A unified network systems approach uncovers a core program underlying T follicular helper cell differentiation Authors: Alisa A. Omelchenko, Syed A. Rahman, Alyson Wang, Tracy Tabib, Vinayak V. Viswanadham, Grace J. Yuen, Perla M. Del Rio Estrada, Valentino D’Onofrio, Yijia Chen, Na Sun, Hamid Mattoo, Swapnil Keshari, Chinmay V. Gadiraju, Gonzalo Salgado, Maribel Soto-Nava, Lady Ruiz-Carbajal, Dafne Díaz-Rivera, Santiago Ávila-Ríos, Sudhir P. Kasturi, Susan P. Ribeiro, Mark J. Shlomchik, Amanda C. Poholek, Shiv Pillai, Rebecca A. Elsner, Vinay S. Mahajan, Jishnu Das Published: 2026 ABSTRACT Characterizing multi-scale processes underlying immune-state progression is critical for defining their function. This becomes pertinent for functionally diverse and plastic immune cells, such as T follicular helper (Tfh) cells. Here, we adopt a multi-scale network-systems approach that incorporates both regulatory and protein-protein interactions. This approach integrates diverse data types, captures regulation across levels of immune system organization, and recapitulates known Tfh differentiation drivers. Further, we present CoreNet, a core Tfh gene set that is conserved between humans and mice, across tissue types and disease contexts, and is consistent across data modalities. Using CoreNet, we implicate NR3C1 and interleukin (IL)-12 in the regulation of Tfh differentiation. Notably, IL-12 is permissive for differentiation of Tfh precursors but blocks differentiation into germinal center Tfh cells. Overall, this work elucidates networks with unexplored roles governing Tfh differentiation across species and tissues, while providing a generalizable framework. CoreNet is accessible through an interactive web server: https://pitt-csi.shinyapps.io/tfhcorenet/. Introduction The immune system relies on its diversity to adjust rapidly and mount responses against the wide spectrum of microbial pathogens it encounters1. Key principles of immune cell diversity have been largely uncovered through analyses centered on lists of molecular markers, including surface markers, cytokines, and transcription factors2–4. However, heterogeneity in cell states is a complex multi-signal process involving epigenetic, transcriptional, and functional variation. Understanding the mechanisms governing lineage commitment and plasticity remains a fundamental question in immunology5–7. Lists of marker genes, although informative in specific contexts, fail to capture the full complexity of tightly interconnected regulatory layers underlying immune state specificity and progression8. Consequently, systems-level approaches such as network analysis are needed to reveal how signals are integrated within immune cells and to uncover the functional mechanisms that drive state definition and transitions. Network approaches provide a unique opportunity to integrate signals from multiple omics datasets, apply prior biological information such as protein interactions or signaling cascades, and present an interpretable representation of genes (nodes) and their influence (edges). By supplying this functional context and directional layers of connectivity, network models allow researchers to formulate mechanistically driven hypotheses offering insights beyond what gene lists alone can reveal. In order to identify networks underlying state progression, we adopted a multi-scale network systems approach that incorporates both regulatory and physical interactions9. Our network-derived core gene approach, CoreNet, integrates diverse data types and captures regulation across multiple levels of immune system organization10,11. At the transcriptional level, gene regulatory networks (GRNs) provide insights into how genes coordinate and execute immune responses, leading to the discovery of new regulatory components12,13. In addition, protein-protein interactions are also critical regulators of differentiation. These physical interactions transmit extracellular contextual cues, such as local inflammatory cytokines and cell-cell interactions, through ligand-receptor associations. Thus, simultaneously incorporating the protein-protein interaction (PPI) network provides orthogonal insights14. The PPI network, built from experimentally validated physical interactions, enables the identification of important protein groups with a specific connectivity structure based on their physical associations (i.e., modules). To demonstrate the power of CoreNet we applied it to the T follicular helper (Tfh) CD4+ subset. Tfh cells are essential for supporting B cell maturation within germinal centers (GCs), driving somatic hypermutation, affinity maturation, and long-lived antibody responses15. They are indispensable for protective humoral immunity to infection. Dysregulated Tfh cell responses have also been implicated in autoimmunity and cancer16–18. Despite their importance, inconsistent findings across species and experimental systems have fueled ongoing debates regarding the core regulatory pathways that govern Tfh differentiation19,20. For example, relying heavily on in vitro culture systems, there have been reports of the necessity of TGF-β and activin A for human, but not murine, Tfh differentiation20,21. There has also been a long standing debate on the role of IL-12 in this process when comparing in vitro human data to in vivo murine data across various infectious models22–27. These inconsistencies are partially due to the lack of in vitro conditions that produce cells representative of Tfh isolated from lymphoid organs19, but further complicated by heterogeneity among Tfh populations that is still poorly defined28,29. Part of our gaps in understanding stems from the fact that many studies have experimentally investigated individual proteins and circuits necessary for Tfh differentiation in specific contexts30–34, each providing only a limited window into the process. For example, recent analyses have aimed to identify a core set of Tfh signatures conserved across tissues and infection35. Computing simple overlaps of genes shared vs different across infection states yields corresponding lists of candidate genes. However, this approach offers limited mechanistic insight to explain why these genes emerge as significant in terms of their underlying connectivity i.e., a bag of genes inherently does not have the connectivity information that a network does. Further, these core Tfh signatures do not explore the regulatory and protein interaction networks implicated in state progression (from naive to GC Tfh cells), or how these networks compare across murine and human models. In the absence of a unified systems approach to analyzing these disparate datasets, it currently remains unclear how much of the observed inconsistencies arise from technical artifacts versus fundamental biological differences (if any) between human and murine Tfh differentiation. Here, we use network approaches to systematically elucidate circuits underlying Tfh state progression by integrating data from a set of comprehensive ‘omic profiles. We then interrogate the identified circuits across data types, disease contexts, and species to evaluate the generalizability of the circuits as well as their conservation across species. Specifically, we first demonstrated that our approach can integrate different data modalities from a single set of experiments and identify known regulators of Tfh differentiation curated from many studies. By applying this approach to more diverse datasets, we then discovered that there is indeed a core Tfh (CoreNet) gene set which is conserved across tissue types and diseases. The CoreNet gene set is also consistent across bulk, single-cell and spatial data modalities. Orthogonal analyses using distinct approaches of single-cell and spatial datasets across contexts also yielded significant overlaps with the CoreNet gene set, demonstrating its robustness and biological relevance. While several components of this set have been individually reported, unique aspects of our work lie in providing functional insight into the core Tfh program by identifying its network connectivity as well as discovering and characterizing targets using an unbiased approach. More importantly, we found that this core set is conserved across humans and mice, addressing a fundamental open question in the field, as different systems (in-vitro, ex-vivo and in-vivo) have often led to different components being discovered. Further, the CoreNet gene set is functionally linked to specific immune signaling pathways. Our findings reveal a function of IL-12 in the regulation of Tfh heterogeneity. IL-12 specifically blocks the differentiation of Tfh precursors into GC Tfh cells, in both humans and mice, consistent with recent studies indicating that the precursor and GC Tfh populations correspond to transcriptionally distinct cell states within the Tfh differentiation trajectory36,37. Overall, we uncover networks with unexplored roles in governing Tfh cell differentiation across species and tissues, paving the way for potential therapeutic interventions. Our CoreNet approach provides a generalizable framework that can be applied not only to immune cells, but across biological contexts. Results An unbiased network systems approach empowers discovery of modules underlying Tfh differentiation We characterized the epigenomic and transcriptomic landscapes of Tfh cells isolated from adult healthy human tonsils at distinct stages of differentiation. We profiled chromatin accessibility with ATAC-Seq38,39, gene expression with RNAseq, and histone H3 modifications with CUT&RUN40 using antibodies against H3K4me3, H3K4me1 and H3K27Ac to map promoters, enhancers and their overall chromatin activation, respectively. Each of these analyses was performed at four stages of Tfh cell differentiation - Naive (CD45RA + CXCR5-PD1-), Early Extra-GC Tfh/Pre-Tfh (CD45RA-CXCR5 + PD1-), Late Extra-GC Tfh/Pro-Tfh (CD45RA-CXCR5 + PD1 + ), and GC Tfh cells (CD45RA-CXCR5 + + PD1 + +) (Fig. 1a). The CD45RA-CXCR5 + PD1- and CD45RA-CXCR5 + PD1+ stages express intermediate levels of CXCR5, PD1 and BCL6 compared to naive and mature Tfh cells, suggesting they represent intermediate stages in the Tfh differentiation pathway41. This resource represents a global epigenomic roadmap of human Tfh cell differentiation states (Fig. 1b). Low-dimensional representations of data from each chromatin assay revealed distinct stage-specific profiles across the four cell subsets that positioned them along a trajectory from the least to the most differentiated state, reflecting the stepwise activation of regulatory programs during Tfh cell differentiation (Supplementary Fig. 1a, 1b).Fig. 1Overview of the CoreNet framework for discovering core programs underlying Tfh differentiation.a Flow cytometry gating strategy to identify Tfh cells and their differentiation state from the 600 million human tonsil cells. b Tfh cell differentiation trajectory of the 4 states identified through flow cytometry. c Schematic displaying multi-omic datasets each providing a specific view of a system and how using networks as biological priors can integrate the views to identify a core program across different conditions such as species. Created in BioRender. Omelchenko, A. (2026) https://BioRender.com/41whr23. d Overview showing how targets were discovered using bulk datasets and refined using single-cell and spatial data. Targets were then validated in a murine in vivo model. Created in BioRender. Omelchenko, A. (2026) https://BioRender.com/360ga2s. e Method for identification of top driver TFs through integration of the epigenomic/transcriptomic datasets with Personalized PageRank network propagation and filtering the output ranks. Created in BioRender. Omelchenko, A. (2026) https://BioRender.com/rauq29k. f Protein-protein interaction network propagation using the gene centric peak proximity score (PPS) to integrate epigenomic datasets and identify significant subnetworks. Created in BioRender. Omelchenko, A. (2026) https://BioRender.com/srebdaj. Validation of bulk data findings with single cell RNA (scRNA) and spatial datasets across contexts like disease and species using single sample gene set enrichment analysis (ssGSEA) and functional microniche analysis. Created in BioRender. Omelchenko, A. (2026) https://BioRender.com/gylqis8. Overview of the CoreNet framework for discovering core programs underlying Tfh differentiation. a Flow cytometry gating strategy to identify Tfh cells and their differentiation state from the 600 million human tonsil cells. b Tfh cell differentiation trajectory of the 4 states identified through flow cytometry. c Schematic displaying multi-omic datasets each providing a specific view of a system and how using networks as biological priors can integrate the views to identify a core program across different conditions such as species. Created in BioRender. Omelchenko, A. (2026) https://BioRender.com/41whr23. d Overview showing how targets were discovered using bulk datasets and refined using single-cell and spatial data. Targets were then validated in a murine in vivo model. Created in BioRender. Omelchenko, A. (2026) https://BioRender.com/360ga2s. e Method for identification of top driver TFs through integration of the epigenomic/transcriptomic datasets with Personalized PageRank network propagation and filtering the output ranks. Created in BioRender. Omelchenko, A. (2026) https://BioRender.com/rauq29k. f Protein-protein interaction network propagation using the gene centric peak proximity score (PPS) to integrate epigenomic datasets and identify significant subnetworks. Created in BioRender. Omelchenko, A. (2026) https://BioRender.com/srebdaj. Validation of bulk data findings with single cell RNA (scRNA) and spatial datasets across contexts like disease and species using single sample gene set enrichment analysis (ssGSEA) and functional microniche analysis. Created in BioRender. Omelchenko, A. (2026) https://BioRender.com/gylqis8. We adopted a multi-scale network approach, CoreNet, to integrate these omic datasets since each individual dataset typically captures only a partial view of the complex and interconnected immune networks underlying immune responses (Fig. 1c). This fragmented perspective can obscure underlying biological mechanisms and fuel scientific disagreements, as researchers interpret only specific slices of the data landscape. Network and systems-based approaches help overcome this challenge by integrating diverse modalities to generate multiscale, connected, directional, unbiased, and interpretable views of immune responses10,11,42. Further, although bulk datasets lack cellular resolution43, when paired with biological networks they can support discovery by identifying functional modules and key regulators (Fig. 1d). Incorporation of single cell and spatial information further refines the core components, allowing for meaningful inference and identification of targets which can be experimentally validated (Fig. 1d). Using these datasets, we first identified prioritized transcription factors (TFs) underlying Tfh state progression. We used a personalized PageRank approach that leverages gene regulatory network (GRN) architecture, where the network is identified by combining the transcriptomic and epigenomic data (Fig. 1e). However, GRNs capture only one dimension of the regulatory landscape and physical protein interactions, such as receptor-ligand associations, also play a critical role in cellular differentiation. We therefore also included in our integrative framework a protein-protein interaction (PPI) network-centric approach that revealed additional components of the core regulatory program not captured by GRN-based approaches alone (Fig. 1f). For the PPI-based approach, we integrated signals across a range of datasets using a unique network propagation technique that combines diverse epigenomic signals (that are potentially individually noisy) into a robust set of modules. We first identified chromatin peaks, i.e., regions enriched for accessible chromatin or histone marks, that significantly varied across all four stages of Tfh differentiation (Supplementary Fig. 1b) and applied a gene-centric approach to integrate these epigenetic signals at the level of individual genes (Supplementary Fig. 1c). Our combined strategy thus prioritizes genes based on both their regulatory impact and their physical interaction context. To validate and refine the core transcriptional programs identified from bulk population data, we analyzed single-cell RNA-seq (scRNA-seq) and spatial datasets spanning diverse tissues, species, and disease contexts (Fig. 1g). Although scRNA-seq provides cellular resolution, it suffers from technical noise and dropout, whereas data from cell populations offers higher signal fidelity, enabling robust identification of core regulatory modules via network propagation. This integrative approach helped us to cross-validate findings from bulk population data across species and platforms, including single-cell and spatial datasets, and to prioritize conserved regulatory genes. We benchmarked our gene sets against established Tfh gene lists to ensure specificity and biological relevance of the CoreNet output. Network-based integration of epigenomic and transcriptomic data uncovers core processes driving Tfh cell differentiation We first applied a framework leveraging cell-state-specific GRNs to identify key TFs driving Tfh differentiation. Specifically, we used a Personalized PageRank approach to construct GRNs from epigenomic data44, and then implicate TFs that underlie the Tfh state progression. From each human tonsil bulk epigenomic dataset that was generated, we selected the top 100 TFs based on their highest rank across all stages, yielding a combined set of 152 unique TFs (CoreNet-GRN) from an initial pool of ~1,200 candidates (Fig. 2a–b, Methods). Notably, the top-ranking TFs showed only weak correlation with gene expression changes (Fig. 2a, y-axis), highlighting a critical deficiency of relying solely on differential gene expression to prioritize core Tfh signatures, a limitation other studies have failed to address35. This disconnect between differential gene expression and regulatory significance may be particularly relevant to the Tfh context, where key regulators of Tfh differentiation including BCL6 are subject to tight post-transcriptional and translational control45. Indeed, only a small number of genes encoding TFs with high mRNA fold change (log2FC) of expression were ranked highly by Taiji and passed standard significance thresholds.Fig. 2CoreNet identifies and validates gene regulatory and protein interactome networks implicated in Tfh differentiation.a The spread of the 1165 TFs analyzed by Taiji given by the mRNA log2FC between the GC and Early Extra-GC state (y-axis) and the Taiji Rank log2FC between the 2 states (x-axis) with all TFs (gray), those which passed the p-value threshold of <0.01 (orange), and the chosen 152 driver TF (CoreNet-GRN) set (blue). b A focused scatter plot on the CoreNet-GRN set without ZNFs analyzed by Taiji given by the mRNA log2FC between the GC and Early Extra-GC state (y-axis) and the Taiji Rank log2FC between the 2 states (x-axis). c All 27 modules identified with the protein-protein interaction network propagation (CoreNet-PPI). d Overlap of size matched gene sets of log2FC RNA and top PPS with their direct interactors with 2 reference Tfh sets. Size matched sets of the top log2FC RNA scores or PPS were identified. 