The systematic review of articles related to music and AD was carried out according to the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) guidelines. Indexed searches were performed in PubMed using the following query “music” AND “Alzheimer,” in title and abstract. The search yielded 323 papers; 217 were excluded after close inspection following the sequential criteria indicated in the PRISMA scheme (Figure 1; Supplementary Table S1); the reasons for exclusions included: duplicated articles, reviews and meta-analysis, articles not focused on AD or music and articles written in a non-English language, among others. We also disregarded a few articles dealing with neuroanatomical studies or music abilities of AD, because the focus of the present review was placed on the benefits and therapeutic effects of music (Figure 2). Of the 107 articles dealing with the therapeutical effect of music, a subset of them were related to music and memory (n = 47), an outcome particularly relevant to AD.

Moreover, a total of 13 additional relevant articles were added to the list by scrutinizing references list of the selected articles and reviews (Figure 1).

Because our findings on the ‘omic’ side include a connection between gene expression and dopamine, we also carried out a PubMed search for the query “music” AND “dopamine*” to investigate the state-of-the-art in this terrain. This search returned 63 papers, of which only 12 of them were retained after close inspection for relevance (see Supplementary Text to check the most relevant information regarding this search).

A total of 334 genes (Supplementary Table S2; Supplementary Figures S1, S2) were selected as candidates to be associated with musical traits. In a first approach, genes from the two main reviews on the matter were selected: Järvelä (2018) proposes a selection of top candidate genes for musical aptitude and music performance, while Navarro et al. (2021) carried out an exhaustive search in the literature related to music (mainly genetic association studies), and included genes that might be related to the impact of music on gene expression. This list was completed by revisiting the original papers studying the effect of music listening and music performance on gene expression (Kanduri et al., 2015a,b; Nair et al., 2019, 2020); and also other studies that, from wider perspectives, also aimed at connecting genes and musical traits. These include: (i) genes that fall in the top 20 regions identified by Liu et al. (2016) under an F_ST scrutiny method of the genome to find signatures of positive selection associated with musical aptitude; (ii) the top genes from Oikkonen’s study (Oikkonen et al., 2016) where convergent evidence for the molecular basis of musical traits was obtained through integration of gene-level data from 105 published studies; and (iii) genes related to the terms “music” or “musical” with relevant publications using geneshots;¹ the query “music*” was performed under “GENERIF” (manually collected gene-term association). From the results of those queries, and after careful manual checking, only those with relevant publications were retained (for instance those related to human studies, candidate genes, etc.)

The association between aberrant epigenetic modifications leading to dysregulation of gene expression and AD progression has been thoroughly investigated (Fenoglio et al., 2018; Stoccoro and Coppede, 2018; Nikolac Perkovic et al., 2021). We analyzed the panel of candidate genes related to music in connection to epigenomics in AD. We used the datasets from Nabais et al. (2021) (Gene Expression Omnibus [GEO]: GSE153712), which contain epigenetic data generated for the Illumina Infinium Human MethylationEPIC Beadchip (Illumina Inc., San Diego, CA) on 161 AD patients (91 females and 70 males), and 471 healthy individuals (272 females and 199 males).

First, the methylation values associated with each of the probes on the MethylationEPIC microarray were converted to Beta values, by calculating the ratio of methylated probe intensity over total intensity (methylated and unmethylated) for each probe. Samples and probes’ quality control was performed with the package minfi, which provides a quality control report on the basis of intrinsic control probes present in the array, in addition to allowing to remove probes and samples according to their signal intensity. Subsequently, to reduce the risk of measurement biases, raw Intensity Data files (*.idat) were filtered using RnBeads package. As usual, cross-reactive probes, probes located within three base pairs of common SNPs, probes with missing values or no variability in methylation, and those located on sex chromosomes, were removed. In the preprocessing step, RnBeads was also used to estimate sample donors’ sex based on their DNA methylation status – an important step to identify possible discrepancies between documented gender and biological sex. Background adjustment of the methylated and unmethylated intensities was performed using the Dasen method (Pidsley et al., 2013), while normalization of Beta values was carried out using the BMIQ normalization method (Teschendorff et al., 2013). To identify differentially methylated positions (DMPs), we used the limma package (Ritchie et al., 2015), which performs a linear model, adjusted for sex, to compare DNA methylation patterns between patients with AD and healthy controls. DMRcate package (Peters et al., 2015) was used to identify significantly differentially or variable methylated regions (DMRs). methylGSA package (Ren and Kuan, 2019) was employed to carry out the gene set testing and pathways analysis of genes associated with DMPs. This method faces the biggest challenge in performing gene set analysis, that is, assigning differentially methylated features to genes, and adjusting for the number of CpGs instead of gene length. For significant DMRs, we performed gene ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis, applying the functions enrichGO and enrichKEGG of the ClusterProfiler package (Wu et al., 2021).

