Genomics: genetic basis of CVDs

Genomic insights are particularly valuable for understanding MI, as they reveal genetic risk factors and mechanisms underlying the disease. Large-scale genome-wide association studies (GWAS) have identified over 200 loci linked to the risk of MI and coronary artery disease (CAD) by impacting lipid metabolism, blood pressure regulation, inflammation, and platelet function [7–12]. Notably, all genes identified—except ABO—were associated with risk for both MI and CAD rather than one condition independently. Recent GWAS analysis of data from approximately 639,000 individuals revealed eight new loci with significant associations with MI; six of these loci exhibited stronger associations with MI than CAD without MI [13]. For instance, the locus on chromosome 1p21.3, which includes the SLC44A3 gene, showed a specific association with MI in the presence of CAD, rather than with coronary atherosclerosis alone, confirmed by independent replication analyses. Functional and bioinformatic studies suggest SLC44A3 may influence plaque rupture and thrombosis directly within the vascular wall, indicated by its expression in MI-relevant vascular tissues and co-localization with atherosclerotic aorta. These findings demonstrate how GWAS can uncover novel biological targets relevant to MI.

Transcriptomics and epigenetics: gene expression and regulation of CVDs

Transcriptomic studies provide crucial insights into the molecular response of various cell types to MI, capturing differentially expressed genes (DEGs) involved in injury and repair processes. These changes reveal acute responses that exacerbate tissue damage and highlight potential therapeutic targets. For example, single-cell RNA sequencing (scRNA-seq) has identified beta-2 microglobulin (B2M) as a key regulator of immune response and tissue repair post-MI, making it a promising therapeutic target [14].

Transcriptomic profiling has also been used to understand ventricular remodeling and systemic effects following MI. Spatial transcriptomics combined with single-nucleus RNA sequencing revealed the expression of mechanical stress response genes like Csrp3 in the infarct border zone, underscoring their role in cardiac remodeling [15]. A genome-wide transcriptomic analysis in murine models highlighted dysregulated biological processes such as immune responses, mitochondrial dysfunction, and RNA and protein processing in tissues like the heart and liver post-MI [16]. In human studies, MI zone transcriptomics identified 1,785 DEGs, with significant upregulation of genes related to apoptosis, necroptosis, and necrosis, providing a detailed view of cell death mechanisms in infarcted tissue [17].

Epigenomics continues to add another layer of complexity to CVD research. Epigenomics examines heritable changes in gene expression that occur without alterations to the DNA sequence, focusing on mechanisms such as DNA methylation, histone modifications, and chromatin remodeling. An epigenome-wide association study (EWAS) identified 211 differentially methylated CpG sites associated with genes linked to cardiac function, CVDs, and recovery post-ischemic injury in individuals with a history of MI [18]. Additionally, nested case-control studies have mapped CpGs to genes involved in signal transduction, metabolism, and gene expression pathways, several of which are associated with MI-related conditions such as HF, stroke, and cardiac aging [19].

High-resolution integrative molecular mapping has further advanced the field. Integrating transcriptomics with epigenomics, such as DNA methylation and histone modifications, further elucidates regulatory mechanisms underlying gene expression shifts, with studies linking disruptions in these processes to pathways associated with muscle contraction and inflammation [20]. A study using single-cell gene expression, chromatin accessibility, and spatial transcriptomics in myocardium from MI patients identified distinct molecular profiles across physiological zones and time points, paving the way for novel therapeutic approaches [21]. Similarly, transcriptome and genome-wide DNA methylation analysis in acute MI mouse models revealed critical stages of differential methylation and gene expression, with the 6-h time point showing drastic changes and enriched pathways associated with the acute phase of MI. These findings were validated through in vitro studies, demonstrating regulation by DNA methylation [22]. These integrative approaches provide a comprehensive understanding of the molecular landscape in MI and HF, offering valuable biomarkers and therapeutic targets.

Proteomics: profiling proteins in CVD research

Proteomics, the large-scale study of proteins, enables the identification of disease-associated biomarkers and the discovery of novel drug targets. By providing insights into protein structure, function, and interactions, proteomics is a powerful tool for understanding disease mechanisms and therapeutic development. In MI, proteomics has revealed key protein changes and associated pathways. A study in male Sprague-Dawley rats identified 169 differentially expressed proteins (DEPs), with 52 upregulated and 117 downregulated, linked to pathways such as complement and coagulation cascades, MAPK signaling, and actin cytoskeleton regulation [23]. Similarly, in a comparison of MI patients and those with stable CAD, 198 DEPs were identified, including proteins involved in GDP and GTP binding, coenzyme A biosynthesis, and AMPK signaling pathways [24]. Proteomics has also been instrumental in studying rare MI complications such as cardiac rupture, identifying 958 DEPs that were enriched in pathways like RNA transport, ribosome function, and necroptosis [25].

