Development of the NeuRGI model

To identify genes important in neutrophil differentiation, we developed a supervised learning algorithm NeuRGI (Neutrophil Regulatory Gene Identifier) (Fig 1A). We reasoned that a functional or driver gene should have all of the following four features: (1) differential or dynamic expressed during neutrophil differentiation. (2) associated with various physiological characteristics of neutrophils. (3) associated with related diseases. (4) conserved in multiple species.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Development of a machine learning method for predicting neutrophil-specific functional genes.

(A) NeuRGI model training workflow involved: 1) extracting gene features from various databases. 2) using genes of neutrophil-related genes as positives and PU-learning as negatives. 3) balancing the training set with under-sampling and training the NeuRGI random forest model with 10-fold cross-validation, then employing a Gaussian Mixture Model (GMM) with NeuRGI scores to identify potential positives. 4) using OntoVAE for in silico knockout of GMM-classified genes to find key regulatory factors for guiding follow-up experiments. (B) AUC and PR curves and the mean AUC value for the NeuRGI model trained from 10-fold cross-validation. The data is split into 10 parts, with 9 parts for training and 1 for testing, repeated 10 times with each part used once for testing. (C) The boxplot illustrates the transformed feature values for 293 known positive genes and 293 PU-learning negative genes in four feature groups. The p value was calculated using the Student’s t-test. (D) Distribution of the relative importance score calculated by average mean decrease Gini for features used in NeuRGI model (color-coded by category).

https://doi.org/10.1371/journal.pcbi.1012877.g001

We collected extensive datasets for feature extraction (S1 and S2 Tables). The expression dynamics were calculated independently using three RNA-sequencing datasets, including bulk RNA-seq data of 8 FACS-sorted continuous hematopoietic populations in mouse bone marrow (BM) [19] and human umbilical cord blood [20], and scRNA-seq data of sorted CD11b⁺Ly6G⁺ mouse neutrophils [21]. Homologous genes between mice and humans were aligned using the BioMart database [22]. We calculated features such as gene expression range, cell specificity index Tau, and gene expression change rate extracted from the pseudotime trajectory inferred from scRNA-seq, to measure the differential and dynamic gene expression fluctuations throughout neutrophil differentiation. For physiological characteristics, we categorized genes according to their associations with neutrophil physiological characteristics, such as neutrophil count and neutrophil measurement, based on Genome-Wide Association Studies (GWAS) information from the Open Targets Platform [23]. For pathological relatedness, we evaluated the Disease Specificity Index (DSI) from DisGeNET [24] and the number of neutrophil-related diseases (NofDisease) such as abnormal neutrophil count based on the MSigDB [25] databases. For conservation, we utilized PhastCons conservation scores from UCSC Genome Browser [26] which are derived from multiple alignment analyses of humans and 99 other vertebrate species.

The training model of NeuRGI is based on the random forest. To train the model, we encompassed a positive gene set (P set), containing 293 protein-coding genes from 19 Gene Ontology Biological Processes (GO-BP) pathways with the keyword ‘neutrophil’ or genes reported having neutrophil-related functions in the literature [8,27–30] (S3 Table). Next, we used the Positive-Unlabeled Learning (PU-learning) Spy algorithm [12] by adding 10% of the positive genes to an unlabeled set as ‘spies’. We then trained a Naïve Bayes classifier with the new set of genes and assigned probabilities to each gene in the unlabeled set. Genes with probabilities below the 10^th percentile were considered ‘negative’ (N set, see Methods). The under-sampling method [31] was used to obtain a balanced training data set, consisting of 293 positive and 293 negative genes (S3 Table). The GO-BP enrichment analysis of the ‘N’ set indicated that the genes were not related to neutrophil differentiation (S1A Fig).

We performed a random forest model fitting on this training set using 10-fold cross-validation. The importance of all features was evaluated based on their relative importance scores of the Gini coefficient (S1B Fig), and feature ablation studies were conducted. We found that the model, after removing 7 bottom-ranked features (Neutrophil to lymphocyte ratio, Sum of neutrophil and eosinophil counts, Neutrophil percentage of leukocytes, Neutrophil percentage of granulocytes, Abnormal neutrophil count, Hom Tau bulk RNA-seq, Cor scRNA-seq and Hom Spearman.p bulk RNA-seq), exhibited the best overall performance (AUC = 0.977, AUC-PR = 0.984, ACC = 0.962, MCC = 0.927, F1 Score = 0.964, S1C Fig and S4 Table). Therefore, this refined model was established as the NeuRGI model (Fig 1B). We further compared the normalized feature scores between the positive and negative sets regarding gene expression, physiological characteristics, pathological relatedness, and gene conservation. The relative scores in the positive genes were significantly higher than those in the negative genes in all four feature categories (Fig 1C). According to the relative importance score of the features, we found that the most predictive feature is pathological relatedness, constituting 68% of the total importance, followed by expression dynamics (26%) (Fig 1D).

Benchmarking NeuRGI for inferring functional genes

We applied NeuRGI to 19,288 protein-coding genes that were not used in model training, and assigned prediction scores for each gene, ranging from 0 to 1, with 1 as most likely to be ‘neutrophil regulator’ (S5 Table). We observed that the distribution of the NeuRGI score exhibited a bimodal distribution, with two peaks around 0.02 and 0.97 identified using R package quantmod (v0.4.26) [32], respectively (Fig 2A). We thus fitted the NeuRGI scores of all predicted genes with a three-component Gaussian Mixture Model (GMM) and categorized the genes into 3 classes: ‘functional’ in neutrophil differentiation, with NeuRGI scores higher than 0.96; ‘non-functional’, with scores lower than 0.02; and ‘uncertain’, with the scores between 0.02 to 0.96. (Fig 2A). In total, we identified 4,786 genes likely to be functional in neutrophil differentiation, and 4,734 genes unlikely to be involved in neutrophil differentiation (S1D Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Machine learning reveals MAP4K4 as a novel regulator of hematopoietic neutrophil differentiation.

(A) Three-component Gaussian Mixture Models (GMM) were applied to the NeuRGI scores across all 19,288 predicted genes. The solid black line represents the overall score distribution, dashed lines indicate the three probability density functions (PDFs). Intersections of the three PDFs define NeuRGI thresholds for classifying genes into function (red shadow), non-function (blue shadow), and uncertain (middle white) categories. (B) The line plot depicts the overlap of functional transcription factors (TFs) predicted by CellOracle and NeuRGI (red line), and by SCENIC and NeuRGI (blue line). The x-axis represents the top N TFs ranked by CellOracle’s perturbation score or SCENIC regulon activity. The y-axis shows the percentage of overlap between the predicted functional TFs. (C) The boxplot illustrates the transformed feature values for 4,786 functional genes and 4,734 non-functional genes in four feature groups. The p value was calculated using the Student’s t-test. (D) The GO-BP term network illustrates the five main clusters enriched from 4,786 predictive functional genes (see Methods). Each dot represents a GO term and the dot size indicates the enrichment score. The dashed oval indicates GO terms with similar functions. (E) Proportion of regulatory (red) and other (blue) genes across three gene sets. Regulatory genes include enzyme (Enz), membrane protein (MP), RNA binding protein (RBP), and transcription factor (TF). The p value was calculated by the proportion test. Numbers within bars represent the gene counts. (F) Boxplot of NeuRGI scores, color-coded by gene type: Enzyme (Enz), membrane proteins (MP), RNA binding proteins (RBP), Transcription factors (TF), and others. The p value was calculated using the Student’s t-test. (G) The scatter plot shows the impact of in silico knockout of 2,569 predictive functional regulatory genes on the “positive regulation of myeloid cell differentiation” pathway and MAGMA Zscore (neutrophil count). The y-axis represents the -log₁₀ (p value) on the pathway after gene in silico knockout (higher values indicate greater impact), and the x-axis represents the gene’s effect on the ‘neutrophil count’ trait (higher Zscores indicate greater impact). Different colors represent different categories of genes. Dot size indicates the NeuRGI score, and contour lines show point density. A cutoff (y = 1.8 and x = 3.6) was set based on the contour lines, dividing the scatter plot into four regions. (H) Expression of top 12 genes in different immune cells from ImmuNexUT, including Naïve CD4 T cells (Naïve CD4), Memory CD4 T cells (Mem CD4), T helper 1 cells (Th1), T helper 2 cells (Th2), T helper 17 cells (Th17), T follicular helper cells (Tfh), Fraction II effector regulatory T cells (Fr. II eTreg), Fraction I naïve regulatory T cells (Fr. I nTreg), Fraction III non-regulatory T cells (Fr. III T), Naïve CD8 T cells (Naïve CD8), CD8+ T effector memory CD45RA+ cells (TEMRA CD8), Effector Memory CD8 T cells (EM CD8), Central Memory CD8 T cells (CM CD8), Naïve B cells (Naïve B), Unswitched memory B cells (USM B), Switched memory B cells (SM B), Double Negative B cells (DN B), Plasmablasts (Plasmablast), Natural Killer cells (NK), CD16 positive monocytes (CD16p Mono), Non-classical monocytes (NC Mono), Intermediate monocytes (Int Mono), Classical monocytes (CL Mono), Myeloid dendritic cells (mDC), Plasmacytoid dendritic cells (pDC), neutrophils (Neu), Low-Density Granulocytes (LDG). (I) Expression of 9 genes in neutrophil differentiation of human UCB. These 9 of the top 12 genes dynamically upregulated, including CREBBP, CYP27A1, JAK2, IFNGR1, MAP4K4, PLCG2, PTPRC, TIGAR, and TYK2. We set ‘time cut’ for cells at different differentiation stages, with HSC set as 1 and Neu as 5, and performed linear regression fitting for the expression of all 9 genes. R represents the Pearson correlation coefficient (R), and the p value (p) was calculated by the Student’s t-test. (J) Expression of top 10 genes in single-cell pseudotime analysis of 2,803 neutrophils in mouse bone marrow. All 12 genes except Cyp27a1 and HLA-DQA1 upregulated during neutrophil maturation. (K) The heatmap displays the log₂ (fold change) in the expression of the top 12 genes in neutrophils from patients with 10 immune-related diseases compared to those from healthy individuals, including ANCA-associated vasculitis (AAV), Takayasu arteritis (TAK), Adult-onset Still’s disease (AOSD), Behçet’s disease (BD), Rheumatoid arthritis (RA), Systemic sclerosis (SSc), Idiopathic inflammatory myopathy (Myo), Sjögren’s syndrome (SjS), Mixed connective tissue disease (MCTD), Systemic lupus erythematosus (SLE). The p value was calculated using the Student’s t-test, and * represents statistical significance (p < 0.05). The histogram represents the number of significant diseases for each gene. (L) Bar plot displaying significantly affected pathways after OntoVAE in silico knockout of MAP4K4 in neutrophils. The blue dashed line represents the significance threshold (p = 0.05).

