Proteome ontology analysis

In our analysis, we identified 4,849 proteins in the brain tissue across the three ApoE-isoform groups (n = 47). From these, we determined protein abundance and turnover rate FCs for comparisons: ApoE2 vs. ApoE3 (E2vsE3) and ApoE4 vs. ApoE3 (E4vsE3). Here, ApoE3 serves as the reference ’normal’ control. We quantified 3,532 abundance FCs for both the E2vsE3 and E4vsE3 comparisons (S2 and S3 Tables). With a smaller number of turnover rate FCs 1,430 for E2vsE3 and 1,405 FCs for E4vsE3 because of the more rigorous statistical filtering criteria.

We focused on ontologies from six databases: GO Function, GO Component, GO Process, WikiPathways, Reactome, and KEGG. We calculated the average abundance FC (Δabundance) and turnover rate FC (Δturnover) for the proteins observed in each ontology (Please refer to the Ontology-level Calculations in the methods section). This yielded ~2700 ontology-level comparisons of average Δabundance and Δturnover calculations for both E2vsE3 and E4vsE3 (S4 Table). The interpretation (Fig 1C) relies on the traditional understanding of protein turnover, contextualizing changes in protein expression. It offers a lens to assess the variances in the steady states of ApoE genotypes [19]. Using a one-sample t-test, we discerned which ontologies deviated significantly from a median Δabundance of 0. In the E2vsE3 comparison, we identified 284 protein ontologies with notable Δabundance (BH-PV < 0.05) (Fig 1D). For the E4vsE3 comparison, 287 protein ontologies had significant Δabundance (BH-PV < 0.05) (Fig 1E).

The box plots in Figs 2–5 summarize the Δabundance and Δturnover for ontologies with significant Δabundance shifts. To maximize visibility and to accommodate for space limitations, these boxplots do not contain outlier points but Figs C-F in S1 File contain outlier points. Given that some ontologies are repetitive, proteins depicted in the box plots might appear in multiple ontologies with analogous names/functions. When faced with such redundancies, we typically chose the ontology with superior coverage (Observed/Total) for representation. As a convention, each ontology is presented in an "ontology name (n)" format, where (n) indicates the count of quantified proteins within that ontology. Overlap between similar ontologies is shown in the heatmap.

Download:

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

Fig 2. ApoE4 expression disrupts endosomal maturation and ApoE2 increases lysosomal capacity.

A-B. Bar plot displaying Δabundance (orange) and Δturnover (blue) of proteins detected in all experimental cohorts for significant* ontologies related to endolysosomal trafficking in the E2vsE3 comparison (A) and in the E4vsE3 comparison (B). C. Heatmap displaying % of proteins shared across the endolysosomal ontologies with significant* Δabundance. (*BH-PV < 0.05). Created with Matplotlib.py and InkScape.

https://doi.org/10.1371/journal.pcbi.1012407.g002

Download:

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

Fig 3. ApoE genotype differentially regulates mitochondrial proteostasis.

A-B. Bar plot displaying Δabundance (orange) and Δturnover (blue) for ontologies of proteins detected in all experimental cohorts related to mitochondrial components with significant* Δabundance in the E2vsE3 comparison (A) and in the E4vsE3 comparison (B). C. Heatmap displaying % of proteins shared across the mitochondrial ontologies with significant* Δabundance. (*BH-PV < 0.05). Created with Matplotlib.py and InkScape.

https://doi.org/10.1371/journal.pcbi.1012407.g003

Download:

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

Fig 4.

ApoE2 and ApoE4 expression drive changes in cellular fuel selection A-B. Bar plot displaying Δabundance (orange) and Δturnover (blue) for ontologies of proteins detected in all experimental cohorts belonging to oxidative phosphorylation with significant* Δabundance in the E2vsE3 comparison (A) and in the E4vsE3 comparison (B). C. Heatmap displaying % of proteins shared across the oxidative phosphorylation ontologies with significant* Δabundance. (*BH-PV < 0.05). Created with Matplotlib.py and InkScape

https://doi.org/10.1371/journal.pcbi.1012407.g004

Download:

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

Fig 5. Proteasomal activity decreases with ApoE4 expression.

A-B. Bar plot displaying Δabundance (orange) and Δturnover (blue) of proteins detected in all experimental cohorts for several ontologies related to proteasomal activity with significant* Δabundance in the E2vsE3 comparison (A) and in the E4vsE3 comparison (B). C. Heatmap displaying % of proteins shared across the ontologies with significant* Δabundance. (*BH-PV < 0.05). Created with Matplotlib.py and InkScape

https://doi.org/10.1371/journal.pcbi.1012407.g005

ApoE Isoforms Modulate Synthesis and Degradation of Endocytic Vesicle Components

We observed that multiple ontologies with significant Δabundance were associated with endocytosis and vesicular processing (Fig 2). Specifically, the general Endocytosis (158) ontology demonstrated increased Δabundance and decreased Δturnover, suggesting reduced degradation in both ApoE2 and ApoE4 compared to ApoE3. In the context of ApoE2, Clathrin-mediated endocytosis (82), Clathrin binding (35), and Clathrin coat (26) mirrored the same ↓degradation effect observed in endocytosis, while SNARE complex (32) showed diminished Δabundance and Δturnover, suggesting a decline in protein synthesis compared to ApoE3. Moreover, ApoE2 expression led to significant alterations in several regulatory ontologies tied to endocytosis and vesicular processes, such as: Endocytic recycling (34) (↑synthesis), Early endosome (↑synthesis), and Regulation of endocytosis (15) (↑synthesis). In ApoE2, proteins related to Lysosome Vesicle Biogenesis (18) have lower degradation while Regulation of Endocytosis (107) had increased synthesis leading to higher abundance of these protein groups and presumably more efficient endolysosomal function.

