Computational and statistical analysis
Preprocessing & Clustering of scRNA-seq Data
1<sup>st</sup> Approach – Classical Methods on Combined Dataset
We began initial analysis following traditional clustering and annotation techniques; however, these methods using manual and at times subjective metrics scaled poorly to the size and scope of our dataset and moreover did not give clear distinction between disease specific cell states and compositional shifts within cell states across disease.
For our first pass at analysis, we grouped the FGID and pediCD datasets together (254,911 cells) and proceeded with the standard Seurat v3.1.5 pipeline (Stuart et al., 2019). We used manual heuristics of gene marker specificity to choose cluster resolution and isolate 9 major cell types (T, B, plasma, epithelial, endothelial, fibroblast, myeloid, mast cell, and glial) and 1 aggregate cluster of T, B, myeloid, and epithelial cells with a strong proliferation signature. We then subclustered the proliferating group and manually merged the proliferating cells with their corresponding cell type based on marker gene expression, and separately re-preprocessed and clustered each cell type annotating based on one vs. rest differential expression (Wilcoxon, fdr < 0.05) within the cell type.
We found several disadvantages to this approach. First, we found it difficult to determine for each cluster whether we should be looking for changes in compositional frequency or gene expression. Particularly within the myeloid major cell group we found extremely disease biased sub-clusters, as much as a 9:1 ratio between pediCD to FGID. It was unclear whether there was massive compositional shift within a conserved cell state or if instead a base cell state was split into multiple clusters based on phenotypic differences in disease and we should perform a differential expression test between it and neighboring FGID biased clusters. Second, after two rounds of manual clustering, some cell types could be clustered more deeply based on pre-existing knowledge, whereas other types that have been less well-resolved may have heterogeneity that has yet to be well-appreciated. Third, having partitioned over 100 distinct clusters, individually supervising each subsets processing and sub-clustering was infeasible. We needed a more systematic method to address these challenges.
2<sup>nd</sup> approach – ARBOL: Iterative Tiered Clustering (ITC) on Separated Disease Conditions
To be able to choose the appropriate future analyses and comparisons, we need a highly accurate representation of cell identity and state. The underlying issue in our first pass at clustering was that in combining the disease conditions together, the variable genes selected at each stage represented a combination of differences between cell identity & disease. This combination could have been manageable if either disease or cell identity were consistently more variable. We could isolate one factor at a specific tier in the hierarchy before sub-clustering to isolate the other. In our case, disease and cell identity both had many overlapping scales of variation. To address this effect, we isolated cell identity by separating our dataset by disease and clustering for cell identity within each disease set (FGID 115,569 cells, pediCD 139,342 cells). This approach required us to perform an additional stage of analysis to find corresponding clusters between the two datasets, but allowed us to more effectively distinguish type, scale, and specificity of disease differences.
Within each disease set we still needed a method to ensure we were reaching the bottom level of biological heterogeneity, and preferably an automated method as our first pass had shown the potential for isolating hundreds of cell states. To efficiently cluster and isolate these cell states we wrote a cloud-based pipeline to systematically optimize parameter selection and stop when biological heterogeneity is exhausted. Homogenous cell subsets were isolated by recursively normalizing, selecting variable genes, and clustering based on silhouette score. We stopped recursing into sub-clusters once we reached one of four end conditions defined as:
Having a group of less than 100 cells (though we did partition many clusters smaller than 100 cells after clustering groups just larger than that cutoff).
Isolating an optimized clustering of only one cluster.
Finding two clusters that have fewer than 25 genes (fdr< 0.05 & |log fold change| > 1.5 & percent expression >= 20% in at least one cluster) differentially upregulated between each cluster using a bimodal test developed in (Shekhar et al., 2016). For this last condition, if reached, we reject the clustering and return back the cells as a single end cluster.
Having reached a max tier limit. Setting this value to 10, we never triggered this condition with either FGID or pediCD datasets, but included it to prevent runaway recursion.
Code for generating this tree of cell clusters is currently available here: (https://jo-m-lab.github.io/ARBOL/ARBOLtutorial.html. Within each recursion, the established steps were processed using Seurat version 3.1.5 (https://github.com/satija.lab/seurat). Normalization and variable gene selection were processed with SCTransform (https://github.com/ChristophH/sctransform) (Hafemeister and Satija, 2019). Clustering for major cell types was performed using Louvain clustering on dimensionally reduced principal components.
Parameters depend on the size of the dataset, and thus must be adjusted based on how many cells are being partitioned for each recursive step. When calculating nearest neighbor graphs, and clustering we set the K parameter to ‘ceiling(0.5*sqrt(N))’. We chose the number of principal components based on the top 15 percentile of calculated improvement of variance explained. For subsets less than 500 cells we used Jackstraw to calculate significant principal components. If neither method succeeded, we chose the first two principal components. This only occurred rarely at late stages where Jackstraw was unable to find significance. We set clustering resolution via a grid search optimizing for maximum average silhouette score (silhouette measures the ratio of intra-cluster distance to inter-cluster distance, where a high score means highly distinct clusters). For stages where we were clustering more than 500 cells a randomized subsample of N cells / 10 was used to calculate the average silhouette score.
Additionally, at each recursive step we output quality metrics and basic plots, such as 1:rest differential expression from the optimal partitioning at each stage and UMAP representations painted by sample metadata (sample ID, cluster number). Our pipeline, saved output as a directory structure matching the tree discovered by this recursive clustering. This tree represents the lower levels of variance of discovered at each tier. At any tier level we are able to extract the cell’s partitioning. Due to the intermixing of patient and cell identity effects at multiple levels of the tree (a fraction of a single patient’s cells might separate out at a high level, but then continue to separate into identifiable cell types, or vice versa), we found the most meaningful levels at the top and bottom of the tree. The clustering tree is useful for understanding the levels of variance in the dataset, but we found it contains too much noise to be easily interpretable. Thus, we later generated a hierarchical clustering of the bottom level clusters based on pairwise differential expression, which is displayed in figures (Figure 3: FGID atlas, Figure 4: pediCD atlas). See the hierarchical clustering of cell subsets section for details.
Cell Type and Subset Annotation from Tiered Clustering
After running the iterative tiered clustering pipeline we manually curated the generated tree of clusters. Specifically, tree generation was reinitiated for the B Cells within the FGID dataset as it had stopped at the first tier on two clusters with < 50 genes differentially expressed, however we could see in this case that there was additional biological stratification based on strong differential expression of CXCR4 (Wilcoxon; logFC=1.22860917, Bonferroni.p=1.0E-300) CD69 (Wilcoxon; logFC=1.27527652, Bonferroni.p=2.99E-151), HMGN1 (Wilcoxon; logFC=-1.1688612, Bonferroni.p=1.62E-227) and HMGA1 (Wilcoxon; logFC=-1.28838294, Bonferroni.p= 1.06E-209) among others. This formed a clear divide between non-proliferating and proliferating B Cells, further validated by a clear separation within the UMAP based on PCA reduced variable genes within the B cells. We further examined each branching point of the tree to determine its splitting cause, noting splits based on spillover, doublet, and singular patient effects. Splits at higher tiers based on doublets often split again allowing us to recover cells that did not have the dual expression profile. Splits that only had patient splits below (measured by having only clusters of single patients) were manually marked as end clusters, thereby merging all clusters below that split. With these manual steps made, we performed pairwise differential expression to ensure each partitioned subset is distinct from its neighbors.
We annotated these final clusters with four methods attempting to balance descriptiveness, ease of understanding, and ease of name generation: The first method, is generated during the hierarchical tiered clustering by following the path from the end cluster up to the original tier. An example annotation is T0C0.T1C3.T2C3.T3C5 marking an end cluster that split at tier 1 into cluster 0 and at tier 2 into cluster 3. These annotations do not provide any biological information to the reader, but do provide a unique ID for the end cluster.