https://doi.org/10.1371/journal.pcbi.1012877.g002

We next compared NeuRGI with CellOracle [9] and SCENIC [10], two state-of-the-art tools for inferring gene regulatory networks (GRN), based on perturbation scores and regulon activity, respectively. Using scRNA-seq data of 2,803 BM neutrophils [21], CellOracle and SCENIC identified 70 and 36 TFs, respectively (S6 Table). The majority of these TFs (68.6% for CellOracle and 75.0% for SCENIC) were classified as functional by NeuRGI (S1E Fig). Additionally, all of the top 20 TFs ranked by CellOracle’s perturbation score and SCENIC’s regulon activity overlapped with NeuRGI’s predicted functional set (Fig 2B), demonstrating accuracy of NeuRGI. Notably, 22 and 9 TFs that were predicted to be important by CellOracle or SCENIC were predicted as ‘uncertain’ by NeuRGI. We further explore the biological distinctions between the ‘uncertain’ and the ‘functional’ TFs called by NeuRGI and one of the other methods. When comparing four feature values across the two groups, we observed that consistently functional TFs showed significantly higher pathological relatedness than uncertain TFs (p < 0.012, S1F Fig), further showing the accuracy of NeuRGI in predicting functional TFs.

We evaluated the predicted functional and non-functional sets regarding the four feature categories scores (gene expression dynamics, physiological characteristics, pathological relatedness and gene conservation), experimental validation, GO-BP pathway, and the enrichment of regulatory genes (defined as TFs, RBPs, enzymes, or membrane proteins).

First, the feature values of the predicted ‘functional’ set were significantly higher than those of the predicted ‘non-functional’ set (Fig 2C), consistent with the observation in the model training. Second, 5 genes with NeuRGI scores higher than 0.995 were reported to play important roles in neutrophil differentiation or function. The 5 genes include 3 membrane proteins FAS, IL17RA, and CD226, 1 TF STAT6, and 1 enzyme GRK6. FAS and STAT6 regulate neutrophil apoptosis [33,34]; IL-17RA plays a crucial role in neutrophil recruitment induced by IL-17 [35]. CD226 (DNAM-1) modulates neutrophil infiltration and inflammation [36]; GRK6 plays an important negative regulatory role in neutrophil-mediated inflammatory responses [37]. Finally, we conducted GO-BP analysis for all 4,786 functional genes and categorized the top 100 GO terms based on their similarity. The terms were clustered into 6 categories (S1G Fig), among which ‘Neutrophil activation involved in immune’, ‘leukocyte proliferation’, ‘leukocyte cell-cell adhesion’, ‘regulation of apoptotic signaling pathway’, ‘Peptidyl-tyrosine phosphorylation’, and ‘homeostasis’ were highly connected (Fig 2D). Notably, among the 4,786 genes identified as likely functional in neutrophil differentiation, 53.7% (2,569) genes were predicted to encode regulatory proteins, including enzymes (1,234), membrane proteins (726), transcription factors (TFs, 397), and RNA-binding proteins (RBPs, 212). This proportion is significantly higher than that in the negative gene set (28.1%, Fig 2E, proportion test, p = 4.1 x 10⁻¹⁴²), suggesting the predicted functional set was enriched with regulatory genes. Interestingly, the NeuRGI score of enzymes was significantly higher than the other protein types (Fig 2F).

Taken together, NeuRGI demonstrated high accuracy in predicting functional regulators particularly those with high pathological relevance. Unlike CellOracle and SCENIC, NeuRGI extends its predictions beyond TFs, capable of identifying a broader range of functional genes. And NeuRGI-predicted functional genes are enriched for regulatory roles, exhibit higher feature scores, and are closely linked to key immune processes. These findings highlight NeuRGI as a comprehensive and robust tool for functional gene inference.

Machine-learning strategy identified MAP4K4 as a novel regulator of hematopoietic neutrophil differentiation

We performed in silico knockout in neutrophils for the 2,569 genes using Ontology-guided Variational Autoencoder (OntoVAE) [13], a deep learning predictive model trained on the 2,040 RNA-seq datasets of immune cells from ImmuNexUT [38] and Atlas Human Lymphocytes [39], with stable loss after 300 epochs (S1H Fig, see S1 Text). OntoVAE is a modified VAE architecture that can incorporate Gene Ontology and provide pathway activities in its latent space and decoder. We retrieved the activities of all pathways and found the ‘neutrophil activation’ term was especially associated with neutrophils, and ‘regulation of mononuclear cell proliferation’ was associated with monocytes, as expected (S1I Fig). In silico knockout was performed by setting the input of a gene as 0, then using paired Wilcoxon tests to identify the affected terms (pre-knockout versus post-knockout). To test the validity, we performed in silico knockouts of the ELANE [40] and SYK [41], two genes have been reported to play important roles in neutrophils and B cells, respectively. We found that pathways related to neutrophils and B cells were significantly affected in consistent with previous studies [42–45] (S2A Fig), suggesting that applying OntoVAE on RNA-seq data of immune cells can predict the pathways alterated upon gene knockouts in neutrophils.

Next, we assessed the impact of the genes on the ‘positive regulation of myeloid cell differentiation’ by OntoVAE in silico knockout and calculated gene-level association Z score of ‘neutrophil count’ based on GWAS using MAGMA [46] (Fig 2G and S5 Table). We then identified 11 genes with top scores of both features and 1 gene with the highest scores of each feature (Fig 2G). All 12 genes have the highest expression in neutrophils in healthy individuals (Fig 2H), among which 9 genes showed increased expression along the neutrophil differentiation (from HSC to neutrophils) in humans (Fig 2I). In addition, the pseudotime trajectory analysis of scRNA-seq data using monocle2 [47] indicated that expression of the 10 genes increased from immature neutrophils (imNeu) to mature neutrophils (mNeu) in mice [21] (Fig 2J). The 12 genes were significantly differential expressed between healthy individuals and patients with immune-mediated diseases (Fig 2K). For example, CLIC1 and JAK2 were significantly differential expressed in 8 types of immune cell-related diseases, followed by PTPRC (7), CREBBP (5), HLA-DQA1(5), and MAP4K4 (5).