1000 permutations of a random size matched sets were also compared for each analysis. One-sided Fisher’s exact test performed with average random overlap, p = 0.212, 0.030, 0.094, 0.039 from left to right. e CoreNet-PPI, CoreNet-GRN, and CoreNet (PPI + GRN) with their direct interactors (DI) overlap with 2 Tfh reference gene sets. 1000 permutations of a random size matched sets were also compared for each analysis. One-sided Fisher’s exact test performed with average random overlap, p = 0.009, 0.014, 0.001, 0.006, 0.010, 0.002 from left to right. f Overlay of CoreNet-DI on the Vinuesa 2016 reference network. Shared genes (circle), analysis only discovered interactors (diamond), and reference only genes (square). g Overlay of the log2FC RNA prioritized set with direct interactors on the Vinuesa 2016 reference network. Shared genes (circle), analysis only discovered interactors (diamond), and reference only genes (square). b–c, f–g Colored by the mRNA log2FC of the GC Tfh and Early Extra-GC Tfh states. Higher expression in GC is shown in purple, and higher in Early Extra-GC Tfh is shown in green. a–g source data are provided as a Source Data file. CoreNet identifies and validates gene regulatory and protein interactome networks implicated in Tfh differentiation. a The spread of the 1165 TFs analyzed by Taiji given by the mRNA log2FC between the GC and Early Extra-GC state (y-axis) and the Taiji Rank log2FC between the 2 states (x-axis) with all TFs (gray), those which passed the p-value threshold of <0.01 (orange), and the chosen 152 driver TF (CoreNet-GRN) set (blue). b A focused scatter plot on the CoreNet-GRN set without ZNFs analyzed by Taiji given by the mRNA log2FC between the GC and Early Extra-GC state (y-axis) and the Taiji Rank log2FC between the 2 states (x-axis). c All 27 modules identified with the protein-protein interaction network propagation (CoreNet-PPI). d Overlap of size matched gene sets of log2FC RNA and top PPS with their direct interactors with 2 reference Tfh sets. Size matched sets of the top log2FC RNA scores or PPS were identified. 1000 permutations of a random size matched sets were also compared for each analysis. One-sided Fisher’s exact test performed with average random overlap, p = 0.212, 0.030, 0.094, 0.039 from left to right. e CoreNet-PPI, CoreNet-GRN, and CoreNet (PPI + GRN) with their direct interactors (DI) overlap with 2 Tfh reference gene sets. 1000 permutations of a random size matched sets were also compared for each analysis. One-sided Fisher’s exact test performed with average random overlap, p = 0.009, 0.014, 0.001, 0.006, 0.010, 0.002 from left to right. f Overlay of CoreNet-DI on the Vinuesa 2016 reference network. Shared genes (circle), analysis only discovered interactors (diamond), and reference only genes (square). g Overlay of the log2FC RNA prioritized set with direct interactors on the Vinuesa 2016 reference network. Shared genes (circle), analysis only discovered interactors (diamond), and reference only genes (square). b–c, f–g Colored by the mRNA log2FC of the GC Tfh and Early Extra-GC Tfh states. Higher expression in GC is shown in purple, and higher in Early Extra-GC Tfh is shown in green. a–g source data are provided as a Source Data file. We examined the stage-specific relevance of key TFs by looking at the difference in rank across the Early Extra-GC and GC Tfh states (Fig. 2b, left to right on the X axis: moves from only relevant for Early Extra-GC state to relevant for both states to relevant for GC Tfh state only). KLF2 and FOSL2 were selectively ranked as only important for the Early Extra-GC Tfh state. This is consistent with prior work that demonstrates that KLF2 is widely expressed on naive mature CD4 + T-cells and repressed in GC Tfh cells46,47. Similarly, miR-155 mediated repression of FOSL2 has been shown to be important for Tfh cell fate commitment48,49. In contrast, we find that POU2F2 and PBX3 were important only in the GC Tfh state. The role of POU2F2 (OCT2) in promoting germinal centers as well as GC B-cells has been extensively studied50,51. Limited studies have shown POU2F2 to be involved in Tfh differentiation through transactivation of the BCL6 and BTLA promoters52,53. This analysis suggests that POU2F2 may play a more profound role specifically in GC Tfh differentiation, as opposed to other Tfh states. We additionally identified a putative factor, PBX3, which has not been studied in GC Tfh differentiation to our knowledge. Other TFs, such as BCL6, were important in both contexts. Although BCL6 shows a strong expression log₂FC favoring the GC Tfh state, its ranking underscores its broader relevance as a lineage-defining TF for Tfh cells54. PRDM1, a well-known repressor of the Tfh program54, exhibits minimal transcript-level changes. Despite this, it was still correctly prioritized as a key regulator of the Tfh program in both states consistent with the well-known requirement that PRDM1 needs to be turned off (along with BCL6 being turned on in a reciprocal loop) for the formation of GC Tfh54. Conventional expression-based analyses would miss key CoreNet-GRN hits like PRDM1, further demonstrating the value of using this regulatory network-informed approach. PPI integration identifies submodules in Tfh cell differentiation Next, to move beyond GRNs, which do not capture physical interactions such as receptor-ligand binding and signaling cascades that also play critical roles in cell differentiation, we integrated a high-confidence human PPI network into our epigenomic analyses. The validity of each protein interaction (edge) in the PPI network is supported by multiple independent experimental assays55,56. However, the PPI network is context agnostic. We added context-specific data to this network by using a score for each protein (node weight for the gene encoding the protein based on its aggregate “peak proximity score” (PPS) computed using all chromatin features)57. Stage-specific epigenetic signals for each chromatin feature were mapped to nearby genes and combined into a single PPS. The PPS was calculated as a distance-weighted sum of the number of variable CUT&RUN or ATAC-Seq peaks relative to a transcription start site, with equal contribution from each chromatin feature. The PPS, a gene-centric measure of stage-specific epigenetic activity, exhibited significant correlation with gene expression levels in both Early Extra-GC and GC Tfh stages, but the magnitudes of the Spearman correlation were modest (Supplementary Fig. 1d, e). The PPS values were then propagated along the PPI network using a random-walk-with-restart approach. There are two key goals achieved by the network propagation approach. First, it identifies high-scoring subnetworks based on propagated PPS signal by taking into account additive effects based on connected protein interactors (functional similarity) rather than focusing on stand-alone high-scoring proteins (nodes). Second, because omic assays may not capture all signals (false negatives), propagations spread the initial PPS signal through the network, allowing connected high-scoring neighborhoods (and thus, genes with weak or missing direct signals) to be detected despite incomplete input data. The discovery of meaningful regulatory modules depends upon both the input scores and the structure of the interaction network. To ensure that only robust and biologically grounded modules were retained, we selected statistically significant subnetworks using stringent permutation testing. Modules identified using randomized gene scores or shuffled networks were excluded ensuring that the chosen subnetworks reflected true biological signals. The prioritized modules were consistently identified across five independent network propagation replicate runs. By leveraging the topology of the PPI network as prior knowledge that encodes biological dependencies (functional similarity), CoreNet-PPI allowed us to identify coherent subnetworks that may underlie Tfh cell differentiation. The resulting 27 modules, comprising 184 genes, map to distinct regions of the interactome and represent <1% of the full network (Supplementary Fig. 2a). We overlaid the mRNA log2FC between GC and Early Extra-GC Tfh populations to distinguish which genes are more highly expressed in each differentiation state (Fig. 2c). Most modules span both states, capturing the dynamic regulation across the Tfh trajectory. Notably, the largest connected module includes several canonical Tfh-associated genes such as CD200 and PDCD1, underscoring its biological relevance (Fig. 2c). In contrast, some modules exhibit clear state-specificity. For example, a module enriched for IL-17 signaling components, including IL-17RC, IL-17F, IL-17RA, IL-17A, and IL-2 shows comparatively elevated expression in the Early Extra-GC Tfh state (Fig. 2c). Building on the identified subnetworks and ranked transcription factors, we integrated the driver TFs and the PPI module genes to define CoreNet and added their direct interactors (DI) within the protein network to define the CoreNet-DI gene set (Methods). To assess the biological relevance of these gene sets, we evaluated their overlap with proteins annotated in multiple experimentally validated Tfh reference datasets58,59. To test if the overlap could be explained by the input transcriptomic and epigenomic signals alone and not benefit from our network integration, we generated size-matched gene sets using either top-ranked genes from RNA-seq log₂FC or from epigenomic PPS values, along with their direct protein interactors (Methods). For all sets, we also generated 1,000 permutations of random size-matched controls to establish a random baseline for expected overlap. We found that the transcriptomic and the epigenomic data alone only weakly recovered biologically relevant genes (Fig. 2d, Supplementary Fig. 2B). These results indicate that while gene-centric scores provide useful starting points, they are incomplete on their own. We found that CoreNet-PPI-DI and CoreNet-GRN-DI showed substantially greater overlap with experimentally validated Tfh reference datasets, with enrichment levels between 2- to 4-fold above the random baseline (Fig. 2e, Supplementary Fig. 2c). Notably, the combined network approach outperformed individual methods, recovering 60–69% of known Tfh-associated genes compared to 45–55% for individual subnetworks (CoreNet-PPI-DI) or TF rankings (CoreNet-GRN-DI) alone. Even when using just the CoreNet genes there was a ~ 35% overlap with the reference sets compared to 0-3% with a random size matched set (Supplementary Fig. 2d). These results highlight the added value of integrating multiple network-derived signals. Either the CoreNet or CoreNet-DI genes could be further evaluated based on what a study prioritizes, a reduction of false negatives (CoreNet-DI) or a reduction of false positives (CoreNet). We also can look at the differential gene expression across states from gene lists of interest. To demonstrate this, we used a list of core T follicular regulator (Tfr) genes that overlapped with CoreNet-DI, and observed that while some Tfr exclusive genes are present in our gene set, they are more highly expressed in the Early Extra-GC state compared to a list of Tfh exclusive genes from the same publication35 (Supplementary Fig. 2e). Together, this demonstrates that network propagation provides a critical framework for identifying higher-order regulatory relationships that are not captured by transcriptomic or epigenomic data alone. Moreover, combining orthogonal (GRN and PPI) network-based strategies further improves the ability to recover biologically meaningful targets. It is not always clear how biologically curated lists of genes from literature are interconnected and higher order information must be provided to understand their connectivity. Here, we use a protein interaction network to provide the higher order functional relationship of these genes. We visualized the Vinuesa 2016 reference network by mapping their connections within the broader PPI interactome. We then overlaid CoreNet-DI onto this network and included any direct interactors of the reference genes that were also part of CoreNet (Fig. 2f). Several key regulators of Tfh differentiation, such as BCL6, PRDM1, STAT1, STAT3, and PDCD1, were successfully recovered (denoted as circles). In addition, many biologically relevant genes not included in the original reference set (shown as diamonds) served as connectors within the network. For example, PTPN11 (SHP2) links PDCD1 to STAT3 and STAT1, highlighting potential intermediates that bridge well-established signaling nodes60. By contrast, overlaying the top mRNA logFC identified genes with their direct interactors resulted in a disjointed network (Fig. 2g). While STAT1, STAT3, and IRF4 were identified in the mRNA logFC genes, their connections to each other through interactors such as STAT6, RELA, or CIITA have been lost. Moreover, we can visualize the rank importance and module expression changes across the differentiation trajectory by comparing GC Tfh with the Late Extra-GC Tfh state (Supplementary Fig. 2f, 2g). This demonstrates the power of using network propagation to give context to higher order organization across the Tfh differentiation trajectory. These results also illustrate that our prioritized gene set can capture both known regulators and less-characterized players in Tfh cell biology. However, some well-known Tfh-associated genes, including IL-21, CXCR5, and ICOS, were not identified through our approach. Since these are not TFs, these genes could only be recovered through the PPI network. We analyzed the shortest PPI paths from the reference set and found that although these genes initially carried strong signals, they were directly connected to interaction hubs. This caused signal dissipation during propagation. For instance, IL-21 and IL-21R are connected to SMARCA2, which interacts with 132 proteins (Supplementary Fig. 3a). Similarly, CXCR5 is connected to GNAI2, which interacts with 72 proteins and is linked to VIRMA, a node with over 2700 connections (Supplementary Fig. 3c). ICOS also interacts directly with VIRMA. These findings underscore an important limitation of PPI-based propagation: while the method effectively controls false positives, it may have some false negatives due to the incompleteness of data, but not methodological, limitations. Still, the interactome has sufficient coverage to perform rigorous systemic network-based approaches on and gain impactful insights61. Single cell RNA-seq datasets align more with CoreNet than with curated Tfh reference sets We hypothesized that if the modules derived from bulk omics datasets are truly reflective of the GC Tfh state, they should consistently exhibit higher enrichment scores in Tfh populations relative to non-Tfh cells across independent single-cell datasets. Therefore, we examined how well the modules identified earlier from bulk data extrapolate to scRNA-seq and spatial datasets spanning diverse tissue types and contexts. We also wanted to compare methods that focus on separating cell states with insights obtained from our approach, which is focused on discovering a core network underlying state progression. We leveraged multi-ome human tonsil atlas data to identify 2 cell populations along the Tfh differentiation pathway, CM Pre-Tfh and GC-Tfh-SAP (resembling Early Extra-GC Tfh and GC Tfh states, respectively) for