DEDUCE: statistical inference on disease-associated genes uncovers tissue–disease associations Authors: Boqi Wang, Jiayi Wang, Ammar Aleem Rashied, Bo Meng, Jesse Zhang, Jun S Liu, Jie Jiang, Zhaohui S Qin Published: 2025 ABSTRACT AbstractAccurate identification of affected tissues of human diseases is important for the derivation of disease etiology and the development of new treatment strategies. In this study, we develop a logistic regression-based method named DEDUCE (disease tissue detection using logistic regression) that combines genomics big data and machine learning to address this important problem. The central hypothesis is that most disease-associated genes are expressed specifically in affected tissues. DEDUCE takes advantage of newly emerged data on disease-related genes as well as tissue-specific gene expression data. The unique feature of DEDUCE is that it takes into account the strength of gene–disease associations. When we applied DEDUCE to a total of 3261, 324 gene–disease associations collected from DisGeNET covering 30,170 diseases and 21,666 genes, we identified 216 significant tissue–disease pairs composed of 120 unique diseases and 37 unique tissues. Many of them shed light on potential explanations for disease pathogenesis. The results showed great consistency with previous findings and were proven effective by empirical plots and gene set enrichment analysis. Overall, DEDUCE has shown great potential in uncovering novel pathogenesis mechanisms of complex diseases. In-depth analysis and experimental validation were required to fully understand these discovered tissue–trait associations and their enriched genes. Introduction The development and rapid adoption of genome-wide association studies (GWASs) over the past two decades have led to the fast accumulation of genetic variants associated with human traits and diseases [1]. Concurrently, large-scale biomedical initiatives, such as the Genotype-Tissue Expression (GTEx) project and the Encyclopedia of DNA Elements [2, 3], have generated vast omics datasets, providing unprecedented opportunities for scientists to extract meaningful biomedical insights from omics big data. It is of great interest to leverage these biomedical big data to address some of the most important questions in health sciences. While the target tissues of many human diseases, such as lung cancer and diabetes, are well established, the knowledge of relevant tissues of many diseases and traits (“diseases and traits” will be referred to as just “diseases” hereafter for simplicity) remains elusive or under-explored. It is widely recognized that identifying the relevant tissues associated with diseases is crucial for better understanding disease etiology, which could lead to more effective and targeted treatment options. In order to address this important question, we explored whether newly emerged data from GWAS can help. Multiple computational methods have been published in recent years to map relevant tissues to diseases [4–7]. However, there remains significant room for improvement, particularly given the increasing availability of genomics data and a deeper understanding of disease-associated genes. As an example, it has long been accepted that gene expression is tissue-specific. Therefore, it is important that we obtain expression profiles from as many tissue types as possible. Fortunately, the GTEx consortium provided a comprehensive catalog of gene expression profiles from 49 tissue types [2]. This fascinating resource has already led to important computational tools and resources such as PrediXcan that study the effects of gene expression variation on human phenotypes [8, 9]. On the other hand, a better curated and more comprehensive collection of disease-associated genes has been reported, such as DisGeNET [10]. We believe that these resources offer an unprecedented opportunity to study affected tissues of complex diseases. In this study, we present disease tissue detection using logistic regression (DEDUCE), a novel computational approach that integrates information from GTEx and DisGeNET to infer disease-relevant tissues. To maximize information contained in the two sources, we introduced a logistic regression-based method that performs tissue type inference leveraging information from both tissue-specific marker genes and disease-associated genes. Compared to existing methods, our method is unique in that for each disease, we consider the relative order of disease-relatedness among all the disease-associated genes using the gene–disease-association (gda) scores provided by DisGeNET. Additionally, DEDUCE calculates a tissue-specific enrichment score for every gene in every tissue based on the gene expression measures across tissues and extracts the top 200 genes with the highest enrichment scores as the tissue-specific marker genes. The logistic regression strategy has been successfully utilized in bioinformatics across various contexts [11, 12]. To evaluate DEDUCE, we applied it to 978 diseases across 47 tissue types. Our results demonstrate that the identified disease–tissue associations are biologically meaningful. The relevance of the top disease–tissue pairs is significantly higher than existing methods. To make the results easier to