Serdar Bozdag

Motivation: Integrating multi-omics data for disease prediction remains challenging due to the complexity of cross-layer biological interactions and the limited use of prior biological knowledge in many machine learning models. We present the Hierarchical Input Neural Network (HINN), a deep learning framework that incorporates biologically informed connections across genomic, epigenetic, and transcriptomic layers to model regulatory relationships in a structured and interpretable manner. HINN was applied to blood-derived multi-omics data to predict cognitive assessment scores, including Mini-Mental State Examination (MMSE), Montreal Cognitive Assessment, Alzheimer’s Disease Assessment Scale–Cognitive Subscale 11 tasks (ADAS11), and Rey Auditory Verbal Learning Test. A biologically guided feature construction pipeline was used to integrate genome-wide association studies-derived single nucleotide polymorphisms (SNPs), promoter-region DNA methylation, and gene expression data through Gene Ontology Biological Process and KEGG pathway mappings. Results: The model was evaluated using a rigorous training strategy that separates feature selection from testing and incorporates cross-validation as well as repeated runs to ensure robustness. Across all cognitive outcomes, HINN achieved the lowest mean normalized mean squared error among the evaluated models, with statistically significant improvements over most baseline methods. Ablation analyses showed that the contributions of individual omics modalities and biologically informed connectivity varied across cognitive outcomes, with the clearest benefits of structured connectivity observed for ADAS11 and MMSE. Feature attribution analysis identified a multi-omics cascade involving SNP (rs116557230), CpG site (cg24041822), and a SOCS4 gene expression probe, which showed consistent associations with multiple cognitive outcomes. These findings highlight the ability of HINN to uncover biologically meaningful cross-omics relationships relevant to cognitive decline. Together, this work demonstrates that integrating deep learning with biological knowledge enables biomarker discovery for complex diseases. Availability and implementation: All code, sample data, and instructions are publicly available via GitHub (https://github.com/bozdaglab/HINN) and archived on Zenodo (https://doi.org/10.5281/zenodo.19197057) to ensure reproducibility and long-term accessibility.

Identifying new therapeutic uses for existing drugs is a major challenge in biomedicine, especially for complex neurodegenerative conditions such as Alzheimer disease and related dementias (ADRD), where treatment options remain limited and relevant data are often sparse, heterogeneous, and difficult to integrate. Although general-purpose Large Language Model (LLM) embeddings encode rich semantic information, they often lack the task-specific biomedical context needed for inference tasks such as computational drug repurposing. We introduce Contextualizing LLM Embeddings via Attention-based gRaph learning (CLEAR), a multimodal representation-fusion framework that aligns LLM embeddings with the topological structure of a context-specific Knowledge Graph (KG). Across five benchmark datasets, CLEAR achieved state-of-the-art results, improving predictive performance (e.g., F1 score) by up to 30% over prior methods. We further applied CLEAR to identify FDA-approved drugs with potential for repurposing for ADRD, including Parkinson disease-related dementia and Lewy Body dementia. CLEAR learned a biologically coherent embedding space, prioritized leading ADRD drug candidates, and accurately summarized known therapeutic relationships for FDA-approved Alzheimer disease drugs. Overall, CLEAR shows that grounding biomedical LLM embeddings with context-specific KG signals can improve drug repurposing in data-sparse, real-world settings. GitHub: https://github.com/bozdaglab/CLEAR

Multi-omics datasets capture complementary aspects of biological systems and are central to modern machine learning applications in biology and medicine. Existing graph-based integration methods typically construct separate graphs for each omics type and focus primarily on intra-omic relationships. As a result, they often overlook cross-omics regulatory signals—bidirectional interactions across omics layers—that are critical for modeling complex cellular processes. A second major challenge is missing or incomplete omics data; many current approaches degrade substantially in performance or exclude patients lacking one or more omics modalities. To address these limitations, we introduce MultiGEOmics , an intermediate-level graph integration framework that explicitly incorporates regulatory signals across omics types during graph representation learning and models biologically inspired omics-specific and cross-omics dependencies. MultiGEOmics learns robust cross-omics embeddings that remain reliable even when some modalities are partially missing. We evaluated MultiGEOmics across eleven datasets spanning cancer and Alzheimer’s disease, under zero, moderate, and high missing-rate scenarios. MultiGEOmics consistently maintains strong predictive performance across all missing-data conditions while offering interpretability by identifying the most influential omics types and features for each prediction task. The source code and the documentation of MultiGEOmics are available at https://github.com/bozdaglab/MultiGEOmics .

