Integrative approach to sporadic Alzheimer’s disease: deficiency of TYROBP in cerebral Aβ amyloidosis mouse normalizes clinical phenotype and complement subnetwork molecular pathology without reducing Aβ burden

Integrative gene network approaches enable new avenues of exploration that implicate causal genes in sporadic late-onset Alzheimer’s disease (LOAD) pathogenesis, thereby offering novel insights for drug-discovery programs. We previously constructed a probabilistic causal network model of sporadic LOAD and identified TYROBP/DAP12, encoding a microglial transmembrane signaling polypeptide and direct adapter of TREM2, as the most robust key driver gene in the network. Here, we show that absence of TYROBP/DAP12 in a mouse model of AD-type cerebral Aβ amyloidosis (APPKM670/671NL/PSEN1Δexon9) recapitulates the expected network characteristics by normalizing the transcriptome of APP/PSEN1 mice and repressing the induction of genes involved in the switch from homeostatic microglia to disease-associated microglia (DAM), including Trem2, complement (C1qa, C1qb, C1qc, and Itgax), Clec7a and Cst7. Importantly, we show that constitutive absence of TYROBP/DAP12 in the amyloidosis mouse model prevented appearance of the electrophysiological and learning behavior alterations associated with the phenotype of APPKM670/671NL/PSEN1Δexon9 mice. Our results suggest that TYROBP/DAP12 could represent a novel therapeutic target to slow, arrest, or prevent the development of sporadic LOAD. These data establish that the network pathology observed in postmortem human LOAD brain can be faithfully recapitulated in the brain of a genetically manipulated mouse. These data also validate our multiscale gene networks by demonstrating how the networks intersect with the standard neuropathological features of LOAD.

We sought to validate in vivo in mice the driver role of TYROBP in LOAD and to demonstrate that the manipulation of Tyrobp leads to changes in subnetworks that mimic the changes observed in the brain in human LOAD. We previously demonstrated that 4-month-old transgenic APP KM670/671NL /PSEN1 Δexon9 (hereafter abbreviated as APP/ PSEN1) mice on a Tyrobp-null (Tyrobp −/− ) background had no apparent changes in the transcription of the constituents of the complement subnetwork [28]. Since all forms of AD are aging-related, we performed the identical battery of molecular, behavioral, electrophysiological, and transcriptomic analyses in 8-month-old mice. As observed in 4-month-old mice, TYROBP deficiency played a protective effect in behavior and electrophysiology in 8-month-old mice. However, unlike the 4-month-old mice, we observed that the levels of the transcripts comprising nearly the entire Tyrobp-driven complement subnetwork were reduced. Therefore, constitutive Tyrobp knockout in the brain of a mouse with Aβ-amyloid deposits, recapitulates, in an agingrelated manner, the complement subnetwork first observed in human LOAD brain [5]. As one would expect, the transcriptomic subnetwork and hub that were associated with increased expression in LOAD brain were associated with reduced representation in the APP/PSEN1;Tyrobp −/− brain.
This is the first direct observation that the complement transcriptome subnetwork molecules predicted from integrative models of human postmortem LOAD brain can be recapitulated in a cerebral amyloidosis mouse model. Moreover, the levels of network component transcripts are reversed in a mouse in which the predicted driver has been knocked out. The ability of a single subnetwork hub/driver to prevent the appearance of the typical behavioral and electrophysiological phenotypes of the APP/PSEN1 mouse while recapitulating the predicted subnetwork characteristics can now serve as a biological template for designing, screening, and repurposing molecular interventions in an effort to identify a biologic or a small molecule drug capable of mimicking the profile of the 8-month-old, Aβ-amyloid-depositing, Tyrobp knockout mouse. Our prediction is that drugs identified in this manner may be useful in the prevention or treatment of LOAD.

Animals
The experimental procedures were conducted in accordance with NIH guidelines for animal research and were approved by the Institutional Animal Care and Use Committee (IACUC) at Icahn School of Medicine at Mount Sinai. All mice were on a C57Bl6/J background. APP KM670/671NL / PSEN1 Δexon9 (=APP/PSEN1) [29], Tyrobp-null or knockout (Tyrobp −/− ) [30] were obtained from Jackson Laboratories and Taconic/Merck Laboratory respectively. APP/PSEN1 were crossed with Tyrobp −/− mice to obtain APP/PSEN1; Tyrobp +/− . APP/PSEN1;Tyrobp +/− were then crossed with Tyrobp +/− to obtain WT, Tyrobp +/− , Tyrobp −/− , APP/ PSEN1, APP/PSEN1;Tyrobp +/− or APP/PSEN1;Tyrobp −/− mice. The 8-month-old male and female mice were sacrificed by decapitation. One hemisphere was collected and immersion-fixed in 4% paraformaldehyde for immunohistochemistry analysis. The other hemisphere was dissected and prefrontal cortex (PFC) was collected for transcriptomic analysis. PFC and hemibrain were then snapfrozen and stored at −80°C prior to RNA isolation or biochemistry analysis. A standard number of mice (4-5 per group and 4-5 per sex) were used for most of the analyses as this size gives us 80% power to see differences of effect size of at least 2.5 between groups at α = 0.05. An increased standard number of mice were used per group and per sex for behavioral and electrophysiological analyses as these analyses are often variable (sample size "n" is specified in the figure legend).