comparison (Fig. 3a). We ran SCENIC + 62 to identify the regulons which separated these two states. These regulons consisted of 33 unique TFs and 2230 unique genes (Fig. 3b). Notably, 20 of the TFs were also identified by the CoreNet-GRN analysis demonstrating a significant overlap (p < 0.0001). Further, the regulons achieved strong separation of the two states in UMAP space (Fig. 3c). We then applied SLIDE63, an interpretable machine learning approach that we recently developed, to the same CM Pre-Tfh and GC-Tfh-SAP populations. SLIDE provides rigorous statistical guarantees regarding inference of necessary and sufficient latent factors (context-specific co-expression gene modules) that discriminate phenotypes (in this case cellular phenotypes in the two states). SLIDE identified 2 significant latent factors and 36 genes (Fig. 3d). Among these genes are known Tfh related genes such as IL-21, PDCD1 and TOX2. The SLIDE model achieved almost perfect discrimination in a rigorous k-fold cross-validation framework with permutation testing (Fig. 3e, AUC = 0.99, p ≤ 0.0001).Fig. 3CoreNet is informative of Tfh biology across single-cell datasets and disease contexts.a Single-cell analysis of Tfh and non-Tfh populations extracted from the Human Tonsil Atlas Multiome67 data set. Created in BioRender. Omelchenko, A. (2026) https://BioRender.com/nkddamk. b SCENIC+ significant regulons separating the CM-PreTfh and GC-Tfh-SAP cell populations (bulk Early Extra-GC Tfh and GC Tfh proxy, respectively). c UMAP visualization of the SCENIC+ regulon CM-PreTfh and GC-Tfh-SAP separation. d Significant SLIDE latent factors separating CM-PreTfh and GC-Tfh-SAP cells. Genes colored by higher expression in CM-PreTfh (blue) or GC-Tfh-SAP (red). e SLIDE performance on cell population separation compared to 20 permuted sets (gray). Mean AUC 0.99 and 0.56, respectively, p < 1e-10. f ssGSEA enrichment of CoreNet-PPI, Vinuesa 2016, and the Hart 2022 gene sets for CM-PreTfh and GC-Tfh-SAP cell populations (Cliff’s Δ = 0.31, −0.21, −0.04; P-value 0.0001, 0.0001, 0.026). g ssGSEA enrichment for single cell separation analysis SCENIC+ and SLIDE gene sets for CM-PreTfh and GC-Tfh-SAP cell populations (Cliff’s Δ = 0.32, 0.98; P-value 0.0001, 0.0001). h Single-cell analysis of Tfh vs. non-Tfh populations with gene set enrichment for modules of interest in scRNA-seq from gut biopsys of people living with HIV (PLWHIV). Created in BioRender. Omelchenko, A. (2026) https://BioRender.com/toqny3o. i UMAP of scRNA clusters from SNN graph construction and Louvain clustering (k.param = 50, resolution = 0.7) on Harmony-reduced space (left); Tfh identification by Tfh Marker Gene Expression Score (right). j Non-Tfh population identified from scRNA clusters by Central Memory Marker Gene Expression Score after filtering. k ssGSEA enrichment for CoreNet-PPI, Vinuesa 2016, and the Hart 2022 gene sets for CM non-Tfh CD4 and Tfh cell populations (Cliff’s Δ = 0.27, −0.1, 0.05; P-value 0.0001, 0.0001, 0.004). l ssGSEA enrichment for single cell separation analysis SCENIC+ and SLIDE gene sets for CM non-Tfh CD4 and Tfh cell populations (Cliff’s Δ = 0.33, 0.98, P-value 0.0001, 0.0001). e–g, k–l Boxplots show the median (center line), 25th and 75th percentiles (box bounds), whiskers extend to 1.5× IQR and p-value is computed using a two-sided Mann–Whitney U test. f–g, k–l Significance given by effect size (Cliff’s Δ) ≥ 0.1 and p-value < 0.05. Gene set sizes: CoreNet-PPI (184), Vinuesa 2016 (42), Hart 2022 (58), SCENIC+ (2230), SLIDE (36). b, d–g, k–l Source data are provided as a Source Data file. CoreNet is informative of Tfh biology across single-cell datasets and disease contexts. a Single-cell analysis of Tfh and non-Tfh populations extracted from the Human Tonsil Atlas Multiome67 data set. Created in BioRender. Omelchenko, A. (2026) https://BioRender.com/nkddamk. b SCENIC+ significant regulons separating the CM-PreTfh and GC-Tfh-SAP cell populations (bulk Early Extra-GC Tfh and GC Tfh proxy, respectively). c UMAP visualization of the SCENIC+ regulon CM-PreTfh and GC-Tfh-SAP separation. d Significant SLIDE latent factors separating CM-PreTfh and GC-Tfh-SAP cells. Genes colored by higher expression in CM-PreTfh (blue) or GC-Tfh-SAP (red). e SLIDE performance on cell population separation compared to 20 permuted sets (gray). Mean AUC 0.99 and 0.56, respectively, p < 1e-10. f ssGSEA enrichment of CoreNet-PPI, Vinuesa 2016, and the Hart 2022 gene sets for CM-PreTfh and GC-Tfh-SAP cell populations (Cliff’s Δ = 0.31, −0.21, −0.04; P-value 0.0001, 0.0001, 0.026). g ssGSEA enrichment for single cell separation analysis SCENIC+ and SLIDE gene sets for CM-PreTfh and GC-Tfh-SAP cell populations (Cliff’s Δ = 0.32, 0.98; P-value 0.0001, 0.0001). h Single-cell analysis of Tfh vs. non-Tfh populations with gene set enrichment for modules of interest in scRNA-seq from gut biopsys of people living with HIV (PLWHIV). Created in BioRender. Omelchenko, A. (2026) https://BioRender.com/toqny3o. i UMAP of scRNA clusters from SNN graph construction and Louvain clustering (k.param = 50, resolution = 0.7) on Harmony-reduced space (left); Tfh identification by Tfh Marker Gene Expression Score (right). j Non-Tfh population identified from scRNA clusters by Central Memory Marker Gene Expression Score after filtering. k ssGSEA enrichment for CoreNet-PPI, Vinuesa 2016, and the Hart 2022 gene sets for CM non-Tfh CD4 and Tfh cell populations (Cliff’s Δ = 0.27, −0.1, 0.05; P-value 0.0001, 0.0001, 0.004). l ssGSEA enrichment for single cell separation analysis SCENIC+ and SLIDE gene sets for CM non-Tfh CD4 and Tfh cell populations (Cliff’s Δ = 0.33, 0.98, P-value 0.0001, 0.0001). e–g, k–l Boxplots show the median (center line), 25th and 75th percentiles (box bounds), whiskers extend to 1.5× IQR and p-value is computed using a two-sided Mann–Whitney U test. f–g, k–l Significance given by effect size (Cliff’s Δ) ≥ 0.1 and p-value < 0.05. Gene set sizes: CoreNet-PPI (184), Vinuesa 2016 (42), Hart 2022 (58), SCENIC+ (2230), SLIDE (36). b, d–g, k–l Source data are provided as a Source Data file. We then implemented single sample gene set enrichment (ssGSEA) analyses to compare methods. ssGSEA is a statistical method designed to measure the composite effect of a predefined gene set at single-cell resolution allowing gene-set activity to be compared across cell populations64. We tested the CoreNet-PPI gene set and benchmarked against two previous reviews summarizing Tfh drivers from a large number of studies (the Vinuesa 201658 and Hart 202259 literature reviews). The signal in each of these sets was evaluated using a combination of a statistical significance and an effect size (Cliff’s Δ, a non-parametric effect size measure used to quantify the difference between two groups) threshold. When using all genes in the gene sets, the CoreNet-PPI gene set was the only one to successfully separate the two cell types when compared to the Tfh reference sets (Fig. 3f, Cliff’s Δ = 0.31, −0.21, −0.04, respectively). Additionally, CoreNet-PPI did equally well as the SCENIC+ analysis with our SLIDE method being the best (Fig. 3g, Cliff’s Δ = 0.32, 0.98, respectively). When the gene sets were filtered for only genes up-regulated in the GC state (Methods), all gene sets saw an increase in performance with the CoreNet-PPI still outperforming the literature curated gene sets and SCENIC+ performing slightly below SLIDE (Supplementary Fig. 4a, 4b, Cliff’s Δ = 0.59, 0.49, 0.52, 0.94, respectively). The superior performance of SLIDE over SCENIC+ with far fewer genes (only 36 as opposed to 1016 GC-activator genes used by SCENIC + ) stems from its unique statistical properties of inference of a necessary and sufficient set of factors. The SLIDE set explains what drives biological phenotypes beyond the use of predictive markers alone while SCENIC+ only uses predictive markers (DEGs). Further, this demonstrates that while we and others have developed methods that can distinguish pairs of cellular states, gene sets/cellular programs underlying such discrimination are inherently different from a gene set or network that explains cellular state progression. We then acquired and processed a scRNA-seq human gut sample dataset for four tissue locations of people living with HIV (PLWHIV)65 (Fig. 3h). The Tfh cell population was identified through expression of Tfh markers and the Azimuth Tonsil reference derived labels (Fig. 3i, Methods). Along with Tfh cells, two non-Tfh populations, Central Memory (CM) non-Tfh CD4+ and Th17, were chosen for comparative analyses (Fig. 3j). When comparing the Tfh and CM non-Tfh CD4+ populations, the CoreNet-PPI (Cliff’s Δ = 0.27) outperformed the Tfh reference sets (Fig. 3k, Cliff’s Δ = −0.1 and 0.05, respectively). CoreNet-PPI also performed similarly to SCENIC+ with SLIDE again achieving the best separation (Fig. 3l, Cliff’s Δ = 0.33, 0.95). The filtering for upregulated genes increased performance of these models as well, with CoreNet-PPI performing equally to the reference sets and SCENIC+ having higher performance but still lower than SLIDE (Supplementary Fig. 4c, 4d, Cliff’s Δ = 0.54, 0.44, 0.57, 0.8, respectively). The evaluation of the Tfh and Th17 populations (Supplementary Fig. 4e) shows comparable effect sizes for all gene sets (Supplementary Fig. 4f). This demonstrates that while curated gene lists can separate the Tfh populations, adding further network context improves performance and interpretability. We further identified the genes that were necessary and sufficient to distinguish distinct Tfh cell states, Tfh1 (CXCR3⁺CCR6⁻), Tfh17 (CXCR3⁻CCR6⁺), and Tfh2 (CXCR3⁻CCR6⁻)66. Owing to limited numbers of Tfh1 and Tfh17 cells, downstream comparisons focused on the Tfh2 population using the single-cell RNAseq Human Tonsil Atlas dataset67. First, Tfh2 cells were compared with CM Pre non-Tfh cells using SLIDE as a control analysis since these populations have no overlap in UMAP space (Supplementary Fig. 5a). Two standalone latent factors were identified (Supplementary Fig. 5b), many of which were enriched for canonical Tfh-associated genes, including CXCR5, ICOS, and PDCD1. These latent factors were sufficient to accurately predict cell identity and performed significantly better than a permuted control (Supplementary Fig. 5c). Notably, latent factor Zsig1 alone was able to robustly stratify the two populations. Consistent with this, the aggregate weights of genes within this factor showed completely non overlapping distributions between the two cell types (Cliff’s Δ = 1; Supplementary Fig. 5d). We next repeated this analysis comparing Tfh2 and GC-Tfh-SAP populations which fully overlap in the UMAP space (Supplementary Fig. 5e). SLIDE identified two latent factors (Supplementary Fig 5f) that effectively separated the two populations relative to a permuted control (Supplementary Fig. 5g). As in the previous comparison, CXCR5 and PDCD1 were more highly weighted on the Tfh2 side and Zsig1 latent factor genes alone enabled strong discrimination between the two cell types (Cliff’s Δ = 0.96; Supplementary Fig. 5h). It is important to note that while SCENIC+ and SLIDE perform well on the separation task they are still lists of genes and do not provide the regulatory or protein network connectivity of the genes of interest. A similar rationale applies to marker genes. They simply reflect predictors or correlates of a cell type or state but don’t provide mechanistic insights into drivers of cellular progression. Thus, to understand the underlying core network, other approaches should be used in combination with these methods. CoreNet is consistent across species While in vitro cultures are necessary for the study of human Tfh differentiation, murine systems are necessary to experimentally interrogate this process in vivo. Current approaches have yielded different results across in-vitro, ex-vivo, and in-vivo systems. This has led to a fundamental question as to whether there are indeed key differences across human and murine Tfh differentiation, or if the observed differences are simply a reflection of artifacts across the different experimental platforms. To address this, we compared our human-derived set of core Tfh modules to a published murine Bcl6 KO reference set (Fig. 4a)31. Intriguingly, CoreNet-DI exhibited a comparable degree of overlap with the mouse reference set as it did with human datasets (Fig. 2e, Fig. 4b). This suggests that there is strong evolutionary conservation between human and murine Tfh differentiation processes. To further test this, we employed a scRNA-seq dataset from a murine model of allergic asthma (where mice were continually exposed to intranasal house dust mite) (Fig. 4a)68. This in vivo model elicits a strong type 2 immune response and induces germinal center formation in the draining lymph nodes, providing a physiologically relevant setting to assess Tfh identity. ssGSEA analyses revealed that CoreNet-PPI and CoreNet-GRN performed as well as the murine Tfh reference in distinguishing Tfh cells from resting T cells within this allergic microenvironment (Fig. 4c). Interestingly, the CoreNet sets were about twice as good at separating these two cell states as the 2016 Tfh reference gene set, further demonstrating that CoreNet captures known and core network modules across species and tissues (Fig. 4d). Together these results suggest that while conventional methods may fail to reconcile differences between various systems, multimodal approaches can enable the identification of gene modules that are truly conserved across species, tissues, and disease contexts. The consistency of these findings across bulk and single-cell transcriptomic datasets supports their robustness and highlights a core gene program that transcends individual experimental platforms.Fig. 4The CoreNet gene program is conserved across species and tissues.a Experimental set up to identify commonalities between the human derived gene set to mouse networks across data modalities and allergy. Created in BioRender. Omelchenko, A. (2026) https://BioRender.com/qc531k6. b CoreNet-PPI, CoreNet-GRN, and CoreNet (PPI + GRN) with direct interactors (DI) overlap with a mouse Tfh reference gene set. 1000 permutations of random size matched sets were also compared for each analysis. P-value = 0.001, 0.002, 0.0009 left to right given by a one-sided Fisher’s exact test. c ssGSEA enrichment for Resting CD4 T-cell and Tfh cell populations for the 2020 murine reference, CoreNet-PPI, CoreNet-GRN (Cliff’s Δ = 0.77, 0.64, 0.73; P = 0.0001 for all). d ssGSEA enrichment for Resting CD4 T-cell and Tfh cell populations for the Vinuesa 2016 reference (Cliff’s Δ = 0.37; P-value = 0.0001). e Visualization of CoreNet-DI across a network derived from a combination of murine Bcl6 KO experiment and Tfh review gene sets. Genes from the 2020, 2016, and 2022 references were pooled and the edges between the genes were provided by the PPI network. Shared reference genes are outlined red while murine reference only or Tfh review only were outlined in blue and orange, respectively. CoreNet-DI was overlaid with this network and direct interactors of reference genes were kept. Shared genes (circle), analysis only discovered interactors (diamond), and reference only genes (square) are colored by mRNA log2FC as previously described. The network was partitioned into 3 sections: well-known common network, less studied network, and reported as different but connected. c–d Boxplots show the median (center line), 25th and 75th percentiles (box bounds), and whiskers extend to 1.5×IQR. Significance given by effect size (Cliff’s Δ) ≥ 0.1 and p-value < 0.05 (*), p > 0.01 (**) computed using a two-sided Mann–Whitney U test. b–e Source data are provided as a Source Data file. The CoreNet gene program is conserved across species and tissues. a Experimental set up to identify commonalities between the human derived gene set to mouse networks across data modalities and allergy. Created in BioRender. Omelchenko, A. (2026) https://BioRender.com/qc531k6. b CoreNet-PPI, CoreNet-GRN, and CoreNet (PPI + GRN) with direct interactors (DI) overlap with a mouse Tfh reference gene set. 1000 permutations of random size matched sets were also compared for each analysis. P-value = 0.001, 0.002, 0.0009 left to right given by a one-sided Fisher’s exact test. c ssGSEA enrichment for Resting CD4 T-cell and Tfh cell populations for the 2020 murine reference, CoreNet-PPI, CoreNet-GRN (Cliff’s Δ = 0.77, 0.64, 0.73; P = 0.0001 for all). d ssGSEA enrichment for Resting CD4 T-cell and Tfh cell populations for the Vinuesa 2016 reference (Cliff’s Δ = 0.37; P-value = 0.0001). e Visualization of CoreNet-DI across a network derived from a combination of murine Bcl6 KO experiment and Tfh review gene sets. Genes from the 2020, 2016, and 2022 references were pooled and the edges between the genes were provided by the PPI network. Shared reference genes are outlined red while murine reference only or Tfh review only were outlined in blue and orange, respectively. CoreNet-DI was overlaid with this network and direct interactors of reference genes were kept. Shared genes (circle), analysis only discovered interactors (diamond), and reference only genes (square) are colored by mRNA log2FC as previously described. The network was partitioned into 3 sections: well-known common network, less studied network, and reported as different but connected. c–d Boxplots show the median (center line), 25th and 