P-values were corrected for multiple testing using the false discovery rate (FDR) method. Only DMPs with an FDR-adjusted p-value <0.05 were considered. DMRs were considered if having a minimum of three CpGs sites inside, and an FDR-adjusted p-value <0.05. The same thresholds were used also for gene-set and pathways analysis.

All the statistical analyses were carried out using R software (v.4.1.2).

We downloaded gene expression data from four independent microarray datasets located at the GEO database. Overall, these datasets include 972 blood samples from AD patients and healthy controls (HC): GSE140829 (n = 453; cases = 204, controls = 249), GSE63061 (n = 273; cases = 139, controls = 134), GSE63060 (n = 246; cases = 142, controls = 104) and GSE97760 (n = 19; cases = 9, controls = 10). The GSE97760 study was finally excluded from the meta-analysis due to: (i) the low number of samples available, and (ii) the fact that it was the only dataset coming from a different array platform (Agilent), reducing the number of common genes with the other datasets (all of them from Illumina Beadchip arrays v.3 and v.4).

For raw data processing, we first performed a normal-exponential background correction following a quantile normalization (after a Log₂-transformation) of the raw data using limma package (Ritchie et al., 2015).

After data normalization, expression data captured by multiple probes belonging to the same gene were averaged. Because of the difference in background measurements for each dataset we used Combat CONormalization Using conTrols (COCONUT) (Sweeney et al., 2016) to correct for batch effects between experiments, and to make the data comparable. COCONUT is an unbiased co-normalization method that assumes that all HC across studies come from the same statistical distribution, estimating first correction factors from each dataset’s HC samples, and then applying them to the AD samples in each dataset. This procedure removes technical differences while still retaining within-dataset differences between HC and AD groups. To correct for internal batches in the GSE140829 dataset, we used the function Removebatcheffect from the limma package before COCONUT co-normalization.

Differential expression (DE) analysis between AD samples and HC was carried out with limma and using gender as covariate to correct the model. Volcano plots of differentially expressed genes (DEGs) were built with EnhancedVolcano (Blighe et al., 2020) and Upsetplots from DEGs, and music and dopamine related genes were generated with the ComplexUpset R package (Lex et al., 2014). We used the R package Enrichmentbrowser (Geistlinger et al., 2016) to collect genes involved in dopamine-related biological processes from GO database and searching for terms including “dopamine” (Supplementary Table S3). Over-representation analysis from music-related genes through GO biological processes was conducted using the Clusterprofiler (Wu et al., 2021) R package. We applied the Benjamini-Hochberg procedure for multiple test correction and thresholds were set to 0.05. Fold enrichment was calculated as the quotient from gene ratio (number of genes of interest which are annotated to the gene-set/total number of genes of interest) and background ratio (size of the gene-set/size of all the unique genes annotated in the reference database).

The flow diagram of Figure 3 provides an overview on the epigenomic, and transcriptomic data used in the present study, together with a schematic representation of the main findings.

The consensus modules from the global co-expression networks represent biologically robust co-expressed gene groups. The analysis carried out on the AD (vs. HC) and the music transcriptomic data reveal commonalities between both datasets, elucidating common coordinated genetic processes behind musical stimulation and AD. We studied the commonalities in network organization of gene expression between the AD dataset and a dataset including whole blood gene expression data from individuals before and after 20 min of classical music stimulation (GSE48624). For this purpose, we used only AD patients from the AD dataset, and samples collected after stimulation from the music dataset.