Our second method is far more descriptive, where we manually annotate the main reason for each particular split. As several cell types contained readily identifiable and meaningful cell subsets, we utilized curation of literature-based markers to provide further guidance within each cell type (Blériot et al., 2020; Cherrier et al., 2018; Dutertre et al., 2019; Guilliams et al., 2018; Robinette and Colonna, 2016). For example, within Tier 1 T cells, we could identify T cells, NK cells and ILCs; within Tier 1 myeloid cells, monocytes, cDC1, cDC2, macrophages and pDCs; within Tier 1 B cells, germinal center, germinal center dark zone and light zone cells; within Tier 1 endothelial cells, arterioles, capillaries, lymphatics, mural cells and venules; and so forth for other cell types. To illustrate this process for one cluster, upon automated hierarchical tiered clustering of T cells, we identified a cluster that was Tier 0: pediCD, Tier 1: T cells, Tier 2: cytotoxic, Tier 3: IEL_FCER1G_NKG7_TYROBP_CD160_AREG. Upon inspection of CD3 genes (CD247, CD3D, etc.), TCR genes (TRAC, TRBC1, etc.), and NK cell genes (NCAM1, NCR1), it became readily apparent these cells were NK cells (Supplemental Figure 6). This still follows the original ranking of variation as found by the hierarchical tiered clustering, while also providing biological interpretation, as an example: CD.Mloid.macrophage_chemokine.S100A8_S100A9_CXCL9_CXCL10_TNF_inflamonocyte.
This method of annotation was particularly useful during analysis as we were immediately able to see how early or late two clusters had split from each other, as well as seeing a number of the subset defining genes. Unfortunately, as is apparent this method also produces extremely long names that are difficult to display and to which to refer. It is also a highly manual process, and difficult to reproduce precisely. To better present our findings and aid others in reproducing our results, our third method automates this annotation. This method is performed by taking each major cell type, which in our case matched the tier one splits, and performing 1:rest differential expression testing (Wilcoxon; adj.p<0.05, only.pos=True) within each major cell type. We then ranked the genes based on the product of ‘-log(Bonferroni.p)’, ‘avg_logFC’, and ‘pct.exp.1/pct.exp.rest’ and took the top 5, forming a name like CD.Mloid_CCL3_CCL4_CCL3L3_TNF_TNFAIP6. This scheme, again was useful, but did not quite meet our demands of recognizability and brevity.
Rank-Score gene selection naming function
To account for genes that might be highly expressed in just a few cells we ranked the marker genes by a score combining their significance, the fold change in expression, and fold change of percent gene positive cells in the subset versus the percent of gene positive cells outside the subset. The collected metrics were multiplied together to provide a single score by which the genes were ranked: (-log(sig+1) * avg_logFC * (pct.in / pct.out)). For most subsets we selected the top 2 of these marker genes. For T/NK/ILC cells and myeloid cells we occasionally chose a slightly lower ranking gene from the top 10 if it was well supported and recognized by the literature. Thus, for T and Myeloid cells we adjusted these names to a finer degree of specificity by visualizing the expression profiles of each subset with a dotplot of canonical marker genes based off of current literature, and limiting to the top 2 genes based off our method 3 rankings and the dotplot of canonical markers, thereby producing the fourth and final annotations in the form: CD.Mono.CXCL10.TNF. Due to the limited nature of current characterization of stromal and epithelial cells we were unable to match the same degree of specificity as the T and Myeloid cells, however we did where possible adjust from the major cell type, to the most specific that we could be confident of. For instance, adjusting “Epith” to “Goblet” based on marker expression of TFF3 and MUC13.
Hierarchical Clustering of Subsets on Unified Gene Space and Removal of Doublets
At this point we had generated a hierarchical representation of the datasets from the top down showing the splits of highest variation at every level. By necessity that means that each level is controlled by and represents different selections of genes, which may have no relation to the genes selected in another branch. To understand the relations of cell subsets and compare across cell type we needed a unified set of genes. For each dataset (FGID and pediCD) we performed pairwise differential expression (Wilcoxon; Bonferroni.p<0.01, max.cells.per.ident=500) and selected the top 50 most significant genes from each test. Gene lists were merged as a union, finding 4445 unique genes for FGID and 1844 unique genes for pediCD that best differentiate the subsets. Subset centers were calculated from these selected genes as the median expression of cells grouped by subset. The resulting table was then hierarchically clustered using correlation distance and complete linkage. Clustering was performed in R using the pvclust package (https://github.com/cran/pvclust).
The resulting tree shows from the bottom up the relationships between cell subsets, and allows cell subsets that were potentially misclassified at a high split in hierarchical tiered clustering to find their biological neighboring subsets. As previously mentioned within our description of hierarchical tiered clustering we did not find any end cluster subsets that met our thresholds for merging. This does not mean that we did not observe shuffling from the initial tiered splits. While overall there was good agreement between the two methods, we noted subsets jumping between major cell types as defined by the first splits of our tiered clustering. We identified the majority of these jumping subsets as doublet clusters by exploring their differential gene results at multiple levels of the tiered clustering tree. We removed these doublet subsets and others based on flipping expression programs at different tiers. For instance, looking like T cells expressing TRAC, IL7R within an epithelial cluster, then at the next tier having significantly higher expression of KRT18 and PIGR than neighboring clusters. After removing doublets, we recalculated subset distances and dimensional reductions, as presented in the main figures.
FGID Atlas Generation and Curation
Within B cells, we identified a strong division between non-cycling and cycling B cells, with those found in the cycling compartment readily identifiable by germinal center markers and further dark zone (AICDA) and light zone (CD83) genes resulting in FG.B/DZ.AICDA.IGKC and FG.B/LZ.CD74.CD83 clusters (Figure 3d) (Victora et al., 2010).
Within myeloid cells, we identified, and confirmed using extensive inspection of literature curated markers, cell subsets corresponding to monocytes (CD14, FCGR3A, FCN1, S100A8, S100A9, etc.), macrophages (CSF1R, MERTK, MAF, C1QA, etc.), cDC1 (CLEC9A, XCR1, BATF3), cDC2 (FCER1A, CLEC10A, CD1C, IRF4 etc.), and pDCs (IL3RA, LILRA4, IRF7) (Figure 3d; Supplemental Table 6) (Blériot et al., 2020; Dutertre et al., 2019; Guilliams et al., 2018). We highlight selected cell states including a migratory dendritic cell state (FG.DC.CCR7.FSCN1), extensive cDC2 heterogeneity relative to cDC1 heterogeneity, and a main distinction between macrophage clusters expressing C1Q*, MMP*, APOE, CD68 and PTGDS (FG.Mac.C1QB.SEPP1, FG.Mac.APOE.PTGDS) and a series of clusters expressing various chemokines including CCL3, CXCL3,and CXCL8 (FG.Mac.CCL3.HES1, FG.Mac.CXCL3.CXCL8, FG.Mac.CXCL8.IL1B).
Within T cells, we followed a similar approach as utilized for Myeloid cells and identified principal cell subsets of T cells (joint expression of CD247, CD3D, CD3E, CD3G with TRAC, TRBC1, TRBC2, or TRGC1, TRGC2 and TRDC), and a combined cluster of cytotoxic cells (FG.T/NK/ILC.GNLY.TYROBP) likely including T cells, NK cells (lower expression of TCR-complex genes with NCAM1, NCR1 and TYROBP), and some ILCs (KIT, NCR2, RORC and low expression of CD3-complex genes) (Figure 3d, Supplemental Table 6) (Cherrier et al., 2018; Robinette and Colonna, 2016). We note that the numerical majority of CD4 T cells (FG.T/NK/ILC.MAF.RPS26) and CD8 T cells (FG.T/NK/ILC.CCR7.SELL) expressed SELL and CCR7 thus identifying them as naïve T cells. However, regardless of clusters expressing CD4 (FG.T.GZMK.GZMA) or CD8A/CD8B (FG.T.GZMK.IFNG, FG.T.GZMK.CRTAM, etc.), most activated T cells were characterized by expression of granzymes (Sallusto et al., 1999).
Within epithelial cells, most cells expressed high levels of OLFM4, identifying them as crypt-localized cells (Moor et al., 2018). We readily identified subsets of stem cells (LGR5), proliferating cells (TOP2A), goblet cells (SPINK4, ZG16, various MUCs), enteroendocrine cells (SCG3, ISL1), Paneth cells (ITLN2, PRSS2, LYZ), tuft cells (GNG13, SH2D6, TRPM5) and enterocytes (APOC3, APOA1, FABP6, etc.) (Figure 3d, Supplemental Table 6) (Barker et al., 2007; van der Flier and Clevers, 2009).
