Neuronal cells similar to muscle cells usually remain in a quiescent cell cycle state once they are differentiated due to their specialized functions. But cell cycle regulatory proteins like cyclins, CDKs, caspases, and p53 continue to be required for axonal migration, maturation, and survival (López-Sánchez et al., 2017; Joseph et al., 2020). One or more of these cell cycle proteins and pathways might get activated in response to various epigenetic or pathologic stimuli. Aberrant cell cycle reentry of neurons with duplication of the genetic material but without subsequent cell division is a well-known phenomenon. However, it is still unknown whether this is deleterious or beneficial. In the early stages of brain development, specific populations of neurons undergo incomplete mitosis and stay in a tetraploid state for the rest of their adult life. Still, a de novo neuronal tetraploidization in the adult brain seems to be an early indicator for neurodegeneration, with cells showing elevated molecular stress response and apoptotic marker proteins due to this aberrant regulation of cell cycle (Frade and Ovejero-Benito, 2015; López-Sánchez et al., 2017; Joseph et al., 2020). The correct assembly, positioning, and working of the mitotic spindle apparatus are essential for the development and differentiation of cells. Spindle orientation defects leading to an imbalance between symmetric and asymmetric divisions have been discussed as being linked to NDD. An asymmetric cell division leads to two daughter cells with different cell fates. An important example is the division of a stem cell, leading to a differentiated brain cell and another stem cell. The preconditions for an asymmetric cell division are the polarization of the content of the mother cell and the alignment of the mitotic spindle along the axis of polarity. If this process is disturbed by an impaired spindle orientation, this will have severe consequences in the development of the brain (Noatynska et al., 2012). During mitosis, motor proteins, together with Huntingtin, localize along spindle microtubules to segregate the chromosomes (Godin and Humbert, 2011). Thus, mutant Huntingtin could lead to defective neurogenesis during embryonic development leading to Huntington's disease later in life (Godin and Humbert, 2011; Wiatr et al., 2018). Defective DNA repair mechanism, together with elevated oxidative stress, is a typical hallmark for aging cells (Joseph et al., 2020). As the brain is the largest consumer of oxygen, reactive oxygen species (ROS) are the main inducers of DNA damage in neurons (Nissanka and Moraes, 2018). The accumulation of unrepaired damaged DNA might contribute to a cell cycle reentry of the neuron to increase polyploidy, thus compensating for the loss of information on some parts of the DNA. Also, general terms like transcription (mitochondrial) translation, RNA modification, spliceosome, and chromatin organization were defined as subprocesses integrated into the hallmark called the cell cycle.

Obviously, transcription, translation, and chromatin remodeling as major contributors to protein homeostasis have a considerable impact on all cellular functions, and genetically or epigenetically misregulated transcription factors can enhance or weaken all the processes that lead to or prevent neurodegeneration like neuroinflammation, oxidative stress, and protein homeostasis (Berson et al., 2018; Jin et al., 2019). Especially neurons face the problem that their locus of transcription is distant from the locus of translation as protein synthesis is decentralized to meet the rapidly changing metabolic needs of axons and dendrites. Nuclear mRNA is transported from the cell body to the periphery as RNA/protein granules linked to motor proteins of the cytoskeleton by adaptor complexes. Conditions, such as mislocalization of the mRNA, a defective cytoskeleton, and deleterious adaptor complexes, lead to synaptic dysfunction, loss of synapses, and, ultimately, to neuronal death (Liu et al., 2017).

Accurate folding and control over levels of many proteins are tightly regulated in neuronal cells by a protein quality control network consisting of molecular chaperones and an intact degradation system of misfolded or discarded proteins, consisting of the ubiquitin-proteasome system and autophagic processes. Chaperone-mediated processes are involved in the promotion of de novo folding of nascent proteins or the refolding of misfolded polypeptide chains, thereby preventing their aggregation (Hartl et al., 2011; Chaplot et al., 2020).

Proteins that cannot be folded correctly or are no longer needed are tagged by the attachment of the small protein ubiquitin (ubiquitinylation) and degraded by a multi-subunit protein complex called the proteasome or by autophagy, which is a universal lysosome-dependent intracellular degradation process for not only protein aggregates but also cell organelles and cytoplasmic constituents to maintain cell homeostasis (Nah et al., 2015; Schmidt et al., 2021). Lysosome-dependent autophagy can either be carried out by lysosomes acting by themselves in a process called micro-autophagy, or in macro-autophagy where cell organelles with a lipid double layer derived from the ER called omegasomes fuse with lysosomes to form autophagosomes (Mizushima and Klionsky, 2007; Li et al., 2012; Nah et al., 2015). In neurons, these autophagosomes are generated constitutively in the cell periphery and mature on their way to the cell body (Xie and Klionsky, 2007). The proteins ensuring this axonal transport are Huntingtin and Huntingtin-associated protein 1. As huntingtin is predominantly cleared by autophagy, the deficiency of intact huntingtin in HD is a double stress factor leading to the accumulation of polyglutamine-expanded huntingtin polyQ-htt and other cytoplasmic constituents that need to be degraded (Wong and Holzbaur, 2014). Besides deficiencies in the protein quality control network leading to the typical protein aggregations in NDD, there are many more factors that impair homeostasis of the cell, causing stress to the cells. Heat, toxins, mechanical damage, infections, starvation, hypoxia, and oxidants all lead to cellular stress responses that aim to control the induced damage and, if possible, to confer resilience to the stressor. The integrated stress response (ISR) is an elaborate signaling cascade, where different stress signals activate different protein kinases that converge in phosphorylating the core of the ISR, the α-subunit of the translation initiation factor 2 (eIF2α). This leads to a reduction of global protein synthesis in favor of ISR-specific mRNAs, such as the activating transcription factor ATF4. ATF4 is the main effector of the ISR. It forms homo- and heterodimers that bind to DNA targets to control the expression of genes involved in cellular adaptation, mainly chaperones that refold denaturated proteins or tag them for immediate degradation (Harding et al., 2003; Fulda et al., 2010; Pakos-Zebrucka et al., 2016). To terminate the integrated stress response, dephosphorylation of eIF2α is required. However, dephosphorylation of eIF2α can also facilitate the production of death-inducing proteins in cases where the cell is so severely damaged that normal functioning and homeostasis cannot be restored (Pakos-Zebrucka et al., 2016). Highly regulated cascades of events will then lead to the decomposition of the cell in a process called apoptosis (Danial and Korsmeyer, 2004; Elmore, 2007; Fulda et al., 2010). Apoptosis is a caspase-mediated programmed cell death that can be triggered either by stress signals from within the cell or by signals that are released by the surrounding cells and that bind to cell-surface death-receptors (Danial and Korsmeyer, 2004; Elmore, 2007). Both intrinsic and extrinsic pathways ultimately induce cell death by activating caspases, enzymes that carry out the degradation of the cell, leading to cell shrinkage, plasma membrane blebbing, chromosome condensation and degradation, mitochondrial death with release of cytochrome C, cellular fragmentation, and formation of membranous apoptotic bodies. At the same time, phosphatidylserine residues are exposed at the cell surface to attract macrophages that phagocytize the cell fragments. The extrinsic pathway is activated by extracellular ligands (e.g., CD95 ligand, tumor necrosis factor α) binding to cell-surface death receptors, which either leads to the formation of the so-called DISC (death-inducing signaling complex) or the activation of nuclear factor kappa B (NFκB) (Wang and El-Deiry, 2003; Lavrik et al., 2005). Caspase 8, which serves as the initiator caspase, is activated within this complex. If the activating stimulus is sufficiently large, caspase 8 begins activating the effector Caspases 3, 6, and 7. They translocate to the mitochondria to trigger pore formation in the outer mitochondrial membrane (Fan et al., 2005), which ultimately leads to mitochondrial permeability and death. The release of cytochrome C promotes caspase-9 activation, which triggers a cascade of proteolytic events. The intrinsic signaling pathway is controlled by the Bcl2-family of proteins and is triggered by cell-internal factors, such as DNA damage, osmotic stress, or growth factor withdrawal. The triggering cell event activates the transcription factor p53, leading to the expression of proapoptotic Bad protein. The Bad protein, in turn, activates the multidomain proapoptotic Bax protein, which triggers permeabilization of the mitochondrial outer membrane. As a result, similar to the extrinsic pathway, cytochrome C escapes from the intermembrane space of the mitochondria and induces a proteolytic cascade (Mcilwain et al., 2013; Wang and Tjandra, 2013; Chen et al., 2018).