RNA isolation
Snap frozen samples from the PFC of male and female mice were homogenized in QIAzol Lysis Reagent (Qiagen). Total RNA purification was performed with the miRNeasy Mini kit (Qiagen). RNA quantification and quality were evaluated by Agilent BioAnalyzer. RNA integrity was checked using the RNA 6000 Nano assay (Agilent, CA, USA). All processed RNA samples had RQN/RIN value of 8.4 or greater.
RT-qPCR 500 ng of total RNAs were reversed transcribed using the High-Capacity RNA-to-cDNA Kit (Applied Biosystem, Foster City, CA, USA). cDNAs were subjected to real-time qPCR in a StepOne Plus system (Applied Biosystem) using The All-in-One qPCR Mix (GeneCopoeia, Rockville, MD, USA). qPCR consisted of 40 cycles, 10 s at 95°C, 20 s at 60°C, and 15 s at 70°C each, followed by dissociation curve analysis. Ct values were normalized to the expression level of L32 and relative gene expression was calculated using the ΔΔCt method [31]. Sequences of primers used:

RNA sequencing
Total RNAs from PFC of 8-month-old male WT (n = 4), Tyrobp +/− (n = 3), Tyrobp −/− (n = 4), APP/PSEN1 (n = 4), APP/PSEN1;Tyrobp +/− (n = 4) or APP/PSEN1;Tyrobp −/− (n = 5) mice were subjected to RNA sequencing. The sequencing library was prepared with the TruSeq RNA Sample Prep Kit v2 protocol (Illumina, CA, USA). Ribosomal RNAs were removed using the Ribo-Zero rRNA Removal Kit (Human/Mouse/Rat) (Illumina, CA, USA). Remaining RNAs were fragmented, and the cDNAs synthesized using random hexamers, end-repaired, and ligated with appropriate adapters for sequencing. The library was processed for size selection and purification using AMPure XP beads (Beckman Coulter, CA, USA). The appropriate Illumina-recommended 6 bp barcode bases were introduced at one end of the adapters during the PCR amplification step. The size and concentration of the RNA-seq libraries were measured by Bioanalyzer and Qubit fluorometry (Life Technologies, NY, USA) before loading onto the sequencer. The rRNA-depleted libraries were sequenced on the Illumina HiSeq 2500 System with 100 nucleotide singleend reads (Illumina, CA, USA).

Differential expression analysis and gene set enrichment analysis
The single-ended raw sequencing reads were aligned to mouse mm10 genome using Star aligner (version 2.5.0b). FeatureCounts was used to quantify gene expression at the gene level based on UCSC gene model. Genes with at least 1 count per million in at least one sample were considered expressed and retained for further analysis. The gene level read count data were normalized using trimmed mean of the M-values normalization method [32] to adjust for sequencing library size difference. Differential gene expression between groups was predicted by a linear model analysis using Bioconductor package limma [33]. To adjust for multiple tests, the false discovery rate (FDR) of the differential expression test was estimated using the Benjamini-Hochberg method [34]. Gene set enrichments were calculated using ingenuity pathway analysis (IPA).

Network analysis using human AD postmortem brain gene-expression data
We utilized large-scale postmortem brain RNA sequencing data from two human AD cohorts (MSBB [6,35] and ROSMAP [36]) to build gene co-expression networks and Bayesian causal networks to model the regulation of geneexpression traits in brain samples. The ROSMAP and MSBB RNA-seq data and associated genomics data are available in the Accelerating Medicines Partnership in AD (AMP-AD) Knowledge Portal at Synapse (https://www. synapse.org/#!Synapse:syn2580853) upon authentication by the AMP-AD Consortium. We downloaded normalized RNA-seq gene-expression abundance data from Synapse. Genes expressed in at least 10% of the samples were selected, and the data were corrected for confounding factors including batch, PMI, sex, and RIN score. The construction of gene co-expression networks and Bayesian networks is detailed in Suppl. methods. Immunohistochemistry 30µm thick sagittal free-floating sections were incubated with anti-Iba1 (1:500, cat#019-19741, Wako, Richmond, VA, USA), 6E10 (1:1000, cat#9320-500, Covance, Princeton, NJ, USA), or C1q (1:1000, cat#ab182451, Abcam, Cambridge, MA, USA) antibodies. Sections probed with Iba1 and 6E10 antibodies were incubated with anti-rabbit Alexa 488 (1:400, cat#A-11008, Thermo Fisher Scientific, Grand Island, NY, USA) and anti-mouse Alexa 594 (1:400, cat#A-11005, Thermo Fisher Scientific) secondary antibodies. Sections probed with C1q were incubated with biotinylated anti-mouse antibody (1:1000, BA-9200, Vector laboratories, Burlingame, CA, USA) and developed with Vectastain ABC Kit (PK-4000, Vector Laboratories) (see Suppl. methods for details). Images were acquired using an Olympus BX61 microscope or a Pannoramic 250 digital scanner (3DHISTECH, Budapest, Hungary).
For measuring percentage of 6E10-immunoreactive areas, 6E10-immunolabeled sections were thresholded and the percentage of 6E10 positive area was calculated using Fiji (v2.0.0) and the "measure" function. For measuring microglial number, Iba1 positive cells in the region of interest (ROI) were manually counted and number was normalized to the area of the ROI. ROI (PFC and hippocampus) were determined and outlined by manual tracing.
To quantify Aβ42 and Aβ40 levels, human/rat Aβ1-40/ 1-42 ELISA kits (Wako, Richmond, VA) were used according to the manufacturer's instructions. Absolute concentrations of total or oligomeric Aβ were normalized to initial tissue weight.