Beyond the heart, proteomics has highlighted the systemic effects of MI, including neuroinflammation. Following permanent ligation of the left anterior descending coronary artery in rats, several proteins associated with brain disease were identified in the brain cortex, including targets linked to elongation of very long chain fatty acids protein 5 (ELOVL5) and ATP-binding cassette subfamily G member 4 (ABCG4). These findings, confirmed through western blotting and immunofluorescence, underscore the interconnected nature of cardiac and neurological events [26]. Collectively, these studies demonstrate the potential of proteomics to uncover critical molecular insights into MI and its complications, paving the way for novel therapeutic strategies.

Metabolomics: exploring metabolic changes in cardiac health

Metabolomics provides a snapshot of small molecules and metabolites within cells and tissues, offering insights into cellular energy status. The heart, composed of ATP-demanding cells, requires substantial energy to function. To sustain its activity, the adult heart metabolizes fatty acids to produce 50–70% of its ATP requirements [27, 28]. Multi-omics approaches have been instrumental in uncovering the mechanisms underlying HF. For example, proteomic and metabolomic characterization of three commonly used HF mouse models revealed downregulation of fatty acid oxidation and ATP metabolism, identifying four metabolite biomarkers associated with HF-related rehospitalization [29]. These findings underscore the potential of metabolomics to improve prognostic assessment.

Although it is well established that the heart shifts its use of carbon substrates for energy production during disease states [27, 30], the mechanisms driving these metabolic changes remain poorly understood. To address this gap, researchers developed a heart-specific genome-scale metabolic network reconstruction, enabling the interpretation of HF-related gene expression data and the identification of metabolic functions associated with these changes [31]. This approach pinpointed nitric oxide and N-acetylneuraminic acid as common metabolic markers of HF, offering new insights into disease progression. Together, these studies highlight the critical role of metabolomics in advancing our understanding of HF. By integrating metabolic data with other omics layers, researchers can uncover novel biomarkers and therapeutic targets, paving the way for precision medicine approaches in CVD.

XAI in oncology

Given the wealth of available omics data, oncology is one of the leading applications of XAI. In breast cancer, researchers have recently developed a new interpretable XAI model to classify breast cancer subtypes using three omics datasets comprising transcriptomics, DNA methylation, and microRNA expression [39]. By considering the biological relationships between the omics data, such as the inhibitory effects of DNA methylation and microRNA, the authors developed moBRCA-net, an interpretable convolutional neural network (CNN). The authors identified features, such as DEGs, associated with meaningful cross-omics layers for further analysis. CNN layers are then used to extract the high-level patterns and spatial relationships from processed omics datasets to determine breast cancer subtypes. Thus, a tool such as this can identify which molecular targets (i.e. regulators of epigenetics, transcriptomics, etc.) to develop drug candidates for based on the subtype of breast cancer.

In oncology, a large amount of bulk-RNA sequencing data has been generated examining the relationship between gene expression and drugs. However, limited data has been generated examining the relationship between single-cell transcriptomics and drugs. Researchers have recently developed the scDEAL framework, which combines two separate encoder-decoder architectures to harmonize the drug-related bulk RNA-seq data with scRNA-seq data to determine the fine-grained drug response predictions at the cellular level [40]. This process provides insights into cell-type-specific drug sensitivity and resistance, advancing approaches for drug response prediction.

To better understand the biological contribution of several features in cancer, researchers recently developed a transformer XAI that takes mutated genetic sequences and combines it with additional omics data, such as transcriptomics to determine oncology-relevant outcomes such as tumor type [41]. In the process, attention maps are employed to interpret the model’s predictions, highlighting which features, such as specific mutations or gene expression loci, contributed most to the model’s decision.

Thus, a tool such as this provides a holistic view of cancer pathology, informing researchers on relevant drug target pathways. Yet current methods only integrate a small number of omic data types, typically transcriptomics and genomics. A recent study developed the Recon8D model, that combines eight different omics networks and their interactions with metabolic networks in cancer cells [42]. This study uncovered the interplay between genetic, transcriptional, epigenetic, proteomic and post-translational regulatory networks and metabolism. It also provided insights into drug response and drug interactions across various omics layers.