The NFkB signaling pathway is activated during apoptosis either to enforce the ongoing apoptotic processes or to navigate the cell back to the survival pathway by stimulating the expression of anti-apoptotic genes, especially TRAF1 and TRAF2, and thus overruling the activities of the caspases (Sheikh and Huang, 2003). NFκB is a family of transcription factors with multiple physiological and pathological functions. The NFκB complex exists in an inactive state in the cytoplasm. It consists of different NFkB dimers and inhibitory kB proteins (IκB) that prevent NFkB to enter the nucleus and bind to the DNA (Israël, 2010). NFkB can be activated by highly variable stimuli, including growth factors, cytokines like tumor necrosis factor alpha and interleukin 1-ß, bacterial and viral antigens like lipopolysaccharides or double-stranded RNA and physicochemical insults like ionizing radiation or free radicals. All these different stimuli activate different phosphorylation cascades that, in the end, lead to an interaction with the IκB kinase (IKK) complex, which then leads to the phosphorylation of IκB in the NFkB complex. Once phosphorylated, IκB gets ubiquitinylated and degraded by the proteasome. The released NFkB can now enter the nucleus to activate target gene expression that regulates immune responses, cell growth, and proliferation but also apoptosis (Qin et al., 2007). The selectivity of the NF-κB response is based on several factors, including dimer composition, timing, organization of chromatin, and cell type (Natoli, 2009; Sen and Smale, 2010). Besides the NFkB signaling pathway, an important low-level term in autophagy and apoptosis, was the mTOR signaling pathway. As mentioned above, an impaired autophagy in NDD leads to the accumulation of toxic protein aggregates that promote cellular stress and death. The kinase mammalian target of rapamycin (mTOR) is a central regulator of metabolism as it senses cellular nutrient, oxygen, and energy levels (Perluigi et al., 2015). It is a potent repressor of autophagy and reacts to food abundancy with cell growth and proliferation, whereas, in times of starvation, it is inhibited, thus allowing the cells to derive energy from autophagic and apoptotic processes (Heras-Sandoval et al., 2020).

The extracellular matrix (ECM) is a three-dimensional network that reaches all cells of the body to provide structural support, segregate tissues, and regulate intercellular communication by many different biochemical processes. It consists of minerals and macromolecules like collagen, elastin, enzymes, glycoproteins, and hyaluronan, which is most abundant in the brain (Bonnans et al., 2014). These compounds contribute to a varying degree of stiffness and elasticity of the ECM throughout the body from hard bone to soft brain tissues. The mechanical properties of the ECM environment can be sensed by the cells, which is the precondition for processes like cell migration, proliferation, differentiation and growth, apoptosis, and, on a higher level, the development of tissues and organs (Bonneh-Barkay and Wiley, 2009; Sethi and Zaia, 2017; Maguire, 2018). These processes also require a direct connection between cells and ECM. The ECM communicates with the cells by connecting to the cytoskeleton through cilia and (focal) adhesion complexes (Geiger and Bershadsky, 2002). The cytoskeleton is a functionally versatile filamental structure that transports and localizes the contents of the cell, it organizes the process of cell division, namely, chromosome segregation and cytokinesis, it connects the cell physically and biochemically to the external environment through interacting with the ECM (cell adhesion, focal adhesion, cell junction), and it enables the cell to move and change shape (cell migration, cell differentiation, growth, and localization) (Wozniak et al., 2004; Fletcher and Mullins, 2010). These functions are mediated by three filament types: microtubules, actin microfilaments, and intermediate filaments, which, in neurons, are called neurofilaments (Herrmann et al., 2007; Hohmann and Dehghani, 2019). Microtubules are hollow tubes consisting of tubulin dimers, which are stabilized by tau proteins. They grow out from the centrosome to the plasma membrane by constantly adding and subtracting tubulin dimers at both ends of the filament (Mitchison and Kirschner, 1984). They are responsible for the intracellular transport together with the motor proteins kinesin and dynein that transport, e.g., vesicles, nutrients, and essential organelles, but also broken-down neurotransmitters and misfolded proteins by extending along the axon and functioning as the cytoskeletal navigation pathways in the transporting process (Guo et al., 2020). The hyperphosphorylation of tau proteins as seen in AD results in microtubule destabilization and cytoskeleton abnormalities (Mandelkow et al., 1995; Barbier et al., 2019). Actin ß- and γ- monomers, together with formin homology proteins, form linear polymers called actin microfilaments, which can then be arranged as a network with the help of the proteins Arp2 and Arp3. The Arp2/3-complex forms dimers that resemble actin dimers, but are much more stable. They attach themselves in a 70° angle and serve as nucleation cores for actin filaments activated by the nucleation-promoting factor proteins of the WASP/WAVE-family (Kurisu and Takenawa, 2009). Both filaments and networks stabilize the plasma membrane of the cells and control the position of the nucleus and other cell organelles (Pfisterer et al., 2017). They maintain and adapt cell shape, thus allowing for cytokinesis, organization of cell junctions, and enhancement of cell adhesion. Cell migration is achieved by the polymerization of new actin filaments in the forward edge of a moving cell that pushes the cell membrane forward in protrusions called lamellipodia. These protrusions form mechanical links between the actin filaments and the ECM called focal adhesions. Once attached, the rear of the cell body contracts, squeezing its contents forward past the adhesion point. Once the adhesion point has moved to the rear of the cell, the cell disassembles it, allowing the rear of the cell to move forward (Ananthakrishnan and Ehrlicher, 2007). The actin filaments also provide the tracks for cellular transport, allowing actin-based motor proteins (myosins) to transport organelles over short distances.