The consensus weighted gene co-expression network was constructed using the WGCNA R package (Langfelder and Horvath, 2008). We used as input normalized expression data (and corrected for differences in gender) from common genes (those represented in both datasets) that showed the most variant expression values between samples (top 75% with the highest variance). Then, independent datasets were integrated into a multi-set format suitable for consensus analysis. We followed the signed network procedure, whereby the similarity between genes reflects the sign of the correlation of their expression profiles. A matrix of correlations between all pairs of selected genes was generated from the expression values, and further converted into an adjacency matrix with a power function. We chose a soft-thresholding power based on the criterion of scale-free topology after testing a set of candidate powers. Considering that the model-fitting index of a perfect scale-free network is 1, we selected a soft-thresholding power of 5 (Supplementary Figure S3A) because it resulted in the maximum model fitting index for both datasets (>0.9). Subsequently, the consensus topological overlap matrix (TOM) from the adjacency matrices and the corresponding dissimilarity (1–TOM) values were computed. Considering the different properties of the datasets, we scaled TOMs to make them comparable (Supplementary Figure S3B). The consensus TOM was calculated with component-wise (‘parallel’) minimum of the TOMs for each set. As co-expression module detection parameters, we chose a minimum module size of 30, a medium sensitivity for cluster splitting, and a 0.2 as dendrogram cut heigh threshold for module merging. The resulting consensus modules or groups of co-expressed genes were labelled by colors and used to calculate module eigengenes (the first principal component of the module). Module membership (MM) was calculated as a measure of intramodular connectivity. The core genes within the most relevant modules were selected using a MM > 0.8.

There is a growing interest in cognitive sciences to explore the beneficial impact of music on AD and other neurodegenerative diseases (Arab et al., 2021; Li et al., 2022). Several fields of research have aimed at analyzing the effects of music in dementia (case studies, randomized clinical trial, quasi-experimental studies…); the number of available studies is high although the methodology employed very heterogeneous. In addition, there are also many ongoing pilot trials and study protocols aiming at evaluating the effectiveness of music in AD patients (Guétin et al., 2009; Belleville et al., 2019; Gulliver et al., 2019; Flo et al., 2022). Recent reviews (e.g., Leggieri et al., 2019) agree that participating in music activities improves behavioral and psychological symptoms.

Our systematic review of the relevant literature unequivocally demonstrates the beneficial effect of music in AD (Figure 2); there is a convergent and vast evidence emerging from the literature indicating an overall beneficial effect of music on rehabilitation and improvement of AD. Only 7 out of the 107 (6.5%) studies that survive the filters of our PRISMA selected criteria did not report a benefit of music in AD. The remaining studies (n = 100; 93.5%) all show some benefit of music, being the main outcome on the enhancing of memory (Figure 2), which represents the main disability of AD.

Memory is a complex cognitive activity, and different types of long-term memory (explicit and implicit) have been studied in regard to dementia (see Section 3.2). A pioneer clinical trial by Arroyo-Anlló et al. (2013) demonstrated that familiar music enhances self-consciousness and awareness, one of the main concerns in AD. Many other studies have highlighted music as a memory enhancer (Simmons-Stern et al., 2010), in which patients with AD demonstrated better recognition accuracy thanks to music mediation. Särkämö et al. (2014) showed that music listening improved remote episodic memory, mood and orientation. Gómez-Gallego and Gómez-Garcia (2017) demonstrated that music therapy improved psychological, social, and cognitive behaviors. Twenty six reviewed studies exposed the improvement of cognition after musical intervention, [see for example (Ceccato et al., 2012; Innes et al., 2018; Lyu et al., 2018)], general cognition and executive function (Doi et al., 2017; Innes et al., 2017; Kim et al., 2022), visuospatial processing (Maguire et al., 2015), language fluency and autobiographical narrations (Thompson et al., 2005; El Haj et al., 2013; Pongan et al., 2017).

There are many other studies focusing on the beneficial effects of music on the psychological or emotional states, specially reducing agitation and anxiety in persons with dementia (Svansdottir and Snaedal, 2006; Raglio et al., 2008; Cooke et al., 2010; Sung et al., 2012; Vink et al., 2013; Cohen-Mansfield, 2014; Narme et al., 2014; Gómez-Romero et al., 2017; Pedersen et al., 2017; Harrison et al., 2021). As a matter of fact, several authors have reported the positive impact of music on wellbeing and reduction of depression in persons with dementia (Guétin et al., 2009; Janata, 2012; de la Rubia Ortí et al., 2018; Ray and Gotell, 2018); while other studies reported the impact of music regarding the reduction of pain and the improvement of quality of life (Pongan et al., 2017). From a psycho-social perspective, music also enhances motivation and reward circuits in AD patients (Simmons-Stern et al., 2010), and stimulates social behavior. Group music participation provides support to caregivers and AD, stimulating meaningful interaction between them (Tamplin et al., 2018).