Field electrophysiology
Coronal brain slices containing the hippocampal formation were prepared as previously described [28,42]. Animals were anesthetized with isoflurane, and brains were rapidly removed from the skull and placed in an ice-cold modified ACSF solution (215 mM sucrose, 2.5 mM KCl, 1.6 mM NaH 2 PO 4 , 4 mM MgSO 4 , 1 mM CaCl 2 , 4 mM MgCl 2 , 20 mM glucose, 26 mM NaHCO 3 , pH = 7.4). Coronal brain slices (400 µm thick) were prepared with a Vibratome VT1000S (Leica Microsystems) and then incubated at room temperature for ≥3 h in a physiologic ACSF (120 mM NaCl, 3.3 mM KCl, 1.2 mM Na 2 HPO 4 , 26 mM NaHCO 3 , 1.3 mM MgSO 4 , 1.8 mM CaCl 2 , 11 mM Glucose, pH = 7.4). The hemislices were transferred to a recording chamber perfused with ACSF at a flow rate of~2 mL/min using a peristaltic pump; experiments were performed at 28.0 ± 0.1°C. Recordings were acquired with a GeneClamp 500B amplifier (Axon Instruments) and Digidata 1440A (Molecular Devices). All signals were low-pass filtered at 2 kHz and digitized at 10 kHz. Field recordings were obtained using a patch-type pipette filled with ACSF and placed in the middle third of stratum radiatum in area CA1. Field excitatory postsynaptic potentials (fEPSPs) were evoked by activating the Shaffer collaterals with a concentric bipolar electrode placed in the middle third of stratum radiatum 150-200 µm away from the recording pipette (Fig. 5a). Square-wave current pulses (60 ms pulse width) were delivered through a constant-current stimulus isolator (Isoflex, AMPI). Input-output curves were generated by a series of stimuli ranging from 30 to 80 µA. Paired-pulse facilitation (PPF) was measured by delivering two stimuli at 20, 50 and 100 ms interstimulus intervals. Each interstimulus interval was repeated three times, and the resulting potentials were averaged. The paired-pulse ratio was calculated by dividing the slope of the second EPSP by the slope of the first EPSP. All results were analyzed by ANOVA followed by Tukey post hoc tests. Baseline recordings (stable for 20 min) were made every 30 s using stimuli that yielded a response equal to 50% of spike threshold. For long-term potentiation (LTP) and long-term depression (LTD) experiments, synaptic strength was measured by EPSP slope with stimuli that yielded a response equal to 30-40% of spike threshold, delivered every 30 s. LTP was induced after 20 min of stable baseline recording with theta burst stimulation, consisting of a series of 10 bursts of 4 stimuli (100 Hz within the burst and 200 ms interburst interval) repeated four times (10 s apart) and delivered at an intensity giving a baseline response equal to 75% of spike threshold. LTD was induced after 20 min of stable baseline recordings by bath application of (RS)-dihydroxyphenylglycine (DHPG; 100 µM), a selective mGluR1/mGluR5 agonist, for 5 min. fEPSPs were collected for at least 60 min during DHPG washout.

Barnes Maze
The Barnes Maze test [43,44] consists of a rotatable circular platform (1.22 m in diameter and 1 m from the floor) with 20 holes in periphery. A removable box was placed underneath one of the holes for escape, and visual cues were placed on the walls of the room. Mice were transported from their cages to the center of the platform via a closed starting chamber where they remained for 10 s prior to exploring the maze for 3 min. Mice failing to enter the escape box within 3 min were guided to the escape box by the experimenter, and the latency was counted as 180 s. Mice were allowed to remain in the escape box for 1 min before the next trial. Two trials per day were performed on four consecutive days. The platform and the escape box were wiped with 70% ethanol after each trial to eliminate the use of olfactory cues to locate the target hole. All trials were recorded by video camera and analyzed with ANY-maze video tracking software (Stoelting Co, Wood Dale, USA).