The interactions of a-actin und myosin but also troponin and tropomyosin are most sophisticated in muscle cells. These highly specialized cells work together to contract and release muscles, thus leading to movement of the whole organism and to the cardiac muscle action potential (Cooper, 2000). Intermediate filaments are the major cytoskeletal proteins, which, in contrast to microtubules and microfilaments, are more heterogeneous, have tensile strength, and do not participate in cell motility (Herrmann et al., 2009). The intermediate filaments primarily provide structural support for the cell, especially for the long neuronal cells where they regulate the diameter of the axon and, thereby, the nerve conduction speed (Helfand et al., 2003).

All cells have intermediate filaments, but the protein subunits of these structures vary. They can be classified into six different types: Types I and II are the keratins, which are expressed in epithelia. Type III contains the proteins vimentin, which frequently colocalizes with microtubules, desmin, peripherin, and glial fibrillary acidic protein (GFAP). Type IV consists of the neurofilament proteins L, M, H, and internexin. Type V consists of the nuclear lamins, and Type VI consists of the protein nestin (Muñoz-Lasso et al., 2020). During embryonic development of the brain microtubules, microfilaments and neurofilaments work together to guarantee proper growth and guidance of axons; during adult life, they maintain neuronal homeostasis and plasticity. Given these fundamental processes rely on an intact cytoskeleton organization, it is not surprising that cytoskeleton defects, including alterations in microtubule stability, in axonal transport as well as in actin dynamics, have been characterized in several NDDs.

Neuronal signal transduction and transport of organelles and vital molecules along the axon maintain neuronal activity and health. In NDD, neuronal signaling is more and more impaired because of synaptic and cytoskeletal dysfunctions. The loss of synapses has detrimental consequences as they are critical to learning, memory, and behavior (Lepeta et al., 2016; Ardiles et al., 2017), and perturbations in axonal transport can lead to neuronal cell death (Perlson et al., 2010).

Cell communication in the brain is achieved by neuronal signaling via synapses. In this process, electrical activity is transferred from one neuron to another through neurotransmitters as chemical mediators. The most important neurotransmitters in the central nervous system are glutamate for the excitatory, and GABA and glycine for the inhibitory synapses, but also dopamine, serotonin, and acetylcholine (Fogarty et al., 2016). Action potentials that travel along axons induce an intracellular Ca²⁺ influx mediated by voltage-gated calcium channels at the presynaptic terminal. Increased Ca²⁺ levels are sensed by the synaptic vesicle protein synaptotagmin I, which induces the Soluble NSF Attachment Protein Receptor (SNARE) protein complex to mediate fusion of the neurotransmitter vesicles with the plasma membrane (SNARE binding), releasing them into the synaptic cleft (neurotransmitter secretion). The neurotransmitters then bind to and thus activate receptors of the postsynaptic cell, which can be either ion channels or G protein-coupled receptors. They transmit signals from the post-synaptic membrane to the cell body by transducing the chemical signal into an electrical signal that depolarizes the postsynaptic cell and is transmitted downstream to the next neuron (Lepeta et al., 2016; Taoufik et al., 2018).

Rather than being just a gap, the synaptic cleft is full of trans-synaptic adhesion molecules that not only physically connect the pre- and post-synaptic compartment but also mediate recognition and signaling processes that are essential for the establishment, specification, and plasticity of synapses. Such synapse-organizing adhesion molecules include neurexins and neuroligins, cadherins, integrins, receptor phosphotyrosine kinases, and phosphatases, such as ephrins and Rho GTPases (Missler et al., 2012; Jang et al., 2017). Deficiencies in the complex synaptic functioning are not only seen in neurodegenerative diseases but also in neuropsychiatric disorders, suggesting that the loss of synapses is not the endpoint incident of the disease but, rather, a triggering event (Taoufik et al., 2018).

Axonal transport, in general, is a process where the motor proteins kinesin and dynein loaded with diverse cargos navigate along the microtubules of the cytoskeleton. Anterograde axonal transport delivers proteins, lipids, mRNA, and mitochondria to the distal synapse, whereas retrograde transport clears misfolded or aggregated proteins and brings distal trophic signals to the soma (Millecamps and Julien, 2013; Sleigh et al., 2019).

Transmembranal transport systems in neurons are the anterograde ER to Golgi transport with COP II-coated vesicles and the retrograde trans-Golgi-Network transport with COP I-coated vesicles (Martínez-Menárguez et al., 2019; Wang L. et al., 2020).

Neurotransmitters or other secretory molecules leave the ER packed in COPII-coated vesicles to be further processed in the Golgi apparatus before being secreted into the synaptic cleft. After an action potential, neurotransmitter and receptors are recycled by the cell. They are protected from lysosomal degradation by COPI-coated vesicles of the retrograde trafficking and redirected to the ER where they are recycled for the next release (Lu and Hong, 2014).

Deficiencies in axonal transport through defects in the cytoskeletal organization, the impairment of motor protein attachment to the microtubules or the destabilization of motor-cargo binding but also aberrant transmembranal anterograde and retrograde transport between ER and Golgi apparatus can induce synaptic failure and ultimately lead to neurodegeneration, as shown in AD, PD, ALS, and HD (Perlson et al., 2010; Guo et al., 2020).

