statistical tests and EC50 determinations can be found in the SI raw data file. FlowJo v.10 was used for analysis and export of flow cytometry data.

Data availability

Request a detailed protocol

All data is available at Dryad. The Jupyter notebook including the channel capacity calculation and noise analysis is available at: https://github.com/imaginationdykim/2022.CC, (copy archived at Fuchsberger, 2023).

Appendix 1

Appendix 1—figure 1
Signal transduction in the reporter cells.

(A) Monoclonal U937 reporter cells expressing macrophage-inducible C-type lectin (mincle), dectin-2, or wild type were stimulated with mannan (n=4). Mannose alone could not stimulate dectin-2 but could inhibit stimulation by Mannan. (B) Dectin-2 downstream signaling is mediated by glycosylation of invertase. (Left) Interaction of Atto647 labeled invertase with dectin-2 or wild type U937 cells. While activation of the downstream signaling is dectin-2 specific (lower panel), the binding of invertase to U937 does not depend only on the interaction with dectin-2 (upper panel). Right: The invertase and U937 cells were treated with α-mannosidase. Such treatment resulted in complete inhibition of nuclear factor kappa-B (NF-κB) activation in dectin-2 cells (lower panel) but did not affect protein binding (upper panel). (C) Histograms of the dose response in Figure 1D and U937 dectin-2 expressing reporter cells react to various concentrations of FurFurMan. Darker histograms were stimulated with higher ligand concentration. (D) Dose response of dectin-2 and DC-specific ICAM-3–grabbing nonintegrin (DC-SIGN) expressing reporter cells stimulated for 16 hr, or stimulated for 2 and 4 hr, washed in fresh media, and incubated to a total of 16 hr. (E) U937 dectin-2 reporter cells were sorted in a GFP high and low population after stimulation for 16 hr with 300 µg/mL Mannan. The sorted cells were the re-stimulated 2 weeks later with 500 µg/mL.

Appendix 1—figure 2
Stimulation time and stimulant dependent signal transduction in the model cells expressing the receptor of interest.

(A) Violin plot of nuclear factor kappa-B (NF-κB) controlled GFP expression of U937 cells in response to 50 ng/mL TNF-α stimulant for various stimulation time. (B) Channel capacity of TNFαR channel to TNF-α stimulant for various stimulation time. Error bars indicate 95% CI. (C) Reporter cell lines expressing various lectins stimulated with TNF-α. (D) THP-1 reporter cells expressing dectin-2 or wild type (WT) were stimulated for 48 hr with FurFurMan (FFM), unstimulated (PBS), or the FurFurMan stimulation was inhibited with 25 mM mannose. Graph shown representative histograms. (E) Geometric means of the experiment done in A in triplicates (n=3) with the error bar representing SD. (F) Representative histograms showing the TNF-α stimulation (13 hr) of U937 and THP-1 reporter cells. THP-1 cells stimulated for 48 hr with TNF-α gave less signal than at 13 hr (data not shown). (G) Quantitation of surface and overall expression of receptors used in this study in U937 reporter cells. Cells were stained either for their surface expression or their overall protein expression with PE coupled antibodies. FI values were transformed into the number of proteins expressed using a PE-quantitation kit. Graph shows geometric mean ± robust SD of the cellular population.

Appendix 1—figure 3
Signal integration of dectin-1 and dectin-2 in the presence of invertase or depleted zymosan.

Monoclonal reporter cells either expressing dectin-2, dectin-1, or both dectin-2 and dectin-1 (n≥3) were stimulated for 16 hr with various concentrations of (A) InvertaseA647, or (B) depleted zymosan respectively. (C) Channel capacities of the data in (A) and (B).

Appendix 1—figure 4
Signal integration between dectin-2 and either MCL or DC-SIGN in the presence of an invertase stimulus.

U937 reporter cells expressing lectins as indicated, representative 2D plots (A) and labeled input binding (B) seen in the main Figure 3B.(C) U937 reporter cells either wild type or DC-specific ICAM-3–grabbing nonintegrin (DC-SIGN) expressing were stimulated with 5 µg/mL LPS-EB (invivogen) and 50 µg/mL InvertaseA647 for 18 hr. (D) 2D dose response and channel capacity of dectin-2 U937 reporter cells stimulated with anti dectin-2 for 16 hr. Channel capacities calculated from different maximum invertase concentrations of dectin-2 expressing cells compared with either DC-SIGN (E) or MCL (F) co-expression (*p<0.05 and **p<0.01, Wilcoxon rank-sum test).

Appendix 1—figure 5
Decomposition of the signaling channels in various receptors and ligand conditions.

(A) Decomposed signal (red) and noise power (blue) of dectin-2, dectin-1, and dectin-1/dectin-2 channel in the presence of FurFurMan stimulant. (B) Signal-to-noise ratio of (A). (C and D) Decomposed signal and noise power of dectin-2 and dectin-2/overexpressed FcRγ channel in the presence of FurFurMan (C) and invertaseA647 (D). (E) Signal-to-noise ratio (C) and (D) (*p<0.05 and **p<0.01 Wilcoxon rank-sum test).

Appendix 1—figure 6
Quantification of the number of receptors and zymosan binding in primary cultured human monocytes.

(A) Comparison of dectin-1 and dectin-2 expression level between our model cell (U937) and primary cultured human monocyte. The number of dectin-2 expression on U937 cells and primary human monocytes are 477±133 and 144±24, respectively. (B) Labeled zymosan binding on dectin-2 expressing U937 and human monocyte.

Appendix 2

Estimation of channel capacity between input doses and reporter GFP

1.1 Data structure

Cells can sense the environment and respond to it. In this work, we quantify the cellular capability to sense carbohydrate information through information theory. The carbohydrate information is given as carbohydrate ligand concentrations in the cell media, and the output is the GFP expression level triggered by NF-κB translocation of individual cells (Appendix 2—figure 1). The inputs (i.e. concentration of ligand) are discrete values covering almost all variability of output distribution. We used 9 or 10 levels of carbohydrate ligand concentrations including the absence of ligand and measured around 100,000 cells for whole doses using flow cytometry. The measured GFP expression levels are integer values ranging from around 0 to 50,000.

Appendix 2—figure 1
Illustrations showing how input and output are measured.

1.2 Mutual information estimation

To estimate the mutual information between carbohydrate inputs and GFP expressions from NF-κB translocation, the input and output array data described in Appendix 2—figure 1 is projected into two-dimensional probability space divided by grids as shown in Appendix 2—figure 2A. The projection allows estimation of joint probability distribution of finite data points. On the other hand, finite sample size together with arbitrariness of binning interval produces over- or under-estimation of joint probability distribution, requiring additional statistical analysis to calculate unbiased channel capacity. The joint probability distribution of individual grid elements is the number of data points in the grid divided by the total number of data points. The joint probability distribution and marginalized input and output distributions is shown in Appendix 2—figure 2 B. The probability distribution for input and output is the marginalized joint probability distribution by output and input index, respectively, as follows:

Px(i)=joutputPxy(i,j)

and

Py(j)=iinputPxy(i,j),

where i and j are the input and output index, respectively, and Pxy is the joint probability of input and output. The mutual information of the given input and output distribution is 0.53 bits, calculated from mutual information definition: MI(input;output)= H(input)+H(output) H(input, output). Note that H(input) , H(output), and H(input, output) are input entropy, input entropy, and joint entropy, respectively, defined as follows:

Hx=iinputPx(i)log2Px(i),
Hy=joutputPy(j)log2Py(j),

and

Hxy=i,jinput and outputPxy(i,j)log2Pxy(i,j).
Appendix 2—figure 2
Mutual information and channel capacity calculation from TNF-α stimulation and nuclear factor kappa-B (NF-κB) reporter.

(A) Grid projection of input and output distribution from dectin-2 communication channel. Inputs are the concentrations and the outputs are the GFP expression level. (B) Probability space describing joint and marginal probability distribution of the input and output distribution shown in (A). (C) Simple joint probability distribution consists of four inputs and two outputs signifying the channel capacity variation depending on the presence of noisy inputs. (D) Joint probability distribution maximizing the mutual information.

1.3 Channel capacity estimation

Suppose there is a joint probability distribution having four inputs and two outputs as described in Appendix 2—figure 2C above. Since the marginal probability distributions are equally distributed for input and output, the input entropy and output entropy yield 2 and 1 bits, respectively. The mutual information is therefore 0.5 bits by subtracting joint entropy from the sum of input and output entropy.

Consider that how much of information can be reliably transmitted from this channel? The input 0 and 3 certainly give output 0 and 1, respectively. On the other hand, the input 1 and 2 give uncertain outputs, distributed equally on all output range. Therefore, by completely suppressing the input channels 1 and 2, one can achieve the maximum information that this channel reliable transmits (i.e. channel capacity). The array multiplication (i.e. element-wise product) between weighting values w(i), [2, 0 , 0, 2], and input distribution Pxi , given in Appendix 2—figure 2C above, [1/4,1/4,1/4,1/4], produces the following weighted input probability distribution that maximize the mutual information:

Px(i)=w(i)Px(i)=[2, 0, 0, 2][Px(0),Px(1),Px(2),Px(3)]=[12,0, 0, 12].

And the changed input probability distribution, Px`i, must satisfy the law of total probability

iinputPx`i=iinputwiPxi=1

under the constraint condition

0Px(i)1.

The modified joint probability distribution yields 1 bit of channel capacity as shown in the below figure of Appendix 2—figure 2C.

Mathematically, the adjusted joint probability distribution by weighting values can be expressed as follows:

Pxy(i,j)=w(i)Pxy(i,j)

And therefore the adjusted input and output marginal probability distributions are

Px(i)=joutputsw(i)Pxy(i,j)=w(i)joutputsPxy(i,j)=w(i)Px(i)

and

Py(j)=iinputsw(i)Pxy(i,j).

Altogether, the mutual information given the weighting values is defined as

MIinput;outputwi=
iinputw(i)Px(i)log wxPx(i)joutputiinput w(i)Pxy(i, j)log iinput w(i)Pxy(i, j)+joutputiinputw(i)P(i, j)log w(i)P(i, j).

Finding input weighting values, wi, that maximize the mutual information subject to iinputwiPxi=1 and 0wiPx(i) is a non-linear optimization problem. Since the direction of the gradient of mutual information is the same as that of those two constraints at the minimum, using Lagrange multiplier method, one can restated the functions as Lagrangian Lwi,λ,σ=MIinput;outputwi-λiinputwiPxi-1-σwiPxi, and find out weightings using numeric approach (Kraft D. 1988. A Software Package for Sequential Quadratic Programming. Wiss. Berichtswesen d. DFVLR). We used sequential least squares programming provided by SciPy Python library (scipy.optimize.minimize, SciPy 1.7.3) to find out the optimizing input weighting values.