In both ApoE2 (E2vsE3) and ApoE4 (E4vsE3) we noted less degradation of general lysosome (146) proteins. Within this general ontology, the lysosomal membrane (70) ontology had diminished Δabundance and Δturnover only in the ApoE4 group, suggesting less synthesis of the membrane components compared to ApoE3. This is consistent with large lysosomal vesicles stored in ApoE4 cells [13,20]. In ApoE4 mice there was higher Δabundance and Δturnover (↑synthesis) of Phosphatidylinositol binding (90) relative to ApoE3. Conversely, there was a decline in both Δabundance and Δturnover (↓synthesis) of SNARE interactions in vesicular transport (19), Synaptobrevin 2-SNAP-25-syntaxin-1a complex (5), and SNARE complex (32).

ApoE isoforms modulate synthesis and degradation of mitochondrial components

Our analysis identified significant Δabundance (BH-PV < 0.05) changes for multiple ontologies related to mitochondrial components (Fig 3). In the E4vsE3 comparison, these ontologies included mitochondrial membranes, protein transport, and morphology (Figs 3A and C in S1 File). Each of these ontologies displayed a negative Δabundance coupled with a positive Δturnover, signifying ↑degradation. We also detected ↓synthesis of mitochondrial calcium ion transmembrane transport (12) and mitophagy (18). In contrast, ApoE4 mitochondrial matrix (159) also had ↑degradation. (Figs 3B and C in S1 File) The percentage of overlapping proteins in each mitochondrial component ontology is displayed in Fig 3C. The key finding from these ontologies is that within the ApoE2 mice there is a coherent increase in the degradation of all mitochondrial components consistent with an increase in mitochondrial degradation as an entire unit. In contrast, the ApoE4 tissues show discordant changes in matrix versus membrane proteins suggesting that mitochondrial maintenance is more piecemeal and that mitophagy may be less efficient as previously suggested in the literature [21]. Both the ApoE2 and the ApoE4 results are synergistic with the changes in lysosome dynamics discussed above. In ApoE2 more efficient lysosomal processing will facilitate mitophagy based quality control while the inhibited lysosomal processing would inhibit mitophagy and make the ApoE4 more reliant upon individual protein replacement strategies.

ApoE4 Disrupts metabolic pathway control

We observed significant changes in Δabundance (BH-PV < 0.05) across multiple ontologies related to energy production (Fig 4). ApoE2 resulted in lower expression of levels of Pyruvate metabolism (32), Citrate cycle (TCA cycle) (26), and Glycolysis/Gluconeogenesis (41). These reductions were primarily attributed to decreased synthesis (↓synthesis), a trend that was also evident in the Oxidative stress and redox pathway (48) proteins which protect the cell from reduced oxygen species. Notably, Fatty acid beta-oxidation (25) demonstrated reduced Δabundance coupled with increased Δturnover (↑degradation) in ApoE2 (E2vsE3), suggesting a potential decrease in fatty acid catabolism and an increase in the use of fatty acids for building complex lipids. In contrast, in the ApoE4 mice, major energy production pathways such as Fructose and mannose metabolism (21), Pyruvate metabolism (32), and Glycogen metabolism (22) all exhibited increased Δabundance and Δturnover, pointing towards enhanced synthesis of enzymes involved in carbohydrate metabolism in the E4vsE3 comparison and an increased reliance on carbohydrates for energy similar to previous observations [17,22].

ApoE Isoforms and Ubiquitin-Proteasome pathway activity

Proteasome-related ontologies exhibited significant changes in regulation due to ApoE isoforms (Fig 5). For both the E2vsE3 and E4vsE3 comparisons, we identified pronounced increases in Δabundance and reductions in Δturnover (↓degradation) associated with the proteasome complex (47) Furthermore, an increased Δabundance and Δturnover (↑synthesis) of proteins involved in the Regulation of ubiquitin-dependent protein catabolic process (62) was statistically significant in both comparisons (BH-PV 0.05).

The proteasome regulatory particle, base subcomplex (11) displayed ↑synthesis in the E2vsE3 comparison and ↓degradation in the E4vsE3 comparison. Meanwhile, proteins within the Proteasome regulatory particle, lid subcomplex (6), demonstrated significant Δabundance due to ↓degradation in the E4vsE3 comparison with ApoE4 expression. However, these changes were not significant in the E2vsE3 comparison. Additionally, in ApoE4 we noted ↑synthesis in the Negative regulation of proteasomal ubiquitin-dependent protein catabolic process (21), ↓degradation Deubiquitination (79), and ↓synthesis Hsp70 protein binding (23). These observations suggest a nuanced regulation of the ubiquitin-proteasome system (UPS) in association with ApoE isoforms. Hsp70 proteins are often deemed pivotal regulators of proteasome activity [23]. These changes suggest a significant reduction in the proteasome-dependent protein quality control for ApoE4 tissue (Fig 5C).

Quantifying ApoE-dependent shifts in liver proteostasis

The liver is the largest producer of ApoE in the body and is also a major receptor of ApoE and its associated cargo [8,18,24]. Therefore, we tested whether the liver tissue from these same experimental mice would show matching ApoE allele-specific shifts in proteome regulation.

In contrast to the brain, ApoE2 liver there was no significant change in any of the endolysosomal processes relative to ApoE3 (S5 Table). Multiple mitochondrial ontologies in the liver changed in significant ways and nearly 60% of their proteostasis changes are equivalent to the brain. Most changes in the mitochondria in the liver with ApoE2 expression involve increased degradation of mitochondrial components, though there is some reduced synthesis for the mitochondrial envelope and transmembrane transport. Liver ApoE2 also did not shift proteasome-related ontologies in the same manner as the brain. In fact, where the brain contains decreased degradation of proteasomal components, the liver increases their synthesis and contains more proteasomal capacity similar to literature reports. [23,25,26]

In ApoE4/E3 liver comparisons there was not a significant change in any of the endocytic processes (S5 Table) with the exception of endosomal protein localization. Both tissues had increased concentration, through increased synthesis in the brain, while the liver decreased degradation. For ApoE4 mitochondrial components 80% of the significant mitochondrial liver ontologies had the same proteostasis changes in the brain (S5 Table). The 20% differences were due to certain NADH and ATP synthesis electron transport chain ontologies that were increased synthesis in the brain and increased degradation in the liver. Similar to the liver comparison of ApoE2 with the proteasome, in ApoE4 liver data there were no shared proteasome changes with the brain (Figs G and H in S1 File). These data suggest that most of the ApoE effects observed in the brain are not global.