access, we cataloged all the significant pairs and developed an R Shiny app, enabling users to explore the data by disease and tissue. We anticipate this resource will enhance our understanding of disease pathogenesis and its relationship with human tissues. Materials and methods DisGeNET For this study, disease-associated genes along with their gda scores were obtained from the DisGeNET database as the input for DEDUCE. The gda scores are calculated based on various factors, including the number and types of sources and the number of publications supporting each gda [10]. A total of 3261 324 gdas were retrieved from DisGeNET, which covers 30,710 diseases and 21,666 genes. Some of the gdas appear more than once in this database. For these cases, we took the median gda score for the gene–disease pair. To ensure proper functionality of DEDUCE, we require each disease to have at least 200 disease-associated genes. There are 978 diseases that pass this filter. A complete list of these diseases, along with detailed information, is summarized in Supplementary Table S1. GTEx In this study, we utilized multitissue gene expression data collected by GTEx (V8 release) to derive tissue-specific marker genes [2]. We excluded tissues with a sample size <70, as well as cell-cultured fibroblasts and Epstein–Barr virus-transformed lymphocyte tissues, which resulted in 47 tissues. The complete list of these 47 tissues and the total number of samples in each tissue are shown in Supplementary Table S2. Tissue-specific marker gene selection To identify tissue-specific marker genes, we adopted a strategy originally developed for defining cell subtype-specific marker genes [12]. For each tissue, an enrichment score was calculated for every gene by multiplying the tissue-specific gene expression level by the fraction of samples that express the gene in this tissue. The tissue-specific gene expression level was determined by dividing the average gene expression level within the tissue by the average gene expression level across all tissues. The fraction of samples expressing the gene serves as a penalty to adjust the enrichment score. More details of this method can be found in the literature [12, 13] and in the Supplementary data. Lists of tissue-specific genes across 47 tissues can be found in Supplementary Table S3. For each tissue, the top genes ranked by the enrichment scores are considered tissue-specific marker genes. For easy and fair comparison, we use the same number of genes for all 47 tissues. In this study, we use 200 marker genes. We have also tested using 100 marker genes for each tissue and found that the results are largely comparable. DEDUCE method overview For each disease, we first compiled all associated genes along with their gda scores from DisGeNET. Next, we calculate the tissue enrichment scores of these genes based on GTEx expression across all tissue types. We then convert these scores into binary indicators, where a gene is considered a tissue-specific marker gene if it ranks among the top 200 in enrichment scores for that tissue. As demonstrated in Fig. 1, DEDUCE treats the gene’s gda score as the explanatory variable and treats the binary indicator of whether the gene is a tissue-specific marker gene as the response variable. Their relationship can be modeled using logistic regression. For each disease-associated gene , we have Overview of the DEDUCE method. In the first step, for a specific disease, all disease-relevant genes according to DisGeNET were ranked from high to low by their gda scores. In the meantime, for each tissue, we identify a list of 200 tissue-specific genes. Therefore, for this tissue, all genes in the genome are dichotomized as either tissue-specific or not. Next, a logistic regression is conducted to test whether there is an association between the gda scores and tissue specificity. The P-value associated with the independent variable β from the logistic regression is used to rank the tissues. The top-ranked tissues are deemed most likely affected by the disease. where In which indicates that the gene is a tissue-specific marker gene and otherwise; is the gda score of the gene . We apply DEDUCE to each tissue type independently. The involvement of a specific tissue in a disease is assessed by testing the null hypothesis that against the alternative that using the Wald test. We use the P-value of the test to measure the strength of association between the disease and the tissue. Performance comparison To evaluate DEDUCE’s effectiveness in identifying and ranking disease–tissue associations, we compared its performance with Tissue-Specific Enrichment Analysis Database (TSEA-DB) [14], a well-established tool for similar analyses. Performance comparison was achieved by comparing the “closeness” between the disease of interest and its most related tissue identified by competing methods. Closer relationships imply better performance. Using word embedding, any word or phrase, such as a disease or tissue name, is mapped to a vector in the embedded space, and vectors corresponding to semantically related phrases are closer than those of unrelated phrases (Fig. 2A). In this study, we use word embeddings provided by Claude 3.5 (Anthropic PBC, San Francisco, CA), a large