A fundamental feature in NDD is the dysregulation of the innate immune response in the central nervous system by chronic inflammation. The innate immune response of the brain represents not only the first line of defense against invading microorganisms (defense response to other organisms) but also responds to endogenous stress like dying or damaged cells after cellular stress. It consists primarily of microglial cells and astrocytes. Microglia are resident tissue macrophages and the principal mediators of inflammation (Frank-Cannon et al., 2009), whereas astrocytes mainly supply neurons with nutrients and energy and form the blood–brain barrier (Blackburn et al., 2009). Both microglia and astrocytes express on their cell-surface pattern-recognition receptors that detect pathogen-associated molecular patterns like residues from bacteria, viruses, fungi but also danger-associated molecular patterns, which can be abnormal endogenous proteins, iron overload, antibodies, cytokines, and chemokines, including toll-like receptors (Akira et al., 2006; Medzhitov, 2007). Also, complement factors bind to these receptors, initiating the complement and coagulation cascade, which enhances the immune response of microglia and astrocytes.

Pathogen- and danger-associated molecular patterns induce neuroinflammation by activating the transcription factor NFkB in microglial cells. NFkB induces the synthesis of proinflammatory cytokines like Interleukin−1ß (IL-1β), IL-6, IL-12, interferon gamma (IFN-γ), chemokines, including the C-C motif chemokine ligand 2 and tumor necrosis factor alpha (TNF-α), but also prostaglandins (Schaefer, 2014; Guzman-Martinez et al., 2019) to promote efficient clearance of cell debris and plaques. Also, these factors, being released into the cell environment, activate the astrocytes to both increase the permeability of the blood–brain barrier for easier recruitment of B- and T-adaptive immune cells (leukocyte activation) in the brain parenchyma and to release intracellular signals like neurotrophic and growth factors and cytokines, promoting neuronal survival, neurite growth, and neurogenesis (Jha et al., 2019). At the same time, microglial cells synthesize and release short-lived cytotoxic factors, such as superoxide radicals, nitric oxide, and reactive oxygen species that help remove pathological agents (Harry and Kraft, 2008; Jurga et al., 2020).

Although an acute insult may trigger oxidative and nitrosative stress and a slightly more permeable blood-brain barrier, this reaction is typically short-lived and unlikely to be detrimental to long-term neuronal survival (Kurutas, 2015). But disease-induced protein aggregations like oligomers of tau and Aβ, aging, and other risk factors can lead to chronic neuroinflammation with a long-standing and, often, self-perpetuating neuroinflammatory response that persists long after the initial injury or insult. Permanently activated microglia may not be able to remove amyloid-beta deposits, which, in turn, contributes to further plaque accumulation as opposed to clearance (Frank-Cannon et al., 2009). Also, the permanently increased oxidative and nitrosative stress leads to damaged neurons, synaptic dysfunction, loss of synapses, and neuronal death (Schain and Kreisl, 2017; Shabab and Zorofchian, 2017; Kinney et al., 2018), aggravated by the fact that astrocytes lose their neurotrophic function during the progressive loss of neurons. This process called astrogliosis affects considerably the neurons and their integrity (Sofroniew and Vinters, 2010; Sofroniew, 2015). Chronic inflammation can also trigger signaling pathways that activate brain tau hyperphosphorylation in residues that are not modified under normal physiological conditions, again leading to degenerated neurons and, finally, the now permeable blood-brain barrier allows much more pathogens to enter the brain than the innate immune response can cope with (Gendelman, 2002).

We associated the highest number of low-level terms across all six hallmarks to the metabolism hallmark as its pathways and biological processes are involved in all basic physiological processes, and the homeostasis of the compounds and their interactions is vital for the well-functioning of the organism. As neurons but also astrocytes, oligodendrocytes, and microglia are highly dependent on a continuous energy supply. One of the most significant hallmarks for NDD is the dysfunction of mitochondria. Mitochondrial metabolism, comprising the oxidation of pyruvate and fatty acids, the citric acid cycle, the oxidative phosphorylation within the respiratory chain, and the formation of antioxidants, was assigned to the hallmark metabolism, which also contains metabolic processes in the cytoplasm like fat metabolism, protein, carbohydrate, and nucleic acid metabolism, but also homeostasis of the cell. Lipid molecules are key components of the brain's complex structure and function, with lipids comprising around 50% of the brain's dry weight (O'brien et al., 1964). Many neurological diseases are caused by mutations in genes that are involved in lipid metabolism (Mesa-Herrera et al., 2019; Petit et al., 2020). As associated low-level terms, we included lipid-, cholesterol-, sphingolipid metabolism, lipoproteins, carboxylic acid, and fatty acid oxidation. Lipids function as key regulators of neurotransmission; cholesterol is a membrane stabilizer and organizer during synaptic vesicle exo- and endocytosis; sphingolipids and their metabolites are regulators of the fluidity of cell membranes and second messengers in signal transduction (Puchkov and Haucke, 2013; Barber and Raben, 2019). Being hydrophobic, lipids and cholesterol must be transported through the blood by lipoproteins. The term carboxylic acid was mostly connected to biological processes concerning the citric acid cycle and was thus assigned to mitochondrial metabolism, like fatty acids, which are degraded in the mitochondria in an energy-generating process called ß-oxidation (Petit et al., 2020).

Main low-level terms of protein metabolism were branched amino acid degradation and proteoglycans. Valine, isoleucine, and leucine are essential branched-chain amino acids (BCAAs). Their catabolism starts in muscle mitochondria and yields NADH and FADH2, which can be utilized for ATP generation. In the brain, BCAAs contribute to the synthesis of excitatory glutamate and inhibitory gamma-aminobutyric acid (GABA) and, at the same time, can reduce the production of serotonin and the catecholamines because they compete for transport across the blood-brain barrier with the aromatic amino acids tryptophan, tyrosine, and phenylalanine, which are the precursors of these neurotransmitters (Fernstrom, 2005). AD metabolomics studies demonstrate that altered branched-chain amino acids metabolism accompanies Alzheimer's disease development. Lower plasma valine levels are correlated with accelerated cognitive decline in AD and the severity of motor dysfunction in HD (Polis and Samson, 2020; Xu et al., 2020). No clear correlation between the levels of BCAA and ALS or PD could be found (Bjornevik et al., 2019; Xu et al., 2020).

Carbohydrates are the main source for energy production in the brain and are broken down by glycolysis, the tricarboxylic acid cycle, and oxidative phosphorylation (Attwell and Laughlin, 2001). The main causes for defective glucose usage in NDD are deleterious mitochondria and a decreased secretion of insulin, coupled with resistance to its actions, problems that link NDD to type II diabetes (Ristow, 2004; Haan, 2006; Seneff et al., 2011; Schaeffer et al., 2021).