Exploring ApoE-genotype Effects Through the Lens of Proteostasis

Compared to the neutral ApoE3 allele, expression of ApoE4 significantly increases the risk for Alzheimers, while the expression of ApoE2 is moderately protective [5,8,12,22,25,27–29]. Here we investigate the etiology of these differences by measuring ApoE-dependent changes in protein homeostasis (proteostasis) in the tissues of human-ApoE transgenic mice (Fig 1A). Proteostasis is the dynamic control of concentration and quality [19,30–32] in the cell. According to our model (Figs 1C and B in S1 File), protein concentration is controlled by the competition between synthesis and degradation while protein turnover rate, or the rate that proteins are replaced, is the average of synthesis and degradation. We present this as simplified equations that relate to synthesis and degradation (see ‘Proteostasis Model and Analysis Rational’ section for more detail). Therefore, changes in protein concentration (Δabundance, Eq 3) paired with changes in protein turnover rate (Δturnover, Eq 10) can highlight the regulatory balance of synthesis versus degradation between conditions (Fig 1B and 1C).

For instance, to increase abundance in the experimental condition (resulting in a positive Δabundance), cells can either synthesize faster or degrade slower. Alternatively, to decrease abundance (leading to a negative Δabundance), cells might reduce synthesis or increase degradation. Using the change in protein turnover rates (Proteostasis Model section, Eq 10) we can deduce whether changes in synthesis or degradation led to changes in abundance (Fig 1B). Therefore, a positive Δabundance indicates increased synthesis when Δturnover is positive or reduced degradation when Δturnover is negative. Conversely, a negative Δabundance signals increased degradation when Δturnover is positive or decreased synthesis when Δturnover is negative (Fig 1C). Applying this logic, we observed a cohesive pattern of ApoE-dependent changes in proteostasis regulation (Fig 6). Below, we discuss how our results unify a diverse set of literature observations for ApoE-dependent modifications of endosome trafficking, as well as lysosomal, mitochondrial, and proteasomal function.

Download:

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

Fig 6. Model comparing the observed changes in proteostasis for ApoE2 and ApoE4.

The arrows are color coded to represent the different pathways impacted in both ApoE2 and ApoE4 when compared to ApoE3. Created with Inkscape and Biorender.

https://doi.org/10.1371/journal.pcbi.1012407.g006

ApoE isoforms modify Endocytic/Endosomal trafficking

Previous research has highlighted the dysregulation of endocytic pathways associated with ApoE4 expression [13,20,27,33–35]. We detected notable ApoE4-dependent changes in several ontologies related to endocytosis (Figs 2 and C in S1 File). This is in line with what is known about how ApoE isoforms modify affinity for cell surface receptors, such as LDLR and APOER2 [3,4], initiating the endocytosis of ApoE along with its content. After this endocytic event, ApoE-laden endosomes undergo various maturation stages, wherein contents are earmarked for recycling, delivery, or degradation.

ApoE4 has a higher affinity binding to receptors [3,4] and induces a trafficking anomaly in the early endosome [33] leading to accumulation and enlargement of lysoendosomal compartments [13,20]. Following an endocytic event, the clathrin coat dismantles, allowing vesicles to fuse into various destinations for cargo release. This fusion mechanism leans heavily on SNARE and SNAP-receptor proteins, also pivotal for exocytosis. In our study, we observed diminished synthesis of SNARE and SNAP ontologies in ApoE4 mice which may disrupt vesicle fusion between organelles [36] and in response to exocytic sequences [37]. Our study revealed a reduced degradation of proteins associated with Clathrin-mediated endocytosis (82), increased synthesis of PICALM (90), and reduced synthesis of the lysosomal membrane (70) (Figs 2B and 2C in S1 File) in the presence of ApoE4. In the literature, ApoE4 expression induced a decline in clathrin-mediated endocytosis in astrocytes [38], dysregulation of early endosome populations in 18- and 25-month-old ApoE4 mice [34] and phosphatidylinositol binding proteins like PICALM [39]. Before undergoing lysosomal degradation, endosomes transition to the late endosomal phase. Our findings suggest that ApoE4 expression reduces the synthesis of both late endosomal and lysosomal membranes. Although the general lysosome (146) ontology exhibits increased degradation with ApoE4 expression, if we look specifically at the membrane components of this ontology then the ApoE4 specifically has less total protein due to lower synthesis. These results are consistent with previous observations of large-volume lysosomes which would have a low membrane surface/volume ratio accumulating in the in ApoE4 cells [13,20]. Collectively, these results underscore multiple points of failure in the ApoE4-associated inhibition of endosomal maturation and stalled lysosomal functions as previously observed [27,33,34].

The E2vsE3 tissue had similarities in vesicle-centric ontologies. Notably, there was a decline in the degradation of endocytosis and clathrin protein-related ontologies, and SNARE complexes saw reduced synthesis. This implies that ApoE2 also modifies vesicle endocytosis. However, the changes suggested a more streamlined regulation of endolysosomal events with ApoE2 (E2vsE3). This again agrees with literature reports of modified receptor binding with ApoE2 having lower affinity while ApoE4 has a higher affinity [3,4]. This coupled with lower degradation of proteins within the lysosome vesicle biogenesis ontology for the E2vsE3 comparison, a process intrinsically tied to endosomal trafficking and central to lysosomal adaptation [40], suggests a tighter control of endocytic events and better lysosomal quality with ApoE2 expression. These observations agree with previous research on astrocytes indicating ApoE2 expression increases lysosomal activity relative to ApoE3 and ApoE4 expression [41].