The 12 genes include 7 enzymes (PLCG2, JAK2, CYP27A1, TIGAR, MAP4K4, TYK2, FIG4), 4 membrane proteins (PTPRC, HLA-DQA1, IFNGR1 and CLIC1), 1 RNA binding protein (CREBBP). Nine genes have been reported to play functional roles in the neutrophils. Specifically, JAK2 influences neutrophil production by regulating progenitor cell proliferation [48]. PTPRC (CD45) regulates neutrophil activation and migration post-myocardial infarction [49]. CLIC1 affects oxidative responses in neutrophils [50]. IFNGR1 enhances immune resistance to bacterial infections [51]. Additionally, TYK2 modulates neutrophil function in acute inflammation [52]. The CREBBP is important for immune responses via oxidative burst regulation [53]. PLCG2 affects neutrophil chemotaxis and inflammation in arthritis [54], CYP27A1 may serve as a biomarker for intervertebral disc degeneration [55], and HLA-DQA1 influences neutrophil-mediated immune processes [56]. The other 3 genes, MAP4K4, FIG4, and TIGAR have not been reported to be involved in neutrophil differentiation or functions, among which MAP4K4 has the highest NeuRGI score (0.9887, S2B Fig). MAP4K4 has the highest expression in neutrophils and highest neutrophil specificity (Tau = 0.75, S2C Fig) and the highest impact on the ‘positive regulation of myeloid leukocyte differentiation’ pathway in knockout simulation (p = 1.79x10⁻⁴, Figs 2L and S2D). The expression is increased along the neutrophil differentiation (S2E Fig, R = 0.73, Pearson correlation) (S7 Table).

MAP4K4 is required for neutrophil differentiation

To investigate the role of MAP4K4 in neutrophil differentiation, HL-60 human leukemia cells were used in the in vitro studies. HL-60 can be differentiated into neutrophil-like cells characterized by CD11b expression upon stimulation with all-trans-retinoic acid (ATRA) [57] (Fig 3A). We found that MAP4K4 specific inhibitor, PF-06260933, significantly impeded the differentiation of HL-60 cells into CD11b⁺ cells in a dose-dependent fashion (Fig 3B). The percentage of CD11b⁺ cells was halved at the concentration of 5 µM and fell to ¼ of the control at 20 µM (Fig 3B), suggesting that MAP4K4 may be involved in neutrophil differentiation.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Loss of MAP4K4 expression causes a decrease in the number of neutrophils.

(A) Scheme of all-trans-retinoic acid (ATRA)-induced neutrophil differentiation of the hematopoietic HL-60 cell models (left); flow cytometry analysis of the neutrophil marker CD11b has validated that ATAR can induce differentiation of HL-60 cells into neutrophils. This is demonstrated by histogram plots showing variations in mean fluorescence intensity (MFI) across different conditions: cells untreated with ATAR, cells treated with ATAR for 72 hours, and cells treated with ATAR for 96 hours (right). (B) Percentage of CD11b⁺ cells in HL-60 cells were treated with 0-20 µM MAP4K4 inhibitor (PF-06260933, PF) for 48 hours, followed by a subsequent 96-hour treatment with both PF and all-trans-retinoic acid (ATRA). (C) The number of granulocyte-macrophage progenitor cells (CFU-GM, CFU-G, CFU-M) colonies formed by 25,000 whole bone marrow cells from control (Ctrl) or Map4k4-cKO (cKO) mice; BM, bone marrow (n=4; mean ± SD). (D) Numbers of neutrophils in the bone marrow of control (Ctrl) or Map4k4-cKO (cKO) mice; BM, bone marrow; Neu, neutrophil. (Ctrl n=7, cKO n=6; mean ± SD). (E) PB WBC numbers of control (Ctrl) or Map4k4-cKO (cKO) mice (Ctrl n=7, cKO n=6; mean ± SD). PB, peripheral blood; WBC, white blood cell; (Ctrl n=7, cKO n=6; mean ± SD). (F) PB Mon, and GR numbers of control (Ctrl) or Map4k4-cKO (cKO) mice; PB, peripheral blood; Mon, monocyte; GR, granulocyte; (Ctrl n=7, cKO n=6; mean ± SD). (G) Numbers of eosinophils in PB of control (Ctrl) or Map4k4-cKO (cKO) mice; PB, peripheral blood; Eso, eosinophils; (Ctrl n=7, cKO n=6; mean ± SD). (H) Numbers of monocytes in PB of control (Ctrl) or Map4k4-cKO (cKO) mice; PB, peripheral blood; Mon, monocytes; (Ctrl n=7, cKO n=6; mean ± SD). (I) Numbers of neutrophils in PB of control (Ctrl) or Map4k4-cKO (cKO) mice; PB, peripheral blood; Neu, neutrophil; (Ctrl n=7, cKO n=6; mean ± SD). (J) Spleen cell numbers control (Ctrl) or Map4k4-cKO (cKO) mice (Ctrl n=7, cKO n=6; mean ± SD). (K) Numbers of eosinophils in the spleen of control (Ctrl) or Map4k4-cKO (cKO) mice; SP, spleen; Eso, eosinophils; (Ctrl n=7, cKO n=6; mean ± SD). (L) Numbers of monocytes in the spleen of control (Ctrl) or Map4k4-cKO (cKO) mice; SP, spleen; Mon, monocytes; (Ctrl n=7, cKO n=6; mean ± SD). (M) Representative FACS analysis of spleen CD11b+ Ly6G+ neutrophils (left); percentage of neutrophils in the spleen of control (Ctrl) or Map4k4-cKO (cKO) mice (right); SP, spleen; Neu, neutrophil; (Ctrl n=7, cKO n=6; mean ± SD). (N) Numbers of neutrophils in the spleen of control (Ctrl) or Map4k4-cKO (cKO) mice; SP, spleen; Neu, neutrophil; (Ctrl n=7, cKO n=6; mean ± SD). Mann-Whitney U test.

https://doi.org/10.1371/journal.pcbi.1012877.g003

To confirm MAP4K4 plays a role in neutrophil differentiation in vivo, we generated transgenic mice whose hematopoietic stem cells and progenitors (HSPCs) lacking the expression of MAP4K4 (Map4k4-cKO, cKO) by crossing Map4k4^fl/fl mice with interferon-inducible Mx1-Cre mice (S3A Fig). RT-PCR analysis confirmed the loss of Map4k4 transcript in HSPCs after poly:IC treatment (S3B Fig). The total cell number in bone marrow was not changed in the cKO mice (S3C Fig). Furthermore, the percentages and the cell numbers of hematopoietic stem cells (HSCs) and progenitors were also not affected compared to controls (Ctrl) (Map4k4^−/− MX1-Cre⁺) (S3D and S3E Fig). However, the colony formation analysis showed significantly reduced colony numbers derived from granulocyte-macrophage progenitor cells (CFU-GM, CFU-G, CFU-M) (Fig 3C), while primitive erythroid progenitor cells (BFU-E) remained unaffected (S3F Fig). In the investigation conducted using flow cytometry analysis, it was observed that there was a 15% reduction in the neutrophil count in cKO mice (Fig 3D), while the levels of eosinophils and monocytes were not affected (S3G and S3H Fig), indicating that the loss of MAP4K4 in HSPCs inhibits neutrophil differentiation.

cKO mice displayed a reduction of white blood cell count in the peripheral blood (Fig 3E), specifically monocyte and granulocyte counts were reduced by 68% and 50%, respectively (Fig 3F). In line with the in-silico KO prediction, lymphocyte, RBC, and platelet counts were not affected (S3I–S3K Fig). Using flow cytometry analysis, we confirmed that the number of neutrophils in peripheral blood was decreased by about 45% in cKO mice eosinophils were reduced by 50%, and monocytes were reduced by 45% (Fig 3G–3I). Similarly, the total splenocyte number also decreased in cKO mice (Fig 3J). Specifically, the numbers of neutrophils were reduced by 40%, and both eosinophils and monocytes were all reduced by 70% in the spleen (Fig 3K–3N). In contrast, the numbers of dendritic cells (DCs) and red pulp macrophages remained unaffected (S3l–S3M Fig).

Collectively, our machine-learning model identified a protein kinase, MAP4K4, which is required for neutrophil differentiation. Subsequent experiments confirmed that the deficiency of MAP4K4 in HSPCs results in a significant reduction in the populations of neutrophils and monocytes.