language model. Specifically, we used the text embedding model voyage-2 from the Python package voyageai, which has 4000 tokens and 1024 embedding dimensions. Meaning each disease or tissue name is converted to a vector of 1024 floating point numbers representing the embedding of the text in the model. Performance comparison between DEDUCE and TSEA. (A) Illustration of how performance comparison is conducted in this study. The whole procedure can be broken down into three steps. In the first step, both disease names and the top tissues identified by either DEDUCE or TSEA were mapped to a common embedded space using the word embedding function provided by Claude 3.5. Then, cosine similarity is calculated between each disease and its most related tissue identified by either method. The two groups of cosine similarity measures (numbers appear in the tables are not actual values, they are just for illustration purpose) were compared using the Wilcoxon signed rank test. (B) A scatter plot demonstrating the correlation between two different measures of the strength of connection for a disease–tissue pair. For all 916 diseases analyzed and their top-ranked tissue of association, we calculated the cosine similarity between the embeddings of the disease and the tissue names, as well as the minus log transformed logistic regression P-values for the predictor variable. (C) A side-by-side boxplot demonstrating the distribution of cosine similarity scores reported by the two competing methods. We use cosine similarity to measure relationships between two vectors. Cosine similarity values range from 0 to 1, where a higher value indicates stronger agreement between the disease name and the predicted tissue. Finally, we apply the Wilcoxon signed-rank test to compare the two sets of cosine similarity measures between diseases and their top-ranked affected tissues identified by DEDUCE and tissue-specific enrichment analysis (TSEA). Results Top significant disease–tissue pairs Using a significant threshold of P-value ≤ .001, we identified 126 disease–tissue associations that passed this threshold and summarized them in Supplementary Table S4. The list includes 120 unique diseases and 37 unique tissues. The table highlights diseases and their most relevant tissues, such as cardiomyopathies with the left ventricle and obesity with visceral adipose tissue. The top 20 most significant disease–tissue associations are listed in Table 1. Table 1.Top 20 most significant disease–tissue associationsDiseaseTissue P-valueCardiomyopathiesLeft ventricle (heart)3.61219E−15CardiomyopathiesAtrial appendage (heart)9.65387E−12Hypertrophic cardiomyopathyLeft ventricle (heart)1.26545E−11Hypertrophic cardiomyopathyAtrial appendage (heart)1.74409E−10ObesityVisceral omentum (adipose)3.53100E−10Kidney failure, acuteLiver6.68106E−10Atrial fibrillationLeft ventricle (heart)7.77857E−10Mental depressionHypothalamus (brain)1.09234E−08Congestive heart failureSubcutaneous (adipose)2.51258E−08EpilepsyFrontal cortex BA9 (brain)3.81446E−08Heart failureSubcutaneous (adipose)4.51008E−08Hypertrophic cardiomyopathySkeletal muscle5.54254E−08ObesityMammary gland (breast)5.7508E−08Mental depressionCaudate basal ganglia (brain)6.86263E−08Cardiovascular diseasesLiver1.02911E−07Liver carcinomaLiver1.42086E−07Depressive disorderHypothalamus (brain)1.42987E−07Cardiomyopathy, dilatedLeft ventricle (heart)2.26647E−07Depressive disorderCaudate basal ganglia (brain)2.49577E−07Prostate carcinomaProstate2.81386E−07 Top 20 most significant disease–tissue associations Among the most statistically significant disease–tissue pairs, many are heart-related diseases, such as cardiomyopathy and congestive heart failure, which show strong associations with heart tissues like the left ventricle and atrial appendages. These associations align well with prior biological knowledge and clinical observations. To highlight the most compelling results, we further examined some of the most prominent disease–tissue pairs for select diseases and tissues. Results from selected diseases Parkinson’s disease Parkinson’s disease (PD) is the second most common neurodegenerative disorder after AD, characterized by uncontrollable movements and impaired balance and coordination. The primary pathological hallmark of PD is the loss of dopaminergic neurons in the substantia nigra pars compacta [15]. Consistent with this, our analysis identified the substantia nigra as the only significantly enriched for PD (Fig. 3A). Among the most enriched overlapping marker genes, SLC6A3 (which encodes dopamine transporter) and tyrosine hydroxylase (TH) play critical roles in dopamine biosynthesis and catecholamine metabolism, both of which are crucial in PD pathogenesis. SLC6A3 and TH were found to be downregulated in a 6-hydroxydopamine rat PD model [16]. A statistically significant under-expression of SLC6A3 and TH was found in the substantia nigra of PD patients compared to controls in single-gene level analysis, further supporting the validity of our approach [17]. DEDUCE results on the top related tissues of selected diseases. (A) Top tissues related to PD. (B) Top tissues related to AD. (C) Top tissues related to Schizophrenia. (D) Top tissues related to bipolar disease. Here, red color indicates a