Within endothelial cells, we readily identified vascular and lymphatic endothelial cells (LYVE1, PROX1), with the vascular cells able to be further identified as capillaries (CA4) or venular endothelial cells (ACKR1, MADCAM1) (Brulois et al., 2020). We also identified a subset of cells (FG.Endth/Peri.FRZB.NOTCH3) expressing high levels of FRZB and NOTCH3, which, rather than being arterioles, likely represent arteriole-associated pericytes or smooth muscle cells given the absence of EFNB2, SOX17, BMX, and HEY1, and the presence of ACTA2 and MYL9, as cluster-defining genes (Figure 3d, Supplemental Table 6) (Travaglini et al., 2020; Whitsett et al., 2019). We highlight that the FG.Endth/Ven.ACKR1.MADCAM1 cluster is characterized by expression of markers for postcapillary venules specialized in leukocyte recruitment (Thiriot et al., 2017).
Within fibroblasts, we identified principal subsets characterized by their structural roles (COL3A1, ADAMDEC1, FBLN1, LUM, etc.), myofibroblasts (MYH11, ACTA2, ACTG2, etc.), and organization of lymphoid cells (CCL19, CCL21 etc.) (Figure 3d, Supplemental Table 6) (Buechler et al., 2021; Davidson et al., 2021). Within the lymphoid-organizing fibroblasts, we draw attention to the FG.Fibro.C3.FDCSP, FG.Fibro.CCL19.C3, and FG.Fibro,CCL21.CCL19 subsets, which appear to have some characteristics of follicular dendritic cells and variable expression of CCL19/CCL21 (T-cell or migratory dendritic cell chemoattractants) and CXCL13 (B-cell chemoattractant) (Das et al., 2017; Heesters et al., 2013). We also identified a separate Tier 1 cluster of glial cells characterized by CRYAB and CLU. Intriguingly within the glial cell Tier 1 cluster, we then recovered a cell subset expressing FDCSP, CXCL13, and CR2, a key complement receptor which allows for complement-bound antigens to be recycled and presented by follicular dendritic cells (Das et al., 2017; Heesters et al., 2013). This highlights the power of iterative tiered clustering to recover discrete cell states that may, through the process of traditional clustering, not be fully resolved. Furthermore, the presence of these discrete cell clusters within larger parent cell clusters will alter the gene expression signatures of the higher-level cell types. For example, the FG.Glial/fDC.FDCSP.CXCL13 in the hierarchical cluster tree then assorts within the lymphoid-organizing stromal cells.
The mast cells recovered did not further sub-cluster in an automated fashion, and were largely marked by TPSB2 and TPSAB1 (>97%), with minimal CMA1 (<20%) expressing cells, suggesting they are largely classical MC-T cells in FGID intestine (Dwyer et al., 2021).
We identified four Tier 1 clusters for plasma cells, which are characterized by their strong expression of IGH* immunoglobulin heavy-chain genes together with either an IGK* (kappa light chain) or IGL* (lambda light chain) genes (Cyster and Allen, 2019; James et al., 2020). This resolved IgA IgK plasma cells, IgA IgL plasma cells, IgM plasma cells, and IgG plasma cells. Iterative tiered clustering identified further heterogeneity within all clusters of IgA and IgG plasma cells, though given the 3’-bias of this dataset, we note that a principled investigation of these clusters would ideally use 5’ sequencing with targeted VDJ amplification.
Together, our treatment-naïve cell atlas from 13 FGID patients captures 138 cell clusters from a non-inflammatory state of pediatric ileum which we annotated and named in a principled fashion.
We note that p044 was overrepresented with more terminally differentiated epithelial cells, likely from incomplete EDTA separation, and thus omit the p044 unique cell clusters from further analyses of composition.
pediCD Atlas Generation and Curation
Within B cells, we also identified a strong division between non-cycling and cycling B cells, with those found in the cycling compartment readily identifiable by germinal center markers and further dark zone (AICDA) and light zone (CD83) genes, as in FGID (Victora et al., 2010). Within cells expressing germinal centers markers, a highly-proliferative branch including clusters such as CD.B/LZ.CCL22.NPW, CD.B/GC.MKI67.RRM2, and CD.B/DZ.HIST1H1B.MKI67 emerged (Figure 4d). The CD.B/LZ.CCL22.NPW was characterized by high levels of MYC, which has been shown to allow for further rounds of germinal center affinity maturation (Dominguez-Sola et al., 2012). More numerous B cell clusters included ones characterized by expression of GPR183, such as CD.B.CD69.GPR183 (also expressing IGHG1) and CD.B.RPS29.RPS21. GPR183 has been shown to regulate the positioning of B cells in lymphoid tissues (Pereira et al., 2009).
Within myeloid cells, we identified, and confirmed using the same extensive inspection of literature curated markers as in FGID, cell subsets corresponding to monocytes (CD14, FCGR3A, FCN1, S100A8, S100A9, etc.), macrophages (CSF1R, MERTK, MAF, C1QA, etc.), cDC1 (CLEC9A, XCR1, BATF3), cDC2 (FCER1A, CLEC10A, CD1C, IRF4 etc.), and pDCs (IL3RA, LILRA4, IRF7) (Figure 4d; Supplemental Figure 6, Supplemental Table 9) (Blériot et al., 2020; Dutertre et al., 2019; Guilliams et al., 2018). We highlight selected cell states including a migratory dendritic cell state (CD.DC.CCR7.FSCN1), extensive cDC2 heterogeneity relative to cDC1 heterogeneity, and a main distinction between macrophages expressing C1Q*, MMP*, APOE, CD68 and PTGDS, (CD.Mac.APOE.PTGDS) and a series of clusters expressing various chemokines including CXCL2, CXCL3, and CXCL8 (CD.Mac.SEPP1.CXCL3, CD.Mono.CXCL3.FCN1, CD.Mono.CXCL10.TNF). Several of the end cell clusters initially clustering with macrophages also expressed monocyte markers (S100A8, S100A9), and expressed detectable, but lower levels of MERTK or AXL relative to bona fide macrophages, potentially indicative of the early stages of the trajectory of monocyte-to-macrophage differentiation (Blériot et al., 2020; Dutertre et al., 2019; Guilliams et al., 2018). We also noted a substantial expansion of clusters characterized by expression of CXCL9, CXCL10, and STAT1, canonical interferon-stimulated genes, observed in clusters such as CD.Mono/Mac.CXCL10.FCN1 (Ziegler et al., 2021, 2020). Moreover, we identified a cluster of inflammatory monocytes, CD.Mono.S100A8.S100A9, characterized by both CD14 and FCGR3A expression.
Within T cells, we followed a similar approach as utilized for FGID T cells and identified cell subsets of T cells (joint expression of CD247, CD3D, CD3E, CD3G with TRAC, TRBC1, TRBC2, or TRGC1, TRGC2 and TRDC), but in pediCD also identified several discrete clusters of NK cells (lower expression of TCR-complex genes with FCGR3A or NCAM1, NCR1 and TYROBP), and ILCs (KIT, NCR2, RORC and low expression of CD3-complex genes) (Figure 4d, Supplemental Figure 6, Supplemental Table 9) (Cherrier et al., 2018; Robinette and Colonna, 2016). We note that T cells and NK cells with a shared expression of GNLY, GZMB and other cytotoxic effector genes cluster almost indistinguishably from each other through iterative tiered clustering and visualization of the hierarchical tree, but that careful inspection of literature-curated markers helped resolve NK cells (CD.NK.CCL3.CD160; CD.NK.GNLY.GZMB) from CD8A/CD8B T cells (CD.T.GNLY.GZMH; CD.T.GNLY.CTSW) (Figure 4d, Supplemental Figure 6, Supplemental Table 9) (Cherrier et al., 2018; Robinette and Colonna, 2016). One of the specific challenges in distinguishing between T cells and NK cells in scRNA-seq data is that NK cells can express several CD3-complex genes, particularly CD247, as well as detectable aligned reads for TRDC or TRBC1 and TRBC2, and thus lower-resolution clustering approaches or datasets with lower cell numbers may miss these important distinctions (Björklund et al., 2016; Renoux et al., 2015). NK cell clusters also expressed the highest levels of TYROBP, which encodes DAP12 and mediates signaling downstream from many NK receptors (French et al., 2006; Lanier, 2001; Lanier et al., 1998). ILC clusters such as CD.ILC.LST1.AREG or CD.ILC.IL22.KIT were characterized by an apparent ILC3 phenotype, with expression of KIT, RORC and IL22, though they also expressed detectable transcripts of GATA3 in the same clusters (Cherrier et al., 2018; Robinette and Colonna, 2016). We detected several clusters expressing CD4 and lacking CD8A/CD8B, including regulatory T cells (CD.T.TNFRSF18.FOXP3), and MAF- and CCR6-expressing helper T cells (CD.T.MAF.CTLA4). Perhaps most strikingly, we resolved multiple subsets of proliferating lymphocytes, including regulatory T cells (CD.T.MKI67.FOXP3), IFNG-expressing T cells (CD.T.MKI67.IFNG), and NK cells (CD.NK.MKI67.GZMA).