Statistics
Graphs represent the mean of all samples in each group ± SEM. n Values and statistical tests are indicated in figure legends. Shapiro-Wilk normality tests were used. Analyses used include one-way ANOVA, two-way ANOVA, and Mann-Whitney tests. Significance is set at p value ≤ 0.05. No mice were excluded for RNA sequencing analysis. For biochemical, histological, behavioral, and electrophysiological analyses, outliers were detected using Grubbs's test (extreme studentized deviate method) with α = 0.05. Investigators were blinded to sample identity for behavioral, electrophysiological, and histological analyses. Sequencing library and RNA sequencing were performed by the Genomics Core Facility at Icahn School of Medicine at Mount Sinai. Samples were assayed with operators blinded to sample identity (e.g., genotype).

Data and software availability
Gene expression data have been deposited electronically to the Synapse Web Portal (https://www.synapse.org, accession number syn10378730) in accordance with data sharing policies established by the NIH AMP-AD consortium. Specific software will also be made available upon request.

Results
In order to validate in vivo in mice the driver role of TYROBP in sporadic LOAD and to demonstrate that the manipulation of TYROBP level leads to molecular changes observed human LOAD, we performed a battery of molecular, behavioral, electrophysiological analyses, and generated transcriptomic profiles in 8-month-old WT and APP/PSEN1 mice that were either WT, heterozygous-or homozygous-null for Tyrobp (Suppl. Figure 1).

TYROBP is upregulated in human sporadic LOAD postmortem samples
We evaluated TYROBP gene expression at different stages of LOAD in a large-scale postmortem brain transcriptomic dataset from the Mount Sinai Brain Bank (MSBB) AD cohort [6,35]. In all four brain regions profiled with RNA-sequencing (RNA-seq) from this cohort, TYROBP mRNA was upregulated by at least 1.2-fold in the demented subjects compared to the nondemented controls (p < 0.05) (Suppl. Figure 2). We confirmed the disease-associated upregulation of TYROBP in brains in a second large postmortem brain RNA-seq dataset from the ROSMAP AD cohort (1.1-fold, p = 0.02) (data not shown).
In an APP/PSEN1 mouse, constitutive absence of TYROBP prevents the expression of proinflammatory and microglial sporadic LOADassociated genes We generated transcriptomic profiles of 24 PFC samples from 8-month-old male WT and APP/PSEN1 mice that were either WT, or heterozygous-or homozygous-null for Tyrobp (n = 3-5 samples per group). At an FDR of 0.05, Tyrobp was the only differentially expressed gene (DEG) in heterozygous-null mice in comparison to WT and APP/ PSEN1 mice, indicating that in a WT or pathological (APP/ PSEN1) background, deletion of one allele of Tyrobp does not induce a significant transcriptomic change at a 5% FDR (Fig. 1a). The Tyrobp homozygous-null transcriptome contained 20 DEGs in comparison to WT mice. Among them, 5 were upregulated and 15 were downregulated (Fig. 1a). These results are consistent with our previous report in 4-month-old mice showing that deficiency or absence of TYROBP in a WT background without immune stimulation perturbs expression of only a limited number of genes [28] (Suppl. Figure 3a).
We next evaluated the changes of the transcriptome in the APP/PSEN1 mouse model and identified 181 DEGs in comparison to WT (Fig. 1a). Over 90% of the identified DEGs were upregulated (165 upregulated and 16 downregulated genes). Among the most significantly upregulated genes, we observed several LOAD-associated genes, including Trem2, Tyrobp, C1q, Itgax (aka CD11c), Cst7, Clec7a, and Gfap (Fig. 1a, see Suppl. Table 1 for full DEGs results) highlighting the induction of immune response components in the 8-month-old APP/PSEN1 mice.
Given the apparent decreased expression of nearly 50% of the upregulated DEGs associated with the APP/PSEN1 mice, we next compared the transcriptome of APP/PSEN1; Tyrobp −/− mice vs. WT mice. We identified 55 DEGs (25 up and 30 down), substantially fewer than the 181 DEGs between APP/PSEN1 vs. WT (Fig. 1a). Notably, expression of several chemokines, transcription factors, and DAM genes increased in APP/PSEN1 mice vs. WT were normalized in APP/PSEN1;Tyrobp −/− mice vs. WT (Fig. 1a, Suppl. Table 1). Accordingly, comparison of the transcriptome of APP/PSEN1;Tyrobp −/− mice vs. Tyrobp −/− mice reported a limited number of DEGs (12 DEGs; 9 up; and 3 down) (Fig. 1a, Suppl. Table 1). While the absence of TYROBP in the APP/PSEN1 did not normalize the expression of the astrocytic genes C4b and Gfap vs. WT or Tyrobp −/− , our data suggest that the absence of TYROBP in the APP/PSEN1 mouse model substantially decreases the induction of the pro-inflammatory immune response observed in the brains of APP/PSEN1 mice.