75th percentiles (box bounds), and whiskers extend to 1.5×IQR. Significance given by effect size (Cliff’s Δ) ≥ 0.1 and p-value < 0.05 (*), p > 0.01 (**) computed using a two-sided Mann–Whitney U test. b–e Source data are provided as a Source Data file. Finally, to visualize the broader architecture of the Tfh cell differentiation landscape, we integrated the two Tfh review references and murine Bcl6 KO gene sets onto the PPI network. Genes present in the Tfh review reference were outlined in orange, those in the Bcl6 KO mouse experiment reference in blue, and those shared between both in red. We then overlaid our CoreNet gene set along with any direct interactors of the reference genes included in the network. This integrated network naturally resolved into three conceptual modules (Fig. 4e). The first is a well-characterized core set of canonical Tfh regulators (including TFs, ligands and receptors) such as PDCD1, STAT1, CD40LG, BCL6, CD200, IL-21, STAT4, and PRDM1. The second is a less-explored region composed of interactors, many linked to transcription factors such as STAT3 and JUN, that have not been highlighted previously as Tfh-associated. Lastly, we identified a “reported as different between species but connected” region which is composed of genes identified in only the reviews or the KO experiment but shown by our human network analysis to be physically and functionally connected in the integrated network. Notably, many murine only nodes (blue) are identified by our human derived network and are interconnected to other identified nodes. These findings underscore the value of a network-based approach in unifying seemingly divergent species-specific results and highlighting evolutionarily conserved components of the Tfh regulatory circuit. Network based approaches capture Tfh modulation in single cell spatial datasets Spatial transcriptomics enables linking molecular function to localized tissue organization and provides context to networks driving cellular progression. We hypothesized that our bulk-derived CoreNet gene set would recapitulate tissue context information from spatial transcriptomic experiments across human and mouse datasets. Moreover, the CoreNet gene set should be enriched with drivers identified by an orthogonal spatially resolved method. A healthy human tonsil spatial transcriptomics (Slide-tag)69 dataset was acquired. As earlier, we defined Tfh and non-Tfh CD4 T-cells using marker genes. CoreNet-PPI and the reference literature-curated gene sets (Vinuesa 201658 and Hart 202259) were then evaluated in terms of their ability to stratify the Tfh and non-Tfh CD4s (Fig. 5a). All three sets provided excellent stratification in this case (Fig. 5b). Further, we also visualized the spatial location of the cells of interest colored by their cell type (Fig. 5c). All gene sets captured relevant spatial microniches of the Tfh cells based on their corresponding enrichment scores (Fig. 5d). The ability of the bulk-derived gene sets to distinguish the Tfh population from other cell types in a single cell spatial context equally or better compared to curated gene sets shows the power of these network approaches. Hence, high-quality bulk datasets in tandem with network approaches can enable capturing complex regulatory processes in the immune system.Fig. 5Spatial and in silico validation of key Tfh modulators identified in CoreNet.a Analysis of the human tonsil spatial dataset. Created in BioRender. Omelchenko, A. (2026) https://BioRender.com/08cbw55. b ssGSEA for CoreNet-PPI, Vinuesa 2016, and Hart 2022 gene sets for non-Tfh CD4+ and Tfh cell populations (Cliff’s Δ = 0.53, 0.41, 0.59, p = 0.0001 for all). Significance given by effect size (Cliff’s Δ) ≥ 0.1 and p < 0.05 computed using a two-sided Mann–Whitney U test. Boxplots show the median (center line), 25th and 75th percentiles (box bounds), and whiskers extend to 1.5×IQR. c 2D spatial coordinates of non-Tfh CD4+ and Tfh cells colored by cell type. d 2D spatial coordinates of non-Tfh CD4+ and Tfh cells colored by enrichment score for the CoreNet-PPI, Vinuesa 2016, and Hart 2022 gene sets. e SpaceTravLR analysis of murine VisiumHD spatial datasets. Created in BioRender. Omelchenko, A. (2026) https://BioRender.com/ohf3mer. f Day 5 Mouse lymph node (LN) SlideSeq spatial data with identified cell types. g Top Modulators promoting Tfh activity using a scaled impact score. h Enrichment of top 30 Tfh modulators in the CoreNet, SCENIC + , and Dalit 2025 gene sets. i Propagation of Bcl6 modulator effect on Il-21 in the SpaceTravLR network. Bcl6 KO yields a significant decrease in Il-21 (log2FC = −0.076, p < 0.001). j Propagation of Nr3c1 modulator effect on Il-21 in the SpaceTravLR network. Nr3c1 KO yields a significant decrease in Il-21 (log2FC = −0.35, p-value < 0.0001). k Day 3 Mouse LN VisiumHD spatial data along with identified cell types. l Top Modulators promoting Tfh activity given by a scaled impact score. m Enrichment of top 30 Tfh modulators in the CoreNet, SCENIC + , and Dalit 2025 gene sets. n Top Ligand-Receptor interactions promoting Tfh activity using a scaled impact score from the pre-trained day 3 Th2 murine viziumHD model. CoreNet-DI pairs shown in orange. o SpaceTravLR identified TGFB1-TGFBR2 ligand-receptor pair (purple) connecting significant modules from CoreNet-GRN (orange) and CoreNet-PPI (green) and shared (red) with 2 degrees of identified interactors. g–j shows the day 5 Tfh murine SlideSeq LN model. l–m shows the day 3 Tfh murine VisiumHD LN model. g–j, l–n Tfh cells are defined by Cxcr5, Bcl6, and Pdcd1 genes. g, l Color denotes CoreNet-only (dark blue), SCENIC + -only (yellow), and shared (green) genes and nodes i–j. h, m Hypergeometric test given p-value. b–h, l–o Source data are provided as a Source Data file. Spatial and in silico validation of key Tfh modulators identified in CoreNet. a Analysis of the human tonsil spatial dataset. Created in BioRender. Omelchenko, A. (2026) https://BioRender.com/08cbw55. b ssGSEA for CoreNet-PPI, Vinuesa 2016, and Hart 2022 gene sets for non-Tfh CD4+ and Tfh cell populations (Cliff’s Δ = 0.53, 0.41, 0.59, p = 0.0001 for all). Significance given by effect size (Cliff’s Δ) ≥ 0.1 and p < 0.05 computed using a two-sided Mann–Whitney U test. Boxplots show the median (center line), 25th and 75th percentiles (box bounds), and whiskers extend to 1.5×IQR. c 2D spatial coordinates of non-Tfh CD4+ and Tfh cells colored by cell type. d 2D spatial coordinates of non-Tfh CD4+ and Tfh cells colored by enrichment score for the CoreNet-PPI, Vinuesa 2016, and Hart 2022 gene sets. e SpaceTravLR analysis of murine VisiumHD spatial datasets. Created in BioRender. Omelchenko, A. (2026) https://BioRender.com/ohf3mer. f Day 5 Mouse lymph node (LN) SlideSeq spatial data with identified cell types. g Top Modulators promoting Tfh activity using a scaled impact score. h Enrichment of top 30 Tfh modulators in the CoreNet, SCENIC + , and Dalit 2025 gene sets. i Propagation of Bcl6 modulator effect on Il-21 in the SpaceTravLR network. Bcl6 KO yields a significant decrease in Il-21 (log2FC = −0.076, p < 0.001). j Propagation of Nr3c1 modulator effect on Il-21 in the SpaceTravLR network. Nr3c1 KO yields a significant decrease in Il-21 (log2FC = −0.35, p-value < 0.0001). k Day 3 Mouse LN VisiumHD spatial data along with identified cell types. l Top Modulators promoting Tfh activity given by a scaled impact score. m Enrichment of top 30 Tfh modulators in the CoreNet, SCENIC + , and Dalit 2025 gene sets. n Top Ligand-Receptor interactions promoting Tfh activity using a scaled impact score from the pre-trained day 3 Th2 murine viziumHD model. CoreNet-DI pairs shown in orange. o SpaceTravLR identified TGFB1-TGFBR2 ligand-receptor pair (purple) connecting significant modules from CoreNet-GRN (orange) and CoreNet-PPI (green) and shared (red) with 2 degrees of identified interactors. g–j shows the day 5 Tfh murine SlideSeq LN model. l–m shows the day 3 Tfh murine VisiumHD LN model. g–j, l–n Tfh cells are defined by Cxcr5, Bcl6, and Pdcd1 genes. g, l Color denotes CoreNet-only (dark blue), SCENIC + -only (yellow), and shared (green) genes and nodes i–j. h, m Hypergeometric test given p-value. b–h, l–o Source data are provided as a Source Data file. We compared CoreNet to the top modulators in two mouse lymph node datasets from a model of allergic asthma70. The modulators were selected by a functional microniche identification method, SpaceTravLR71 (Fig. 5e). SpaceTravLR combines deep learning with statistical machine learning to discover spatially resolved regulatory and signaling circuits underlying microniches linked to functional outcomes. Initially, we used a Day 5 mouse lymph node profiled using Slideseq. There was a large number of bona fide Tfh (and GC Tfh) due to continued exposure to the allergen house dust mite (HDM) (Fig. 5f). After training SpaceTravLR, we used it for in-silico functional screening to identify the top gene modulators of Tfh identity (where Tfh cells were defined by Bcl6, Cxcr5, and Pdcd1). The top modulators identified by SpaceTravLR aligned highly with CoreNet (Fig. 5g ) and did not change significantly when adding Il-21 as a marker gene (Supplementary Fig. 6a). When compared to SCENIC+ and the Tfh signature from a recent publication35, the Dalit 2025 gene set, CoreNet was significantly more highly enriched (Fig. 5h, Supplementary Fig. 6b). Additionally, using SpaceTravLR we performed in-silico knockouts of a common marker, Bcl6, and an uncommon marker, Nr3c1, to examine their effects on Il-21 (Fig. 5i, 5j). SpaceTravLR leverages network propagation to model how genetic perturbations spread through signaling networks and spatial cellular neighborhoods. While it is the first method, to our knowledge, to combine spatial transcriptomics with network propagation, the underlying principle is straightforward but powerful. For instance, a Bcl6 knockout may affect Il-21 either directly or indirectly through downstream target genes, allowing SpaceTravLR to quantify the total propagated effect of Bcl6 on Il-21 expression along all possible paths. By accumulating these cascading effects along all possible paths, we can quantify the total influence of Bcl6 on Il-21 expression (Fig. 5i). Further, SpaceTravLR’s use of group lasso regularization helps prune spurious gene-gene edges, increasing confidence that the identified propagation paths are biologically meaningful. This orthogonal approach demonstrates that the knockout of Bcl6 leads to an overall negative effect on (i.e., decrease of) Il-21 (logFC −0.076, p < 0.001) consistent with prior literature demonstrating a feedback loop between Bcl6 and Il-2150,54,72,73 that maintains Tfh identity. We repeated the knockout experiment with an interesting hit, Nr3c1, which also led to a decrease in Il-21 (log2FC −0.350, p-value < 0.0001, Fig. 5j). Notably, the SpaceTravLR mouse LN single cell spatial networks show a high amount of overlap with our human bulk-derived network genes (Fig. 5i, 5j). Further, very few genes overlap with the SCENIC+ analysis in both knock out analyses. This demonstrates that while SCENIC+ is good at separating two states, it generates large unspecific regulons which fail to provide the connectivity and context necessary for mechanistic insight and understanding the underlying biology. We then analyzed a second lymph node dataset70 from the same context which was harvested at Day 3 (Fig. 5k, Methods). This set contains less Tfh cells (earlier timepoint, fewer Tfh and GC Tfh formed), but it is profiled using VisiumHD providing a higher resolution. Similarly to the day 5 analysis, the top modulators identified by SpaceTravLR in the day 3 dataset significantly overlapped with CoreNet (Fig. 5l, Supplementary Fig. 6C). CoreNet showed a higher fold enrichment for these modulators than SCENIC+ and the Dalit 2025 gene set, the latter of which was not significantly enriched (Fig. 5m, Supplementary Fig. 6d). This clearly demonstrates the power of our CoreNet at capturing biologically meaningful relationships that can help quantify functional alterations to Tfh cells upon perturbation. On the other hand, approaches relying on pairwise comparisons (SCENIC + ) or bag-of-gene approaches fail partly or completely to capture these relationships. We also used a model trained on the full Day 3 VisiumHD mouse lymph node from the original SpaceTravLR manuscript, which included only 2116 total genes, 67 CoreNet, and 535 CoreNet-DI genes. The higher resolution enabled us to identify the top ligand-receptor interactions promoting Tfh identity. We utilized the full Th2 model from the SpaceTravLR paper to identify Tfh fate specific ligand-receptor pairs. This analysis also overlapped highly with CoreNet-DI (Fig. 5n). The highest scoring ligand-receptor pair was Tgfb1-Tgfbr2, both of which are interactors of our CoreNet gene set. The ligand receptor pair connects two modules from CoreNet-PPI as well as some modules from CoreNet-GRN analysis (Fig. 5o). These modules have been previously characterized and have critical roles in Tfh differentiation20,74,75. Further, all methods (CoreNet, SCENIC + , SLIDE, and SpaceTravLR) identified NR3C1 as a gene of interest which has yet to be characterized in the context of Tfh differentiation. Importantly, we show that independent and orthogonal network approaches converge on core connected Tfh modules which provide insight into potential targets. Marker genes, DEGs or similar correlates are inadequate at capturing the functional relationships captured by CoreNet as they simply reflect predictors or correlates of a cell type or state. They fail to provide mechanistic insight into drivers of cellular progression that are offered by CoreNet. Mitigation of false negatives Due to a stringent false positive control and missing links in the PPI network, a few known Tfh markers were not incorporated into CoreNet (Supplementary Fig. 3a, 3b). Importantly, these missing markers (CXCR5, ICOS and IL-21) do not imply a contradiction with existing literature. Rather, because we employ stringent false positive control, we have a small number of false negatives. To mitigate this, we leveraged network propagation to recover biologically important genes that did not meet the initial stringent threshold to reduce the false negative rate (Supplementary Fig. 7a). Two complementary approaches were used (Methods). In the first, genes were ranked by their peak proximity scores. Genes that scored highly (top 5%) but lacked sufficient connectivity to be included in the high-scoring modules were selected. This allowed for the recovery of candidate genes whose network edges may be absent due to the gaps in the PPI data. In the second, we traced the direct transcriptional targets of the driver transcription factors across the GC Tfh GRNs. For each epigenomic dataset we generated, we ranked candidates by a composite score integrating network-derived edge weights and TF importance. We selected the top 200 scoring genes per set to represent the genes downstream of important regulatory nodes. The union of these recovered genes with CoreNet defines CoreNet-Extended, a set of approximately 1400 genes. Key Tfh markers, such as CXCR5, ICOS, and IL-21 were recovered in the CoreNet-Extended gene set (Supplementary Fig. 7b). Further, CoreNet-Extended demonstrated a higher overlap than CoreNet on reference datasets58,59 (Supplementary Fig. 2d, 7c). However, this broader coverage comes with important trade-offs. CoreNet-Extended is substantially larger than CoreNet and consequently less enriched for Tfh-relevant genes. It also does not improve the identification of top modulators from SpaceTravLR, an orthogonal network-based approach (Supplementary Fig. 7d, 7e). Critically, it incorporates genes that are disconnected or fragmented from the network (Supplementary Fig. 7b), which dilutes mechanistic interpretability and makes the set closer to an undifferentiated gene list. CoreNet-DI, which has the largest overlap with the reference sets (Fig. 2e, Supplementary Fig. 2c) also retains the important network connectivity. However, it is a larger set than CoreNet-Extended. Both CoreNet-DI and CoreNet-Extended represent distinct post-processing strategies that supplement CoreNet, and the appropriate choice depends on the biological question at hand. Where breadth and recapitulation of known markers are prioritized, these extended sets offer clear utility. However, for the discovery analyses in this work, we rely on the tightly interconnected CoreNet and CoreNet-DI, whose functional coherence and network grounding make them best suited for identifying and prioritizing previously uncharacterized Tfh state progression drivers. Proteomics analysis reveals common and uncommon markers Validating network-derived markers at the protein level is critical, as cell-cell communication and migratory behavior are ultimately governed by surface protein abundance. Since CoreNet was derived from bulk RNA-seq and chromatin accessibility data, we sought to assess whether the network-identified markers are also reflected at the protein level. To this end, we performed two independent proteomic validations using tonsillar datasets (Supplementary Fig. 8a). First, we leveraged the human Tonsil Atlas CITE-seq dataset67,76, which provides simultaneous measurement of gene expression and surface protein abundance via Antibody Derived Tags (ADTs). CITE-seq panels are inherently constrained by the number and