significant disease–tissue association, and blue color indicates a nonsignificant disease–tissue association. Alzheimer’s disease Alzheimer’s disease (AD) is the most common cause of dementia, leading to a slow but progressive decline in memory and cognition. It affects 6.9 million individuals aged 65 and older in the United States and 55 million people globally [18]. Interestingly, among all the tissues we tested, visceral omental adipose tissue was the only one passing the significance threshold for AD (Fig. 3B and Supplementary Table S4). The visceral adipose tissues have been shown to participate in the pathogenesis of AD in recent studies. Notably, the proinflammatory adipokine leptin, secreted by the visceral white adipose tissue, has been shown to reduce amyloid-beta (Aβ) accumulation in the hippocampus when applied to AD model mice [19]. Since Aβ levels in different brain regions are indicative of AD progression, leptin’s ability to decrease plaque formation and neurodegeneration provides insights into potential strategies for AD prevention [19, 20]. Other hormones secreted by visceral white adipose tissues, such as adiponectin and resistin, have also shown associations with AD. Adiponectin positively correlates with Aβ deposition and cognitive decline, while resistin appears to protect neurons from Aβ-induced oxidative stress [21, 22]. Additionally, population studies indicate that individuals with higher levels of visceral adipose tissue exhibited a relatively higher risk of developing AD [23]. Collectively, these findings support the significant association between AD and the visceral omental adipose tissue. Schizophrenia Schizophrenia is a complex psychiatric disorder characterized by hallucinations, delusions, and disorganized thinking and behavior. Its origins are thought to result from genetic and environmental disruptions in brain development [24]. Our analysis identified 10 significant tissues for Schizophrenia, all of which are brain tissues. Among them, the putamen tissue was the most significant (Fig. 3C). Schizophrenia patients have been observed to exhibit a larger putamen compared to controls, with better-functioning patients having even larger putamen sizes than those requiring others’ assistance in fulfilling basic needs [25]. This suggests that the putamen may contribute to medical resistance in poor outcome patients and plays different roles at various disease stages. Moreover, multiple genetic variants associated with schizophrenia have been linked to putamen volume, with such associations being attenuated in schizophrenia patients [26, 27]. A functional Magnetic Resonance Imaging (fMRI) study found that first-episode schizophrenia patients with auditory verbal hallucinations had a significantly lower amplitude of low fluctuation in their left putamen compared to those without this symptom, indicating a possible connection between the tissue and symptom manifestation [28]. Additionally, patients experiencing both auditory and visual hallucinations exhibited stronger functional connectivity between bilateral putamen and nucleus accumbens than those with only auditory hallucinations [29]. These findings collectively underscore a strong connection between the putamen and schizophrenia. Bipolar disorder Bipolar disorder is a psychiatric condition characterized by alternating periods of depression and abnormally elevated mood [30–32]. Our results demonstrate that the brain’s frontal cortex tissue has the most significant association with bipolar disorder (Fig. 3D). Studies revealed the increase of messenger RNA and protein levels of arachidonic acid metabolic cascade markers like COX-2 and mPGES, which regulate signal transduction, apoptosis, and protection of oxidative stress in the frontal cortex of bipolar disorder patients [33]. Notably, four FDA-approved mood stabilizers for bipolar disorder have been shown to repress the arachidonic acid pathway in brain tissues, further indicating its involvement in bipolar disorder’s pathogenesis at the frontal cortex [33]. Epigenetic studies have found unique hypermethylation at the promoter regions of the drebrin-like protein gene in frontal cortex samples from bipolar disorder patients. Functionally, bipolar disorder patients exhibit reduced frontal cortex activation when controlling negative facial expressions, potentially contributing to their inability to regulate emotions [34]. Results from selected tissues While analyzing disease-centric associations provides valuable insights into how specific disorders manifest in different tissues, a complementary approach is to examine the perspective from individual tissues. This allows us to identify which diseases are most significantly linked to specific tissues, highlighting key biological interactions that may not be immediately apparent in a disease-focused analysis. Below, we explore notable findings for several tissues, emphasizing their strongest disease associations. Kidney cortex Top tissues associated with the kidney cortex are shown in Fig. 4A. Beyond kidney diseases, diabetes shows significant enrichment in the kidney cortex tissue, a connection that has been well established in the literature. In diabetic patients, mitochondrial glucose oxidation in kidney cells is reduced