Identification of genes associated with smell dysfunction in Parkinson’s disease Authors: Md Zainul Ali, Manjari Kiran, Pankaj Singh Dholaniya Published: 2026 ABSTRACT Highlights•Machine learning-based classification identified nine important non-motor symptoms for PD prediction.•ML based feature selection identified olfactory dysfunction as the most discriminatory non-motor feature for PD.•Blood transcriptomics revealed 32 genes associated with olfactory dysfunction and disease status.•Olfactory-associated gene modules and pathways overlap with PD diagnosis and core disease biology.•Monocyte-specific expression changes further link smell loss and PD.Keywords: Parkinson’s disease, Non-motor symptoms, Feature selection, Olfactory dysfunction, Transcriptomics, Machine learning analysis Introduction Parkinson’s disease (PD), a prevalent neurodegenerative disorder, primarily affects individuals aged over 50 years and remains without a definitive cure. Clinically, PD manifests prominently through motor symptoms such as bradykinesia, rigidity, tremors, and impaired balance. Typically, these motor indicators emerge only after significant neuronal degeneration, with nearly half of the brain’s dopamine-producing neurons already lost, irreversibly reducing dopamine levels [1], [2]. Consequently, relying solely on motor symptoms substantially limits early diagnosis and therapeutic intervention. Researchers have increasingly acknowledged non-motor symptoms (NMS) as preceding motor dysfunction [3], [4]. Notably, symptoms such as olfactory dysfunction (OD), constipation, REM sleep-related symptoms, fatigue, and mood alterations frequently manifest before the onset of motor deficits [5], [6], [7], [8], [9], [10], [11]. Despite clinical relevance, most studies historically isolate these non-motor indicators, often classifying them as secondary features. This fragmented approach limits the integration of symptom manifestation with underlying molecular mechanisms, particularly regarding DNA methylation dynamics. Consequently, the potential of NMS as diagnostic and mechanistic indicators of PD remains significantly underexplored [4], [12]. Epigenetic changes like DNA methylation (DNAm) are now seen as important in PD development. For example, DNAm affects genes like alpha-synuclein, which is central to PD [13]. Other studies demonstrated that DNAm changes with age may affect the onset and progression of the disease [14], [15]. Advancements in computational methodologies, especially machine learning (ML), have substantially enriched PD research. Nonetheless, existing ML applications predominantly emphasize motor symptoms or general clinical attributes [16], [17], [18], [19], [20]. These studies have addressed diagnostic classification, disease progression, and subtype stratification. However, most have not specifically prioritized individual NMS as early discriminative markers, nor have they linked such clinical signals to transcriptomic. We therefore sought to address this gap through a systematic approach that integrates clinical symptomatology with molecular level modulations. Parkinson’s Progression Markers Initiative (PPMI) is a prominent research resource supported by the Michael J. Fox Foundation, provides extensive clinical and omics datasets, including genomic, RNA sequencing, DNA methylation, and detailed symptom evaluations from both PD patients and healthy individuals [21]. The machine learning and systems biology approaches, such as weighted gene co-expression network analysis (WGCNA), provide opportunities to systematically evaluate the discriminative potential of NMS and to uncover transcriptomic alterations associated with PD. Exploring these associations may help bridge clinical observations with molecular mechanisms, offering novel insights into disease pathways and potential biomarkers. In this study, we aimed to integrate the insights from clinical, and transcriptomic data from PPMI, using ML and network-based approaches to investigate the role of NMS in PD. We integrated WGCNA, functional enrichment, DNA methylation profiling, and independent single-cell RNA sequencing analysis to investigate the transcriptomic and epigenetic signatures associated with the top-ranked NMS. Overall, this study may enrich our understanding of PD-related mechanisms linked to non-motor features and facilitate future biomarker-oriented research leading to early-stage diagnosis of the disease. Methods Data acquisition and preprocessing All clinical, transcriptomic, and methylation data for this study were downloaded from the PPMI database. PD and HC labels were assigned according to the PPMI cohort definitions available in the source database. All participants included from the PD cohort had a clinical diagnosis of Parkinson’s disease according to PPMI cohort assignment. Because this was a secondary analysis of harmonized PPMI data, the present study did not perform independent clinical re-adjudication and did not separately reapply biomarker- or DAT-based diagnostic criteria. The PPMI PD cohort includes multiple protocol-defined subgroups, and eligibility criteria have varied across recruitment phases. Further details of sample processing and cohort inclusion criteria are available through the PPMI study documentation [21]. The processed methylation data (850 K Illumina EPIC platform) from whole blood samples of the study participants was obtained from the PPMI database. This dataset initially encompassed a total of 524 participants. However, to streamline the analysis, focus was placed exclusively on patients and healthy controls, excluding other categories such as SWEDD (Scans Without Evidence of Dopaminergic Deficit) or missing categories. This refinement resulted in a dataset comprising 458 participants' data. Furthermore, the database has data collected from questionnaires for both motor and non-motor symptoms. For this study only NMS (also referred as features in this article) were considered. The focus was on 45 features collected through 21 clinical tests, consisting of NMS (Tables S1–S2). These assessments were selected to capture a broad spectrum of non-motor and related clinical domains in PD, including sleep disturbances, autonomic dysfunction, mood symptoms, anxiety, impulsive-compulsive behavior, cognition, gastrointestinal and urinary symptoms, daytime sleepiness, and olfactory function (Supplementary Tables S1–S3). A total of 684 participants has