Namshik Han

Co-expression networks are used to find prognostic lncRNA programmes, but lncRNA quantification depends on unprocessed RNA in the library. In renal cancer (1,062 patients), the leading lncRNA axis tracked the non-exonic (non-feature) read fraction, a routine alignment metric, at Spearman rho 0.90. Without adjustment, most of the lncRNA network merged into one module. Adjustment preserved agreement with single-cell tumour-normal differences. Under rules fixed with the renal results known, 9 of 25 TCGA cancer types had a lncRNA-specific axis and 4 had network collapse (8 and 2 of 23 non-renal types). Intronic lncRNAs tracked the metric more than intergenic ones in all 25, and the metric was not consistently prognostic. All five renal modules that survived adjustment kept their direction in an external cohort, two nominally significant and none at a 5% false discovery rate. Three localised to malignant or epithelial cells. Together they added no discrimination over TNM and grade.

Partial epithelial-to-mesenchymal transition (pEMT) was defined in single malignant cells of head and neck squamous cell carcinoma but is usually scored in bulk tumours. In three single-cell datasets, its laminin-332 and hemidesmosome genes were expressed by malignant cells (the malignant arm) and its fibrillar collagens and vimentin by fibroblasts (the stromal arm). In bulk tumours the stromal arm dominated co-expression, so standard pEMT scores mainly measure stroma. We recovered the malignant arm with a classifier trained on single cells and a 12-gene core. In 40 patients treated with platinum and cetuximab, both malignant measures were higher in those with long progression-free survival (area under the curve 0.78 and 0.82), and the stromal arm was not. The malignant arm was inseparable from the Basal subtype, whose score did as well or better in these patients (0.93) and in two xenograft panels. Bulk pEMT scores should be read by cell of origin.

Brown adipose tissue (BAT) increases energy expenditure but is scarce in adults, yet most strategies focus on activating existing thermogenic adipocytes rather than recruiting new ones. We integrate human single-nucleus transcriptomics with protein-interaction and knowledge-graph analyses to identify stage-specific, druggable programmes governing brown adipogenesis. This pipeline resolved four ordered transcriptional states and three transitions, revealing an amyloid-β/APP regulatory layer and nominating APP as a central signalling hub. Most prioritised compounds altered differentiation or metabolic phenotype, several substantially. Phenserine (APP-lowering) and sacubitril (neprilysin inhibitor) promoted differentiation through distinct mechanisms and stages and could be combined sequentially for benefit. Phenserine increased differentiation across three human models and, in mice, expanded BAT mass and increased thermogenic capacity. Together, these findings reveal targetable control points in human brown adipocyte recruitment, nominate repurposable drugs for BAT expansion, and establish a generalisable framework coupling single-cell transcriptomics with network-based drug repurposing across staged biological systems.

The growing landscape of Alzheimer’s disease (AD) datasets creates opportunities to integrate heterogeneous evidence and systematically discover disease effectors. We present BRIDGE-AD, an interpretable network medicine framework that transforms multimodal data into a unified, disease-specific gene representation for AD effector prioritisation. We integrated more than 30 datasets and curated resources spanning omics, functional, genetic and prior disease knowledge layers. BRIDGE-AD outperformed recently published pretrained and modality-specific gene embeddings in recovering AD-associated genes and produced a genome-wide resource of candidate AD effectors. Established and newly prioritised effectors formed 19 functional clusters, revealing a global molecular landscape of AD biology. BRIDGE-AD supported an SPP1-centred cross-compartment hypothesis and nominated SCARB2, a poorly characterised candidate, for functional validation. SCARB2 rewired lysosomal, lipid-handling and autophagic programmes in microglia, whereas disrupted SCARB2 glycosylation in AD implicated altered SCARB2 processing and function. The accompanying website, https://explore-bridgead.com , enables users to trace the curated evidence and generate mechanistic hypotheses.