selection of antibodies included, meaning that only a subset of candidate markers can be evaluated in any given dataset. Despite this limitation, overlap between the CoreNet gene set and the available ADT panel identified 7 common markers of which all were expressed in the Tfh population. Azimuth was used to define GC Tfh and non-Tfh CD4 + T cell populations prior to comparative analysis. The established Tfh markers PDCD1 and BTLA were significantly enriched in the GC Tfh population, validating the approach (Supplementary Fig. 8b). Notably, CCR3 also emerged as significantly elevated in GC Tfh cells relative to non-Tfh CD4 + T cells (Supplementary Fig. 8b, 8c, p < 0.0001). This finding was not anticipated from transcriptomic data alone, given the absence of detectable CCR3 transcript across RNA expression datasets and highlights the value of proteomic validation for uncovering surface markers that operate below the threshold of RNA detection. CCR3, a canonical marker of Th2-polarized CD4 + T cells77, arose as an unexpected and potentially important feature of GC Tfh in our analysis. While CCR3 transcripts were undetectable across RNA expression datasets, it was significantly enriched at the protein level in GC Tfh cells relative to non-Tfh CD4 + T cells in the Tonsil Atlas (Supplementary Fig. 8b), and was independently recovered as a high-confidence Tfh-associated node following network propagation of epigenomic data (Fig. 2c). The concordance between proteomic enrichment and epigenomic accessibility, in the absence of detectable transcription, suggests that CCR3 may be subject to post-transcriptional regulation in GC Tfh cells. Supporting the existence of a CCR3-expressing Tfh population in vivo, CCR3 + CXCR5+ circulating Tfh cells have been identified as a distinct peripheral subset in humans, one that is selectively depleted in severe COVID-1978. This suggests that the population carries functional significance in the context of disease. Together, these findings position CCR3 as a candidate marker of a GC Tfh subpopulation. To extend this analysis with greater marker coverage, we performed a second independent validation using the spatial CITE-seq dataset from scProAtlas79, which utilizes the Human Cell Atlas76 tonsil data. Tfh cells were identified using a multi-criteria scoring strategy (Methods) and were compared against non-Tfh CD4 + T cells in downstream differential protein abundance analyses. The validity of this classification was supported by identifying negative controls not in the marker gene set. For example, a repressor of Tfh should be more highly expressed in non-Tfh CD4 cells and a random gene should have no difference in protein abundance. The analysis demonstrated the expected behavior of the negative controls. Genes with broad expression like SLC23A2 and S100A10 showed equal protein abundance (Supplementary Fig. 8d). Further, known Tfh repressors such as SATB180 and TGFB181 were significantly higher in the non-Tfh CD4+ population, collectively lending confidence to the specificity of the approach (Supplementary Fig. 8d). There was significant enrichment of multiple well-established Tfh markers in the identified Tfh population, including PDCD1, STAT3, BTLA, and TCF7, relative to non-Tfh CD4 + T cells providing a strong reference point (Supplementary Fig. 8e). Beyond known markers, two CoreNet-derived candidates without established roles in GC Tfh cells showed significantly higher protein abundance in the Tfh population. Each marker was held out from the classification to validate that it is more highly abundant in the Tfh population. NR3C1, which has now been independently recovered across six orthogonal analyses in this study, was again identified here, further reinforcing its candidacy as a functionally relevant Tfh factor. Additionally, as previously mentioned, POU2F2, has been described in GC B cells50,51, but has not been well characterized in GC Tfh cells52,53. Its protein enrichment here further suggests a potential role that warrants investigation. Taken together, the two independent proteomic validations presented here demonstrate that CoreNet derived markers are not only transcriptionally and epigenomically relevant but are also reflected at the protein level in GC Tfh cells. The consistent enrichment of known Tfh markers across both datasets, alongside the expected behavior of negative controls, lends strong confidence to the approach. Beyond validation of known biology, these analyses surface two candidates, NR3C1 and POU2F2, that have not been characterized in GC Tfh cells and whose repeated identification across orthogonal modalities argues against spurious discovery. The emergence of CCR3 as a proteomic hit in the absence of detectable transcript further underscores that transcriptomic atlases alone are insufficient to fully resolve the molecular identity of GC Tfh cells. Collectively, these findings establish CoreNet as a tool capable of nominating biologically meaningful candidates and motivate functional investigation of these markers in the context of GC Tfh biology. Identified modules correspond to specific immune signaling pathways Next, we evaluated how our bulk-derived CoreNet genes related to known biological pathways82,83. This provides another orthogonal way to contextualize our findings as the modules are discovered based on the structures of the underlying gene regulatory and protein interaction networks, but do not consider known pathways or gene sets. We began by assessing overlaps of the modules with the Hallmark gene sets that represent well-defined biological states and processes84. To determine significance of overlaps, the observed intersections to those expected from size-matched random sets of network proteins were compared. This provides a rigorous null model (baseline) for expected overlap by chance. The results were visualized using a volcano plot. Each point represents a specific pathway, with the effect size (log odds ratio) plotted on the X-axis and significance (–log10Q) on the Y-axis, reflecting both the strength and statistical support of each enrichment (Methods). Three of 50 pathways passed the thresholds (-log10Q > 0.7 and log odds ratio > 0.58) - IL-6/JAK/STAT3 signaling, MYC signaling and TNFα signaling via NF-κB (Fig. 6a).Fig. 6Identification of pathways enriched in CoreNet.a Volcano plot visualizing effect sizes and significance of functional overlaps between the CoreNet-DI set and Hallmark pathways. Significant pathways colored in green (-log(Q-value) > 0.7, log2(odds_ratio) > 0.58). Adjusted p-value (FDR) < 0.2, and 1.5-fold enrichment. b Human scRNA top 15 Hallmark pathways given by the top effect size between the Tfh and Th17 cell population enrichment scores. Pathways that were also identified in the volcano plot are highlighted in green. c ssGSEA enrichment of the TNFA-SIGNALING-VIA-NFKB hallmark pathway for Th17 and Tfh cell populations (Cliff’s Δ = −0.56, p-value < 0.01). d Volcano plot visualizing effect sizes and significance of functional overlaps between the CoreNet-DI set and PID pathways. Significant pathways colored in green (-log(Q-value) > 0.7, log2(odds_ratio) > 0.58). Adjusted p-value (FDR) < 0.2, and 1.5-fold enrichment. e Human ssGSEA top 15 PID pathways given by the top effect size between the Tfh and Th17 cell population enrichment scores. Pathways that were also identified in the volcano plot are highlighted in green. f ssGSEA enrichment of the IL-27 PID pathway for Th17 and Tfh cell populations (Cliff’s Δ = −0.40, p-value < 0.01). g PID IL-27 pathway PPI network visualization with the CoreNet-DI gene set overlaid. Shared genes (circle), analysis only discovered interactors (diamond), and reference only genes (square) are colored by mRNA log2FC as previously described. h PID IL-23 pathway PPI network visualization with the CoreNet-DI gene set overlaid. Shared genes (circle), analysis only discovered interactors (diamond), and reference only genes (square) are colored by mRNA log2FC as previously described. c, f p-value computed using a two-sided Mann–Whitney U test. Boxplots show the median (center line), 25th and 75th percentiles (box bounds), and whiskers extend to 1.5× IQR. a–h Source data are provided as a Source Data file. Identification of pathways enriched in CoreNet. a Volcano plot visualizing effect sizes and significance of functional overlaps between the CoreNet-DI set and Hallmark pathways. Significant pathways colored in green (-log(Q-value) > 0.7, log2(odds_ratio) > 0.58). Adjusted p-value (FDR) < 0.2, and 1.5-fold enrichment. b Human scRNA top 15 Hallmark pathways given by the top effect size between the Tfh and Th17 cell population enrichment scores. Pathways that were also identified in the volcano plot are highlighted in green. c ssGSEA enrichment of the TNFA-SIGNALING-VIA-NFKB hallmark pathway for Th17 and Tfh cell populations (Cliff’s Δ = −0.56, p-value < 0.01). d Volcano plot visualizing effect sizes and significance of functional overlaps between the CoreNet-DI set and PID pathways. Significant pathways colored in green (-log(Q-value) > 0.7, log2(odds_ratio) > 0.58). Adjusted p-value (FDR) < 0.2, and 1.5-fold enrichment. e Human ssGSEA top 15 PID pathways given by the top effect size between the Tfh and Th17 cell population enrichment scores. Pathways that were also identified in the volcano plot are highlighted in green. f ssGSEA enrichment of the IL-27 PID pathway for Th17 and Tfh cell populations (Cliff’s Δ = −0.40, p-value < 0.01). g PID IL-27 pathway PPI network visualization with the CoreNet-DI gene set overlaid. Shared genes (circle), analysis only discovered interactors (diamond), and reference only genes (square) are colored by mRNA log2FC as previously described. h PID IL-23 pathway PPI network visualization with the CoreNet-DI gene set overlaid. Shared genes (circle), analysis only discovered interactors (diamond), and reference only genes (square) are colored by mRNA log2FC as previously described. c, f p-value computed using a two-sided Mann–Whitney U test. Boxplots show the median (center line), 25th and 75th percentiles (box bounds), and whiskers extend to 1.5× IQR. a–h Source data are provided as a Source Data file. To further validate our findings, we used the human PLWHIV scRNA-seq dataset to identify Hallmark pathway differences between the Tfh and Th17 cell populations. We used the top Cliff’s Δ to identify the 15 most different pathways across these cell types (Fig. 6b). IL-6/JAK/STAT3 and TNFα signaling via NF-κB also ranked among the top 15 pathways. Interestingly, recent findings have linked aberrant extra-follicular TNFα accumulation in COVID-19 to loss of Tfh cells and germinal centers85. We also note the decrease of TNFα in PLWHIV Tfh populations compared to Th17 populations (Fig. 6c). Finally, we evaluated overlaps of the Tfh modules with the PID pathways, which represent human signaling and regulatory processes86. The bulk analysis identified a few pathways including SMAD2, which is a well-known signaling pathway in Tfh differentiation20,21 (Fig. 6d). Additionally, we see enrichment for the IL-27 and IL-23 pathways in both the bulk and single cell analysis, which have been less studied in the context of Tfh (Fig. 6d–e). IL-12, IL-23 and IL-27 belong to the same heterodimeric family87 and there has been some evidence that they regulate Tfh differentiation22,88,89. Since IL-23 is well known to enhance Th17 differentiation, it is expected that the IL-23 pathway enrichment would be relatively lower in Tfh cells than in Th17 cells (Fig. 6e). Further, both IL-12 and IL-27 are shown to be relatively less enriched in Tfh cells when compared to Th17 (Fig. 6e–f). We visualized our prioritized gene set and direct interactors with the IL-27 (Fig. 6g) and IL-23 (Fig. 6h) PID pathways. As shown, there is a lot of overlap between the genes in these two pathways including STAT1, STAT3, and IL-12 related genes. We also observed higher expression levels of genes encoding this cytokine family in the Early Extra-GC Tfh state. IL-12 fine tunes Tfh differentiation We noted that the IL-12 family cytokines were consistently identified throughout our unbiased analyses as top hits. These were among the top PID pathways from the scRNA pathway enrichment analysis (Fig. 7a, Fig. 6d–f) and were predicted to be important regulators when comparing early Extra-GC Tfh and GC Tfh (Fig. 2c). It has been previously reported that IL-12 signaling through STAT4 promotes Tfh differentiation23 whereas other studies suggest that IL-12 and STAT4 can inhibit Tfh differentiation27,90. While the analysis indicates the IL-12 pathway is highly important, expression of key signaling proteins STAT4 and IL-12RB1 vary across states, with expression downregulated in GC Tfh as compared to Extra-GC Tfh (Figs. 7b, 2c). The expression of the IL-12 receptor subunits, IL-12RB1 and IL-12RB2, are highest in Early Extra-GC Tfh Late Extra-GC Tfh (Supplementary Fig. 9 A). SpaceTravLR analysis shows an overall positive effect on (i.e., increase of) Il-12rb1 following Stat4 knock out (log2FC = 0.048, p < 0.0001), suggesting that high STAT4 signaling may repress expression of the IL-12 receptor. Several in vitro studies of human Tfh differentiation suggest that IL-12 enhances Tfh differentiation, but these Tfh are in fact CXCR5low PD-1low and therefore may be more representative of an early Extra-GC Tfh stage. Combining these observations, we hypothesized that IL-12 may play a stage-specific role in Tfh differentiation, enhancing early development towards the Tfh fate but suppressing full differentiation to GC Tfh.Fig. 7IL-12 pathway blockade promotes the GC-Tfh fate.a ssGSEA enrichment of the IL-12 PID pathway for Th17 and Tfh cell populations. Cliff’s Δ = 0.56, p < 0.0001 computed using a two-sided Mann–Whitney U test. Boxplots show the median (center line), 25th and 75th percentiles (box bounds), and whiskers extend to 1.5×IQR. b. PID IL-12 pathway PPI network visualization with the CoreNet-DI gene set overlaid. Red dashed line indicating interaction identified in submodule 1 (Fig. 2c). Shared genes (circle), analysis only discovered interactors (diamond), and reference only genes (square) are colored by mRNA log2FC as previously described. c–f B18 + /- mice were given either heat killed (HK) STm control or STm infection and immunized with NP-CGG in alum on day 0. Mice were additionally treated with neutralizing anti-IL-12 antibodies (or vehicle control) beginning at day −1 or 3 (e), or day 7 or 11 (f). Tfh were analyzed at 12 days after infection/immunization. C. Experimental design for panels d and e. d Representative FACS plots pre-gated on [singlet, live, CD45R-, CD4 + , FoxP3-, CD44hi] activated T conventional cells. Shown are gates for total Tfh (CXCR5high PD-1high), and GC Tfh (CD90low) or non-GC Tfh (CD90high). e Numbers of GC Tfh and non-GC Tfh per spleen. Data are combined from two experiments, n = 8 Ctrl, n = 12 Inf no block, n = 14 Inf day −1 block, n = 15 Inf day 3 block. f Numbers of GC and non-GC Tfh were quantified 12 days after infection as in d. Data are combined from two experiments, n = 8 Ctrl, n = 10 Inf no block, n = 9 Inf day 7 block, n = 9 Inf day 11 block. For e–f charts, data were combined from two individual experiments each, circles represent individual mice and bars the mean +/− SD. NS, not significant, *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. P values were calculated using two-tailed t tests. a–b, e–f Source data are provided as a Source Data file. IL-12 pathway blockade promotes the GC-Tfh fate. a ssGSEA enrichment of the IL-12 PID pathway for Th17 and Tfh cell populations. Cliff’s Δ = 0.56, p < 0.0001 computed using a two-sided Mann–Whitney U test. Boxplots show the median (center line), 25th and 75th percentiles (box bounds), and whiskers extend to 1.5×IQR. b. PID IL-12 pathway PPI network visualization with the CoreNet-DI gene set overlaid. Red dashed line indicating interaction identified in submodule 1 (Fig. 2c). Shared genes (circle), analysis only discovered interactors (diamond), and reference only genes (square) are colored by mRNA log2FC as previously described. c–f B18 + /- mice were given either heat killed (HK) STm control or STm infection and immunized with NP-CGG in alum on day 0. Mice were additionally treated with neutralizing anti-IL-12 antibodies (or vehicle control) beginning at day −1 or 3 (e), or day 7 or 11 (f). Tfh were analyzed at 12 days after infection/immunization. C. Experimental design for panels d and e. d Representative FACS plots pre-gated on [singlet, live, CD45R-, CD4 + , FoxP3-, CD44hi] activated T conventional cells. Shown are gates for total Tfh (CXCR5high PD-1high), and GC Tfh (CD90low) or non-GC Tfh (CD90high). e Numbers of GC Tfh and non-GC Tfh per spleen. Data are combined from two experiments, n = 8 Ctrl, n = 12 Inf no block, n = 14 Inf day −1 block, n = 15 Inf day 3 block. f Numbers of GC and non-GC Tfh were quantified 12 days after infection as in d. Data are combined from two experiments, n = 8 Ctrl, n = 10 Inf no block, n = 9 Inf day 7 block, n = 9 Inf day 11 block. For e–f charts, data were combined from two individual experiments each, circles represent individual mice and bars the mean +/− SD. NS, not significant, *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. P values were calculated using two-tailed t tests. a–b, e–f Source data are provided as a Source Data file. To test this experimentally, we turned to a murine model of infection