During nutritional uptake, the insulin pathway secures the release of the hormone insulin into the blood, which then binds to insulin receptors on the cell surface and thus regulates the uptake of glucose into the cells. But insulin also controls divergent signaling pathways like the Act-pathway or the MAP-kinase pathway, which are involved in cell growth and differentiation (Saltiel and Pessin, 2002). Studies with mouse models have shown that insulin is also involved in the excitability of (hippocampal) neurons (Dai et al., 2014). AD neurons show a reduced action potential due to soluble Aß, an effect that is aggravated by missing insulin and that can be partially reversed by the anti-cancer agent bexarotene (Cramer et al., 2012; Dai et al., 2014).

Another important task of carbohydrates is the glycosylation of proteins and lipids in the ER. In the brain, glycosylated proteins participate in a myriad of processes, from electrical gradients to neurotransmission (Conroy et al., 2021), and glycolipids are crucial for the interaction of cells and for signal transduction (Reily et al., 2019). Thus, an intact N-glycosylation and O-glycosylation of lipids and proteins are important for the well functioning of the brain, and disrupted glycosylation can lead to NDD (Moll et al., 2020), e.g., O-GlcNAcylation of CNS proteins important for axonal and synaptic function is significantly reduced in AD, ALS, HD, and PD patients' tissue (Liu et al., 2004; Lüdemann et al., 2005; Kumar et al., 2014).

Nucleotides are not only the building blocks of DNA and RNA but also are involved in cellular metabolism, with, especially, ATP and NADH as energy carriers and cAMP, a second messenger molecule, as a translator of signals from outside to cellular processes. Thus, these molecules are involved in all basic physiological processes critical to the proper function of the organism (Huang et al., 2021). Nucleotides are either built de novo or, as this process takes much energy, recovered during the degradation of DNA and RNA in a nucleotide salvage pathway. A disbalance in the synthesis pathways again disturbs energy homeostasis and many cellular processes like proliferation, differentiation, migration, and apoptosis. The respiratory chain complex in the mitochondria not only efficiently produces energy but also reactive oxygen species that can harm the mitochondrial DNA. Antioxidants like glutathione protect mitochondria and cells against damage, but, if this homeostasis is disturbed, oxidative stress induces the destruction of mitochondria and apoptosis of neuronal and other brain cells (Kausar et al., 2018).

In this meta-study, we used datasets of four different omics-layers: Genomics (SNP data), Transcriptomics, Proteomics, and Methylomics. In the following, we describe where these data were acquired from.

The SNP data are based on the genome-wide association studies (GWAS) Catalog (Buniello et al., 2019), which has been updated on November 18, 2021. The experimental factor ontology (EFO) numbers for the exact search pattern were EFO_0000249 (Alzheimer's disease), EFO_0002508 (Parkinson's disease), Orphanet_399 (Huntington's disease), and EFO_0000253 (Amyotrophic Lateral Sclerosis). For all NDD but HD, a child term was given, which was also included in the data acquisition step. In total, 153 studies with genomic data were used for the analyses (Maraganore et al., 2005; Fung et al., 2006; Coon et al., 2007; Reiman et al., 2007; Schymick et al., 2007; van Es et al., 2007; Van Es et al., 2008, 2009; Abraham et al., 2008; Bertram et al., 2008; Cronin et al., 2008, 2009; Li et al., 2008, 2021; Webster et al., 2008; Beecham et al., 2009, 2015; Carrasquillo et al., 2009; Harold et al., 2009; Lambert et al., 2009, 2013a,b; Landers et al., 2009; Latourelle et al., 2009; Pankratz et al., 2009, 2012; Satake et al., 2009; Simón-Sánchez et al., 2009; Edwards et al., 2010; Feulner et al., 2010; Hamza et al., 2010; Heinzen et al., 2010; Laaksovirta et al., 2010; Naj et al., 2010, 2011; Seshadri et al., 2010; Shatunov et al., 2010; Stein et al., 2010; Antúnez et al., 2011; Do et al., 2011; Furney et al., 2011; Hollingworth et al., 2011, 2012; Hu et al., 2011, 2016; Kim et al., 2011; Kramer et al., 2011; Liu et al., 2011, 2021; Logue et al., 2011; Nalls et al., 2011, 2014, 2019; Saad et al., 2011; Spencer et al., 2011; Wijsman et al., 2011; Chung et al., 2012, 2018; Cummings et al., 2012; Kamboh et al., 2012a,b; Kwee et al., 2012; Lill et al., 2012; Meda et al., 2012; Cruchaga et al., 2013; Davis et al., 2013; Deng et al., 2013; Jonsson et al., 2013; Langefeld, 2013; Martinelli-Boneschi et al., 2013; Miyashita et al., 2013; Reitz et al., 2013; Diekstra et al., 2014; Fogh et al., 2014, 2016; Goris et al., 2014; Hill-Burns et al., 2014, 2016; Kauwe et al., 2014; Nelson et al., 2014; Pérez-Palma et al., 2014; Ramanan et al., 2014; Ramirez et al., 2014; Sherva et al., 2014, 2020; Vacic et al., 2014; Xie et al., 2014; Hirano et al., 2015; McLaughlin et al., 2015; Tosto et al., 2015; Wang et al., 2015, 2021; Biernacka et al., 2016; Chen et al., 2016; Deming et al., 2016; Herold et al., 2016; Jun et al., 2016, 2017; Pickrell et al., 2016; Schott et al., 2016; Traylor et al., 2016; Van Rheenen et al., 2016; Watanabe et al., 2016; Benyamin et al., 2017; Chang et al., 2017; Foo et al., 2017, 2020; Lee et al., 2017; Mez et al., 2017; Moss et al., 2017; Siitonen et al., 2017; Sims et al., 2017; Chao et al., 2018; Gusareva et al., 2018; Marioni et al., 2018; Miron et al., 2018; Mukherjee et al., 2018; Nicolas et al., 2018; Pottier et al., 2018; Wallen et al., 2018; Yashin et al., 2018; Bandres-Ciga et al., 2019; Blauwendraat et al., 2019, 2020; Dekker et al., 2019; Jansen et al., 2019; Kunkle et al., 2019, 2021; Lo et al., 2019; Moreno-Grau et al., 2019; Nazarian et al., 2019a,b; Wei et al., 2019; Zhu et al., 2019; Cha et al., 2020; Hong et al., 2020; Nakamura et al., 2020; Ryu et al., 2020; Alfradique-Dunham et al., 2021; Bone et al., 2021; DeMichele-Sweet et al., 2021; de Rojas et al., 2021; Kang et al., 2021; Loesch et al., 2021; Park et al., 2021; Reddy et al., 2021; Rodrigo and Nyholt, 2021; Sakaue et al., 2021; Schwartzentruber et al., 2021; Shigemizu et al., 2021; Smeland et al., 2021; Tan et al., 2021; Wightman et al., 2021).