ApoE-dependent changes in Mitochondrial Proteostasis

We observed ApoE-dependent changes in mitochondrial proteostasis that were consistent with the modified autophagy and lysosomal function discussed above. In ApoE4 (E4vsE3) mice, we measured elevated degradation in mitochondrial membrane (406), mitochondrial inner membrane (303) (Figs 3 and D in S1 File), Cristae formation (11), Mitochondrial fusion (13), and mitochondrial transport (72) with no accompanying change in mitochondrial matrix (159) and decreased synthesis of mitophagy (18). Mitochondrial membrane complexes play critical functions in cellular homeostasis—such as energy production, calcium level modulation, apoptosis, and the regulation of reactive oxygen species (ROS) [42]. Prior research has documented dysfunction in the mitochondrial membrane’s integrity as a promoter of Alzheimer’s disease [43,44].

Most mitochondrial proteins are encoded on the nuclear DNA and are transported into the mitochondria through translocases (TIM and TOM) [45,46]. These translocases interact with the many inner mitochondrial membrane (IMM) folds that make up the cristae via the mitochondrial cristae organizing system (MICOS).[47] Our observations indicate a change in mitochondrial protein import, especially evident in the higher degradation of the TOM complex (10), cristae formation (11), and the MICOS complex (7) ontologies. (See Figs 5B and 5E in S1 File). The MICOS also plays a vital role in cristae organization and the function of respiratory complexes.[48] Disruptions in MICOS have been documented to modify cristae structure[49], and recent studies associate altered MICOS protein expressions with ApoE4 manifestation[44]. The MICOS literature also report evidence of mitochondrial fusion and fission imbalance in Neuro-2a cells expressing ApoE4.[44] Previous analysis of AD brains indicated diminished protein levels connected to mitochondrial fusion/fission[50], which our data supports as a degradation driven loss of fusion proteins (See Figs DA and DB in S1 File).

Mitochondria and the endoplasmic reticulum (ER) collectively form the mitochondria-associated membrane (MAM), which has implications in AD pathology [51,52]. These MAMs regulate oxidative phosphorylation, calcium levels, protein degradation, and mitochondrial membrane organization. Our dataset elucidates an ApoE4-induced MAM (57), marked by increased degradation contrasted against ApoE3 (Figs 3 and D in S1 File). Our results support ApoE4-related MAM instability by diminished synthesis of chaperone complexes, mitophagy, and calcium transport.

These results suggest that ApoE4-tissue relied on replacement of individual components within the mitochondria to maintain homeostasis. In contrast, ApoE2 mice increased degradation of mitochondrial membrane ontologies with a matching increase in the degradation of the matrix proteins. Although both ApoE2 and ApoE4 mice revealed changes in the mitochondrial membrane and transport, our ApoE2 findings suggests that there is a cohesive organelle-wide response involving both membrane and matrix proteins. Combined with the endolysosomal systems results, we postulate this might be evidence of superior mitophagy in ApoE2. Additionally, we theorize, as described in the literature [17,21,22,44,53], that the disparate attempts at protein replacement in ApoE4 mitochondria culminate in dysfunction.

ApoE disrupts ATP production

There is an increasing body of research on ApoE genotype-specific effects in ATP production [54–56] which links compromised bioenergetic pathways to neurodegeneration [57–60]. Several studies highlight an ApoE4-related shift towards glycolysis and diminished oxygen consumption in brain tissues. [22,57,61] Our data align with these observations, with increased synthesis of carbohydrate metabolism ontologies in ApoE4 (see Fig E in S1 File). These ontologies include Fructose and mannose metabolism (21), Pyruvate metabolism (32), and Glycogen metabolism (22). We posit that this increased reliance on carbohydrate metabolism is a consequence of the lack of cohesive mitochondrial maintenance.

While ApoE4 expression is associated with diminished ATP production [54,62–64], ApoE2 expression has been shown to enhance ATP production and glycolytic activity [57,61]. In ApoE2 tissue, we observed decreased abundance in ontologies such as the TCA cycle (25), Pyruvate metabolism (32), and Glycolysis / Gluconeogenesis (41) due to diminished synthesis and augmented degradation. These regulatory changes are different than expected based on the literature. This might be because our study averages together all cell types and regions of the brain and therefore may diverge from cell type-specific experiments. Collectively, our data accentuates the isoform-specific alterations in diverse metabolic pathways and suggests that isolating single cell types or regions from the brain may be important to test for metabolic changes in response to ApoE isoforms.

Linkages between the Proteasomal and Mitochondrial Homeostasis

The proteasome also plays a major role in the mitochondrial quality control, by degrading excess or misfolded proteins that disrupt mitochondrial activity [65], and saving cells from misfolding stress response. [66] Reduced proteasome activity has consistently been implicated as a major player in the pathophysiology of neurodegeneration [21,25,26,67,68]. ApoE2 has been shown to directly inhibit the cleavage of amyloid precursor protein (APP) to decrease Aβ peptide formation. [69]while ApoE4 may promote its accumulation. [10,13,70–72] This buildup of Aβ has been shown to directly inhibit proteasome function leading to increased accumulation of amyloid plaques. [73,74]

Proteasome function can also be disrupted by mitochondrial dysfunction. For example, oxidation of the 26S subunit of the proteasome due to increased mitochondrial oxidative stress has been shown to increase the accumulation of ubiquitinated substrates and decrease proteasomal activity [75]. Notably, we discerned a significant reduction in the synthesis of HSP70 proteins in E4 which latch onto misfolded or compromised proteins before proteasomal degradation [23]. This interconnection of the proteasome and mitochondria as well as their consistent implication in neurodegenerative disease has led some researchers to suggest that dysfunction in either the proteasome or mitochondria are “two sides of the same coin” leading to a futile cycle of mitochondrial and proteasomal insult. [76]

ApoE2 also reduced the degradation of the proteasome, paralleling the ApoE4 response (see Figs 5 and F in S1 File). While both ApoE2 and ApoE4 increase synthesis of ‘ubiquitin-dependent catabolism’ ontologies, the regulatory complex of the proteasome (responsible for facilitating the unfolding and admission of ubiquitinated polypeptides into the proteasome’s degradation chamber [77]) the ‘negative regulation of catabolism’ had increased synthesis in ApoE4. This implies larger proteasome pool in ApoE4, albeit with more regulation.