with the bacteria Salmonella enterica serovar Typhimurium (STm), which elicits a strong IL-12 response that suppresses GCs and Tfh formation27,91. In this system, the IL-12 suppression mechanism is dominant, meaning that it also suppresses the GC response to a concurrently administered but unrelated antigen, which we utilize to track the effect of STm infection on the GC response. In this case, we chose nitrophenol haptenated chicken gamma globulin (NP-CGG) for the unrelated antigen. In this system, only infection with live STm, not heat killed (HK) bacteria, is able to suppress GC responses27,91. To examine the stage-specific effects of IL-12 on Tfh differentiation, we administered IL-12-neutralizing antibodies beginning one day prior to, or 3 days after, STm/NP-CGG immunization in this experimental scheme (Fig. 7d–f, orange). Blocking IL-12 starting prior to infection ensures T cells neither receive IL-12 at priming nor throughout GC formation. At day 3 T cells have been primed and pre-Tfh have moved to the T/B border to interact with pre-GC B cells, but the GC structure has yet to form and GC Tfh differentiation has yet to occur36,92,93. We then examined the effect of either early or day 3-initiated IL-12 blocking on total Tfh (CXCR5high PD-1high), and GC Tfh or non-GC Tfh, which are CD90low or CD90high, respectively (Fig. 7d, Supplementary Fig. 9B)36 at 12 days after infection/immunization. Consistent with prior publications27, total Tfh were suppressed; here we additionally and critically show that this effect is primarily driven by the selective suppression of GC Tfh (Supplementary Fig. 9c, d). Indeed, compared to HK STm control, STm infection strongly suppressed the formation of GC Tfh, but minimally affected non-GC Tfh (Fig. 7e, Supplementary Fig. 9e, f). GC B cells were also increased (Supplementary Fig. 9g, h), recapitulating genetic studies. Blocking IL-12 at either day −1 or day 3 fully restored GC Tfh quantity (Fig. 7e). Blocking IL-12 also enhanced non-GC Tfh, even though STm infection did not suppress non-GC Tfh, potentially suggesting further heterogeneity among non-GC Tfh or additional roles of IL-12. Given that the effect of blocking IL-12 at day −1 or day 3 was similar, and that both of these times are prior to the formation of bona fide GC T and B cell populations, we next asked whether blocking IL-12 after GCs have developed can still rescue GC Tfh formation. For this, we administered blocking antibodies once at day 7 or 11 after infection, both times that we have previously shown Tfh are already suppressed27. Strikingly, blocking at these later times did not rescue GC Tfh formation, indicating that suppression of GC Tfh is not reversible. However, blocking at day 7 did somewhat enhance the non-GC Tfh population (Fig. 7f), indicating that there is an early window that extends to day 7 in which IL-12 also inhibits non-GC Tfh development, also consistent with a similar effect in the earlier blocking window experiments (Fig. 7e). Collectively, our experimental analysis demonstrates that the effects of IL-12 differ with respect to time and stage of Tfh differentiation, confirming and validating a key aspect of our systems analysis. We also confirm that the IL-12-family member implicated in suppression of Tfh development is IL-12 itself. Discussion Our systems-level approach provides a different perspective on both established and previously unrecognized mechanisms underlying Tfh cell differentiation. A high-quality, well-designed bulk dataset forms the foundation of our analysis, enabling the identification of meaningful biology. We refine these findings using scRNA-seq and spatial datasets. Critically, we validate key predictions in an in vivo murine model, demonstrating physiological relevance. While traditional differential expression analysis has been the standard for interpreting RNA-seq datasets, it often fails to capture complex regulatory relationships. In contrast, network-based approaches incorporate the topology of biological interactions to reveal hidden structure in the data, identifying functionally important genes that may not exhibit strong differential expression on their own. Using a PPI network to incorporate higher-order relationships, additive effects, and functional dependencies among the genes allows us to capture a stronger biological signal. Genes involved in Tfh cell progression are more likely to encode interacting proteins that form coherent network submodules. In addition, network propagation serves to filter out genes that may exhibit similar epigenetic signatures but are unlikely to play a functional role in Tfh biology. Crucially, this highlights the power of combining several data types through network analysis. Even if the gene encoding the TF is not highly expressed, it can be identified as a critical regulator of the Tfh program. While we sought to identify key regulators of Tfh differentiation, it is notable that the top CoreNet modules are conserved across species when using this approach. Previous studies have highlighted the significance of several of these modules, but these studies were limited to assessing Tfh in isolation, using isolated data modalities, in vitro systems, or either human or murine systems. A significant advancement of this work is that the predicted prioritizations are enhanced by combining modalities, not diluted. For instance, the core Tfh modules identified from human tonsil samples also recapitulated known regulators of Tfh differentiation in mice, suggesting the core Tfh gene set is conserved. Our analysis also unifies the previously identified modules by illustrating how these seemingly individual networks are interconnected at the protein-protein interaction (PPI) level. Fully differentiated GC Tfh cells from humans and mice share similar surface protein phenotypes and nearly identical gene expression profiles. However, prior in vitro studies attempting to induce Tfh differentiation have often suggested that there are functional differences between species. Since the methods available for human studies are limited to ex vivo analysis and in vitro approaches, the cells generated in vitro may not correspond to bona fide in vivo GC Tfh. Thus, they may not fully reflect GC Tfh differentiation and apparent discrepancies between mice and humans may reflect variability in experimental systems and limitations of analysis, rather than true species differences. In vitro systems may not be able to recapitulate complex cognate interactions with GC B cells that GC Tfh rely on for their differentiation and maintenance. Our network approach, however, highlighted conserved regulators between species, enabling us to formulate hypotheses about Tfh heterogeneity from human data sets and directly test them at a mechanistic level using murine models. An intriguing candidate emerged from this perspective: the IL-12 signaling pathway. Our analysis of human T cells ranked it as highly important but downregulated among GC Tfh compared to other pre-Tfh states. Several prior studies utilizing in vitro differentiation of human T cells have reported that IL-12 can enhance or even be necessary for human Tfh differentiation. However, it has also been shown that IL-12 can suppress Tfh and GC formation in multiple murine models of infection. Combining these observations with our own data, we hypothesized that IL-12 may enhance early stages of Tfh differentiation, but block progression to the fully mature GC Tfh state. To test this, we examined the recently defined murine GC Tfh cells and found that IL-12 strongly suppresses their advanced differentiation. Interestingly, non-GC-Tfh cells were minimally repressed by STm infection, although blocking IL-12 did enhance their differentiation, potentially through other inflammatory signals induced by STm infection. In vitro differentiation, therefore, likely represents an immature Tfh differentiation state similar to pre-Tfh, and for both human and murine cells IL-12 signaling blocks the most mature stage of GC Tfh differentiation. Intriguingly, Yeh et al.36 demonstrated that GC Tfh development relies on B cell antigen presentation, while non-GC Tfh can be generated via interaction with DCs. The initial observation that IL-12 could produce Tfh-like cells in vitro in human cells was through co-culture with DCs25, so B cell-derived signals could be a missing link for the in vitro differentiation of GC Tfh. Thus, at present, in vitro systems fail to accurately represent ex vivo GC Tfh, and studies such as ours, combining high quality human cell analysis with murine experiments, remain crucial for dissecting molecular mechanisms and cellular interactions that are difficult to model in vitro. In addition to IL-12 signaling, our analysis predicted several modules that have been demonstrated to play roles in Tfh differentiation, but which have been less well-studied. One intriguing protein is PTPN11 (SHP-2). PTPN11 has been implicated in PD-1 signaling in GC Tfh, and PD-1 contributes to the localization of Tfh within the GC. Consistent with this and human data60, GC Tfh were the highest expressors of PD-1 (Fig. 7d), and our network connects PTPN11 and PD-1 (PDCD1, Fig. 2c). However, the molecular mechanism of SHP-2 function in this was not investigated. Other modules, such as HIF1A, have mixed reports of their role in Tfh differentiation94–96. Our analysis predicts that HIF1A is important but is not differentially expressed between non-GC Tfh and GC Tfh (Fig. 2f). IRF4 was also identified (Fig. 2c, 2f, 2g), which has been shown to direct Tfh versus Th1 differentiation through reading out TCR signaling strength97. Consistent with this, our analysis shows IRF4 connected to NFATC2 and a hub of other molecules that regulate TCR signaling (Fig. 4e). While some reports highlight the importance of TCR signaling in Tfh60,97–101, our analysis reveals a sub-network around the NFATC2 hub, suggesting that intriguing and important biology remains to be discovered in how TCR signaling regulates Tfh differentiation. The integrative analysis reported here succeeds in capturing and encompassing over a decade of Tfh-related discoveries while at the same time revealing multiple candidate genes and pathways. Although curated gene sets may help define Tfh identity in some contexts, they often lack the ability to resolve cross-context differences. The CoreNet gene sets identified in this work converge on a core regulatory network that is conserved across species, tissues, disease states, and experimental systems, an outcome only possible if the underlying biology is truly consistent. The identified modules and genes offer a rich resource for future functional studies and therapeutic development and are web accessible through our web resource. Notably, our findings help resolve a long-standing debate in the field regarding the role of IL-12 in Tfh differentiation, with important implications for vaccine strategies that utilize IL-12–based adjuvants. Further, this method could be applied to other biological contexts outside of Tfh differentiation and provides a roadmap for dissecting immune cell differentiation which has broad relevance for immunotherapy and vaccine design. Methods Ethics statement All human tonsil samples were excess surgical tissue obtained from patients at Massachusetts General Hospital. No further demographic stratification was performed. Institutional Review Board at Massachusetts General Hospital provided ethics oversight (approval number 2020P003521). All tissue samples from people living with HIV (PLWHIV) were procured with explicit written informed consent from participants prior to donation, adhering strictly to the principles outlined in the Declaration of Helsinki. The utilization of remnant samples was formally sanctioned by both the Research Committee and the Ethics in Research Committee of the National Institute of Respiratory Diseases “Ismael Cosío Villegas” (INER), Mexico City as part of the “C71-18” protocol. Under a material transfer agreement this project was conducted at Emory University with ethical approval STUDY00006200. Human Flow cytometry and sorting 600 million cells from healthy adult human tonsils were resuspended in Miltenyi magnetic-activated cell sorting (MACS) buffer and stained with biotinylated anti-human CD19 (clone: HIB19; BioLegend, #555411) on ice for 30 min. Cells were washed once in MACS buffer and incubated with antibiotin microbeads (Miltenyi Biotec) for 30 min on ice. Cells were then loaded on Miltenyi LS columns at 300 million cells per column, and the flow-through was collected as the B cell–depleted fraction. These cells were spun down and subjected to Fc block (Human BD Fc Block BD Biosciences, #564219), washed with 1% BSA in PBS, and stained with LIVE/DEAD Fixable Blue Dead Cell Stain (ThermoFisher Scientific, #L23105), before they were surface stained with antibodies against CD3 (clone: SK7; BioLegend, #344817), CD4 (clone: OKT4; BioLegend, #317420), CD45RA (clone: H100; BioLegend, #304122), CXCR5 (clone: J252D4; BioLegend, #356920), and PD1 (clone: EH12.2H7; BioLegend, #329924). The following populations were sorted using FACSAria2 (Becton Dickinson), as shown in the gating strategy present in Fig. 1a: (i) naive CD4 T cells (CD4 + CD45RA + CXCR5-PD1-), (ii) Early Extra-GC Tfh (CD4 + CD45RA-CXCR5 + PD1-), (iii) Late Extra-GC Tfh (CD4 + CD45RA-CXCR5 + PD1 + ), and (iv) GC Tfh cells (CD4 + CD45RA-CXCR5 + + PD1 + +). All antibodies were used at a dilution of 1:40. Bulk RNA-seq and analysis Total RNA was isolated from the FACS-sorted healthy tonsillar cells using the RNeasy plus Micro Kit, according to the manufacturer’s instructions (Qiagen). RNA-sequencing libraries were prepared as previously described102. Briefly, whole transcriptome amplification and tagmentation-based library preparation were performed using the SMART-Seq2 protocol. Libraries were assessed for quality using High Sensitivity DNA chips on the Agilent Bioanalyzer, quantified using Qubit fluorometer (Thermo Fisher Scientific), as well as KAPA Library Quantification kit (KAPA Biosystems), followed by 35-bp paired-end sequencing on a NextSeq 500 instrument (Illumina). 5–10 million reads were obtained from each sample and aligned to the University of California, Santa Cruz hg38 transcriptome. Gene expression was calculated using RSEM as previously described, with alignments to ENSEMBL hg38, and differential expression analysis was conducted using EBSeq103, with differentially-expressed genes determined by selecting genes that differed across the 4 states with a PPDE (posterior probability of differential expression) of at least 0.95. CUT&RUN CUT&RUN libraries were prepared using rabbit monoclonal antibodies against H3K4me1, H3K4me3 and H3K27Ac (Cell Signaling Technology catalog numbers 5326, 9751, and 8173, respectively) as previously described. The protein-A Mnase fusion protein was a gift from Steven Henikoff. ATAC-seq ATAC-Seq libraries were prepared as originally described38. Fifty thousand freshly sorted cells of each population were pelleted and washed with 50 μL chilled 1× PBS, and with 50 μL lysis buffer (10 mM Tris-HCl pH 7.4, 10 mM NaCl, 3 mM MgCl2, 0.1% IGEPAL CA-630). The nuclei were pelleted in a cold micro-centrifuge at 550 × g for 10 min and resuspended in a 50 μL transposition reaction with 2.5 μL Tn5 transposase (FC-121-1030; Illumina) to tagment open chromatin. The reaction was incubated at 37 °C in a Thermomixer (Eppendorf) at 300 rpm for 30 min. Tagmented DNA was purified using a QIAGEN MinElute kit and amplified with 7 or 11 cycles of PCR, based on the results of a test qPCR. ATAC-Seq libraries were then purified using a QIAGEN PCR cleanup kit and quantified using KAPA library quantification kit (KAPA Biosystems, Roche) and sequenced on the Nextseq 550 platform. Personalized PageRank identification of driver TFs To identify candidate driver transcription factors (TFs), we applied the personalized PageRank algorithm implemented in Taiji v1.2.044. A sample configuration file is provided in the accompanying codebase. Briefly, both epigenomic datasets (ATAC-seq, H3K4me1, H3K4me3, H3K27Ac) and bulk RNA-seq data were organized by biological replicate and by cell state (Naive, Early Extra-GC Tfh, Late Extra-GC Tfh, and GC Tfh). A separate Taiji run was performed for each of the four epigenomic modalities, each paired with the bulk RNA-seq dataset. For the ATAC-seq runs, input consisted of four replicate pairs of narrowPeak and BED files per state, with the file format specified as ‘NarrowPeak’. For the CUT&RUN datasets (H3K4me1, H3K4me3, H3K27Ac), four replicate BAM files were used per state, with the tag set to ‘PairedEnd’. In all runs, four replicate TSV files of bulk RNA-seq per state were provided, with the input tag set to ‘GeneQuant’. The human genome assembly GRCh38 was used throughout. Taiji’s output GeneRanks files were filtered to retain only genes with a p < 0.01. Within each epigenomic dataset, genes were ranked per state, and for each gene, the minimum rank across the four states was determined. The 100 genes with the lowest minimum ranks were selected from each dataset. These top-ranked TFs were then combined across the four epigenomic datasets, yielding a final set of 152 unique transcription factors which make up CoreNet-GRN. To visualize the results, Taiji ranks were averaged across the four epigenomic datasets. We then calculated log2FC in both average Taiji rank and gene expression between