For the transcriptomic and proteomic layer, we used the studies that we have identified previously (Ruffini et al., 2020), using the Gene Expression Omnibus (GEO) (Barrett et al., 2013), the Expression Atlas (Papatheodorou et al., 2018) databases, and doing literature research (Ruffini et al., 2020). For the transcriptomic data, the keywords used in the GEO database were <name of disease> AND (“microarray” OR “RNAseq”) AND “human,” The Expression Atlas was used in Release 35 (May 2020, https://www.ebi.ac.uk/gxa/home) and scanned for Alzheimer's, Parkinson's, Huntington, and Amyotrophic Lateral Sclerosis, using the filter “Homo sapiens” in the section “Differential Experiments.”

For the proteomic data, literature search was conducted in PubMed and Google Scholar, with the keywords: (“neurodegenerative diseases” OR “Alzheimer's^* disease” OR “Parkinson's^* disease” OR “Huntington^* disease” OR “Amyotrophic Lateral Sclerosis”) AND (proteomics OR “quantitative proteomics” OR “differentially expressed proteins” OR biomarkers) AND human NOT mice for publications from 2010 to 2020.

Of the included experiments, 63% of the transcriptomic experiments were performed with brain material, while 67% of the proteomic experiments were conducted with brain material. Among the non-control patients, 84% were classified as having a moderate or severe condition in the transcriptome experiments and 90% in the proteome experiments. In total, transcriptomic data of 39 studies (Blalock et al., 2004, 2011; Zhang et al., 2005; Dunckley et al., 2006; Lesnick et al., 2007; Liang et al., 2007; Scherzer et al., 2007; Simunovic et al., 2009; Cox et al., 2010; Elstner et al., 2011; Dumitriu et al., 2012, 2016; Feyeux et al., 2012; Berchtold et al., 2013; Hokama et al., 2014; Riley et al., 2014; Calligaris et al., 2015; Dijkstra et al., 2015; Labadorf et al., 2015; Magistri et al., 2015; Prudencio et al., 2015; Raman et al., 2015; Ring et al., 2015; Kapeli et al., 2016; Lin et al., 2016; Scheckel et al., 2016; Lim et al., 2017a,b; Gagliardi et al., 2018; Mehta et al., 2018; Stopa et al., 2018; Mathys et al., 2019; Meyer et al., 2019; Otake et al., 2019; Swindell et al., 2019; Switońska et al., 2019; Al-Dalahmah et al., 2020; Dols-Icardo et al., 2020; Higginbotham et al., 2020) and proteomic data of 22 studies were acquired (Fang et al., 2009; van Dijk et al., 2012; Varghese et al., 2013; McQuade et al., 2014; Riley et al., 2014; Collins et al., 2015; Dumitriu et al., 2016; Hondius et al., 2016; Ratovitski et al., 2016; Lachén-Montes et al., 2017, 2019; Seyfried et al., 2017; Umoh et al., 2018; Zhang et al., 2018; Higginbotham et al., 2019, 2020; Iridoy et al., 2019; Bader et al., 2020; Johnson et al., 2020; Oeckl et al., 2020; Rotunno et al., 2020; Wingo et al., 2020) (see also Table 1). A table giving the number of samples per condition, the severity state of the included samples, and the used technology is given in Supplementary Table 4.

For the methylomic data, we used all data from the PDMethDB (Wang C. et al., 2020) (last accessed December 13, 2021), the EWAS Atlas (Li M. et al., 2019) (accessed 01.02.21) for the four respective diseases, and other data we found through literature research using PubMed and Google Scholar. Queries for “epigenome-wide” OR “epigenome wide” OR “EWAS” OR “genome-wide” AND “methylation” OR “genome wide” AND “methylation,” as well as “metastudy” AND “neurodegenerative disease” AND “epigenetic,” were used to find studies published after 2010. Combined, the databases and literature research resulted in the inclusion of data of 20 studies (Bakulski et al., 2012; Kaut et al., 2012; Masliah et al., 2013; de Jager et al., 2014; Lunnon et al., 2014; Sanchez-Mut et al., 2016; Watson et al., 2016; Zhang et al., 2016; Young et al., 2017, 2019; Gasparoni et al., 2018; Smith et al., 2018, 2019; Zadel et al., 2018; Altuna et al., 2019; Li P. et al., 2019; Tarr et al., 2019; Go et al., 2020; Lu et al., 2020; Marshall et al., 2020), one of which contained data for both AD and PD. As the higher availability of methylome data for Alzheimer's and Parkinson's diseases made it possible, we chose to include exclusively brain tissue-derived data for those two diseases. However, as the data availability for HD and ALS were much poorer, only blood tissue data were included for those NDDs. In total, we gathered data from 234 different studies for the four respective diseases and omics-layers (see Table 1).

The raw gene lists obtained from the transcriptomic, proteomic, and methylomic data were filtered for a false discovery ratio (FDR) < = 0.05, the proteomic data were mapped to gene names, and all the obtained gene names (from genomic, transcriptomic, proteomic, and methylomic data) were then mapped to the list of current protein-coding gene symbols (22 March 2021) from the HUGO Gene Nomenclature Committee (Tweedie et al., 2021) by mapping all alias names to the original gene symbol. Those names that were not found as any gene symbol or alias were discarded. The following data management step consisted of removing those genes that appeared as significantly altered only once in our transcriptomic, proteomic, and methylomic data. However, since the database for our methylomic and proteomic data in ALS and for the proteomic data in HD was very sparse, we decided to retain all significant genes for these gene lists to interpret a potentially noisy signal instead of hardly any signal at all.

The transcriptomic, proteomic, and methylomic data also contained information about the regulation of these significant genes/proteins. This information was contained as G-fold change (FC), log2 FC or other in transcriptomic and proteomic experiments, and as beta values in some methylomic experiments. To make this information from the different data sources as comparable as possible, we derived the mean direction of regulation (MeanRegDir) for all significantly conspicuous genes in each of the three omics levels, which equals to 1 if a gene/protein was overexpressed/hypomethylated in the disease group in every instance in which it occurred. Conversely, MeanRegDir indicates a value of −1 if a gene in the disease group was downregulated/hypomethylated in all experiments in which it appeared as significantly striking. If the signal is ambiguous, MeanRegDir will give a value between −1 and 1, as indicated in Equation (1).