Our data on proteasomal and lysosomal regulation in ApoE4 are in agreement with existing literature [33,78], suggesting that suboptimal endocytic regulation might directly impact autophagy and the proteasomal maintenance of mitochondrial proteostasis. Our findings also substantiate the ApoE4-associated dysregulation of MAM structures [22,51,52], which potentially results in disrupted mitochondrial morphology and impaired energy production. Our working hypothesis postulates that ApoE4 expression reduces autophagy and proteasomal activity, diminishing capacity to eliminate dysfunctional mitochondria.

Liver Proteostasis Changes Compared to Brain

ApoE is an important lipid transporter to many parts of the body, not just the brain. Although ApoE2 protects against Alzheimer’s Disease, it increases risk for cardiovascular disease [18]. This leads us to question if the mechanism by which ApoE2 protects against Alzheimer’s, may be detrimental and disease causing to other tissues. To understand if ApoE elicits a global response across all tissue, we tested and analyzed the liver tissue from our experimental mice in the same manner as the brain.

The changes in ApoE2/E3 liver and brain proteostasis were not equivalent. Although many trends were similar between brain and liver, few changes in the liver were statistically significant. Mitochondrial protein localization, transportation, and organization had similar proteostasis trends between tissues, that suggest mitochondrial turnover to be reduced in the liver compared to the brain (S3 Table compared to S5 Table).

Similarly, ApoE4/E3 liver and ApoE4/E3 brain proteostasis changes in the endosome, metabolism, and proteasome pathways in the brain are not similar between tissues (see S4 and S5 Tables). However, significant effects of ApoE4 regarding the mitochondria between the liver and brain contained several of the same proteostasis changes, suggesting there may be some degree of shared effects due to ApoE4.

Thus, while some cellular pathways may be affected similarly, a global effect specific to ApoE allele is not observed in our data. We propose the lack of a global ApoE effect (Figs G and H in S1 File) is because most tissues have a large number of apolipoproteins (~16) participating in lipid transport [8,18]. The brain is limited to the apolipoproteins that it creates (primarily ApoE and ApoJ), and ApoA1 which can cross the blood-brain-barrier [16]. Since lipid trafficking in the liver involves many more apoliproproteins, this may dilute the ApoE effect. Future experiments investigating peripheral tissues that only receive ApoE from the liver could help distinguish if the ApoE-dependent changes in the brain are due to endocytosis and internalization of the ApoE-specific cargo.

Ethics statement

All animal handling experiments were reviewed and authorized by the Brigham Young University Institutional Animal Care and Use Committee (IACUC protocol #191102).

Experimental design and statistical rationale

Proteostasis model and analysis rational.

A protein homeostasis model must account for common sources and sinks of protein mass (Fig B in S1 File). In this model, we assume there is a large circulating pool of free amino acids affected by diet, metabolites, and waste expulsion. Amino acids become the precursors for protein synthesis in an initial tRNA charging step, which then polymerize in an mRNA-dependent step before folding into functional proteins. Multiple competing processes subsequently influence the resulting protein concentration. First, degradation returns the protein to the constituent amino acids in the free pool. The functional proteins may also transition (reversibly) into an aggregate/condensate state that undergoes a separate degradation process. Finally, protein concentration may be affected by importing or exporting proteins.

Our goal is to use chemical kinetics and translate this diagram into a mathematical model with a few tunable parameters identifiable from experimental data. Unfortunately, a complete mathematical translation of this system leads to a model with too many parameters to draw meaningful conclusions. We, therefore, make several simplifying assumptions about which processes are dominant to restrict the parameters to an identifiable subset.

First, the mice in this study are healthy adults, so we assume the protein concentrations are in steady-state with no protein aggregate. We assume that the pool of free amino acids is large so that the rate bottleneck in the tRNA-charging/synthesis/folding steps is synthesis and that import/export is negligible. We also assume the protein pool is well mixed, assuring the random selection of protein for degradation (unregulated). Although reasonable for the present study, these assumptions are bad for surprisingly large sections of the proteome where reversible aggregation [81,82], multistage regulation of synthesis rates [83–85], regulated exchange of protein subunits in complexes [86, 87] and nonrandom degradation [32,67,88,89] are biologically important, but we maintain them as a starting point for the modeling. Previous literature reports have presented mathematical modeling of protein turnover rates using similar assumptions. [31,90–96] This then results in Eq 1 where the time-dependent change in a protein concentration is the difference between the synthesis and degradation rates.

(1)

This model assumes that the concentration of an individual protein ([P]) in every location is under the control of a zero-order synthesis rate (k_syn) and a concentration-dependent degradation (k_deg) step. The assumption of zero-order synthesis suggests that the precursor is stable and unresponsive to protein concentration, while a first-order rate for degradation suggests that there is no regulation of degradation other than protein concentration. In general, the rates k_syn and k_deg need not be constant as they are under the control of numerous exogenous factors that may vary in time. However, we now formalize our final assumption: protein homeostasis. This assumption is that the multiple processes regulating each protein concentration are in dynamic equilibrium so that these rates are constant for a given experimental condition. Synthesis and degradation for a given protein are equal, ensuring that the number of proteins produced is equal to the number of proteins lost. [19] Therefore, during the measurements d[P]/dt = 0, leading to the relationship: (2)

Our hypothesis is that the expression of the ApoE polymorphisms (ApoEx = ApoE2 or ApoE4) creates a unique steady state or proteostasis across the proteome that can differ from the concentration of the human wild-type control (ApoE3). The change in protein abundance (Eq 3) between the two conditions allows us to infer how the ratio of the rates is changed in the experimental cohorts, but neither k_syn nor k_deg is individually identifiable.