The input weighting values that maximize the mutual information given in Appendix 2—figure 2B is shown in Appendix 2—figure 2D. wi is 53,53,53,0,0,0,5,0; hence, Px`i becomes 16,16,16,0,0,0,12,0. Note that other weighing values, for example, 5,0,0,0,0,0,5,0 and 0,103,53,0,0,0,5,0, can be the solution due to the same output distribution for inputs 0, 1, and 2. The optimized input distribution yields around 0.71 bit of mutual information between input and output distribution.

1.4 Channel capacity calculation from modalized weighting values

Since we use optimization algorithms, we do not predefine the weighting values to find out the maximum mutual information. On the other hand, estimating mutual information under various Gaussian-shaped input weighting values can give intuition of physiologically relevant input distribution and cellular response (Cheong et al., 2011).

Appendix 2—figure 3 shows the estimation of channel capacity between TNF-α doses and NF-κB reporter under various unimodal and bimodal Gaussian input distributions. Note that several superposition of two different Gaussian distribution that forms a unimodal distribution, having single maximum peak, is excluded (Appendix 2—figure 3 B). Appendix 2—figure 3 C and D, show the calculated mutual information values from the unimodal and bimodal input distributions, respectively. Since the input range from 0 to 5 yields the same output response, the variation of input distribution within those range does not affect the mutual information. Therefore, discrete increases of mutual information are pronounced if the mutual information values are sorted in ascending order, particularly, in the case of bimodal input distribution (Appendix 2—figure 3D). Appendix 2—figure 3 E and F show the probability space given from the mutual information maximizing unimodal and bimodal input distribution, respectively. Maximum mutual information from bimodal input distribution yields around 10% higher value than that of the unimodal distribution and less than 1% of lower value compared to the optimized input distribution described in the previous section.

1.5 Influence of the number of output binning on channel capacity

Projection of input and output distribution onto probability space is described in Appendix 2—figure 2 A and B. Since the input and output data points are finite, relatively large number of output binning will produce discontinuous joint probability distribution throughout different output indexes. On the other hand, insufficient number of output binning cannot capture the original probability distribution from the input and output but average out the local variation of joint probability distribution throughout the output indexes. Therefore, the number of output binning significantly influence the mutual information and channel capacity of input and output distribution. Note that the number of input binning is the same as the number of input doses.

Appendix 2—figure 4 describes the changing mutual information and channel capacity values in different output binning numbers. Since the input binning is given as the input doses, there is no variation in input entropy in the mutual information calculation. On the other hand, the increase of binning increases the output entropy and joint entropy. As increasing the output binning number, the increased output entropy than that of the joint entropy is bigger (Appendix 2—figure 4 D). Therefore, mutual information and channel capacity increase as increasing output binning.

Appendix 2—figure 3
Mutual information calculation under unimodal and bimodal input distributions.

(A) Examples of unimodal input distributions. The parameter σ is the SD of the Gaussian function selected from 0.5, 1, 2, 4 and 8. There are 60 cases of input distributions. (B) Examples of bimodal input distribution containing the same σ parameters of the unimodal distributions. The number of bimodal combinations of the distribution is 1496. Vertically sorted various unimodal (C) and bimodal (D) input marginal probability distribution by the mutual information yields of the distribution. The probability space for the maximum mutual information given from unimodal (E) and bimodal (F) input distributions.

Appendix 2—figure 4
Mutual information and channel capacity variation depending on the number of output binning.

(A–C) Probability space of different output binning number of random distribution (upper) and the optimized probability distribution maximizing the mutual information of the given distribution (lower). The number of random data points in each input is 10, and thereby, the total is 100. (D) Probability space of the same input and output distribution in different output binning. (E) Output binning number dependence of mutual information and channel capacity values given total 10,000 random data points distributed equally on 10 input indexes. The lines represent the linear regression.

1.6 Influence of the number of samples on channel capacity

As described in the previous section, it is important to consider the ratio between the number of output binning and the number of samples to estimate the channel capacity. If the number of samples is relatively smaller than the number of binning, the joint probability space become sparse and generate one to one input and output relationship which in turn increases the calculated channel capacity.

Appendix 2—figure 5 describes changing mutual information and channel capacity with respect to the total number of samples. Since the samples are random distribution, the ground truth mutual information and channel capacity are 0. On the other hand, the distribution yields the more mutual information and channel capacity as decreasing the total number of samples (Appendix 2—figure 5A-C). The mutual information and channel capacity values deviating from 0 are a bias since the ground truth mutual information and channel capacity are 0.

Appendix 2—figure 5
Mutual information and channel capacity variation depending on the sample size.

(A–C) Probability space of different sample sized random distribution (upper) and the optimized probability distribution maximizing the mutual information of the given distribution (lower).

Appendix 2—figure 6
Explanation of bootstrapping procedure to estimate the channel capacity at infinite sample size.

(A) Original random distribution having 100 data points in each input and therefore 1000 data points for total. (B) Subsampled data from (A) with replacement in various subsampling percentage. (C) Probability spaces of (B) maximizing the mutual information (i.e. channel capacity). (D) Channel capacity values calculated in various subsampling percentages. Note that the x-axis is the inverse of subsampled sample number. Therefore, the range from 0.001 to 0.002 represents the 100–50% subsampling.

1.7 Bootstrapping method to estimate channel capacity at infinite sample size

As shown in the previous section, the size of sample determines the degree of bias in the calculated channel capacity. Furthermore, the size of sample is always finite, therefore the calculated channel capacity is biased. On the other hand, using linear regression of subsampled datasets, the channel capacity value at infinite sample size can be estimated as followed (Appendix 2—figure 6). The sample distribution given in Appendix 2—figure 6 A is subsampled into various subsampling percentages shown in Appendix 2—figure 6 B. The subsampling uses random sampling with replacement in every drawing, therefore the original sample distribution shown in Appendix 2—figure 6 and 100% subsampled distribution from the original data show difference in distribution points (Appendix 2—figure 6 B). These subsampled distributions tend to yield higher channel capacity in smaller subsample size (Appendix 2—figure 6 C). By plotting channel capacity value with respect to the inverse of sample sizes, the channel capacity value at infinite sample size can be estimated (Appendix 2—figure 6 D). This bootstrapping method alleviates the degree of bias in the calculated channel capacity, but still the estimated channel capacity of random distribution at infinite sample size is around 0.15 bits, 0.15 bits higher than the ground truth channel capacity (i.e. 0 bit).

Appendix 2—figure 7
Channel capacity estimation in various output binning numbers and total sample numbers.

(A) Extrapolated channel capacity values of random dataset at infinitely subsampled distribution under various total number of samples and output binning. The white line in the heatmap represents the channel capacity value at 0.01 bits. (B) Extrapolated channel capacity values at infinite subsample size of the input (TNF-α doses) and output (GFP reporter) of the dataset. (C–D) Line graphs showing the channel capacities of random and experimental dataset.

1.8 Channel capacity bias map depending on the output binning and total number of samples

As shown in the series of previous section, choosing an appropriate output binning and the total number of samples are essential to calculate the unbiased channel capacity. Since the two factors dependently influence the bias of channel capacity, it is required to estimate the channel capacity in various total sample numbers and binning numbers. Appendix 2—figure 7 above and B describe the bias map depending on the total number of samples and the number of binning calculated from random (A) and experimental dataset (B). Note that all channel capacities are estimated using bootstrapping method to interpolate the channel capacity value at infinite sample size.

In the case of random dataset, most of the regions spanning in range 0–160,000 total sample size and 0–1000 output binning number exhibit less than 0.01 bit of channel capacity. The white line in Appendix 2—figure 7 A indicate the contour line having 0.01 bit of estimated channel capacity. Therefore, the output binning number and total number of samples having the values above the white line exhibit less bias than 0.01 bits of channel capacity. On the other hand, the channel capacity values below the white line exhibit the value higher than 0.01 bits of channel capacity. In this work, the allowed bias is either 0.01 or 0.05 bits depending on the input and output layer (see section 123).

Appendix 2—figure 7 B is the bias map calculated from between TNF-α doses and reporter GFP of U937 cells. In this example, the total number of samples is 170,472 and subsampled without replacement in different y-axis of the heatmap (Appendix 2—figure 7). Overall, as shown in the widely spreading orange and red color, the calculated channel capacity fluctuates near 1.4 bits. The exceptions are either the case where the number of output binning is less than 200 or the coordination of total number of samples and binning numbers are below the white line shown in Appendix 2—figure 7 A. In this work, the minimum sample number in whole dataset is 63,816 which is the above of the line in Appendix 2—figure 7 A. Therefore, expected maximum bias in channel capacity is less than 0.01 bits even in the case of 1000 output binning.

We determine the channel capacity value as the highest channel capacity values calculated from output binning numbers ranging from 10 to 1000. Appendix 2—figure 7C, D shows channel capacity values depending on the output binning number for 25,570 and 51,141 total sample sized random distribution and Doses-GFP response data. In the case of 25,570 total sample size, around 600 output binning number, bias start to increase. On the other hand, in the case of 51,141 total sample size, there is no noticeable increase of bias in the given output binning range. The highest channel capacity values for 51,141 sample size is 1.41 bits at 985 output binning number. In the original sample size (i.e. 170,472), the maximum channel capacity is 1.41 bits at 510 output binning number. Therefore, in this example dataset, the estimated channel capacity is 1.41 bits.

In this work, if the input is discrete dose information, we estimate channel capacity using bootstrapping method described in Appendix 2—figure 6 with multiple output binning numbers ranging from 10 to 1000. The maximum value of channel capacity in the binning range is the final channel capacity value of the calculation. Appendix 2—figure 8 shows the results of channel capacity calculation from experimental datasets. The y-intercepts of individual lines are the unbiased channel capacity. And the maximum value of those y-intercepts in the individual dataset is selected as the final channel capacity value.

Appendix 2—figure 8
Channel capacity estimation of experimental data using bootstrapping in various y-binning number.

The y-intercept values of the regression line are the estimated channel capacity in the given y-binning number. The number of subsampled data points in each inverse sample side is 30.

Appendix 3

Decomposition of signaling channel

2.1 Definition of signal and noise power

The signal power of a signaling channel is the variance of the average output distribution of individual input responses. Therefore, signal power can be written as E[R-i-E(R-i)]2 hence ER-i2-ER-i2, where E and R-i are the expectation operator and average output distribution at ith input dose, respectively. Since R-i=jRjPRjSi, where Rj and PRjSi are marginal output value at jth index and conditional probability of output at jth index given the ith input signal Si, substituting R-i into ER-i2-ER-i2 provides the signal power as σr2=iP(Si)[jRjP(Rj|Si)]2[i,jRjP(Rj|Si)]2, where PSi is the input probability at ith index.

In the case of noise power, it is defined as the average of the variance of the output distribution of individual input responses. Therefore, noise power can be written as E[R¯i2R¯i2] and can be further expanded as σn2=iP(Si)[jRj2P(Rj|Si)[jRjP(Rj|Si)]2].