Constitutive absence of TYROBP prevents dysregulation of the major AD-related complement transcriptomic subnetwork
To identify biological pathways that may be dysregulated in WT and APP/PSEN1 mice WT or KO for Tyrobp, we performed gene set enrichment analysis (GSEA) using IPA (Fig. 2). We first sought to identify the pathways that were specifically dysregulated in the 8-month-old amyloidosis mouse model and performed GSEA on the list of genes that were differentially expressed in the APP/PSEN1 mice in comparison to WT mice (FDR < 0.05). At −log(FDR) > 1.3, 56 canonical pathways were dysregulated in the APP/ PSEN1 vs. WT mice, several of which were associated with immune response and glial cell activation (Fig. 2a for selected canonical pathways, see Suppl. Table 2 for full results of canonical pathway analyses). Top dysregulated pathways included phagosome formation and maturation, neuroinflammation signaling, complement system, dendritic cell maturation, production of nitric oxide (NO) and reactive oxygen species (ROS), and IL-10 and IL-8 signaling. A list of DEGs associated with selected dysregulated pathways is presented in Fig. 2b. Prediction of the activation state revealed activation of several pathways including neuroinflammation signaling, TREM1 signaling, production of NO and ROS, and complement system in the APP/PSEN1 mice vs. WT mice (−log(FDR) > 1.3 and z-score > 2) (Fig. 2c, Suppl. Table 2). The apoptosis signaling pathway was predicted to be decreased in the APP/PSEN1 mice (−log(FDR) > 1.3 and z-score < −2). This pathway was driven by an increased expression of antiapoptotic genes (Naip3, Naip5, and Bcl2a1b and d) implicated in inflammasome activation or upregulated during inflammatory processes [48,49].
We next performed GSEA on the list of DEGs in APP/ PSEN1;Tyrobp −/− vs. APP/PSEN1 mice (FDR < 0.05).  Figure 3b). Complement system, phagosome formation and production of NO and ROS were the most significantly dysregulated in the APP/PSEN1 KO for Tyrobp vs. APP/PSEN1 WT for Tyrobp (−log (FDR) = 5.32, 4.58, and 3.45, respectively). Contrary to the previous comparison, prediction of the activation state showed down-regulation of pathways associated with immune response and glial cell activation (e.g., neuroinflammation signaling pathway, Trem1 signaling, dendritic  Table 2).
We analyzed whether these apparent beneficial effects of the absence of TYROBP on the cerebral amyloidosis transcriptome could decrease the number of perturbed canonical pathways observed in the APP/PSEN1 mice vs. WT. Only seven pathways were significantly dysregulated in APP/ PSEN1;Tyrobp −/− vs. WT (−log(FDR) > 1.3), five of them being related to cholesterol metabolism (superpathway of cholesterol biosynthesis, cholesterol biosynthesis I-III, and epoxysqualene biosynthesis) and two of the dysregulated pathways were involved in inositol biosynthesis and metabolism (D-myo-inositol (1,4,5)-trisphosphate biosynthesis, superpathway of inositol phosphate compounds) (Fig. 2a, Suppl. Table 2). Importantly, all canonical pathways activated in the brains of APP/PSEN1 mice vs. WT (z-score > 2) were normalized in APP/PSEN1;Tyrobp −/− vs. WT (Fig. 2c). GSEA in the APP/PSEN1;Tyrobp −/− mice vs. Tyrobp −/− mice reported nine dysregulated pathways. However, considering the limited number of DEGs in this comparison (12 DEGs), the importance of the dysregulations is unclear. No canonical pathways were significantly dysregulated in the Tyrobp −/− vs. WT mice.
These results demonstrate that the absence of TYROBP reduces the perturbations of the transcriptome observed in the brains of APP/PSEN1 mice and alleviates the induction of the immune/pro-inflammatory response observed in the APP/PSEN1 mouse brain.

