Mild cognitive impairment (MCI), a transitional and intermediate state between normal aging and Alzheimer’s disease (AD), however, may have a significantly higher risk of converting to probable AD than the normal population. The conversion rate of MCI patients to AD was at an average of 10–17% per year (Petersen et al., 2009; Landau et al., 2010; Davatzikos et al., 2011) and approximately 60% within 10 years (Mitchell and Shiri-Feshki, 2009). A recent follow-up study even reported that the majority (45.5%) of those MCI individuals subsequently developed AD for an average of 26.6 months (Espinosa et al., 2013). The high risks make it highly important to involve the early prodromal stage, especially MCI, in exploring neurobiological and pathogenic substrates of AD.

Gray matter volume (GMV) atrophy is one of the main cardinal signs of neurodegeneration in AD and is irreversible. There is a long preclinical stage of AD, in which no obvious symptoms but subtle structural changes in specific brain regions can be detected (Tondelli et al., 2012). In the early stage (MCI), marked localized atrophy could occur in many cortical regions and certain sub-cortical regions. During the progression from MCI to AD, global and local GMV atrophy was reported mainly in the temporal neocortex, parahippocampal cortex, and cingulate gyrus (Spulber et al., 2012). Subsequently, this atrophy spreads aggressively to affect most of the brain in clinical AD (Pini et al., 2016). Although the neurobiological and pathogenic substrates underlying particular structural changes across AD spectrum have been investigated using MRI-based genetic study, such as the associations between genetic variations within PARP1 and CARD10 and a more rapid rate of hippocampal volume loss (Nho et al., 2013a,b), between the TREM2 variant and fronto-basal gray matter loss (Luis et al., 2014), between APOE and longitudinal change in the hippocampus (Apostolova et al., 2014; Andrawis et al., 2012), between expression level of ABCA7 and GMV changes in post-central gyrus, between superior frontal gyrus and ZCWPW1, and between right post-central gyrus and APOE (Roshchupkin et al., 2016) were identified, much more AD risk variants have been reported (Guo et al., 2017; Lacour et al., 2017; Shen and Jia, 2016). For example, a large meta-analysis on GWAS involved 74,046 people and identified 20 risk genes and 11 new susceptibility loci associated with AD (Lambert et al., 2013). Since a fraction of MCI patients are in the pre-AD stages, AD risk alleles, as well as additional genetic factors specifically influencing MCI progression, have received extensive attention (Moreno-Grau and Ruiz, 2016). Indeed, some major pathogenic genes were identified in MCI using open gene expression data sets (Tao et al., 2020). Recently, a meta-analysis also revealed several abnormally regulated genes, shared pathways, and transcription factors in MCI and AD (Bottero and Potashkin, 2019). However, one previous study identified that gene expression patterns in MCI are neither an extension of aging nor an intermediate between aged controls and AD (Berchtold et al., 2014). In the Chinese population, the SORL1 genetic variants, especially polymorphism rs985421, were identified to reduce the risk of converting from MCI to AD (Gao et al., 2014; Jin et al., 2014). Considering this evidence, it is really important to investigate whether structural changes in the AD spectrum were driven by similar gene variants.

Recently, the Allen Human Brain Atlas (AHBA¹) microarray dataset provided an indirect way for relating brain-wide transcriptomic data to neuroimaging data (Arnatkeviciute et al., 2019). The practical pipeline has been verified in various brain disorders, such as major depressive disorders (Anderson et al., 2020; Tan et al., 2021; Ji et al., 2021), schizophrenia (Liu et al., 2019), and migraine (Eising et al., 2016). Moreover, the expression level of genes involved in mitochondrial respiration and metabolism of proteins was found to be associated with regional GMV patterns across AD-memory, AD-executive, AD language, and AD-visuospatial subgroups (Groot et al., 2021). Therefore, it is ideal to investigate the relationship between transcriptional data and GMV changes in prodromal and clinical AD using this method, which will open up a new view to advance our understanding of the biological mechanisms underlying structural changes in AD.

In the current study, major steps were performed according to the schematic summary of the processing pipeline (Figure 1): (1) GMV changes were analyzed using two-sample t-tests for each patient’s group compared to NC; (2) gene expression levels were obtained from the AHBA data and processed using the new pipeline; (3) regional GMV changes and regional gene expression level for interesting genes were calculated for each sample locations; (4) cross-sample spatial correlations between gene expression levels and GMV alterations were performed using partial least square regression (PLS) for each group; (5) obtain consistent genes among three group; (6) Spearman correlations between gene expression levels of consistent genes and GMV changes were performed for each group; and (7) functional enrichment analysis was conducted using Metascape analysis² to explore ontological pathways of the consistent genes.