answered these questionnaires but many of the questions in the questionnaire are unanswered by the study participants. The data was preprocessed by eliminating participants with missing values. This resulted in a final dataset of 251 individuals, 165 with PD and 86 healthy controls, with complete responses for all 45 features, using the earliest available measurement for each feature across visits. Clinical characteristics of the data (distribution of the participants across different criteria) have been shown in detail in supplementary figure S1. The analysis began by categorizing the scores, marks, or values obtained from the questionnaires into three distinct groups, each representing the severity level of the respective feature (Table S3). Categorizing the data into severity levels helps simplify complex, continuous data into manageable classes, making it easier to identify patterns, improve model performance, and allow for better clinical interpretation of symptom impact. For instance, the assessment of OD was conducted using the UPSIT, with a total scoring range from 0 to 40 [22]. A score of 0 indicated the poorest smelling capability, whereas a score of 40 indicated normal olfactory function. These scores were further categorized into Normal (Score > 33), Mild or Hyposmia (Score 19 to 33), and Severe or Anosmia (Score < 19) [23], [24]. Similar categorization methods were applied to other features, except for two: years of education and the dominant hand used. Years of education were categorized based on the level of higher education achieved, while dominant hand was categorized as right (0), left (1), or ambidextrous (2). More detail is provided in Supplementary Table S3. Age categorization was performed to group participants into four distinct categories: 'Group_01′ for individuals under 45 years, 'Group_02′ for those between 45 and 59 years, 'Group_03′ for those aged 60 to 74 years, and 'Group_04′ for participants aged 75 years or older. Machine learning-based feature selection and classification modeling The primary aim of our classification approach is to accurately distinguish between Clinical Groups, that is, PD patients and healthy controls, based on NMS profiles. We aim to build a strong ML classifier to identify the symptoms that most effectively predict Clinical Groups, enabling earlier and more reliable detection of PD. This analytical framework serves two main purposes: 1) Feature Prioritization: To rank NMS based on their importance in classification, helping clinicians focus on the most informative signs and reduce reliance on less relevant or noisy predictors or features. 2) Diagnostic Discrimination: To develop a predictive model that can separate Clinical Groups with high accuracy, sensitivity, and specificity, ensuring that individuals at risk are correctly identified. This approach supports accurate diagnosis and helps in guiding biological analysis by identifying the key symptoms that contribute most to prediction. Feature selection: approaches to identify key non-motor symptoms Four different feature selection (FS) techniques were used to identify the most predictive NMS: Decision Tree (DT), Random Forest (RF), Boruta, and Normalized Mutual Information (NMI) [25], [26], [27], [28], [29]. Only those features identified by at least two FS methods were retained to enhance robustness and reduce the likelihood of including spurious predictors. Each method was selected for its unique analytical strength: Decision Tree was used for its interpretability and ability to hierarchically rank features based on their discriminatory power through recursive binary partitioning [25]. Random Forest, an ensemble method, provided robust feature importance estimates by aggregating outputs over multiple decision trees and accounting for nonlinear interactions [26], [27]. Boruta offered a rigorous wrapper-based approach by comparing actual features against shadow (random) features to retain only statistically significant predictors, thus minimizing the inclusion of noise [28]. NMI quantified nonlinear dependencies between features and the target variable, offering a mutual information-based perspective often missed by tree-based methods [29]. Mutual Information (MI) between two variables X and Y was computed using the formula. where p(x) and p(y) are the probability distributions of X and Y respectively, and p(x,y) is the joint probability distribution of X and Y. Normalized Mutual Information (NMI) is then obtained by normalizing MI: where H(X) and H(Y) are the entropies of X and Y respectively, defined as: This normalized metric allows for the comparison of mutual information across different datasets, providing a standardized measure of feature relevance. Feature importance was visualized using elbow plots to assess relative rankings across methods (Supplementary fig. S3). This aided in selecting the most informative features for model inclusion. Applying predefined importance thresholds ensured that the selected features were both statistically relevant and interpretable, thereby improving model robustness. Evaluating classifiers and interpreting feature impact Various ML classification algorithms were explored for the study (Supplementary Table S4) [25], [26], [27], [30], [31], [32], [33], [34], [35], [36]. Subsequently, RF classifier was chosen for final classification due to its superior balance of accuracy, computational efficiency, interpretability, and robustness. Parameters for the RF model was carefully configured based on established methodological standards to ensure accurate and generalizable outcomes. Specifically, the number of trees (n_estimators) was set to 100 to strike a balance between accuracy and computational efficiency. The maximum depth (max_depth) was limited to 5 to control model complexity and mitigate overfitting. The Gini criterion was employed to minimize node impurity effectively. Model performance and generalizability were rigorously evaluated through repeated 10-fold cross-validation (100 repetitions), ensuring reliability of predictions. The predictive accuracy of each feature was evaluated individually using a linear Support Vector Machine (SVM). SVM with a linear kernel is particularly effective for evaluating the predictive power of individual features. The linear SVM finds the optimal hyperplane that maximizes the margin between classes, making it ideal for assessing whether a single feature can