Absence of TYROBP decreased the induction of C1q expression in APP/PSEN1
The complement system plays an important role in the pathogenesis of AD [26,27,50]. In our transcriptomic analysis, absence of TYROBP in APP/PSEN1 mice alleviates the increased expression of several central complement components associated with APP and/or PSEN1 mutations. We assessed by qPCR, western blot and immunostaining the expression of C1q, the initiating protein of the classical complement cascade, in male and female APP/PSEN1 mice WT or null for Tyrobp at 8 months of age (Suppl. Figure 4). Consistent with the RNAseq results, mRNA level of C1q was increased in the APP/PSEN1 mice as compare to WT mice and was normalized in APP/ PSEN1;Tyrobp −/− (Suppl. Figure 4a). Western blot and immunostaining confirmed the decreased C1q expression in APP/PSEN1;Tyrobp −/− mice as compared to that of APP/ PSEN1 mice (Suppl. Figure 4b-d).
Brain network signatures in APP/PSEN1 mice are enriched for human sporadic LOAD network signatures We intersected the APP/PSEN1 mouse signature (APP/ PSEN1 vs. WT) with multiple LOAD brain gene expression signatures previously identified in human brains. Suppl. Table 3 summarizes the source of the human LOAD gene signatures and the intersection analysis results obtained from the FET. Genes upregulated in the brains of APP/ PSEN1 mice were significantly enriched for genes upregulated in CA1 of the hippocampus (4.1-fold, adjusted p value 6.1 × 10 −4 ) or CA3 (3.3-fold, adjusted p value 8.1 × 10 −4 ) of LOAD brains as identified by Miller et al. [51]. Meanwhile, genes upregulated in APP/PSEN1 mouse brains were also significantly enriched for genes positively correlated with the Braak score in the PFC region of the Harvard Brain Tissue Resource Center (HBTRC) cohort (2.4-fold, adjusted p value 1.6 × 10 −3 ) [6]. The genes downregulated in the brains of APP/PSEN1 mice were not enriched in human AD signatures, perhaps at least partially attributable to the limited number of downregulated genes identified in the current study.
TYROBP knockout signatures are enriched in TYROBP regulated networks in human LOAD brains TYROBP was identified as a network key driver implicated in LOAD based on the Bayesian causal regulatory network analysis in the HBTRC expression dataset [5]. We overlaid the Tyrobp-null signature detected in APP/PSEN1 mice (APP/PSEN1;Tyrobp −/− vs. APP/PSEN1) onto the HBTRC Bayesian network and found enrichment of downregulated genes in the first four layers downstream of TYROBP (up to 29.4-fold; FET p value 3.5 × 10 −7 -1.2 × 10 −33 ), thus validating the TYROBP-centered regulatory network (Fig. 3).
One key output of the NIH AMP-AD program has been RNA-seq datasets from over 1000 human brain tissues across multiple brain regions, generated from individual LOAD cohorts, including the aforementioned MSBB and ROSMAP. This allowed us to build new Bayesian networks from each individual brain region and combine them with the HBTRC networks to create a more comprehensive causal network, termed union Bayesian network (see Suppl. Text). In this union Bayesian network, we observed a stronger enrichment of downregulated genes by deletion of Tyrobp in the first four layers of the network neighborhood of TYROBP (FET p value 3.6 × 10 −15 , 5.0 × 10 −48 , 2.7 × 10 −51 , and 7.7 × 10 −23 accordingly). For example, in the first-layer neighborhood of TYROBP, there was a 70-fold enrichment and one third of the 27 first-layer neighbor genes were downregulated by the deficit of TYROBP in APP/PSEN1 mice (FET p value 3.6 × 10 −15 ). Figure 3 shows the network topology of the up to 3-layer network neighborhood around TYROBP. The validated network genes included TREM2, a prominent LOAD GWAS gene, and complement component subunits includingC1QA and C1QB. In addition to TREM2, we found 5 other LOAD GWAS hits in this subnetwork including APOE, the MS4A locus, CD33, INPP5D, and the HLA locus. Unlike TREM2, the expression of these GWAS genes was not altered by deletion of Tyrobp. As expected, the upregulated genes in the brains of APP/PSEN1 mice were also significantly enriched in the neighbors of TYROBP (9.9-fold, FET p value 4.3 × 10 −71 ; Suppl. Figure 5). In summary, this causal network analysis shows that the human LOAD causal networks are highly predictive of the Tyrobp-null murine signature.
We further intersected the present mouse gene signatures with co-expression network modules identified across multiple human LOAD datasets. As summarized in Suppl. Table 4, the genes upregulated in APP/PSEN1 mice were significantly enriched in human immune response modules. Meanwhile, Tyrobp −/− signatures in APP/PSEN1 mice were also enriched in the same human immune response modules. Many of these immune response modules were highly associated with AD pathology, including the top module "yellow" in the PFC co-expression network from the HBTRC data in which TYROBP was a module member [5]. The result suggests that Tyrobp modulates the immune response pathways which are activated by transgenic expression of APP/PSEN1.