Although GWAS, whole-exam sequencing approaches, and meta-analyses showed many AD risk genes, little was known about which genes are associated with GMV changes in the AD spectrum. To narrow this gap, we explored structural changes across the full clinical AD spectrum and performed spatial correlations between these changes and the expression level of AD risk genes. As a result, we identified significant volume atrophy in the left thalamus, left cerebellum, and bilateral middle frontal gyrus across the AD spectrum. These structural changes were consistently associated with expression levels of 12 genes for all three groups and all four sub-groups, which were mainly involved in cellular protein localization, cellular component organization, and regulation of iron transport.

As expected, we identified significant volume atrophy in the left thalamus, left cerebellum, and bilateral middle frontal gyrus across the AD spectrum, suggesting a consistent structural changing pattern. These results could also be used as potential biomarkers for early diagnosis of AD. Moreover, our results were consistent with previous studies reporting GMV reduction in the AD spectrum (Pini et al., 2016; Spulber et al., 2012). A study suggested that GMV reduction might be associated with various micro-structural factors, such as alterations in size, morphology, and number of the cellular and non-cellular components, as well as microglial cells in the cerebral cortex and sub-cortical nuclei (Ji et al., 2021). Indeed, such microstructural changes were widely reported in studies on AD (Davies et al., 2017; Gasparoni et al., 2018; Ishunina et al., 2019; Nicastro et al., 2020). However, the genetic mechanisms resulting in these microstructural changes and GMV reduction remain largely unknown. Moreover, as we know, no direct relation was identified since there is no such large sample of patients with AD who simultaneously have brain-wide transcriptomic and neuroimaging data. Although it is an indirect method of assessing the similarity of spatial distribution patterns between them, our results offered a better understanding of biological mechanisms underlying structural changes in prodromal and clinical AD.

All gene expression values in this study were positive, but the t-statistic values were negative (indicating reduced GMV) or positive (indicating increased GMV). Thus, the negative correlations indicated that brain regions with greater GMV reduction showed higher gene expression, and positive correlations meant that brain regions with greater GMV reduction showed lower gene expression. For instance, we identified the strongest negative association between gene expression of ECHDC3 and GMV changes, whereas the strongest positive association between gene expression of SORL1 and GMV changes across the AD spectrum. The ECHDC3 is a gene that mainly encodes the enzyme enoyl-CoA hydratase domain containing 3, which has been found to be associated with brain neurodegeneration (Tan et al., 2021), especially in AD (Desikan et al., 2015). Moreover, a previous study showed that the pleiotropy at ECHDC3 may be related to the association finding at this locus among persons lacking the APOE ε4 allele (Jun et al., 2017). The SORL1, belonging to both the low-density lipoprotein receptor family and the vacuolar protein sorting-10 domain protein family (Lane et al., 2010), is a key protein involved in the processing of the amyloid-beta (Aβ) precursor protein and the secretion of the Aβ peptide (Campion et al., 2019). It has been observed with a deficiency in the brains of patients suffering from MCI (Sager et al., 2007) and AD (Shen et al., 2014; Dodson et al., 2006) and was supported by results from meta-analyses (Campion et al., 2019). Moreover, the SORL1 genetic variants were reported to modulate or confer the risk of aMCI to probable AD in the Han Chinese population (Chou et al., 2016). Similar to our results, one study also showed that risk variants in SORL1 were associated with less gray-matter tissue in sub-cortical regions, such as the putamen, thalamus, and pallidum (Roshchupkin et al., 2016).

Most importantly, the gene expression level of APOE showed a significantly negative association with GMV changes across the AD spectrum. However, the gene expression levels of APOE sub-types were not available in this study, which mainly included three major polymorphic alleles in humans, namely APOE2, APOE3, and APOE4. Among them, APOE4 remains by far the strongest and most prevalent genetic risk of AD since it has a great influence on two hallmark pathological proteins by modulating the formation of amyloid-β peptide (Aβ) plaques and neurofibrillary tangles containing hyperphosphorylated tau protein (Serrano-Pozo et al., 2021). Similar to our results, many previous studies report significant independent effects of APOE4 genotype on hippocampal volume in MCI and AD (Wang et al., 2015; Saeed et al., 2018; Veldsman et al., 2021; Andrawis et al., 2012), especially in those who progressed to AD (Fang et al., 2019). Other studies described the effects of APOE4 on CA1 (Kerchner et al., 2014), CA3/DG (Mueller and Weiner, 2009), and subiculum (Donix et al., 2010) at sub-regional level. Mean adjusted hippocampal atrophy rates in APOE4 carriers were significantly higher in MCI converter, MCI stable, and AD compared with non-carriers (Manning et al., 2014). In addition to hippocampal atrophy, GMV loss in temporal and parietal lobes, right caudate nucleus, insula, right parietal operculum, the right precuneus, and the cerebellum bilaterally were also reported in APOE4 carriers in patients with MCI (Goni et al., 2013; Spampinato et al., 2011). Moreover, APOE4 has been shown to modify the association between cerebral morphology and cognitive performance in healthy middle-aged individuals (Cacciaglia et al., 2019) and between smaller volumes in the left hippocampus and a tendency to retrieve earlier acquired words in the category fluency task in MCI (Venneri et al., 2011).