XAI in neurology

Despite having applications in practically all domains of engineering and bioinformatics, CNNs excel in handling visual tasks. Researchers have taken advantage of this by recently developing an interpretable CNN model that is trained on classifying amyotrophic lateral sclerosis (ALS) samples by analyzing images generated from RNA expression data, where RNA values are transformed into pixel representations [43]. The authors then apply SHAP to identify the most influential pixels for ALS classification. This mapping not only enables the classification of ALS, including minority classes, but also highlights genes potentially implicated in ALS molecular pathways. These insights are highly relevant to drug discovery, as identifying genes associated with ALS can inform the development of targeted therapies.

Researchers developed NeuroPpred-Fuse and NeuroPred-FRL to predict neuropeptide structure by encoding peptides as sequence-derived features which capture structural and functional properties relevant to neuropeptide prediction [44, 45]. Both models encode peptide sequences using diverse sequence-derived features, ensuring a comprehensive representation of structural and functional properties. Then, both models incorporate feature selection methods to identify the most relevant features and employ advanced ML classifiers to improve prediction accuracy. Additionally, both models emphasize interpretability by analyzing the weight contribution of selected features, facilitating a better understanding of the biological factors influencing neuropeptide prediction. These approaches collectively advance neuropeptide discovery for therapeutic applications.

Understanding the spatial distribution and heterogeneity of cells is crucial for drug discovery as it reveals how cell types interact within their microenvironment. Researchers developed ACTIONet to better understand cell diversity and interactions from single-cell transcriptomic data by clustering cells into archetypes and mapping how these archetypes relate to each other in a geometric space that reflects similarities or differences between cells [46]. The method then builds adaptable connections between cells based on their local density and relationships. For example, denser regions of cells may form tighter groups, while sparser regions allow more flexibility in how cells are linked. This adaptability ensures that the network accurately reflects the natural variation in cell populations, making it easier to explore both detailed and large-scale cell-state patterns. By applying this method to human brain data, ACTIONet was able to combine information from multiple datasets, define common cell types, and identify important genes tied to these types. This approach makes it easier to study complex tissues like the brain and potentially the heart by providing a clearer picture of cell behavior and function. This enables the identification of targeted therapies that can disrupt disease-specific cell networks or enhance tissue regeneration.

XAI in cardiology

Cardiology has been steadily integrating AI tools over the past decade, particularly in clinical laboratory settings such as cardiovascular imaging and electrophysiology [47]. Despite the application of advanced XAI tools for understanding the molecular mechanisms in oncology and neurology, there are strikingly fewer XAI tools developed for the investigation of cardiovascular molecular mechanisms [5].

Adams et al. [48] developed an ML-based framework using Random Forest Iterative Feature Reduction and Selection (RF-IFRS) to predict the risk of major adverse cardiovascular events (MACE) in statin users by analyzing genome-wide association data. The model identified six genetic variant networks, including genes involved in vasculogenesis, angiogenesis, and vascular stability. This approach incorporates XAI by using decision tree-based visualizations to illustrate the relationships between genetic variants and their contributions to MACE risk. These insights not only predict MACE in statin users with improved accuracy but also identify biologically relevant pathways for CVD progression, offering potential targets for drug development to mitigate MACE risk and advance cardiovascular therapies.

Pezoulas et al. [49] developed a customized XGBoost model to predict carotid artery disease severity by analyzing metabolomics data from the Young Finns Study cohort. XAI is implemented through SHAP which enables researchers to understand how specific vascular and metabolic features contribute to carotid artery disease severity, represented by increased intima-media thickness. The model demonstrates high accuracy and sensitivity in detecting early carotid artery disease stages while providing valuable insights into the metabolic and structural factors underlying disease progression. By highlighting these specific factors, the model not only predicts disease severity with high accuracy but also provides actionable insights into the biological mechanisms driving the disease, which can be targeted in drug development to prevent MI and subsequent HF.

Recently, Dr. Yilmaz [50] developed a Gradient Boosted Trees (GBT) model that can distinguish between acute MI and control groups based on a metabolomics dataset. XAI is implemented through the LIME method, which highlights the contribution of specific metabolites, such as creatinine, cytosine, and myo-inositol, to the model’s predictions. This interpretability allows researchers to understand why certain metabolites are significant for classifying acute MI, enabling the identification of key biomarkers. This approach highlights the biological role of these metabolites in infarct-related metabolic processes, offering potential applications in diagnostics and therapeutic development.