Within epithelial cells we identified substantial heterogeneity in CD. Most cells expressed high levels of OLFM4, identifying them as crypt-localized cells (Moor et al., 2018). We readily identified subsets of stem cells (LGR5), proliferating cells (TOP2A), goblet cells (SPINK4, ZG16, various MUCs), enteroendocrine cells (SCG3, ISL1), Paneth cells (ITLN2, PRSS2, LYZ), tuft cells (GNG13, SH2D6, TRPM5) and enterocytes (APOC3, APOA1, FABP6, etc.) (Figure 4d, Supplemental Table 9) (Barker et al., 2007; van der Flier and Clevers, 2009). Amongst several clusters characterized by CCL25 and OLFM4 expression, we identified a subset marked by LGR5 expression, characteristic of intestinal stem cells (CD.EpithStem.LINC00176.RPS4YA1) (Barker et al., 2007). We identified several subsets expressing CD24, indicative of crypt localization, with expression of REG1B (CD.Secretory.GSTA1.REG1B; CD.Secretory.REG1B.REG1A) (Moor et al., 2018). We also identified early enterocyte cluster CD.EC.ANPEP.DUOX2, characterized by FABP4 and ALDOB and expressing DUOX2 and MUC1. We resolved several clusters of enteroendocrine cells, including CD.Enteroendocrine.TFPI2.TPH1 and CD.Enteroendocrine.NEUROG3.MLN. We also found two clusters we labeled as M cells based on expression of SPIB (CD.Mcell.CCL23.SPIB; CD.MCell.CSRP2.SPIB) (Beumer et al., 2020; Mabbott et al., 2013). Paneth cells did not further sub-cluster despite forming an independent Tier 1 cluster (CD.Epith.Paneth). Most strikingly, we identified a diversity of goblet cells recovered across multiple patients including CD.Goblet.HES6.COLCA2 expressing REG4 and LGALS9, and CD.Goblet.TFF1.TPSG1 expressing TFF1 and ITLN1, amongst others. We also identified a cluster of Tuft cells: CD.EC.GNAT3.TRPM5.
Within endothelial cells, we also readily identified vascular and lymphatic endothelial cells (LYVE1, PROX1), with the vascular cells able to be further identified as capillaries (CA4) or venular endothelial cells (ACKR1, MADCAM1) (Figure 4d, Supplemental Table 9). We also identified a subset of cells (CD.Endth/Mural.HIGD1B.NDUFA4L2) expressing high levels of FRZB and NOTCH3, which, rather than being arterioles, likely represent arteriole-associated pericytes or smooth muscle cells given the absence of EFNB2, SOX17, BMX, and HEY1, and the presence of ACTA2 and MYL9, as cluster-defining genes. In pediCD, we also identified a cluster of arteriolar endothelial cells, CD.Endth/Art.SEMA3G.SSUH2, identified by expression of HEY1, EFNB2, and SOX17. We also highlight that the endothelial venules characterized by expression of markers for postcapillary venules specialized in leukocyte recruitment, such as CD.Endth/Ven.ADGRG6.ACKR1 and CD.Endth/Ven.POSTN.ACKR1, exhibited greater diversity than in FGID with multiple end cell clusters identified (Thiriot et al., 2017).
Within fibroblasts, we identified principal subsets characterized by their structural roles (COL3A1, ADAMDEC1, FBLN1, LUM, etc.), myofibroblasts (MYH11, ACTA2, ACTG2, etc.), and organization of lymphoid cells (CCL19, CCL21 etc.) (Figure 4d) (Buechler et al., 2021; Davidson et al., 2021). The principal hierarchy in fibroblasts in pediCD was between FRZB-, EDRNB- and F3-expressing subsets such as CD.Fibro.LY6H.PAPPA2 and CD.Fibro.AGT.F3, which were also enriched for CTGF and MMP1 expression, and ADAMDEC1-expressing fibroblasts, which were enriched for several chemokines such as CXCL12, and in some specific clusters CXCL6, CXCL1, CCL11, and other chemokines. Amongst three fibroblast subsets marked by C3 expression, we identified follicular dendritic cells (CD.Fibro/fDC.FCSP.CXCL13), along with fibroblasts expressing CCL21, CCL19, and the interferon-stimulated chemokines CXCL9 and CXCL10 (CD.Fibro.CCL21.CCL19; CD.Fibro.TNFSF11.CD24) (Das et al., 2017;
Heesters et al., 2013). Distinct from the FGID atlas, within the pediCD atlas, glial cells clustered within fibroblasts, but were also marked by S100B, PLP1 and SPP1 expression.
The mast cells recovered in pediCD did further sub-cluster in an automated fashion, were largely marked by TPSB2 (>90%), with minimal CMA1 (<16%) expressing cells, suggesting they are largely classical MC-T cells in pediCD intestine (Figure 4d) (Dwyer et al., 2021). Intriguingly, some subsets (CD.Mstcl.AREG.ADCYAP1) were enriched for IL13-expression. We also detected a small cluster of proliferating mast cells from several patients (CD.Mstcl.CDK1.KIAA0101).
We also identified four Tier 1 clusters for plasma cells, which are characterized by their strong expression of IGH* immunoglobulin heavy-chain genes together with either a IGK* (kappa light chain) or IGL* (lambda light chain) genes. This resolved IgA IgK plasma cells, IgA IgL plasma cells, IgM plasma cells, and IgG plasma cells. Iterative tiered clustering identified further heterogeneity within all clusters of IgA plasma cells, though given the 3’-bias of this dataset, we note that a principled investigation of these clusters would ideally use 5’ sequencing with targeted VDJ amplification.
Together, our treatment-naïve cell atlas from 14 pediCD patients captures 305 cell clusters from an inflammatory state of the pediatric ileum suggesting an increase in the number and diversity of cell states present in the intestine during overt inflammatory disease.
Association of Cell Subsets to anti-TNF Response
Compositional differences are an important metric for understanding the baseline differences that prognose a patent’s response to treatment. We measure these differences with proportional enrichment of particular cell subsets within each patient, and finding the significantly reproducible enrichments across disease. As an extreme toy example, we might find that subset A cells comprise as much 80% of cells sampled in one condition whereas they might only comprise 30% in a different condition. This type of compositional analysis is highly affected by the number and choice of subsets included, and the sampling depth per patient (how may cells are collected). The first factor is controlled by the confidence in our clustering and using computationally optimized parameters. We further control this factor by limiting analysis of compositional shifts of cell states to within major cell types. This isolates the chance of error from affecting the entire analysis and allows us to gain a more direct biological insight of the rise and fall of particular cell states in the context of similar subsets. We control the second factor of sampling depth differences by computing a normalized cell count score per patient of the form (ncells in subset / ncells in patient’s major cell type) * 1e6. This score provides us with the number of cells expected per million.
Mann-Whitney tests
We input our cells per million score into a two-sample Wilcoxon test in base R, which is equivalent to the Mann-Whitney rank score test. We set a significance threshold of p_value < 0.05. We made 5 different pairwise comparisons (FGID vs FR, FGID vs PR, NOA vs FR, NOA vs PR, FR vs PR). Comparisons between FGID and pediCD groups were determined by finding maximum correspondence between the disease conditions for each subset.
