Disturbance of phylogenetic layer-specific adaptation of human brain gene expression in Alzheimer's disease

Alzheimer's disease (AD) is a progressive neurodegenerative disorder with typical neuropathological hallmarks, such as neuritic plaques and neurofibrillary tangles, preferentially found at layers III and V. The distribution of both hallmarks provides the basis for the staging of AD, following a hierarchical pattern throughout the cerebral cortex. To unravel the background of this layer-specific vulnerability, we evaluated differential gene expression of supragranular and infragranular layers and subcortical white matter in both healthy controls and AD patients. We identified AD-associated layer-specific differences involving protein-coding and non-coding sequences, most of those present in the subcortical white matter, thus indicating a critical role for long axons and oligodendrocytes in AD pathomechanism. In addition, GO analysis identified networks containing synaptic vesicle transport, vesicle exocytosis and regulation of neurotransmitter levels. Numerous AD-associated layer-specifically expressed genes were previously reported to undergo layer-specific switches in recent hominid brain evolution between layers V and III, i.e., those layers that are most vulnerable to AD pathology. Against the background of our previous finding of accelerated evolution of AD-specific gene expression, here we suggest a critical role in AD pathomechanism for this phylogenetic layer-specific adaptation of gene expression, which is most prominently seen in the white matter compartment.


Results
The present study was performed on the temporal cortex of three healthy controls (HC) and three age-matched patients with AD. High throughput RNA-sequencing was performed on samples spanning the entire cortical depth, i.e., comprising all layers, including subcortical white matter (ALL), and samples where the supragranular layers (SUP), infragranular layers (INF) or the subcortical white matter (SWM) were individually dissected from surrounding tissue. After removing the adapter sequence and low-quality reads, an average of ~ 107,166,000 reads per sample was obtained (Supplementary Table S1).
Layer-specific differences in gene expression within healthy controls and AD. To establish whether we could identify layer-specific differences in gene expression with our approach, we first analysed the differentially expressed genes (DEG) between SUP, INF and SWM separately for both HC and AD.
In HC, the number of DEG comparing SUP and INF was relatively small, amounting to a total of 15 (Table 1 and Fig. 1i), comprising protein-coding genes, lncRNAs and other non-protein-coding genes to about one third each. Expectedly, the amount of DEG was much higher between grey matter layers and SWM, amounting to a total of 634 and 62 comparing SUP with SWM and INF with SWM, respectively. In comparing INF with SWM, DEG classes comprise equal amounts of protein-coding genes, lncRNAs and other non-coding genes, while in the comparison between SUP and SWM, coding genes make up half of the total DEG.
In AD, the expression pattern throughout the cortex was more heterogeneous than in HC, resulting in a higher number of DEG for all comparisons made (Table 1 and Fig. 1ii). 121 DEG were identified comparing SUP with INF, an eightfold increase, while 2.135 and 788 DEG could be detected comparing SUP with SWM and www.nature.com/scientificreports/ genes, which together with the higher overall number of differences reflects a drastic increase in the number of differentially expressed coding genes. Of note, many of the 62 protein-coding DEG identified in the SUP with INF comparison were already previously linked to AD or reported in a brain layer-specific context, such as, e.g., AQP1 10,11 , CUX2 12-14 , HCN4 14 , PVALB 13 , RORB 12,13,15 , SLC17A6 16 , SYT2 17 or TPH2 18,19 (Table 2). Moreover, our analysis also shows that the differential expression of protein-coding genes between SUP and INF is paralleled by significant differences between SUP and SWM expression, while differences between INF and SWM are less pronounced.
Layer-specific differences in gene expression between healthy control and Alzheimer's disease. To better understand the physiological and AD-related pathophysiological characteristics of layer-specific gene expression, we next sought to compare the expression patterns between HC and AD. To first assess the overall magnitude of DEG between HC and AD, we started the comparison using samples comprising all cortical layers, including subcortical white matter (ALL). Only 17 DEG between HC and AD were identified in ALL ( Table 3). Out of these, 6 were downregulated and 11 up-regulated (Supplementary material Fig. SF1 and Supplementary Table S2). Amongst these DEG, the majority codes for proteins, while only 3 are annotated as pseudogenes, 1 lncRNA, 1 miRNA, and 1 transcript still to be confirmed (TEC). Thus, for all comparisons made, most DEGs comprise protein-coding genes, followed by lncRNAs ( Table 3).