linearly separate classes [30], [37], [38]. Feature contributions at the individual prediction level were interpreted using SHAP, a unified framework based on cooperative game theory. SHAP values quantify the impact of each feature on the model's output, offering both global and local interpretability. In our analysis, SHAP was used to rank the features based on their influence on the classification outcome, providing further validation of the most important predictors identified through feature selection. This approach enabled us to assess the consistency and strength of each feature's contribution across the dataset [39], [40]. Model evaluation metrics and statistical analyses Evaluating model predictive performance involved using the Area Under the Receiver Operating Characteristic Curve (AUROC). AUROC quantitatively assesses a model’s capability to distinguish between classes effectively. Predicted probabilities from each model underwent validation through repeated 10-fold cross-validation, enhancing prediction reliability and robustness. Constructing ROC curves required calculating true positive rates (TPR) and false positive rates (FPR) across varying thresholds, facilitating comprehensive visual comparisons among model performances. AUROC evaluations were carried out across multiple model configurations, each based on different sets of features obtained through various feature selection methods. Specifically, we compared models using all 45 initial features, individual feature sets selected by DT, RF, Boruta, and NMI, as well as the consensus-based feature set referred to as TOP9F, which includes features common to at least two of the four methods. ROC curves plotted for each configuration, along with their respective AUROC values, provided a clear comparative view of the predictive performance of each feature set. Further statistical analyses focused on examining gene expression level and methylation level differences between low smell score vs high smell score participants. Two-group comparisons utilized Wilcoxon test, with significance indicated through p-adjusted value thresholds [41], [42]. Boxplots visualized methylation distributions at significantly altered CpG sites, created using the R package ggplot2. These visualizations clearly differentiated methylation patterns between groups, with statistical significance levels explicitly marked by asterisks (* for p-adjusted < 0.05, ** for p-adjusted < 0.005, and *** for p-adjusted < 0.0005). A few important libraries and packages used in this study is mentioned in the supplementary table S5. Gene co-expression network construction and module-trait association analysis WGCNA systematically identifies groups of genes (modules) with similar expression patterns, helping unravel complex biological relationships. This analysis moves beyond examining isolated genes, effectively identifying clusters significantly associated with specific biological conditions or clinical traits. WGCNA constructs a weighted correlation network from expression data, utilizing gene pair correlations to establish network connectivity. Subsequent module detection algorithms organize interconnected genes into coherent modules, facilitating trait association analyses [43], [44]. We used 2 data variables UPSIT and Clinical Groups in combination to identify the gene modules using WGCNA. The RNA Seq data obtained from PPMI database was used for WGCNA analysis using R package (version 1.73). Gene IDs were set as row names, and sample names as column names. Quality control was performed using the goodSamplesGenes function to detect outlier genes and samples. A total of 1368 genes failing QC were removed, while all 503 samples initially passed. Hierarchical clustering (average linkage) and principal component analysis (PCA) were used to identify outlier samples; hence seven samples were excluded. Phenotypic traits were matched to expression profiles using sample ID. For normalization and filtering, count data were processed using DESeq2 (version 1.42.1). Genes with fewer than 150 counts in the samples were removed, retaining 4710 genes for downstream analysis. Variance-stabilizing transformation (VST) was applied, and normalized counts were transposed for downstream analysis. For network construction, a signed co-expression network was built using the WGCNA framework. The soft-thresholding power (β) was determined via the pickSoftThreshold function. Power 14 was selected to achieve scale-free topology (R2 > 0.8) while minimizing mean connectivity (Supplementary fig. S3). Modules were identified using blockwise clustering (blockwiseModules) with parameters: maxBlockSize = 5000, mergeCutHeight = 0.25, and TOMType = signed. Modules received distinct color labels, facilitating clear visualization and differentiation. Meaningful associations were identified by performing Pearson correlation analyses between module eigengenes and two key data variables: UPSIT and Clinical Groups. Heatmaps were used to visually highlight significant module-trait correlations, with statistical significance adjusted using Student asymptotic tests. This structured analysis helped uncover gene clusters that showed strong links to OD, offering important insights into potential molecular patterns relevant to PD detection. Investigating the functional relevance of identified OD associated genes in PD To explore the molecular basis of OD, we analyzed WGCNA-identified gene modules that were significantly associated with UPSIT. Within these modules, individuals with low and high UPSIT scores were compared using the Wilcoxon test with adjusted p-values, and differentially expressed genes (DEGs) were identified. Functional enrichment analyses were performed on these genes, followed by the literature review to highlight genes previously linked to olfactory pathways. Furthermore, from this curated list of OD-associated genes, the corresponding CpG sites were annotated using the R package IlluminaHumanMethylationEPICanno.ilm10b4.hg19 (version 0.6.0). The beta normalized values for CpG were obtained from the PPMI dataset. The CpGs with significant change in the methylation levels within the individuals with low vs. high UPSIT scores were reported. Assessment of AROMA gene expression using single-cell RNA sequencing Single-cell RNA sequencing (scRNA-seq) data of peripheral blood mononuclear cells (PBMCs) from healthy controls and PD patients (early PD or EPD, and