(3)

Using metabolic isotope labeling we can add rate information that will distinguish between changes in synthesis and degradation. Assume that at t = 0 our model simplifications are true, but that the amino acid pool is replaced with a deuterated version. Proteins synthesized after t = 0 are isotopically labeled, and we can measure the time-dependent replacement of old unlabeled for new labeled proteins. To make this mathematically explicit, we denote the concentration of normal proteins by [P] and the concentration of deuterated proteins by [P^D]. These two concentrations now satisfy the initial value problems. (4) and (5)

These are true because normal proteins are no longer being synthesized (Eq 4) and [P^D] have no initial concentration (Eq 5). These ordinary differential equations can be solved in closed-form using standard techniques. The solutions are: (6) and (7)

Notice that these equations satisfy [P]_t + [P^D]_t = k_syn/k_deg, which is independent of time as it must be in homeostasis. However, the measurable fraction of deuterated protein over time is given by.

(8)

Eq 8 seems to suggest that the degradation rate is the measurable driving force behind the turnover of old protein and the replacement by labeled protein. However, because the processes of synthesis and degradation are exactly balanced in the proteostasis condition, we can just as easily identify the turnover rate as the per-molar synthesis rate: k_syn/[P] or fractional synthesis [97] as it is commonly called. It is important to emphasize that these rates are only properly defined in homeostasis. Because there is no assurance that homeostasis is equally applied to all proteins simultaneously[85, 88, 98], we find it conceptually preferable to define the turnover rate as the mean of the per-molar synthesis and degradation rates: (9)

As stated above each experimental mouse cohort will have a unique homeostasis with a protein-specific synthesis and degradation rate. Using ApoE3 as our normal control we can assess how the average of the synthesis and degradation rates have changed with the E2 and E4 polymorphisms (ApoEx).

(10)

This means that if, for example, the ApoE polymorphism increases a protein concentration (+ΔAbundance, Eq 3) the ΔTurnover (Eq 10) will highlight whether the change in proteostasis was driven by an increase in k_syn/[P](ApoEx) or a decrease in k_deg(ApoEx) because the sign of ΔTurnover will be different for each possibility. Together the experimental ΔAbundance and ΔTurnover for each protein identifies whether the differences in proteostasis are primarily due to changes in synthesis or degradation. Graphing these values produces a plot where each quadrant (Fig 1C) has meaning. For example, a positive x-axis (+Δabundance) and y-axis (+Δturnover) suggest that synthesis increases (Syn↑). Conversely, a protein with lower expression levels (-Δabundance) between ApoE genotypes, could result from less synthesis (Syn↓) if turnover rate decreases (-Δturnover) or more degradation (Deg↑) if the protein turnover rate increases (+Δturnover). Since, each measurement has independent noise, a nonrandom enrichment of multiple proteins from a functionally related ontology within a quadrant is an important metric of confidence that the cell is regulating protein expression to change biochemical functions (Fig 1D and 1E).

Mouse handling

The mouse model employed for this study consists of ApoE knockout C57BL/6 transgenic mice that express one of the three human ApoE alleles (ApoE2, ApoE3, ApoE4, n = 24/allele, see S1 Table) homozygously under the GFAP promoter (JAX# 004632, 004633, 004631). Notably, this model has provided valuable insight into genotype-specific effects of ApoE in a large number of other experiments [70,71,99–108]. This study does not encompass differences from wild-type mice. The transgenic mice were selected with deliberate focus on ApoE isoforms rather than wild-type conditions, or age differences. The findings reported in this publication use fold change relative to ApoE3 to minimize the GFAP promoter variable as reported previously [10,11,109]. Because mouse ApoE has a low sequence identity (77%) and a different transcription promoter, the human ApoE3 model is the best control for comparison. While we recognize the limits of a transgenic model, this study provides valuable identification of in vivo patterns which can be confirmed in future ApoE knock-in mice models and human studies. These results refer solely to the effects of ApoE isoform differences, rather than Alzheimer’s Disease. Similarly, these results represent the weighted average across the entire mouse brain. We presume that some brain regions will have differential sensitivity. [14,22,34] Any claims regarding Alzheimer’s disease are made solely to highlight similarities between current ApoE/AD research and our observations to create a holistic mechanistic hypothesis.

Mice were randomly selected for replicate designation and timepoint based on availability. They were all 6–8-month-old, retired breeders with no signs of disease or neurological dysfunction. There were no exclusions among this group. Specific cohort denominations and animal numbers can be found in Fig A in S1 File. Blinding was not used during any portion of this experiment as it was necessary to compare groups at each point. Mice were housed together in the same room of the facility at the same time. Mice had ad libitum access to water and standard nutritional rodent feed (Teklad 8604) while housed in a temperature-controlled environment of ~24°C. This environment included a 12-hr circadian cycle. To initiate turnover rate measurements, mice received an intraperitoneal (IP) injection of sterile D₂O 0.9% w/v saline (35 μl/g body weight) calculated to increase internal D₂O concentrations to an initial 5% of overall water weight (w/w). Mice were then given 8% D₂O as the sole hydration source for the remainder of the experiment. This was done to maintain overall internal water at 5% D₂O enrichment. Mice were sacrificed according to the following timepoints post IP injection: day 0 (no D₂O injection), hour 6, day 1, day 4, day 16, and day 32. Mice were euthanized via CO₂ asphyxiation followed by bilateral thoracotomy. Blood was collected via cardiac puncture for D₂O enrichment calculations. Brains were divided sagittally into respective hemispheres. Brain and liver were flash frozen on blocks of solid CO₂. Tissues were stored frozen at -80°C until processing.