The only lncRNA found differentially expressed was AC119673, which is located antisense to the PM20D1 gene, known as a quantitative trait locus (QTL) of AD and proposed as a potential blood-based biomarker for AD 20 . Among the protein-coding genes (Table 4), most were previously identified to be associated with AD, such as PTGER3, reported as an AD-associated hub gene 21 . SNPs in the LIPG gene are associated with AD and cardiovascular diseases 22 . In addition, KIF25 was found hypermethylated in AD 23 , HLA-DQB1 has alleles showing association to AD 24,25 , CHRM5, a cholinergic receptor, is up-regulated in AD 26 , HLA-DRB5 was identified as a risk gene for AD 25 , MTRNR2L12 has been proposed as a candidate blood marker of early AD-Like Dementia in adults with Down Syndrome 27 , and ADAMTS18 possesses secretase activity, a relevant feature in AD development/progression. Similarities and differences in layer-specific DEG in healthy control and Alzheimer's disease. To better characterise the specific effects of AD in each brain layer, we first identified those DEG that were found in both HC and AD comparisons (red dots in Fig. 1ii), a total of 521 DEG, with most of them coding for proteins. Out of all the DEG shared between comparisons, only the miRNA hsa-miR-3687 follows an oppo- Table 2. Protein-coding brain layer associated DEG in AD in SUP versus INF layers. I: increased in the first term of the comparison; D: decreased in the first term of the comparison; -: unchanged. Underline genes were also repeatedly associated with AD. $ : values referent to the comparison between SUP and INF. Despite the significant number of DEG in common, both HC and AD conditions still present distinct expression profiles when considering unique DEG ( Fig. 2 a and b). While most of the unique DEG are up-regulated in the HC brain SWM compared to ALL and the layers (SUP, INF) (Supplementary Tables S5, S7, and S8), in AD, the opposite holds, where DEG are mostly down-regulated in SWM compared to ALL and layers (SUP, INF) (Supplementary Tables S11, S13, and S14). Consolidated, the SWM layer of both HC and AD comprise more distinct profiles than SUP and INF.

Layer-specific biological pathways within healthy controls and AD. To identify physiological and
AD-related features of the layer-specific expression pattern, we next performed a Gene set enrichment analysis (GSEA) for each comparison made.
Concerning HC, only those comparisons that involved SWM showed significant GO terms (Supplementary  Tables S20 to S22). Here, we identified several Biological Processes (BPs) related to synaptic function, such as "signal release", "cell-cell signalling", "regulation of action potential", and brain functions such as "cognition" and "behaviour". Also, the comparisons involving the INF were related to BPs involved in secretion, such as "neurotransmitter secretion", "hormone secretion", "peptide secretion", and exocytosis, besides "learning" and "memory".
All those GO-terms found in HC could be replicated in AD, but additional BPs could be identified (Supplementary Tables S15 to S19 and Supplementary Tables S25 to S27). Four AD-specific GO clusters reproducibly present in all comparisons made (Supplementary Tables S25 to S27) were identified applying NaviGO 28 and Resnik's similarity approach 29 (Fig. 3). Presynaptic transport processes were clustered most prominently, with 'synaptic vesicle release' representing the most prominent cluster (Fig. 3A), which is in line with a common hypothesis that synaptic dysfunction is the major correlate of cognitive decline 30 , precedes neuronal cell death and is already detectable in pre-symptomatic stages of the disease 31,32 . Of note, recent data underline the early role of presynaptic vulnerability 33 in contrast to postsynaptic compartments. In addition, 'inner ear development' 34 was also detected as a cluster (Fig. 3D), supporting the growing evidence that several genes affecting proper inner ear development, e.g. alpha-synuclein (SNCA) 35 or beta-secretase (BACE1) 36 , are also involved in Table 3. Biotype of the DEG between HC and AD in each layer.  www.nature.com/scientificreports/  www.nature.com/scientificreports/ neurodegenerative disorders including AD. This could reinforce the proposal of hearing loss as an interplay of 'peripheral' and ' central' hearing dysfunction, which is linked to cognitive decline and has been attributed increasingly a modifiable risk factor for AD 37 . "Blood circulation" and "Circulatory system process" were found significant in four out of the six comparisons (Supplementary Table S23).