Appendix 3—figure 1A, B describes how different input and output distributions contribute to signal and noise power. Increasing the variance of output response of individual input does not influence the signal power but only increase the noise power. Likewise, increasing the mean output while keeping the variance of individual output response for each input provides increased signal power without affecting the noise power (Appendix 3—figure 1 B).

Appendix 3—figure 1
Signal and noise power calculated under the constraint of mean (A) and variance (B) of the output distribution.

Data availability

We have uploaded the raw data of the study to Dryad at https://doi.org/10.5061/dryad.18931zd2g. Our scripts for data evaluation are also linked to GitHub and stated in the manuscript.

The following data sets were generated
    1. Rademacher C
    (2023) Dryad Digital Repository
    Data from: Information transfer in mammalian glycan-based communication.
    https://doi.org/10.5061/dryad.18931zd2g

References

  1. Book
    1. Cover TM
    2. Thomas JA.
    (2012)
    Elements of Information Theory
    John Wiley & Sons.

Decision letter

  1. Andre Levchenko
    Reviewing Editor; Yale University, United States
  2. Aleksandra M Walczak
    Senior Editor; CNRS, France

Our editorial process produces two outputs: (i) public reviews designed to be posted alongside the preprint for the benefit of readers; (ii) feedback on the manuscript for the authors, including requests for revisions, shown below. We also include an acceptance summary that explains what the editors found interesting or important about the work.

Decision letter after peer review:

Thank you for submitting your article "Information transfer in mammalian glycan-based communication" for consideration by eLife. Your article has been reviewed by 2 peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Aleksandra Walczak as the Senior Editor. The reviewers have opted to remain anonymous.

The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.

Essential revisions:

1) Please, address the comments from Reviewer 1 regarding the physiological relevance of the cell model used in this study. In particular, it is imperative to relate the results to physiological levels of receptor expression and address the relevance of the signal processing in the cell type used in the experiments to cellular information transfer in vivo.

2) Both reviewers raise the key issue of how the information transfer was assessed vs previously published studies. Please, expand on the description of the information analysis, put it more explicitly into the context of the prior studies, consider supplying further details and justification for the specific capacity calculation algorithm used in the study.

3) It is important to expand on the discussion of the physiological relevance of having this complex receptor system, particularly in the context of different information transfer properties.

Reviewer #1 (Recommendations for the authors):

Overall, this is a fairly interesting piece of work, extending the framework of information theory to a new class of signaling networks that has not, at least to my knowledge, been considered before in this context. It is also quite interesting that there are significant differences in the information transmitted by dectin-2 and other, very closely related receptors; this suggests that even signaling systems with very similar downstream pathways can have very different noise properties. Understanding exactly why this happens mechanistically, and why these networks may have evolved this way, are very interesting future directions suggested by this work.

All of that being said, I have a number of important suggestions for the authors that I feel would need to be addressed before the manuscript could be considered for eLife. These are:

1) I think the physiological relevance of the results are difficult to evaluate as the manuscript is currently written. The fundamental issue is that the U937 cells considered here are not evolved to sense the glycans considered as "input" signals by the authors. Since this is an artificial cell line, the cells in question have not even evolved to sense anything; they were developed from cancer cells and have likely adapted to growing in the lab. So, while these are derived from a (potentially?) relevant immune cell population, they are clearly not representative of the actual human cells that have evolved to sense and respond to potential fungal infections. As far as I can tell, these cells don't even express the relevant receptors, since the authors have to introduce those receptors exogenously.

The authors may argue that the cells in question are at least immunological in origin, and that cell line work of this type is common in the field due to the difficulty of engineering the appropriate reporters into, say, a mouse in order to do work in primary cells. That may be true, but it still does not alleviate a very critical set of concerns regarding these results. In particular, it is currently difficult to interpret the findings the authors have about dectin-2 having a lower 'channel capacity' than other receptors like dectin-1. As the authors (basically) state, there are two hypotheses regarding this particular result. It does not seem that the difference in channel capacity between dectin-2 and dectin-1 is due to differences in average expression levels or due to differences in the distribution of protein levels themselves. Indeed, compared to TNF-α, there actually seems to be less noise in dectin-2 binding vs. TNF-α binding (more on that question below, see Figure 4A). So, the differences in channel capacity here are likely either due to fact that dectin-2 is somehow just more "noisy" as a receptor in and of itself, or the fact that there are differences in the downstream signaling network in between the receptor and NFκB.

This second hypothesis seems far more likely, given that it is highly unclear that intrinsic differences in, say, the conformational changes that drive signaling processes downstream signaling could be that significantly different between dectin-1 and dectin-2. In other words, it seems extremely unlikely that an intrinsic, biophysical difference between the receptors could be sufficient to explain the difference in channel capacity. As such, it is likely that the real difference here is due to some difference in downstream interactions. The authors seem to imply that any such differences are currently unknown, since these two receptors seem to interact with the same downstream pathways. Regardless, it seems much more likely to me that differences in the signaling networks induced by these two receptors (or dectin-2 and any of the receptors studied here) are driving differences in the observed information transfer.

It has been shown that the channel capacity depends in a critical way on the distribution of the number of downstream signaling molecules present (see the Suderman et al. (2018) Interface Focus 8 (6), 20180039). In other words, if you have a network that signals through a molecule with low abundance, that will tend to increase noise levels and decrease channel capacity. Or, if there is a molecule in the signaling pathway with a particularly broad distribution across the cell population, that can also lead to low channel capacities.

The problem in this case is that the cells in question have not evolved to use dectin-2 to sense anything in the environment. So, a particular protein downstream of dectin-2 may be expressed at a low level, or expressed with a broad distribution, simply because there is no need for these cells to sense their environment efficiently. As a result, it is very difficult to interpret the physiological relevance of the results presented here.

As currently written, the authors seem to make the argument that their findings have revealed a fundamental difference between dectin-2 and the other signaling molecules considered here. I am not myself convinced of that fact, simply because the physiological relevance of exogenously expressing these receptors in an immortal cancer cell line is unclear. I would suggest the authors think about what their results truly mean for our understanding of the underlying biology. Currently, I think it is impossible to infer that the authors would see the same kind of results in primary cells that actually perform the job of sensing these molecules in the body. That being said, the results are still interesting-dectin-1 and dectin-2 are very closely related, and to have such different noise properties is really rather intriguing. It suggests that receptors can have very different dose-response distributions even if they are very similar and are thought to signal through (essentially) identical downstream systems. The fact that this can happen is surprising, I think, and sets up the two alternative hypotheses discussed above. But the authors need to acknowledge this and be more reasonable regarding the extent to which these results can be extrapolated into a more meaningful biological context.

2) One of the most interesting aspects of this work is the fact that the authors consider not just the downstream signaling response, but also directly measure the statistical properties of the first step in the signaling process: namely, receptor binding. I think there is a significant opportunity to understand a critical aspect of cell signaling here in a new and interesting way, and so I have some suggestions for analyses the authors could do that would, I think, really increase the impact of the work.

I would be really interested to see what the channel capacity would be for the information flow between ligand concentration and the level of bound receptor. In other words, the authors can calculate the channel capacity between the dose of the ligand that they give the cells, and the distribution of bound receptor that they observe on the cell surface.

Looking at the data (e.g. Figure 4A), my intuitive sense is that the channel capacities will be relatively low for that calculation: maybe 2 bits or something like that. Interestingly, that is much, much lower than the theoretically possible value we would expect for receptors expressed at around 1000 copies per cell (which is around 3.2-4 bits, see the Suderman Interface Focus paper mentioned above). This may simply be because the receptors are being expressed exogenously and thus have lower absolute numbers, and broader distributions across the population, than would be observed in a more physiological context. That being said, this is to my knowledge the first actual measurement of the amount of bound receptor on the surface of individual cells. This is thus a great opportunity to calculate the information flow between ligand concentration and bound receptor levels.

I should note here that it might be intuitive for the authors to see this as an "upper bound" on the channel capacity for their system; in other words, the channel capacity between the ligand concentration and the bound receptor concentration should set an upper bound for the channel capacity for the entire system. The data processing inequality from Shannon seems, at least on the face of it, to guarantee that. It is important to note, however, that the signaling networks in question do not form a Markov chain; the amount of reporter expression can easily average over the entire signaling history in a particular cell. So, whatever the authors find here, it is not technically an upper bound on the downstream information flow. That being said, it would still be very interesting to know!

It should be relatively easy for the authors to just make this calculation based on the data they already have. I do, however, have suggestions for a few experiments that would improve the rigor of this calculation and make the results even more interesting.

Firstly, it is unclear how representative the bound protein numbers are of the actual number of binding events we would have in the cell culture when the cells are exposed to the ligand. This is because the authors have to wash the cells (at least I assume they do that) and then put them through the FACS instrument to measure the fluorescence. It would be really cool of the authors could do a time course to measure how fast the ligand "falls off" of the receptor. As long as the measurements made by the authors occur before significant unbinding has happened, the channel capacity they calculate should be very representative of what actually happened in the culture.

The other experiment that would be really interesting would be to repeat the experiment with labeled ligands but using whatever primary cells are known to express dectin-1, dectin-2, etc. In other words, the fluorescence here is coming from the ligand, so there is no need to engineer any reporters into the cell. If the authors see similar distributions of bound receptor on the surface of primary cells, and similar channel capacities between ligand concentration input and the bound receptor as output, that would be very intriguing. In any case, doing this experiment would I think begin to address some of the concerns with physiological relevance raised above. It would also be extremely interesting as a contribution to the field.

3) As mentioned in my public comments, another issue in this work is the fact that the authors are relatively reticent about how their calculations are actually performed. While the authors present the relevant equations for calculating the Mutual Information and claim to follow the approach of Cheong et al., the methods are not sufficiently detailed to understand exactly what they have done here. This is particularly important because their approach is clearly not identical to that used by Cheong, since their algorithm for maximizing the MI across input distributions is different. Cheong et al. approach this problem by trying a limited set of possible input distributions and using the one that gives the highest MI, while the authors here use a built-in optimization algorithm in python. It is not clear if this is the only point at which the authors deviate from the approach laid out by Cheong et al., or if they have modified other aspects of the calculation.

The authors need to explain how they approach the problem of correcting for finite size effects. Do they use the approach of Cheong et al., bootstrapping the data at different levels and then using a linear extrapolation of MI vs. 1/N (where N is the number of data points/cells) to extrapolate to the case of an infinite population size (1/N \to 0). Did they use the same approach to choosing the number of bins to calculate the probability distribution over outputs? Cheong et al. did this by visual inspection, if I recall correctly, looking for a "plateau" in bin numbers where the MI calculated from the data was non-zero but the MI calculated for randomized data was 0. Based on the statements made by the authors, and Figure S6, they seem to have performed both of these steps-otherwise it is completely unclear where the confidence intervals on their MI/channel capacity estimates come from, and how they chose the number of bins to use. Although it is important to note that the randomized controls seem to be missing from Figure S6. Regardless, since the authors have evidently re-implemented the code, they need to explain exactly what they have done.

