Umeå University's logo

Deep learning (DL) models require extensive data to achieve strong performance and generalization. Deep generative models (DGMs) offer a solution by synthesizing data. Yet current approaches for tabular data often fail to preserve feature correlations and distributions during training, struggle with multi-metric hyperparameter selection, and lack comprehensive evaluation protocols. We address this gap with a unified framework that integrates training, hyperparameter tuning, and evaluation. First, we introduce a novel correlation- and distribution-aware loss function that regularizes DGMs, enhancing their ability to generate synthetic tabular data that faithfully represents the underlying data distributions. Theoretical analysis establishes stability and consistency guarantees. To enable principled hyper-parameter search via Bayesian optimization (BO), we also propose a new multi-objective aggregation strategy based on iterative objective refinement Bayesian optimization (IORBO), along with a comprehensive statistical testing framework. We validate the proposed approach using a benchmarking framework with twenty real-world datasets and ten established tabular DGM baselines. The correlation-aware loss function significantly improves the synthetic data fidelity and downstream machine learning (ML) performance, while IORBO consistently outperforms standard Bayesian optimization (SBO) in hyper-parameter selection. The unified framework advances tabular generative modeling beyond isolated method improvements. Code is available at: https://github.com/vuhoangminh/TabGen-Framework.

Multiple myeloma is preceded by monoclonal gammopathy of undetermined significance (MGUS). Only a minority of patients with MGUS will develop multiple myeloma, but precise prediction of progression is impossible using routine clinical biomarkers. Changes in the levels of blood immune markers can help predict disease progression. Data remain inconsistent for some markers of interest such as monocyte chemotactic protein-3 (MCP-3), macrophage inflammatory protein-1 alpha (MIP-1α), fibroblast growth factor-2 (FGF-2), vascular endothelial growth factor (VEGF), fractalkine, and transforming growth factor-alpha (TGF-α). We aimed to investigate the associations between the prediagnostic serum levels of these candidate biomarkers and future multiple myeloma risk, as well as to assess marker changes over time. We performed a nested case-control study using prospective samples from the Janus Serum Bank in Norway to investigate associations between multiple myeloma risk and prediagnostic serum levels of MCP-3, MIP-1α, FGF-2, VEGF, fractalkine, and TGF-α. The study included 293 future multiple myeloma cases with serum samples collected 20 years (median) before multiple myeloma diagnosis and 293 matched cancer-free controls. Patients with multiple myeloma had an additional prediagnostic sample collected up to 42 years before diagnosis to identify marker changes over time. Markers with >60% detection rate (MIP-1α, VEGF, and TGF-α) were included in the statistical analysis. We observed no statistically significant associations between multiple myeloma risk and serum levels of MIP-1α, VEGF, or TGF-α in samples collected 20 years before diagnosis. However, TGF-α levels decreased significantly closer to the diagnosis in patients with multiple myeloma (P < 0.001). The decrease in TGF-α levels may reflect subtle microenvironmental changes related to multiple myeloma progression.

PREVENTION RELEVANCE: This study observed a decline in TGF-α serum levels closer to multiple myeloma diagnosis, which may aid in predicting multiple myeloma progression and early detection, although validation in other longitudinal cohorts is needed.