Tissue preparation

Singular brain hemispheres and liver sections were homogenized in lysis buffer (25mM Ammonium Bicarbonate treated with diethylpyrocarbonate and ThermoScientific Halt Protease & Phosphatase Inhibitor Cocktail) for 60 sec using a MP FastPrep-24 homogenizer. Homogenized samples were centrifuged for 15 minutes at 14,000xg to separate them into cytosolic and membrane isolates. The membrane pellet was resuspended in lysis buffer and centrifuged for 15 minutes at 14,000xg a total of three times to remove cytosolic components. Each fraction was resuspended in 5% SDS. Aliquot concentration was measured via a Pierce BCA Protein Assay Kit purchased from ThermoFisher Scientific, and 50 μg of protein were prepared according to S-Trap documentation (cytosol and membrane fractions were prepared separately). Proteins were digested with trypsin Lys/C overnight at 36°C. Resultant peptides were dehydrated in a ThermoScientific Savant SPD131DDA SpeedVac Concentrator and resuspended at a final concentration of 1 μg/μL in buffer A (3% acetonitrile, 0.1% formic acid).

LC-MS

Samples were separated and measured via liquid chromatography-mass spectrometry (LC-MS) on an Ultimate 3000 RSLC in connection with a Thermo Easy-spray source and an Orbitrap Fusion Lumos. Peptides were pre-concentrated with buffer A (3% acetonitrile, 0.1% formic acid) onto a PepMap Neo Trap Cartridge (particle size 5 μm, inner diameter 300 μm, length 5 mm) and separated with an EASY-Spray HPLC Column (particle size 2 μm, inner diameter 75 μm, length 25 mm) with increasing buffer B (80% acetonitrile, 0.1% formic acid) gradient:

0–5 min, 0 to 5% B; 5–87 min, 5 to 22% B; 87–102 min, 22 to 32% B; 102–112 min, 32 to 95% B; 112–122 min, 95% B; 122–125 min, 95 to 2% B; 125 to 127 min, 2% B; 127–129 min, 2 to 100% B; 129–132 min, 100% B; 132–133 min, 100 to 2% B; 133–135 min, 2% B; 135–137 min, 2 to 100% B; 137–140 min, 100% B; 140–142 min, 100 to 0% B; 142–144 min, 0% B.

The MS-based data-dependent acquisition method was set to a 3 second cycle time. MS1 scans were acquired by the Orbitrap at a resolution of 120,000. Precursors with a charge > 1 and < 6 were selected for MS2 fragmentation. MS2 scans of CID precursor fragments were detected with the linear ion trap at a scan rate of 33.333 Da/sec with a dynamic injection time. CID collisions were set to 30% for 10ms. A 60 second dynamic exclusion window was enabled; isotopes and unassigned charge states were excluded. The deuterium labeling information was collected separately in an MS1-only acquisition with the Orbitrap at a resolution of 60,000 as previously described by Naylor et al. [110]

Raw Data Processing for Peptide Identification and Label-free Quantitation

Raw files were searched against the 2022 Uniprot/Swissprot mus musculus FASTA (containing 17144 entries) using Peaks Studio v.11 (Bioinformics Solutions Inc.) for label-free quantitation (LFQ) analysis. During the data refinement step, the feature “associate feature with chimera [DDA]” was selected to deconvolute scans with co-eluted isobaric peptides. The parent mass error tolerance was set to ±15 ppm and the fragment mass error tolerance was set to 0.5 Da. Cysteine carbamidomethylation was set as a fixed modification, and both methionine oxidation and pyro-glu from glutamine were set as variable modifications in the search. Digest mode was set to semi-specific for the trypsin-lysC enzyme mix allowing for ≤ 3 missed cleavages and the peptide length range was set to 6–45 amino acids. The false discovery rate (FDR) for peptide matches was set to 1%, and protein ID significance was set to -10log(P-value) ≥ 15 for each identified protein.

Peaks Studio (Bioinformics Solutions Inc.) was also used to search raw files for use in Deuterater [110] software. The raw files were searched against the 2021 Uniprot/Swissprot mus musculus FASTA (containing 17144 entries). Peptide searches were performed using trypsin/lysC semi-specific digest with a tolerance of ±20ppm and missed cleavages ≤ 3. Carbamidomethylation was set as a fixed modification and pyro-glu from glutamine and methionine oxidation were set as variable modifications. Within the Peaks Studio DB module, proteins were identified with two or more unique peptides at an FDR of 2% and significance was set to -10log(P-value) ≥ 15 for each identified protein.

Protein ΔAbundance analysis

The group of mice used in this paper were divided into two male and two female groups for analysis [79]. Each group produced a dataset for cytosolic proteins and another dataset for membrane proteins. Please refer to the Experimental Design and Statistical Rationale section for more information resulting in a total of eight datasets.

Data filtering, normalization, and quantitative calculations were performed independently for each dataset following standardized metrics for data quality and analysis following the process described by Aguilan et al. [111] Each Peaks Studio DB protein.csv output dataset contains the proteins identified in the analysis and the expression values (relative abundance) for each protein in each sample are labeled as “Area”. This output was filtered to retain only the top proteins in each protein group and proteins with at most one missing protein “Area” value per genotype (i.e., n– 1/genotype/dataset). Subsequently, protein “Area” values in the dataset underwent log₂ transformation. The distribution of these protein “Area” values was mean centered by subtracting by the average protein “Area” from each protein “Area” within the sample. To ensure comparability across samples, the distribution width was also normalized between samples by calculating the correlation slope between these total average protein “Area” values across all samples and the individual sample values. Each protein “Area” in a sample was then divided by the corresponding sample slope. For samples with a missing protein “Area” value, imputation was carried out using the scikit-learn KNN imputer function module in python with the two closest neighbors. [112]

Protein fold change (FC) values, which represent the relative change in protein abundance values (“Area”) compared to a reference, were calculated, and used as a metric of change in abundance (Δabundance). For this study, FCs were calculated for protein expression values in ApoE2 mice and ApoE4 mice with ApoE3 expression values as reference, respectively. As per Aguilan et al.’s methodology, an F-test was employed to assess the variance between protein expression values before performing p-value calculations for statistical significance. To evaluate the statistical significance of expression levels in each comparison, a two-sample independent t-test (homoscedastic) was employed for proteins with an insignificant F-test result and a two-sample independent t-test (heteroscedastic) for proteins with a significant F-test result. Both the F-test and t-test calculations were conducted with the Scipy python package. [113]