MAP4K4 regulates neutrophil differentiation in bone marrow

To further delineate the role of MAP4K4 in neutrophil differentiation, we performed scRNA-seq analysis on BM cells from cKO mice. Given the low abundance of HSPCs, we performed a sample enrichment for c-Kit⁺ HSPCs and mixed them with total BM cells at a 2:3 ratio for scRNA-seq (Fig 4A). Ctrl and cKO datasets were merged and anchor genes [58] were used to remove the batch effect (S4A Fig). After quality control, 9,051 cells from wildtype and 5,674 cells from cKO mice were included for further analysis. After removing mitochondrial and ribosomal genes, a total of 54,838 genes were detected with 11,261 reads (S4B Fig), an average of 2,865 genes per cell (S4C Fig), highlighting the high quality of our dataset (S8 Table). We next used Seurat [58] to process the 14,725 cells for normalization, batch effect correction, and cell-type clustering. The cells were visualized by Uniform Manifold Approximation and Projection (UMAP) and clustered into 16 distinct cell types based on the expression of canonical cell markers [7] (S4D Fig and S9 Table), including HSC and multipotent progenitors (HSC/MPP), granulocyte monocyte progenitors (GMP), neutrophil progenitors (ProNeu), neutrophil precursors (PreNeu), immature neutrophils (imNeu), mature neutrophils (mNeu), megakaryocytic-erythroid progenitors (MEP), erythrocyte progenitors (ProEry), monocyte progenitors (ProMon), B progenitors (ProB), B cells, natural killer cells (NK), natural killer T cells (NKT), eosinophils, macrophages and conventional dendritic cells (cDC) (Fig 4B and S4E Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Loss of MAP4K4 expression impairs the differentiation process of bone marrow neutrophils.

(A) Overview of study design, created using Biorender. (B) UMAP Plots for single-cell gene expression pooled across bone marrow samples, clusters visualized and labeled by cell type. The plot is split by control (Ctrl) vs Map4k4-cKO (cKO). (C) Proportions of the six neutrophil clusters in control (Ctrl) and Map4k4-cKO (cKO) samples. The p value was calculated by the adjusted proportion test. (D) Numbers of neutrophil progenitor cells in the bone marrow of control (Ctrl) or Map4k4-cKO (cKO) mice; c-Kit^hiLy6G^neg (myeloblasts, MB); c-Kit^intLy6G^neg (promyelocytes, PM); c-Kit^negLy6G^low (myelocytes, MC); c-Kit^negLy6G^int (metamyelocytes, MM); and c-Kit^negLy6G^hi (band cells and segmented neutrophils, BC/SC); BM, bone marrow (Ctrl n=7, cKO n=6; mean ± SD). Mann-Whitney U test. (E) ROS generation of neutrophil progenitor cells in BM of control (Ctrl) or Map4k4-cKO (cKO) mice; MFI, mean fluorescent intensity; ROS, reactive oxygen species; MB, myeloblasts; PM, promyelocytes; MC, myelocytes; MM, metamyelocytes; BC/SC, band cells and segmented neutrophils; (n=6; mean ± SD). Mann-Whitney U test. (F) The number of DEGs in control (Ctrl) vs Map4k4-cKO (cKO) of six neutrophil subpopulations. (G) GO-BP analysis of cluster-based DEGs between control (Ctrl) and Map4k4-cKO (cKO) HSC to neutrophils. Selected GO terms with Benjamini-Hochberg-corrected p values < 0.05 (one-sided Fisher’s exact test) are shown. The dot size represented the number of genes. The color scale represented the adjusted p value. Fig 4A was created using Biorender.

https://doi.org/10.1371/journal.pcbi.1012877.g004

To model the neutrophil differentiation process, we applied Monocle2 [47] to 11,031 cells including HSC/MPP, GMP, and neutrophils of different maturation stages, and identified a continuously increasing pseudotime from HSC/MPP to mNeu, consistent with the neutrophil differentiation process [7] (S4F–G Fig). During this neutrophil differentiation, we found that the fractions of proNeu and preNeu in the cell division stage of neutrophil differentiation were increased in cKO mice. However, imNeu was decreased (Fig 4C). Aiming to validate these results from scRNA-seq, we analyzed the maturation stages of neutrophils in BM by flow cytometry. In line with the scRNA-seq data, we found a significant increase in promyelocytes (proNeu) [7], myelocytes (mixture of GMP and proNeu) [7], and metamyelocytes (preNeu) [7] in cKO mice BM (Fig 4D) and a reduction of band cells and segmented neutrophils (imNeu/mNeu) [7] in cKO mice BM (Fig 4D). To investigate the whether MAP4K4 affects the function of neutrophils, we assessed the reactive oxygen species (ROS) production in neutrophils and the precursors in BM. No significant difference in ROS levels were found among all of these populations in the bone marrow of MAP4K4-deficient mice compared to the control mice (Fig 4E). Similarly, neutrophils isolated from the spleen and peripheral blood of MAP4K4-deficient mice exhibited no significant alterations in ROS production (S4H Fig). Furthermore, the phagocytosis of E. coli by neutrophils from both the spleen and peripheral blood also showed no significant differences in the MAP4K4-deficient mice (S4I Fig). These results suggest that the deletion of MAP4K4 primarily affects the differentiation of neutrophils, with no significant impact on their function, such as ROS production and phagocytic activity.

MAP4K4 regulates cell apoptosis during neutrophil differentiation process

To investigate the downstream regulatory network of MAP4K4, we performed differential expression gene (DEG) analysis for neutrophil subpopulations. ProNeu and GMP were the most affected populations with the largest number of DEGs (Fig 4F and S10 Table). Next, we compared the differences in pathways in Ctrl and cKO mice. Notably, the cKO up-regulated DEGs in HSC/MPP, PreNeu, and imNeu were specifically enriched with ‘regulation of intrinsic apoptotic signaling pathway’, while those up-regulated genes in GMP, ProNeu, imNeu, and mNeu of Ctrl mice were uniquely enriched with ‘myeloid cell differentiation’ (Fig 4G), suggesting MAP4K4 may potentially affect apoptotic signaling pathway.

Next, we further performed consensus co-expression network analysis using hdWGCNA [59] in scRNA-seq. Five co-expression modules (M) were identified (Figs 5A and S5A–S5E). Notably, the hub genes of M1 are mainly expressed in HSC/MPP, while M2 for GMP, ProNeu, and PreNeu, M3 in PreNeu, M4 in imNeu, and M5 in mNeu. For instance, classical marker genes of mNeu including Retnlg and Mmp8 were highly expressed in M5 (Fig 5B). These results showed that these modules corresponded to the hieratical differentiation neutrophil maturation in hematopoiesis, with M1 representing the start, and M5 representing the end of pseudotime in scRNA-seq (Fig 5C and 5D).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Loss of MAP4K4 expression affects apoptosis during the process of neutrophil differentiation.

(A) UMAP plot of the neutrophil co-expression network. Each node represents a single gene, and edges represent co-expression links between genes and module hub genes. Point size is scaled by eigengene-based connectivity (kME). Nodes are colored by co-expression module assignment. The top five hub genes per module are labeled. Network edges were down-sampled for visual clarity. (B) Heatmap of scaled gene expression for the top 10 hub genes by kME in each module. (C) UMAP colored by Pseudotime. (D) UMAP colored by harmonized module eigengenes (hMEs) of each module. (E) Module eigengenes (MEs) as a function of pseudotime for each co-expression module. For each module, a separate loess regression line is shown for each condition. The p value was calculated by distance correlation t-test (dcorT.test). dcor is a transformation of a bias-corrected version of distance correlation. The smaller the dcor, the greater the distance between control (Ctrl) and Map4k4-cKO (cKO) conditions. (F) Differentially expressed hub genes between control (Ctrl) and Map4k4-cKO (cKO) parasites involved in positive regulation of apoptotic process determined by Gene Set Enrichment Analysis (GSEA). RES represents the running enrichment score. RLM represents the ranked list metric. (G) Violin plots of apoptosis scores of control (Ctrl) against Map4k4-cKO (cKO) for each cluster, the p value was calculated by the Student’s t-test. (H) Percentages of annexin V⁺ and SYTOX⁻ cells in bone marrow neutrophil progenitor cells of control (Ctrl) or Map4k4-cKO (cKO) mice; MM, metamyelocytes; BC/SC, band cells and segmented neutrophils; BM, bone marrow (n=10; mean ± SD). Mann-Whitney U test. (I) Percentages of annexin V⁺ and SYTOX⁻ cells of HL-60 cells treated with 0-20 µM MAP4K4 inhibitor (PF-06260933, PF) for 48 hours, followed by a subsequent 96-hour treatment with both PF and all-trans-retinoic acid (ATRA) (n=4; mean ± SD). Mann-Whitney U test. (J) Quantification of HL-60 cell numbers after a 48-hour treatment with a 10 µM concentration of the MAP4K4 inhibitor (PF-06260933, PF), followed by a subsequent 96-hour treatment with both PF and all-trans-retinoic acid (ATRA) (n=4; mean ± SD). Mann-Whitney U test.