Advance PD or APD) were obtained from an independent publicly available dataset [45]. The dataset was analyzed using Seurat (v5.0) with Harmony integration for batch correction. After quality filtering (based on mitochondrial gene content, total counts, and gene features), cells were clustered using the FindNeighbors and FindClusters functions with the UMAP algorithm for dimensional reduction. The identified transcriptionally distinct clusters were annotated into canonical peripheral blood cell types using established marker genes. Heatmaps representing Z-score–normalized average expression across groups were produced using pheatmap package. The Z-scores were clipped between –1 and + 1 to emphasize directionality. AROMA genes showing significant differential expression were annotated with significance markers (*, *, *** corresponding to adjusted p < 0.05, 0.01, and 0.001, respectively). Results Machine learning based feature selection reveals nine key NMS To begin our analysis, we assessed the predictive strength of NMS for the detection of PD using the computational framework outlined in the Methods section. Initially, several ML algorithms were tested for classification performance using the full set of 45 clinical features. Among these, the RF classifier outperformed others in terms of overall accuracy and robustness (Fig. 1a) and was therefore selected for further analysis. To refine the model and reduce redundancy, four different FS methods were used: DT, RF, Boruta, and NMI. Each method has its own strengths and focuses on different aspects of feature importance, as explained in Methods section 2.2. To ensure consistency and avoid method-specific bias, we retained only those features that were commonly selected by at least two of the four methods. This strategy improves model stability and reduces the inclusion of noisy predictors. The final set of nine features, selected based on this intersection approach, is referred to as TOP9F. The commonality of these features is illustrated in the Upset plot (Fig. 1b). To evaluate how these features performed in comparison to other sets, we built separate classification models using feature sets obtained from each FS method individually, as well as the combined TOP9F set. Among all, the model using TOP9F gave the highest accuracy, specificity, and balanced accuracy (see Table 1 and Fig. 1c and 1d). The RF model built on the TOP9F set showed a small but meaningful improvement in accuracy, increasing from approximately 81.5 percent to 82.6 percent. This suggests that selecting a stable and focused set of features can enhance the model’s performance while also making it more interpretable. The nine features identified were apathy (MDS_Apathy), fatigue (MDSP_Fatigue), four cognition-related measures [Montreal_Cognitive, Symbol_Digit, Trail_Making_A, and Trail_Making_B], REM sleep questionnaire-derived dream-related symptoms (REM_Dream), gastrointestinal autonomic symptoms from the SCOPA-AUT domain (SCOPA_Gastro), and smell loss (UPSIT). Together, the TOP9F features represent a reliable and informative subset of non-motor and related clinical features that may aid PD identification and guide downstream biological analyses. Overview of feature selection and model performance: (a) Accuracy distribution across various ML algorithms; (b) Upset plot showing common features across four FS methods; (c) FS based accuracy comparison; (d) ROC curves for selected models. Table 1Comparison of classification performance using different feature sets selected by DT, RF, Boruta, and NMI, along with their intersection (TOP9F). Metrics include number of features, sensitivity, specificity, balanced accuracy, and overall accuracy using the RF classifier on the NMS dataset.S. No.DetailsAll FeaturesDecision TreeRandom ForestBorutaNMITOP9F1Number of Features4551091092Sensitivity (in %)74.0086.0076.0076.0072.0076.003Specificity (in %)72.0070.0080.0074.0074.0080.004Balanced Accuracy (in %)73.0078.0078.0075.0073.0078.005Accuracy (in %)81.5081.2082.6080.8082.5082.60 Comparison of classification performance using different feature sets selected by DT, RF, Boruta, and NMI, along with their intersection (TOP9F). Metrics include number of features, sensitivity, specificity, balanced accuracy, and overall accuracy using the RF classifier on the NMS dataset. Loss of smell is the most distinctive feature among NMS in PD To identify which individual NMS contributed most strongly to disease prediction, we performed the piecemeal analysis using linear SVM model and ranked all the features based on their accuracy. Among all features, the UPSIT, which measures olfactory function, showed the highest classification accuracy (Fig. 2a). This indicates that smell loss is the most powerful standalone predictor for early PD. Furthermore, we applied SHAP to quantify the contribution of each feature to the model's predictions. UPSIT consistently showed the highest SHAP value in both male and female participants, as well as in combined analyses, confirming its dominant and independent influence on model outcomes (Fig. 2b). Additionally, we observed a decline in UPSIT scores across the three stages of Hoehn and Yahr Scale (HYS), indicated by a reduced proportion of individuals with high UPSIT scores as disease severity increased (Fig. 2c). This pattern was further confirmed by comparing the distribution of high versus low UPSIT scores across different age groups in both healthy controls and PD patients. Our analysis clearly shows that the proportion of individuals with low UPSIT scores increased significantly and progressively with age in the PD group (Fig. 2d). (a) Classification accuracy using each of the 9 selected non-motor features individually; (b) SHAP-based feature importance for disease classification; (c) Distribution of UPSIT categories across Hoehn and Yahr stages, HYS 0 individuals were healthy controls, while those with HYS scores of 1 or 2 were taken from the PD patients.; (d) Distribution of UPSIT categories across different age ranges. WGCNA reveals gene modules associated with smell loss & PD To explore molecular underpinnings of smell loss in PD, WGCNA analysis was performed. It identifies 9 distinct modules associated with UPSIT and Clinical Groups (Fig. 3a). From these modules, 6 were significantly correlated with Clinical Groups among which 3 modules; black (n = 102 genes), brown (n = 454 genes), and red (n = 126 genes) were