Knowledge Graph Framework Uncovers Drug Repurposing Targets for Aplastic Anaemia

A hybrid computational framework combining literature-based discovery with curated database validation has mapped multi-level molecular interactions in aplastic anaemia, identifying critical disease-associated genes, key signaling cascades, and several drug repurposing candidates. Aplastic anaemia is a serious haematological disorder marked by the immune-mediated destruction of haematopoietic stem cells and progressive bone marrow failure, resulting in pancytopenia across peripheral blood lineages. While clinical immunosuppressive regimens are routinely deployed to manage the condition, their general lack of target specificity highlights a continuous demand for systematic methodologies capable of pinpointing precise biological targets and evaluating candidate compounds for repositioning. To address this challenge, investigators constructed a multi-layered biomedical knowledge graph that integrates extracted semantic predications from published literature with pathway-level annotations and structured database resources, providing an exploratory, hypothesis-generating platform designed to illuminate the complex biological mechanisms linking pharmacological interventions with molecular dysregulation in the bone marrow microenvironment.

The pathogenesis of aplastic anaemia involves intricate immunological and genetic underpinnings that complicate clinical management and therapy development. The disorder can arise from diverse aetiologies, encompassing environmental exposures to industrial chemicals, insecticides, pesticides, and pharmaceuticals, alongside genetic mutations, inherited predispositions, autoimmune reactions, and occasional pregnancy-related presentations. Earlier genomic investigations highlighted approximately two hundred differentially expressed genes, while mutational profiling has documented alterations in regulatory genes such as TP53, DNMT3A, and ASXL1 that contribute directly to haematopoietic stem cell dysfunction. Simultaneously, immune regulatory factors such as IL2RA and CTLA4 exhibit altered expression patterns that undermine self-tolerance and normal immune control. Standard immunosuppressive therapies, notably calcineurin inhibitors such as cyclosporine and T-cell depleting antibodies like antithymocyte globulin or alemtuzumab, have formed the traditional backbone of clinical intervention, yet the molecular mechanisms governing their broad activity and the potential for more selective pharmacological perturbagens require deeper, network-wide characterization across complementary biomedical datasets.

To assemble the literature-derived foundation for their knowledge network, the investigators retrieved structured semantic predications from the Semantic MEDLINE Database, known as SemMedDB, which processes PubMed-indexed publications using the SemRep natural language processing system. Utilizing version 43 of SemMedDB under permission from the National Institutes of Health, the researchers analyzed relational predicate tables using statistical computing packages in R, including dplyr, data.table, stringr, tidytext, and tidyr. Semantic predications represent biological statements as standardized subject-predicate-object triplets grounded in the Unified Medical Language System Metathesaurus and Semantic Network. Extraction queries were organized across two targeted exploratory strategies: a disease-specific approach capturing aplastic anaemia alongside related anatomical and clinical descriptors such as bone marrow, bone marrow cells, and bone marrow stem cells, and a drug-specific approach compiling established therapeutics. Predications were categorized into core semantic layers representing genes, drugs, and diseases, alongside extended contextual layers encompassing cellular, physiological, and pathological attributes to preserve biological granularity.

Building upon the foundational literature-based discovery ABC paradigm, the researchers established a tripartite network linking disease-associated entities to intermediate molecular concepts in order to infer potential therapeutic associations. Drug-gene interactions and gene-disease associations were extracted independently and harmonized into a consolidated edge list, where duplicate assertions were removed based on exact subject-predicate-object matching and edge weights were assigned according to the reported frequency of identical predications. Bidirectional semantic scrutiny was applied during extraction to capture both subject-to-object and object-to-subject relationships, thereby reducing directional bias and maximizing overall edge coverage. Topological analysis was conducted across both a core network restricted to primary biological classes and an extended network incorporating contextual semantic categories. Network centrality properties were quantified using degree, in-degree, out-degree, closeness, and betweenness metrics, with degree and betweenness serving as primary indicators to reflect local network connectivity and intermediary positioning across distinct biological modules.

To expand the network beyond discrete pairwise predications, the investigators integrated functional pathway annotations derived through enrichment analysis using Bioconductor packages, including clusterProfiler, pathview, enrichplot, ReactomePA, and biomart. Entrez identifiers for extracted gene sets were mapped to Gene Ontology terms, Kyoto Encyclopedia of Genes and Genomes pathways, and Reactome pathways, defining statistical significance at a Benjamini-Hochberg-adjusted probability threshold below 0.05. The resulting pathway enrichment profile highlighted prominent immunological and haematopoietic systems, including haematopoietic cell lineage, JAK-STAT signaling, cytokine-cytokine receptor interaction, T-cell receptor signaling, cell adhesion molecules, and antigen processing and presentation. Furthermore, the analysis revealed enrichment in alternative regulatory cascades, such as regulation of MAPK/ERK cascades, monocyte and macrophage differentiation, and negative regulation of telomere maintenance, illustrating how cytokine imbalance and intracellular signaling perturbations converge to suppress haematopoietic lineage progression and alter the supportive bone marrow stroma.

