{"paper_id":"3ce855bf-db1c-41c1-aea0-07389cb4e1f2","body_text":"Velociraptor: Cross-Platform Quantitative Search  \nUsing Hallmark Cell Features \n \nClaire E. Cross1,2,6, Cass Mayeda1,2, Stephanie Medina1,2,6, Madeline J. Hayes 1,2,6, Saara Kaviany3,6, \nJames A. Connelly 3,6, Jeffrey C. Rathmell 2,5,6, Kyle D. Weaver 4,5, Reid C. Thompson 4,5, Lola B. \nChambless4,5, Rebecca A. Ihrie1,4,5, Jonathan M. Irish1,2,5,6* \n \n1 Department of Cell and Developmental Biology, Vanderbilt University, Nashville, TN, USA. \n2 Department of Pathology, Microbiology and Immunology, Vanderbilt University Medical Center, \nNashville, TN, USA. \n3 Department of Pediatrics, Vanderbilt University Medical Center, Nashville, TN. \n4 Department of Neurological Surgery, Vanderbilt University Medical Center, Nashville, TN, USA. \n5 Vanderbilt-Ingram Cancer Center, Vanderbilt University Medical Center, Nashville, TN, USA. \n6 Vanderbilt Human Immunology Discovery Initiative of the Vanderbilt Center for Immunobiology. \n*Corresponding author(s) \n \nEmail: jonathan.irish@vanderbilt.edu (J.M.I.) \n \nRunning Title: A framework for identifying and seeking cellular hallmarks using quantitative cell type \nlabels readable by human experts and machines \n  \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nAbstract \nA key challenge for single cell discovery analysis is to identify new cell types, describe them \nquantitatively, and seek these novel cells in new studies often using a different platform. Over the last \ndecade, tools were developed to address identification and quantitative description of cells in human \ntissues and tumors. However, automated validation of populations at the single cell level has \nstruggled due to the cytometry field’s reliance on hierarchical, ordered use of features and on \nplatform-specific rules for data processing and analysis. Here we present Velociraptor, a workflow \nthat implements Marker Enrichment Modeling in three cross-platform modules: 1) identification of \ncells specific to disease states, 2) description of hallmark features for each cell and population, and 3) \nsearching for cells matching one or more hallmark feature sets in a new dataset. A key advance is \nthat Velociraptor registers cells between datasets, including between flow cytometry and quantitative \nimaging using different, overlapping feature sets. Four datasets were used to challenge Velociraptor \nand reveal new biological insights. Working at the individual sample level, Velociraptor tracked the \nabundance of clinically significant glioblastoma brain tumor cell subsets and characterized the cells \nthat predominate in recurrent tumors as a close match for rare, negative prognostic cells originally \nobserved in matched pre-treatment tumors. In patients with inborn errors of immunity, Velociraptor \nidentified genotype-specific cells associated with GATA2 haploinsufficiency. Finally, in cross-platform \nanalysis of immune cells in multiplex imaging of breast cancer, Velociraptor sought and correctly \nidentified memory T cell subsets in tumors. Different phenotypic descriptions generated by algorithms \nor humans were shown to be effective as search inputs, indicating that cell identity need not be \ndescribed in terms of per-feature cutoffs or strict hierarchical analyses. Velociraptor thus identifies \ncells based on hallmark feature sets, such as protein expression signatures, and works effectively \nwith data from multiple sources, including suspension flow cytometry, imaging, and search text based \non known or theoretical cell features.  \n \nKeywords \nCell identification, machine learning, glioblastoma, inborn errors of immunity, single cell biology \n  \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nIntroduction \nA cell can be classified based on its morphology, location, protein expression, or a combination of \nthese features 1. The method of cell identification depends on the technology used to measure \ncellular features. In the field of flow cytometry, the classic approach to identify cells is manual biaxial \ngating, which uses strict thresholds of marker positivity and negativity that are used to distinguish cell \npopulations \n2,3. Hierarchical filtering for populations using manual gating excludes cells that don’t \nperfectly match a phenotype of interest, even if a cell falls just short of a threshold for a single protein. \nThis approach is subjective and labor intensive \n4, especially when analyzing high-parameter mass \ncytometry datasets that can contain measurements of greater than 40 markers 5,6. Additionally, \nmanual gating is biased towards identification of well-established populations that have known \npatterns of protein expression 7.  \nUnsupervised machine learning cell identification approaches aim to overcome the limitations of \nsubjectivity and bias towards known populations that are endemic to manual biaxial gating. Cell \npopulations can be distinguished in low-dimensional projections of high-dimensional data, which \nincludes all cells and not just known populations \n4,7,8. To identify different populations within a sample, \nunsupervised clustering methods, such as k-means clustering and FlowSOM 9, group cells into a pre-\ndefined number of non-overlapping subpopulations that ideally represent distinct cell types. However, \nthese disjoint clustering methods have often been limited by the requirement of a priori knowledge of \nthe number of distinct populations that exist in a sample. Additionally, stochasticity can result in \nvarying assignment of cell types across separate analyses. The resulting disjoint clusters can also \ncontain heterogeneous mixtures of cells that belong to distinct cellular states that should not be \ngrouped into a single phenotype. Alternatively, we and others have established the use of cell-specific \napproaches that utilize overlapping local phenot ypic neighborhoods to identify populations \n10-12. This \nlocal phenotypic neighborhood approach continuously tracks phenotypic space, which enables \ndetection of rare cells (<5% of a sample) and subtle phenotypic shifts that could have been \noverlooked if that cell were to be placed into a larger, disjoint cluster \n10. Individual populations can \nthen be tested for association with external variables, such as infection status or overall survival time, \nusing various unsupervised machine learning algorithms 13-16.  \nHistorically, identifying the phenotype of a subpopulation of cells that is associated with an external \nvariable has required manual review and expert annotation of markers that are expressed by that \ngroup of cells. The development of Marker Enrichment Modeling (MEM) enabled automatic \nquantification of population phenotypes \n17. MEM generates a label that includes the markers that are \nenriched in different populations as well as the relative level of enrichment of each marker. These \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nlabels are quantitative and both human- and machine-readable, meaning that not only is a \npopulation’s identity summarized in a way that can be interpreted by a scientist, but a MEM label can \nalso be used as the starting point for subsequent computational analyses. For example, the \nphenotypes of multiple populations can be compared by computing a \nΔ MEM label that quantifies \npairwise differences in feature enrichment between two populations 18. Thus, cell identity can be \nquantified based on the expression or enrichment of various features, such as proteins. However, a \nhistorical lack of standardization and use of differing approaches to define cell identity across fields \nhas led to the generation of multiple independent methods to identify the same population of cells \n1. \nEven in a well-established field such as immunology, different groups may use different features to \nidentify the same cell population (e.g., using CD127, CD25, FOXP3, or a combination of these \nfeatures to identify regulatory T cells). Thus, identification of similar populations in additional datasets \nbased on a learned phenotype still requires a high degree of manual supervision to identify cells that \nmay have been characterized using different features in order to validate findings.  \nPopulations of interest identified in analyses should be validated for stability and generalizability using \napproaches such as k-fold cross validation and leave-one-out cross validation \n19,20. A stable cluster \nwill be composed of the same, or similar cells, across multiple runs of a machine learning pipeline on \nrepeated samples of the dataset. Population identific ation in newly collected samples will reveal the \ngeneralizability of initial results. The gold standard to bi ologically validate a result is to use a different \nplatform to ensure the observed population is not a technological artefact. However, there are many \nchallenges in validating automatically identified cell populations across experiments and modalities. \nInherent technological differences cause single cell platforms to collect data that can span widely \ndifferent ranges. As a result, a common cross-platform cell identification approach is to develop a \nmanual gating scheme to select cells based on marker positivity and negativity \n10,16. Computational \nalgorithms have also been developed to integrate distinct data types by combining the data into a \ncommon phenotypic space \n21-23. However, these methods require (1) a large overlap between marker \npanels as they were developed to integrate scRNAseq data, (2) extensive data pre-processing steps \ncorrect for technological differences and to transform data onto similar scales, and often (3) a \n“matchability test” to ensure that the two datasets are amenable to a joint analysis \n24. While a single \npooled analysis better captures the natural variability that exists among patients, the number of cells \nanalyzed per patient is often limited to the fewest number of cells collected from a donor to prevent a \nsingle sample from dominating an analysis \n14,25-27. This equal sampling limits the total number of cells \nanalyzed, which can reduce the overall power of the analysis. Another limitation of a pooled analysis \ncan be that upon collection of an additional sample to be included in the cohort, samples must be re-\nanalyzed. Algorithms such as UMAP and SCAFFOLD \n8,28 allow additional samples to be added into a \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\npreviously generated low dimensional embedding but assume that there are no sample-specific cell \ntypes or patterns of expression that were not present in the original analysis. Ideally, a cross-dataset \nidentification algorithm will require minimal overlap between feature panels, analyze samples \nindividually rather than performing a pooled analysis in a common phenotypic space, and control for \npotential differences in data scales due to minor batch effects or different data types. \nHere, we present Velociraptor, a novel cell identification workflow that reveals clinically relevant cell \npopulations, identifies the phenotypes of those populations, and seeks out similar cells across \nexperiments and single cell platforms. We use Velociraptor to address biological challenges that are \nregularly encountered across diverse settings of disease, including automated cell identification in \ncases of limited sample or cell numbers and matching of cell populations across technologies. The \ndiscovery and seeking of populations across datasets ultimately allowed for identification of rare and \nabundant cells associated with external variables, determination of the most essential proteins that \ndistinguish a population from other cells, and automated validation of cell subsets across single cell \nplatforms without forcing distinct data into a common phenotypic space.  \n \nResults \nVelociraptor overview \nThe Velociraptor workflow consists of two novel cell identification tools that can be used in tandem or \nindividually: Velociraptor-Claw (VR-Claw) and Velociraptor-Eye (VR-Eye) ( Figure 1A\n). VR-Claw \nutilizes local phenotypic neighborhood identification to reveal cells that are associated with \ncontinuous variables (e.g., overall survival time). MEM quantifies the phenotype of a population of \ninterest, and that MEM label is used to seek similar cells in new samples with VR-Eye. This cell \nidentification workflow addresses four biological challenges: 1) de novo cell identification and \nalignment with external labels, 2) cell identification and patient classification in new samples, 3) \nidentifying cells across batches collected using the same platform, and 4) identifying cells across \nsingle cell platforms ( Figure 1B\n). Each analysis was statistically validated using methods such as k-\nfold cross validation, whole sample analysis, repeated down-sampling, and leave-one-out cross \nvalidation (LOOCV) (Figure 1C\n). VR identified clinically relevant cell populations across experiments \nand single cell platforms (Figure 1D) as described in the analyses below. \n \nVelociraptor identified clinically relevant cell populations in primary GBM tumors \nVelociraptor-Claw revealed prognostic GBM populations \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nThe goal of the first analysis was to validate that Velociraptor could accurately identify cell \npopulations that were known to be associated with overall survival times. The algorithm was tested \nusing a published 36-dimensional mass cytometry dataset that characterized a cohort of 28 primary \nhuman glioblastoma tumors (Dataset 1\n) 14. This dataset has been shown to contain two distinct \nprognostic cell populations: 1) Glioblastoma Negative Prognostic (GNP) cells that co-express \nastrocytic marker S100B and stem-like marker SOX2 and are associated with shorter overall survival \ntimes (median survival of 144 days), and 2) Glioblastoma Positive Prognostic (GPP) cells that are \ncharacterized by high expression of EGFR and are associated with longer overall survival times \n(median survival of 836 days).  \nVR-Claw identified one stable GNP population (p<0.05, HR>1) and four stable GPP populations \n(p<0.05, HR<1; Figure 2A\n). Here, stable means that a similar cluster was significantly (p<0.05) \nassociated with overall survival in each of the 5 rounds of cross validation. GNP cells in cluster 1 co-\nexpressed S100B and SOX2 and displayed basal phosphorylation of multiple signaling proteins \n(MEM: S100B+6 SOX2+5 p-STAT3+4, CyclinB1+3 p-STAT5+3, p-S6+3, p-AKT+2, and p-NFkB +2; Figure \n2B). GPP cells in clusters 2, 3, 4, and 5 occupied four distinct expression profiles ( Figure 2B ). \nCluster 2 expressed a high level of EGFR (MEM: EGFR+7 SOX2+3 GFAP+3 p-NFkB+2 CD44+2). Cluster \n3 co-expressed astrocytic markers S100B and GFAP as well as EGFR (MEM: S100B +3 EGFR +3 \nGFAP+2). Cluster 4 exhibited high basal signaling and lacked expression of canonical neural cell \nsurface markers (MEM: p-AKT+10 p-STAT1+3). Cluster 5 displayed high levels of phosphorylation of p-\nSTAT5 and p-NFkB along with EGFR expression (MEM: p-STAT5 +9 EGFR +5 p-NFkB +4 S100B +2 \nCD56+2 SOX2+2 CD44+2 GFAP+2).  \nAcross five folds of cross validation, patients were consistently classified as being either GNP-High or \nGPP-High based on their abundance of these prognostic subsets ( Figure 2C\n and Supplementary \nFigure S1). Notably, patients with a large abundance of either GNP or GPP cells lacked the other \nprognostic subset ( Figure 2C ). Kaplan-Meier analyses confirmed that GNP-High patients have a \nshorter overall survival time and that GPP-High patients have a longer overall survival time compared \nto patients that had a low abundance of these two populations, respectively ( Figure 2D\n and Figure \n2E). VR-Claw accurately classified patients as being GNP-High or GPP-High compared to the \nprevious classification reported by RAPID (GNP-High F1-measure=0.94, GPP-High F1-\nmeasure=0.72; Supplementary Figure S1C\n). These results indicated that VR-Claw successfully \nidentified prognostic populations of GBM cells in Dataset 1. \n \nVelociraptor-Eye accurately identified known populations of GBM cells based on learned phenotype \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nThe next analysis goal was to establish that the novel cell seeking approach used by VR-Eye \naccurately identified cell populations based solely on their phenotype. MEM labels learned by VR-\nClaw were used as reference phenotypes for VR-Eye to seek in both the training data that was used \nto learn the phenotype and in the testing data that was withheld from VR-Claw and MEM analyses \n(Figure 2F\n and Figure 2G). Patient-level abundance of GNP-like cells by VR-Eye in the training data \nand testing data was highly correlated (R 2=0.98; Figure 2H). Patient-level abundance of GPP-like \ncells by VR-Eye in the training data and testing data was highly correlated (R2=0.99; Figure 2I).  \nTo determine the cellular features that best distinguished each population from other cells, a series of \nVR-Eye searches was run using different combinations of markers to compare cell identity to the \npopulation of interest. Populations within Dataset 1 were used as “knowns” to be sought based on \ntheir multidimensional patterns of protein expression Supplementary Figure S2A . Specifically, \nRAPID-identified GNP and GPP populations were used to test VR-Eye. F1-measures were calculated \nto compare VR-Eye cell classification with RAPID cell classification. Optimization of search input \nrevealed that the key cellular features that distinguished GNP cells from other cells were co-\nexpression of SOX2 and S100B as well as lack of GFAP and p-NFkB expression (F1-measure=0.80; \noptimized GNP label: SOX2\n+7 S100B+4 GFAP+1 p-NFkB+0; Supplementary Figure S2B). GPP cells \nwere distinguished by high expression of EGFR and lack of CD49F, p-p38, CD44, and B3TUB (F1-\nmeasure=0.90; optimized GPP label: EGFR +7 CD49F +2 p-p38 +2 CD44 +1 B3TUB +0. Population \nabundances for GNP and GPP cells identified by VR-Eye with the optimized search input were \nstrongly correlated with population frequencies identified by RAPID ( Supplementary Figure S2C ; \nGNP R2=0.94, GPP R2=0.99).  \nTo benchmark Velociraptor performance with other methods of population identification, VR-Eye and \nRAPID were each applied to ten different t-SNEs generated using the same sample of 131,880 GBM \ncells. Additionally, four cytometry experts manually gated for GNP and GPP cells based on a \npublished gating scheme \n14. VR-Eye identified GNP-like cells with a median F1-measure of 0.77 and \nGPP-like cells with a median F1-measure of 0.84 across ten runs ( Supplementary Figure S2D ). \nRAPID identified GNP-like cells with a median F1-measure of 0.63 and GPP-like cells with a median \nF1-measure of 0.61 across ten runs. Biaxial gating identified GNP-like cells with a median F1-\nmeasure of 0.35 and GPP-like cells with a median F1-measure of 0.57 across ten runs. Taken \ntogether, these analyses established that VR-Eye accurately and reproducibly identified cell \npopulations by seeking a specified phenotype.  \n \nVelociraptor-Eye accurately revealed cell populations in individual analyses \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nThe next goal was to compare a multi-sample pooled analysis to individual sample analyses. In this \ntest of VR-Eye, all GBM cells from each tumor in Dataset 1  were analyzed in a patient-specific \nmanner rather than in a down-sampled, combined analysis. This type of analysis also determines \nhow representative a subset of cells is to an entire tumor sample. For each tumor, all GBM cells were \nused to generate a tumor-specific t-SNE ( Figure 3A). The number of GBM cells in individual tumors \nranged from 4,710 cells to 329,650 cells with a median cell count of 68,761. Next, GNP and GPP \npopulations were sought with VR-Eye using the optimized population phenotypes (Figure 3B\n). The \nmajority of GBM cells were not similar to either prognostic phenotype. Individual tumors ranged in \nGNP-like abundance from 0.00% to 54% and in GPP-like abundance from 0.00% to 88%. Tumors \nwith greater than 5% abundance either GNP-like or GPP-like cells contained less than 2.5% of the \nother population. The difference between each cell’s GNP similarity score and GPP similarity score \nwas calculated for each sample as a surrogate of population homogeneity ( Figure 3C\n). Population \nabundances for GNP and GPP cells identified by VR-Eye in full tumor samples were strongly \ncorrelated with population frequencies identified by RAPID in the down-sampled dataset ( Figure 3D; \nGNP R2=0.96, GPP R 2=0.98). These results show that VR-Eye accurately identifies cell populations \nin individual samples as opposed to a pooled analysis.   \n \nVelociraptor-Eye identified prognostic cell populations in recurrent GBM tumors \nHaving verified that VR accurately and reproducibly identified prognostic cell populations that have \nbeen reported in a published dataset, the goal of the next analysis was to use Velociraptor to identify \nprognostic cells in newly collected data. To address this goal, three recurrent GBM tumors were \nresected, dissociated, and stained with a 33-dimensional antibody panel to measure protein \nexpression using mass cytometry as previously described ( Dataset 2\n) 14,29. All 33 markers in the \nantibody panel used to characterize the recurrent tumors were present and measured in the same \nchannels in Dataset 1\n, and each recurrent tumor had a corresponding primary tumor that was \nincluded in Dataset 1. For each recurrent tumor, VR-Eye sought GNP and GPP populations using the \noptimized phenotypes for each population, respectively ( Figure 3E). Each recurrent tumor contained \na greater percentage of GNP-like cells than its corresponding primary tumor, and in the two primary \ntumors that originally contained GPP-like cells, the corresponding recurrent tumors each showed a \ndecrease in GPP-like cell abundance (Figure 3F\n).  \n \nVelociraptor identified genotype-specific cells in patients with Inborn Errors of Immunity \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nThe next test of Velociraptor was designed to assess how well the entire workflow performs in the \ncase of limited sample numbers and limited cell numbers. To address this question, we performed \ndeep T cell profiling on PBMCs from six sets of patients with Inborn Errors of Immunity and healthy \ndonors using a 43-dimensional mass cytometry panel ( Dataset 3\n). Inborn Errors of Immunity \nencompass a variety of rare, monogenic mutations that lead to diverse clinical presentations of \nautoimmunity, autoinflammation and immunodeficiencies 30,31. The first set of donors included data \nfrom 1 healthy adult donor and 2 IEI patients with STAT1 GOF mutations that have been previously \nreported \n32 as well as unpublished data from 2 patients with GATA2  haploinsufficiency (Set 1, N=5). \nDonor sets 2-6 included an additional 38 IEI patients, 6 healthy adult donors, and 2 healthy pediatric \ndonors (N=46). In total, all donors included in this study (N=51) represented mutations in 26 unique \ngenes. The inclusion of multiple patients with GATA2 haploinsufficiency in two batches that were \ncollected on different dates presented the opportunity to identify GATA2 haploinsufficiency-specific \ncell populations in Set 1 that could then be sought in Set 2 for validation.  \nT cells from donors in Set 1 were embedded into a two-dimensional t-SNE based on protein \nexpression (Figure 4A\n). A modified version of VR-Claw was created to identify cells associated with a \ncategorical variable (e.g., genotype) using local phenotypic neighborhood-based identification. This \nversion of VR-Claw was inspired by T-REX 10, but differs in two key aspects: 1) VR-Claw tests each \ncategory represented in a dataset for cell enrichment rather than being restricted to a binary \ncomparison, and 2) VR-Claw includes a series of filtering steps that ensure findings are not specific to \na single patient within a cohort analysis. VR-Claw identified three populations that were greatly \nenriched in (>95%) and three populations that were greatly lacking from (<5%) the patients with \nGATA2 haploinsufficiency ( Figure 4B\n). Further inspection of these populations confirmed that \npopulations were consistent in both GATA2 haploinsufficient patients (Figure 4C). MEM was used to \nquantify the phenotypes of each cell population ( Figure 4D). CD4+ and CD8+ T cells populations that \nwere statistically lacking from the two patients with GATA2  haploinsufficiency each expressed high \nlevels of CD27 and CCR7 (MEM scores ≥  +6). Conversely, the CD4+ and CD8+ populations that were \nstatistically enriched in the two patients with GATA2 haploinsufficiency expressed very high levels of \nCD57 (MEM scores ≥  +6). VR-Claw was repeated an additional nine times using independently \nsampled cells with replacement to verify population stability (Supplementary Figure S3). \n VR-Eye was then used to seek similar cell populations in donor samples from Sets 2-6. All T cells \nfrom each donor were embedded into donor-specific t-SNE spaces to maximize the number of cells \nincluded in this analysis ( Figure 4E ). Each population revealed with VR-Claw was first sought in \nsamples from Set 1 to validate findings and to determine appropriate similarity thresholds that \ndistinguish each population (Figure 4F\n). Each population was then sought in the remaining IEI patient \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nand healthy donor samples ( Figure 4F). For each patient, the abundance of each population was \nquantified ( Figure 4G ). Taken together, these analyses showed that Velociraptor successfully \nidentified clinically relevant cells in the case of limited samples and limited cell numbers. \n \nCross-platform identification \nHaving shown that VR-Eye accurately identifies cells that were characterized using mass cytometry, \nthe goal of the next analysis was to establish whether VR-Eye could be used to identify cells \ncharacterized with other single cell platforms. This question was addressed using a 41-dimensional \nimmune-focused imaging mass cytometry dataset that characterized tumor immune \nmicroenvironments of human breast cancer ( Dataset 4\n) 33. In this dataset, the authors annotated 18 \ndifferent cell types. To first establish that VR-Eye could accurately identify cells in a new data type, \neach cell type identified by Tietscher et al. was sought and compared to the original identification \n(Figure 5A). Each population was sought using a 41-feature phenotypic quantification that included \nevery protein measured in the dataset as well as with a filtered phenotypic quantification that only \nincluded proteins that were specifically enriched or lacking in that population according to relative \nMEM on a common t-SNE embedded with equally sampled cells from each patient (n=257,076 cells \ntotal; Supplementary Figure S4A\n). Filtered search input ranged in size from 4 features to 28 \nfeatures. Full 41-feature labels resulted in more accurate identification than the filtered labels in seven \nout of ten cell types ( Supplementary Figure S4B\n). To compare the populations identified by the \noriginal authors and by VR-Eye, a phenotypic homogeneity score was calculated. This score \ncompared how similar a single neighborhood’s phenotype was to the phenotype of its corresponding \npopulation. Populations identified by VR-Eye using full 41-dimensional labels had the highest \nhomogeneity scores in nine out of ten cell types compared to the original populations and the \npopulations identified by VR-Eye with a filtered label (Supplementary Figure S4C\n). Cell identification \naccuracy was confirmed using LOOCV. These results showed that VR-Eye accurately identifies cell \npopulations in an external imaging mass cytometry dataset. \nTo test whether quantified reference populations can be used to identify similar cells across data \ntypes, we sought memory T cells using a variety of search inputs that were generated using different \ntechnologies. Memory CD4 and CD8 T cells subsets were manually gated as a reference population \n(Supplementary Figure S5A\n). Theoretical memory T cell labels were generated based on known \npatterns of protein expression. Minimal theoretical labels including CD3, CD45R0, and CD4 or CD8a \nmarkers were designed as the minimal feature set that could be used to identify subsets of memory T \ncells, and maximum theoretical labels including CD3, CD4, CD8a, CD45R0, CD45RA, CD20, HLA-\n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nDR, CD68, CD11c, CD303, LAMP3, CD31_vWF, and panCK markers were designed to exclude \nadditional cell types present in the tissue. To assess how well reference phenotypes calculated from \northogonal platforms could identify memory T cells, MEM was performed on a biaxially gated memory \nT cell populations from a healthy adult donor in Dataset 3\n to yield a CyTOF-generated label, and a \n40-dimensional spectral flow cytometry dataset (Dataset 5 ) 34 was used to calculate a spectral-\ngenerated MEM label on memory T cell populations. Each of these labels were used to seek memory \nT cells with VR-Eye (Figure 5B, Figure 5C, and Supplementary Figure S5B).  \nMemory CD4 and CD8 subsets were identified with VR-Eye using similarity thresholds that were \noptimized to best capture the gated population. The phenotype of each VR-identified population was \ncompared to biaxially gated memory T cell populations with a \nΔ MEM analysis (Figure 5D, Figure 5E, \nand Supplementary Figure S5C). Features of VR populations identified using the maximum theory-\ngenerated label and the CyTOF-generated label did not differ from the MEM label of the gated \nmemory CD4 T cell population by more than 1. This indicates high phenotypic similarity between the \nVR-identified population and the gated population. The greatest difference observed was elevated \nCD8a expression on the VR population identified using the minimum theory-generated label (\nΔ MEM: \nCD8a+2). To compare the VR-identified populations to the gated populations at the individual cell \nlevel, the precision, recall, and F1-measure for each VR-Eye analysis was calculated \n(Supplementary Figure S5D\n).  \nUpon inspection of the data, we observed cells that shared high similarity to both memory CD4 and \nCD8 T cell populations in addition to cells that strongly matched a single phenotype, and these cells \nwere confirmed with a VR-Eye search for CD4\n+CD8+ double positive memory T cells \n(Supplementary Figure S5E). The difference between each cell’s memory CD8 similarity score and \nmemory CD4 similarity score was calculated to distinguish cells that matched only a single phenotype \n(Supplementary Figure S5F ). To exclude cells that matched multiple memory T cell subsets, we \nfiltered the VR-identified memory CD4 T cell population such that each cell (1) surpassed the \noptimized memory CD4 T cell similarity threshold, and (2) had a difference in memory CD4 and \nmemory CD8 similarity of greater than 5. This filtering step increased the percentage of VR-identified \nmemory CD4 T cells falling into the CD4\n+ biaxial gate from 66% to 90% (Supplementary Figure \nS5G).  Exclusion of cells that matched both memory CD4 and CD8 phenotypes also improved \npopulation recall compared to gated populations (memory CD4 recall = 0.93, memory CD8 recall = \n0.97). These results show that VR-Eye enables cross-platform cell identification and accurately \nidentifies sought populations using different feature sets.  \n \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nDiscussion \nVelociraptor identifies cells based on association with external variables or similarity to a reference \nphenotype. The modular nature of the workflow follows best practices developed for cytometry \nanalysis in prior decades \n35,36 and allows for independent use of a single tool or a complete analysis \nusing VR-Claw and VR-Eye linked by MEM. The use of local phenotypic neighborhoods that track \nacross continuous phenotypic space rather than placing each cell into a single cluster allows \nVelociraptor to identify both abundant and rare populations of cells at a granular level and to discern \nsubtle changes in protein expression across populations. VR-Claw identified expected GNP-like and \nGPP-like populations in Dataset 1\n that were identified in previous studies using the RAPID algorithm \n14 with exact and reproducible boundaries around prognostic clusters that were significantly \nassociated with patient outcome (p<0.01; Figure 2A ). Notably, the local neighborhood approach \nenabled VR-Claw to identify specific regions of GPP-like cells more significantly associated with \noutcome (p<0.01, HR<1, [95% CI 0.039-0.493]) compared to those identified with RAPID, as well as \nan additional GPP cluster (cluster 3) that co-expresses S100B and EGFR and was not found by \nRAPID due to limitations in RAPID’s clustering approach. Therefore, VR-Claw can be used to reveal \nrare prognostic cells that have historically been overlooked. \nVR-Eye offers a novel approach to defining cell identity that is based on phenotypic comparisons with \nwell-established cell identities to assess if a population strongly matches a single cell type or if it has \nsimilarity with multiple populations. Additionally, VR-Eye simultaneously considers an entire \nphenotype when calculating a continuous similarity score rather than performing sequential binary \nfiltering based on strict thresholds that distinguish marker negativity and positivity; this allows VR-Eye \nto tolerate a small degree of variability in the expression of a protein. These features of VR-Eye allow \nfor multidimensional definitions of cell identity rather than binary cell classification. VR-Eye \noutperformed RAPID and biaxial gating cell identification strategies by reproducibly identifying the \nsame GNP and GPP cells across 10 t-SNEs generated using the same sample of 131,880 GBM cells \n(Supplementary Figure S2\n). VR-Eye also enables cell identification within an individual sample \nrather than within a pooled analysis, which bypasses the need to re-analyze historical cohorts upon \ncollection of a new sample. Additionally, VR-Eye interrogates every cell in a sample rather than \nrequiring equal sampling based on the sample with the fewest cells in a grouped analysis, which \nincreases the total number of cells included in an analysis and thus the overall power of the analysis.  \nNotably, Velociraptor revealed that GBM cell populations associated with shorter survival times were \nmore abundant in GBM tumor recurrences compared to the same patient’s primary GBM tumor \n(Figure 4E and Figure 4F\n). Conversely, populations of positive prognostic GBM cells were greatly \ndiminished or completely lacking in matched recurrent tumors. These shifts in population frequency \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\noccurred regardless of the original frequency of GNP and GPP cells in the primary tumor. One \npotential explanation for this change in tumor makeup is that GPP cells are more susceptible to \nchemotherapy and radiation treatment, whereas GNP cells may persist such that the same GNP cells \nand their progeny remain in the recurrent tumor. An alternative explanation is that another population \nof cells survives treatment and later gives rise to a new population of GNP cells in the recurrent tumor \neither due to plasticity or differentiation. Interestingly, the phenotype of GNP cells share similarity with \nmultiple healthy neuronal cells due to aberrant co-expression of S100B, a marker of astrocytic cells, \nand SOX2, a Yamanaka factor that is highly expressed in pluripotent stem cells \n1,37.  Due to the \nincreased expression of the transcription factor SOX2, GNP cells can potentially be considered more \n“stem-like” than their GPP counterparts. This feature of cell identity may contribute to their \npersistence and increased abundance in recurrent GBM tumors compared to matched primary GBM \ntumors, as cancer stem cells have been implicated as major players driving tumors and tumor \nrecurrence across cancer types \n38-41.  \nAs a proof of concept, Velociraptor next identified genotype-specific cells across sets of patients with \nIEIs. VR-Claw first revealed multiple CD57 + T cell subsets that were statistically enriched in IEI \npatients using samples of only 2000 cells per patient. VR-Eye then sought similar cell populations in \nother healthy donors and IEI patients, including one additional patient with GATA2 haploinsufficiency. \nResults suggested a lack of naïve T cell subsets in IEI patients with GATA2 haploinsufficiency and a \ncorresponding enrichment of CD57 T cell subsets; however, additional samples are required to \nconfirm these findings. Notably, CD57 cells have been observed in other disease settings. After \nobserving these findings from Velociraptor, we noted a study that also reported that patients with \nGATA2 haploinsufficiency have reduced numbers of naïve T cells and increased numbers of CD57\n+ T \ncells compared to healthy donors that correlated with clinical severity 42. This study supports the \nVelociraptor findings shown here, which confirms the usefulness of this workflow even in the case of \nlimited patient samples (N = 3 patients with GATA2 haploinsufficiency) and limited cell numbers (n = \n2000 cells per patient in the VR-Claw analysis). \nFinally, we used VR-Eye to identify cells across single cell platforms. Reference population labels \ngenerated using different platforms successfully identified memory CD4 T cells in the example shown \nhere. Interestingly, a minimum theoretical phenotypic description that contained only markers thought \nto be highly expressed in the sought population did not accurately identify memory CD4 T cells. \nInstead, searching labels containing a balance of features that are highly expressed (+10) and not \nexpressed (+0) performed well across datasets and single cell modalities. This feature of a robust \nsearching label was also observed in the optimization of reference phenotypes in Dataset 1 and \nDataset 3, suggesting that cell identity can be described in terms of both positive and negative \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nenrichment. Reference phenotypes generated using immune-focused panels from CyTOF and \nspectral flow cytometry successfully identified memory CD4 T cells despite only sharing 12 and 8 \nmarkers, respectively, with the IMC antibody panel. Inclusion of additional markers, such as panCK to \ndiscriminate against malignant cells, in the maximum theoretical label did increase the difference \nbetween identified memory CD4 T cells and distinct cell types. This analysis also revealed a small \npopulation of CD4\n+ cells that had been excluded from the CD4 T cell memory compartment with \nexpert gating. Upon manual inspection, these cells closely matched the memory CD4 label, but also \nexpressed CD8a which was likely why they were excluded from the gated population. A subsequent \nsearch found these cells matched a CD4\n+CD8+ double positive memory label greater than a CD4 + \nsingle positive label. Overall, these results (1) helped resolve a population that had not been \ndescribed in the original research or by manual gating, and (2) demonstrated the value of an objective \nmatch, of including both positive and negative features, and of detecting matches to more than one \nknown identity. \nAdditional algorithms have been developed by others to integrate data collected by multiple single cell \nmodalities by placing each cell into a common phenotypic space following data transformation, \nregardless of the origin of the data. However, this method may require many shared features between \ndatasets as it was developed for single cell transcriptomics \n22,43 and relies on the presence of similar \ncell types and cell states in each dataset 24. VR-Eye differs in that it does not require a priori  \nknowledge that similar cell populations are present in two datasets, but rather tests if there are similar \ncells present in two datatypes. Through utilization of MEM, orthogonal data types need not be forced \ninto common phenotypic space. Thus, VR-Eye can be used to seek similar populations of cells across \ndata types, disease states, and types of tissue.  \nBoth tools in the Velociraptor workflow perform phenotypic neighborhood definition on dimensionally \nreduced latent spaces. Dimensionality reduction preserves high-dimensional data structure while \ncondensing data into a space that is less sparse and more amenable to clustering \n44. One type of \ndimensionality reduction that was not compared here is a graphical representation of the data 12. One \nbenefit of using a graphical data structure is that many downstream analysis tools are designed to \nuse graphical data structures as input \n45-47. However, dimensionally reduced data spaces generated \nvia t-SNE or UMAP can be converted to graph representation if so desired, where each cell is a node \nand edges are drawn between each cell and its phenotypic neighbors.  \nAlthough the Velociraptor workflow successfully identified clinically relevant cell populations across \nhuman disease datasets, there are some limitations to the approach. First and foremost, rigorous \nexperimental design and execution remain paramount to accurate cell identification using \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nunsupervised machine learning. As with pooled analysis, proper data QC and batch normalization \nimprove results from VR-Claw. Batch effects are less of a concern when performing individual \nanalyses using VR-Eye so long as the overall pattern of protein expression remains consistent across \nbatches. Prior to data collection, careful consi deration and design of antibody panels will determine \nhow well Velociraptor can perform. A panel must have sufficient overlap in features with any dataset it \nis to be compared to, so that populations can be identified across datasets. Additionally, a MEM label \nis designed to capture the full range of data that can be collected on a given single cell platform. \nPerforming MEM on a small subset of cells that does not represent the entire data scale of a platform \nmay inflate the enrichment levels of proteins which could lead to inaccurate downstream cross-\ndataset identification. VR-Eye uses a reference phenotype to find similar cells across datasets, and \nthe specification of features in the searching label is the most crucial step when using this algorithm. \nInclusion of too many features may be over-optimized to the specific set of cells used to generate that \nlabel. We therefore recommend careful optimization and validation of a search label to ensure its \ngeneralizability. Standard statistical validation approaches used here, including cluster stability testing \nand cross validation, should be performed to confirm any Velociraptor findings.  \nWith the rapid increase in high-dimensional single cell technologies that can generate distinct \ninformation about cell populations, there is a need to match similar cells across datasets to learn a \nmore detailed understanding of cell identity. Therefore, robust measurements of cell identity are \nrequired to accurately identify cells across data types and thus to hypothesize about their role in \ndisease. Multiple continuous cell comparisons to well-established populations using Velociraptor will \nprovide a new perspective on cell identity that could be used to implicate similar cell populations \nacross disease settings. \n \nMethods \nGBM Tissue Collection and Processing for Dataset 2 \nRecurrent GBM tumors were surgically resected at Vanderbilt University Medical Center between \n2014 and 2017. All samples were collected with written informed consent under Institutional Review \nBoard protocol #131870 and in accordance with the Declaration of Helsinki. Tumor samples were \ndissociated into single cell suspensions as previously described \n29. \n \nIEI and Donor PBMC Collection for Dataset 3 \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nBlood was collected from IEI patients with written informed consent under Institutional Review Board \nprotocol #182228, from healthy donors with written informed consent under Institutional Review \nBoard protocols #131311 and #191562, and in accordance with the Declaration of Helsinki. 100mL of \nblood per donor were collected via venipuncture into heparin tubes (Becton Dickinson). Blood was \ndiluted 1:4 with PBS, placed into a Ficoll-Paque Plus density gradient (GE Lifesciences), and \ncentrifuged at 400 x g for 30 min. Next, buffy coat s were isolated, washed with PBS, and centrifuged \nat 500 x g for 10 min. Cell pellets were resuspended in ACK lysis buffer for 5 min, washed with X, and \ncryopreserved at 1 x 10\n7 cells/mL in 10% DMSO in FBS at 80°C. \n \nMetal-conjugated antibodies \nAll antibodies used for mass cytometry analysis are listed in Supplementary Table 1 and \nSupplementary Table 2 . Pre-conjugated, metal-tagged antibodies were purchased from Fluidigm, \nand unconjugated pre-conjugated purchased in purified form and custom conjugated using the \nMaxparX8 Antibody Labeling Kit (Fluidigm) according to manufacturer’s protocol.  \n \nCell Preparation and Mass Cytometry \nAntibody staining and mass cytometry analyses were performed as previously described 48,49. \nCryopreserved samples were rapidly thawed in a 37°C water bath and resuspended in complete \nRPMI 1640 supplemented with 10% FBS and 50 U/ml penicillin-streptomycin (Thermo Scientific \nHyClone). Cells were washed with serum-free RPMI 1640 and rested for 15 min.  Next, cells were \nwashed with serum-free RPMI 1640 and stained with a 103Rh Cell-ID intercalator (Fluidigm) at a final \nconcentration of 1 \nμ M for 5 min at RT. Cells were then washed with PBS + 1% BSA to quench \nstaining. Samples were then stained with the appropriate surface antibody cocktail ( Supplementary \nTable 1 and Supplementary Table 2) at RT for 30 min. Next, cells  were washed with PBS and fixed \nwith 2% formaldehyde at room temperature for 20 min. Cells we re washed with PBS and \npermeabilized with ice cold methanol overnight at 20°C. The following morning, cells were washed \nwith PBS then washed with PBS + 1% BSA. Next, samples were stained with intracellular antibodies \nat RT for 30 min and washed with PBS. Cells were intercalated wi th IridiumCell-ID at a final \nconcentration of 125 nM in PBS + 1.6% formaldehyde at 4°C overnight. On the day of data collection, \ncells were washed once with PBS and was hed once in ultrapure deionized water. Samples were \nresuspended in ultrapure deionized water with 10% EQ four element calibration beads (Fluidigm) and \nfiltered through a 40-mm FACS filter tube. Data were collected on a Helios CyTOF 3.0 (Fluidigm) and \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nstored in FCS files. Instrument quality control and tuning processes were performed following the \nguidelines for the daily instrument operation.  \n \nData Processing \nRaw mass cytometry files were normalized using the MATLAB bead normalization tool 50 prior to \nupload to the cloud-based analysis platform Cytobank 51. Data were arcsinh transformed with a \ncofactor of 5. Cell doublets were excluded using Gaussian parameters, beads were excluded, and \nintact live cells were selected for downstream analysis based on DNA content measured via Iridium \nand Rhodium intercalation. In Dataset 2 , CD45 -CD31- GBM cells were selected for downstream \nanalyses. In Dataset 3, CD45+ cells from all donors were plotted on a common t-SNE, and T cells \nwere selected based on CD3 expression. Donors with at least 2,000 T cells were included in \nsubsequent analyses. \n \nVelociraptor-Claw Algorithm \nThe Velociraptor-Claw algorithm includes dimensionality reduction, definition of local phenotypic \nneighborhoods, iterative testing of cell neighborhoods for association with an external variable, and \nphenotypic quantification of identified cells of interest. Here, dimensionality reduction was performed \non using t-SNE either in R or on Cytobank with a perplexity of 60 and 10,000 iterations. Local \nphenotypic neighborhoods were defined as the k-nearest neighbors (KNN) for every cell in the low \ndimensional embedding of the dataset using the fast nearest neighbors (FNN) package in R. The \nvalue of k was defined to be the square root of the total number of cells included in the analysis \n52. \nWithin each neighborhood, patients are categorized as Low or High for that neighborhood based on \ntheir abundance of cells residing in that neighborhood compared to the interquartile range of patient \ncell abundance for that neighborhood as described \n14. A univariate Cox proportional hazards model \nwas used to investigate the association of each neighborhood with patient overall survival time. The \nindex cell of each neighborhood was annotated with its neighborhood’s effect size (hazard ratio) and \nstatistical significance (p-value). Cells are then clustered into prognostic populations using DBSCAN \n(from the dbscan package in R). A univariate Cox proportional hazards model is then performed on \neach prognostic population to ensure that it is still associated with overall survival following DBSCAN \nclustering. Prognostic populations can be filtered by the number of cells in that cluster or by the \nnumber of samples in that cluster. Finally, the phenotypes of prognostic populations are quantified \nusing MEM \n17. \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nThe modified version of VR-Claw used to analyze Dataset 3  identified cells that were enriched \n(>95%) or lacking (<5%) from categories of patients within a dataset (i.e., genotype) following cell \nneighborhood definition. Genotype-specific cells were clustered using DBSCAN and filtered such that \nfinal clusters contained at least 25 cells and were consistently enriched/lacking across all patients \nwithin that genotype class. All steps downstream of dimensionality reduction in both versions of VR-\nClaw are deterministic. \n \nVelociraptor-Eye Algorithm \nThe Velociraptor-Eye algorithm includes dimensionality reduction, definition of local phenotypic \nneighborhoods, and phenotypic similarity calculation of each neighborhood compared to a reference \nphenotype. A reference phenotype that is written in the form of a MEM label is supplied as input. As \ndescribed above, dimensionality reduction was performed on using t-SNE either in R or on Cytobank \nwith a perplexity of 60 and 10,000 iterations. Local phenotypic neighborhoods were defined as the k-\nnearest neighbors (KNN) for every cell in the low dimensional embedding of the dataset using the fast \nnearest neighbors (FNN) package in R. The value of k was defined to be 60. MEM was used to \nquantify the phenotype of each cell neighborhood, and the similarity of each neighborhood to the \nreference population was calculated using the root-mean-square deviation (RMSD) between \nneighborhood and reference MEM labels (Equation 1\n).  \n/g1855/g1857/g1864/g1864 /g1871/g1861/g1865/g1861/g1864/g1853/g1870/g1861/g1872/g1877 /g1871/g1855/g1867/g1870/g1857 /g3404 100 /g3398  /g3497 1\n/g1839 /g3533/g4666/g1871/g1855/g1867/g1870/g1857 /g3015/g3003/g3009/g3036/g3398 /g1871/g1855/g1867/g1870/g1857 /g3019/g3006/g3007/g3036/g4667 /g2870\n/g3014\n/g3036/g2880/g2869\n,                              /g46661/g4667  \nIn Equation 1, M represents the number of markers shared between the reference MEM label and the \nMEM label of the cell neighborhood, score\nNBH denotes the cell neighborhood’s MEM score for a \nparticular marker, and score REF denotes the reference label’s MEM score for the same marker. \nFollowing dimensionality reduction, each step of VR-Eye is deterministic. \n \nData Availability  \nDatasets analyzed in this manuscript are online at FlowRepository \n53, and will be made public upon \nacceptance. Transparent analysis scripts for datasets in this manuscript will be made available on the \nCytoLab Github page (https://github.com/cytolab/) with open-source code and commented \nRmarkdown analysis walkthroughs upon acceptance.   \n \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nAcknowledgements \nWe thank Vanderbilt’s Cancer and Immunology Core as well as all the surgeons, patients, and \nfamilies that supported this work. We thank Hum an Immunology Discovery Initiative collaborators, \nincluding Todd Bartkowiak, for helpful discussions of the inborn errors of immunity data, and we thank \nCaroline Roe for discussions of cell identification algorithms. Research was supported by the \nfollowing funding resources: NIH/NCI grants R01 NS096238 (RAI, JMI), R01 CA226833 (JMI, CEC, \nSM, MJH), R01 NS118580 (RAI), U01 AI125056 (JMI), U54 CA217450 (JMI, MJH), T32GM137793 \n(CEC), the Vanderbilt-Ingram Cancer Center (VICC, P30 CA68485), the Michael David Greene Brain \nCancer Fund (RAI, JMI), the Southeastern Brain Tumor Foundation (RAI, JMI), a gift from Daniel F \nHewins (RAI), the Ben & Catherine Ivy Foundation (RAI, JMI), and by the Human Immunology \nDiscovery Initiative of the Vanderbilt Center for Immunobiology. \n \nAuthor Contributions \nCEC and JMI designed the study and conceptualized the velociraptor workflow. MJH and SM \ncollected data. CEC and CM coded data analysis scripts. CEC, CM, and JMI performed mass \ncytometry data analysis and interpretation. SK and JAC provided clinical care to patients with IEIs \nand identified GATA cases. RAI, JMI, LBC, and RCT developed Vanderbilt’s human glioblastoma \nresearch program. RAI, MJH, and SM coordinated intraoperative tissue collection by research teams. \nLBC, RCT, and KDW provided freshly resected glioblastoma tissue specimens, including recurrence \nsamples. CEC and JMI wrote the manuscript. JCR, RAI, and JMI provided financial support. All \nauthors contributed to reviewing and editing the manuscript. \n \nDeclaration of Interests \nJCR is a founder and scientific advisory board member of Sitryx Therapeutics. \n  \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nReferences \n1 Medina, S., Ihrie, R. A. & Irish, J. M. Learning cell identity in immunology, neuroscience, and cancer. \nSemin Immunopathol 45, 3-16 (2023). https://doi.org/10.1007/s00281-022-00976-y \n2 Loken, M. R. Multidimensional data analysis in immunophenotyping. Curr Protoc Cytom Chapter 10, \nUnit 10 14 (2001). https://doi.org/10.1002/0471142956.cy1004s00 \n3 Givan, A. L. Flow cytometry: an introduction. Methods Mol Biol 699, 1-29 (2011). \nhttps://doi.org/10.1007/978-1-61737-950-5_1 \n4 Qiu, P.  et al. Extracting a cellular hierarchy from high-dimensional cytometry data with SPADE. Nat \nBiotechnol 29, 886-891 (2011). https://doi.org/10.1038/nbt.1991 \n5 Greenplate, A. R., Johnson, D. B., Ferrell, P. B., Jr. & Irish, J. M. Systems immune monitoring in cancer \ntherapy. Eur J Cancer 61, 77-84 (2016). https://doi.org/10.1016/j.ejca.2016.03.085 \n6 Spitzer, M. H. & Nolan, G. P. Mass Cytometry: Single Cells, Many Features. Cell 165, 780-791 (2016). \nhttps://doi.org/10.1016/j.cell.2016.04.019 \n7 Amir el, A. D.  et al. viSNE enables visualization of high dimensional single-cell data and reveals \nphenotypic heterogeneity of leukemia. Nat Biotechnol 31, 545-552 (2013). \nhttps://doi.org/10.1038/nbt.2594 \n8 Becht, E.  et al. Dimensionality reduction for visualizing single-cell data using UMAP. Nat Biotechnol \n(2018). https://doi.org/10.1038/nbt.4314 \n9 Van Gassen, S.  et al. FlowSOM: Using self-organizing maps for visualization and interpretation of \ncytometry data. Cytometry A 87, 636-645 (2015). https://doi.org/10.1002/cyto.a.22625 \n10 Barone, S. M.  et al. Unsupervised machine learning reveals key immune cell subsets in COVID-19, \nrhinovirus infection, and cancer therapy. eLife (2021). https://doi.org/10.7554 \n11 Burns, T. J., Nolan, G. P. & Samusik, N. Continuous visualization of differences between biological \nconditions in single cell data. bioRxiv (2018). https://doi.org/10.1101/337485 \n12 Dann, E., Henderson, N. C., Teichmann, S. A., Morgan, M. D. & Marioni, J. C. Differential abundance \ntesting on single-cell data using k-nearest neighbor graphs. Nat Biotechnol 40, 245-253 (2022). \nhttps://doi.org/10.1038/s41587-021-01033-z \n13 Bruggner, R. V., Bodenmiller, B., Dill, D. L., Tibshirani, R. J. & Nolan, G. P. Automated identification of \nstratifying signatures in cellular subpopulations. Proc Natl Acad Sci U S A 111, E2770-2777 (2014). \nhttps://doi.org/10.1073/pnas.1408792111 \n14 Leelatian, N. et al.  Unsupervised machine learning reveals risk stratifying glioblastoma tumor cells. Elife \n9 (2020). https://doi.org/10.7554/eLife.56879 \n15 Arvaniti, E. & Claassen, M. Sensitive detection of rare disease-associated cell subsets via \nrepresentation learning. Nat Commun 8, 14825 (2017). https://doi.org/10.1038/ncomms14825 \n16 Kramer, K. J. et al. Single-cell profiling of the antigen-specific response to BNT162b2 SARS-CoV-2 \nRNA vaccine. Nat Commun 13, 3466 (2022). https://doi.org/10.1038/s41467-022-31142-5 \n17 Diggins, K. E., Greenplate, A. R., Leelatian, N., Wogsland, C. E. & Irish, J. M. Characterizing cell \nsubsets using marker enrichment modeling. Nat Methods 14, 275-278 (2017). \nhttps://doi.org/10.1038/nmeth.4149 \n18 Greenplate, A. R.  et al. Computational Immune Monitoring Reveals Abnormal Double-Negative T Cells \nPresent across Human Tumor Types. Cancer Immunol Res 7, 86-99 (2019). \nhttps://doi.org/10.1158/2326-6066.CIR-17-0692 \n19 Levine, J. H.  et al. Data-Driven Phenotypic Dissection of AML Reveals Progenitor-like Cells that \nCorrelate with Prognosis. Cell 162, 184-197 (2015). https://doi.org/10.1016/j.cell.2015.05.047 \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\n20 Kuchroo, M.  et al. Multiscale PHATE identifies multimodal signatures of COVID-19. Nat Biotechnol 40, \n681-691 (2022). https://doi.org/10.1038/s41587-021-01186-x \n21 Welch, J. D.  et al. Single-Cell Multi-omic Integration Compares and Contrasts Features of Brain Cell \nIdentity. Cell 177, 1873-1887 e1817 (2019). https://doi.org/10.1016/j.cell.2019.05.006 \n22 Yang, Z. & Michailidis, G. A non-negative matrix factorization method for detecting modules in \nheterogeneous omics multi-modal data. Bioinformatics 32, 1-8 (2016). \nhttps://doi.org/10.1093/bioinformatics/btv544 \n23 Stuart, T. et al. Comprehensive Integration of Single-Cell Data. Cell 177, 1888-1902 e1821 (2019). \nhttps://doi.org/10.1016/j.cell.2019.05.031 \n24 Zhu, B.  et al. Robust single-cell matching and multimodal analysis using shared and distinct features. \nNat Methods 20, 304-315 (2023). https://doi.org/10.1038/s41592-022-01709-7 \n25 Barsch, M. et al. T-cell exhaustion and residency dynamics inform clinical outcomes in hepatocellular \ncarcinoma. J Hepatol 77, 397-409 (2022). https://doi.org/10.1016/j.jhep.2022.02.032 \n26 Benjamini, Y. & Hochberg, Y. Controlling the Fals e Discovery Rate A Practical and Powerful Approach \nto Multiple Testing. Jornal of the Royal Statistical Society (1995).  \n27 Olin, A.  et al. Stereotypic Immune System Development in Newborn Children. Cell 174, 1277-1292 \ne1214 (2018). https://doi.org/10.1016/j.cell.2018.06.045 \n28 Spitzer, M. H.  et al. IMMUNOLOGY. An interactive reference framework for modeling a dynamic \nimmune system. Science 349, 1259425 (2015). https://doi.org/10.1126/science.1259425 \n29 Leelatian, N. et al.  