also significantly (p < 0.05) correlated with USPIT (Fig. 3b). The total number of genes associated with 6 correlated modules is 1127, among which 682 genes belong to the 3 modules of UPSIT category. Furthermore, from these 682 gene 340 genes showed direct expression correlation (p < 0.05) with both UPSIT and Clinical Groups status (Fig. 3c) and are henceforth referred to as PD UPSIT-Linked Significantly Correlated (PULSE) genes in this study. These results indicate that smell loss and PD may share common genes and pathways. Further, pathway enrichment analysis of the PULSE genes revealed involvement in key biological processes such as apoptosis, oxidative stress, autophagy, chemokine signaling, and inflammation, reflecting their substantial role in the molecular pathology underlying OD in PD (Fig. 3d, Table S6). The complete list of the enriched pathways and GO terms is provided in Supplementary Table S7. (a) Hierarchical clustering dendrogram showing gene modules before and after merging; (b) Heatmap depicting correlations between modules and the two features, highlighting significant associations; (c) Venn diagram showing shared genes between UPSIT and Clinical Groups; (d) Dot plot showing key pathways (obtained from GO Terms, KEGG pathway, Reactome & Disease Pathway analysis) involved in smell loss & PD. Shared molecular signatures confirm the potential link between OD and PD Following the extensive literature search of PULSE genes, we identified 32 genes to be strongly associated with olfaction, we called these genes as Anosmia-Related Olfactory Mechanism-Associated (AROMA) genes (Table S8). We further examined the gene expression changes of AROMA genes in individuals with low and high smell scores using the Wilcoxon test, we found significant change in gene expression (Fig. 4a, b). In addition, to explore the epigenetic regulation of the PULSE genes, we carried out CpG methylation analysis on participants with low and high UPSIT scores. The result demonstrated that 20 CpG sites were significantly differentially methylated, which includes 3 hypermethylated and 17 hypomethylated sites, in individuals with low olfactory scores (Fig. 4c, Table S9). Furthermore, we investigated the role of AROMA genes in the olfactory transduction pathway to better understand their involvement in OD. Although the pathway is well-defined in the literature, we structured it into four main stages and a subsequent maintenance phase based on our analysis. These stages are not rigidly classified in prior works, and the division into these phases is a heuristic applied for clarity in our study, rather than a universally established framework: (1) odorant binding to olfactory receptors (ORs) on olfactory sensory neurons (OSNs), (2) signal transduction via Golf, cAMP, and ion channels, (3) action potential generation and axon projection to the OB, (4) synaptic transmission and local circuitry, followed by (5) maintenance, regeneration, and immune defense. The AROMA genes, mapped to each of these stages based on our literature search, are shown in Fig. 4d. (a) Heatmap & (b) Boxplots of the AROMA genes implicated in olfaction among PULSE genes showing significant expression across UPSIT low vs high score; (c) Boxplots of 20 CpGs associated with PULSE genes showing significant differences in methylation level. (d) Schematic representation of the olfactory signaling pathway highlighting key genes involved in odor detection, signal transduction, synaptic transmission, and sensory neuron maintenance. OR: Olfactory Receptor. OSN: Olfactory Sensory Neurons. AC: Adenylate Cyclase. Part of this image), licensed under CC BY 4.0 ( adapted from Servier Medical Art (https://smart.servier.com/https://creativecommons.org/licenses/by/4.0/). Expression pattern of AROMA genes in single-cell profiles of PD To further investigate the expression landscape of AROMA genes at the single-cell level, we analyzed an independent scRNA-seq dataset derived from PBMCs of PD patients and healthy controls. The clustering analysis identified ten transcriptionally distinct cell populations (Fig. 5a), cluster annotation based on canonical marker genes revealed six cell types, including CD4+ T cells, CD8+ T cells, B cells, NK cells, monocytes, and granulocyte–monocyte progenitors (GMPs). The UMAP visualization further illustrated distinct clustering of these cell types, highlighting clear separation across populations (Fig. 5b). Analysis of AROMA gene expression across five major cell types showed that only monocytes exhibited significant alterations, compared to healthy control 12 genes displaying upregulation in EPD (PREX1, ITPK1, PIK3CG, STAT5B, NEDD9, ATP6V1B2, CAP1, KAT6A, NCOA2, PLCG2, and STAT3), or in APD (CREB5). (a) Clustering analysis identifying ten transcriptionally distinct cell populations from PBMCs based on single-cell transcriptomic profiles. (b) UMAP visualization depicting distinct clusters, including CD4+ T cells, CD8+ T cells, B cells, NK cells, and monocytes. Heatmaps showing significant expression changes of AROMA genes across various cell types: B cells, CD4+ T cells, CD8+ T cells, Monocytes, and NK cells. Discussion In this study, we aimed to explore the role of NMS in PD and their molecular underpinnings. NMS have been increasingly recognized as significant indicators of PD, often preceding motor symptoms by several years. The integration of clinical symptom data with gene expression and DNA methylation analysis provided important insights into the biological mechanisms associated with these manifestations. Our analysis identified RF as the best-performing ML classification model and, through ML-based feature selection, defined nine key non-motor features termed TOP9F. Autonomic dysfunction is also a well-recognized non-motor feature of PD. In the present study, autonomic symptoms were represented in the initial feature set through SCOPA-AUT and MDS-UPDRS-related questionnaire-derived variables. From the autonomic dysfunction domain, only the gastrointestinal symptom feature derived from SCOPA-AUT was identified as one of the features in the TOP9F. Linear SVM classification and SHAP analysis further showed that OD was the most discriminative non-motor feature. It outperformed other non-motor features, including cognitive and sleep-related measures, in its predictive contribution. SHAP analysis further corroborated UPSIT as the most influential feature, with lower UPSIT scores