due to the inhibition of pyruvate dehydrogenase activity, with preferential oxidation of fatty acids and ketone bodies. This metabolic shift increases reactive oxygen species levels, primarily due to enhanced H2O2 production by the nicotinamide adenine dinucleotide phosphate (NADPH) oxidase Nox4 [35]. Additionally, diabetic tubule mitochondria exhibit low pyruvate and malate oxidation, likely due to decreased pyruvate transporter expression [35]. There was a high level of lipid substrate oxidization in diabetic tubule mitochondria regardless of the rate-controlling fatty acid oxidizing enzyme CPT1, indicating that CPT1 has a lower level of control in fatty acid oxidation within diabetic kidneys [35]. DEDUCE results on the top related diseases of selected tissues. (A) The top diseases related to the kidney cortex. (B) The top diseases related to the spleen. (C) The top diseases related to the pancreas. Here, red color indicates a significant disease–tissue association, and blue color indicates a nonsignificant disease–tissue association. A key finding from our kidney cortex analysis is that transcription factors hepatocyte nuclear factor 1 homeobox A (HNF1A) and homeobox B (HNF1B) have the highest gda score within the overlaps between disease-associated genes and kidney-specific marker genes. These two genes play critical roles in maturity-onset diabetes of the young (MODY), where their mutations contribute to renal phenotypes independent of diabetic nephropathy. Notably, 20% of the type I diabetes patients carry a mutation in either the HNF1A or HNF1B gene [36]. Previous studies have shown that heterozygous mutations in HNF1A at chromosome 12q can cause MODY type 3, an autosomal dominant noninsulin-dependent diabetes mellitus [37, 38]. HNF1A regulates megalin and cubilin transcription, and its mutations could lead to the loss of these two proteins in proximal tubule cells, impairing receptor-mediated endocytosis [38]. Similarly, 86% of subjects that carry HNF1B mutations experience renal impairment, with 58% also diagnosed with diabetes [39]. Spleen For the spleen tissue, as shown in Fig. 4B, four diseases surpassed the significance threshold, with myositis ranking the highest. Previous studies have demonstrated multiple associations between spleen tissues and various forms of myositis. In mice, injection with syngeneic muscle antigens induced allergic myositis, leading to significant blast transformation responses in spleen cells [40]. Additionally, patients with polymyositis and dermatomyositis exhibited antibodies that precipitate with antigens in human spleen tissue extracts [41]. Further supporting this connection, Tamir et al. found a strong positive correlation between urinary nitrate levels and muscle inflammation in 10-month-old Swiss Jim Lambert (SJL)/J mice, an inbred mouse strain, in which the source of the nitric oxide was the inducible nitric oxide synthase (iNOS) found in spleen cells [42]. It is probable that the iNOS originates from the invaded macrophage, based on morphological evidence of cells exhibiting iNOS [42]. Additionally, mice infected by the chikungunya virus, which could induce muscle inflammation and lead to viral myositis, have shown a high level of lymphoid proliferation and hemorrhage in the spleen compared to controls, further strengthening the connection between immune response in spleen tissue and myositis [43]. However, CD8+ cytotoxic T cells, a type of white blood cell contributing substantially to the inflammatory response of inclusion body myositis, did not demonstrate a statistically significant difference between the muscle and spleen tissues of inclusion body myositis patients [44]. This suggests that the spleen may participate in the development of myositis as much as muscle tissues do give similar levels of strongly involved immune-related cells. Hepatitis ranks as the second most relevant disease linked to the spleen. Among patients with both acute and chronic hepatitis B, viral DNA and RNA were detected in the spleen tissue, albeit at a relatively lower level compared to the liver tissue, indicating the spleen may serve as a viral reservoir reinfection source [45, 46]. Shear-wave elastography studies have further demonstrated a strong positive association between spleen stiffness and fibrosis stage in chronic hepatitis B patients [47]. This correlation is relatively weak during early stages from F1 to F3 but becomes strong at stage F4, which suggests that the spleen involvement in the development of chronic hepatitis B may occur after fibrosis forms nodules and changes liver architecture [47, 48]. In addition, Yan et al. have built a model using liver/spleen volume ratio and other spleen morphological variables to predict hepatic venous pressure gradient among hepatitis B patients and yielded high accuracy, demonstrating a significant association between spleen and hepatitis B [49]. Pancreas As expected, all diseases surpassing the significance threshold for the pancreas tissue are directly related to pancreatic function (Fig. 4C). Pancreatitis emerged as the most significant disease association, with high significance in the pancreas, with chronic pancreatitis ranking 1st and general pancreatitis ranking 7th. Chronic pancreatitis is primarily