Genome-wide association studies (GWAS) have contributed to our understanding of glioma susceptibility. To date, 25 risk loci for development of any of the glioma subtypes are known. However, GWAS studies reveal little about the molecular processes that lead to increased risk, especially for non-coding single nucleotide polymorphisms (SNP). A particular SNP in intron 2 of LRIG1, rs11706832, has been shown to increase the susceptibility for IDH1 mutated low-grade gliomas (LGG). Leucine-rich repeats and immunoglobulin-like domains protein 1 (LRIG1) is important in cancer development as it negatively regulates the epidermal growth factor receptor (EGFR); however, the mechanism responsible for this particular risk SNP and its potential effect on LRIG1 are not known. Using CRISPR-CAS9, we edited rs11706832 in HEK293T cells. Four HEK293T clones with the risk allele were compared to four clones with the non-risk allele for LRIG1 and SLC25A26 gene expression using RT-qPCR, for global gene expression using RNA-seq, and for metabolites using gas chromatography-mass spectrometry (GC–MS). The experiment did not reveal any significant effect of the SNP on the expression levels or splicing patterns of LRIG1 or SLC25A26. The global gene expression analysis revealed that the risk allele C was associated with upregulation of several mitochondrial genes. Gene enrichment analysis of 74 differentially expressed genes in the genome revealed a significant enrichment of type I interferon response genes, where many genes were downregulated for the risk allele C. Gene expression data of IDH1 mutated LGGs from the cancer genome atlas (TCGA) revealed a similar under expression of type I interferon genes associated with the risk allele. This study found the expression levels and splicing patterns of LRIG1 and SLC25A26 were not affected by the SNP in HEK293T cells. However, the risk allele was associated with a downregulation of genes involved in the innate immune response both in the HEK293T cells and in the LGG data from TCGA.

Hip fracture is the clinically most important fracture, but the genetic architecture of hip fracture is unclear. Here, we perform a large-scale hip fracture genome-wide association study meta-analysis and Mendelian randomization study using five cohorts from European biobanks. The results show that five genetic signals associate with hip fractures. Among these, one signal associates with falls, but not with bone mineral density (BMD), while four signals are in loci known to be involved in bone biology. Mendelian randomization analyses demonstrate a strong causal effect of decreased femoral neck BMD and moderate causal effects of Alzheimer's disease and having ever smoked regularly on risk of hip fractures. The substantial causal effect of decreased femoral neck BMD on hip fractures in both young and old subjects and in both men and women supports the use of change in femoral neck BMD as a surrogate outcome for hip fractures in clinical trials.

BACKGROUND: Gliomas are complex tumors with several genetic aberrations and diverse metabolic programs contributing to their aggressive phenotypes and poor prognoses. This study defines key metabolic features that can be used to differentiate between glioma subtypes, with potential for improved diagnostics and subtype targeted therapy.

METHODS: Cross-platform global metabolomic profiling coupled with clinical, genetic, and pathological analysis of glioma tissue from 224 tumors - oligodendroglioma (n=31), astrocytoma (n=31) and glioblastoma (n=162) - were performed. Identified metabolic phenotypes were evaluated in accordance with the WHO classification, IDH-mutation, 1p/19q-codeletion, WHO-grading 2-4, and MGMT promoter methylation.

RESULTS: Distinct metabolic phenotypes separate all six analyzed glioma subtypes. IDH-mutated subtypes, expressing 2-hydroxyglutaric acid, were clearly distinguished from IDH-wildtype subtypes. Considerable metabolic heterogeneity outside of the mutated IDH pathway were also evident, with key metabolites being high expression of glycerophosphates, inositols, monosaccharides and sugar alcohols and low levels of sphingosine and lysoglycerophospholipids in IDH-mutants. Among the IDH-mutated subtypes, we observed high levels of amino acids, especially glycine and 2-aminoadipic acid, in grade 4 glioma, and N-acetyl aspartic acid in low-grade astrocytoma and oligodendroglioma. Both IDH-wildtype and mutated oligodendroglioma and glioblastoma were characterized by high levels of acylcarnitines, likely driven by rapid cell growth and hypoxic features. We found elevated levels of 5-HIAA in gliosarcoma and a subtype of oligodendroglioma not yet defined as a specific entity, indicating a previously not described role for the serotonin pathway linked to glioma with bimorphic tissue.

CONCLUSION: Key metabolic differences exist across adult glioma subtypes.

Background: Understanding the trajectory and development of disease is important and the knowledge can be used to find novel targets for therapy and new diagnostic tools for early diagnosis.