I would suggest that the authors also adopt several improvements to the Cheong et al. approach that have been developed in subsequent works. In particular, Suderman et al. (2017, PNAS 114 (22), 5755-60) implemented a broader range of bootstrap samples across which to do the linear extrapolation to infinite population size, automated the choice of bin sizes, and developed a bootstrap approach to estimating confidence intervals that is much more realistic (and statistically appropriate) than simply using the confidence intervals from the linear extrapolation step. The authors can refer to the Suderman paper on how this was done; it should be relatively easy to implement these improvements, if they have not been done already.

I appreciate the author's innovation of using an optimizer in python to maximize the MI. I would expect this actually produces higher "channel capacities" than the previous approaches mentioned above. That being said, it is unclear how this optimization algorithm works. The authors may expect readers to go to the python documentation, but there are problems there. For one, the algorithm implanted in python might change, so the documentation may describe an approach that is different from the one the authors actually used to perform their calculations. How various packages are implemented in python can change from time to time, and something so generic as "optimize" may be changed as better optimization algorithms become available in the future. So the authors should describe exactly what they did. Also, readers are, in general, not going to go to python documentation to understand the methods of a paper like this, nor should they have to. The authors should describe what the algorithm is doing as currently implemented, in terms that readers can readily understand.

Finally, the authors should also perform the calculation using the exact same distributions used by Cheong et al. and Suderman et al. This will allow us to compare their results more directly to the results from those previous authors, as well as allow us to determine if the approach that they used results in similar channel capacities to the optimization algorithm used here. While I expect the optimization algorithm employed by the authors will yield larger numbers for the channel capacity (and thus better estimates, since the channel capacity is formally a supremum), it would be extremely useful if the authors checked this.

4) Another issue related to the calculation of the MI itself is the fact that the authors use logarithmically spaced bins, rather than linear bins, to perform the calculation. The authors describe this approach as if the matter of how to construct the bins is simply a matter of personal preference or convenience for the calculation, and choose logarithmic bins because they give somewhat larger values for the channel capacity. For one, this is not surprising; as is often the case, the FACS data obtained by the authors has a log-normal-like character; in other words, the authors see normal-ish distributions on a log scale (e.g. Figure 2A). As such, it is natural that using logarithmic bins will result in higher channel capacities, because it will tend to equalize (as much as possible) the effective number of observations within the bins.

While this may seem reasonable at first, it is actually, in my view, hard to justify. The reason is that the choice of how to generate the bins is not arbitrary, but actually corresponds to a very strong statement regarding what the authors actually consider the relevant "output" of the system to be. Using logarithmic bins is obviously equivalent to first taking the log of the data, and then using linear bins on that log-transformed data (Figure S2A should make this fact abundantly clear). The authors are thus envisioning that the channel is not a channel whose input is ligand concentration and output is GFP level, but rather a channel where the input is the ligand concentration and the output is the log of the GFP level. This is not an arbitrary decision, but rather one that has real consequences for how we construe the calculation in a physiological context. If the output is the log of the GFP level, that means that the cell is sensitive not to linear changes in protein concentration, but rather fold-changes.

For certain transcription factors, authors have argued that it is actually this kind of relative, fold change that matters (see Lee et al. (2014) Mol Cell 53 (6) 867-79). I would argue, however, that the conclusions of that work are hardly so robust as to suggest that, for every input-output transcriptional regulatory system within cells, it is the fold change of the "output" protein, rather than its linear change, that matters to the cell. Certainly, if the output is an enzyme, a transporter, or a protein that performs a whole host of other functions, then it is natural to assume that the appropriate output is the protein concentration itself, not the fold change in protein concentration.

If the authors can argue that every single gene whose transcription is controlled by NFκB has a fold-change impact on its downstream function, then I could see a strong argument being made for logarithmic bins. As it stands, however, I think the most natural way to interpret the output of the channel is on a linear scale, and I would say the authors should focus on that, rather than a logarithmic scale. If the authors want to include their logarithmic calculations in this work, I would suggest moving them to the supplement, and making clear the fact that such a calculation construes the output as having a fold-change impact on whatever is downstream of the protein level measured by the authors.

5) In this work, the authors focus on the measurement of GFP levels at a certain time. As written, it is not clear if the authors are considering the steady-state response of the cells that they are treating, a peak response, or something else. As far as I can tell, the authors stimulate the cells for 16 or 13 hours, depending on the ligand in question. I am not sure how these numbers were chosen-did the authors look at the GFP expression dynamics and choose this number based on those measurements? From an information-theoretic standpoint, it is best if the results are based on steady-state protein levels, but the author's don't seem to explain their rationale for choosing the time points that they choose. If the GFP expression levels do not reach a steady-state, it would be great if the authors had a particular reason for choosing the times they choose. I would also suggest that the authors consider looking to see if their channel capacity estimates are robust to the time they choose by looking at other time points and calculating the channel capacities to see if they get similar results. I am not sure what times to choose, but perhaps 24 hours or 36? It is hard to say without looking at the average dynamics over time, so providing that kind of data would be extremely helpful.

Also, the authors should acknowledge that a large body of literature has emerged that makes the claim that it is not the individual time points that matter, but rather the entire time series of the response that should be construed as the output. The authors can refer to the Zhang et al. paper in Cell Systems, the paper by Selimkhanov et al. (Science (2014) 346 (6215) 1370-3), papers by the Hoffman group (notably Tang et al. (2020) Nat Commun 12 (1) 1272), and a host of others. I am not suggesting that the authors adopt this worldview; there are actually serious mathematical and conceptual issues with construing the output of a communication channel to be a function like a time series. But, this idea is out there in the literature, and the authors should address this point and discuss how looking at GFP time series, rather than individual time point measurements, might yield different (and undoubtedly higher) channel capacity estimates.

5) As a final, rather important but also rather technical point, the authors in Figure 3C do statistical tests on their channel capacity estimates, comparing dectin-1 and dectin-2 cells to those that express dectin-1 and dectin-2 together. The issue here is that they use a t-test to estimate statistical significance, but it is really unclear where the dispersion (e.g. error bars or standard deviations) in the channel capacity estimates come from. I expect, since the authors day "n = 9" in the legend, that they split their data into 9 groups (maybe 9 sets of cells whose results were collected in different FACS runs or something?), performed the channel capacity estimate on each independently, then used the means and standard deviations in those estimates to do the t-test. That is completely made up by me, however; the authors really need to explain where these error estimates (and their +/- X values for channel capacity estimates in the various tables) actually come from.

In any case, the t-test of course only works if the data in question is drawn from actual Gaussian distributions. This does not mean that the distributions "look Gaussian" or "seem okay;" in the application of a t-test, you make extremely strong assumptions that only real honest-to-goodness Gaussian distributions actually satisfy. So the data really needs to actually be drawn from a Gaussian distribution for the results of the test to be interpreted appropriately.

My first suggestion here is that the authors need to use a non-parametric statistical test. That is, unless there is a true, theoretical reason to expect that the data will be drawn from a Gaussian (I know of no such reason, but perhaps the authors have one). The Wilcoxon rank-sum test was made for problems like this one, so I would suggest the authors use that.

A larger problem here, however, is that taking n = 9 (however it was done in the end) is not going to give a great estimate of the uncertainty in the channel capacity estimate. I would instead suggest that the authors merge their data into one data set (unless there are some kind of batch effects that prevent them from doing that) and then using a bootstrap/resampling approach, as in the Suderman et al. PNAS paper, to estimate confidence intervals and generate distributions for statistical tests. These bootstraps can be done hundreds or thousands of times, allowing for even more non-parametric tests (like permutation tests) to be used to estimate statistical significance.

This is honestly a fairly minor point, because nothing terribly important in the conclusions of the paper rest on the statistical tests here. But, if the authors want to argue that there are real differences between dectin-1, dectin-2 and dectin-1+dectin-2 cells, they should do these statistics in a more non-parametric way.

Reviewer #2 (Recommendations for the authors):

I think this paper would be greatly improved if the authors were more precise and detailed in the description of their capacity calculation. To address this, it's likely the authors were following the method from Cheong et al., but some of those details should be described here.

https://doi.org/10.7554/eLife.69415.sa1

Author response

Essential revisions:

1) Please, address the comments from Reviewer 1 regarding the physiological relevance of the cell model used in this study. In particular, it is imperative to relate the results to physiological levels of receptor expression and address the relevance of the signal processing in the cell type used in the experiments to cellular information transfer in vivo.

We thank the reviewer for their concern and updated the revised manuscript accordingly. We agree that for future work such a correlation between model cell lines and primary cells will benefit most of the biophysical and synthetic biology work focused on the interpretation of the signaling transmission. However, direct implementation of the NF-κB reporter into primary cells is very challenging. Hence, we decided to validate the choice of the model cell line by direct, side-by-side, comparison of dectin-1 and dectin-2 positive U937 reporter cells with the primary human monocytes and have mentioned and included these data in manuscript (line 382) and the Supporting Information (Appendix 1 Figure 6), respectively. These experiments revealed that human monocytes to the large extent overlap with the dectin-2 positive U937 cells in their expression level of dectin-2 receptor and efficiency to interact with zymosan particles.

Additionally, at this early stage of our understanding for the complexity of glycan-mediated information transmission, we see several advantages of our approach: (i) U937 cells originate from human monocyte and expresses immunological receptors such as mincle, CD200, CD200R, Siglec-1, and Siglec-3, therefore resemble physiological relevance (Byrareddy et al. PLOS ONE 10, e0140689 (2015)). Model cells must contain all necessary down streaming molecules connecting our receptors of interest to NF-κB. By this the model cell line we chose here will also be applicable for future studies on other glycan binding proteins associated with the immune system. (ii) We used monoclonal U937 cells, thus our results are independent from the heterogeneity of cells.

On the other hand, we agree that the U937 cells may differ in their expression levels of these downstream components of the signaling pathways and in turn, influence the activation of NF-κB, which may alter the channel capacity. Hence, in the revised manuscript, we do not compare the channel capacity between two different receptors that have different downstream pathways such as TNF-a and dectin-2.

2) Both reviewers raise the key issue of how the information transfer was assessed vs previously published studies. Please, expand on the description of the information analysis, put it more explicitly into the context of the prior studies, consider supplying further details and justification for the specific capacity calculation algorithm used in the study.

In the revised manuscript, we added a general introduction on channel capacity calculation method that used in this study and its comparisons to prior studies as described in Appendix 2 and 3.

3) It is important to expand on the discussion of the physiological relevance of having this complex receptor system, particularly in the context of different information transfer properties.

Thank you for raising this concern. We agree that the physiological relevance needs to be discussed further and we expanded the Discussion to explain why we choose U937 and its physiological relevance as follows:

Line 382: “Finally, it is important to take into consideration that our conclusions came from model cell lines, which were used as a surrogate for cell-type-specific lectin expression patterns of primary immune cells. Human monocytes and dectin-2 positive U937 cells have comparable receptor densities and respond similar to stimulation with zymosan particles (Appendix 1 Figure 6A and B).”