In addition to APOE, we also identified that gene expression levels of CD33 and ECHDC3 were negatively associated with GMV changes in MCI and AD. Together, these genes were primarily involved in the positive regulation of iron transport. As we all know, iron is essential for neurons and glia in many aspects (Reinert et al., 2019), such as electron transport, reductase activity of nicotinamide adenine dinucleotide phosphate (NADPH), and myelination of axons. It is available to neurons and glia by transporting from the basolateral membrane of endothelial cells to the cerebral compartment (McCarthy and Kosman, 2015). Although the direct relationship between brain iron and AD remains largely unknown (Tripathi et al., 2017), improper iron transport mechanisms are speculated to lead to the accumulation in various cortical regions and the hippocampus in AD (Mezzaroba et al., 2019; Li et al., 2019). Once exceeding, it might produce reactive oxygen species and pro-inflammatory proteins (Yauger et al., 2019), which cannot be optimally handled in MCI and AD patients (Mills et al., 2010). Indeed, increased iron in the brain has been considered to be one of the primary causes of neuronal death in several neurodegenerative diseases, especially in AD (Lu et al., 2017). In addition, iron is able to induce tau and Aβ aggregation (Sayre et al., 2000) and enhance the toxicity of Aβ (Squitti, 2012). Previous AD models further support the interlinkage between iron metabolism and AD by showing that iron chelators may prevent neuronal loss (Ward et al., 2015). Therefore, it is reasonable to speculate that gray matter atrophy might be related to the dysregulation of iron transport, possibly iron deposit, and neuronal death.

Gene-related changes in GMV were also observed in different brain regions of AD mouse models. For example, a recent study found significant GMV reductions in several brain regions, including the insular cortex (left), basal forebrain (left), subiculum (right), and others, in AAV/APP transgenic mice (Lin et al., 2024). The 3 × Tg-AD mice model, which carries three human mutant genes, has also demonstrated GMV reductions, particularly in the visual cortex (Chiquita et al., 2019). Observations in amyloid beta precursor protein/presenilin 1 (APP/PS1) transgenic AD model mice revealed that the volume of the left hippocampus and right olfactory were reduced (Xu et al., 2024). The APOE gene, which was significantly associated with GMV changes in our study, has also been implicated in exhibiting phenotypes related to cognitive decline in mouse models (González et al., 2023). While the literature predominantly focuses on the APOE gene, the interplay between other genes and GMV in mouse brains remains less explored. Although current literature lacks direct evidence for some genes, we posit that investigating their functions and phenotypes in mouse models could yield significant insights.

Several major methodological limitations are worth mentioning in the current study. First, the gene expression level was calculated from six postmortem brains, whereas the neuroimaging data were obtained from the ADNI dataset. Moreover, we only used gene expression data from the left hemisphere since only two right hemisphere data were available in the AHBA. Therefore, a large sample across the AD spectrum with both brain-wide transcriptomic and neuroimaging data of the same individuals is needed to further verify our results. In future research, we plan to expand the scope of our gene set to include genes from NIAGADS, Alzforum, and ADGC databases, which will increase the likelihood of identifying significant correlations between gene expression and GMV. Second, although most previous studies used surface-based morphology to gray matter segmentation, several previous studies successfully adopted VBM and performed transcriptome-neuroimaging spatial correlation analyses to explore gene expression profiles associated with gray matter volume changes in epilepsy (36444721), children with persistent stuttering (34041495), schizophrenia (37607339), and Alzheimer’s disease (28,105,773, 27,718,423). Due to the algorithm potentially involving brain sites in close spatial proximity but not closely anatomically connected, its potential effects on our results cannot be excluded. Finally, the associations of these structural changes with alterations in the expression of specific genes are indirect and notably weak. Although corrections for multiple comparisons were performed, more attention should be paid to the interpretation of our results.

In conclusion, this exploratory study linked structural brain changes to gene expression levels by assessing the similarity of spatial distribution patterns. It showed eight genes positively associated and four genes negatively associated with GMV alterations across the AD spectrum, which were validated in four MCI subgroups. These genes were mainly enriched in biological processes related to cellular protein localization, cellular component organization, and iron transport regulation. Collectively, these findings provide a deeper understanding of the biological mechanisms underlying structural changes in both prodromal and clinical Alzheimer’s disease.