Fisher’s Exact tests
A similar compositional analysis to that done with the Mann-Whitney was performed with a Fisher’s Exact test. We input for each subset the number of cells for that subset against the number of cells not of that subset within the major cell type split on rows by pairwise comparisons (NOA vs FR, NOA vs PR, FR vs PR). We computed FDR correction of p_values at major cell type and entire dataset levels and found significance subsets at both levels. But, most interestingly in comparing the two tests we found that the Mann-Whitney discovered as significant (pval < 0.05) the portion of cell subsets with largest effect sizes. Understanding our limited patient number at these within pediCD comparisons and wanting to only report results most likely to be reproducible biology, we determined to only follow those subsets reported as significant within both Mann-Whitney and Fisher’s exact tests.
Discrete cell cluster changes across the pediCD clinical severity and response spectrum
When comparing FR/PRs to NOAs, two subsets with significantly increased frequency in FR/PR patients amongst T cells, NK cells, and ILCs were identified. These were CD.NK.MKI67.GZMA and CD.T.MKI67.IL22 (Figure 5c; Supplemental Figure 7b; Supplemental Table 15). Beyond the strong proliferation signature, CD.NK.MKI67.GZMA were enriched for genes such as GNLY, CCL3, KLRD1, IL2RB and EOMES, and CD.T.MKI67.IL22 were enriched for IFNG, CCL20, IL22, IL26, CD40LG and ITGAE. This indicates that with increasing pediCD clinical severity, there is increasing local proliferation of cytotoxic NK cells and proliferation of tissue-resident T cells with the capacity to express anti-microbial and tissue-reparative cytokines, and molecules to interface with antigen-presenting cells and B cells. Alongside this increase, there was a significant decrease amongst fibroblasts of CD.Fibro.CCL19.IRF7, and amongst epithelial cells of CD.EC.SLC28A2.GSTA2 clusters in the FR/PR patients compared to NOA (Supplemental Figure 7b). The CD.Fibro.CCL19.IRF7 were enriched for CCL19, CCL11, CXCL1, CCL2, and very specifically for OAS1 and IRF7. The CD.EC.SLC28A2.GSTA2 cluster was characterized by its two namesake markers, involved in purine transport and glutathione metabolism (Moor et al., 2018).
We also detected significant decreases in FRs relative to NOAs in certain cell types, particularly within Epithelial cells including CD.EpithStem.LINC00176.RPS4Y1, CD.MCell.CSRP2.SPIB, CD.EC.FABP6.PLCG2, and CD.EC.FABP1.ADIRF (Supplemental Figure 7c; Supplemental Table 15). We note that the relative decrease in M cells is in stark contrast to the “ectopic” M-like cells that were detected in adult ulcerative colitis (Smillie et al., 2019).
We next focused on those cell subsets that were significantly changed only between PRs and NOAs (Figure 5c; Supplemental Figure 7d; Supplemental Table 12). Here we note several distinct clusters within the lymphocyte cell type, including increases in the PR patients compared to NOA patients of CD.T.MKI67.IFNG, CD.T.MKI67.FOXP3, CD.T.GNLY.CSF2, and CD.NK.GNLY.FCER1G. The two MKI67 clusters again highlighted an increase in proliferative cells, specifically cells enriched for IFNG, GNLY, HOPX, ITGAE and IL26 (CD.T.MKI67.IFNG), and IL2RA, BATF, CTLA4, TNFRSF1B, CXCR3, and FOXP3 (CD.T.MKI67.FOXP3), the latter of which may be indicative of proliferating regulatory T cells. The two GNLY clusters emphasized cytotoxicity as they were both enriched for GNLY, GZMB, GZMA, PRF1 and more specifically for IFNG, CXCR6, and CSF2 (CD.T.GNLY.CSF2), or AREG, TYROBP, and KLRF1 (CD.NK.GNLY.FCER1G). Amongst myeloid cells, there was an increase in CD.Mac.CXCL3.APOC1, CD.Mono/Mac.CXCL10.FCN1, and CD.Mono.FCN1.S100A4 in PR versus NOA. The CD.Mac.CXCL3.APOC1 cluster was enriched for a variety of chemokines including CCL3, CCL4, CXCL3, CXCL2, CXCL1, CCL20, and CCL8. It was also enriched for TNF and IL1B. The CD.Mono/Mac.CXCL10.FCN1 cluster was enriched for CXCL9, CXCL10, CXCL11, GBP1, GBP2, GBP4, GBP5, suggestive of activation by IFN, and more specifically Type II IFNγ, based on the GBP gene cluster (Ziegler et al., 2020). CD.Mono.FCN1.S100A4 was characterized by S100A4, S100A6, and FCN1 expression. These two immune clusters were paralleled by increases in certain clusters within endothelial cells in PR versus NOA patients (CD.Endth/Ven.LAMP3.LIPG) and epithelial cells (CD.Goblet.TFF1.TPSG1).
Several clusters of cells were decreased in PR versus NOA, including CD.T.LAG3.BATF, CD.T.IFI44L.PTGER4, and CD.T.IFI6.IRF7 amongst lymphocytes (Supplemental Figure 7d) (Roncarolo et al., 2018). Amongst myeloid cells, CD.cDC2.CLEC10A.FCGR2B were decreased, and amongst fibroblasts CD.Fibro.IFI6.IFI44L were decreased. In epithelial cells, CD.Tuft.GNAT3.TRPM5 cells were decreased. Alongside the decrease in Tuft cells amongst epithelial cells, two more clusters closely related to the aforementioned CD.EC.GSTA2.SLC28A3 cluster, also marked by GSTA2 expression, were significantly decreased (CD.EC.GSTA2.CES3, and CD.EC.GSTA2.TMPRSS15).
We assessed the compositional differences between FRs and PRs and only identified one cell cluster which was significantly increased in PRs: CD.B/DZ.HIST1H1B.MKI67, which are proliferating dark zone B cells. CD.T.EGR1.TNF T cells were significantly decreased in PR versus FR (Supplemental Figure 7e; Supplemental Table 15). These data suggest that at the time of diagnosis of pediCD, there are a series of changes in multiple cell types that encapsulate the distinctions between NOA and FR or PR patients.
We provide an example of specific tiers, subclusters, and modules of co-expressed genes with proliferating T cells as a case study. We compare our approach of iterative clustering within a cell state (such as proliferating T cells), and then testing for significant differences in the composition between disease severity groups (Supplemental Figure 8a) vs. performing differential expression within the proliferating T cell state between disease severity groups (Supplemental Figure 8b). Importantly, differential expression at this level fails to recover the critical substructure present in our dataset (Supplemental Figure 8a,b). This is further validated through the additional use of topic modeling: an orthogonal method to clustering for cell program identification (Supplemental Figure 8c) (Bielecki et al., 2021). For example, while GZMA alone is not differentially expressed, the cell subset CD.NK.MKI67.GZMA proliferating cytotoxic NK cells, identified by ARBOL and Topic modeling, is significantly increased by compositional (i.e. percentage of the parent cell subset, as typically performed in flow cytometry) analysis with increasing disease severity (Figure 5c). This highlights the need to examine how co-variation in multiple genes defines cell clusters beyond traditional differential expression at a coarser level of clustering(Shalek et al., 2014, 2013).
Principal component analysis of cell frequencies and correlation to clinical metadata
Cell frequencies were calculated per patient for cell subsets (i.e. end clusters) within parent cell types and cell subsets (i.e. end clusters) within all cells as CPM = ((count/sum(count)) * 1e6.
Principal component analysis (PCA) was performed on the resulting patient x CPM matrices using the R package stats::prcomp(., scale=TRUE). Variance explained per PC was calculated as std^2/sum(std^2). PCA loadings per patient and per cell subset were extracted from the prcomp() result. PCA1 and PCA2 from the total PCA x patient and from each celltype’s PCA x patient matrix were correlated with clinical metadata using Spearman rank correlation as calculated by the R package stats::cor.test(., method=’spearman’). P values were recorded from the cor.test() call, and FDR was calculated using R’s fdrtool::fdrtool(p.values, statistic=”pvalue”). For combined celltype PCA’s, patient x CPM tables were concatenated before PCA.