Reviewer #1 (Recommendations for the authors):

Overall, this is a fairly interesting piece of work, extending the framework of information theory to a new class of signaling networks that has not, at least to my knowledge, been considered before in this context. It is also quite interesting that there are significant differences in the information transmitted by dectin-2 and other, very closely related receptors; this suggests that even signaling systems with very similar downstream pathways can have very different noise properties. Understanding exactly why this happens mechanistically, and why these networks may have evolved this way, are very interesting future directions suggested by this work.

All of that being said, I have a number of important suggestions for the authors that I feel would need to be addressed before the manuscript could be considered for eLife. These are:

1) I think the physiological relevance of the results are difficult to evaluate as the manuscript is currently written. The fundamental issue is that the U937 cells considered here are not evolved to sense the glycans considered as "input" signals by the authors. Since this is an artificial cell line, the cells in question have not even evolved to sense anything; they were developed from cancer cells and have likely adapted to growing in the lab. So, while these are derived from a (potentially?) relevant immune cell population, they are clearly not representative of the actual human cells that have evolved to sense and respond to potential fungal infections. As far as I can tell, these cells don't even express the relevant receptors, since the authors have to introduce those receptors exogenously.

The authors may argue that the cells in question are at least immunological in origin, and that cell line work of this type is common in the field due to the difficulty of engineering the appropriate reporters into, say, a mouse in order to do work in primary cells. That may be true, but it still does not alleviate a very critical set of concerns regarding these results. In particular, it is currently difficult to interpret the findings the authors have about dectin-2 having a lower 'channel capacity' than other receptors like dectin-1. As the authors (basically) state, there are two hypotheses regarding this particular result. It does not seem that the difference in channel capacity between dectin-2 and dectin-1 is due to differences in average expression levels or due to differences in the distribution of protein levels themselves. Indeed, compared to TNF-α, there actually seems to be less noise in dectin-2 binding vs. TNF-α binding (more on that question below, see Figure 4A). So, the differences in channel capacity here are likely either due to fact that dectin-2 is somehow just more "noisy" as a receptor in and of itself, or the fact that there are differences in the downstream signaling network in between the receptor and NFκB.

This second hypothesis seems far more likely, given that it is highly unclear that intrinsic differences in, say, the conformational changes that drive signaling processes downstream signaling could be that significantly different between dectin-1 and dectin-2. In other words, it seems extremely unlikely that an intrinsic, biophysical difference between the receptors could be sufficient to explain the difference in channel capacity. As such, it is likely that the real difference here is due to some difference in downstream interactions. The authors seem to imply that any such differences are currently unknown, since these two receptors seem to interact with the same downstream pathways. Regardless, it seems much more likely to me that differences in the signaling networks induced by these two receptors (or dectin-2 and any of the receptors studied here) are driving differences in the observed information transfer.

It has been shown that the channel capacity depends in a critical way on the distribution of the number of downstream signaling molecules present (see the Suderman et al. (2018) Interface Focus 8 (6), 20180039). In other words, if you have a network that signals through a molecule with low abundance, that will tend to increase noise levels and decrease channel capacity. Or, if there is a molecule in the signaling pathway with a particularly broad distribution across the cell population, that can also lead to low channel capacities.

The problem in this case is that the cells in question have not evolved to use dectin-2 to sense anything in the environment. So, a particular protein downstream of dectin-2 may be expressed at a low level, or expressed with a broad distribution, simply because there is no need for these cells to sense their environment efficiently. As a result, it is very difficult to interpret the physiological relevance of the results presented here.

As currently written, the authors seem to make the argument that their findings have revealed a fundamental difference between dectin-2 and the other signaling molecules considered here. I am not myself convinced of that fact, simply because the physiological relevance of exogenously expressing these receptors in an immortal cancer cell line is unclear. I would suggest the authors think about what their results truly mean for our understanding of the underlying biology. Currently, I think it is impossible to infer that the authors would see the same kind of results in primary cells that actually perform the job of sensing these molecules in the body. That being said, the results are still interesting-dectin-1 and dectin-2 are very closely related, and to have such different noise properties is really rather intriguing. It suggests that receptors can have very different dose-response distributions even if they are very similar and are thought to signal through (essentially) identical downstream systems. The fact that this can happen is surprising, I think, and sets up the two alternative hypotheses discussed above. But the authors need to acknowledge this and be more reasonable regarding the extent to which these results can be extrapolated into a more meaningful biological context.

We agree on Reviewer #1’s comment regarding the difference of cell type can influence the amount of signaling molecules and thereby the channel capacity. Therefore, there is an arbitrariness in the calculated channel capacities. Indeed, we made model cell lines in THP1 and showed different NF-κB response but relatively weak compared to U937 cell. At this point we can only speculate why this is and assume that U937 cell has more downstream signaling molecule compared with the other cell lines. In the corrected manuscript, we do not compare the channel capacity between two different receptors that have different downstream pathway such as TNF-a and dectin-2.

The expression level of various downstream molecules of U937 cell might be different to the macrophages or dendritic primary cells. But U937 has all essential signaling molecules to translocate NF-κB with our receptors of interest. Therefore, we could investigate how two different carbohydrate receptors compromise the incoming carbohydrate information to the NF-κB response; This is hardly done by primary cells.

To address the concern, as was suggested, in the updated version of the manuscript we directly compared the receptors expression level of CD14+ human monocytes primary cells and our model synthetic cell lines (Appendix 1 Figure 6). Overall, it is challenging to introduce NF-κB reporter in the primary cells and one advantage of our system is that U937 reporter cell line was expanded from a single clone. Thus, our results are not affected by the cell-to-cell variability, providing more defined mechanistic conclusions. Since we could not introduce the reporter into the primary cells, we correlated the response of dectin-1 and dectin-2 positive U937 cells with the primary classical monocytes by comparing: i. binding of glycan input and ii. quantification of the receptors number. As a result, we observed that dectin-2 positive U937 cells and classical monocytes can interact comparably with the labeled glycan input. We also found that the receptor density on the transduced cell line was one order of magnitude higher in comparison to primary monocytes. We believe that such discrepancy may dependent on the primary cell type (unfortunately we could not run similar studies on multiple immune cell samples (1000€ each)), or variability in their isolation protocol. Thus, we consider the choice of our model system physiological and relevant for the behavior of classical human monocytes, at least based on the parameters we quantified in this direct side-by-side comparison.

We also would like to emphasize that our conclusions on the noisiness of dectin-2 receptor are coming from the direct comparison between dectin-2 and mincle receptors. Although dectin-1 and dectin-2 are receptors of the same lectin family, they have low structural homology and completely different mechanism of the downstream signal transmission and originate from different gen clusters (Saijo and Iwakura, 2011). One major difference for instance, is the presence of the ITAM-domain on dectin-1 and the absence of such on dectin-2. In addition, mincle cannot be phosphorylated on its own, similar to dectin-2, and relies on the downstream interaction with FcRy receptor. Thus, we also agreed with the review, that the different noise for signaling via these two receptors is an interesting observation which deserves further focus. This is ongoing research in the lab.

Currently, we assume that the noisiness of dectin-2 might be hidden in the sequence of its intracellular domains. In the follow-up work we plan to swap intracellular sequences of dectin-2 and mincle and check how the construction of such chimera receptors affects signaling efficiency. To summarize, in our work we concluded on the dectin-2 noisiness by comparing two structurally homologous receptors, expressed at the same density, in the same cell line (expanded from a single clone).

To further improve our manuscript and acknowledging that our results are derived in a synthetic model system. We try to make the reader aware of the limited physiological implication of our results. Still, we highlighted, an advantage of using synthetic approaches to dissect precise mechanisms of glycan-mediated signaling as follows:

Line 382: “Finally, it is important to take into consideration that our conclusions came from model cell lines, which were used as a surrogate for cell-type-specific lectin expression patterns of primary immune cells. Human monocytes and dectin-2 positive U937 cells have comparable receptor densities and respond similar to stimulation with zymosan particles (Appendix 1 Figure 6A and B).”

2) One of the most interesting aspects of this work is the fact that the authors consider not just the downstream signaling response, but also directly measure the statistical properties of the first step in the signaling process: namely, receptor binding. I think there is a significant opportunity to understand a critical aspect of cell signaling here in a new and interesting way, and so I have some suggestions for analyses the authors could do that would, I think, really increase the impact of the work.

I would be really interested to see what the channel capacity would be for the information flow between ligand concentration and the level of bound receptor. In other words, the authors can calculate the channel capacity between the dose of the ligand that they give the cells, and the distribution of bound receptor that they observe on the cell surface.

Looking at the data (e.g. Figure 4A), my intuitive sense is that the channel capacities will be relatively low for that calculation: maybe 2 bits or something like that. Interestingly, that is much, much lower than the theoretically possible value we would expect for receptors expressed at around 1000 copies per cell (which is around 3.2-4 bits, see the Suderman Interface Focus paper mentioned above). This may simply be because the receptors are being expressed exogenously and thus have lower absolute numbers, and broader distributions across the population, than would be observed in a more physiological context. That being said, this is to my knowledge the first actual measurement of the amount of bound receptor on the surface of individual cells. This is thus a great opportunity to calculate the information flow between ligand concentration and bound receptor levels.

I should note here that it might be intuitive for the authors to see this as an "upper bound" on the channel capacity for their system; in other words, the channel capacity between the ligand concentration and the bound receptor concentration should set an upper bound for the channel capacity for the entire system. The data processing inequality from Shannon seems, at least on the face of it, to guarantee that. It is important to note, however, that the signaling networks in question do not form a Markov chain; the amount of reporter expression can easily average over the entire signaling history in a particular cell. So, whatever the authors find here, it is not technically an upper bound on the downstream information flow. That being said, it would still be very interesting to know!

It should be relatively easy for the authors to just make this calculation based on the data they already have. I do, however, have suggestions for a few experiments that would improve the rigor of this calculation and make the results even more interesting.

We agree on Reviewer #1’s great suggestion on the additional analysis of information transmission in the presence of labelled input. We tried to include our new results regarding this (please see Author response image 1), but we could not fully explain the results due to the following reasons:

The measured labelled inputs signal of the cell at a single time point cannot reflect the total amount of interaction between the ligand and receptor. There are several processes that prevent us from directly connecting detected stimulants and signal initiation from such a snapshot. By the time we record the data, several ligands will be left receptor occupancy, whilst initiating signaling. Additionally, labelled stimulants will be degraded in the lysosome after the endocytosis. Therefore, the channel capacity between the ligand concentrations and bound ligands, in this experiment, cannot be the “upper bound” of the channel capacity between ligand concentrations and GFP expression. Similarly, we see that the channel capacity between bound ligand and GFP expression is lower than the channel capacity between ligand concentrations and GFP expression.