Methods: Large cohorts from different parts of the world are unique assets for research as they have systematically collected plasma and DNA over long-time periods in healthy individuals, sometimes even with repeated samples. Over time, the population in the cohort are diagnosed with many different diseases, including brain tumors.

Results: Recent studies have detected genetic variants that are associated with increased risk of glioblastoma and lower grade gliomas specifically. The impact for genetic markers to predict disease in a healthy population has been deemed low, and a relevant question is if the genetic variants for glioma are associated with risk of disease or partly consist of genes associated to survival. Both metabolite and protein spectra are currently being explored for early detection of cancer.

Conclusions: We here present a focused review of studies of genetic variants, metabolomics, and proteomics studied in prediagnostic glioma samples and discuss their potential in early diagnostics.

No strong aetiological factors have been established for glioma aside from genetic mutations and variants, ionising radiation and an inverse relationship with asthmas and allergies. Our aim was to investigate the association between pre-diagnostic immune protein levels and glioma risk. We conducted a case–control study nested in the Northern Sweden Health and Disease Study cohort. We analysed 133 glioma cases and 133 control subjects matched by age, sex and date of blood donation. ELISA or Luminex bead-based multiplex assays were used to measure plasma levels of 19 proteins. Conditional logistic regression models were used to estimate the odds ratios and 95% CIs. To further model the protein trajectories over time, the linear mixed-effects models were conducted. We found that the levels of sVEGFR2, sTNFR2, sIL-2Rα and sIL-6R were associated with glioma risk. After adjusting for the time between blood sample collection and glioma diagnosis, the odds ratios were 1.72 (95% CI = 1.01–2.93), 1.48 (95% CI = 1.01–2.16) and 1.90 (95% CI = 1.14–3.17) for sTNFR2, sIL-2Rα and sIL-6R, respectively. The trajectory of sVEGFR2 concentrations over time was different between cases and controls (p-value = 0.031), increasing for cases (0.8% per year) and constant for controls. Our findings suggest these proteins play important roles in gliomagenesis.

I och med utveckling inom storskalig analys av blodprover har man idag insett nyttan av att omvandla biobanker med lagrade humanprover till data-banker där forskare snabbt kan få tillgång till data för att svara på forsknings-frågor. Problemet är att många av teknikerna för att skapa storskaliga data är semikvantitativa, värdena går inte att relatera till en absolut koncentration och är därmed svåra att slå samman och jämföra över tid. Randomisering, det vill säga att proverna analyseras i slumpvis inbördes ordning, är en av de viktigas-te aspekterna för att skapa data som går att slå samman och återanvända för många forskningsfrågor. Detta underlättar korrigering av oönskade analysva-riationer över tid. Utöver detta kan man använda sig av bryggningsprover, QC-prov (kvalitetskontrollprov) eller ankarprover, som analyseras upprepat både inom och mellan analystillfällen, vilket underlättar att lägga samman dataset som analyseras vid olika tillfällen.

Många kommersiella analysplattformar inkluderar ett eget QC-prov i analysen och vissa delar med sig av data för dessa prover. Det vore värdefullt om alla plattformar delade dessa data för kvalitetsutvärdering och eventuell korrige-ring av analysvariationer över tid. För alla semikvantitativa plattformar som undersöktes (Olink, Somalogic, Metabolon och Biocrates) var den tekniska variabiliteten mellan QC-proverna betydligt lägre än variabiliteten mellan ana-lyserade plasmaprover. Detta var tydligast för proteomikplattformarna, vilket antyder att förutsättningarna att upptäcka biologiska skillnader är bättre i pro-teomikdata. Undantaget från detta är en femte plattform, Nightingale, en kvan-titativ men smalare metabololmikmetod som anses generera stabila mätningar.