Single cell analysis of human tissues and solid tumors with mass cytometry. \nCytometry B Clin Cytom 92, 68-78 (2017). https://doi.org/10.1002/cyto.b.21481 \n30 Picard, C. & Casanova, J. L. Novel primary immunodeficiencies. Adv Exp Med Biol 568, 89-99 (2005). \nhttps://doi.org/10.1007/0-387-25342-4_6 \n31 Bousfiha, A.  et al. The 2022 Update of IUIS Phenotypical Classification for Human Inborn Errors of \nImmunity. J Clin Immunol 42, 1508-1520 (2022). https://doi.org/10.1007/s10875-022-01352-z \n32 Kaviany, S.  et al. Systems Immunology Analyses of STAT1 Gain-of-Function Immune Phenotypes \nReveal Heterogeneous Response to IL-6 and Broad Immunometabolic Roles for STAT1. \nImmunohorizons 6, 447-464 (2022). https://doi.org/10.4049/immunohorizons.2200041 \n33 Tietscher, S.  et al. A comprehensive single-cell map of T cell exhaustion-associated immune \nenvironments in human breast cancer. Nat Commun 14, 98 (2023). https://doi.org/10.1038/s41467-022-\n35238-w \n34 Park, L. M., Lannigan, J. & Jaimes, M. C. OMIP-069: Forty-Color Full Spectrum Flow Cytometry Panel \nfor Deep Immunophenotyping of Major Cell Subsets in Human Peripheral Blood. Cytometry A 97, 1044-\n1051 (2020). https://doi.org/10.1002/cyto.a.24213 \n35 Diggins, K. E., Ferrell, P. B., Jr. & Irish, J. M. Methods for discovery and characterization of cell subsets \nin high dimensional mass cytometry data. Methods 82, 55-63 (2015). \nhttps://doi.org/10.1016/j.ymeth.2015.05.008 \n36 Saeys, Y., Van Gassen, S. & Lambrecht, B. N. Computational flow cytometry: helping to make sense of \nhigh-dimensional immunology data. Nat Rev Immunol 16, 449-462 (2016). \nhttps://doi.org/10.1038/nri.2016.56 \n37 Takahashi, K. & Yamanaka, S. Induction of pluripotent stem cells from mouse embryonic and adult \nfibroblast cultures by defined factors. Cell 126, 663-676 (2006). \nhttps://doi.org/10.1016/j.cell.2006.07.024 \n38 Al-Hajj, M., Wicha, M. S., Benito-Hernandez, A., Morrison, S. J. & Clarke, M. F. Prospective \nidentification of tumorigenic breast cancer cells. Proc Natl Acad Sci U S A 100, 3983-3988 (2003). \nhttps://doi.org/10.1073/pnas.0530291100 \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\n39 Huang, E. H.  et al. Aldehyde dehydrogenase 1 is a marker for normal and malignant human colonic \nstem cells (SC) and tracks SC overpopulation during colon tumorigenesis. Cancer Res 69, 3382-3389 \n(2009). https://doi.org/10.1158/0008-5472.CAN-08-4418 \n40 Lapidot, T. et al. A cell initiating human acute myeloid leukaemia after transplantation into SCID mice. \nNature 367, 645-648 (1994). https://doi.org/10.1038/367645a0 \n41 Singh, S. K.  et al. Identification of human brain tumour initiating cells. Nature 432, 396-401 (2004). \nhttps://doi.org/10.1038/nature03128 \n42 Ruiz-Garcia, R. et al. Acquired Senescent T-Cell Phenotype Correlates with Clinical Severity in GATA \nBinding Protein 2-Deficient Patients. Front Immunol 8, 802 (2017). \nhttps://doi.org/10.3389/fimmu.2017.00802 \n43 Stein-O'Brien, G. L.  et al. Decomposing Cell Identity for Transfer Learning across Cellular \nMeasurements, Platforms, Tissues, and Species. Cell Syst 8, 395-411 e398 (2019). \nhttps://doi.org/10.1016/j.cels.2019.04.004 \n44 Altman, N. & Krzywinski, M. The curse(s) of dimensionality. Nat Methods 15, 399-400 (2018). \nhttps://doi.org/10.1038/s41592-018-0019-x \n45 Hamilton, W. L., Ying, R. & Leskovec, J. in 31st Conference on Neural Information Processing Systems    \n(Long Beach, CA, 2017). \n46 Scarselli, F., Gori, M., Tsoi, A. C., Hagenbuchner, M. & Monfardini, G. The graph neural network model. \nIEEE Trans Neural Netw 20, 61-80 (2009). https://doi.org/10.1109/TNN.2008.2005605 \n47 Zitnik, M. & Leskovec, J. Predicting multicellular function through multi-layer tissue networks. \nBioinformatics 33, i190-i198 (2017). https://doi.org/10.1093/bioinformatics/btx252 \n48 Ferrell, P. B., Jr.  et al. High-Dimensional Analysis of Acute Myeloid Leukemia Reveals Phenotypic \nChanges in Persistent Cells during Induction Therapy. PLoS One 11, e0153207 (2016). \nhttps://doi.org/10.1371/journal.pone.0153207 \n49 Greenplate, A. R.  et al. Myelodysplastic Syndrome Revealed by Systems Immunology in a Melanoma \nPatient Undergoing Anti-PD-1 Therapy. Cancer Immunol Res 4, 474-480 (2016). \nhttps://doi.org/10.1158/2326-6066.CIR-15-0213 \n50 Finck, R.  et al. Normalization of mass cytometry data with bead standards. Cytometry A 83, 483-494 \n(2013). https://doi.org/10.1002/cyto.a.22271 \n51 Kotecha, N., Krutzik, P. O. & Irish, J. M. Web-based analysis and publication of flow cytometry \nexperiments. Curr Protoc Cytom Chapter 10, Unit10 17 (2010). \nhttps://doi.org/10.1002/0471142956.cy1017s53 \n52 Rashidi, H. H., Tran, N. K., Betts, E. V., Howell,  L. P. & Green, R. Artificial Intelligence and Machine \nLearning in Pathology: The Present Landscape of Supervised Methods. Acad Pathol 6, \n2374289519873088 (2019). https://doi.org/10.1177/2374289519873088 \n53 Spidlen, J., Breuer, K., Rosenberg, C., Kotecha, N. & Brinkman, R. R. FlowRepository: a resource of \nannotated flow cytometry datasets associated with peer-reviewed publications. Cytometry. Part A : the \njournal of the International Society for Analytical Cytology 81, 727-731 (2012). \nhttps://doi.org/10.1002/cyto.a.22106 \n  \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nFigure Legends \nFigure 1 – Overview of Velociraptor.  \nA) The Velociraptor workflow includes two novel cell identification tools: (1) Velociraptor-Claw (VR-Claw), \nwhich reveals condition-specific cell populations (e.g., cells associated with overall survival), and (2) \nVelociraptor-Eye (VR-Eye), which seeks cells bas ed on a user-defined phenotype. VR-Claw utilizes local \nphenotypic cell neighborhoods to identify clinically relevant cell populations. The upper right panel in the red \nbox highlights cells associated with shorter (red) or longer (blue) survival times. Marker Enrichment Modeling \n(MEM) 17 can be used to automatically quantify population phenotypes, and MEM labels can then be used as \ninput to VR-Eye. VR-Eye quantifies and plots similarity to a specified phenotype of interest (i.e., with a MEM \nlabel) with purple indicating low similarity and red indicating high similarity. B) Biological challenges and \ndatasets explored in this manuscript. C) Key findings from Velociraptor for each biological challenge. D) \nStatistical validation approaches used to validate findings include k-fold cross validation, interrogation of cells \nnot used to learn a phenotype of interest, repeated down-sampling, and leave-one-out cross validation.  \n \nFigure 2 – Velociraptor refines GNP and GPP cell identification and patient classification.  \nA) VR-Claw analysis of CD31 -CD45- cells from the RAPID cohort. Cell neighborhoods are color coded based \non HR, p-value determined with a Cox proportional hazards model, and cluster stability across 5 iterations.  B) \nProtein enrichment of GNP and GPP cell subsets determined using absolute Marker Enrichment Modeling \n(MEM). C) Overall survival of GNP-High (red line) and GNP-Low (black line) patients. D) Overall survival of \nGPP-High (blue line) and GPP-Low (black line) patients. E) The leftmost portion of the plot shows abundance \nof GNP and GPP cells identified by VR-Claw across 5 folds of identification on training data. GNP abundance \nis shown in red, and GPP abundance is shown in blue. GNP-High and GPP-High abundance thresholds are \nshown with red and blue dotted lines, respectively. The two columns in the middle of the plot show patient \nstatus (GNP-High in red, GPP-High in blue, and Low for both populations in grey) as determined by \nVelociraptor (VR, middle left column) and RAPID (R, middle right column). The rightmost portion of the plot \nshows overall patient survival. The dotted black line indicates a median overall survival of 389 days. Censored \npatients are indicated with *. F) VR-Eye analysis seeking Population 1 using the reference MEM label shown in \n(B). A spectrum intensity scale indicates cell similarity with red indicating high similarity and purple indicating \nlow similarity. G) VR-Eye analysis seeking Population 2 using the reference MEM label shown in (B). A \nspectrum intensity scale indicates cell similarity with red indicating high similarity and purple indicating low \nsimilarity. H) Correlation between each patient’s percentage of GNP-like cells identified in Training Data and \nTesting Data by VR-Eye in the first round of 5-fold cross-validation. Pearson correlation tests were conducted. \nI) Correlation between each patient’s percentage of GPP-like cells identified in Training Data and Testing Data \nby VR-Eye in the first round of 5-fold cross-validation. Pearson correlation tests were conducted. \n \nFigure 3 – Velociraptor-Eye quantifies opposing cell identities and reveals shifts in glioblastoma tumor \ncell composition at recurrence.   \nA) Three representative t-SNEs each created using a single patient’s CD45\n-CD31- GBM cells (individual cell \ncounts ranged from 4,710 to 329,650 per patient for a total of 2,363,915 live GBM cells). A magma intensity \nscale represents cell density on the t-SNE axes. B) Three representative VR-Eye analyses of individual, \npatient-specific t-SNEs generated from CD45 -CD31- GBM cell populations. Plots display each patient’s cells’ \nsimilarity to GNP cells on the y-axis (GNP searching label: SOX2 +7 S100B+4 GFAP+1 p-NFkB+0) and GPP cells \non the x-axis (GPP searching label: EGFR +7 p-p38+2 CD49F+2 CD44+1 TUJ1+0). The dotted red horizontal line \nindicates the GNP-like similarity threshold of 83%, such that every cell above this line is considered “GNP-like”. \nThe dotted blue vertical line indicates the GPP-like similarity threshold of 82%, such that every cell to the right \nof this line is considered “GPP-like”. C) Patient-level VR-Eye summary. Red dots represent GNP-like cells, and \nblue dots represent GPP-like cells as determined by VR-Eye. Patients are ordered by decreasing range of \n%GPP cells to %GNP cells as determined by RAPID. Bars to the right represent overall survival time. Patient \ncodes are colored based on RAPID status, where red indicates a GNP-High patient and blue represents a \nGPP-High patient according to RAPID. D) Correlation between each patient’s percentage of GNP-like or GPP-\nlike cells, respectively, identified by RAPID in Dataset 1 and by VR-Eye on the entire GBM cell population per \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\ntumor. E) VR-Eye analyses of individual patient t-SNEs generated from CD45 -CD31- GBM cell populations \nfrom three matched primary and recurrent GBM (individual t-SNEs range from 7,498 to 134,876 cells per \npatient). Density plots show the difference between each cell’s GNP and GPP similarity. Cells colored blue had \na GPP similarity value greater than 82% and are considered “GPP-like”. Cells colored red had a GNP similarity \nvalue greater than 83% and are considered “GNP-like”. Primary tumors are labeled with “_P”, and recurrent \ntumors are labeled with “_R”. F) Quantification of GNP-like and GPP-like cell abundance in paired primary and \nrecurrent tumors. \n \nFigure 4 – Velociraptor reveals, characterizes, and confirms CD57 + T cells that are enriched in patients \nwith GATA2 haploinsufficiency.   \nA) Cell density on t-SNE axes generated using T cells from donors in Set 1. Set 1 includes one adult healthy \ndonor, two IEI donors with GATA2  haploinsufficiency, and two IEI donors without GATA2 haploinsufficiency. \nThe left plot shows equally sampled T cells from all patients in Set 1 (N=5 donors, n=10,000 cells). The right \nplot shows equally sampled T cell from the GATA2  haploinsufficiency patients in Set 1 only (N=2 donors, \nn=4,000 cells). A magma color scale indicates cell density on the t-SNE axes where purple indicates low \ndensity and yellow represents high density.  B) VR-Claw analysis on equally sampled T cells from IEI patients \nand healthy donors in Set 1 (N=5 donors, n=10,000 cells). Red and orange denote populations enriched in \nGATA2 haploinsufficiency patients; blue and purple denote populations lacking from GATA2 haploinsufficiency \npatients. C) Quantification of cell abundance with respect to all T cells for each VR-Claw population. D) Filtered \nMEM labels show proteins enriched in VR-Claw Populations. Proteins with a relative MEM score of > 2 were \nincluded. E) Cell density on t-SNE axes. The top plot was generated using all T cells from a healthy pediatric \ndonor included in Set 2 (n=\n 55,201 cells). The bottom plot was generated using all T cells from an IEI patient \nwith GATA2 haploinsufficiency included in Set 1 (n=  278,870 cells). A magma color scale indicates cell density \non the t-SNE axes where purple indicates low density and yellow represents high density.  