Finding Corresponding Cell Subsets Between Diseases
Separating the data on disease condition into two datasets was important as it allowed us to isolate the axis of cell identity within each disease and be confident in the homogeneity of each subset. But does present the problem of how to make comparisons across the disease condition:
1<sup>st</sup> Approach – KNN classifier
Our first attempt to find corresponding clusters followed the methods of Tasic et al. 2018. We used the best differentiating genes sets created for our unified gene space clustering to as the mapping space for a nearest-neighbor classifier. For each cell within the disease condition, we could map it to the nearest cell subset within the other disease condition. As a trial run, we created this gene space for each major cell type of the FGID disease condition and performed 5-fold cross-validation. Unfortunately, we could only achieve accuracies of up to 55%.
Rethinking approach
From this trial run we realized that we needed a more automated system to choose genes as the most significantly differentially expressed genes did not create enough separation between cluster centers to effectively classify new cells. We chose to use an RF classifier as it allowed us to train for the optimal selection of genes, required little to no preparation of data, and provided probabilities of each cell being predicted to each class. These probabilities for each class proved particularly useful do to our second realization. Because the number of subsets differs between disease conditions, we could not assume that there was a one-to-one relationship between conditions. Nor could we assume that the many-to-one relationships were unidirectional with one base subset splitting into many states only from FGID toward CD. A single classifier would not allow us to distinguish between these many types of relationships. However, we realized that by creating a classifier for both directions (FGID to pediCD & pediCD to FGID) we could take advantage of the difference in confidence between the two classifications to discover the direction and type of relationship. In a perfect world of 1-to-1 relationships, we would expect all cells of subset A in condition X to match with 100% confidence to subset B in condition Y. In that particular case the summed probability equal 2 and there would be zero difference in confidence of one classifier to its matching classifier. In our imperfect world we might instead see 90% of cells of subset A in condition X to matching with > 85% confidence to subset B in condition Y, and only 30% of subset B in condition Y matching with > 85% confidence. From this discrepancy we can infer that subset A may be a cell state in condition X that is layered on top of a base state B in condition Y. Low confidence in both directions indicates subsets unique to a particular condition.
2<sup>nd</sup> Approach – Training Random Forest Model
We trained these models in scikit-learn with 5-fold cross validation and params: min_samples_leaf=1, oob_score=True, criterion=“gini”, max_depth=200, n_estimators=700, max_features=“sqrt” The training set (but not the test set) was sampled with replacement such that all classes contained as many samples as the maximum proportioned class. This up-sampling procedure provided the largest gain to our test accuracy, sensitivity, and specificity scores, increasing accuracy ∼7-10% across each cell type.
We trained random forest classifiers for each cell type in each disease condition using SciKit-Learn v0.22.2, with the intent to classify each cell to the subset in the opposed dataset the cell is most similar to (Pedregosa et al., 2011). For each cell type we optimized a classifier for accuracy using grid-based search tuning number of trees, depth, number of features, criterion, and min samples per leaf with 5-fold cross validation for each set of tuning parameters. We never observed full overfitting where the accuracy on test folds began to drop with increased size of model, but we did quickly find diminishing returns as we increased model size. For simplicity and because optimal tuning parameters were robust to overfitting, we chose to use the same largest model parameters for all models (number of trees = 500, depth = 200, number of features = sqrt, criterion = gini, min samples per leaf = 1, oob_score=True) (Pedregosa et al., 2011). Our initial training rounds found accuracies in the mid 60%. A definite increase from the NN classifier, but not high enough for us to be confident in the results. The main issue we eventually determined to be the uneven class distributions (far more cells in subset A than subset B). This caused the smaller subsets to be under trained. To compensate we up-sampled with replacement each subset within the training fold to contain at least the 75th quantile number of cells. This single change improved accuracy on the unmodified test fold the most, on average providing a ∼7-10% improvement of accuracy, precision and recall across each cell type, and provided accuracies ranging from high 70 to low 90 percent per major cell type.
Applying Random Forest Model
Satisfied with the validation results within each disease condition, we ran the random forest model across the disease conditions. We trained each random forest model with optimized parameters on all folds of its dataset, then proceeded to get probability predictions for each cell from the disease condition to the trained disease condition. With these class probabilities per cell we could aggregate for each disease condition by taking the mean class probabilities for each group, leaving us with 2 n by m table where n equals the number of subset groups and m equals the number of subset classes in the opposing disease condition. Using the mean probabilities for the group allowed more information from the cell level to rise to the aggregated levels than using the individual class prediction alone (computed as the class with max confidence of cell membership). These tables also provide confidences to all classes which is important for understanding the transverse confidence in both directions.
It is especially important to understand the many-to-one relationships between disease conditions and find where a base cell state becomes layered in additional expression profiles, as these are the exact cases where we can infer the underlying signaling patterns that diversify or concentrate cell state profiles. In diverse splitting of a subset across disease we can start to understand the heterogeneity of patient response to treatment as it becomes clear which particular cell profiles are correlated with strong and poor response. To gain insight to these changes, we care about where there is strong confidence in both directions and where there is strong confidence in only one direction. The simplest method to calculate these is to separately take the sum of the pairwise prediction confidences and the difference. We call the sum of confidences the correspondence of a subset, and the difference the bias.
Visualizing Correspondence and Bias
We plot these metrics on a dot plot where each possible connection is laid out on a grid. For each dot we set the size to match the correspondence, and color the dot based on the bias, such that a perfect match would appear as a large white circle. A more unidirectional match would be tinted darker in the color matching the disease condition with more confidence. Matches with more bias tend to indicate a subset matching a base cell state but also expressing some additional gene modules. To aid the human eye on picking up the major patterns we filter to only show the top 10% highest correspondences. This parameter was chosen after looking at the distribution of correspondence scores and selecting the majority of the right tail of the distribution. It keeps the strongest matches in both ways and keeps the strongest in highly biased matches. To also aid the human eye we perform a hierarchical clustering using cosine distance and complete linkage on the prediction confidences and compute an optimal ordering based on the cosine distances using the “cba” package in R: https://cran.r-project.org/web/packages/cba/index.html. This allows us to sort subsets on the rows and columns such that subsets that get predicted similarly are next to each other. From this visualization we are able to easily discern which are the subsets FGID that split into many phenotypes within pediCD from high correspondence and bias, which subsets don’t change phenotype much at all based on high correspondence and low bias, and which are the subsets are potentially unique to a disease condition based on very low correspondence and bias.
Random Forest classifiers relative to joint clustering and integration approaches
While integration methods continue to improve (Hao et al., 2021; Hie et al., 2019; Korsunsky et al., 2019; Pliner et al., 2019), even the latest methods are still benchmarked largely on broad cell type or subset integration, and often struggle (<50% classification accuracy) with fine cell states (Sikkema et al., 2022). Thus, we employed a random forest (RF) classifier-based approach, which has been applied successfully in work to identify correspondence in fine sub-clusters in the mammalian retina (Peng et al., 2019; Shekhar et al., 2016). Specifically, we employed paired RF models (one trained on FGID, the other trained on pediCD) to obtain cross dataset predictions per cell using an up-sampling procedure that enhanced accuracy ∼7-10% across each cell type (Supplemental Figures 10-12). With our final model, we attained cross-dataset predictions (pediCD to FGID & FGID to pediCD) for each cell, giving a probability score of a cell belonging to a subset in the other disease condition (Supplemental Figure 5c; Methods).
Some macrophage clusters characterized by STAT1 activation did not demonstrate significant correspondence to any FGID cluster. We also generally noted substantially increased cluster diversity in pediCD end clusters relative to their correspondence in FGID. This emerged from more patient-specific clusters found in pediCD, and an overall decrease in Simpson’s index of diversity considering the patient composition of each end clusters (Supplemental Figure 10b).