Author response image 1
Using labelled input adds measurable intermediate layer between ligand dose and GFP expression that can be directly related to ligand-receptor bindings.

(A) Illustration showing response of cell in different doses of labelled inputs. (B) Illustration showing information transmittance from labelled input doses to ligand-receptor binding and from ligand-receptor binding to NF-κB translocation (left). And schematics representing measurable information including receptor-bound labelled input of cells (right). (C) Response of dectin-1 expressing U937 cells on labelled zymosan in different doses.

We would like to share our experimental results regarding the channel capacity calculation in the presence of labelled input as follows:

Before we introduce our new results regarding this, we want to point out our previous results. As shown in the plot from Appendix 1 Figure 4A, WT cells exhibit non-specific binding to labelled invertase and there is no big difference in channel capacity between WT and dectin-2 calculated from ligand concentration and the level of bound ligands. Hence, it is difficult to delineate the non-specific binding of labeled invertase underlying the binding detected for dectin-2 expressing cells. However, only in presence of dectin-2, we detect downstream signaling and expression of GFP.

Due to the presence of non-specific binding of labeled invertase to dectin-2 positive cells, we had to ensure that the activation of NF-κB via dectin-2 is indeed specific to protein glycosylation (Appendix 1 Figure 1B). In the subsequent experiments, we remove dectin-2 epitopes from the stimulants using α-mannosidase and as a result, could not detect any activation of the GFP expression, although such non-mannosylated protein was still non-specifically recognized by the cells. These results confirm that the overall quantified response is glycan-specific and is not affected by non-specific binding of invertase to U937 cells. The findings are included in Appendix 1 Figure 1B and the main text as follows:

However, for the suggested analysis, we had to circumvent the non-specific binding of invertase to the cells and replaced it with label zymosan as a stimulant. The outcomes of these experiments are the following:

The corresponding channel capacity values in each layer is written in the plots in Author response image 2. The data indicate that the channel capacity between bound ligands to GFP expression is almost always higher than that of between ligand concentration and GFP. Therefore, the downstream pathway connecting bound ligand to GFP expression information has more capability in transmitting information than the receptor itself. In case of dectin-2 stimulation via anti-dectin-2 antibody (Author response image 2D), we suspect that lower channel capacity between bound ligand and GFP expression (0.42 bit) than ligand concentration to GFP (0.59 bits) is due to degradation of the APC dye over the time course of 16h stimulation.

Author response image 2
Single cell resolved scattered plot showing distribution of GFP expression and bound-labelled input on WT (A), dectin- (B)1, and dectin-2 (C) cells with various concentrations of labelled-zymosan or dectin-2 antibodies (D).

In the case of channel capacity comparison between ligand concentrations to bound ligand and ligand concentrations to GFP, dectin-1 shows less channel capacity in ligand concentration to bound ligands than the other. Please see more data points shown in Author response image 3.

Author response image 3
Channel capacity values in the presence of labelled input information.

According to Reviewer #1’s comment and general intuition, it is hard to understand that the channel capacity of a full communication channel is higher than a fraction of it. We think there is a dissipation of information in the ligand bound state coming from stimulant digestion or off-rate. Several ligands can bind to dectin-1 and initiate downstream signaling, but some may detach from the receptor prior to internalization, but still initiating signaling.

Note that the channel capacity calculation from bound ligand to GFP expression need extra care on the bias generation since the binning is also applied on the input variable. Therefore, the bias increases more rapidly as increasing the number of binning on both input and output (Author response image 4). On the other hand, insufficient number of binning can underestimate the calculated channel capacity. To address this problem, we employed equal frequency binning. The equal frequency binning partitions the data set such that the individual binning region contains the same number of data points. The equal frequency binning, compared with linear binning, is faster approaches the maximum channel capacity with increasing number of binning (see also Author response image 5).

Author response image 4
Channel capacity calculation including labelled zymosan information.

A) Schematic diagram representing both the layers where the corresponding information is measured (rectangular box) and the transmissions of information between layers (colored allows). B-D) Channel capacity values calculated from varied binning number and binning method for dectin-1, dectin-2 and WT U937 cells. The information of individual colors in the plot is given in A). The error bars in the plots represent the 95% confidence interval of the channel capacity value.

Author response image 5
Output binning number and binning method dependence of channel capacity value for experimental dataset.

The inset plots show the relative difference of channel capacity value to the maximum channel capacity value in the entire binning range (i.e., from 10 to 1000) of the corresponding binning method.

Taken together, even we could calculate the transmitted information through a receptor bound state, the measured labelled input information does not fully capture the total amount of receptor ligand interactions and therefore we could not quantitatively explain the results. Hence, we provide out experiments and analysis for the Reviewers only. We think implementation of measuring temporal dynamics of labelled ligand binding and unbinding with pH stable fluorophores, can provide actual information of ligand receptor interactions. This is ongoing research in the lab and we again thank the reviewer for his/her input.

Firstly, it is unclear how representative the bound protein numbers are of the actual number of binding events we would have in the cell culture when the cells are exposed to the ligand. This is because the authors have to wash the cells (at least I assume they do that) and then put them through the FACS instrument to measure the fluorescence. It would be really cool of the authors could do a time course to measure how fast the ligand "falls off" of the receptor. As long as the measurements made by the authors occur before significant unbinding has happened, the channel capacity they calculate should be very representative of what actually happened in the culture.

We agree with the reviewer, unbinding of ligand could obscure the data analysis (see also above). In our assay, ligands were always washed before flow experiments. We assume that due to multivalent nature of glycan-lectin interactions the desorption rate for the studied system is very low, which is why the washing step does not affect binding. Our assumption stems from working with these receptors in isolated biophysical assays (e.g. SPR) in previous work (Aretz et al. ACIE 2017). This effect will be even more pronounced if multiple copies of these receptors are present on a cellular surface. Thus, the measured channel capacity is likely to be representative of what is happening in cell culture. We also would like to note that the system presented here is not ideal to investigate dynamics of lectin-glycan interactions. In our follow-up studies, we plan to establish a kinetic microscopy assay to probe such interactions on a single-receptor level. We also plan to use not the end-point reporter, as NF-κB-GFP, but real-time signalling sensors, like ERK-KTR.

The other experiment that would be really interesting would be to repeat the experiment with labeled ligands but using whatever primary cells are known to express dectin-1, dectin-2, etc. In other words, the fluorescence here is coming from the ligand, so there is no need to engineer any reporters into the cell. If the authors see similar distributions of bound receptor on the surface of primary cells, and similar channel capacities between ligand concentration input and the bound receptor as output, that would be very intriguing. In any case, doing this experiment would I think begin to address some of the concerns with physiological relevance raised above. It would also be extremely interesting as a contribution to the field.

According to Reviewer #1’s suggestion we compared labelled zymosan binding and dectin-2 expression level between our transfected U937 cell and primary cultured human monocytes as shown in Appendix 1—figure 6. We found that our model cells express one order of magnitude higher dectin-2 expression relative to the human monocytes (Appendix 1—figure 6B A) and this resulted in the similar labelled zymosan binding trend on both cell types as increasing the zymosan concentration (Appendix 1—figure 6B). We include the following sentence in Appendix 1 Figure 6 and main text, respectively:

Line 382: “Finally, it is important to take into consideration that our conclusions came from model cell lines, which were used as a surrogate for cell-type-specific lectin expression patterns of primary immune cells. Human monocytes and dectin-2 positive U937 cells have comparable receptor densities and respond similar to stimulation with zymosan particles (Appendix 1 Figure 6A and B).”

3) As mentioned in my public comments, another issue in this work is the fact that the authors are relatively reticent about how their calculations are actually performed. While the authors present the relevant equations for calculating the Mutual Information and claim to follow the approach of Cheong et al., the methods are not sufficiently detailed to understand exactly what they have done here. This is particularly important because their approach is clearly not identical to that used by Cheong, since their algorithm for maximizing the MI across input distributions is different. Cheong et al. approach this problem by trying a limited set of possible input distributions and using the one that gives the highest MI, while the authors here use a built-in optimization algorithm in python. It is not clear if this is the only point at which the authors deviate from the approach laid out by Cheong et al., or if they have modified other aspects of the calculation.

According to Reviewer #1’s comment we extended the channel capacity calculation procedure as described in Appendix 2 and 3 for the clear explanation. In the case of modulating input distribution to maximize the channel capacity, we compared the method given in Cheong et al. and our approach as shown in Appendix 2 Section 4.

Mutual information calculation under unimodal and bimodal input distributions. (A) Examples of unimodal input distributions. The parameter s is the standard deviation of the Gaussian function selected from 0.5, 1, 2, 4 and 8. There are 60 cases of input distributions. (B) Examples of bimodal input distribution containing the same s parameters of the unimodal distributions. The number of bimodal combinations of the distribution is 1496. Vertically sorted various unimodal (C) and bimodal (D) input marginal probability distribution by the mutual information yields of the distribution. The probability space for the maximum mutual information given from unimodal (E) and bimodal (F) input distributions.

As shown in Appendix 2—figure 3, we compared between channel capacity values calculated from predefined unimodal/bimodal distribution and that of our optimization approach using built in Python function. We found that our approach gives the same (at least in given significant figures) channel capacity, 1.01 bit, for both cases, if we use bimodal input distribution.

The authors need to explain how they approach the problem of correcting for finite size effects. Do they use the approach of Cheong et al., bootstrapping the data at different levels and then using a linear extrapolation of MI vs. 1/N (where N is the number of data points/cells) to extrapolate to the case of an infinite population size (1/N \to 0).

In the original manuscript we did not consider the bias originated from the finite sample side because the sample number is very high (~150,000). Please see the channel capacities evaluated from experimental dataset (Appendix 2—figure 8).

Channel capacity estimation of experimental data using bootstrapping in various y-binning number. The y-intercept values of the regression line are the estimated channel capacity in the given y-binning number. The number of subsampled data points in each inverse sample side is 30.

The maximum corrected bias using bootstrapping in the entire dataset is less than 0.03 bits. The slopes in the linear regression lines in the plot are inverse proportional to the sample size. Therefore, our relatively bigger sample size does not significantly affect the corrected channel capacity. But we admit that in the case of relatively smaller sample size, using bootstrapping is required and gives more reliable result. Therefore, to retain a generality in channel capacity calculation of this work, we included bootstrapping calculation in all channel capacity estimates. The detailed bootstrapping procedures are shown in Appendix 2 Section 7.

Did they use the same approach to choosing the number of bins to calculate the probability distribution over outputs? Cheong et al. did this by visual inspection, if I recall correctly, looking for a "plateau" in bin numbers where the MI calculated from the data was non-zero but the MI calculated for randomized data was 0. Based on the statements made by the authors, and Figure S6, they seem to have performed both of these steps-otherwise it is completely unclear where the confidence intervals on their MI/channel capacity estimates come from, and how they chose the number of bins to use. Although it is important to note that the randomized controls seem to be missing from Figure S6. Regardless, since the authors have evidently re-implemented the code, they need to explain exactly what they have done.