the GC and Early Extra-GC Tfh or Late Extra-GC states. Calculation of gene-centric peak proximity scores Differential signal in epigenomic datasets was quantified using EBSeq to identify genes that differed across the 4 states. For this gene subset, peak proximity scores (PPS’s) for a gene (“gene centric PPS”) were calculated by computing a weighted sum of the number of peak clusters from the pattern in the gene’s vicinity using distance-based weights to reflect the varying contributions of promoter (± 10 kb) [3x weight], proximal enhancer (± 25 kb) [2x weight] and distal enhancers (± 250 kb) [1x weight] to gene regulation. Gene-centric PPS values were first calculated individually for each epigenomic dataset. These were then combined into a final PPS value using equal weights for the different datasets. Network propagation using random walk with restart To integrate PPS values with protein-interaction topology, we propagated PPS-derived node weights across a human PPI network using the random walk with restart algorithm, implemented in HotNet257. The HotNet2 code is available at https://github.com/raphael-group/hotnet2. The PPI network was constructed as the union of high-quality reference human binary and co-complex protein-protein interaction networks from the HINT database56. The influence matrix required for HotNet2 network propagation were prepared for the HINT-derived PPI network (input network) and permuted versions of this network were prepared using the makeNetworkFiles.py script provided with HotNet2. For a robust analysis, we generated 500 permuted networks and 500 heat-vector permutations. All files were saved in HDF5 format. Network propagation was run using the default HotNet2 parameters: beta = 0.4 and four delta values automatically selected by the software. A sample configuration file is provided in the accompanying codebase. Modules were considered significant if they had a P < 0.05 and contained more than three genes. We performed five independent runs and retained only the subnetworks that appeared in all five runs. These reproducible subnetworks are referred to as the CoreNet-PPI gene set. The CoreNet-PPI networks were visualized using Cytoscape104 v. 3.10.2 and colored according to log fold change (log2FC) values across the Tfh differentiation trajectory, where fold change was calculated as log2(GC Tfh / Early Extra-GC Tfh). A random sample of 2000 edges from the HINT-derived PPI network was also visualized alongside the subnetworks using the R package qgraph105. The NetworkX (v. 2.5)106 steiner_tree python package was used to identify the shortest path between the genes in the 2016 reference set from the PPI network. Then, all direct interactors of a gene of interest were added to the graph and visualized in Cytoscape. Genes belonging to the 2016 reference set were outlined in orange and the node color corresponds to the node PPS values prior to the network propagation with dark blue being the highest expression and gray indicating no value. Mitigation of false negatives CoreNet was defined as the union of the CoreNet-GRN and CoreNet-PPI gene sets. CoreNet-DI was defined by adding direct interactors of CoreNet genes from the HINT-derived PPI network. To generate CoreNet-Extended, a set with fewer false negatives than CoreNet, we used two (PPI and GRN based) recovery strategies to identify genes with strong propagated or regulatory support, but not enough to make it into CoreNet. We removed mitochondrial content, ribosomal components, hemoglobin, sex chromosomes, MTRN, TRAV, TRBV, IGKV, IGHV, and IGLV, and LOC genes to reduce technical and tissue-specific bias. Equally, for the GRN-based recovery, genes already in CoreNet-GRN were excluded and we ranked downstream targets using GC Tfh GRN edge weights and TF importance scores. For the PPI-based recovery, genes already in CoreNet-PPI were excluded, and the remaining were ranked by their heat scores to identify the top 5% highest scoring genes (564). Similarly, for the GRN-based recovery, genes already in CoreNet-GRN were excluded and we ranked downstream targets using GC Tfh GRN edge weights and TF importance scores for each epigenomic set. The top 200 genes were taken from each of the 4 networks and combined into one set (592). These genes were combined with CoreNet to generate CoreNet-Extended for a total of 1406 genes. The NetworkX (v. 2.5)106 steiner_tree python package was used to identify the shortest path between the genes in the 2016 reference set from the PPI network and visualized in Cytoscape. Genes belonging to the 2016 reference set were outlined in orange and the node color corresponds to the node heat score post network propagation with dark blue being the highest expression and gray indicating no value. The overlap of CoreNet-Extended and Vinuesa gene set was depicted using an oval node shape for overlap and a rectangle for no overlap. Functional overlap between Tfh cell subnetworks and published datasets We evaluated the functional overlap between our CoreNet and published Tfh gene sets, where functional overlap was defined as either exact gene matches or genes that directly interact. Established Tfh gene sets were derived from three reviews: Vinuesa et al.58, Hart and Laufer59, and a review of mouse Tfh genes by Crotty et al.31. We compared the functional overlap of genes prioritized by our combined network propagation approach (CoreNet) to those identified by protein interaction network (CoreNet-PPI) propagation alone, Taiji (CoreNet-GRN) alone, and to the average overlap from 1000 random network-derived gene sets matched in size for each dataset (baseline). To test the importance of the network approach, we also generated size-matched comparison sets by selecting the top 184 genes by RNA-seq log2FC between GC Tfh and Early Extra-GC Tfh cells, or by epigenomic PPS. The RNA-seq log2FC and PPS gene sets represent top-scoring genes based solely on individual gene-level metrics. Functional overlap for these gene sets was similarly compared to 1,000 total gene set size-matched random gene sets and average random overlap was calculated. P values were calculated using a Fisher’s exact test. We also created another version of the random baseline for testing. 1000 size-matched sets to the prioritized genes (184) were randomly generated and their direct interactors were included. As adding the direct interactors led to a varying size of the random set, the average random overlap and average random gene size were used for the binomial proportions test p-value calculation. The CoreNet sets and CoreNet-Extended were also tested. For these analyses, random gene sets size-matched to CoreNet and CoreNet-Extended were sampled from the PPI network and overlapped with the literature gene sets. Significance was measured using a binomial proportions and Fisher’s exact test respectively. The PPI network overlap between the reference Tfh gene sets and the CoreNet-DI set was created by extracting the corresponding nodes and edges from the HINT database associated with the reference genes, as well as including additional direct interactors of the reference genes that were present in the CoreNet-DI set. The PPI network overlap was visualized using Cytoscape. We colored these overlapping modules based on the log2FCs across the trajectories, where fold change was defined as log2(GC Tfh/Early Extra-GC Tfh) or log2(GC Tfh/Late Extra-GC Tfh). SCENIC+ analysis For construction of enhancer driven gene regulatory networks (GRNs), SCENIC + 62 was utilized on the downloaded human tonsil atlas67 R object and associated fragment files. The clusters “CM-PreTfh” and “GC-Tfh-SAP” which most closely resemble the Early Extra-GC Tfh and GC Tfh states were subsequently processed according to SCENIC+ guidelines with no further cell filtration. The code can be found on the github: https://github.com/aertslab/scenicplus. Specifically, the snRNA portion was preprocessed to ensure that the non-normalized and non-scaled matrix was retained in the raw slot. The multiome snATAC was processed following pcisTopic recommendations, performing MACS2 peak calling on each of the clusters “CM-PreTfh” and “GC-Tfh-SAP” and generating a consensus region set from both. This consensus region set was used to create a custom cisTarget database with the Aerts lab’s publicly available motif collection and motif-to-TF annotations. For preparing a FASTA file, sequences for all the consensus peaks were pulled along with an additional 1 kb of background padding to be used as the background sequence for cluster-buster (per SCENIC+ recommendations). Databases containing the motif rankings and scores were subsequently constructed from this FASTA with cluster-buster to be used in the snakemake pipeline. Topic modeling with parallel latent dirichlet allocation (LDA) using a Collapsed Gibbs Sampler was performed on the snATAC. Model selection was performed through evaluation of the provided pycisTopic metrics (Minmo_2011, Log-likelihood, Arun_2010, and Cao_Juan_2009). Based on these metrics 20 topics were chosen for the subsequent analyses (clustering, topic binarization with the ‘otsu’ method, imputation, and assessment of differentially accessible regions between cell types). Differentially accessible regions (DARs) were assessed using a Wilcoxon rank-sum test between each group. Region sets and an snATAC object were saved following processing. Finally, to run the SCENIC+ snakemake pipeline, the snATAC cisTopic object, associated snRNA data, region_sets, and cisTarget databases were provided to generate direct and extended enhancer GRNs along with a final MuData file containing the eRegulons and enrichment scores. SLIDE analysis Two single-cell populations of interest were chosen for each analysis. SEACells107 was used to cluster the cells per cell type. The code can be found on GitHub: https://github.com/dpeerlab/SEACells. The following parameters were used: convergence_epsilon=1e-5, n_SEACells=75, n_waypoint_eigs=10, and build_kernel_on = ’X_pca’. The 2000 highest variable genes per cell type were chosen after removing mitochondrial content, ribosomal components, hemoglobin, sex chromosomes, MTRN, TRAV, TRBV, IGKV, IGHV, and IGLV genes to reduce technical and tissue-specific bias. The variable genes for each cell population were combined and used as features for SLIDE63. The SLIDE code is available on GitHub: https://github.com/jishnu-lab/SLIDE. As an example of parameters, for the GC-Tfh-SAP and CM Pre-Tfh analysis the following parameters were used: delta=0.05, lambda=4, spec=0.1, y_factor=TRUE, y_levels: (0, 1), eval_type=auc, SLIDE_iter=1000, SLIDE_top_feats=10, sampleCV_iter=500, sampleCV_K = 4, do_interacts=TRUE, fdr=0.1. Only significant latent factors were chosen based on the permutation control for all runs. Additionally, the Cliff’s delta was calculated to see how well a single latent score in the Z matrix separates the distribution of two populations. scRNA analysis of Tfh sub-populations scRNA-seq data for human tonsils was acquired from the human tonsil atlas67. To identify Tfh sub-populations, cells which pass the marker criteria CD44 + CXCR5 + PDCD1 + LY6G6C-SELPLG- were kept35. Tfh1 (CXCR3⁺CCR6⁻), Tfh17 (CXCR3⁻CCR6⁺), and Tfh2 (CXCR3⁻CCR6⁻) were further subset. There were less than 100 cells which were included in the Tfh1 and Tfh17 population with 3681 cells for Tfh2. For the comparison the CM Pre-non-Tfh and GC-Tfh-SAP clusters from the Tonsil Atlas were chosen. Cells labeled as Tfh1, Tfh17, or Tfh2 were excluded from these clusters if present. SEACells and SLIDE were performed as previously described. For the SLIDE analysis lambda was set to 2 and 1 (CM Pre-non-Tfh and GC-Tfh-SAP, respectively) with the rest of the parameters as described. scRNA enrichment of Tfh cell subnetworks and published datasets For the scRNAseq analyses, ssGSEA was performed using the escape R package108 (min.size = 2, normalize = FALSE) to identify the cell enrichment score for each gene set of interest and cell type. Different gene sets were evaluated: two references derived from published review articles, CoreNet-PPI, SCENIC + , and SLIDE. Enrichment scores were compared across cell types to assess differential pathway activity. For the GC up-regulated genes in the Vinuesa dataset only the upregulating and promoting genes were kept. For the GC-upregulated genes in SCENIC+ the GC activator TFs and their regulons were kept. For all other sets the gene lists were filtered based on a positive log2 fold change of GC vs Early Extra-GC Tfh mRNA expression. Statistical significance was determined using p-values calculated from the Wilcoxon rank-sum test and effect sizes were quantified using Cliff’s Δ to capture the magnitude of enrichment differences. The data was visualized using boxplots with a dot overlay. Only 70% of the population was visually represented by dots and was randomly sampled. Functionally different datasets between cell types were considered significant by passing both Wilcoxon rank-sum test p-value (<0.05) and effect size given by the Cliff’s Δ ≥ abs(0.1). Human PLWHIV Gut scRNA-seq dataset Nine PLWHIV receiving suppressive antiretroviral therapy (ART) for over one year donated gut tissue biopsies. Ileal, colonic, and rectal samples were collected during lower endoscopy procedures (Olympus CF-HQ190). Twenty tissue snips per anatomical site were collected using standard biopsy forceps (Boston Scientific, Radial Jaw), immediately transferred into complete RPMI medium (cRPMI, RPMI 1640 [Cytiva, Marlborough, Massachusetts] supplemented with 10% fetal bovine serum [Biowest, Bradenton, Florida], 2mM L-glutamine [Cytiva], and 100 U/ml penicillin and 100ug/ml of streptomycin [Cytiva]), and transported to the laboratory within 15 minutes of collection. No adverse events were reported during the procedures. All participants were scheduled for a follow-up consultation one week after the surgical procedure. None of the participants of this study had opportunistic infections, and all were HBV- and HCV-negative. Clinical and demographic data of all donors has been reported earlier65. Gut single-cell emulsions were prepared within 30 minutes after cell sorting. Single cell RNA sequencing (Chromium Next GEM Single Cell Gene Expression, 10X Genomics) was performed. Thawed cells were counted, and the viability was determined by staining the cells with Trypan blue. Isolated CD45+ cells were centrifuged and resuspended at a concentration of 800–1200 cells per microliter in RPMI with 2% (v/v) human AB serum and loaded onto 10X Chromium instrument and Single Cell Gene Expression (GEX) libraries were prepared with the following kit: Chromium Single Cell 5’ V1 (10X Genomics). 10,000 cells were targeted for each sample. Isolated cells were transposed and partitioned into Gel Beads-in-emulsion (GEMs) using the 10x Chromium Controller and Next GEM Chip J. GEMs were visually inspected, and only samples with an opaque and uniform aspect were used for library preparation. Representative traces and quantitation of libraries were determined using Bioanalyzer High Sensitivity DNA Analysis (Agilent, Santa Clara, CA). Sequencing was done on Illumina NovaSeq S4 targeting 20,000 reads per nucleus for gene expression. Fastq files were processed with 10x Genomics Cell Ranger v5.0.1 using 10x Genomics Cloud Analysis109,110. Reads were mapped to the GRCh38 human reference genome and counted without depth normalization. The filtered count matrix was then analyzed using the Seurat (v5.1.0)111 in R. Low quality cells were identified and removed based on the following criteria: more than 100 RNA unique molecular identifiers, less than 25% mitochondrial read fraction. Doublet cells were identified using the DoubletFinder package in R112. Only CD4+ population cells were analyzed further. We performed single-cell RNA-seq analysis using Seurat113 (v5.0.0) R version (4.4.0), followed by batch correction with Harmony to integrate data across experimental conditions. Outlier cells were removed prior to analysis to ensure high-quality clustering. Genes associated with mitochondrial content, ribosomal components, hemoglobin, sex chromosomes, and the surfactant protein family were excluded to reduce technical and tissue-specific bias. Gene expression values were then normalized. We identified the top 5,000 highly variable features using the variance-stabilizing transformation (vst) method and scaled gene expression values across cells. Principal component analysis (PCA) was performed using these variable genes, and the resulting principal components were corrected for batch effects using the Harmony algorithm. An elbow plot of the Harmony components was used to determine the number of informative dimensions, and the first 20 components were selected for downstream analysis. Clustering was performed using shared nearest neighbor (SNN) graph construction and the Louvain algorithm, with parameters k.param= 50 and resolution = 0.7 based on the Harmony-reduced space. We identified cluster 2 as primarily Tfh cells by using a composite expression score of Tfh markers: BCL6, CXCR5, CXCL13, CD200, PDCD1. Further, the Azimuth Human Tonsil v2 reference67 was used to label the gut CD4+ cells. The cells in the overlap of cluster 2 with either Azimuth GC-Tfh-SAP or Tfh T:B border labels were chosen for the Tfh group. We also identified 2 non-Tfh populations for comparison: Th17 and Central Memory (CM) CD4 + . First, we removed all Tfh, Naive, Cycling, Treg/Eff-Tregs-IL32, and