F) Representative \nVR-Eye analyses seeking VR-Claw Populations 1- 6 in T cells from a pediatric healthy donor (PHD1, top row) \nand a GATA2 haploinsufficient patient (IDDI-040, bottom row). MEM labels shown in Panel D were used as the \nsearching label for each population. A rainbow intensity scale indicates similarity to the sought population with \nred representing high and purple representing low similarity. G) Quantification of T cell population abundances \nin all IEI patients and healthy donors (N=51) as determined by VR-Eye. Gold stars indicate high priority \npopulations to continue studying, and red triangles indicate populations that require further validation using \nadditional GATA2 haploinsufficiency samples. \n \nFigure 5 - VR-Eye accurately identifies cell types in an external imaging dataset using reference \nphenotypes generated by different single-cell platforms.   \nA) Plots along the top of the panel show the expression of selected proteins on an arcsinh scale. Grey \nindicates low expression and purple indicates high expression. The bottom left plot shows cell types as \ndetermined by Tietscher et al on a representative slide. The bottom right plot shows VR-Eye analyses for each \ncell type shown in the upper left plot. VR similarity was plotted as a gradient for each cell type as indicated. B) \nAll IMC cells (n=257,076) embedded in a common t-SNE. Cells identified as memory CD4 T cells by biaxial \ngating are plotted in dark green. C) VR-Eye analyses seeking memory CD4 T cells based on reference \npopulations measured with various platforms. The theory-generated label was written based on known protein \nexpression patterns of memory CD4 T cells. The CyTOF-generated label was calculated on memory CD4 T \ncells that were gated in a healthy donor from Dataset 3. The spectral-generated label was calculated on \nmemory CD4 T cells that were gated in Dataset 5. A rainbow intensity scale indicates similarity to the sought \npopulation with red representing high and purple representing low similarity. D) MEM label generated from the \ngated memory CD4 population. E) \nΔ MEM analyses comparing MEM labels from each VR-identified memory \nCD4 T cell populations to the MEM label generated from the gated memory CD4 T cell population. \n \nSupplementary Figure S1 – Velociraptor-Claw repr oducibly identifies prognos tic cells and classifies \npatients in Dataset 1.  \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nA) VR-Claw analyses of CD45 -CD31- GBM cells across each round of 5-fold cross validation. Cell \nneighborhoods are color coded based on HR and p-value. B)  Patient classification according to prognostic \npopulation abundance by RAPID and VR-Claw. \n \nSupplementary Figure S2 – Optimization of prognostic GBM population identification. \nA) Expression levels of protein markers on common t-SNE axes for 131,880 CD45\n-CD31- GBM cells in Dataset \n1. A spectrum intensity scale indicates expression levels with blue representing low expression and red \nrepresenting high expression.  B) VR-Eye analysis seeking GNP-like cells (dark red in the middle plot) and \nGPP-like cells (dark blue in the middle plot). GNP-like cells, identified in the left plot, were sought with an input \nof SOX2+7 S100B+4 GFAP+1 p-NFkB+0).  GPP-like cells, identified in the right plot, were sought with an input of \nEGFR+7 p-p38+2 CD49F+2 CD44+1 TUJ1+0) in the RAPID cohort (4,710 CD45 -CD31- GBM cells per patient). A \nspectrum intensity scale indicates cell similarity with red indicating high similarity and purple indicating low \nsimilarity. C) Correlation between each patient’s percentage of GNP-like or GPP-like cells, respectively, \nidentified by RAPID and VR-Eye in Dataset 1. Pearson correlation tests were conducted. D) Comparison of cell \nidentification approaches across multiple iterations of each approach. For VR-Eye and RAPID, each algorithm \nwas applied to 10 different t-SNEs generated from the same sampling of Dataset 1 to identify GNP-like and \nGPP-like populations. Biaxial gating was performed by four different cytometry experts. \n \nSupplementary Figure S3 – VR-Claw reproducibly reveals GATA2 haploinsufficiency-specific cells in \nDataset 3. \nA) VR-Claw analyses across 10 runs using 2,000 randomly sampled T cells per donor in Set 1 (N=5, n=10,000 \ncells total). Each t-SNE was independently generated for each run using that run’s sample of T cells. Red and \norange denote populations enriched in GATA2 haploinsufficiency patients; blue and purple denote populations \nlacking from GATA2 haploinsufficiency patients. B) RMSD analysis on MEM labels generated from GATA2 \nhaploinsufficiency-specific populations identified in each of the 10 runs of VR-Claw. Similarity values were \ncalculated by subtracting normalized RMSD values from 100. A rainbow intensity scale indicates population \nsimilarity with red indicating high similarity and purple indicating low similarity. Stable populations that appear \nin at least 5 out of 10 runs are marked to the right of the heatmap with either a blue or a dark red line. Blue \nlines indicate populations that were lacking from GATA2 haploinsufficiency patients, and dark red lines indicate \npopulations that were enriched in GATA2 haploinsufficiency patients as determined by VR-Claw. Phenotypes \nof stable populations are summarized to the far right.  \n \nSupplementary Figure S4 – 41-dimensional phenotypic labels outperform filtered phenotypes in \nDataset 4.  \nA) Expression levels of protein markers on common t-SNE axes for 257,076 cells in Dataset 4. A spectrum \nintensity scale indicates expression levels with blue representing low expression and red representing high \nexpression. Markers not included as t-SNE input are labeled in blue. B) Comparison of VR-Eye cell \nidentification using full 41-dimensional MEM labels or filtered MEM labels (filtered labels range from 4-\ndimensional to 28-dimensional) based on population protein enrichment as search input. C) Comparison of \npopulation homogeneity across different cell identification methods. The homogeneity score equals the inverse \nof the median RMSD with respect to the median protein expression for each marker for each population. \n \nSupplementary Figure S5 – Velociraptor identifies CD4\n+CD8+ double positive memory T cells present in \nhuman breast cancer tumors.  \nA) Gating scheme to select memory CD4 T cells shown on a representative patient. B) VR-Eye analysis \nseeking memory CD4 T cells using a minimum marker label. C) Δ MEM analysis comparing the VR-identified \nmemory CD4 T cell population to the gated population. Δ MEM labels were calculated by subtracting the \nabsolute MEM label of the VR-identified population from the absolute MEM label of the manually gated \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\npopulation. D) Quantification of search accuracy for each memory CD4 T cell label sought. Precision is plotted \nin purple, recall is plotted in gold, and F1-measures are plotted in green. Similarity thresholds were optimized \nto best capture the gated population. E) Biaxial plots of CD4 and CD8a expression are shown for cells included \nin Velociraptor-identified memory CD4 (left plot) or CD8 (middle plot) populations. Contour indicates density. A \nspectrum intensity scale indicates each cell’s similarity to the memory CD4\n+CD8+ double positive reference \nphenotype (right plot) as determined by maximum theory searches. F) Density plots show the difference \nbetween each cell’s memory CD4 similarity and memory CD8 similarity to the maximum theory label. Biaxially \ngated memory CD4 T cells are shown in light blue, and biaxially gated memory CD8 T cells are shown in dark \nblue. Ungated cells are shown in light grey. G) Biaxial plots of CD4 and CD8a expression are shown for cells \nwith a memory CD4 T cell similarity value greater than or equal to 65 and a memory CD8 – CD4 similarity \ndifference less than or equal to -5. Contour indicates density. \n \n  \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nCross et al. – Figure 1 \n \nFigure 1 – Overview of Velociraptor.  \nA) The Velociraptor workflow includes two novel cell identification tools: (1) Velociraptor-Claw (VR-Claw), which reveals \ncondition-specific cell populations (e.g., cells associated with overall survival), and (2) Velociraptor-Eye (VR-Eye), which \nseeks cells based on a user-defined phenotype. VR-Claw utilizes local phenotypic cell neighborhoods to identify clinically \nrelevant cell populations. The upper right panel in the red box highlights cells associated with shorter (red) or longer (blue) \nsurvival times. Marker Enrichment Modeling (MEM) \n17 can be used to automatically quantify population phenotypes, and \nMEM labels can then be used as input to VR-Eye. VR-Eye quantifies and plots similarity to a specified phenotype of \ninterest (i.e., with a MEM label) with purple indicating low similarity and red indicating high similarity.  B) Biological \nchallenges and datasets explored in this manuscript.  C) Key findings from Velociraptor for each biological challenge. D) \nStatistical validation approaches used to validate findings include k-fold cross validation, interrogation of cells not used to  \nlearn a phenotype of interest, repeated down-sampling, and leave-one-out cross validation.   \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nCross et al. - Figure 2 \n \nFigure 2 – Velociraptor refines GNP and GPP cell identification and patient classification.  \nA) VR-Claw analysis of CD31 -CD45- cells from the RAPID cohort. Cell neighborhoods are color coded based on HR, p-\nvalue determined with a Cox proportional hazards model, and cluster stability across 5 iterations. B) Protein enrichment of \nGNP and GPP cell subsets determined using absolute Marker Enrichment Modeling (MEM). C) Overall survival of GNP-\nHigh (red line) and GNP-Low (black line) patients. D)  Overall survival of GPP-High (blue line) and GPP-Low (black line) \npatients. E) The leftmost portion of the plot shows abundance of GNP and GPP cells identified by VR-Claw across 5 folds \nof identification on training data. GNP abundance is shown in red, and GPP abundance is shown in blue. GNP-High and \nGPP-High abundance thresholds are shown with red and blue dotted lines, respectively. The two columns in the middle of \nthe plot show patient status (GNP-High in red, GPP-High in blue, and Low for both populations in grey) as determined by \nVelociraptor (VR, middle left column) and RAPID (R, middle right column). The rightmost portion of the plot shows overall \npatient survival. The dotted black line indicates a median overall survival of 389 days. Censored patients are indicated \nwith *. F) VR-Eye analysis seeking Population 1 using the reference MEM label shown in (B). A spectrum intensity scale \nindicates cell similarity with red indicating high similarity and purple indicating low similarity. G) VR-Eye analysis seeking \nPopulation 2 using the reference MEM label shown in (B). A spectrum intensity scale indicates cell similarity with red \nindicating high similarity and purple indicating low similarity. H) Correlation between each patient’s percentage of GNP-like \ncells identified in Training Data and Testing Data by VR-Eye in the first round of 5-fold cross-validation. Pearson \ncorrelation tests were conducted. I) Correlation between each patient’s percentage of GPP-like cells identified in Training \nData and Testing Data by VR-Eye in the first round of 5-fold cross-validation. Pearson correlation tests were conducted.\n \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nCross et al. - Figure 3 \n \nFigure 3 – Velociraptor-Eye quantifies opposing cell identities and reveals shifts in glioblastoma tumor cell \ncomposition at recurrence.   \nA) Three representative t-SNEs each created using a single patient’s CD45 -CD31- GBM cells (individual cell counts \nranged from 4,710 to 329,650 per patient for a total of 2,363,915 live GBM cells). A magma intensity scale represents cell \ndensity on the t-SNE axes. B) Three representative VR-Eye analyses of individual, patient-specific t-SNEs generated from \nCD45-CD31- GBM cell populations. Plots display each patient’s cells’ similarity to GNP cells on the y-axis (GNP searching \nlabel: SOX2+7 S100B+4 GFAP+1 p-NFkB+0) and GPP cells on the x-axis (GPP searching label: EGFR +7 p-p38+2 CD49F+2 \nCD44+1 TUJ1+0). The dotted red horizontal line indicates the GNP-like similarity threshold of 83%, such that every cell \nabove this line is considered “GNP-like”. The dotted blue vertical line indicates the GPP-like similarity threshold of 82%, \nsuch that every cell to the right of this line is considered “GPP-like”. C) Patient-level VR-Eye summary. Red dots represent \nGNP-like cells, and blue dots represent GPP-like cells as determined by VR-Eye. Patients are ordered by decreasing \nrange of %GPP cells to %GNP cells as determined by RAPID. Bars to the right represent overall survival time. Patient \ncodes are colored based on RAPID status, where red indicates a GNP-High patient and blue represents a GPP-High \npatient according to RAPID. D) Correlation between each patient’s percentage of GNP-like or GPP-like cells, respectively, \nidentified by RAPID in Dataset 1 and by VR-Eye on the entire GBM cell population per tumor. E) VR-Eye analyses of \nindividual patient t-SNEs generated from CD45\n-CD31- GBM cell populations from three matched primary and recurrent \nGBM (individual t-SNEs range from 7,498 to 134,876 cells per patient). Density plots show the difference between each \ncell’s GNP and GPP similarity. Cells colored blue had a GPP similarity value greater than 82% and are considered “GPP-\nlike”. Cells colored red had a GNP similarity value greater than 83% and are considered “GNP-like”. Primary tumors are \nlabeled with “_P”, and recurrent tumors are labeled with “_R”. F) Quantification of GNP-like and GPP-like cell abundance \nin paired primary and recurrent tumors.\n  \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nCross et al. - Figure 4 \n \nFigure 4 – Velociraptor reveals, characterizes, and confirms CD57 + T cells that are enriched in patients with \nGATA2 haploinsufficiency.   \nA) Cell density on t-SNE axes generated using T cells from donors in Set 1. Set 1 includes one adult healthy donor, two \nIEI donors with GATA2 haploinsufficiency, and two IEI donors without GATA2 haploinsufficiency. The left plot shows \nequally sampled T cells from all patients in Set 1 (N=5 donors, n=10,000 cells). The right plot shows equally sampled T \ncell from the GATA2 haploinsufficiency patients in Set 1 only (N=2 donors, n=4,000 cells). A magma color scale indicates \ncell density on the t-SNE axes where purple indicates low density and yellow represents high density.  B) VR-Claw \nanalysis on equally sampled T cells from IEI patients and healthy donors in Set 1 (N=5 donors, n=10,000 cells). Red and \norange denote populations enriched in GATA2 haploinsufficiency patients; blue and purple denote populations lacking \nfrom GATA2 haploinsufficiency patients. Populations that were statistically enriched in or lacking from both GATA2 \nhaploinsufficient patients are outlined in magenta.  C) Quantification of cell abundance with respect to all T cells for each \nVR-Claw population. D) Filtered MEM labels show proteins enriched in VR-Claw Populations. Proteins with a relative \nMEM score of > 2 were included. E) Cell density on t-SNE axes. The top plot was generated using all T cells from a \nhealthy pediatric donor included in Set 2 (n=\n 55,201 cells). The bottom plot was generated using all T cells from an IEI \npatient with GATA2 haploinsufficiency included in Set 1 (n= 278,870 cells). A magma color scale indicates cell density on \nthe t-SNE axes where purple indicates low density and yellow represents high density.  F) Representative VR-Eye \nanalyses seeking VR-Claw Populations 1- 6 in T cells from a pediatric healthy donor (PHD1, top row) and a GATA2 \nhaploinsufficient patient (IDDI-040, bottom row). MEM labels shown in Panel D were used as the searching label for each \npopulation. A rainbow intensity scale indicates similarity to the sought population with red representing high and purple \nrepresenting low similarity. G) Quantification of T cell population abundances in all IEI patients and healthy donors (N=51) \nas determined by VR-Eye. Gold stars indicate high priority populations to continue studying, and red triangles indicate \npopulations that require further validation using additional GATA2 haploinsufficiency samples.\n \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nCross et al. - Figure 5 \n \nFigure 5 - VR-Eye accurately identifies cell types in an external imaging dataset using reference phenotypes \ngenerated by different single-cell platforms.   \nA) Plots along the top of the panel show the expression of selected proteins on an arcsinh scale. Grey indicates low \nexpression and purple indicates high expression. The bottom left plot shows cell types as determined by Tietscher et al on \na representative slide. The bottom right plot shows VR-Eye analyses for each cell type shown in the upper left plot. VR \nsimilarity was plotted as a gradient for each cell type as indicated. B) All IMC cells (n=257,076) embedded in a common t-\nSNE. Cells identified as memory CD4 T cells by biaxial gating are plotted in dark green. C) VR-Eye analyses seeking \nmemory CD4 T cells based on reference populations measured with various platforms. The theory-generated label was \nwritten based on known protein expression patterns of memory CD4 T cells. The CyTOF-generated label was calculated \non memory CD4 T cells that were gated in a healthy donor from Dataset 3. The spectral-generated label was calculated \non memory CD4 T cells that were gated in Dataset 5. A rainbow intensity scale indicates similarity to the sought \npopulation with red representing high and purple representing low similarity. D) MEM label generated from the gated \nmemory CD4 population. E)  \nΔ MEM analyses comparing MEM labels from each VR-identified memory CD4 T cell \npopulations to the MEM label generated from the gated memory CD4 T cell population.  \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nCross et al. – Supplementary Figure S1 \n \nSupplementary Figure S1 – Velociraptor-Claw reproducibly identifies prognostic cells and classifies patients in \nDataset 1.  \nA) 5-fold cross validation design. B) VR-Claw analyses of CD45 -CD31- GBM cells across each round of 5-fold cross \nvalidation. Cell neighborhoods are color coded based on HR and p-value. C) Patient classification according to prognostic \npopulation abundance by RAPID and VR-Claw. \n \n  \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nCross et al. – Supplementary Figure S2 \n \nSupplementary Figure S2 – Optimization of prognostic GBM population identification. \nA) Expression levels of protein markers on common t-SNE axes for 131,880 CD45 -CD31- GBM cells in Dataset 1. A \nspectrum intensity scale indicates expression levels with blue representing low expression and red representing high \nexpression.  B) VR-Eye analysis seeking GNP-like cells (dark red in the middle plot) and GPP-like cells (dark blue in the \nmiddle plot). GNP-like cells, identified in the left plot, were sought with an input of SOX2 +7 S100B+4 GFAP+1 p-NFkB+0).  \nGPP-like cells, identified in the right plot, were sought with an input of EGFR +7 p-p38+2 CD49F+2 CD44+1 TUJ1+0) in the \nRAPID cohort (4,710 CD45 -CD31- GBM cells per patient). A spectrum intensity scale indicates cell similarity with red \nindicating high similarity and purple indicating low similarity. C) Correlation between each patient’s percentage of GNP-like \nor GPP-like cells, respectively, identified by RAPID and VR-Eye in Dataset 1. Pearson correlation tests were conducted. \nD) Comparison of cell identification approaches across multiple iterations of each approach. For VR-Eye and RAPID, \neach algorithm was applied to 10 different t-SNEs generated from the same sampling of Dataset 1 to identify GNP-like \nand GPP-like populations. Biaxial gating was performed by four different cytometry experts.  \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nCross et al. – Supplementary Figure S3 \n \nSupplementary Figure S3 – VR-Claw reproducibly reveals GATA2 haploinsufficiency-specific cells in Dataset 3. \nA) VR-Claw analyses across 10 runs using 2,000 randomly sampled T cells per donor in Set 1 (N=5, n=10,000 cells total). \nEach t-SNE was independently generated for each run using that run’s sample of T cells. Red and orange denote \npopulations enriched in GATA2 haploinsufficiency patients; blue and purple denote populations lacking from GATA2 \nhaploinsufficiency patients. B) RMSD analysis on MEM labels generated from GATA2 haploinsufficiency-specific \npopulations identified in each of the 10 runs of VR-Claw. Similarity values were calculated by subtracting normalized \nRMSD values from 100. A rainbow intensity scale indicates population similarity with red indicating high similarity and \npurple indicating low similarity. Stable populations that appear in at least 5 out of 10 runs are marked to the right of the \nheatmap with either a blue or a dark red line. Blue lines indicate populations that were lacking from GATA2 \nhaploinsufficiency patients, and dark red lines indicate populations that were enriched in GATA2 haploinsufficiency \npatients as determined by VR-Claw. Phenotypes of stable populations are summarized to the far right.  \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nCross et al. – Supplementary Figure S4 \n \nSupplementary Figure S4 – 41-dimensional phenotypic labels outperform filtered phenotypes in Dataset 4.  \nA) Expression levels of protein markers on common t-SNE axes for 257,076 cells in Dataset 4. A spectrum intensity scale \nindicates expression levels with blue representing low expression and red representing high expression. Markers not \nincluded as t-SNE input are labeled in blue. B) Comparison of VR-Eye cell identification using full 41-dimensional MEM \nlabels or filtered MEM labels (filtered labels range from 4-dimensional to 28-dimensional) based on population protein \nenrichment as search input. C)  Comparison of population homogeneity across different cell identification methods. The \nhomogeneity score equals the inverse of the median RMSD with respect to the median protein expression for each \nmarker for each population. \n \n  \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint \n\nCross et al. – Supplementary Figure S5 \n \nSupplementary Figure S5 – Velociraptor identifies CD4 +CD8+ double positive memory T cells present in human \nbreast cancer tumors.  \nA) Gating scheme to select memory CD4 T cells shown on a representative patient. B) VR-Eye analysis seeking memory \nCD4 T cells using a minimum marker label. C) Δ MEM analysis comparing the VR-identified memory CD4 T cell population \nto the gated population. Δ MEM labels were calculated by subtracting the absolute MEM label of the VR-identified \npopulation from the absolute MEM label of the manually gated population. D) Quantification of search accuracy for each \nmemory CD4 T cell label sought. Precision is plotted in purple, recall is plotted in gold, and F1-measures are plotted in \ngreen. Similarity thresholds were optimized to best capture the gated population. E) Biaxial plots of CD4 and CD8a \nexpression are shown for cells included in Velociraptor-identified memory CD4 (left plot) or CD8 (middle plot) populations. \nContour indicates density. A spectrum intensity scale indicates each cell’s similarity to the memory CD4\n+CD8+ double \npositive reference phenotype (right plot) as determined by maximum theory searches. F) Density plots show the \ndifference between each cell’s memory CD4 similarity and memory CD8 similarity to the maximum theory label. Biaxially \ngated memory CD4 T cells are shown in light blue, and biaxially gated memory CD8 T cells are shown in dark blue. \nUngated cells are shown in light grey. G) Biaxial plots of CD4 and CD8a expression are shown for cells with a memory \nCD4 T cell similarity value greater than or equal to 65 and a memory CD8 – CD4 similarity difference less than or equal to \n-5. Contour indicates density. \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted May 4, 2024. ; https://doi.org/10.1101/2024.05.01.591375doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}