Equation 1: Calculation of mean regulation of direction

with

By intersecting the four analyzed omics-layers per disease, we were able to test whether the number of shared genes between each pair of omics-layers per disease was significantly increased. We used a hypergeometric test for the overlapping sets, with the total amount of 19,207 gene symbols of protein-coding genes (HUGO Gene Nomenclature Committee, 22 March 2021) (Braschi et al., 2019) as the total population. To also have a measure for the effect size of the significantly high numbers of genes in the overlapping sets and thus being able to differentiate between the differently significantly found large gene sets, we calculated the effect size of the hypergeometric test results by setting the difference between observed and expected value in relation to the standard deviation of the hypergeometric distribution. This led to the following formula and the value eff_size that gives the number of standard deviations by which our expected value is beyond/below the expected value:

Equation 2: Effect size of the hypergeometric test results

with:

with:

Intersections were performed and visualized using the R (version 4.0.2) (Urbanek et al., 2014) package venn (version 1.10) (Dusa, 2016).

The significantly distinctive genes and proteins that emerged in our analysis represented a variety of processes connected to the respected NDD. Consequently, the functional analysis of all genes that emerged for an individual omics-layer for a particular NDD would lead to results showing a combination of different related processes and would, therefore, lower the statistical power of the hypergeometric test and interpretability of the results.

To better separate the distinct processes related to our gene sets, we constructed protein-protein interaction (PPI) graphs for each omics-layer of each NDD and isolated communities of high modularity within that graph with the louvain algorithm (Blondel et al., 2008). These communities of high modularity were then used for the functional analyses and showed clearly distinct processes for our different gene sets.

The intersection of the four analyzed omics-layers per disease showed significantly high numbers of shared genes between the transcriptomic and proteomic levels for all four analyzed NDDs (Figure 1, Table 2). All other pairwise intersections show a significantly high number of shared genes in two to three of the analyzed NDDs, while AD shows significantly high numbers of shared genes for all pairwise intersections of omics-layers. AD also showed the highest number of genes in total across all omics-layers with 14,970 genes. PD followed with 6,026, and then HD with 5,795 and ALS with 2,725 genes (Figure 1). The mean direction of regulation was close to 0 for all omics-layers in AD (0.13, −0.1, −0.06 for transcriptomic, proteomic, and methylomic data, respectively); in PD, the mean direction of regulation was −0.59 for the transcriptomic level, 0.42 for the proteomic level, and 0.33 for the methylomic level. For HD, the methylomic layer showed a mean direction of regulation of −0.8, while the other omics-layers were closer to 0 (0.04 transcriptomic, −0.18 proteomic). In ALS, the transcriptomic and proteomic layers showed a positive mean direction of regulation (0.12 and 0.16), while the methylomic layer showed a strong negative mean direction of regulation (−0.75). For the genomic level, there are no mean directions of regulation as these data show an association between the presence of specific SNPs and the respective NDDs. The calculated effect sizes for the hypergeometric test results show the highest values in the transcriptomic-proteomic intersection for all NDDs, with differences between expected and observed numbers of genes of more than 10 standard deviations (stds) except for PD (second highest value with 9.9 stds). In PD, the highest effect size was achieved in the transcriptomic-methylomic intersection with 13 stds difference. All other comparisons showed much weaker effect sizes, with values below 5 stds.

The correlation between the transcriptomic and proteomic mean direction of regulation of the genes in the transcriptomic and proteomic layers' intersection in AD, HD, and ALS was significantly high (all p < 10⁻⁵) (see Figure 2). The correlation for this layer intersection in PD was not significantly high. However, the correlation was slightly positive, even though the mean direction of regulation for the whole transcriptomic PD data was highly negative (−0.59), while it was highly positive (0.42) for the whole proteomic PD data. All other correlations of mean direction of regulation in the pairwise intersections between transcriptomic, proteomic, and methylomic layers per disease were neither significantly high nor low (p ≥ 0.05).

The network construction step led to a PPI network for each analyzed omics-layer in all four analyzed NDDs. However, the constructed communities of the networks based on the genomic layer in HD and ALS, as well as the methylomic layer in ALS, showed a little functional signal. Consequently, we focused on the remaining 13 PPI networks for the analysis.

These remaining networks varied between 7 (PD proteome) and 15 (PD methylome) communities with sufficient functional annotation (any biological process with FDR < = 0.05) to be further considered in our analysis. The modularity of these 13 PPI networks varied after community creation between 0.484 (PD proteome) and 0.588 (ALS proteome), thus showing a small deviation in modularity values across the different NDDs and omics-layers. The modularity value is positive if the communities in the network show higher interconnectedness than randomly expected; negative values if there are less connections within the communities than randomly expected are strictly lower than 1 (Equation 3: Modularity). Thus, the 13 networks that we further consider show more connections within their communities than expected by chance (as all have positive modularity values). Examples of those network plots for PD are given in Figures 3–6. The networks for AD, HD, and ALS can be found in the Supplementary Presentation 1—Network Figures.

The hallmarks of neurodegenerative diseases that we defined based on the functional analyses of the communities are a deregulated Cell Cycle, defects in autophagy and apoptosis on the cellular level, a defective Extracellular Matrix Organization and misguided cell and organ development, the impaired Signal Transduction/Transport of vital molecules within and between cells within the local tissue environment, an excessive Immune System response, and defective Metabolic Processes in the whole systemic environment.

The results of the X²-test showed that there is a significant difference between the expected and observed values for the four omics-layers across all six hallmarks (p < 0.05), but not for the four NDDs across all six hallmarks (see Supplementary Table 3—Low-Level Term Associations).

The most striking results we found when analyzing the accumulated numbers of associations between our communities and low-level terms across the four NDDs and also across the four omics-layers were the low number of associations found in PD for the Immune System hallmark and the low number of associations in the methylomic layer in the Metabolic Processes hallmark while having a high number of associations in the proteomic layer.

A more detailed description of the results is presented in the following;

The intersection of genes found in genomic, transcriptomic, proteomic, and methylomic data per disease showed that, while the number of shared genes between the transcriptomic and proteomic level was significantly high in all four investigated NDDs, the significance of the intersections was more heterogenous in the other comparisons.