In the original manuscript, we found appropriate binning number using visual inspection as Cheong et al. did. In the case of bias estimation using randomized dataset, we could not find any noticeable bias generation in the entire binning range from 0 to 1000. But we did not include the data. In the revised manuscript, we employed more rigorous binning method as described below:

Appendix 2—figure 7 shows channel capacities calculated from bootstrapping for various binning and total subsample numbers. In the case of experimental data set (TNF-a stimulation on TNFAR), the total number of samples represent the number of subsampled dataset (without replacement) from the original dataset. The white line in the random dataset represents the 0.01 bits of contour line. This indicates that the bias of channel capacity calculation is not only dependent on the number of output but also the number of measured samples. Note that the smallest number of experimental dataset is around 99,000. Therefore, we can expect that the maximum bias in our channel capacity calculation is less than 0.01 bits in the binning range from 10 to 1000 according to the random dataset. Indeed, as shown in Appendix 2—figure 7D, channel capacity values for subsampled (51151 sample size) experimental and random dataset exhibit stable line in the given binning range.

Accordingly, we decided to choose channel capacity by which the maximum channel capacity value calculated from 10 to 1000 of output binning range. And we included this procedure in Appendix 2—figure 7.

I would suggest that the authors also adopt several improvements to the Cheong et al. approach that have been developed in subsequent works. In particular, Suderman et al. (2017, PNAS 114 (22), 5755-60) implemented a broader range of bootstrap samples across which to do the linear extrapolation to infinite population size, automated the choice of bin sizes, and developed a bootstrap approach to estimating confidence intervals that is much more realistic (and statistically appropriate) than simply using the confidence intervals from the linear extrapolation step. The authors can refer to the Suderman paper on how this was done; it should be relatively easy to implement these improvements, if they have not been done already.

According to Reviewer # 1’s suggestions, we have implemented additional statistical methods to improve the channel capacity estimates of the work as follows: We implemented bootstrapping method, and by using linear regression we extrapolated channel capacity value at infinite sample size (Appendix 2 Figure 6). On the other hand, the confidence intervals, having less than the minimum value of significant figure, are not present in this work. Instead, we compared different outcomes using non-parametric statistical test as Reviewer #1’s suggestion from comment #2.

In the case of binning size selection, we decided to choose channel capacity by which the maximum channel capacity value calculated from 10 to 1000 of the output binning range as described in Appendix 2—figure 7.

I appreciate the author's innovation of using an optimizer in python to maximize the MI. I would expect this actually produces higher "channel capacities" than the previous approaches mentioned above. That being said, it is unclear how this optimization algorithm works. The authors may expect readers to go to the python documentation, but there are problems there. For one, the algorithm implanted in python might change, so the documentation may describe an approach that is different from the one the authors actually used to perform their calculations. How various packages are implemented in python can change from time to time, and something so generic as "optimize" may be changed as better optimization algorithms become available in the future. So the authors should describe exactly what they did. Also, readers are, in general, not going to go to python documentation to understand the methods of a paper like this, nor should they have to. The authors should describe what the algorithm is doing as currently implemented, in terms that readers can readily understand.

We agree to Reviewer #1’s comment regarding the lacking explanation of mutual information maximization procedures. The explanation should be more general and independent. Therefore, we have corrected Appendix 2 Section 3. describing mutual information procedures as follows:

“Finding input weighting values, w(i), that maximize the mutual information subject to iinputw(i)Px(i)=1 and 0w(i)Px(i) is a non-linear optimization problem. Since, the direction of the gradient of mutual information is the same as that of those two constraints at the minimum, using Lagrange multiplier method, one can restated the functions as Lagrangian L(wi,λ,σ)=MI(input;output|w(i))λ[iinputw(i)Px(i)1]σ[w(i)Px(i)], and find out weightings using numeric approach (Kraft D. 1988. A Software Package for Sequential Quadratic Programming. Wiss. Berichtswesen d. DFVLR). We used Sequential Least Squares Programming (SLSQP) provided by SciPy Python library (scipy.optimize.minimize, SciPy 1.7.3) to find out the optimizing input weighting values.”

Finally, the authors should also perform the calculation using the exact same distributions used by Cheong et al. and Suderman et al. This will allow us to compare their results more directly to the results from those previous authors, as well as allow us to determine if the approach that they used results in similar channel capacities to the optimization algorithm used here. While I expect the optimization algorithm employed by the authors will yield larger numbers for the channel capacity (and thus better estimates, since the channel capacity is formally a supremum), it would be extremely useful if the authors checked this.

According to Reviewer #1’s suggestion, we compared the channel capacities calculated from predefined bimodal and unimodal input distribution with our systematic approach as described in Appendix 2—figure 3 and Appendix 2 Section 4. We found that there is no noticeable channel capacity difference between those two methods in our significant figure range.

4) Another issue related to the calculation of the MI itself is the fact that the authors use logarithmically spaced bins, rather than linear bins, to perform the calculation. The authors describe this approach as if the matter of how to construct the bins is simply a matter of personal preference or convenience for the calculation, and choose logarithmic bins because they give somewhat larger values for the channel capacity. For one, this is not surprising; as is often the case, the FACS data obtained by the authors has a log-normal-like character; in other words, the authors see normal-ish distributions on a log scale (e.g. Figure 2A). As such, it is natural that using logarithmic bins will result in higher channel capacities, because it will tend to equalize (as much as possible) the effective number of observations within the bins.

While this may seem reasonable at first, it is actually, in my view, hard to justify. The reason is that the choice of how to generate the bins is not arbitrary, but actually corresponds to a very strong statement regarding what the authors actually consider the relevant "output" of the system to be. Using logarithmic bins is obviously equivalent to first taking the log of the data, and then using linear bins on that log-transformed data (Figure S2A should make this fact abundantly clear). The authors are thus envisioning that the channel is not a channel whose input is ligand concentration and output is GFP level, but rather a channel where the input is the ligand concentration and the output is the log of the GFP level. This is not an arbitrary decision, but rather one that has real consequences for how we construe the calculation in a physiological context. If the output is the log of the GFP level, that means that the cell is sensitive not to linear changes in protein concentration, but rather fold-changes.

For certain transcription factors, authors have argued that it is actually this kind of relative, fold change that matters (see Lee et al. (2014) Mol Cell 53 (6) 867-79). I would argue, however, that the conclusions of that work are hardly so robust as to suggest that, for every input-output transcriptional regulatory system within cells, it is the fold change of the "output" protein, rather than its linear change, that matters to the cell. Certainly, if the output is an enzyme, a transporter, or a protein that performs a whole host of other functions, then it is natural to assume that the appropriate output is the protein concentration itself, not the fold change in protein concentration.

If the authors can argue that every single gene whose transcription is controlled by NFκB has a fold-change impact on its downstream function, then I could see a strong argument being made for logarithmic bins. As it stands, however, I think the most natural way to interpret the output of the channel is on a linear scale, and I would say the authors should focus on that, rather than a logarithmic scale. If the authors want to include their logarithmic calculations in this work, I would suggest moving them to the supplement, and making clear the fact that such a calculation construes the output as having a fold-change impact on whatever is downstream of the protein level measured by the authors.

We understand Reviewer #1’s comment. Please see the public comment #3 and our answer including Author response image 5. We found whether we choose logarithmic, linear or equal frequency binning, there is no significant difference in the calculated channel capacity values if the binning number is more than 200. The binning method, however, should not be arbitrary but has to have strong relationship with the context. Therefore, in the revised manuscript, we used linear binning and corrected all the channel capacity values accordingly. Please note that the new calculation does not affect our conclusion.

5) In this work, the authors focus on the measurement of GFP levels at a certain time. As written, it is not clear if the authors are considering the steady-state response of the cells that they are treating, a peak response, or something else. As far as I can tell, the authors stimulate the cells for 16 or 13 hours, depending on the ligand in question. I am not sure how these numbers were chosen-did the authors look at the GFP expression dynamics and choose this number based on those measurements? From an information-theoretic standpoint, it is best if the results are based on steady-state protein levels, but the author's don't seem to explain their rationale for choosing the time points that they choose. If the GFP expression levels do not reach a steady-state, it would be great if the authors had a particular reason for choosing the times they choose. I would also suggest that the authors consider looking to see if their channel capacity estimates are robust to the time they choose by looking at other time points and calculating the channel capacities to see if they get similar results. I am not sure what times to choose, but perhaps 24 hours or 36? It is hard to say without looking at the average dynamics over time, so providing that kind of data would be extremely helpful.

Thank you for bringing this to our attention. We noticed that the stimulation time written in Materials and methods part is not correct. We stimulated 16h in every experiment except TNF-a stimulation as described in the original main text. The time at which the mean value of the GFP expression level reaches its maximum was determined as the stimulation time. As shown in (Appendix 1—figure 2A), U937 cells exhibit maximum mean GFP response at around 12 hours of TNF-a stimulation (50 ng/mL) while the expression level stay in steady state. And at around 20 hours later, the GFP expression level decreases to the basal level. Therefore, we selected 13 hours as the stimulation time if the TNF-a is the stimulant. For the other stimulants, we used 16 hours as the stimulation time according to the maximum GFP response. Author response image 6 shows the channel capacity estimate of dectin-2 expressing U937 cell using mannan as the stimulant at different time points. We include these data into our manuscript as follows:

Line 168: “Note that we choose the stimulation time, the period of incubation time of the cell with the input ligands, as the time point when GFP response and channel capacity reaches the maximum and steady state value (Appendix 1 Figure 2A and B).”

Author response image 6
Channel capacity of TNFαR channel to TNF-α stimulant for various stimulation time.

Error bars indicate 95 % confidence interval.

Also, the authors should acknowledge that a large body of literature has emerged that makes the claim that it is not the individual time points that matter, but rather the entire time series of the response that should be construed as the output. The authors can refer to the Zhang et al. paper in Cell Systems, the paper by Selimkhanov et al. (Science (2014) 346 (6215) 1370-3), papers by the Hoffman group (notably Tang et al. (2020) Nat Commun 12 (1) 1272), and a host of others. I am not suggesting that the authors adopt this worldview; there are actually serious mathematical and conceptual issues with construing the output of a communication channel to be a function like a time series. But, this idea is out there in the literature, and the authors should address this point and discuss how looking at GFP time series, rather than individual time point measurements, might yield different (and undoubtedly higher) channel capacity estimates.