Tfh adjacent (Tfh mem, GC-Tfh-OX-40 and Tfh-LZ-GC) labeled cells from the dataset. A composite score of Th17 markers (CCR6, IL-23R, and RORC) was used to select cluster 5 as the Th17 cells. For the CM non-Tfh CD4+ group, the dataset was reclustered using the Azimuth Human PBMC reference114 and used markers to visualize the cells (CCR7, SELL). The cells labeled central memory and effector memory were selected. Murine scRNA-seq datasets We processed publicly available mouse lymph node scRNA-seq data from He et al.70 allergy model. We followed a standard Scanpy115 (v1.10.3) workflow. Raw counts were normalized per cell and log-transformed with default parameters. Top 2000 highly variable genes were selected. We performed principal component analysis (PCA) and embedded neighborhoods in UMAP space. Potential cell doublets were identified and removed using Scrublet116 implemented in Scanpy with an expected doublet rate of 0.05. Major immune cell types were annotated based on the expression of canonical marker genes: Cd19 and Ms4a1 for B cells; Cd3d and Cd4 for CD4 + T cells; Cd3d and Cd8a for CD8 + T cells; Klrb1c and Klrk1 for NK cells; Itgax for DCs. Cells co-expressing Ms4a1 (CD20, B cell marker) and Cd3d (T cell marker) were identified as potential T/B doublets and removed. We further subsetted the CD4 + T cell compartment and identified CD4 + T cell subtypes using established lineage-defining markers: Bcl6 and Pdcd1 for Tfh; Gata3 and Ccr4 for Th2; Il-2ra and Foxp3 for Treg. CD4 + T cells negative for Tfh, Th2, and Treg markers and exhibiting low expression of activation markers (Cd69, Icos) were designated as resting CD4 + T cells. ssGSEA was performed on the Tfh and Resting CD4 + T cell populations using the escape R package (min.size = 2, normalize = FALSE) to identify the cell enrichment score for each gene set of interest and cell type. Human spatial datasets Cells with Tfh or CD4 + T-cell labels were extracted from human tonsil slide-tag snRNA-seq data69 with the R package Seurat. From the CD4 + T-cell group, naive T-cells were filtered out using a composite score of the following markers (‘CD3E’, ‘CCR7’, ‘LEF1’, ‘SELL’, ‘GATA3’,‘CD4’, ‘LTB’, ‘RCAN3’,‘CD40LG’, ‘PIK3IP1’, ‘TCF7’) with a threshold of >10. ssGSEA was performed using the escape R package to identify the cell enrichment score for each gene set of interest and cell type. The cells were visualized in 2D space and colored by enrichment. SpaceTravLR and mouse spatial datasets SpaceTravLR leverages convolutional neural networks to generate spatially varying coefficients for a linear model, composed of transcription factor, ligand, and receptor terms. The spatial datasets used to train SpaceTravLR were obtained from a murine model of allergic asthma as described earlier70. The first mouse lymph node, harvested at Day 5 and profiled using SlideseqV2, was subsampled to 12,681 (33%) spots and trained using 3354 genes to reduce computational intensity. The second mouse lymph node, harvested at Day 3 and profiled using Visium HD, was subsampled to 22,324 (25%) spots and trained using 2859 genes for the same reason. For both mouse lymph nodes trained using SpaceTravLR, we utilized the scanpy framework to perform cell and gene filtering, library size normalization, and log1p transformation. The log transformed data was also smoothed for each cell type using MAGIC. We then removed housekeeping genes and identified the top 3000 highly variable genes for the SlideseqV2 lymph node and the top 2000 highly variable genes for the VisiumHD lymph node. We included highly expressed genes encoding known ligands or receptors and ensured that the genes relevant to the small populations of CD4 T cells were included, along with the CoreNet mouse orthologs. Both models were then trained over 150 epochs, a learning rate of 5e-3, and a batch size of 512. The amount of ligand received by each cell was computed by considering all cells within a radius of ~300 µm. The genes that SpaceTravLR was trained on included 291/271 CoreNet and 366/403 CoreNet-Extended mouse orthologs, 231/338 SCENIC+ orthologs, and 143/218 genes from Dalit et al. 2025 for VisiumHD/SlideSeq models, respectively. The number of genes in the intersection of the trained genes were then used to compute the enrichment scores when quantifying overlap for each method. The model trained and published along with the original SpaceTravLR manuscript71 followed a similar preprocessing pipeline, but did not include many of the relevant Tfh genes. Only 38 of the CoreNet-PPI, 30 of the CoreNet-GRN, and 535 of the CoreNet-DI were included. The full Day 3 VisiumHD mouse lymph node included all 89,296 spots but only 2116 total genes. This model was trained in 4 chunks, each over 100 epochs, a learning rate of 5e-3, and a batch size of 512. The amount of ligand received by each cell was computed by considering all cells within a radius of ~700 µm. Path reconstruction is described in SpaceTravLR71. Briefly, it leverages the gradient information utilized during forward network propagation. Starting from a target gene (e.g., Il-21), we iteratively identify the most influential upstream modulators: (1) select the top 5 genes contributing to Il-21’s expression change, (2) for each of these genes, select their top 5 modulators, (3) repeat for up to 2 additional hops or until reaching the perturbed gene (Bcl6). Since Bcl6 is the only knockout, all accumulated expression changes trace back to this single source, creating an interpretable regulatory cascade. Human Tonsil CITE-seq analysis CD4 + T cells were isolated from the Human Tonsil Atlas CITE-seq dataset and loaded as a Seurat object. Antibody-derived tag (ADT) identifiers from the CITE-seq panel were harmonized to HGNC-approved gene symbols and the remaining unmatched ADTs were resolved through manual curation. TCR chain markers, isotype controls, and identifiers without clear gene-level mappings were excluded from downstream scoring. Tfh-associated ADT targets were identified through an overlap with CoreNet. Cell type annotation was performed using Azimuth with the human tonsil reference atlas (tonsilref) via the RunAzimuth function from the Azimuth R package. Differential ADT protein abundance between the Azimuth labeled GC-Tfh-SAP cells and CM Pre-non-Tfh cells was assessed using a Wilcoxon rank-sum test implemented via Seurat’s FindMarkers function, with no minimum fold-change or abundance percentage thresholds applied so as to retain all proteins for comparison. P-values were corrected for multiple testing using the Benjamini-Hochberg procedure, and proteins with an adjusted p-value below 0.05 were considered significant. Results were visualized as a volcano plot, with proteins from the Tfh target gene set and an absolute log2 fold-change greater than 0.5 labeled. Per-protein mean CLR-normalized ADT abundance and standard error were additionally plotted as a bar chart. Human Tonsil spatial CITE-seq analysis The spatial CITE-seq dataset was acquired from the scProAtlas pre-processed AnnData object (.h5ad format). CoreNet was filtered to retain only those upregulated in GC T cells relative to early T cells, as determined by RNA-seq log2 fold-change data. Raw count data were normalized to a library size of 10,000 counts per cell and log1p-transformed. Missing protein values (NaN) arising from undetected proteins were imputed to zero. Dimensionality reduction was performed sequentially: highly variable genes (top 2000) were identified using the Seurat flavor where applicable, followed by PCA, k-nearest neighbor graph construction, and UMAP embedding, all implemented in Scanpy. Tfh cell scoring was performed using Scanpy’s score_genes function applied to the CoreNet gene set. Cells were classified as Tfh if they simultaneously satisfied three criteria: (1) a Tfh score at or above the 85th percentile across all cells, (2) detectable abundance of at least 15 CoreNet genes (or all available CoreNet genes if fewer than 15 were present in the dataset), and (3) positive CD4 abundance. This strategy yielded 102 Tfh cells from a total of 2439 profiled cells (4.18%). Downstream comparisons were restricted to CD4-positive cells (n = 1375), comprising 102 Tfh and 1273 non-Tfh CD4 T cells. Differential protein abundance between Tfh and non-Tfh CD4 T cells was assessed for individual marker genes using two-sided Mann–Whitney U tests. P-values were corrected for multiple testing using the Benjamini-Hochberg procedure, and proteins with an adjusted p-value below 0.05 were considered significant. Mean abundance and standard error of the mean were computed per group, and fold-change was calculated as the ratio of mean Tfh to mean non-Tfh abundance. Markers NR3C1 and POU2F2 were then held out one at a time, and the classification and significance were recalculated as described. Pathway enrichment calculations Pathways were extracted using the msigdb R package (v7.5.1)117. Top pathways were identified for bulk RNA and scRNA datasets from two different databases: Hallmark and PID, across species and tissues. The PPI network overlap between the pathway gene sets and the identified Tfh gene set was created and visualized as previously described. For the bulk RNA-seq pathways, enriched for the CoreNet-PPI gene set were detected by calculating the functional overlap between the CoreNet-PPI gene set and all pathways in the two databases. The significance of gene set overlap was assessed by comparing to the average overlap of 1000 random, size-matched sets of network genes. Functionally overlapping pathways were considered significant if the odds ratio (defined as the odds of overlap between the identified gene set and a given pathway divided by the odds of overlap from a random gene set) exceeded 1.5, and the false discovery rate (FDR) was below 0.2. P-values were calculated using a binomial proportions test and adjusted for multiple comparisons using the Benjamini–Hochberg correction. Significant modules were visualized as described previously, with color representing the log2FC across states. For the scRNA-seq analysis, ssGSEA was performed using the escape R package (min.size = 2, normalize = FALSE) to identify the cell enrichment score for each pathway. Functionally different pathways between cell types were considered significant by passing both Wilcoxon rank-sum test p < 0.05 and effect size given by the Cliff’s Δ ≥ abs(0.1). The top 15 pathways with the highest abs(Cliff’s Δ) value were reported. Murine salmonella experiments Mice B1-8 + /-118 BALB/c mice, containing elevated frequencies of NP-specific B cells, were used for all mouse experiments. Mice were bred and housed in specific pathogen-free conditions in a room at 20–26°C with 30–70% relative humidity. Mice were euthanized using carbon dioxide (CO2) inhalation followed by a secondary method of cervical dislocation in accordance with institutional guidelines. All experiments were conducted under protocols approved by the University of Pittsburgh Institutional Animal Care and Use Committee (IACUC). All mice were between the age of 8-16 weeks at the start of experiments. Males and females were both used and no differences were noted. Bacterial strains Salmonella enterica serovar Typhimurium (STm) aroA attenuated strain SL3261119, provided by Roy Curtiss III, Arizona State University) was used for all experiments. Glycerol stocks were prepared by growing bacteria overnight in Luria broth, then mixed 1:1 with sterile 50% glycerol and stored at −80 °C. Infection, Immunization, and anti-IL-12 Treatment To prepare STm, one day prior to injection, stabs of frozen STm glycerol stock were inoculated into Luria broth (Fisher) containing 100 μg/mL streptomycin and grown at 37 °C overnight with shaking. The day of injection, heat killed bacteria were prepared as follows: cells were pelleted at 1800g, washed twice in PBS, resuspended in PBS, transferred to a fresh tube, and incubated at 56 °C for 1 h. For infectious inoculum, bacterial cultures were split 1:25 into fresh media and grown for an additional 2 h at 37 °C to allow bacteria to reach mid-log phase, then washed 3 times in 20 °C sterile PBS. Bacterial concentration was estimated by absorbance at OD600, and approximately 5 × 105 colony forming units were injected i.p. in 200 μL PBS. Colony forming units of infectious inoculum, and absence of live bacteria in heat killed inoculum, were confirmed by overnight growth at 37 °C of serial dilutions of inoculum on Luria broth agar (BD Difco) plates with streptomycin. For NP-CGG immunization, mice were immunized with 50 μg NP-CGG precipitated in alum i.p. in 200 μL PBS. The ratios of NP to CGG ranged from 31-33. For IL-12 blocking experiments, mice were infected with STm or given HK STm, then immunized with NP-CGG 3-5 h later. All mice were given an initial loading dose of 200 μg anti-IL-12 in PBS by intraperitoneal injection (BioXcell catalog number BE0051, clone C17.8). Mice in day −1 treatment groups received an additional 100 μg at days 3 and 7. Mice in day 3 treatment groups received another 100 μg at day 7. Mice in day 7 and 11 treatment groups only received the loading dose. Flow cytometry analysis for mouse experiments Spleen pieces were weighed and made into single cell suspensions by mechanical disruption in STm media (PBS + 5% bovine serum + 2 mM EDTA). RBC were lysed with ACK lysis buffer for 60 s, washed and resuspended in STm media, and live cells enumerated with trypan blue on a hemacytometer. Cells were pelleted and washed once in PBS before staining. Five million cells per sample were stained at 100 million cells/mL with the dead cell discriminator Ghost Violet 510 (Tonbo, per manufacturers protocol) in PBS, and washed with staining media (SM; PBS with 3% bovine serum (Gibco), 1 mM EDTA, and 0.02% sodium azide). Cells were stained with fluorescently conjugated antibodies as labeled in figures in SM for 20 min on ice, then washed twice with staining media, and fixed with 1% PFA in PBS for 20 minutes on ice. For intracellular stains, cells were instead fixed/permeabilized with FoxP3 Transcription Factor staining Kit (Invitrogen) for 30 minutes on ice. After fixation, all wash steps pelleted cells at 750 g. Cells were washed twice with FoxP3 Kit permeabilization buffer, resuspended with 25 μL SM containing 10% mouse and 10% rat serum, and 25 μL SM containing 2x concentration of intracellular stain antibodies (CXCR5, T-bet, Bcl-6, and FoxP3) was added, and the mixture incubated overnight at 4 °C. Cells were then washed twice with FoxP3 Kit permeabilization buffer, resuspended in SM, and data was collected using BD Biosciences LSRII or Fortessa or Cytek Aurora. All wash steps used 500 g to pellet cells here and for all remaining sections unless otherwise specified. Antibodies used: CD4-BV605 clone GK1.5, BD Biosciences catalog number 563151, used at dilution 1:400. CD45R-PerCPCy5.5 clone RA3-6B2, Biolegend catalog number 103236, used at dilution 1:50. CD44-APCCy7 clone IM7, Biolegend catalog number 103028, used at dilution 1:200. PD-1-Biotin clone G4, grown and conjugated in house, used at dilution 1:400. SA (streptavidin)-BUV395, BD Biosciences catalog number 564167, used at dilution 1:200. FoxP3-FITC clone FJK-16s, Invitrogen/ThermoFisher catalog number 11-5773-82, used at dilution 1:400. CXCR5-PE clone 2G8, BD Pharmingen catalog number 551959, used at dilution 1:100. CD90.2-BV421 clone 53-2.1, Biolegend catalog number 140327, used at dilution 1:500, CD19-BUV395 clone 1D3, BD Biosciences catalog number 563557, used at dilution 1:200, CD38-Pacific Blue clone 90, conjugated in house, used at dilution 1:200, CD44-BV605 clone IM7, BD Biosciences catalog number 563058, used at 1:400, CD95-PECy7 clone Jo2, BD Biosciences catalog number 557653, used at 1:800. Quantification and statistical analysis for murine studies No statistical methods were used to pre-determine sample sizes, but our sample sizes are similar to those reported in our previous publications. Data collection and analysis were not performed blind to the conditions of the experiments. Mice were assigned to experimental groups based on genotype. The experiments were not randomized. For Salmonella infected mice, mice that showed poor infection as determined by >50x lower than average splenic bacterial CFU at necropsy compared to others in the group were removed. Data distribution was assumed to be normal, but this was not formally tested. Statistical significance was quantified with GraphPad Prism software (version 10) by two-tailed t-test or as indicated in each figure. A P-value <= 0.05 was considered significant, with the following symbols used to denote P-value ranges: p < =0.05(*), p < 0.01 (**), p < 0.001 (***), p < 0.0001 (****). Cell number was calculated per spleen using a combination of FACS analysis and trypan blue live cell counts as follows: the percent of cell of interest among live cells was determined by FACS, then the ratio of spleen processed for FACS analysis by weight as a portion of the total weight was used to calculate the total number of cells of interest per spleen. All flow cytometry data was analyzed using FlowJo version 10 (BD), spectral flow cytometry data was unmixed using SpectroFlo (version 3, Cytek Biosciences), and figures prepared using Adobe Illustrator version 2025. Supplementary information Supplementary Information Transparent Peer Review file Supplementary Information Transparent Peer Review file Source data Source Data Source Data