corresponding to greater disease severity, suggesting its potential for both diagnosis and monitoring disease progression. This aligns with prior research emphasizing the close association of olfactory pathways with neurodegenerative processes, particularly in PD [46]. The strong predictive value of UPSIT underscores the importance of OD in risk stratification for PD [5]. Epidemiological data suggest PD patients exhibit reduced olfactory function, distinguishing PD from similar neurodegenerative conditions such as multiple system atrophy (MSA) and progressive supranuclear palsy [47], [48]. Reduced smell (hyposmia) is also linked to memory problems and may help in grouping PD types [48]. In the clinic, smell tests can help both in diagnosis and in spotting risk in people with no symptoms [22], [47]. Moreover, smell tests represent cost-effective, minimally invasive clinical tools for early detection and risk stratification, enhancing their clinical utility [49]. Recent advancements in digital olfactometry and wearable sensors are enhancing the precision and practicality of smell tests, strengthening their potential in PD diagnosis, particularly during its early stages [50], [51]. Smell testing remains clinically useful in PD. However, in the post-COVID era, OD may also arise from persistent post-viral effects, with a subset of individuals exhibiting prolonged or incomplete recovery of smell function [52], [53], [54]. Therefore, smell test results should be interpreted together with clinical history and other diagnostic features rather than in isolation. At the molecular level, OD correlate strongly with pathological processes such as alpha-synuclein aggregation within the OB [55], [56]. Our findings reinforce OD as a reliable early indicator of PD, with over 90% of PD patients exhibiting some degree of smell impairment long before motor symptoms emerge, as shown in previous studies [4]. The ability to link clinical symptoms with molecular data enhances diagnostic accuracy, facilitating the identification of at-risk individuals even before the onset of motor symptoms. UPSIT, when coupled with such technologies, may offer a promising, cost-effective approach for early PD detection, thus improving the accuracy of diagnoses and facilitating early intervention. In addition, the WGCNA of peripheral blood transcriptomics data revealed three gene modules associated with OD, overlapping with six modules linked to PD, which are mentioned as PULSE genes in this study. This overlap suggests shared genetic program between OD and PD, further supporting the idea that OD reflects molecular changes paralleling the development of motor symptoms [6], [57]. Additionally, the methylation pattern of 20 CpG sites associated with these genes show a significant change in methylation level in PD vs control group. Notably, the OB, an early site of PD-related pathology, strengthens the biological connection between OD and PD [57], [58]. Identifying common biological pathways such as apoptosis, oxidative stress, and autophagy highlights the molecular overlap between OD and PD. We further narrow down the PULSE genes to 32 AROMA genes associated with OD, based on their involvement in olfactory transduction and sensory neuron maintenance. These findings emphasize the integral role of OD within the molecular disruptions typical of PD. CpG methylation analysis identified 20 significant methylation sites, with hypomethylation observed in individuals with lower olfactory scores, suggesting that the change in expression of certain genes is possibly by epigenetic changes. These insights highlight the potential of methylation-based markers in PD detection, however further experimental validations are needed to support these findings [14], [59]. Furthermore, mapping the AROMA genes to the different stages of the olfactory transduction pathway highlights their functional influence. In examining the genetic components that contribute to OD in PD, it becomes evident that certain genes play crucial roles at different stages of olfactory transduction. For instance, during the initial phase of OSN differentiation, genes like CREB5 [60], HSPA6 [61], PREX1 [62], and SLC6A6 [63] are pivotal. These genes influence processes such as receptor trafficking and taurine transport, which are foundational for the initial binding of odorants. Interestingly, these early processes set the stage for how olfactory deficits might manifest, especially in neurodegenerative conditions like PD. Once the initial binding occurs, the signal transduction phase is initiated, where genes such as BEST1 [64], ITPK1 [65], MAPK1 [66], PIK3CG [67], and STAT5B [68] come into play. These genes are involved in regulating critical pathways such as the cAMP signaling pathway and ion channel activity, both of which are essential for effective odor detection. Disruptions in these pathways can lead to significant impairments in the ability to detect smells, a hallmark of early PD. Moreover, the process of axon targeting and synaptic formation, which is necessary for the proper projection of OSNs to the OB, relies heavily on genes like APLP2 [69], DIP2B [70], NEDD9 [71], NDST1 [72], and RTN3 [73]. These genes ensure that the axons of OSNs are accurately directed, allowing for proper communication between OSNs and OB interneurons. Without the coordination of these genes, the structural integrity of the olfactory system could be compromised, leading to dysfunctions. Once the signals reach the OB, synaptic transmission takes over, facilitated by genes such as GABARAPL1 [74], NDEL1 [75], [76], PLXNC1 [77], and RAB3D [78], [79]. These genes play key roles in synaptic vesicle exocytosis and receptor trafficking, ensuring that the communication between OSNs and OB interneurons is precise. Any disruption here could lead to improper signaling, further impairing olfactory function. Lastly, the survival and long-term maintenance of OSNs depend on genes like APAF1 [80], ATP6V1B2 [81], BCL6 [82], CAP1 [83], [84], [85], [86], IFNAR1 [87], IGF2R [88], IL1R2 [89], IL6R [90], IRF1 [91], KAT6A [92], [93], NCOA1 [94], [95], [96], NCOA2 [95], [96], PLCG2 [97], and STAT3 [98]. These genes are involved in immune response regulation, olfactory inflammation, and OSN regeneration, ensuring that the olfactory system remains functional over time. In PD, altered expression of these genes has been implicated in the regulation of olfactory dysfunction, revealing