Importantly, when we jointly clustered macrophages from FGID and pediCD together, we identified that several of the original end clusters identified through ARBOL in pediCD were divided across the UMAP: being split into distinct clusters of cells (Supplemental Figure 10c-f). This reinforces the need to quantitatively approach the choice of clustering parameters and number of iterations used. We also employed the STACAS package to integrate T cells between FGID, confirming the higher percentage of proliferating T cells in pediCD patients compared to FGID (Supplemental Figure 11b) (Andreatta and Carmona, 2021). Integration of T cells from the FGID and pediCD datasets (n = 29640 and 38031, respectively) was performed using the STACAS package (v1.1.0) (Andreatta and Carmona, 2021) Sankey plot was created using RAWGraphs 2.0 beta (https://github.com/rawgraphs) (Mauri et al., 2017). However, as compared to ARBOL, the integration approach on T cells resulted in lower heterogeneity, thus masking important disease-associated differences revealed by the ARBOL approach.
Pseudotime analysis of expression landscape on cells
The micrograin structure found through hierarchical tiered clustering is vital for being able to directly compare like cells across disease conditions, and find significant changes in phenotype and composition within individual subsets. It is also vital to understand how those like subsets relate to each other within a disease condition and how the larger macrograin structure differs across conditions. This macrograin structure can be explored through the gradients of gene expression among cells of a major type. Pseudotime and RNA-velocity are both excellent tools for exploring these gradients. For both tools, the choice of genes directly determines the structure found within the dimensional reduction, and thus what genes are chosen as significantly location specific within the resulting landscape of cells. for our purposes, as we knew we would be exploring a single cell lineage, and exploring the relationships of cell states within that space, we required for our dimensional reduction the genes common to that space. We selected genes by performing differential expression between the major cell type and all other cell types within that disease. We took the outer union of those genes. Then removed genes from the list found to be differentially expressed between disease conditions at the major cell type level. From these genes we performed PCA to 50 principal components and then computed a UMAP reduction to 2 components. This selection process allows the dimensional reduction to find smooth gradients between cells and provided a common space for cells of multiple disease conditions to exist.
From this common expression landscape we utilized Monocle3 https://cole-trapnell-lab.github.io/monocle3/ (Cao et al., 2019) to extract a best estimate linear path through the space, which we calculated a diffusion pseudotime on allowing use to numerically estimate the distribution of cells within the expression landscape. We utilized a list of genes that were cell-type defining genes in either FGID or pediCD, but removed genes that were differentially expressed between FGID and pediCD, to allow for cell type/subset to drive placement on the UMAP (Luo et al., 2022; Ordovas-Montanes et al., 2018). This allowed us to place our fine-grained clusters within a joint gene-expression space related to underlying cell types in FGID and pediCD. We note that caution must be taken in interpreting these findings as UMAP distances represent non-linear distances unlike PCA space. To compute the significance of changes in that distribution we used a permutation test of Hellinger distance between distributions. At each of 10,000 permutations we shuffled the group ordering within the comparison pair. We performed this test five times for comparisons between FGID vs FR, FGID vs PR, NOA vs FR, NOA vs PR and FR vs PR. Our threshold was set as Bonferroni corrected p_value < 0.05.
Bulk RNA-seq library generation from PREDICT patients
Population RNA-seq was performed as follows. In brief, tissue was collected directly into RLT Plus lysis buffer (Qiagen) and stored at −80°C until RNA isolation. RNA was isolated from 30 patients using a Qiagen Qiashredder protocol followed by total RNA isolation with the RNeasy Plus Mini Kit (Qiagen). Total RNA was quality controlled via Agilent high-sensitivty RNA Tapestation assay (Agilent). Total cDNA synthesis was performed using Takara SMART-Seq v4 Ultra Low input RNA kit with 10ng of RNA input from samples with RIN values >8 (Takara). cDNA was amplified and cleaned using a 0.6x SPRIselect protocol (Beckman Coulter). Sequencing libraries were generated following the Illumina NXTR XT DNA SMP Prep Kit and Index Kit (Illumina). Libraries were pooled post-Nextera and cleaned using Agencourt AMPure SPRI beads with successive 0.7X and 0.8X ratio SPRIs. Sequencing was performed on Illumina HiSeq®2500 (Illumina) by multiplexed single-end read run with 80 cycles (split from HiSeq SBS Kit V4 250 cycle kit, FC-401-4003). Samples were sequenced at an average read depth of 15 million reads per sample.
Tissue bulk RNA-seq Data Analysis
Tissue and sorted basal cell samples were aligned to the GRCH38-2020-A genome and transcriptome using STAR and RSEM. One sample (p022) was excluded as an outlier based on PCA space so we retained 26 samples for further analyses. Differential expression analysis was conducted using DESeq2 package for R. Genes regarded as significantly differentially expressed were determined based on an adjusted p-value using the Benjamini-Hochberg procedure to correct for multiple comparisons with a false discovery rate <0.05.
92-gene signature derived from pediCD-TIME
In order to generate a gene signature from pediCD scRNA-seq data to apply to bulk RNA-seq data, we took the 25 end clusters associated with FR and PR to anti-TNF in the pediCD-TIME-negative (PC2 negative) direction (Figure 5), and for each of those clusters we selected the top 10 best-scoring markers using the rank-score function (described above).
Duplicate hits were only accounted once, and genes that appear as hits in both the positive and negative directions were filtered out. This gene signature is in Supplemental Table 17.
Gene set enrichment analysis
Fold changes between patients responding or not responding to anti-TNF therapy from RISK and E-MTAB-7604 cohorts were calculated with Seurat (v4.0.3) (Haberman et al., 2014; Hao et al., 2021) and DESeq2 (v1.30.1) (Love et al., 2014) packages, respectively. GSEA analysis was performed using the fgsea R package (v1.16.0) (Korotkevich et al., 2021). Genes with similar fold changes were preranked in a random order. The code for this analysis can be found in the GitHub repository https://jo-m-lab.github.io/3p-PREDICT-Paper/4_GSEA/PREDICT_GSEA_final.html.
Differential expression testing
To calculate differential expression between FR and PR groups, for each subset with a least 50 cells in each condition we used a Wilcoxon test thresholded to 0.05 Bonferroni corrected p-value and down sampled using the “max.cells.per.ident” argument within Seurat’s ‘FindMarkers’ function to a maximum of 10000 cells. The limits on minimum and maximum number of cells were chosen to mitigate issues with comparisons between disproportionate populations and computational efficiency. There does still exist 2 orders of magnitude between the minimum and maximum; however, the subsets most of interest and reported through the manuscript are within the same order of magnitude There are noted spillover effects within the expression tests. We observe ubiquitous contamination of genes such as IGHA1, IGHG1 and DEFA5 across all cell types and subsets. These genes are routinely found as enriched within more severe inflammation, beyond even this dataset. This is a real effect, but less than useful for understanding driving factors within individual cell subsets. So, we focus on significant differentially expressed genes that also have a high pct.cells.expressing.in/ pct.cells.expressing.out ratio.
Understanding how additional outgroups influences cluster diversity and stability through entropy testing
By nature of clustering over two distinct diseases in Figure 6, the specific Simpsons’ Diversity for each cluster is less than what was observed for pediCD and FGID alone, but still highlights that despite these clusters representing ∼0.05-0.1% of cells, they are remarkably composed of cells sampled from most patients within a disease category (Supplemental Figure 15). In particular, we decided to measure the degree to which end clusters maintain their composition, or if cells are split or shuffled into different clusters across trees. To answer this question, we calculate the log base 2 Shannon entropy per curated end cluster to compare the joint ARBOL to the original separate ARBOL’s to determine how many groups each end cluster is split into in an alternate tree. A lower entropy corresponds a majority of the same cells composing a similar cluster in the alternate tree, providing a quantification of the preservation of label-level information, per label, between two trees. These entropies are compared to simulations of randomly scrambled end clusters of the same sizes mapping to the same number of end clusters in the new tree, with a fixed percentage mapping to a single new cluster, while the other cells map randomly.