https://doi.org/10.1371/journal.pcbi.1012877.g005

To identify the modules affected by MAP4K4 during neutrophil differentiation, we analyzed the pseudotime trajectory of Ctrl and cKO data. The differential module eigengene (DME) analysis revealed significant differences between Ctrl and cKO in all modules (Fig 5E), and M3 and M4 were the most affected by MAP4K4-deficient with the lowest distance correlation score (dor) (Fig 5E) followed by M2. Notably, genes in M2, M3, and M4 were all enriched in pathways related to apoptosis (S5F Fig and S11 Table). Furthermore, GSEA analysis revealed that ‘Positive regulation of apoptotic process’ pathway was significantly enriched in Ctrl for ProNeu (p.adjust = 0.049), cKO for PreNeu (p.adjust = 0.047), and cKO for Immature-Neu (p.adjust = 0.043), indicating that the process of apoptosis was inhibited in ProNeu of cKO, while promoted in PreNeu and imNeu of cKO (Fig 5F, Methods). We measured the ‘apoptosis score’ based on the averaged gene expression level in the ‘positive regulation of apoptosis process’ pathway (Fig 5G and S12 Table). We found that the apoptosis score for ProNeu cells decreased the most in cKO mice, in contrast, the apoptosis score increased in cKO-imNeu and mNeu (Fig 5G). Apoptosis-related genes such as Rpl11 and Prelid1 upregulated in all cKO cell types from HSC/MPP to PreNeu except for ProNeu, while Myc, Acin1, and Pdcd4 were only downregulated in cKO-ProNeu (S5G Fig). We validated these results by assessing neutrophil cell death rates at different differentiation stages using a combination of annexin V and SYTOX staining. This analysis indicated reduced apoptosis rates among cKO-promyelocytes (proNeu) [7], cKO-myelocytes (mixture of GMP and proNeu) [7], and cKO-metamyelocytes (preNeu) [7]. In contrast, apoptosis rates increased by 25% in cKO-band cells and segmented neutrophils (imNeu/mNeu) [7] (Figs 5H and S5H). These findings underscore the differential apoptotic responses across various stages of neutrophil maturation in the cKO model.

To further assess the role of MAP4K4 in regulating neutrophil apoptosis, we conducted in vitro studies using HL-60 cells. Notably, the treatment of MAP4K4 inhibitors (PF-06260933) in HL-60 cells did not increase cell apoptosis unless at a high (20 µM) concentration (S5I and S5J Fig). However, when the cells were treated with neutrophil-like state inducer (ATRA), MAP4K4 inhibitors resulted in cell apoptosis even at low (5 µM) concentration. And this apoptotic-inducing effect was dose-dependent. To confirm this phenotype, we treated the cells continuously with the inhibitor at a moderate (10 µM) concentration, 24 hours after treatment of ATRA, the cell numbers reduced by 50% compared to the control group (Fig 5I and 5J). These results demonstrate the significant impact of MAP4K4 on cell viability during the neutrophil differentiation process.

MAP4K4 modulates the phosphorylation level of apoptosis-related genes

As a serine/threonine protein kinase, MAP4K4 plays a pivotal role in modulating protein phosphorylation [60], our in silico knockout also predicted it affects phosphorylation (Fig 2L). To investigate how MAP4K4 influences downstream protein phosphorylation levels, we performed mass spectrometric analysis on protein expression and phosphorylation level in a cell line lacking expression of MAP4K4 (Fig 6A). Most peptides ranged between 7 and 20 amino acids (S6A and S6B Fig) and the principal components analysis (PCA) results showed samples were clustered by sample types, indicating good sample quality (S6C and S6D Fig). Compared to the controls, the phosphorylation level of 1911 proteins, and the expression of 1606 proteins were significantly changed in MAP4K4-KO cells (Fig 6B and S13 and S14 Tables). Notably, MAP4K4-KO has more impact on protein phosphorylation level than the expression level (Fig 6B and 6C).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Knocking out MAP4K4 reduces phosphorylation levels in the apoptosis pathway, and overexpression STAT5A rescued MAP4K4 deficiency-caused neutrophil apoptosis.

(A) Schematic of mass spectrometry analysis, created using Biorender. (B) The number of proteins with significant differences in protein and phosphorylation levels after MAP4K4 knockout. (C) Comparison heatmap of log ₂ (fold-change) in protein and phosphorylation levels after MAP4K4 knockout. (D) Differential phosphorylation level between control and MAP4K4 KO parasites is involved in the positive regulation of the apoptotic process, as determined by Gene Set Enrichment Analysis (GSEA). RES represents the running enrichment score. RLM represents the ranked list metric. (E) Volcano plot of the apoptosis-related proteins with differentially expressed phosphorylation levels. The x-axis shows the phosphorylation difference of log₂ (fold-change) between MAP4K4 KO and Ctrl. The y-axis shows –log₁₀ (adjusted p value) of the t-test between these two groups. Red indicates proteins with up-regulated phosphorylation after MAP4K4 KO, and blue, proteins with down-regulated phosphorylation after MAP4K4 KO. The top 20 proteins with the most significant differences in phosphorylation among the apoptosis-related proteins are visualized. (F) Western blot (WB) for MAP4K4, p-STAT5A, STAT5A, p-STAT5B, STAT5B, and GAPDH in sgAASV1 and sgMAP4K4 cells. (G) Representative FACS analysis of HL-60 cells were first infected with overexpression of STAT5A or STAT5B or control plasmids and then treated with 10 µM MAP4K4 inhibitor (PF-06260933, PF) for 48 hours, followed by a subsequent 96-hour treatment with both PF and all-trans-retinoic acid (ATRA) for staining of anti–annexin V and SYTOX was shown (left). Percentages of annexin V⁺ and SYTOX⁻ cells (right) (n=4; mean ± SD). Mann-Whitney U test. Fig 6A was created with Biorender.

https://doi.org/10.1371/journal.pcbi.1012877.g006

Then, we performed GO-BP analysis of differential phosphorylation level proteins between Ctrl and MAP4K4-KO samples. Notably, proteins with downregulated phosphorylation levels were specifically enriched with ‘Positive regulation of apoptotic process’ and ‘Positive regulation of myeloid leukocyte cytokine production involved in immune response’, while those up-regulated proteins were uniquely enriched with ‘Mitochondrial ATP synthesis coupled electron transport’ and ‘Nucleotide phosphorylation’ (S6E Fig). GSEA analysis showed that ‘Positive regulation of apoptotic process’ was significantly enriched for proteins with downregulated phosphorylation level (p.adjust = 0.015, Normalized Enrichment Score (NES) = -1.53) (Fig 6D). These results suggest that MAP4K4 may play a critical role in modulating phosphorylation levels of genes involved in apoptotic pathways, thereby influencing cell apoptosis during neutrophil differentiation.

Among the apoptosis-related proteins (S12 Table) [61] with downregulated phosphorylation levels, STAT5A and STAT5B exhibited the most downregulation level in phosphorylation with fold changes of 15 and 9, respectively (Fig 6E). We confirmed that both the protein expression and phosphorylation levels of STAT5A were significantly decreased in MAP4K4-KO k562 cells using Western blot analysis (Fig 6F), while the protein levels of STAT5B remained largely unaffected. Due to the absence of a specific antibody for phosphorylated STAT5B, its phosphorylation status could not be evaluated. To further delineate the distinct roles of STAT5A and STAT5B in the MAP4K4-mediated apoptosis during neutrophil differentiation, we overexpressed STAT5A, STAT5B, or control 3xFLAG proteins in HL-60 cells treated with MAP4K4 inhibitor (S6F Fig) and analyzed cell apoptosis level during ATRA-induced neutrophil differentiation. We found that overexpression of STAT5A but not STAT5B rescued the cell apoptosis induced by MAP4K4 inhibitor (Fig 6G). These findings suggest by regulating the phosphorylation level of apoptosis-related genes, such as STAT5A, MAP4K4 regulates apoptosis during the process of neutrophil differentiation.

Ethics statement

All animal experiments followed the protocols approved by the Institutional Animal Care and Use Committee of West China Second University Hospital [(2018) Animal Ethics Approval No. 004].

Overview of neutrophil regulatory gene identifier (NeuRGI)

To build a machine learning-based Neutrophil Regulatory Gene Identifier (NeuRGI) model, we used a workflow consisting of several steps: (1) Data collection. (2) Feature extraction. (3) Defining the positive and negative sets using PU-learning [75]. (4) 10-fold cross-validation [76]. (5) Model training. (6) Scoring and ranking. (7) Gaussian Mixture Model (GMM) [77] classified NeuRGI score. We implemented and tested our workflow in R (version 4.0), our code is open source and available on GitHub (https://github.com/LuChenLab/NeuRGI).

The framework of NeuRGI consists of 7 key steps:

Step 2: Feature extraction.

We extracted 28 features of genes covering 4 aspects, including expression dynamics, physiological characteristics, pathological relatedness, and conservation score (S2 Table).