Motivation: Electronic Health Records (EHRs) contain vast amounts of longitudinal patient medical history data, making them highly informative for early disease prediction. Numerous computational methods have been developed to leverage EHR data; however, many process multiple patient records simultaneously, resulting in high memory consumption and computational cost. Moreover, these models also often lack interpretability, limiting insight into the factors driving their predictions. Efficiently handling large-scale EHR data while maintaining predictive accuracy and interpretability therefore remains a critical challenge. To address this gap, we propose DyGraphTrans, a dynamic graph representation learning framework that represents patient EHR data as a sequence of temporal graphs. In this representation, nodes correspond to patients, node features encode temporal clinical attributes, and edges capture patient similarity. DyGraphTrans models both local temporal dependencies and long-range global trends, while a sliding-window mechanism reduces memory consumption without sacrificing essential temporal context. Unlike existing dynamic graph models, DyGraphTrans jointly captures patient similarity and temporal evolution in a memory-efficient and interpretable manner. Results: We evaluated DyGraphTrans on Alzheimer’s Disease Neuroimaging Initiative (ADNI) and National Alzheimer’s Coordinating Center (NACC) for disease progression prediction, as well as on the Medical Information Mart for Intensive Care (MIMIC-IV) dataset for early mortality prediction. We further assessed the model on multiple benchmark dynamic graph datasets to evaluate its generalizability. DyGraphTrans achieved strong predictive performance across diverse datasets. We also demonstrated interpretability of DyGraphTrans aligned with known clinical risk factors. The source code and datasets are available at https://github.com/bozdaglab/DyGraphTrans . Contact: Serdar.Bozdag@unt.edu Supplementary: Uploaded as an attachment.

Single-cell RNA sequencing (scRNA-seq) allows transcriptomic profiling at single-cell resolution, providing valuable insights into cellular diversity across tissues, developmental stages, and diseases. However, accurately identifying cell types remains challenging due to the high dimensionality, sparsity, and noise inherent in scRNA-seq data. To address these challenges in cell type identification in scRNA-seq data, we introduce scAURA (single c ell A lignment- and U niformity-based Graph Debiased Contrastive R epresentation A rchitecture), a unified framework that integrates graph debiased contrastive learning with self-supervised clustering. We evaluated scAURA on 18 real single-cell datasets collected from six sequencing platforms spanning diverse tissue and cell types in human and mouse. scAURA outperformed all state-of-the-art (SOTA) methods in nine and eight datasets in Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI), respectively. On average, scAURA obtained average ranks of 2.28 (ARI) and 2.39 (NMI) across all 13 SOTA methods, demonstrating its consistent superiority across datasets. scAURA also exhibited strong robustness to dropout noise by maintaining stable clustering performance even under increasing sparsity levels. Furthermore, in an external single-cell Alzheimer’s disease dataset, scAURA accurately clustered different cell types, identified novel cell type–specific marker genes, and inferred their potential transcriptional regulators. The source code and datasets are available at https://github.com/bozdaglab/scAURA . Contact: Serdar.Bozdag@unt.edu

Motivation: Early prediction of Alzheimer’s disease (AD) progression from mild cognitive impairment (MCI) remains a major challenge, particularly when relying on non-invasive biomarkers. Identifying individuals with progressive MCI (pMCI) before conversion to AD could improve intervention strategies and clinical management. In this study, we developed a machine learning (ML) framework integrating blood-based multi-omics and demographic data to distinguish pMCI from stable MCI (sMCI). Results: We trained ML models using combinations of single nucleotide polymorphism (SNP), DNA methylation, gene expression, lipid, and bile acid metabolite data using both early and late data integration strategies. Late integration consistently outperformed early integration, with the L1-regularized logistic regression model achieving the highest F1 score of 90.7% when combining SNP and lipid data. Feature interpretability analyses using LIME and SHAP identified reproducible biomarkers across omic layers, including SNPs in MYH11, FOXP1, MAPK10, and SYN3; methylation changes in CDX2 and DHX58; and altered ceramide lipid CER.D19.1.24.0 levels, all previously associated with AD-related pathology. These findings demonstrate that combining multi-omics and demographic data can improve early prediction of AD progression and support the feasibility of blood-based, interpretable biomarkers for precision diagnostics. Availability and implementation: The data analyzed in this study were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (https://adni.loni.usc.edu/) and are available to qualified researchers upon application. The code used in this study, together with instructions and a toy dataset, is available at: https://github.com/bozdaglab/AD_sMCI_vs_pMCI.

Alzheimer’s disease (AD) is a major global health concern, expected to affect 12.7 million Americans by 2050. Machine learning (ML) algorithms have been developed for AD diagnosis and progression prediction, but the lack of racial/ethnic diversity in clinical datasets raises concerns about their generalizability across demographic groups, particularly underrepresented populations. Studies show ML algorithms inherit biases from data, leading to biased AD predictions. This study investigates the fairness of ML models in AD diagnosis. We hypothesize that models trained on a single racial/ethnic group perform well within that group but poorly in others. We employ feature selection and model training techniques to improve fairness. Our findings support our hypothesis that ML models trained on one group underperform on others. We also demonstrated that applying fairness techniques to ML models reduces their bias. This study highlights the need for racial/ethnic diversity in datasets and fair models for AD prediction.