Also, some investigations aimed at studying the possibilities and limitations of different musical interventions, contrasting listening activities vs. more active musical intervention. Sakamoto et al. (2013) suggested that musical interactive interventions exhibited stronger beneficial emotional effect than music listening in individuals with severe dementia. Recent studies corroborate the large effect of music active intervention in cognition, behavior, and functional state in AD (Gómez-Gallego and Gómez-Garcia, 2017), and some of them based on singing or choral showed relevant results in AD patients (Pongan et al., 2017; Lyu et al., 2018; Pongan et al., 2020; Flo et al., 2022; McDowell et al., 2022).

The benefits of individualized music or person-centered music interventions in AD is a recurrent finding (Gerdner, 2000; El Haj et al., 2015; Ihara et al., 2019), being familiar music a powerful tool with AD patients. Overall, these studies demonstrate that music has the power to enhance the recall of their past personal history more than other activities (Lord and Garner, 1993). Clements-Cortes et al. (2016), in the context of music therapy, studied the potential of 40 Hz sensory brain stimulation in AD and Parkinson’s Disease patients. More than 100 reviewed studies as well as the most recent reviews (Leggieri et al., 2019; Matziorinis and Koelsch, 2022) highlight the beneficial effect of music therapy for AD management and the slowdown of neurodegeneration (Matziorinis and Koelsch, 2022). In the same line, a recent meta-analysis (Lai et al., 2020) proposes music therapy as the best treatment for Mild Cognitive Impairment (MCI).

Music constitutes a distinct domain of non-verbal knowledge but shares certain cognitive organizational features with other brain knowledge systems (Omar et al., 2010). In line with a recent review (Baird and Samson, 2015), it should be considered that various forms of musical memory exists, and they may be differentially impaired in AD. In a systematic search aimed at analyzing the main findings related to different type of memories in studies focused on musical intervention in AD patients, we found 34 relevant articles (Figure 1). Main findings from this review are summarized in Table 1.

Nine articles studied autobiographical memory, considered to be of a mainly episodic nature (related to personal past events), and that is generally deteriorated in AD patients. Clinical trials, reports pre-post intervention, and case studies, have demonstrated the power of music to stimulate autobiographical memory (Irish et al., 2006; El Haj et al., 2012). Music constitutes a powerful stimulus that retrieves autobiographical, involuntary, and spontaneous memories in AD patients (Irish et al., 2006; El Haj et al., 2012, 2013, 2015; Fraile et al., 2019), reinforcing the sense of identity (Platel et al., 2021). Autobiographical memory has been considered an island of preservation during the progression of AD (Baird et al., 2018), while other authors have shown the power of popular songs (Basaglia-Pappas et al., 2013) or favorite songs (Fraile et al., 2019) to improve this kind of memory.

Some studies focused on how familiar music has been relatively spared by AD, and the patients’ intact recognition of familiar music (Vanstone et al., 2009). Other studies based on unfamiliar music found impaired episodic musical memory in AD patients (Quoniam et al., 2003; Menard and Belleville, 2009; Vanstone et al., 2012). Also, music is being used as a mnemonic strategy to enhance verbal episodic memory (Simmons-Stern et al., 2012; Moussard et al., 2014; Palisson et al., 2015; Ratovohery et al., 2019).

Semantic memory is a long-term memory related to more general knowledge and facts (grammar, name of colors, etc). In the context of music memory, it has been defined as “the information accessed by sense of familiarity for a melodic progression, regardless of timbre or starting pitch, and stripped from any contextual information” (Groussard et al., 2019); and also, as the ability to distinguish distorted melodies. Several studies have shown that musical semantic memory is highly preserved in AD (Cuddy and Duffin, 2005; Gagnon et al., 2009; Vanstone et al., 2009; Vanstone and Cuddy, 2010; Kerer et al., 2013; Cuddy et al., 2015). Kerer et al. (2013) showed that MCI and AD patients were impaired in tasks requiring verbal memory for music (recalling composers name, titles…) but, they performed better than healthy subjects at discriminating musical excerpts or remembering melodic lines. Omar et al. (2010) reported that AD patients recognized musical emotions and musical instruments but showed impaired recognition of compositions and musical symbols.