The tripartite network was subsequently expanded into a comprehensive heterogeneous knowledge graph incorporating four principal node categories: diseases, genes, drugs, and functional pathways. To prioritize potential drug repurposing candidates in an unbiased, quantitative manner, the team formulated a composite Drug Repurposing Score combining standardized metric inputs. Individual variables, including literature-derived predication frequency, total network degree, and betweenness centrality, were standardized using z-score normalization to account for differences across measurement scales. The scoring model assigned weights of thirty percent to literature publication frequency, thirty percent to total degree centrality, and forty percent to betweenness centrality, placing intentional emphasis on betweenness to prioritize compounds that bridge disparate topological regions rather than those simply reflecting dense prior publication history. Robustness testing across four alternative weighting configurations confirmed consistent ranking performance with a Spearman correlation coefficient of 1.00, establishing a stable computational hierarchy among candidate molecules.

Analysis of the integrated knowledge graph illuminated several key genes operating at the convergence of immune dysregulation and haematopoietic failure. Among the primary nodes, colony stimulating factor 2 (CSF2) emerged as an influential cytokine governing the proliferation, differentiation, and functional activation of granulocytes and macrophages, wherein dysregulated expression can prompt neutropenia or aberrant myeloid proliferation. Additional central genetic components included erythropoietin (EPO), essential for erythroid maturation, interleukin-6 (IL6), a pro-inflammatory cytokine driving Th17 cell differentiation and bone marrow suppression, and the major histocompatibility complex class II allele HLA-DRB1, which mediates antigen presentation to helper T cells. Cross-referencing extracted gene-disease assertions against the DisGeNET curated database confirmed five overlapping genes, reinforcing biological plausibility. The network also highlighted vascular endothelial growth factor A (VEGFA), thyroid peroxidase (TPO), cell surface markers CD34 and CD4, and non-coding RNA regulator MIR335, alongside less-characterized factors like LAMC2, RBM45, RAP1A, TBPL1, and chromosomal fragile site regions.

Topological evaluation of pharmacological agents positioned the calcineurin inhibitor cyclosporine as a densely interconnected central hub across the knowledge graph, confirming its foundational role as an immunosuppressive agent while demonstrating extensive computational connections to secondary drugs, signaling cascades, and molecular targets such as cyclophilin and P-glycoprotein. Although cyclosporine is widely used clinically in aplastic anaemia, curated entries in DrugBank and PubChem annotate its primary formal indications around solid organ transplant rejection and autoimmune disorders, classifying its aplastic anaemia application as an established repositioning success. Further interrogation using predicates such as COMPARED_WITH, COEXISTS_WITH, and STIMULATES identified overlapping immunosuppressive agents and candidate perturbagens, including tacrolimus, methotrexate, mycophenolate mofetil, mizoribine, everolimus, prednisone, and camostat mesylate. External validation tools including DGIdb for drug-gene interactions and DrugBank for pharmacological targets supported the biological plausibility of several immunosuppressive compounds that modulate purine synthesis, mTOR signaling, and inflammatory protease activation.

Beyond conventional immunosuppressive agents, the computational framework highlighted experimental perturbagens and pathway-level modulators that provide mechanistic insight rather than immediate clinical options. Tetradecanoyl phorbol acetate, also known as phorbol 12-myristate 13-acetate, surfaced as a literature-derived candidate due to its potent activation of protein kinase C, a serine and threonine kinase involved in haematopoietic stem cell proliferation, cell survival, and immune regulation. However, because tetradecanoyl phorbol acetate possesses well-documented tumor-promoting properties, the authors explicitly noted that it cannot be considered a clinically suitable drug candidate, serving instead as a preclinical probe alongside functionally related analogues such as prostratin, bryostatin-1, or ingenol mebutate. Similarly, the ionophore ionomycin was classified as an experimental probe targeting intracellular calcium mobilization and subsequent calcineurin-NFAT signaling. In contrast, camostat mesylate, an established serine protease inhibitor approved for chronic pancreatitis, was highlighted as a hypothesis-generating repurposing prospect warranting investigation based on its protease-modulating properties in inflammatory settings.

The authors underscored several methodological limitations, reiterating that network-derived centrality metrics and composite prioritization scores serve strictly as computational hypothesis-generating tools rather than direct measures of therapeutic efficacy or clinical safety. The initial extraction relied on predefined domain keywords and SemMedDB semantic types without automated ontology-driven hierarchical expansion, which may have constrained the initial breadth of entity capture. Furthermore, external cross-referencing against databases like DisGeNET, DGIdb, and DrugBank provided supportive rather than confirmatory biological evidence, as literature-mined predications inevitably inherit reporting biases from original biomedical publications. Consequently, the researchers emphasized that the candidate perturbagens and novel genetic associations identified through the heterogeneous knowledge graph—including non-canonical factors like LAMC2, RBM45, and genomic fragile sites—remain exploratory in nature. Future investigations incorporating high-throughput transcriptomic profiling, disease-specific animal models, and rigorous preclinical pharmacokinetic evaluations will be required to validate these computational insights before translational applications can be realized.

Subject of Research: Computational knowledge graph integration and literature-based drug repurposing for aplastic anaemia

Article Title: A knowledge graph–driven literature-based discovery framework for drug repurposing in aplastic anaemia

Article References: Sankar, D., Bhasuran, B., Oviya, I. R., & Saranya, G. (2026). A knowledge graph–driven literature-based discovery framework for drug repurposing in aplastic anaemia. Frontiers in Bioinformatics, 6, Article 1905251. https://doi.org/10.3389/fbinf.2026.1905251

Image Credits: AI Generated

DOI: 10.3389/fbinf.2026.1905251

Keywords: Aplastic Anaemia, Knowledge Graph, Drug Repurposing, SemMedDB, Bioinformatics, Immunology, Haematology, Literature-Based Discovery, knowledge, graph, driven, literature-based