To calculate representative label preservation, Shannon entropy was calculated as where x1…xk are the set of end cluster labels in a new ARBOL, and p(x) is the fraction of cells of that curated cluster in each of the k new labels. Simulations of 75% label preservation, 50% label preservation, and random shuffling were performed 100 times per pediCD atlas T/NK/ILC cluster size with 57 possible labels. In simulations with preservation, unpreserved labels are shuffled randomly among the other 56 new labels. The average of the 100 simulations per cluster size are displayed in Supplemental Figure 15. Entropies are calculated per pediCD atlas cluster to determine preservation of CD severity vector cell states in new atlases, which contain a higher number of labels. We fixed cluster size and n possible labels to pediCD atlas sizes in simulations, as maximum Shannon entropy scales with both cluster size and number of new labels, such that calculation of joint atlas labels to pediCD alone finds a lower distribution of entropies than the reverse, as there are fewer possible labels in the curated pediCD atlas than the raw joint FGID-pediCD ARBOL.
Simulation of 75% label retention finds entropies between 1 for an end cluster of 8 cells, and 2.1 for an end cluster of >2000 cells. Similarly, simulation of 50% label retention finds entropies of 1.9 to 3.9. Random shuffling of cell labels finds an average of 5.7. We first mapped the pediCD atlas end clusters to the joint atlas and identified that half of end clusters fall under an entropy of 2 with most clusters below the maximum 50% label retention simulation of 3.8 (Supplemental Figure 15). We then took the joint atlas end clusters and mapped them to the pediCD atlas end clusters and recovered a similar distribution. Importantly, these values are all significantly less than when labels are randomly shuffled, whereby entropy values rapidly approach 5.8 with increasing label size. We find that our high-resolution pediCD clusters vary between near perfect label preservation, as many did not mix with FGID cells (Supplemental Figure 15 right NK.MKI67.CD38), to a state where they are combined into one larger cluster (Supplemental Figure 15 right T.STMN1.RRM2), and below 50% label retention, as some larger clusters were broken up and mixed with FGID (Supplemental Figure 15 left T.CCL4.GZMK). It is of interest to note how an “outgroup” condition may influence the resulting entropy of clusters. In these cases, recombination of end clusters based on sample diversity may reduce entropy between ARBOL results.
Joint FGID-CD ARBOL cluster frequencies analysis
We performed analysis of the compositional PCA space of cell type abundances of T, Myeloid, and Epithelial subsets in the joint ARBOL to determine key similarities between the NOA anti-TNF response category and FGID patients. Celltypes were named using our automated naming method. The top 20 cell types in the positive direction of PC1 were selected and visualized using a heatmap of z-score normalized per-patient frequencies. Samples were manually ordered according to disease and antiTNF_response (NM [Not Measured / FGID], NOA, FR, PR) to facilitate direct comparisons across groups. This method enabled clear visual inspection of compositional similarities and differences among patient categories based on hierarchical clustering, revealing subsets which overlap between NOA and FGID patients.
The automated joint-ARBOL names were then compared to the original curated celltype names in both the FGID and CD atlases separately with alluvial plots to give context to the overlapping clusters.
Pediatric-to-adult disease trajectory analysis in integrated atlas cell-state composition space
We performed Harmony integration on adult CD, pediCD, and pediCD longitudinal samples using these 3 categories as the harmony variable. Harmony and SCTransform normalization was performed at each iteration of subclustering to create the resulting integrated ARBOL atlas. Compositions of T, B, Myeloid, Fibroblast, and Plasma cells were calculated per sample by normalizing ARBOL cluster abundance per sample by total number of these celltypes combined in that sample. PCA was performed on the resulting sample x composition matrix, resulting in Figure 8d (left). We plugged this matrix into a Seurat object and used Seurat functions to perform nearest neighbors embedding and UMAP, resulting in Figure 8d (right). This Seurat object was then converted to a Monocle CellDataSet object with a negBinomial.size() expression family. A pseudotime graph was then learned in UMAP space with geodesic_distance_ratio 0.5 and minimal_branch_len 5, with no partitions. The graph was rooted in the NOA cells to begin the pseudotime trajectory at the earliest timepoint in disease trajectory.
To discover cell states varying along the pseudotime trajectory, we used Monocle 3’s graph_test() algorithm, an implementation of the Moran’s I test of spatial autocorrelation. Fig 8f counts the top 50 autocorrelated cell states along pseudotime per celltype ranked by q-value. Cell states’ changes over pseudotime were then visualized using ggplot geom_smooth with a linear cubic model to display cell state presence patterns over pseudotime. This model is estimated with geom_smooth(method = “lm”, formula = y ∼ splines::bs(x)).
To develop a deeper understanding into which cell types varied concordantly and which were more sample or condition-dependent, we used the TradeSeq (https://github.com/statOmics/tradeSeq) package (Supplemental Figure 17). TradeSeq provides an association test which we used in a leave-one-out cross validation experiment, testing the association of each cell state with pseudotime across bootstraps. We then used TradeSeq’s trajectory clustering method to reveal trends in trajectories, named by trend (Supplemental Figure 17b), and overlaid the bootstrap results onto them. Bootstrap tests showed that T, Myeloid, and Epithelial (EC) subsets are robustly associated with the disease pseudotime, and that they generally follow downward-hill, downward-slope, quick-climb, and upward-hill trends.
Multiplex Immunohistochemistry
A fully automated Multiplex Immunohistochemistry assay was performed on the Ventana Discovery ULTRA platform (Ventana Medical Systems, Tucson, AZ), and as previously described(Marron et al., 2022). The assays were optimized for HCC. Optimal concentrations of each antibody were determined, and they were applied in the following sequence and detected with the indicated fluorophore.
Pan-Immune Panel:
Rabbit Anti-CD20 (Clone SP263, Abcam, ab64088) was detected with DISCOVERY Rhodamine 6G (Roche, Part number: 7988168001).
Rabbit anti-CD3 (Clone SP162, Abcam, ab135372) was detected with DISCOVERY DCC (Roche, Part number: 7988192001).
Rabbit anti-CD8 (Clone SP239, Abcam, ab178089) was detected with DISCOVERY RED 610 (Roche, Part number: 7988176001).
Rabbit Anti-CD68 (Clone SP251, Abcam, ab192847) was detected with DISCOVERY Cy5 (Roche, Part number: 7551215001).
Rabbit Anti-FOXP3 (Clone SP97, Abcam, ab99963) was detected with DISCOVERY FAM (Roche, Part number: 7988150001).
Following staining, the tissue was counter-stained and cover slipped with Invitrogen ProLong Gold Antifade Mountant with NucBlue. Whole slide imaging was performed on the Zeiss Axioscan which was equipped with a Colibri light source and appropriate filters for visualizing these specific fluorophores. Where applicable, an additional marker, Mouse anti-Pan Keratin (Clone AE1/AE3/PCK26, Roche, Part number: 5266840001) was added. Coverslips were removed by immersing the slides overnight in distilled water, staining with the PanCK antibody and detecting with Goat anti-mouse IgG (H+L)-Cy7 (AAT Bioquest, 16856). Slides were re-cover slipped and scanned as described above.
Quantitative image analysis
Quantitative image analysis was preformed using HALO Indica Labs Hyperplex module (IndicaLabs, Albuquerque, NM). For each sample, images from the pan-immune panel and the pan-CK were fused to generate one single image. A first classifier was applied to detect the tissue on each section and an automatic annotation was generated as “whole section”. A second classifier was applied using panCK and DAPI channels to define the epithelium layer (panCK positive) and the lamina propria (panCK negative). Automated annotations were generated as “epithelium” and as “lamina propria”. Immune cells were detected in each area of interest (whole section, epithelium, lamina propria). T cells were defined as CD3+, B cells as CD20+, myeloid cells as CD68+. For each T cell subset, CD8 T cells were defined as CD3+CD8+FOXP3-, CD4 as CD3+CD8-, Tregs as CD3+CD8-FOXP3+ and CD4 conventional as CD3+CD8-FOXP3-. Numbers of positive cells for each immune subset were counted and their density measured.
General Statistical Testing
Parameters such as sample size, number of replicates, number of independent experiments, measures of center, dispersion, and precision (mean +/- SEM) and statistical significances are reported in Figures and Figure Legends. A p-value less than 0.05 was considered significant. Where appropriate, a Bonferroni or FDR correction was used to account for multiple tests, as noted in the figure legends or Methods. All statistical tests corresponding to differential gene expression are described above and completed using R language for Statistical Computing.