Only AD showed a significant number of shared genes in all pairwise intersections of the four analyzed omics-layers. The fact that ALS methylomic and HD genomic data consisted of only two studies and thus had a very small database (61 and 35 genes) may have contributed to the non-significance of pairwise comparisons that considered the methylomic level for ALS and the genomic layer for HD. However, even though the database was much broader for HD methylomic data (885 genes), these genes also stem from only two studies, but showed significant overlaps in all pairwise intersections, except for the one with genomic data. Summarizing, especially those pairwise comparisons with a low number of expected overlapping genes, seemed to not show a significantly large pairwise overlap. All pairwise intersections that are based on two large groups of genes and, therefore, showed an expected number of shared genes > = 10 also showed a significantly high number of overlapping genes in these hypergeometric tests (Supplementary Table 2—Hypergeom Test Per Disease).

The fact that the effect size was highest for the transcriptomic-proteomic intersection in AD, ALS, and HD and, also, very high in PD (9.9) seems as if these layers shared the most striking relation in terms of overlapping genes. However, in PD, the effect size is even higher between the transcriptomic and methylomic layers. PD is also the only disease in which not proteomic and transcriptomic layers showed the largest database in terms of including the most genes but the transcriptomic and methylomic data did. Thus, in all NDDs, especially those intersections, showed the largest effect sizes that were based on the omics-layers, with the most genes per NDD. The statement that the transcriptomic and proteomic layers of our data showed high significantly enriched numbers of genes in their overlap in all analyzed NDDs holds true; however, it is, to this point, unclear if the high effect sizes in contrast to the other intersections are due to biological facts or to the broader database for these omics-layers in our data.

A more surprising point is the fact that the mean direction of regulation was vastly unequal to zero for both, the transcriptomic and the proteomic data in PD. This was also the case for the methylomic layer of HD and ALS, but, given that the transcriptomic and proteomic layers of PD are based on much more studies (11 and 7, while 2 studies for HD and ALS methylome), this high absolute mean direction of regulation, in combination with the dichotomy of these values, constitutes an interesting finding. The finding of a mainly (70%) negative direction of regulation for PD transcriptomic differentially expressed genes was already described in a meta-study, focusing only on microarray data (Kelly et al., 2019). However, the reason for this phenomenon remains unclear.

The finding of significant overlaps between the transcriptomic and proteomic data for only three out of four NDDs seems to be surprising and to hint at an extraordinary phenomenon in the posttranscriptional or translational processes in PD. However, given that the mean direction of regulation for the PD transcriptomic data was highly negative (−0.59), while the PD proteomic data were, overall, highly positive (0.42), it is striking that the genes in the intersection of these two omics-layers are still positively correlated, even though not significantly. Thus, this result might show a signal for some altered posttranscriptional behavior in PD in contrast to the other NDDs, but this could also be an artifact of the striking negative mean direction of regulation in the transcriptomic and the similarly strong positive mean direction of regulation in the proteomic data that were discussed earlier.

However, even the highly significant correlations between the transcriptomic and proteomic data in AD, HD, and ALS reached values of r = 0.283 (AD) to r = 0.344 (ALS), which might seem surprisingly small, following James Watson's simplified version of the central dogma of molecular biology (Watson et al., 1988).

Even though proteins are produced from mRNA, the correlation between mRNA levels and protein abundances is moderate. This non-trivial mRNA–protein relation is a general phenomenon, ranging from yeast to humans (De Sousa Abreu et al., 2009), and can have biological or, simply, technical reasons. Biological reasons can be found both on the posttranscriptional and on the translational levels, where many different mechanisms enhance or repress the synthesis of proteins from a certain copy number of mRNA molecules (Crick, 1958; Varshavsky, 1997; Fire et al., 1998; Kozak, 1999; Glickman and Ciechanover, 2002; Pillai et al., 2007; Kurreck, 2009; Maier et al., 2009; Chursov et al., 2013; Komar, 2016; Liu et al., 2016; McShane et al., 2016; Lau et al., 2018).

We attempted to limit the data included in this study to those derived from brain tissue whenever possible. For this reason, and because of the generally greater availability of data, we analyzed data in this study mainly from patients in the late stages of their respective NDD. This results in better comparability and less noise within the data but should be considered in the analysis and conclusions drawn from these studies, as effects occurring mainly in early stages of disease progression may not be apparent in this analysis.

It should also be noted that we included only data of studies that were published and, furthermore, available in PubMed, Google Scholar, the GWAS Catalog (Buniello et al., 2019), the GEO database (Barrett et al., 2013), the Expression Atlas (Papatheodorou et al., 2018), PDMethDB (Wang C. et al., 2020) or the EWAS Atlas (Li M. et al., 2019). Therefore, as pointed out in Pan et al. (2005), this introduces some language bias, as our methodology did not include searching for results published in other than the English language in, e.g., local journals, which was shown to influence the strength of p-values between, e.g., Chinese and non-Chinese studies (Pan et al., 2005; Tang, 2005). Also, we did not consider unpublished as well as unsignificant data. However, as our approach was mainly aimed for giving a comprehensive and broad overview and, therefore, mainly focused on those genes that were found as significant in, at least, two experiments, regardless of the exact strength of the actual p-value, we consider this problem of publication bias as a minor one for our approach as the number of overexaggerated and false-positive disease associations should be strongly reduced by this methodology.

For some approaches used in this study, it was necessary to choose between different possible approaches, e.g., for the hallmarks to be used or the definition of hub genes. In addition, the network creation of the louvain algorithm is non-deterministic and, therefore, subject to a random influence in the creation of the networks. To minimize this random influence, we applied the louvain algorithm with the input shuffled 100 times anew for each of the analyzed networks, and then chose the iteration that produced the highest modularity. With this approach, we found that the modularity between the different iterations did not change much; however, taking the best of 100 runs makes the results of the louvain community creation more reliable.

The definition of hallmarks and hub genes involves some flexibility, so we have attempted to define hallmarks in light of previous definitions of hallmarks for neurodegeneration (e.g., Ramanan and Saykin, 2013) and of our actual observed subprocesses in the functional analyses. As there are myriads of centrality measurements for nodes in a network, we took centrality measurements into account that were connected to the number of direct links between a node and other nodes, the mediation role of a node in a network, and to the sum of the shortest paths to all other nodes in the network, respectively (Freeman, 1977, 1978; Freeman et al., 1979). In addition, we chose to include the eigenvector centrality of each node (Newman, 2010; Bonacich, 2015), which gives a centrality score to each node that is proportional to the sum of the centrality scores of each node's neighbors to cover some varieties of ways for defining centrality of a node in a network.