caused by progressive inflammation of the pancreas that impairs its structure and function [50]. One major risk factor is mutations in the cystic fibrosis transmembrane conductance receptor (CFTR) gene, which disrupt the formation of cyclic AMP-regulated chloride channels [50]. For both chronic pancreatitis and general pancreatitis, CFTR has shown to overlap between pancrea-specific marker genes and their disease-associated genes and demonstrated relatively high gda scores, confirming a strong role they play in the pathogenesis of pancreatitis and high functionality of DEDUCE. The pancreas also produces glucagon and insulin, the two hormones controlling carbohydrate metabolism in the human body. Since diabetes is a disease in which the body fails to respond to or produce sufficient insulin, the pathology of both type I and type II diabetes involves deep pancreatic malfunction. In type I diabetes, the loss of islet beta cells and inflammation of pancreatic endocrine tissues drive disease progression [51]. In contrast, type II diabetes is characterized by a reduction in beta cell mass and increased pancreatic fat content [51]. These strong associations support the high enrichments of diabetes mellitus and hyperglycemia in our results. An unexpected finding is the association between dwarfism and the pancreas. Although the enrichment score fails to pass the significance threshold, dwarfism still ranks the 10th most significant, following multiple forms of pancreatitis and diabetes, vascular disease, and neoplasm (Fig. 4C). Several case studies support this connection. In an early study of prenatal ultrasound diagnosis, researchers found a case in which polycystic pancreas and short-limb dwarfism coexist [52]. Furthermore, 30% of infants with congenital pancreatic cysts exhibit body anomalies such as short-limb dwarfism and Jeune syndrome, a type of congenital dwarfism with a deformed chest wall [53]. The autopsy of an infant with Saldino–Noonan short-ribbed polydactyly disorder, a rare and lethal dwarfism, revealed enlarged cystic pancreatic formations, strengthening the connection between dwarfism and pancreatic abnormalities [54]. Another case report of an infant with asphyxiating thoracic dysplasia and short-limbed dwarfism also confirmed such a connection with multiple congenital pancreatic cysts found in necropsy [55]. Additionally, the co-occurrence of dwarfism and exocrine pancreatic insufficiency has been observed as well, indicating a more complex relationship between pancreas tissue and dwarfism [56]. Comparing with TSEA To assess the predictive accuracy of DEDUCE, we compared its results with those from other competing methods such as TSEA-DB, which employs the deTS method [14]. To quantify the relationships between diseases and affected tissues, as described in the “Materials and methods” section, we mapped all disease and tissue names to an embedded space, and then calculated cosine similarity scores between diseases and their most affected tissues. To justify the validity of using cosine similarity to measure disease tissue connection, we investigated the relationship between the logistic regression goodness-of-fit measure and the cosine similarity measure. To be specific, given a disease, we hypothesized that if a tissue is affected by the disease, the odds of being a tissue-specific gene is more correlated with the gene’s gda score. Additionally, the cosine similarity between the embeddings of the disease and the tissue is higher, compared to a tissue that is unaffected by the disease. To check this, we calculated cosine similarity between a disease and its top-ranked tissue and retrieved the logistic regression P-values for all 916 diseases we analyzed. Indeed, we found that there is a significant positive correlation between cosine similarity scores and the negative log transformed logistic regression P-values with R2 of 0.074 and P-value ≈ 5.561e−17. (Fig. 2B). The distribution of cosine similarity measures between diseases and their most affected tissues across all diseases for both DEDUCE and TSEA methods is presented in Fig. 2C. We found that overall the cosine similarity measures returned by DEDUCE are higher than those returned by TSEA. When applied the Wilcoxon signed-rank test to compare the two sets of cosine similarity measures returned by DEDUCE and TSEA, respectively. The result is highly significant with P-value 1.408e−06, suggesting that DEDUCE provides a more accurate match between disease and relevant tissues than TSEA, reinforcing its effectiveness in tissue–trait association analysis. The DEDUCE Shiny App To make the DEDUCE analysis results easily accessible, we developed an interactive R Shiny web interface. Users can pick diseases or tissues of interest and see results from the DEDUCE analysis in real time, including logistic regression, gene enrichment, and statistics. The app covers over 30,170 diseases, 19,296 genes, and 49 tissues, showing more than 3.2 million gdas. All graphs and summary statistics are publicly accessible and help users understand which tissues are most affected by diseases and which genes are involved. All graphs and data can be accessed at https://arashied92.shinyapps.io/shiny-app/. At the top of the page, users can choose a disease