Similarly, DeGroat et al. [51] used a multimodal AI/ML framework to predict CVDs by integrating transcriptomic and genomic data from a clinically integrated dataset. Using XGBoost with Bayesian optimization, the model identified transcriptomic features and pathogenic single nucleotide polymorphisms (SNPs) as predictors, achieving perfect classification in its test cohort. SHAP then provided interpretability, revealing biomarkers such as RPL36AP37 and HBA1 as the most significant contributors. Unlike Yilmaz [50], this model leveraged multi-omics data and identified not just metabolic but also genetic biomarkers. Both studies demonstrate the utility of XAI in cardiovascular biomarker discovery, paving the way for personalized treatments, though they differ in their focus on metabolomic vs. multi-omics integration and the specific cardiovascular conditions they address.

It is important to note that the above examples demonstrate that current XAI applications in cardiology typically involve one or two omics modalities, rather than fully integrated multi-omics. Of the four studies highlighted, three used a single data type (either genomics or metabolomics), and only one [51] combined two data types. This underscores that multi-omics integration in cardiovascular AI studies is still in an early phase of development. In other words, truly harmonized multi-omics analyses for heart disease are a work in progress and represent a promising area for future growth.

To summarize the state-of-the-art methods and their outcomes, Table 1 provides a brief comparison of representative XAI applications in cardiovascular research, including the data types used, AI approaches, and performance achieved.

AI-driven drug repurposing in oncology

Kinases are integral to numerous pathological processes, particularly in cancer, making them attractive targets for therapeutic intervention. Considerable attention has therefore been devoted to the discovery and development of small-molecule kinase inhibitors. For example, a computational platform called VirtualKinomeProfile has been used to profile compound-kinase interactions and predict the inhibitory activity of repurposed drugs, ultimately identifying 19 small-molecule inhibitors of epidermal growth factor receptor (EGFR), hematopoietic cell kinase (HCK), vascular endothelial growth factor receptor 1 (VEGFR1), and mitogen- and stress-activated kinase 1 (MSK1)—key kinases implicated in cancer progression [57]. Another ML-based effort, employing the XGBoost algorithm, identified promising Janus kinase 2 (JAK2) inhibitors from the ZINC database, of which 13 compounds underwent experimental validation and 6 demonstrated IC50 values below 100 nM [58]. Further expanding the scope, a one-drug-multi-target approach was applied to repurpose existing drugs for dual cancer indications, revealing that levosimendan, a phosphodiesterase (PDE) inhibitor effective against HF, also exhibits activity against multiple cancers, including lymphoma, through the direct inhibition of RIOK1 and other kinases [59].

Efforts in drug repurposing have not been limited to kinase inhibitors. A deep learning-driven pipeline integrating transcriptomic data and chemical structures identified pimozide—an anti-dyskinesia agent used for controlling Tourette’s disorder symptoms—as a strong candidate for treating non-small cell lung cancer [60]. Additionally, a ML-based in silico screening technique applied to FDA-approved compounds from DrugBank pinpointed candidates that inhibit the REarranged during Transfection (RET) receptor, an important factor in ligand-independent kinase activation and carcinogenesis [61]. Together, these computational strategies underscore the potential of drug repurposing, not only for identifying effective kinase inhibitors but also for revealing entirely new anticancer applications of existing therapeutics.

AI-driven drug repurposing in neurology

Researchers have increasingly employed ML-based approaches to identify promising candidates for drug repurposing in the context of Alzheimer’s disease (AD). For instance, Rodriguez et al. [62] leveraged computational models to associate disease severity with underlying molecular mechanisms, screening 80 FDA-approved compounds to produce a ranked list of repurposing candidates. Among these, ruxolitinib, a JAK1/2 inhibitor originally developed as a targeted therapeutic for myeloproliferative neoplasms, emerged as a top hit, highlighting the potential of data-driven pipelines to pinpoint existing drugs with new therapeutic applications for AD.

Moreover, dynamic molecular descriptors derived from molecular dynamics simulations were integrated into ML models to identify compounds targeting caspase-8, a protease implicated in AD pathology [63]. Other methodologies have demonstrated similar potential. The use of graph CNNs for identifying inhibitors of the beta-secretase enzyme 1 (BACE1), a key player in amyloid-beta production, has allowed for the capture of subtle structural nuances that traditional descriptors may overlook [64]. These approaches refined candidate prioritization, ultimately guiding the selection of promising drug candidates for experimental validation.

Additional strategies have employed Bayesian ML models, combined with extended connectivity fingerprints, to systematically scan large collections of bioactive compounds—including regulatory-approved drugs available in databases like ChEMBL—for inhibitors of glycogen synthase kinase 3β (GSK3β), an enzyme central to tau phosphorylation in AD [65]. Through this process, ruboxistaurin, originally investigated for other conditions, emerged as a potent GSK3β inhibitor with an IC50 of 97.3 nM. These findings reinforce the utility of computational methods in identifying new therapeutic applications for existing drugs, streamlining the pathway to effective AD interventions.