Constitutive deletion of Tyrobp does not alter amyloid burden in 8-month-old APP/PSEN1 mice
We assessed whether the absence of TYROBP would modulate Aß deposition in the PFC and hippocampus (HC) of male and female APP/PSEN1 mice. The area covered by 6E10-immunoreactive plaques was unchanged by the deletion of Tyrobp in APP/PSEN1 mice (Fig. 4a-c). We and others previously reported decreased microglial clustering around the Aß plaques in Tyrobp −/− mice with Aß pathology [28,52]. We observed this same phenomenon in the 8month-old APP/PSEN1;Tyrobp −/− as compared to APP/ PSEN1 mice with normal levels of TYROBP (Fig. 4a, b). We then asked whether reduced clustering of microglia around plaques in the APP/PSEN1 mice null for Tyrobp was due to a reduction in the total number of brain microglia, as observed in Trem2 −/− mice with Aß pathology [52,53]. Iba1 immunostaining in 8-month-old APP/PSEN1 mice null for Tyrobp did not show differences in the total number of microglia in cortex or hippocampus as compared to APP/ PSEN1 mice with normal levels of TYROBP (Fig. 4d).
Constitutive absence of TYROBP is associated with elevated levels of soluble Aß but reduced levels of insoluble Aß in 8-month-old APP/PSEN1 mice We assessed whether the absence of TYROBP would modulate levels of Aß species in APP/PSEN1 mice. We measured levels of Aß40 and Aß42 in TBS, Triton-X, and formic-acid soluble Aß fractions from brains of 8-monthold male and female is Type="Italic">APP/PSEN1 mice WT or null for Tyrobp (Suppl. Figure 6a). In both males and females, absence of TYROBP was associated with increased levels of Aß40 and Aß42 in the most soluble fraction (TBS) (Suppl. Figure 6b,c). While not statistically significant, a trend toward a decreased Aß42/40 ratio was observed in the APP/PSEN1;Tyrobp −/− mice as compared to APP/PSEN1 mice. This effect was more obvious in females than in males (p value = 0.0635 and p value = 0.2587, respectively) (Suppl. Figure 6d). Conversely, levels of both Aß40 and Aß42 were reduced in the Triton-X fraction as compared to APP/PSEN1 mice with normal levels of TYROBP (Suppl. Figure 6e, f) without any change in the Aß42/40 ratio (Suppl. Figure 6g). The levels of Aß40 and Aß42 did not change in the formic-acid fraction (Suppl. Figure 6h,i), but Aß40 levels trended lower in APP/ PSEN1;Tyrobp −/− mice as compared to sex-matched APP/ PSEN1 mice leading to an increased Aß42/40 ratio in the APP/PSEN1;Tyrobp −/− mice. Notably and in accordance with our previous results [28,54], female APP/PSEN1 mice WT or null for Tyrobp had higher levels of Aß40 and 42 in the formic-acid fractions when compared to genotypematched males. We next assayed oligomeric Aß peptides using antibodies NU-4, A11, and OC antibodies to distinguish among Aß conformers (Suppl. Figure 6k-m), which have been correlated with impaired cognitive performances in humans and mice [55]. At 8 months of age, the absence of TYROBP played no obvious role in determining levels of NU-4-, A11-, or OC-epitope-containing oligomeric Aß.
Deletion of Tyrobp ameliorates aberrant synaptic plasticity in APP/PSEN1 mouse hippocampus Synaptic plasticity is abnormal in the hippocampus of APP/ PSEN1 mice [56], and in 4-month-old mice, we found that Tyrobp deletion partially rescued the normal synaptic phenotype [28]. In hippocampal slices from 8-month-old mice, basal transmission at CA3-CA1 synapses was significantly enhanced in Tyrobp −/− and APP/PSEN1;Tyrobp −/− mice as compared to WT controls, while APP/PSEN1 mice did not differ from WT controls (Fig. 5b). We examined a shortterm form of presynaptic plasticity, PPF, which is inversely related to glutamate release probability [57]. At 4 months, PPF was depressed in APP/PSEN1 mice and was fully rescued by deletion of Tyrobp [28]. At 8 months, PPF was depressed in APP/PSEN1 mice at an interpulse interval of 20 ms and was normalized by deletion of Tyrobp (Fig. 5c). Absence of TYROBP in WT background did not affect the PFF. We also studied two long-term forms of synaptic plasticity: synaptically induced LTP, and DHPG-induced LTD. LTP was similar in the four genotypes [WT, Tyrobp −/− , APP/PSEN1, and APP/PSEN1;Tyrobp −/− ] (Fig. 5d).
However, LTD was markedly reduced in slices from APP/ PSEN1 mice, and this deficit was fully rescued by deletion of Tyrobp (Fig. 5e). Of note, absence of TYROBP in WT background did not affect LTD. These results establish that as late as 8 months of age, deletion of mouse Tyrobp prevents the development of electrophysiological abnormalities that are present in a model of cerebral amyloidosis.
Constitutive deletion of Tyrobp prevents impaired performance in Barnes maze learning and memory in APP/PSEN1 mice We previously concluded that deletion of Tyrobp can slow or delay the progression of learning deficits in APP/PSEN1 transgenic mice during the early stages of amyloid deposition [28]. In the current study, we probed the effect of the absence of TYROBP on spatial learning and memory using the Barnes Maze Test in 8-month-old WT, Tyrobp −/− , APP/ PSEN1, and APP/PSEN1;Tyrobp −/− mice (Fig. 5f). The percentage of distance traveled in the target quadrant (TQ) and the latency to find the hidden zone were identical between Tyrobp −/− mice and WT mice suggesting that deficiency of TYROBP did not have a deleterious effect on learning and memory in a WT background. APP/PSEN1 mice showed impairment of learning and memory relative to WT mice with a shorter distance traveled in the TQ and increased latency to find the hidden zone as compare to WT mice (Fig. 5g, h). Absence of TYROBP in APP/PSEN1 mice was associated with improved learning and memory relative to APP/PSEN1 with WT TYROBP levels. Thus, percentage of distance traveled in the TQ and latency to find the hidden zone were similar in APP/PSEN1;Tyrobp −/− and WT littermates. These data are consistent with a beneficial effect of Tyrobp deletion on learning and memory deficits associated with APP/PSEN1 mutations.