The most prominent GOs were identified in AD comparisons that involved SWM, amounting to an overlap of 79 terms, which comprised "synaptic vesicle exocytosis", "regulation of dendrite development", "sensory perception", "neurotransmitter secretion", "synaptic vesicle transport", "regulation of neuronal synaptic plasticity", "regulation of neurotransmitter levels", "neuron-neuron synaptic transmission", "membrane depolarisation", "regulation of neuronal synaptic plasticity", and "adult behavior". This clearly points towards a substantial synaptic dysfunction as a critical early event in AD as suggested previously 31,32 . Further, the term "demyelination" exclusively found in AD comparisons, was an additional indicator of fibre degeneration in AD 4 .
Next, we evaluated the expression patterns of the unique DEG in the most significant BP for each comparison. Although some pathways are present in more than one comparison, such as "extracellular matrix organisation" and "extracellular structure organisation" in both SUP with ALL and INF with ALL comparisons, they are represented by different genes, such as SFRP2 (in SUP with ALL) and MYH11, COL5A2, and COL8A1 (in INF with ALL), all up regulated in the ALL samples when compared to the others. This is also consistent with the cellular environment alterations caused by the accumulation of β-amyloid 38 .
The comparisons involving the SWM samples all show BP involved in signal transmission, such as "signal release", "regulation of synaptic transmission", and "generation of a signal involved in cell-cell signalling. Again, most of the DE genes in significant BP are downregulated. Interestingly, exceptions are the genes KLK6 and TNC, which are both up-regulated in SWM in the ALL with SWM comparison. The KLK6 gene is part of the "regulation of nervous system" pathway, while the TNC is part of the "neuron projection morphogenesis" and "axonogenesis". KLKL6 is suggested as a microglia marker 39 , and TNC is an extracellular protein involved in developing and repairing neural tissues 40 . In the SUP with SWM comparisons, the genes CD38, GAB2, NGFR, NTN1, PSEN1, P2RX7, PRAM1, and SEMA4D are all up-regulated in SWM and part of BP such as "exocytosis", "axon guidance", "learning memory", and "regulation of synaptic activity". Lastly, in the INF with SWM comparison, the gene STC2 is up-regulated in SWM and part of the "generation of a signal involved in cell-to-cell signalling" and "cation transport" BPs. These findings underscore the importance of SWM as a most critical compartment in AD pathogenesis.

Discussion
Brain regional vulnerability has been the subject of previous investigations 8, [41][42][43][44] . Some early studies examined gene expression in selected human neocortical laminae using microarrays to address aspects such as the corticocortical network architecture 45 , schizophrenia 46,47 or frontotemporal lobar degeneration 48 . More recent studies used RNA-sequencing to compare gene expression in other primates 49 and single-nucleus RNA-sequencing to obtain layer-enriched expression 50 . However, to the best of our knowledge, the aspect of layer-specific vulnerability has not been addressed for AD at the molecular level using RNA-sequencing, so far [45][46][47][48][49][50] .
Our present intention to take a layer-specific approach has also been stimulated by recent functional cognitive brain imaging developments allowing for ultra-high structural resolution, demonstrating layer-specific activity patterns in mental events 43 . Thus, with the present study, we hope to add a new molecular dimension to those differences that have already been established between supra-and infragranular cortical layers in AD concerning cytoarchitectonics, connectomics and vulnerability.