We appreciate Reviewer #1’s suggestion for discussing channel capacity calculation from dynamics datasets. We understand there are lots of variability of selecting input and output structure. For example, one might use oscillating ligand concentration as an input and collect temporal NF-κB translocation into nucleus as an output (Kellogg, R. A. & Tay, S. Cell 160, 381–392 (2015)), which seemingly, even the authors did not consider information theory, yields more higher channel capacity than static input. Indeed, there are a lot of ways of selecting input and output for channel capacity calculation in the same signaling pathway. In this work, we were concerned with the information transmission via carbohydrates via lectin signaling pathways in a cell population level at a fixed time point. And we agree that describing channel capacity measurement using single cell and time resolved cellular output brings more broader insight on this work. This is ongoing research in the lab. We added following sentence in the Discussions and Conclusions part as follows:

Line 174: “In addition, this channel capacity can be further increased if one can measure the temporal evolution of output dynamics instead of static output dataset (Selimkhanov et al., 2014).”

6) As a final, rather important but also rather technical point, the authors in Figure 3C do statistical tests on their channel capacity estimates, comparing dectin-1 and dectin-2 cells to those that express dectin-1 and dectin-2 together. The issue here is that they use a t-test to estimate statistical significance, but it is really unclear where the dispersion (e.g. error bars or standard deviations) in the channel capacity estimates come from. I expect, since the authors day "n = 9" in the legend, that they split their data into 9 groups (maybe 9 sets of cells whose results were collected in different FACS runs or something?), performed the channel capacity estimate on each independently, then used the means and standard deviations in those estimates to do the t-test. That is completely made up by me, however; the authors really need to explain where these error estimates (and their +/- X values for channel capacity estimates in the various tables) actually come from.

In any case, the t-test of course only works if the data in question is drawn from actual Gaussian distributions. This does not mean that the distributions "look Gaussian" or "seem okay;" in the application of a t-test, you make extremely strong assumptions that only real honest-to-goodness Gaussian distributions actually satisfy. So the data really needs to actually be drawn from a Gaussian distribution for the results of the test to be interpreted appropriately.

My first suggestion here is that the authors need to use a non-parametric statistical test. That is, unless there is a true, theoretical reason to expect that the data will be drawn from a Gaussian (I know of no such reason, but perhaps the authors have one). The Wilcoxon rank-sum test was made for problems like this one, so I would suggest the authors use that.

A larger problem here, however, is that taking n = 9 (however it was done in the end) is not going to give a great estimate of the uncertainty in the channel capacity estimate. I would instead suggest that the authors merge their data into one data set (unless there are some kind of batch effects that prevent them from doing that) and then using a bootstrap/resampling approach, as in the Suderman et al. PNAS paper, to estimate confidence intervals and generate distributions for statistical tests. These bootstraps can be done hundreds or thousands of times, allowing for even more non-parametric tests (like permutation tests) to be used to estimate statistical significance.

This is honestly a fairly minor point, because nothing terribly important in the conclusions of the paper rest on the statistical tests here. But, if the authors want to argue that there are real differences between dectin-1, dectin-2 and dectin-1+dectin-2 cells, they should do these statistics in a more non-parametric way.

Since it is unclear whether the channel capacities are distributed Gaussian distribution across different batches of experiment, we have changed statistical test from t-test to Wilcoxon rank-sum test, a non-parametric statistical test, if we compare channel capacities in the manuscript.

Figure 2B and D include the significances from Wilcoxon rank-sum test, indicating the preserved statistical significance (*p>0.05, **p>0.01) that we had in the original manuscript.

In the revised manuscript, we do not compare the other pairs of receptors due to the possible difference in downstream molecule number of U937. And also please note that the pooling the dataset that measured in different dates usually generate high bias in the channel capacity since the baseline of the GFP signal is significantly different due to the variation of laser condition of flow cytometer day-to-day.

Reviewer #2 (Recommendations for the authors):

I think this paper would be greatly improved if the authors were more precise and detailed in the description of their capacity calculation. To address this, it's likely the authors were following the method from Cheong et al., but some of those details should be described here.

We thank Reviewer #2’s suggestion of improving the description of channel capacity calculation. Therefore, in the revised manuscript we added the following contents in Appendix 2.

https://doi.org/10.7554/eLife.69415.sa2

Article and author information

Author details

  1. Felix F Fuchsberger

    1. Department of Pharmaceutical Sciences, University of Vienna, Vienna, Austria
    2. Department of Biomolecular Systems, Max Planck Institute of Colloids and Interfaces, Potsdam, Germany
    3. Department of Microbiology, Immunology and Genetics University of Vienna, Max F. Perutz Labs, Vienna, Austria
    Contribution
    Conceptualization, Formal analysis, Validation, Investigation, Visualization, Writing – original draft, Writing – review and editing
    Contributed equally with
    Dongyoon Kim
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-9379-9792
  2. Dongyoon Kim

    1. Department of Pharmaceutical Sciences, University of Vienna, Vienna, Austria
    2. Department of Biomolecular Systems, Max Planck Institute of Colloids and Interfaces, Potsdam, Germany
    3. Department of Microbiology, Immunology and Genetics University of Vienna, Max F. Perutz Labs, Vienna, Austria
    Contribution
    Conceptualization, Resources, Data curation, Software, Formal analysis, Investigation, Visualization, Methodology, Writing – original draft, Writing – review and editing
    Contributed equally with
    Felix F Fuchsberger
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-9162-8730
  3. Natalia Baranova

    1. Department of Pharmaceutical Sciences, University of Vienna, Vienna, Austria
    2. Department of Microbiology, Immunology and Genetics University of Vienna, Max F. Perutz Labs, Vienna, Austria
    Contribution
    Conceptualization, Formal analysis, Investigation, Writing – review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-3086-9124
  4. Hanka Vrban

    1. Department of Pharmaceutical Sciences, University of Vienna, Vienna, Austria
    2. Department of Microbiology, Immunology and Genetics University of Vienna, Max F. Perutz Labs, Vienna, Austria
    Contribution
    Investigation
    Competing interests
    No competing interests declared
  5. Marten Kagelmacher

    Department of Biomolecular Systems, Max Planck Institute of Colloids and Interfaces, Potsdam, Germany
    Contribution
    Validation, Investigation
    Competing interests
    No competing interests declared
  6. Robert Wawrzinek

    1. Department of Pharmaceutical Sciences, University of Vienna, Vienna, Austria
    2. Department of Biomolecular Systems, Max Planck Institute of Colloids and Interfaces, Potsdam, Germany
    3. Department of Microbiology, Immunology and Genetics University of Vienna, Max F. Perutz Labs, Vienna, Austria
    Contribution
    Formal analysis, Investigation, Writing – review and editing
    Competing interests
    No competing interests declared
  7. Christoph Rademacher

    1. Department of Pharmaceutical Sciences, University of Vienna, Vienna, Austria
    2. Department of Biomolecular Systems, Max Planck Institute of Colloids and Interfaces, Potsdam, Germany
    3. Department of Microbiology, Immunology and Genetics University of Vienna, Max F. Perutz Labs, Vienna, Austria
    Contribution
    Conceptualization, Supervision, Funding acquisition, Investigation, Writing – original draft, Project administration, Writing – review and editing
    For correspondence
    Christoph.rademacher@univie.ac.at
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-7082-7239

Funding

European Research Council (716024)

  • Christoph Rademacher

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. Open access funding provided by Max Planck Society.

Acknowledgements

This project (GLYCONOISE) has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation program (Grant agreement No. 716024). We thank Max Planck Society for support and Prof. Dr. Peter H Seeberger for helpful discussions. We also thank the Deutsches Rheuma Forschungszentrum (DRFZ) for providing access to their cell sorting facility. The computational results presented were obtained using the CLIP cluster (https://clip.science).

Senior Editor

  1. Aleksandra M Walczak, CNRS, France

Reviewing Editor

  1. Andre Levchenko, Yale University, United States

Version history

  1. Received: April 14, 2021
  2. Preprint posted: May 10, 2021 (view preprint)
  3. Accepted: February 19, 2023
  4. Accepted Manuscript published: February 20, 2023 (version 1)
  5. Version of Record published: March 14, 2023 (version 2)
  6. Version of Record updated: March 31, 2023 (version 3)

Copyright

© 2023, Fuchsberger, Kim et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 1,216
    Page views
  • 232
    Downloads
  • 2
    Citations

Article citation count generated by polling the highest count across the following sources: PubMed Central, Crossref, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

  1. Felix F Fuchsberger
  2. Dongyoon Kim
  3. Natalia Baranova
  4. Hanka Vrban
  5. Marten Kagelmacher
  6. Robert Wawrzinek
  7. Christoph Rademacher
(2023)
Information transfer in mammalian glycan-based communication
eLife 12:e69415.
https://doi.org/10.7554/eLife.69415

Share this article

https://doi.org/10.7554/eLife.69415

Further reading

    1. Cell Biology
    2. Genetics and Genomics
    Guanxiong Yan, Yang Ma ... Wei Miao
    Research Article

    Although most species have two sexes, multisexual (or multi-mating type) species are also widespread. However, it is unclear how mating-type recognition is achieved at the molecular level in multisexual species. The unicellular ciliate Tetrahymena thermophila has seven mating types, which are determined by the MTA and MTB proteins. In this study, we found that both proteins are essential for cells to send or receive complete mating-type information, and transmission of the mating-type signal requires both proteins to be expressed in the same cell. We found that MTA and MTB form a mating-type recognition complex that localizes to the plasma membrane, but not to the cilia. Stimulation experiments showed that the mating-type-specific regions of MTA and MTB mediate both self- and non-self-recognition, indicating that T. thermophila uses a dual approach to achieve mating-type recognition. Our results suggest that MTA and MTB form an elaborate multifunctional protein complex that can identify cells of both self and non-self mating types in order to inhibit or activate mating, respectively.

    1. Cell Biology
    2. Neuroscience
    Anna Kádková, Jacqueline Murach ... Jakob Balslev Sørensen
    Research Article

    SNAP25 is one of three neuronal SNAREs driving synaptic vesicle exocytosis. We studied three mutations in SNAP25 that cause epileptic encephalopathy: V48F, and D166Y in the synaptotagmin-1 (Syt1)-binding interface, and I67N, which destabilizes the SNARE complex. All three mutations reduced Syt1-dependent vesicle docking to SNARE-carrying liposomes and Ca2+-stimulated membrane fusion in vitro and when expressed in mouse hippocampal neurons. The V48F and D166Y mutants (with potency D166Y > V48F) led to reduced readily releasable pool (RRP) size, due to increased spontaneous (miniature Excitatory Postsynaptic Current, mEPSC) release and decreased priming rates. These mutations lowered the energy barrier for fusion and increased the release probability, which are gain-of-function features not found in Syt1 knockout (KO) neurons; normalized mEPSC release rates were higher (potency D166Y > V48F) than in the Syt1 KO. These mutations (potency D166Y > V48F) increased spontaneous association to partner SNAREs, resulting in unregulated membrane fusion. In contrast, the I67N mutant decreased mEPSC frequency and evoked EPSC amplitudes due to an increase in the height of the energy barrier for fusion, whereas the RRP size was unaffected. This could be partly compensated by positive charges lowering the energy barrier. Overall, pathogenic mutations in SNAP25 cause complex changes in the energy landscape for priming and fusion.