Many studies agreed in that implicit musical memory is well-preserved in AD patients (Baird et al., 2017; Deason et al., 2019), and specially procedural memory (Jacobsen et al., 2015). According to Groussard et al. (2019) “musical procedural memory is the ability to perform a previously learned musical motor sequence in a fluid manner.” Some pioneering case studies have focused on exploring the preservation of musical procedural memory to play an instrument (Beatty et al., 1988, 1994, 1997, 1999; Crystal et al., 1989; Cowles et al., 2003; Fornazzari et al., 2006). Also, other studies have reported the preservation of musical procedural memory in advanced AD (Cuddy et al., 2015; Jacobsen et al., 2015). Jacobsen et al. (2015) stated that “…[long term] musical [procedural] memory is surprisingly robust,” because the two brain regions that predetermine long-term memory encoding musical memory (caudal anterior cingulate gyrus and ventral supplementary motor area) have been preserved in advanced AD patients compared with other brain regions.

A total of 334 genes were selected through a thorough scrutiny of the literature and datasets (Supplementary Table S2). A narrow selection of 127 candidate genes was highlighted based on transcriptomic/epigenomic analyses (see below). Six of those were particularly prioritized as candidate genes, namely, SNCA, SLC6A4, ASCC2, FTH1, PLAUR, and ARHGAP26, because: (i) they were repeatedly found to be associated with musical aptitude in previous work and have emerged in transcription-based studies, and (ii) they occur consistently in the literature on AD.

Among the biological processes in which the 334 music-related genes are involved, we found that learning / memory was the most significant pathway (P-adjusted = 3.3E-9). Other important processes related to cognition (P-adjusted = 4.4E-9), rhythm (P-adjusted = 2.5E-7) and neuron death regulation (P-adjusted = 2.3E-7) were present among the top 10 significant pathways (Supplementary Table S4; Figure 4A).

Some dopaminergic terms from GO were also significantly over-represented in the music-related gene-set, such as GO:0001963 (dopaminergic synaptic transmission, P-adjusted = 4.2E-5), GO:0042417 (dopamine metabolic process, P-adjusted = 0.002) or GO:0007212 (dopamine receptor signaling pathway, P-adjusted = 0.01). Interestingly, from the above list of the top six candidate genes, SNCA and SLC6A4 are both involved in neuron–neuron synapse mediated by dopamine pathway (GO:0032227), as inferred by the protein–protein interaction network analysis.² SNCA gene function is related to the regulation of dopamine release and transport, whereas SLC6A4 is implicated in the regulation of serotonergic signaling by regulating serotonin reuptake into the presynaptic neuron, affecting emotions and stress responses.

There are 8,497 DMPs (FDR p-value <0.05) when contrasting AD patients vs. controls. 6,432 (75.7%) CpGs appeared as hypomethylated and 2,065 (24.3%) as hypermethylated in AD samples.

A total of 5,928 (69.8%) of these 8.4 k CpGs could be annotated to genes, resulting in 4,097 unique genes. Of the positions annotated to genes, 345 (4%) CpGs are annotated and related to more than one gene. For most of these genes, there are multiple DMPs associated (2.76 ± 1.28) gene-set and pathways enrichment analysis with the 8.4 k DMPs points to the following pathways depending on the database used: (a) KEGG: cytokine-cytokine receptor interaction, JAK/STAT signaling pathway, osteoclast differentiation, chemokine signaling pathway; (b) Reactome: neutrophil degradation, signaling by interleukins, nucleotide-binding domain, leucine rich repeat containing receptor (NLR) signaling pathways, platelet activation, signaling and aggregation, and (c) GO: leukocyte activation, cell activation, positive regulation of immune system process, immune effector process, cell migration, cell motility, localization of cell, immune response, locomotion. In addition, 853 DMRs (FDR p-value <0.05), containing at least three CpGs positions inside could be annotated to a total of 976 unique genes (only one DMR annotated to more than one gene). Of these regions, 645 (75.6%) appeared as hypomethylated in patients with AD, while the remaining 207 (24.3%) appeared as hypermethylated in cases vs. controls.

A total of 81 genes out of the 334 related to music (24.3%) were also found to be differentially methylated in the comparison of AD patients vs. HC, either considering genes associated with DMPs or those overlapping DMRs. Of these genes, 55 show an hypomethylation pattern in patients with AD when compared to HC, and 21 exhibit an opposite pattern of methylation; the remaining 5 genes, containing multiple CpGs inside, display both hypo- and hypermethylation patterns. The hypo and hypermethylation pattern, for most of these genes, resides within the gene body, while there are a small number of genes, such as SLC6A4 and AVPR1A, where the aberrant methylation is observed within the promoter. The pathways analysis performed considering the 117 CpGs associated with the 81 music-related genes, revealed an enrichment for lysosome pathway in KEGG, for biological processes involved in nervous system development in GO, and for cadherin and Wnt signaling pathways, included in memory formation and synaptic plasticity and maintenance, in Panther.