Understanding complex diseases requires models that can integrate diverse layers of biological data while yielding insights that are biologically interpretable. Although multi-omics integration with machine learning (ML) has advanced disease prediction and biomarker discovery, most existing approaches overlook the hierarchical and regulatory relationships that connect these molecular layers. Here, we present the Hierarchical Input Neural Network (HINN), a deep learning framework that incorporates known cross-omics relationships directly into its architecture, capturing the flow of information from genomics to epigenomics, transcriptomics, and downstream biological processes. By embedding these relationships, HINN improves both predictive performance and biological interpretability. We applied HINN to blood-derived multi-omics data from individuals with Alzheimer’s disease or mild cognitive impairment to predict cognitive scores from standardized assessments. HINN outperformed both baseline and state-of-the-art models and pinpointed multi-omics biomarkers—including SNPs and promoter-region CpG sites in ATP6V1C1 and RCHY1 that were significantly correlated with plasma p-Tau181 levels. These features map to biologically relevant processes with potential implications for cognitive decline. Our findings demonstrate how combining deep learning with biological knowledge can uncover interpretable, blood-based biomarkers for cognitive decline due to complex diseases such as Alzheimer’s. All code and data are openly available at https://github.com/bozdaglab/HINN.

Motivation: The accurate prediction of drug–target interactions (DTI) is a crucial step in drug discovery, providing a foundation for identifying novel therapeutics. Traditional drug development is both costly and time-consuming, often spanning over a decade. Computational approaches help narrow the pool of compound candidates, offering significant starting points for experimental validation. In this study, we propose a Top-DTI framework for predicting DTI by integrating topological data analysis (TDA) with large language models (LLMs). Top-DTI leverages persistent homology to extract topological features from protein contact maps and drug molecular images. Simultaneously, protein and drug LLMs generate semantically rich embeddings that capture sequential and contextual information from protein sequences and drug SMILES strings. By combining these complementary features, Top-DTI enhances predictive performance and robustness. Results: Experimental results on the public BioSNAP and Human DTI benchmark datasets demonstrate that the proposed Top-DTI model outperforms state-of-the-art approaches across multiple evaluation metrics, including AUROC, AUPRC, sensitivity, and specificity. Furthermore, the Top-DTI model achieves superior performance in the challenging cold-split scenario, where the test and validation sets contain drugs or targets absent from the training set. This setting simulates real-world scenarios and highlights the robustness of the model. Notably, incorporating topological features alongside LLM embeddings significantly improves predictive performance, underscoring the value of integrating structural and sequence-based representations. Availability and implementation: The data and source code of Top-DTI are available at https://github.com/bozdaglab/Top_DTI under the Creative Commons Attribution NonCommercial 4.0 International Public License.

The widespread availability of multi-omics tumor profiling has enabled detailed molecular characterization of individual tumors, paving the way for more effective, less toxic, and patient-specific therapies. However, widespread compound screens in human patients are constrained by ethical and logistical challenges, underscoring the need for computational models capable of predicting in vivo drug response. Here, we introduce PDDRNet-MH , a multiplex heterogeneous network-based framework that integrates genomic, transcriptomic, and epigenomic tumor profiles with drug chemical structures, pharmacological activity, and side-effect data to infer personalized drug responses. PDDRNet-MH constructs an integrated network connecting patients, cell lines, drugs, and genes, with each component represented by four biologically and pharmacologically informed similarity layers. This design enables the systematic propagation of drug-biomarker associations across modalities. We applied PDDRNet-MH to breast cancer patient data from The Cancer Genome Atlas and benchmarked its predictive performance against state-of-the-art methods on eleven FDA-approved breast cancer drugs. PDDRNet-MH achieved consistently high accuracy, with perfect prediction scores for gemcitabine and vinorelbine (Area Under the Receiver Operating Characteristic Curve [AUC-ROC] = 1.00; Area Under the Precision-Recall Curve [AUC-PR] = 1.00) and near-perfect scores for methotrexate and zoledronate (AUC-ROC = 0.95; AUC-PR = 0.99), demonstrating its ability to robustly distinguish sensitive from resistant patients. Biologically, PDDRNet-MH accurately prioritized established clinical biomarkers, including HER2 (ERBB2) for lapatinib and BRCA1/2 for doxorubicin and cyclophosphamide. Beyond known associations, the model identified additional genes within the HER2 amplicon on chromosome 17q12, including STARD3, MIEN1, and PPP1R1B, whose amplification was significantly associated with elevated drug response scores, suggesting potential roles in HER2-targeted therapy. These findings highlight the ability of PDDRNet-MH to recover and extend clinically relevant drug-biomarker associations, supporting its utility in guiding precision oncology.