1. 1. Expression dynamics

We extracted 16 features from bulk and scRNA-seq data separately. For mouse and human bulk RNA-seq data, the Spearman correlation, Spearman p value, range, slope, fit p value, and cell specificity Tau value were extracted. For mouse scRNA-seq data, the Spearman correlation, Spearman p value, range, and slope were extracted. The bulk RNA-seq represented the transcriptome from Fluorescence-activated Cell Sorting (FACS-sorted) cells, while the cell types of scRNA-seq were assigned by the expression of gene markers. For bulk RNA-seq [19,20], cell types were sorted based on lineage commitment and assigned consecutive numbers starting from 1 to represent pseudotime. For example, HSC, MPP, GMP, ProNeu, and PreNeu were set as 1, 2, 3, 4, and 5, respectively. For scRNA-seq, the R package monocle (v2.22.0) [47] was used to infer the pseudotime of 2,803 immature and mature neutrophils from the bone marrow dataset [21] with default parameters. Then features were extracted the same way from bulk RNA-seq and scRNA-seq separately using their respective pseudotime. Homologous genes between mice and humans were aligned using R package biomaRt (v2.46.3) [22]. The Mouse and Human BioMart databases were extracted using the useMart function, followed by the alignment of homologous genes using the getLDS function.

Feature. Spearman.cor and Spearman.p

The correlations between gene expression and pseudotime were calculated. The correlation (Spearman.cor) was calculated by the function cor with parameter ‘method = spearman’ from the R package stats, and the correlation test (Spearman.p) was performed using the function cor.test with parameter ‘method = spearman, alternative = two.side’ from the R package stats (v4.0.2).

Feature. Range

The range of value (Range) was the expression difference value between maximum and minimum values, which means the measures of max variation. The range was calculated as the TPM values for each gene during neutrophil differentiation. Furthermore, the range of value was scaled to 0~1 by redefining the maximum value and labeled as Range.

Feature. Slope and fit p value

Linear regression analyses between gene expression and pseudotime were conducted. The regression coefficient (Slope), symbolizing the extent of influence of the independent variable on the dependent variable, and the p value of the regression variable (fit p value) indicating the statistical significance of this influence, were extracted from the regression model.

Feature. Tau

Cell specificity index (Tau) [79], is a quantitative, graded scalar measure of the specificity of an expression profile. Tau interpolates the range between 0 for housekeeping to 1 for strictly stage or cell-type-specific.

where N is the number of neutrophil lineage commitment cell types, and is the gene expression normalized to the maximal expression.

1. 2. Physiological characteristics

From Open Targets Platform [23], we collected single nucleotide polymorphisms (SNPs) associated with neutrophil-related physiological phenotypes includingneutrophil count, neutrophil measurement, neutrophil percentage of granulocytes, neutrophil percentage of leukocytes, neutrophil to lymphocyte ratio, and sum of neutrophil and eosinophil counts. Furthermore, we conducted a statistical analysis of the host genes for these SNPs. Genes with one SNP were labeled as 1, those with two SNPs were labeled as 2, and genes without SNPs were labeled as 0.

1. 3. Pathological relatedness

First, we searched for neutrophil-related pathological phenotype ‘abnormal neutrophil count’ from MsigDB [25]. The values were transformed into binary form, where genes related to the pathological phenotype were labeled as significant (1), while the rest as nonsignificant (0).

Second, we collected disease features below from DisGeNET [24]

Feature. DSI.

The Disease Specificity Index (DSI) ranges from 0 to 1 and is inversely proportional to the number of diseases associated with a particular gene. A gene associated with a large number of diseases (e.g., TNF, associated with more than 1500 diseases) will have a DSI close to zero, while a gene associated with only one disease, is more ‘specific’ for that disease and has a DSI of 1.

Feature. DPI

The Disease Pleiotropy Index (DPI) ranges from 0 to 1 and is proportional to the number of different (MeSH) disease classes a gene is associated with. Thus, a gene associated with diseases of diverse classes (such as APOE, associated with Cardiovascular Diseases, Mental Disorders, Neoplasms, Respiratory Tract Diseases, etc), will have a DPI close to 1. Conversely, the PSCA, associated with 58 diseases, most of which are neoplasms has a relatively low DPI.

Feature. No of diseases

The number of diseases associated with the gene

Feature. PLI

The probability of the gene being loss-of-function intolerant (PLI) is provided by the gnomAD consortium. PLI closer to 1 indicates that the gene or transcript cannot tolerate protein truncating variation (nonsense, splice acceptor, and splice donor variation). The gnomAD team recommends transcripts with a PLI>= 0.9 for the set of transcripts extremely intolerant to truncating variants. PLI is based on the idea that transcripts can be classified into three categories: 1) null: heterozygous or homozygous protein-truncating variation is completely tolerated; 2) recessive: heterozygous variants are tolerated but homozygous variants are not; 3) haploinsufficient: heterozygous variants are not tolerated. An expectation-maximization algorithm was then used to assign a probability of belonging in each class to each gene or transcript. PLI is the probability of belonging to the haploinsufficient class.

1. 4. Conservation score

Mean phastcons was the only feature in the aspect, which represented the conservation of a gene across different species. The mean phastcons was calculated based on the conservation data phastCons100way.UCSC.hg38 of 99 vertebrates, with the data obtained from the UCSC Genome Browser (https://hgdownload.cse.ucsc.edu/goldenpath/hg38/phastCons100way/) [79].

Step 7: Gaussian mixture model (GMM) classified NeuRGI score.

The ranges of the two peaks in the NeuRGI score density were determined using the findPeaks function from the R package quantmod (v0.4.26) [32]. A Gaussian Mixture Model (GMM) was applied directly to the NeuRGI scores to characterize the distribution of predicted gene scores. The analysis was performed using the mclust package in R (v6.0.0) [77]. The GMM was fitted using the function Mclust with the parameters G = 3 and modelNames = ‘V’, which models the data as originating from three Gaussian distributions with variable variances for each cluster. The model is defined as follows:

Where represents the mixture proportions, and is the Gaussian distribution with mean and covariance for each component k. The parameters , and were estimated from the data using the Expectation-Maximization (EM) algorithm.

Evaluation of NeuRGI performance

1. Area under the receiving operator characteristic curve (AUC).

The area under the receiving operator characteristic curve (AUROC) was used to assess the performance of NeuRGI using the R package pROC (v1.18.5) [83]. The sensitivity and specificity of the model were extracted from the results of the roc function (Fig 1B).

Where TP, FP, TN, and FN represent true positives, false positives, true negatives, and false negatives, respectively.

2. Area under the precision-recall curve (AUC-PR).

The Area Under the Precision-Recall Curve (AUC-PR) is an alternative to AUC, particularly useful when dealing with imbalanced datasets. This metric evaluates the trade-off between precision and recall at different thresholds, focusing more on the positive class. The AUC-PR is the area under the Precision-Recall curve, which plots precision against recall. A higher AUC-PR value indicates better model performance in terms of precision and recall balance. We used pr.curve function from the R package PRROC (v1.3.1) to calculate AUC-PR (Fig 1B).

3. Accuracy (ACC).

Accuracy is the proportion of correct predictions out of the total number of predictions. It is calculated as the ratio of the sum of True Positives (TP) and True Negatives (TN) to the total number of instances. Accuracy is suitable when the dataset is balanced but may be misleading when the dataset is imbalanced.

4. Matthews correlation coefficient (MCC).

The Matthews Correlation Coefficient (MCC) is a more balanced metric for evaluating binary classification performance, taking into account true positives, true negatives, false positives, and false negatives. The MCC value ranges from -1 (perfect disagreement) to +1 (perfect agreement), with 0 indicating random classification performance.

5. F1 score.

The F1 Score is the harmonic mean of precision and recall, providing a single metric that balances both. It is particularly useful when dealing with imbalanced datasets, as it emphasizes both false positives and false negatives. The F1 Score is calculated as follows:

Feature importance

We applied the mean decrease coefficient in the Gini Coefficient [84] to evaluate the importance of the features. This coefficient indicates how well features split the data at each node in the trees, reflecting its substantial role in improving node purity and classification accuracy. Gini impurity reduction was calculated for each split in decision trees. The decrease in Gini from each feature was summed up across all nodes and trees. This sum was divided by the total number of splits of features, producing the Mean Decrease Gini index for each feature. The Mean Decrease Gini index was calculated as follows:

where G(t) represents the Gini impurity at node t, pi is the proportion of class i samples at node t, and J is the total number of classes. N is the number of samples at node t, and nL and nR are the number of samples at the left and right child nodes, respectively.