or tissue from dropdown menus. The app shows bar plots of related diseases or tissues, making it easy to explore these relationships. The disease plot page presents a horizontal bar graph of all 47 tissues with the x axis being negative-log transformed logistic regression P-values. Bars of tissues passing the significance threshold are painted red. In the disease table page, the table illustrates enrichment information of the selected disease as shown in the disease plot. For each tissue, we present a horizontal bar graph of the top 30 most significantly enriched diseases and a table of all diseases’ enrichment information on the tissue table page. Documentation on DEDUCE and contact information are also provided. Discussion Identifying the affected tissues of human diseases is crucial for understanding disease etiology and progression. The advent of biomedical big data has enabled new approaches to address this challenge. In this study, we present DEDUCE, a novel strategy that links two sources of genomics big data using a logistic regression framework. By integrating disease-association information from DisGeNET and tissue-specific gene expression data from GTEx, DEDUCE accurately identifies key tissues involved in complex diseases, providing insights into disease pathogenesis and potential therapeutic targets. For instance, in the case of AD, although thousands of genes are implicated, DEDUCE was able to refine the analysis and pinpoint only a few relevant tissues. This generally enhances the precision of disease research and provides clearer guidance for therapeutic strategies, making it a valuable tool for precision medicine. Several existing methods leverage tissue-specific expression data and disease-associated genes to infer relevant tissues for diseases [14]. A common strategy is to use the enrichment method, similar to the one used in Gene Ontology (GO) term enrichment analysis [57]. While such an approach is capable of revealing connections between tissues and diseases, it does not take into account the strength of the gdas, potentially missing important insights. DEDUCE overcomes this limitation by incorporating gda scores, representing the association of each gene with the disease as a continuous variable, thus providing a more quantitative and nuanced evaluation. This means that DEDUCE is able to offer a more quantitative advantage in handling tissue–disease associations. This quantitative response allows us to prioritize tissues based on their level of involvement in the disease, rather than merely categorizing them as “significant” or “nonsignificant.” This principle is akin to LRpath [11], which demonstrated improved performance over enrichment-based method in pathway and gene set analysis. By applying a logistic regression model, DEDUCE treats gda scores as explanatory variables and the binary status of tissue-specific marker genes as the response variable, allowing for more nuanced modeling. This approach enables us to capture gene enrichment within specific tissues and quantify the strength of these associations, providing more reliable estimates of tissue–disease relationships. Logistic regression is particularly well-suited for handling binary response data, making the model more applicable to large-scale gene expression analyses. This results in a more robust statistical inference process with P-values, reducing errors caused by data noise or sample size imbalances, and enhancing the interpretability and credibility of the identified tissue–disease associations. When comparing DEDUCE with competing methods, we observed similarities for some disease–tissue associations. However, DEDUCE is capable of further refining these associations, capturing subtle differences that the binary enrichment approach might overlook. By incorporating the strength of gdas, our approach offers a more detailed and quantitative perspective on tissue involvement in disease. The DEDUCE method is built on the assumption that for a disease/tissue pair, there is a linear relationship between the log odds of being a tissue-specific marker gene and the gene’s gda score. This assumption is the basis for using logistic regression. We have tested a few diseases and found the assumption is reasonable for the tissue that is affected by a disease. We acknowledge that this assumption is perhaps oversimplified, because we ignore more complex relationships between the log odds of being a tissue-specific marker gene and the gene’s gda score. For future work, we will explore more flexible models. One potential idea is that similar to Spearman’s correlation, we can use ranks of the gda scores, instead of their actual values in the logistic regression. For a small proportion of diseases, the gda scores are identical for all disease-associated genes. Among the 978 diseases we studied, 62 (6.3%) of them (Supplementary Table S1B) belong to this category. For such diseases, the model described earlier cannot be applied. An alternative strategy is to use the binary indicator of whether a gene is associated with the disease as the response variable and treat the gene’s enrichment score for the tissue as the explanatory variable. In essence, we reversed the variable’s role in the model. Note that this is a temporary