Protein FC values were averaged across all datasets for each respective comparison. This produced a single set of “Area” (expression value) FCs for each comparison. Please note that both the ApoE2 vs ApoE3 (E2vsE3) comparison and the ApoE4 vs ApoE3 (E4vsE3) use the same list of quantified proteins. As outlined by Van den Berg [114], protein FC values from individual comparisons were range scaled using the following formula prior to ontology exploration: (11)

Where x’_ij, x_ij, , x_imax, and x_imin are the scaled FC value, non-scaled FC value, mean FC, largest FC, and smallest FC, respectively. Range scaling was selected because it captures relative change in protein expression while considering the full range of values specific to the dataset. These scaled FC values will be utilized in functional analyses as described in the Ontology-level Calculations section below. The python script created for the steps outlined in this quantitative analysis can be found in the GitHub repository JC-Price/PublicProteomeDataAnalysis: scripts for proteome data wrangling.

Protein ΔTurnover rate calculation

Protein turnover rate values were calculated using Deuterater[110] v5. This software uses an accurate mass and time database to extract peptide isotope patterns from LC-MS (MS1) centroided data utilizing feature identifications (e.g. retention time, mass, peptide ID, etc.) obtained from MS/MS data (refer to the Raw Data Processing for Peptide Identification and Label-free Quantitation section above).

Isotope patterns were extracted from MS1 raw data with an extraction retention time window of 1.5 min and an m/z error limit of ≤ 30 ppm. The n-value represents the number of available positions on a peptide where deuterium can replace hydrogen. In the theory generation step, peptides with data missing in an extracted file are removed, and the n-value is calculated for remaining extracted peptides based on known quantities for each amino acid [115,116]. Subsequently, Fraction New measures the amount of turnover rate for each peptide in a file by calculating changes in neutromer abundance and spacing [110]. These calculations were performed using the average between M0 and the highest isotope peak for peptides meeting specified criteria, including peptide n-values greater than 5, a minimum peptide sequence length of 6, and a minimum allowed M0 change of 0.04. In the Rate Calculation stage, the data from the Fraction New step is fitted to a kinetic rate curve using Eq 8 from our proteostasis model. Turnover rates were calculated for peptides that met a specified criterion, including a minimum of 3 non-zero peptide timepoints, and measurement deviation of less than 0.1, as previously reported [86]. The asymptote value is assumed to be 1 in the first iteration of analysis for proteins, but not for lipids where multiple pools of the same lipid are frequently observed [117].

After the Deuterater [110] analysis, all proteins with a valid turnover rate value (Rsq ≥ 0.6, combined unique peptides > 1, combined rate > 0) grouped by allele cohort, and the average turnover rate value was calculated for each protein in the cohort, respectively. These protein turnover rate values were log₂ transformed, and the protein turnover rate FC was calculated as the difference of the log₂ rates. The ApoE2 mice and ApoE4 mice were compared to the reference ApoE3 mice, resulting in a single set of protein turnover rate FCs for each comparison. To standardize the protein turnover rate FCs, auto scaling [114] was applied, where x’_ij, x_ij, , and s_i are the scaled turnover rate FC value, non-scaled turnover FC rate value, mean turnover FC rate and turnover rate FC standard deviation, respectively: (12)

Auto scaling was implemented because it considers both the mean and standard deviation to reduce the effects of outliers and variation in the data while preserving the ability to focus on small changes. It is important to note that because of the signal to noise requirements fewer proteins had valid turnover rate FC values than quantifiable abundances. Consequently, proteins with turnover rate FCs represent a smaller subset population in comparison proteins with expression value FCs calculated from “Area” values. These protein turnover rate values were used for ontology analysis as outlined in the Ontology-level Calculations section below. For further reference, the python script created to process the calculated_rates output from Deuterater [110] can be found in the GitHub repository JC-Price/PublicProteomeDataAnalysis: scripts for proteome data wrangling

Ontology-level calculations

The StringDB [80] multiprotein tool was employed to calculate group FC values for functionally-related protein groups (ontologies) regardless of statistical significance. To streamline the analysis and reduce the number of redundant term ID (ontologies), ontologies were selected only from the following established: GO Process, GO Function, GO Component, KEGG, Reactome, and WikiPathways. To quantify the representation of each ontology, the “observed gene count” was divided by the “background gene count” to calculate the “Ontology_coverage (%)” for each ontology. Only ontologies with ≥25% were included in this analysis. This latter criterion ensures that the identified ontologies are sufficiently represented in the data (S4 and S5 Tables).

The “matching proteins in your network (labels)” was used to associate each observed protein in the ontology with the calculated “Area” FC and turnover rate FC, respectively, for both the E2vsE3 and the E4vsE3 comparison. Next, the average protein “Area” FC and turnover rate FC was calculated for each identified ontology by averaging the FC values of proteins within that category. This step summarized the collective expression and turnover rate changes of proteins within specific functional groups for each comparison.

To assess the statistical significance of the FC values within each ontology, a one-sample t-test with null hypothesis (H0) stating the average ΔAbundance (Eq 3) for the ontology is equal to 0, and the alterative hypothesis (Ha) indicating that it is not equal to 0. This statistical test is used to determine whether the observed changes in protein expression for the ontology, as a whole, were statistically significant. To account for multiple comparisons and maintain a controlled false discovery rate (FDR), the Benjamini-Hochberg correction (BH-PV) was calculated for the resulting p-values (FDR = 0.25). Ontologies with a BH-PV < 0.05 were considered statistically significant. In the case of highly similar ontology with significant changes, the ontology with the most proteins was selected to represent the results. The Python code used to analyze StringDB and calculate the FCs can be found in the GitHub repository JC-Price/PublicProteomeDataAnalysis: scripts for proteome data wrangling. https://github.com/JC-Price/PublicProteomeDataAnalysis