Methods comparison with CellOracle and SCENIC

To compare NeuRGI with CellOracle [9] and SCENIC [10], 2,803 immature and mature neutrophils from mouse bone marrow scRNA-seq data were used [21].

GO network analysis of predictive functional genes

GO-BP enrichment analysis of predictive functional genes was performed using enrichGO from R package clusterProfiler (v4.2.2) [86]. The top 100 significant (p.adjust < 0.05) GO terms were reserved to calculate the GO semantic similarity matrix using the function GO_similarity from the R package simplifyEnrichment (v1.0.0) [87]. The similarity matrix was then clustered using the function simplifyGO with parameter method = ‘kmeans’ from R package simplifyEnrichment (v1.0.0). The GO network was finally visualized using Cytoscape [88] (v3.8.2).

MAGMA

To screen for genes related to neutrophil counts, MAGMA [46] was used to compute gene p value using GWAS summary statistics data. Firstly, we used a reference panel based on individuals of European ancestry in the 1000 Genomes Project [89]. Then we used a 10-kb window around the gene body to map SNPs to genes. And GWAS summary statistics ‘neutrophil count’ (ebi-a-GCST004629) was downloaded from the European Bioinformatics Institute (EBI) database (https://www.ebi.ac.uk/gwas/downloads/summary-statistics). GWAS summary statistics were next used as input for MAGMA (v 1.10) [46] to compute the Z score with default parameters.

Predictive modeling of genetic perturbations using biologically informed variational autoencoders

See S1 Text.

Mice

The C57BL/6 wildtype mice and Map4k4^f/− mice were purchased from Gem Pharmatech Co. Ltd. Map4k4^f/− mice were backcrossed against C57BL/6 mice, then intercrossed to obtain Map4k4^f/f mice. The Chen Chong laboratory kindly provided mx1-cre mice and similarly backcrossed onto C57BL/6. Map4k4^f/f mice were mated to Mx1-Cre Map4k4^f/f mice to generate Mx1-Cre Map4k4^f/− mice, then intercrossed to obtain Mx1-Cre Map4k4^−/− (control, Ctrl) and Mx1-Cre Map4k4^f/f (Map4k4-cKO, cKO) littermates for phenotypic analysis. Floxed alleles were deleted in Mx1-Cre Map4k4^f/f mice 5-6 weeks before experiments by five intraperitoneal injections (300μg per mouse) of poly I:C (Sigma, #P1530-100MG) in PBS every other day. Mice were housed and bred under SPF conditions (Specific Pathogen Free) at the West China Second University Hospital Laboratory Animal Center. They were allowed access to diet and water ad libitum. All animal experiments followed the protocols approved by the Institutional Animal Care and Use Committee of West China Second University Hospital [(2018) Animal Ethics Approval No. 004].

Cell culture

K562, HL-60, and HEK293T cell lines were initially purchased from the American Type Culture Collection (ATCC). K562 cells were maintained in IMDM medium (Gibco, #C12440500BT) containing 10% fetal bovine serum (FBS, YEASEN, #40130ES76) and 1% penicillin-streptomycin (Hyclone, #SV30010). HEK293T cells were maintained in DMEM medium (Gibco, #C11995500BT) supplemented with 10% FBS and 1% penicillin-streptomycin. HL-60 cells were maintained in RPMI 1640 medium (Gibco, #C11875500BT) supplemented with 10% FBS and 1% penicillin-streptomycin. HL-60 was induced by supplementing 1 µM all-trans-retinoic acid (MCE, #HY-14649).

Viral production and transduction

The human STAT5A and STAT5B cDNA sequences were synthesized by Sangon Biotech and were reconstituted into the mammalian expression plasmid Plenti-EF1a-MCS-FLAG-6×His-CMV-mCherry-Puro (MiaoLing Biology, P59992). For human genes, sgRNAs targeting MAP4K4 or sgAAVS1 as control were connected into Plenti-CRISPR V2, donated by Wang Yuan Laboratory, via the Bsmb1 restriction site.

Viral transduction

HEK293T cells were seeded at 6 × 10⁶ cells per 10 cm culture dish and maintained in high-glucose DMEM (Gibco, #C11995500BT) supplemented with 10% FBS for 12–24 hours (h). Subsequently, a co-transfection was performed with 12 μg of the lentiviral construct (sgAASV1 and 6 μg of psPAX2, 3 μg of pMD2.G packaging plasmid, using Liposomal Transfection Reagent (YEASEN, #40802ES03) according to the manufacturer’s instructions. After 48 h of transfection, the supernatants were collected, and the culture was replaced with a fresh medium. Supernatants were again collected at 72 h after transfection. The collected supernatants were pooled, filtered through a 0.45 μm strainer, and centrifuged at 1000 g at 4 °C for 90 min. The concentrated viral particles were added to the cultured K562 cells and maintained for 36–48 h.

Western blotting

K562 cells were lysed in RIPA lysis Buffer (Thermo Fisher Scientific, #89900) containing 10mM Protease Inhibitor Mini Tablets, EDTA-free (Thermo Fisher Scientific, #A32955) and phosphatase inhibitors (Beyotime, P1045) on ice for 30 min and centrifuged. Protein content in the supernatant was quantified using a Pierce BCA Protein Assay Kit (Thermo Fisher Scientific, #23227). Protein extracts were separated by 10% SDS polyacrylamide gel electrophoresis (PAGE) and transferred to nitrocellulose membranes (Millipore, #ISEQ00010). The blots were probed with the primary antibodies including rabbit anti-MAP4K4 antibody (CST, #3485), rabbit anti-p-STAT5A (Tyr699) antibody (Abcam, # ab32043), rabbit monoclonal (E289) anti-STAT5A antibody (Abcam, #ab32043), rabbit anti-STAT5B antibody (Abcam, #ab30648), rabbit monoclonal (C11C5) anti-p-STAT5 (Tyr694) antibody (CST, #9359), rabbit monoclonal (D2O6Y) anti-STAT5 antibody (CST, #94205), rabbit monoclonal (14C10) anti-GAPDH antibody (Cell Signaling Technology, #2118) in the universal antibody diluent (NCM biotech, #WB500D) for overnight at 4°C, washed three times with TBST, and then incubated with the HRP conjugated goat anti-rabbit antibodies (HuaBio, #HA1001). The chemiluminescence signal was detected using a ChemiDOCTMMP Imaging System (Bio-Rad Laboratories).

Single-cell suspension

Bone marrow cells (BMCs) were collected from the tibia, femur, and ilium, lysed with red blood cell lysis buffer (Solarbio, #R1010), washed with PBS, and filtered through a 70 µm cell strainer (biosharp, #BS-70-CS).

Place the spleen on a 70 µm cell strainer (biosharp, #BS-70-CS), and use a sterile syringe plunger to gently break apart the spleen tissue. Rinse the strainer with PBS to collect the cells into a tube. Remove the supernatant, treat with a hemolytic solution to lyse red blood cells, then wash and resuspend the cells in PBS.

Cell collection, RNA extraction, and qPCR

RNA was extracted using a Cell Total RNA Isolation Kit (FORE GENE, #RE-03113). Complementary DNA (cDNA) was synthesized using the RevertAid first-strand cDNA synthesis kit (Thermo Scientific, #K1621). PCR was performed using the 2 × Phanta Max Master Mix (Vazyme, P515). Real-time qPCR reactions were performed using the Blue qPCR SYBR Master Mix (YEASEN, #11184) on a Bio-Rad (CFX Connect) thermocycler. Primers were listed in the S15 Table.

Flow cytometry assay

Single-cell suspension of BMCs, splenocytes, peripheral blood cells, or HL-60 cells were incubated with Fixable Viability Dyes (FVS, Biosciences) in PBS at room temperature in the dark for 20 minutes. The cells were next washed with staining buffer and incubated with fluorochrome-conjugated antibodies listed below in staining buffer at 4 °C for 30 minutes. After thorough washing to remove excess antibodies, BMCs, splenocytes, and peripheral blood cells were analyzed using an Aurora CS Flow Cytometer (CYTEK), and Hl-60 cells were analyzed using an Attune Nxt (Life Technologies) flow cytometer. The data were analyzed using FlowJo software (BD).

To characterize hematopoietic stem and progenitor cells, cells were stained with BV785-conjugated anti-Sca-1 (BD Biosciences, #563991), BV711-conjugated anti-CD16/32 (BioLegend, #101337), BV605-conjugated anti-CD71 (BD Biosciences, #563013), BV421-conjugated anti-CD135 (BioLegend, #135315), PerCP-Cy5.5-conjugated anti-CD34 (BioLegend, #119328), PE-CF594-conjugated anti-CD127 (BioLegend, #135315), PE-conjugated anti-CD3e (BD Biosciences, #553062), PE-conjugated anti-Gr-1 (BioLegend, #108408), PE-conjugated anti-CD11b (BD Biosciences, #557397), PE-conjugated anti-B220 (BioLegend, #103208), PE-conjugated anti-TER119 (BD Biosciences, #553673), APC-Cy7-conjugated anti-CD117 (Biolegend, #105826), FITC-conjugated anti-CD41 (BD Biosciences, #553848).