Vid all utveckling av biomarkörpaneler för att prediktera sjukdom behöver man göra upptäcktsanalyser, sedan valideringsstudier och därefter tester i den situation man tänker att testet ska fungera. De breda omikplattformarna läm-par sig för upptäckt och eventuellt validering, men för det faktiska kliniska tes-tet behövs en kvantitativ analys för att verkligen utvärdera att de proteiner eller metaboliter man vill använda är stabilt uppmätbara och fungerar för att pre-diktera sjukdom eller risk för sjukdom.

Background: Changes in immune marker levels in the blood could be used to improve the early detection of tumor-associated inflammatory processes. To increase predictiveness and utility in cancer detection, intraindividual long-term stability in cancer-free individuals is critical for biomarker candidates as to facilitate the detection of deviation from the norm.

Methods: We assessed intraindividual long-term stability for 19 immune markers (IL10, IL13, TNFa, CXCL13, MCP-3, MIP-1a, MIP-1b, fractalkine, VEGF, FGF-2, TGFa, sIL2Ra, sIL6R, sVEGF-R2, sTNF-R1, sTNF-R2, sCD23, sCD27, and sCD30) in 304 cancer-free individuals. Repeated blood samples were collected up to 20 years apart. Intraindividual reproducibility was assessed by calculating intraclass correlation coefficients (ICC) using a linear mixed model.

Results: ICCs indicated fair to good reproducibility (ICCs ≥ 0.40 and < 0.75) for 17 of 19 investigated immune markers, including IL10, IL13, TNFa, CXCL13, MCP-3, MIP-1a, MIP-1b, fractalkine, VEGF, FGF-2, TGFa, sIL2Ra, sIL6R, sTNF-R1, sTNF-R2, sCD27, and sCD30. Reproducibility was strong (ICC ≥ 0.75) for sCD23, while reproducibility was poor (ICC < 0.40) for sVEGF-R2. Using a more stringent criterion for reproducibility (ICC ≥ 0.55), we observed either acceptable or better reproducibility for IL10, IL13, CXCL13, MCP-3, MIP-1a, MIP-1b, VEGF, FGF-2, sTNF-R1, sCD23, sCD27, and sCD30.

Conclusions: IL10, IL13, CXCL13, MCP-3, MIP-1a, MIP-1b, VEGF, FGF-2, sTNF-R1, sCD23, sCD27, and sCD30 displayed ICCs consistent with intraindividual long-term stability in cancer-free individuals. Impact: Our data support using these markers in prospective longitudinal studies seeking early cancer detection biomarkers.

Introduction: Medulloblastoma is a malignant embryonal tumor of the cerebellum that occurs predominantly in children. To find germline genetic variants associated with medulloblastoma risk, we conducted a genome-wide association study (GWAS) including 244 medulloblastoma cases and 247 control subjects from Sweden and Denmark.

Methods: Genotyping was performed using Illumina BeadChips, and untyped variants were imputed using IMPUTE2.

Results: Fifty-nine variants in 11 loci were associated with increased medulloblastoma risk (p < 1 × 10–5), but none were statistically significant after adjusting for multiple testing (p < 5 × 10–8). Thirteen of these variants were genotyped, whereas 46 were imputed. Genotyped variants were further investigated in a validation study comprising 249 medulloblastoma cases and 629 control subjects. In the validation study, rs78021424 (18p11.23, PTPRM) was associated with medulloblastoma risk with OR in the same direction as in the discovery cohort (ORT = 1.59, pvalidation = 0.02). We also selected seven medulloblastoma predisposition genes for investigation using a candidate gene approach: APC, BRCA2, PALB2, PTCH1, SUFU, TP53, and GPR161. The strongest evidence for association was found for rs201458864 (PALB2, ORT = 3.76, p = 3.2 × 10–4) and rs79036813 (PTCH1, ORA = 0.42, p = 2.6 × 10–3).

Conclusion: The results of this study, including a novel potential medulloblastoma risk loci at 18p11.23, are suggestive but need further validation in independent cohorts.