Background: Early diagnosis of paediatric lower respiratory tract infection (LRTI) in intensive care is difficult because clinical features overlap with non-infectious respiratory failure and conventional microbiology has limited sensitivity. We aimed to integrate airway microbial and host-response signals to identify early diagnostic biomarkers in mechanically ventilated children. Methods: We re-analysed airway metagenomic sequencing and tracheal aspirate transcriptomics from 261 ventilated paediatric intensive care unit (PICU) patients using differential expression, protein–protein interaction network analysis, and machine-learning feature selection. Findings were evaluated in an exploratory assessment of biologically related cytokine markers within an independent prospective cohort (RASCALS; n = 100 enrolled, n = 78 analysed) using non-bronchoscopic mini-bronchoalveolar lavage (mini-BAL), plasma cytokines, and clinician-adjudicated diagnoses. Results: Respiratory syncytial virus and Haemophilus influenzae were the only pathogens consistently enriched in LRTI. Host transcriptomics showed activation of interferon, cytokine, and chemokine signalling. Network-guided feature selection identified a seven-gene panel (IRF7, FFAR3, GZMB, FABP4, FN1, CXCL5, BCAR1) with high diagnostic performance in cross-validation (median F1 0.94), comparable to a published 14-gene model. In RASCALS, a logistic model using mini-BAL IL-1β, IL-4, and IL-8 classified bacterial LRTI with 65% accuracy (sensitivity 70%, specificity 63%), while blood IL-6 and TRAIL achieved 82% accuracy (sensitivity 79%, specificity 83%). Conclusions: Integrating airway microbial profiles with host-response biomarkers supports earlier and more specific LRTI diagnostic classification in ventilated children. The seven-gene panel and cytokine combinations are candidates for rapid PCR- or immunoassay-based bedside tests to inform antimicrobial stewardship.

Inflammatory bowel diseases (IBD) remain a therapeutic challenge due to their heterogeneous nature and the absence of clinically actionable biomarkers to guide precision treatment. While multi-omics studies have advanced our understanding of the disease, meaningful translation into practice has been hindered by lack of validation and a need to move from association to a more generalizable signal. Epigenetics, particularly DNA methylation, offers a promising lens to capture disease-relevant regulatory states that reflect both genetic predisposition and environmental influences. However, the path to clinical adoption has been limited by persistent computational hurdles. Here we synthesize the current landscape of IBD biomarker discovery and highlight the conceptual advantages of epigenetic signatures. We outline the main obstacles that have limited clinical translation to date, including data heterogeneity, batch effects, and the challenge of distinguishing functional “driver” from non-functional “passenger” epigenetic changes. We then discuss how recent advances in computational methodology—spanning data harmonization, integrative modeling, and interpretable machine learning—can help bridge the gap between complex datasets and reliable, deployable biomarkers. Finally, we propose a forward-looking roadmap for study design and validation aimed at moving the field toward routine clinical implementation, thereby realizing the full potential of epigenetics in IBD.

The high cost and attrition rate of drug development underscore the need for more effective strategies for therapeutic target discovery. Here, we present a network medicine-based machine learning framework that integrates single-cell transcriptomics, bulk multi-omic profiles, genome-wide CRISPR perturbation screens, and protein-protein interaction networks to systematically prioritise disease-specific targets. Applied to clear cell renal cell carcinoma, the framework successfully recovered established targets and predicted five therapeutic candidates, with subsequent in vitro validation demonstrating that among these, ENO2 inhibition had the strongest anti-tumour effect, followed by LRRK2 , a repurposing candidate with phase III Parkinson’s disease inhibitors. The proposed approach advances target discovery by moving beyond single-feature, single-modality heuristics to a scalable, machine learning-driven strategy that is generalisable across diseases.

Single-cell foundation models (scFMs) have shown promise as transferable representations of cellular state, but recent zero-shot evaluations suggest that they do not consistently outperform simpler baselines. We asked whether this apparent limitation reflects an intrinsic weakness of scFMs or instead the difficulty of using them without task-specific adaptation. To test this, we fine-tuned two widely used scFMs, Geneformer and scGPT, on common tumour subtypes from renal, lung, and breast cancer, and compared them with a LightGBM baseline on within-domain validation cohorts and on out-of-domain rarer, unseen cancer subtypes. Across all three organs, the models achieved near-perfect within-domain discrimination (AUROC 0.98-1.00), but differences emerged under subtype shift. On chromophobe RCC, scGPT and Geneformer achieved AUROC 0.88 and 0.92 respectively versus 0.64 for LightGBM; on SCLC, Geneformer reached 1.00 versus 0.82 for LightGBM; and on TNBC, scGPT achieved 0.80 versus 0.49 for LightGBM. To determine whether this generalisation reflected meaningful adaptation rather than arbitrary feature drift, we applied Integrated Gradients, an interpretability technique, to the fine-tuned scFMs and SHAP to LightGBM. LightGBM showed highly stable gene-importance rankings across datasets, whereas the foundation models were substantially more context-sensitive. However, this flexibility was not random: all models converged on a shared within-domain core, while scFMs acquired larger rare-subtype-specific gene sets and pathway programmes during transfer. Pathway enrichment further supported the biological relevance of these attributed genes. Together, these results suggest that fine-tuned scFMs can bridge clinically relevant domain shifts in cancer single-cell analysis and that interpretability provides a practical route to distinguishing biologically grounded adaptation from rigid reuse of training-era rules.

Developing an efficient and safe therapy necessitates a mechanistic understanding of the complex underlying pathology and manipulation of the multiple path