In addition, 15 out of the 337 music-related genes (4.5%) match those associated with the DMRs that emerge when comparing AD patients and HC; 10 as hypermethylated in patients, and the remaining 5 as hypomethylated. The mean number of DMPs for these 15 genes was 3. The pathways analysis of the genes associated with DMRs reveal the significant enrichment of 356 unique GO terms, and 12 KEGG pathways. The enriched pathways in both categories include those related to T cell signaling pathways, myeloid lymphocyte differentiation, leukocyte mediated immunity, chemokine signaling pathways, etc.

After data normalization and merging, a total of 19,145 genes were common when examining the three datasets of AD meta-analyzed in the present study. A quality control check of the normalized data through a principal component and clustering analysis highlighted 12 outlier samples belonging to the dataset GSE63061 that were finally disregarded for downstream analysis (GSM1539649, GSM1539651, GSM1539652, GSM1539653, GSM1539655, GSM1539644, GSM1539646, GSM1539647, GSM1539648, GSM1539650, GSM1539722, GSM4187601).

A total of 1,837 genes were found to be upregulated and 1,717 downregulated, all statistically significant in AD patients when compared against healthy controls (Figures 4B,C; Supplementary Table S5). The expression changes shown by DEGs are moderate but never too large [Log₂FC ranging from −0.60 and 0.34; comparable to those reported in the literature (Wang et al., 2021; He et al., 2022)]; probably because the main physiopathological alterations in AD patients occur in the brain. The high number of DEGs detected, each with subtle gene expression alterations, therefore mirrors the pathological condition at systemic level.

304 out of the full list of 334 music-related candidate genes (91%) were also present in the AD meta-analysis. Notably, 60 of the 304 (20%) were among the DEGs observed when comparing AD patients and control samples (34 were upregulated and 26 downregulated; Figure 4D). Music-related genes showing the lowest P-adjusted value and higher log₂FC were downregulated in AD vs. HC.

There is suggestive evidence indicating that dopaminergic pathways may play a role in the interplay between AD, memory, and music (see above). In our meta-analysis, when comparing transcriptomes from AD patients vs. healthy controls, the expression analysis showed a few DEGs involved in dopaminergic pathways (13 upregulated and 4 downregulated in AD vs. HC). In addition, there are dopaminergic genes among those in the music gene-set (n = 10), with SNCA (Synuclein Alpha) being the gene that best represents the AD-dopamine-music connection (Figure 4C; Supplementary Table S3).

To further investigate possible molecular links between AD and music-related genes, we generated a consensus co-expression network using AD patients and the only transcriptomic dataset related to musical stimuli available, aiming at detecting common conserved co-expression modules.

After applying a variance gene filtering (see Methodology section), 14,358 and 15,685 genes were retained in AD and music datasets, respectively, from which 11,920 genes were shared between both datasets and therefore available for follow-up consensus analysis. A total of 84 out of the 303 music-related gene-set in the meta-analysis (27.7%) did not pass the variance filtering; therefore, a reduced subset of 219 (72.3%) music-related genes was among the 11,920 genes used as input.

The consensus co-expression analysis revealed 20 co-expressed gene modules of very different sizes (ranging from 61 genes of the ‘darkred’ module, to 3,482 of the turquoise module; Figure 5A; Supplementary Figure S4A,B). The overall preservation of the eigengene networks was moderately high (D = 0.8; Supplementary Figure S4C) and the inter-module relationships in the two data sets has similarities. The highest eigengene preservation measurement was in the salmon and yellow modules (Supplementary Figure S4C).

The module with more music-related genes relative to the size of the module was the salmon one (18/195; Figure 5B); interestingly, most of them belong to the core of the module (11 genes in music dataset and 14 in the AD dataset with MM > 0.8; Figure 5C) suggesting a major role within it. Among the top 5 most connected genes within the module, we found 3 music-related genes: GMPR (top hub gene in both datasets with MM values of 0.95 and 0.93 in AD and music datasets respectively), SELENBP1 and ADIPOR1 (Supplementary Table S6).