Discussion
Despite decades of intensive research, the pathogenesis of LOAD remains elusive, and therapeutic interventions focused on reduction of Aß accumulation have failed to provide meaningful clinical benefits. Multiscale network modeling of large-scale functional genomics data creates novel opportunities to identify key targets causally linked to LOAD, thereby offering new insights for drug-discovery programs. We previously constructed a molecular network based on whole-genome gene-expression profiling and genotyping data on 1647 autopsied brain tissues from LOAD and nondemented subjects and reported the first LOAD causal network centered on TYROBP [5]. In the computational model, TYROBP lies at the center of the complement system subnetwork, and, based on its predicted role as a hub and driver of the complement subnetwork, we predicted that changes in the level of TYROBP would lead to similar changes in the expression level of genes in that network.
Herein, we report the first biological validation of the TYROBP-centered LOAD network in the amyloidogenic APP/PSEN1 mouse model. Our data indicate that constitutive deletion of Tyrobp prevents behavioral and electrophysiological deficits in 8-month-old Aß-amyloiddepositing mouse. These results are associated with a recapitulation of the expected complement system subnetwork in APP/PSEN1 mice, followed by its reversal in the absence of TYROBP, predicted from the multiscale gene networks of human postmortem LOAD brain. Tyrobp deletion in the APP/PSEN1 mice repressed the induction of inflammatory cytokines and numerous other genes that are upregulated in the switch from homeostatic microglia to disease-associated microglia (DAM) [45], including Trem2, Clec7a, Cst7, Itgax, and Ctsd. DAM are found in AD at the proximity of the Aß plaques and express genes associated with lysosomal-phagocytic, lipid metabolism, and AD risk factors including Apoe and Trem2. While we did not detect differences in ApoE expression, the decreased expression of numerous DAM genes highlights the role of TYROBP as an important early activator of DAM. Moreover, the absence of APP/ PSEN1 associated behavioral and electrophysiological impairments in the absence of TYROBP suggests that at least in this mouse model, decreasing DAM activation might play a beneficial role in the progression of AD.
The complement system is robustly activated in our dataset. This system plays a crucial role in physiologic synaptic pruning [50], is activated in human and mouse AD brains [27,[58][59][60] and is associated with early synapse loss in AD [26,27]. Amyloidogenic APP/PSEN1 mouse models with reduced components of the complement system, e.g., C3, C1q, or CR3, showed improved performance on learning and memory tasks despite having increased Aß deposition [26]. Similarly, herein and in our previous report [28], we show that deletion of Tyrobp in the APP/PSEN1 mouse model did not play a prominent role in Aß deposition but reversed the behavioral and electrophysiological alterations associated with the APP/PSEN1 mutations. We also observed that deletion of Tyrobp in a tauopathy mouse model reduced the expression of C1q and improved learning behavior and synaptic function despite a paradoxical increased in the spread and the phosphorylation state of tau [61]. The regulation of the complement subnetwork by TYROBP, including a decrease in several of the same components previously reported further supports a beneficial effect of decreased complement pathway activation in AD pathology. Notably, APP/PSEN1 mice with deletion of Progranulin (Grn) demonstrate overexpression of some TYROBP network genes, including C1q [62], and while they have reduced diffuse Aß and axonal dystrophy with partially improved behavioral deficits, they also manifest increased neuronal injury. We report that deletion of Tyrobp in the APP/PSEN1 model also reduces Grn expression level in APP/PSEN1 alone.
This study represents the first in vivo validation of a hub/ driver gene and subnetwork predicted by integrative network analyses of human postmortem sporadic LOAD brain and suggests that reduction of TYROBP level or inhibition of its activity could represent a potential therapeutic target for AD prevention and/or treatment. Moreover, this study validates the multiscale network for identifying gene subnetworks and key drivers in complex diseases and for providing new insights for drug discovery programs.

Compliance with ethical standards
Conflict of interest The authors declare that they have no conflict of interest.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons. org/licenses/by/4.0/.