The comparison of gene expression between the temporal cortex of HC and AD on samples comprising all brain layers and subcortical white matter (ALL) allowed to detect a small number of DEG, with most of them already being reported concerning AD. Since the ALL-samples comprise different cell types such as neurons, including their processes, astrocytes and oligodendrocytes, only the most prominent differences beyond cell type and spatial location might be expected to show up.
In contrast, all comparisons between different layers, including SWM, revealed layer-specific DEG in the HC brain, which were even more frequent in the corresponding AD brain comparisons. Protein coding genes were found differentially expressed most often, followed by ncRNAs, with both together constituting a fraction of more than 55% and 88% of all DEG in HC and AD, respectively. Of note, the number of both coding and non-coding RNAs found differentially expressed is at least more than three times increased in the corresponding AD comparisons. This adds further evidence to a critical role of non-coding RNAs in AD pathomechanism, which remains an understudied aspect of the disease. Along this line, we recently could show that AD-associated protein-coding RNA but even more so, non-coding RNAs show accelerated evolution 51 , indicating a link between most recent hominid brain evolution and vulnerability towards AD. This aspect might well be linked to the process of phylogenetic development of a six-layered cortical structure.
Specific synapse-related GO terms are uniquely affected in AD. For HC, significant layer-specific GO terms were found for comparisons involving SWM, which are often related to synaptic function, including regulation of action potentials and signal release. In AD, additional GOs were identified; they form a network containing "synaptic vesicle transport", "synaptic vesicle exocytosis", and "regulation of neurotransmitter levels". In addition, the GO term "glutamate secretion" was represented prominently, which against the background of the critical involvement of subcortical white matter, as documented in the present study, and in agreement with previous findings, clearly identifies presynaptic changes of long axons, mainly glutamatergic neurons, as most early and most constant finding in AD pathomechanism 31 www.nature.com/scientificreports/ Subcortical white matter as a critical compartment in AD pathogenesis. For all comparisons made, most DEG were found in comparisons that involved SWM. While in HC, this might simply reflect genuine differences in cellular composition between grey and white matter, it indicates a prominent role for the subcortical fibre compartment together with associated cellular elements such as oligodendrocytes in AD pathology. While white matter changes have not been in the direct focus of AD research until recently, it becomes clearer that they represent an earliest and most critical aspect of the disease 44 that correlates very well with clinical/ cognitive traits of the disease 54 . Still, the phenotypic link between the progression of neurofibrillary changes in AD and the myelination patterns, suggesting some dysfunction in oligodendrocytes as a potential cause or early trigger of AD, had been reported by Braak and Braak quite some time ago 55 . A recent study examining AD cases without co-morbitities also points to an early environment of altered oligodendrocytes into AD pathology 56 .
Our present findings in the changes in RNA expression pattern, mainly associated with a reduction of transcript expression in SWM, might well be seen against the background of disease-associated changes in axonal transport systems and its potential involvement in the spreading of tau pathology, which, according to recent findings by Braak and Tredici 57 , takes a cortico-cortical top-down route. Moreover, it is in line with the recent result of relatively low white matter integrity in subjective cognitive decline (SCD), which is assumed as one of the earliest stages on the continuum towards AD 58 .
Our recent observation on accelerated evolution of AD-associated genes 51 and the recent findings of noticeable changes in the subcortical white matter is in direct agreement with a report on an about two-fold higher evolutionary rate of human white matter genes compared to pyramidal cell layers 49 . Human-specific organisation of layer-specific gene expression patterns might thus play an important role in the pathological process of AD development.
He and collaborators 49 and Zheng and collaborators 59 show that recent hominid brain evolution is associated with brain layer-specific switches in gene expression when comparing humans to chimpanzees. Two genes (ULBP2, MGAT5) out of the 18 reported by He and collaborators. (2017) that undergo a human-specific expression transition from layer V to layer III were detected in our AD specific DEG analysis. Furthermore, another five genes (CHRNB3, CNTNAP4, AQP1, and NGB), reported by this group as human brain layer-specific, were also found differentially expressed in our AD comparisons. In the work of Zhen and collaborators 59 , four out of the seven genes (CRY , PRSS12, SCN4B, and SYT2) show changes in expression patterns between layer V and layer III, when comparing mouse to human brain, were also detected amongst our AD-associated DEG. As layers III and V are the layers most typically involved in neurofibrillary degeneration, our data suggest a link between the layer-specific gene expression and its phylogenetic switch and AD pathology.