To characterize bone marrow neutrophils, cells were stained with PerCP-Cy5.5-conjugated anti-CD34 (BioLegend, #119328), APC-Cy7-conjugated anti-CD117 (Biolegend, #105826), PE-Cy7-conjugated anti-Ly6G (BioLegend, #127618), PE-conjugated anti-CD3e (BD Biosciences, #553062), PE-conjugated anti-B220 (BioLegend, #103208), PE-conjugated anti-TER119 (BD Biosciences, #553673).

To characterize innate immune cells, BMCs were stained with BV711-conjugated anti-CD11b (Biolegend, #101242); BV605-conjugated anti-Ly6C (Biolegend, #128036), BV421-conjugated anti-F4/80 (Biolegend, #123137), APC-Cy7-conjugated anti-MHCII (BioLegend, #107628), PE-Cy7-conjugated anti-Ly6G (BioLegend, #127618), PE-conjugated anti-TER119 (BD Biosciences, #553673), PE-conjugated anti-CD3e (BD Biosciences, #553062), PE-conjugated anti-19 (BioLegend, #115508), PE-conjugated anti-Nkp46 (BioLegend, #137604), PE-conjugated anti-CD117 (BioLegend, #105808). For splenocytes, cells were stained with the same antibody panel except CD117. For peripheral blood cells, cells were stained with BV711-conjugated anti-CD11b (Biolegend, #101242), PE-Cy7-conjugated anti-Ly6G (BioLegend, #127618), PE-conjugated anti-TER119 (BD Biosciences, #553673). HL-60 cells were stained with PerCP-Cy5.5-conjugated anti-CD11b (BD Biosciences, #561114).

Apoptosis assay

Single-cell suspension of BMCs or HL-60 cells was washed with PBS and subsequently resuspended in 1× Annexin V Binding Buffer (BD Biosciences, #556454) at a concentration of 1 × 10⁶ cells/mL. The cell suspension (400 µL) was treated with Annexin V-BV421 (1:100 dilution; BD Biosciences, #563973) and SYTOX Green Dead Cell Stain (1:1000 dilution; Invitrogen, #S34860). Following incubation at room temperature for 30 min in the dark, the stained cells were immediately subjected to flow cytometric analysis.

Neutrophil phagocytosis

E. coli OP50-GFP was cultured overnight at 37°C, then washed in PBS and counted. A suspension of E. coli was prepared to a concentration of 10⁶ colony-forming units (CFU) and was incubated with 5 × 10⁵ neutrophils in a total of 1000 μL of RPMI media. The mixture was rotated at 8 rpm at 37°C for 30 minutes. Following the incubation, the sample was transferred to ice to halt phagocytosis. Neutrophils were then collected and stained with BV711-conjugated anti-CD11b (Biolegend, #101242) and PE-Cy7-conjugated anti-Ly6G (BioLegend, #127618) at 4°C. After washing with cold PBS 3 times, the phagocytosis percentages of gated CD11+ Ly6G+ cells were determined using a FACS scanner.

Reactive oxygen species

A single-cell suspension of BMCs, spleen, or peripheral blood cells was incubated with 1 μM CellROX reagent (A) at 37°C in a 5% CO₂ atmosphere for 1 hour (Thermo Fisher Scientific, #C10492). During the final 15 minutes of incubation, 1 μL of 5 μM SYTOX Red Dead Cell stain solution (B) was added per 1 mL of the suspension. The stained cells were immediately analyzed by flow cytometry. Neutrophils were identified using anti-CD11b and anti-Ly6G monoclonal antibodies.

Colony-forming unit assay

Bone marrow cells (10⁵) were plated in methylcellulose culture medium (MethoCult M3434, STEMCELL Technologies) and incubated at 37°C with 5% CO₂. Colonies derived from CFU-E were counted at 48 h after plating, and the colonies derived from primitive erythroid progenitor cells (BFU-E) and granulocyte-macrophage progenitor cells (CFU-GM, CFU-G, and CFU-M) were counted 10-14 days after plating.

Single-cell RNA-sequencing data analysis of HSPCs of Ctrl and Map4k4-cKO mice

Co-expression analysis of neutrophils in Ctrl and Map4k4-cKO samples

Co-expression analysis was carried out using the R package hdWGCNA (v0.2.2) [59] and the R package WGCNA (v1.71) [91]. 11031 neutrophil-related cells including HSC/MPP, GMP, ProNeu, PreNeu, imNeu, and mNeu were selected for co-expression network analysis. We retained 54838 genes which were expressed in at least 5% of cells from any cluster. Metacell transcriptomic profiles were constructed separately for each of the 2 conditions and each cell type using the hdWGCNA function MetacellsByGroups with parameters ‘k = 25, max_shared = 10, min_cells = 0’. We selected a soft-power threshold β = 9 based on the parameter sweep performed with the TestSoftPowers function. The co-expression network was computed with the ConstructNetwork function with the following parameters: networkType= ‘signed’, TOMType= ‘signed’, soft_power=9, deepSplit=4, detectCutHeight=0.995, minModuleSize=50, mergeCutHeight=0.2. Module eigengenes were computed using the ModuleEigengenes function, and we applied Harmony to correct MEs based on the sequencing batch. Eigengene-based connectivity for each gene was computed using ModuleConnectivity. The co-expression network was embedded in two dimensions using UMAP with the RunModuleUMAP function with the top five genes (ranked by eigengene-based connectivity) per module as the input features. Differential condition module eigengene analysis was performed using R package energy (v1.7.7). The p value was calculated by the function dcorT.test. dcor is a transformation of a bias-corrected version of distance correlation calculated by the function dcorT. Hub genes of each module were extracted using the function GetHubGenes. Apoptosis score [92] was calculated using the function AddModuleScore from the Seurat R package [58] (v4.3.0) with apoptotic-related gene sets including GO:0043065 (S12 Table).

Protein and protein phosphorylation mass spectrometry analysis

Approximately 1 x 10⁸ MAP4K4 KO K562 cells, or control cells, were harvested to facilitate the identification and quantification of proteins and protein phosphorylation using proteomics mass spectrometry.

Selection of proteins with different protein expression or protein phosphorylation levels after MAP4K4 KO.

The fold change (FC) is defined as the ratio of the mean relative quantification values between MAP4K4-KO and Ctrl samples. The log₂FC is calculated as follows:

Where R represents the relative quantification value of the modified site, i denotes the sample, and k represents the modified site. A t-test was performed to test the significance. Proteins with P < 0.05 and abs(log₂FC) > 0.2 were identified as differentially expressed proteins. Phosphorylation sites with P < 0.05 and abs(log₂FC) > 0.2 were identified as differentially expressed phosphorylation sites. Since a protein may have multiple differentially phosphorylated sites, the change in the phosphorylation level of a protein is represented by the site with the largest change.

GO-BP and GSEA analysis

GO-BP was performed using enrichGO and compareCluster from R package clusterProfiler [86] (v4.2.2). Gene Set Enrichment Analysis (GSEA) was performed using the function GSEA of R package clusterProfiler (v4.2.2) [86] with pvalueCutoff = 1. Genes were initially sorted based on log₂ (FC). The enrichment was then performed using the GSEA function. The pvalueCutoff parameter was set to 0.5, and the cutoff of the p value was 0.05. Visualization was achieved with function gseaplot2 from the R package enrichplot (v1.10.2) [93]. Gene sets used to perform enrichment analysis were downloaded from org.Mm.e.g.,db (v3.12.0) package.

Statistical analysis

Statistical analyses were performed by the R software 4.0.2 (https://www.R-project.org/). One-tailed Wilcoxon test was used to identify pathways that were significantly dysregulated upon knockout by OntoVAE. The two-sided Student’s t-test was used to identify significantly differentially expressed genes in scRNA-seq data. Distance correlation t-test was used for hdWGCNA analysis. The two-sided Student’s t-test was performed to identify proteins with different protein expression or protein phosphorylation levels after MAP4K4 KO. For in vivo and in vitro experiments, data were presented as the mean ± SD of at least four independent experiments. The Mann-Whitney U test was utilized to analyze differences between experimental groups, except the RT-qPCR data, which were analyzed using a two-sided Student’s t-test. The result with p value < 0.05 was considered statistically significant unless otherwise specified.