AI-driven drug repurposing in infectious diseases

As drug-resistant pathogens continue to emerge, both the lack of strong market incentives and diminishing returns on investment in new drug development have made drug repurposing an increasingly attractive strategy in the field of infectious diseases. For example, deep learning models trained on data from 2,335 compounds with their activity against Escherichia coli compounds have been used to virtually screen over 107 million molecules, leading to the discovery of halicin, a compound originally investigated for diabetes treatment, as a potent antibacterial agent [66]. Halicin demonstrated low minimum inhibitory concentrations and broad-spectrum activity, including efficacy against Acinetobacter baumannii in murine models. Similarly, an AI-driven screening of approximately 7,500 compounds identified abaucin as a narrow-spectrum antibacterial candidate targeting A. baumannii, which also effectively controlled infection in mice models [67].

In addition to identifying single-agent therapies, there is increasing interest in designing combination therapies using approved drugs to combat antibiotic resistance. Chemogenomics data has been integrated into ML frameworks to predict synergistic or antagonistic interactions of antibiotic combinations for inhibiting E. coli growth [68]. Transcriptomic data analysis has been employed to discover synergistic drug pairs of repurposed drugs against Mycobacterium tuberculosis [69]. Further refinements have accounted for temporal and sequence-dependent factors influencing drug combination effects [34, 70]. A recent study powered by a transparent ML approach screened all FDA-approved drugs for synergistic activity with existing antibiotics against E. coli. This study discovered the cerebrovascular drug fasudil as a potential synergizer of antibiotics and provided insights into mechanisms of synergy [71]. These advancements illustrate the growing sophistication and promise of computational strategies in optimizing combination therapy design.

A similar trend in drug repurposing emerged during the COVID-19 pandemic. Knowledge-graph-based strategies identified baricitinib, originally developed for the treatment of rheumatoid arthritis, as a candidate for treating COVID-19 by blocking JAK1, JAK2, adaptor associated kinase 1 (AAK1), and cyclin G-associated kinase [72]. In another investigation, a combination of cheminformatics and protein interaction analyses revealed that σ1 and σ2 receptor modulators, such as clemastine and cloperastine, exert antiviral activity against SARS-CoV-2 [73]. ML frameworks incorporating drug-target and protein-protein interaction data singled out ritonavir and rifampicin as potential repurposed agents for COVID-19 [74]. Moreover, virtual screening using hydroxychloroquine as a template and comparing structural similarities against nearly 4,000 approved drugs identified amodiaquine, zuclopenthixol, and nebivolol as effective in vitro agents in Vero E6 cells infected with SARS-CoV-2 [75].

Collectively, these efforts underscore the value of computational and ML methodologies in accelerating drug repurposing and combination therapy design, providing viable alternatives for managing resistant infections and emerging pathogens.

AI-driven drug repurposing in cardiology and extrapolations

Antidiabetic agents have gained traction for their ability to lower cardiovascular mortality and MACE, underscoring a growing interest in leveraging metabolic therapies for cardiovascular benefit [76]. Beyond metabolic modulation, inflammation is increasingly recognized as a key driver of CVD, prompting investigations into the repurposing of anti-inflammatory drugs as cardioprotective agents. For instance, canakinumab, a monoclonal antibody targeting interleukin-1β initially developed for treating autoinflammatory conditions, was tested in patients with a history of MI and elevated C-reactive protein levels [77]. After 48 months of treatment, reductions in inflammatory biomarkers and improved cardiovascular outcomes were observed, demonstrating the promise of an anti-inflammatory approach. Similar findings were reported for colchicine, a drug traditionally used for gout attacks. In individuals with chronic coronary disease, treatment with colchicine led to a significant decrease in cardiovascular events, and when administered shortly after MI, reduced the risk of subsequent ischemic events [78, 79].

Collectively, these studies underscore the potential of repurposed anti-inflammatory agents for addressing the inflammatory underpinnings of CVD. Although AI approaches have demonstrated efficacy in other therapeutic areas, including neurology and oncology, their application in cardiovascular drug repurposing is only beginning to emerge. By integrating ML algorithms with large-scale clinical, genomic, and proteomic data, AI-driven frameworks could efficiently uncover novel drug-disease associations, predict therapeutic responses, and refine dosing strategies. Such capabilities promise to streamline the drug development pipeline, reducing both time and cost. As research continues to elucidate the intricate interplay between metabolic and inflammatory pathways in CVD, leveraging AI-enhanced bioinformatics for drug repurposing may catalyze the translation of preclinical insights into safe, effective cardiovascular therapies.