Though we did not find any differential expression of MAPT (microtubule associated protein tau) or APP (amyloid beta precursor protein), which gene products contribute to neurofibrillary tangles and Aβ plaque deposition in cortical AD layers, our data provide indications of layer specific modified expression of genes potentially affecting APP metabolism and Aβ generation. ADAMTS4 was increased in SUP in AD brain compared to INF as well as to SWM (Table 4). This gene codes for a disintegrin-like and metalloproteinase with trombospondin type 1 motif and can cleave Aβ peptide sequence between Glu-3 and Phe-4 60 . Coexpression of ADAMTS4 and APP resulted in HEK293 cells in the secretion of Aβ 4-40 peptides 60 supporting a contribution of this enzyme to the Aβ amyloid pathology in AD. Their elevation in SUP is paralleled by the most early formation of plaques compared to INF 9 . Moreover, the observed reduction of cortexin (CTXN3) in SUP of AD compared to INF (supplementary Table S12) could also contribute to the increased Aβ generation 61 in this AD brain layer. This data are well in agreement with the hypothesis, that layer specific alterations of gene expression contribute to specific AD pathology.
Evolutionary alteration of layer specific gene structure and expression might affect neuronal plasticity and allow for a higher degree of cellular individuality, resulting in a remodelling of layer specific function, paving the way to a wider spectrum of cognitive abilities, potentially at the expense of increased vulnerability.
This process might be closely linked to a shift or redefinition of neuronal identity, since several coding genes, differentially expressed between SUP and INF in AD (Table 2) define specific cell types. For example, PVALB is coding for parvalbumin, a marker for a subset of interneurons, TPH2 is coding for tryptophan hydroxylase, a marker for serotonergic neurons 62 , while RORB is known as a developmental driver of neuronal subtype identity in the neocortex 13,63 . Finally, alterations of CUX2 expression (Table 2), a SUP layer specific marker 64 , might be linked to neurodevelopmental disorders such as autism and schizophrenia 65 , or epileptogenesis 64 .
We used layer-specific RNA sequencing to understand better molecular correlations of systematic differences towards AD-specific vulnerabilities in this work. Our findings indicate the importance of non-coding genes in layer-specific physiology of the human brain and AD-pathology. Furthermore, the DEG found corroborate the findings of pre-symptomatic molecular alterations in white matter that might be intrinsic to AD's origin. Lastly, many AD-associated layer-specifically expressed genes were previously reported to undergo layer-specific switches in recent hominid brain evolution, thus suggesting an evolutionary pattern critical to the genesis and development of AD.

Methods
Human brain tissue. Brain tissue of 3 AD patients and 3 healthy controls (Table 5) [69][70][71] , and the NIA-AA guidelines for the neuropathological assessment of AD 55 72 , for Aβ/amyloid plaque score according to Thal et al. 7 and for neuritic plaque score according to CERAD 1 . Neurofibrillary tangles and Aβ/amyloid plaques were detected by immunocytochemical labelling of phospho-tau (anti-human PHF-tau monoclonal antibody AT8; Thermo Scientific) and Aβ (beta-amyloid monoclonal antibody, 6E10; BioLegend), respectively. The severity of AD pathology was scored following the consensus guidelines for the neuropathologic evaluation of AD according to Hyman et al. 17 and Montine et al. 73  Bioinformatics sample pre-processing. The adapter sequence and reads presenting sequencing quality lower than 20 were removed using TrimGalore! version 0.6.3 74 (https:// www. bioin forma tics. babra ham. ac. uk/ proje cts/ trim_ galore/). All comparisons were performed in R 4.0.0. 0n statistical environment, and all plots were done using the R packages ComplexUpset version 0.5.17 75