critical molecular pathways behind the loss of smell. As we map the AROMA genes to the different stages of olfactory transduction, it becomes clear that these genetic factors are not only critical to the basic functioning of the olfactory system but also to the progression of smell loss in PD. This mapping illustrates the functional roles of these genes in various stages of olfactory signaling, from initial odorant binding to long-term maintenance of olfactory sensory neurons, offering significant insights into how genetic disruptions contribute to olfactory dysfunction in PD. Consistent with these observations, the scRNA-seq analysis of peripheral blood cells supports the association between smell loss and PD, as several AROMA genes were specifically upregulated in monocytes. Earlier studies have demonstrated that elevated chemokines such as CCL2, CCL3, and CCL5, in inflamed olfactory tissue can enter circulation and recruit CCR2+ monocytes from the blood, leading to immune infiltration and persistent neuroinflammation in olfactory regions [99], [100], [101], [102], [103]. Furthermore, monocyte activity is reported to be linked with early olfactory impairment, with peripheral immune alterations correlating with reduced olfactory performance [104]. Activated monocytes can migrate across the blood–brain barrier into the substantia nigra, where they amplify neuroinflammatory signaling and contribute to dopaminergic neuron loss [105]. Monocytes have an increasingly recognized role in PD, as PD is characterized by chronic peripheral inflammation and enhanced trafficking of activated monocytes into the central nervous system, where they contribute to neuroinflammatory signaling and dopaminergic neuron stress [106], [107]. In parallel, monocyte–microglia crosstalk and type-I interferon–driven inflammatory responses are known to influence olfactory circuit integrity [108], [109], suggesting that dysregulated monocyte function may contribute to both smell loss and PD pathology. Recent studies also suggest that monocytes themselves may take part in pathways related to olfactory receptors [110], [111], [112], [113]. These findings provide a coherent biological context for the monocyte-specific gene expression alterations observed in our single-cell analysis. Twelve AROMA genes are significantly altered in monocytes of PD patients. These genes map to different stages of the olfactory pathway and together indicate that peripheral immune changes may influence early OD. At the first stage, CREB5 is a transcriptional regulator highly active in monocytes [114], [115], [116]. When upregulated it can sustain pro-inflammatory and survival programs in myeloid cells that are likely to cause tissue damage and contribute to olfactory dysfunction [117], [118], [119]. PREX1, which regulates Rac1-dependent cytoskeletal remodeling, plays key roles in monocyte recruitment [120], migration [121], [122], and neuronal morphology [123], [124]. PREX1 overexpression also enhances mTOR signaling [125], a pathway reported to be elevated in PD brains and linked to defective axonal autophagy [126]. These functions suggest that PREX1-driven signaling may affect both monocyte behavior and the structural integrity of OSNs. Genes involved in the second stage of signal transduction, including ITPK1, PIK3CG, and STAT5B also highlight the connection between immune signaling and smell loss. ITPK1 is highly expressed in monocytes [127] and regulates inositol-phosphate signaling and inflammatory stress responses [128]. PIK3CG encodes PI3Kγ, an essential checkpoint that limits TLR-driven cytokine production [129] while supporting monocyte migration and activation under chronic inflammatory conditions [130]. Increased PI3Kγ expression observed in PD is consistent with the heightened inflammatory state reported in circulating monocytes [106], [107], [131]. STAT5B is a transcription factor in the JAK–STAT pathway that is activated in monocytes/macrophages by cytokines (e.g., GM-CSF, IL-2 family) and controls genes for survival, proliferation and inflammatory programs [132], [133], [134]. The maintenance and immune-defense stage include several genes whose altered expression indicates persistent immune activation. ATP6V1B2 is a V-ATPase component enriched in monocytes [135] and linked to lysosomal acidification [136], and its elevation may reflect increased lysosomal demand associated with α-synuclein and autophagy stress in PD. CAP1 promotes NF kB activation and cytokine release in monocytes [137] and is enriched in intermediate monocytes that interact with T cells. KAT6A, a histone acetyltransferase, plays a critical and complex role in monocytes and their differentiation into macrophages, primarily in regulating pro-inflammatory responses and influencing myeloid differentiation pathways [138]. PLCG2, essential for monocyte to macrophage differentiation [139], [140], and STAT3, which regulates anti-inflammatory and angiogenic programs [141], [142], [143], [144], further indicate a shift toward an activated and dysregulated monocyte state. Together, these genes illustrate how peripheral monocytes in PD show coordinated changes across multiple olfactory pathway stages. Their altered expression suggests that chronic immune activation, disrupted trafficking, and increased lysosomal and signaling stress may contribute to early olfactory impairment. These findings support the idea that smell loss reflects broader molecular disruptions linked to PD pathophysiology. The role of peripheral blood as a surrogate for central pathology is also supported by our results, which show that blood-based transcriptomic and epigenetic assays can capture systemic signatures of OD, offering a minimally invasive method to detect early neurodegenerative changes [13], [14], [15], [145], [146], [147]. Our results align with existing studies showing that peripheral omics profiles mirror oxidative stress, immune activation, and apoptotic pathways active in the OB [1], [49]. While peripheral tissues cannot fully replicate the brain’s microenvironment, the concordance between peripheral omics profiles and PD pathology strengthens their potential for use in early PD screening. Combining blood-derived biomarkers with olfactory assessments thus enhances the feasibility of screening at-risk populations [2], [4]. Unlike prior studies that predominantly focused on clinical assessments or ML models based on motor or voice