stop-gap solution. With more studies to be conducted, we anticipate that we will likely see some differences among these gda scores in the future release of the DisGeNET data. One limitation of DEDUCE is its reliance on gda scores, meaning its accuracy depends on the completeness and quality of databases like DisGeNET. For less-studied diseases or those with limited data, DEDUCE results might not fully reflect the actual tissue–disease relationships. This means that our quantitative analysis is constrained by the range and quality of current data, potentially limiting the comprehensiveness of the findings. Moreover, noise in gene expression data, such as the presence of pseudogenes, could affect the accuracy of the analysis, particularly in the selection of tissue marker genes where sensitivity to threshold settings plays a role. The DEDUCE method has broad potential applications in future research, particularly in analyzing the impact of gene changes on diseases and organs. For example, we have used DEDUCE to identify that COVID-19-related genes are primarily concentrated in lung tissues, which brings insights that are crucial for understanding the respiratory impacts of COVID-19. However, as the virus evolves, other organs such as the heart or kidneys might become increasingly involved. DEDUCE can help predict these changes and identify which tissues are potentially under impact in disease progression. In this study, we chose P-value 0.001 as the significance threshold. We chose the threshold empirically to balance discovery power and type I error. An effective way to adjust for multiple testing is the false discovery rate (FDR) [58]. In this study, we used R package qvalue [59] to perform FDR estimation. The q-values for the 126 top disease–tissue associations are included in Supplementary Table S4. Lastly, we used 200 marker genes for each tissue in our model. We have tested other options such as 100, the results are similar, and 200 marker genes are marginally better. To accommodate different research needs, we have included both options in the DEDUCE Shiny app for flexibility in tissue-specific analysis. Supplementary Material Acknowledgements We thank members of the Qin and Jiang research groups for helpful comments and suggestions. Author contributions: Boqi Wang (Data curation [lead], Formal analysis [lead], Investigation [lead], Methodology [lead], Software [supporting], Writing – original draft [supporting], Writing – review & editing [supporting]), Jiayi Wang (Data curation [supporting], Formal analysis [supporting], Investigation [supporting], Methodology [supporting], Software [supporting], Validation [lead], Visualization [lead], Writing – original draft [lead], Writing – review & editing [lead]), Ammar Aleem Rashied (Resources [lead], Software [lead], Visualization [lead]), Bo Meng (Investigation [supporting], Validation [supporting], Visualization [supporting], Writing – review & editing [supporting]), Jesse Zhang (Formal Analysis [supporting], Investigation [supporting], Validation [supporting], Visualization [supporting], Writing – review & editing [supporting]), Jun S. Liu (Conceptualization [supporting], Supervision [supporting]), and Jie Jiang (Conceptualization [lead], Funding acquisition [supporting], Investigation [supporting], Methodology [supporting], Project administration [supporting], Resources [supporting], Supervision [supporting], Validation [supporting], Writing – original draft [supporting], Writing – review & editing [lead]) Contributor Information Boqi Wang, Bioinformatics and System Biology Graduate Program, University of California at San Diego, La Jolla, CA 92093, United States. Jiayi Wang, College of Arts and Sciences, Emory University, Atlanta, GA 30322, United States. Ammar Aleem Rashied, Department of Biostatistics and Bioinformatics, Rollins School of Public Health, Emory University, Atlanta, GA 30322, United States. Bo Meng, Department of Pharmacology and Chemical Biology, School of Medicine, Emory University, Atlanta, GA 30322, United States. Jesse Zhang, College of Arts and Sciences, Emory University, Atlanta, GA 30322, United States. Jun S Liu, Department of Statistics, Harvard University, Cambridge, MA 02138, United States. Jie Jiang, Department of Cell Biology, School of Medicine, Emory University, Atlanta, GA 30322, United States. Zhaohui S Qin, Department of Biostatistics and Bioinformatics, Rollins School of Public Health, Emory University, Atlanta, GA 30322, United States. Supplementary data Supplementary data is available at NAR Genomics & Bioinformatics online. Conflict of interest None declared. Funding JJ and ZSQ are partially supported by NIH R21NS133960, and ZSQ is partially supported by NIH R01GM141392. Data availability The original GTEx data used for this study can be found at https://gtexportal.org/. Disease-associated genes along with their gene-disease associations (gda) scores were obtained from the DisGeNET https://disgenet.com/. References Associated Data Supplementary Materials Data Availability Statement The original GTEx data used for this study can be found at https://gtexportal.org/. Disease-associated genes along with their gene-disease associations (gda) scores were obtained from the DisGeNET https://disgenet.com/.