features, our study integrates transcriptomic, epigenetic, and clinical data (Table S10). This systems-level approach uncovers the molecular underpinnings of PD, offering new insights into the disease’s molecular disruptions and advancing our understanding of PD as a multi-stage, systemic disorder. Combining olfactory assessments with gene expression and methylation assays could provide a robust framework for identifying individuals at risk of PD and for stratifying them based on molecular signatures. The single-cell transcriptomic analysis revealed that alterations in AROMA gene expression were specifically concentrated in monocytes, not broadly across PBMC subsets. This cell-type-specific signature strengthens the biological relevance of our findings, suggesting a distinct role for monocytes in connecting smell loss to PD pathophysiology. Together, these convergent observations strengthen the rationale for integrating molecular and olfactory measures in future diagnostic and mechanistic studies. The strength of our study lies in uniting clinical, transcriptomic, single-cell, and epigenetic perspectives, providing a comprehensive view of the molecular changes associated with smell loss in PD. Limitation This study proposes a strong molecular level correlation between smell dysfunction and PD; however, it has a several limitations. First, it is a secondary analysis of PPMI data, and PD-HC status was inherited from source-cohort definitions rather than being independently adjudicated in the present study. Second, Alzheimer-related co-pathology was not explicitly assessed in the PPMI-derived dataset used here; therefore, although smell loss emerged as the highest-ranked feature in our PD-versus-HC framework, its specificity relative to other neurodegenerative causes of OD could not be formally evaluated. Third, OD may also arise from non-PD causes, including persistent post-viral smell loss in the post-COVID era, which may reduce specificity when smell testing is interpreted in isolation. Finally, the modest sample size and the fact that the study was conducted in limited geographic regions restrict generalizability and preclude causal inference, and the functional interpretation of the identified genes should be considered hypothesis-generating needing further experimental validation. Conclusion This study integrates clinical and transcriptomic analyses to reposition OD as both the most discriminative non-motor symptom and a biologically grounded early feature of PD. By combining machine learning–based symptom prioritization with gene expression and co-expression network, we identify molecular signatures that link smell loss to key PD-related pathways. These findings suggest that early alterations in peripheral blood gene networks may reflect broader pathogenic processes contributing to olfactory impairment and PD progression. A key observation is the specific alteration of AROMA genes in monocytes, linking peripheral immune activation to OD and suggesting a novel role for these cells in PD pathophysiology. Functional studies targeting these key molecular pathways, including monocyte-specific mechanisms, could provide new therapeutic options for delaying or preventing PD onset. Overall, this work positions smell loss not merely as a clinical symptom but as an integrated clinical–molecular indicator of early PD, offering new opportunities for mechanistic exploration, early diagnosis, and the development of targeted interventions aimed at modifying disease trajectories. Code availability Scripts for the data analysis github.com/iam-zain/PD_and_Non_Motor_Symptoms. Author contributions M.Z.A.: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Visualization, manuscript writing, review & editing. M.K.: Methodology, Resources, manuscript review & editing. P.S.D.: Conceptualization, Funding acquisition, Methodology, Project administration, Resources, Supervision, manuscript review & editing. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgment We wish to express our gratitude to University Grant Commission (UGC) Special Assistance Program (SAP) for their support to the Department of Biotechnology & Bioinformatics and the Bioinformatics Infrastructure Facility (BIF) at the School of Life Sciences, University of Hyderabad, India, for providing essential support and infrastructure. Additionally, we also acknowledge CSIR- HRDG, India, for granting a fellowship to Md Zainul Ali during his studies. Data used in the preparation of this article were obtained [on 22nd February 2022] from the Parkinson’s Progression Markers Initiative (PPMI) database (www.ppmi-info.org/access-data-specimens/download-data), RRID:SCR 006431. This analysis used Tier 1 (open access) PPMI data ppmi-publication-policy. For up-to-date information on the study, visit www.ppmi-info.org “PPMI – a public-private partnership – is funded by the Michael J. Fox Foundation for Parkinson’s Research and funding partners, including [A complete list of PPMI funding partners is available on the PPMI website].” Ethical considerations Data used in the preparation of this article were obtained from the PPMI database, which is publicly available and de-identified. The PPMI study has obtained the necessary ethical approvals for the collection and use of the data, in accordance with institutional and regulatory guidelines. No additional ethical approval was required for this secondary data analysis. Declaration of generative AI and AI-assisted technologies in the writing process All substantive intellectual content, analysis, and interpretations remain the original work of the authors. Portions of the manuscript text were revised and refined with the assistance of AI-based language tools, and any AI-generated suggestions were critically reviewed and approved by the author team prior to submission. Funding sources This work was supported by the funding received from the Anusandhan National Research Foundation (ANRF) IRG (ANRF/IRG/2025/000561/LS), Department of Biotechnology (DBT-BT/PR25787/GET/119/96/2017), Govt. of India, and IoE-UoH research grant (UoH-IoE-RC2-21-016) to PSD. Footnotes Supplementary data to this article can be found online at https://doi.org/10.1016/j.prdoa.2026.100453. Supplementary data The following are the Supplementary data to this article: Supplementary Data 1 Supplementary Data 2 References Associated Data Supplementary Materials Supplementary Data 1 Supplementary Data 2