Cross-platform normalization enables machine learning model training on microarray and RNA-seq data simultaneously

Machine learning and related approaches in transcriptomics

Data acquisition for transcriptomic studies used to be the bottleneck in the transcriptomic analytical pipeline. However, recent developments in transcriptome profiling technologies have increased researchers' ability to obtain data, resulting in a shift in focus to data analysis. Incorporating machine learning to traditional analytical methods allows the possibility of handling larger volumes of complex data more efficiently. Many bioinformaticians, especially those unfamiliar with ML in the study of human transcriptomics and complex biological systems, face a significant barrier stemming from their limited awareness of the current landscape of ML utilisation in this field. To address this gap, this review endeavours to introduce those individuals to the general types of ML, followed by a comprehensive range of more specific techniques, demonstrated through examples of their incorporation into analytical pipelines for human transcriptome investigations. Important computational aspects such as data pre-processing, task formulation, results (performance of ML models), and validation methods are encompassed. In hope of better practical relevance, there is a strong focus on studies published within the last five years, almost exclusively examining human transcriptomes, with outcomes compared with standard non-ML tools.

Investigating heart rate variability measures during pregnancy as predictors of postpartum depression and anxiety: an exploratory study

Perinatal affective disorders are common, but standard screening measures reliant on subjective self-reports might not be sufficient to identify pregnant women at-risk for developing postpartum depression and anxiety. Lower heart rate variability (HRV) has been shown to be associated with affective disorders. The current exploratory study aimed to evaluate the predictive utility of late pregnancy HRV measurements of postpartum affective symptoms. A subset of participants from the BASIC study (Uppsala, Sweden) took part in a sub-study at pregnancy week 38 where HRV was measured before and after a mild stressor (n = 122). Outcome measures were 6-week postpartum depression and anxiety symptoms as quantified by the Edinburgh Postnatal Depression Scale (EPDS) and the Beck Anxiety Inventory (BAI). In total, 112 women were included in a depression outcome analysis and 106 women were included in an anxiety outcome analysis. Group comparisons indicated that lower pregnancy HRV was associated with depressive or anxious symptomatology at 6 weeks postpartum. Elastic net logistic regression analyses indicated that HRV indices alone were not predictive of postpartum depression or anxiety outcomes, but HRV indices were selected as predictors in a combined model with background and pregnancy variables. ROC curves for the combined models gave an area under the curve (AUC) of 0.93 for the depression outcome and an AUC of 0.83 for the anxiety outcome. HRV indices predictive of postpartum depression generally differed from those predictive of postpartum anxiety. HRV indices did not significantly improve prediction models comprised of psychological measures only in women with pregnancy depression or anxiety.

Feature-specific quantile normalization and feature-specific mean-variance normalization deliver robust bi-directional classification and feature selection performance between microarray and RNAseq data

BACKGROUND

Cross-platform normalization seeks to minimize technological bias between microarray and RNAseq whole-transcriptome data. Incorporating multiple gene expression platforms permits external validation of experimental findings, and augments training sets for machine learning models. Here, we compare the performance of Feature Specific Quantile Normalization (FSQN) to a previously used but unvalidated and uncharacterized method we label as Feature Specific Mean Variance Normalization (FSMVN). We evaluate the performance of these methods for bidirectional normalization in the context of nested feature selection.

RESULTS

FSQN and FSMVN provided clinically equivalent bidirectional model performance with and without feature selection for colon CMS and breast PAM50 classification. Using principal component analysis, we determine that these methods eliminate batch effects related to technological platforms. Without feature selection, no statistical difference was identified between the performance of FSQN and FSMVN of cross-platform data compared to within-platform distributions. Under optimal feature selection conditions, balanced accuracy was FSQN and FSMVN were statistically equivalent to the within-platform distribution performance in multivariable linear regression analysis. FSQN and FSMVN also provided similar performance to within-platform distributions as the number of selected genes used to create models decreases.

CONCLUSIONS

In the context of generating supervised machine learning classifiers for molecular subtypes, FSQN and FSMVN are equally effective. Under optimal modeling conditions, FSQN and FSMVN provide equivalent model accuracy performance on cross-platform normalization data compared to within-platform data. Using cross-platform data should still be approached with caution as subtle performance differences may exist depending on the classification problem, training, and testing distributions.

Genetic and transcriptomic analyses of diffuse large B-cell lymphoma patients with poor outcomes within two years of diagnosis

Despite the improvements in clinical outcomes for DLBCL, a significant proportion of patients still face challenges with refractory/relapsed (R/R) disease after receiving first-line R-CHOP treatment. To further elucidate the underlying mechanism of R/R disease and to develop methods for identifying patients at risk of early disease progression, we integrated clinical, genetic and transcriptomic data derived from 2805 R-CHOP-treated patients from seven independent cohorts. Among these, 887 patients exhibited R/R disease within two years (poor outcome), and 1918 patients remained in remission at two years (good outcome). Our analysis identified four preferentially mutated genes (TP53, MYD88, SPEN, MYC) in the untreated (diagnostic) tumor samples from patients with poor outcomes. Furthermore, transcriptomic analysis revealed a distinct gene expression pattern linked to poor outcomes, affecting pathways involved in cell adhesion/migration, T-cell activation/regulation, PI3K, and NF-κB signaling. Moreover, we developed and validated a 24-gene expression score as an independent prognostic predictor for treatment outcomes. This score also demonstrated efficacy in further stratifying high-risk patients when integrated with existing genetic or cell-of-origin subtypes, including the unclassified cases in these models. Finally, based on these findings, we developed an online analysis tool ( https://lymphprog.serve.scilifelab.se/app/lymphprog ) that can be used for prognostic prediction for DLBCL patients.

Refining breast cancer biomarker discovery and drug targeting through an advanced data-driven approach

Breast cancer remains a major public health challenge worldwide. The identification of accurate biomarkers is critical for the early detection and effective treatment of breast cancer. This study utilizes an integrative machine learning approach to analyze breast cancer gene expression data for superior biomarker and drug target discovery. Gene expression datasets, obtained from the GEO database, were merged post-preprocessing. From the merged dataset, differential expression analysis between breast cancer and normal samples revealed 164 differentially expressed genes. Meanwhile, a separate gene expression dataset revealed 350 differentially expressed genes. Additionally, the BGWO_SA_Ens algorithm, integrating binary grey wolf optimization and simulated annealing with an ensemble classifier, was employed on gene expression datasets to identify predictive genes including TOP2A, AKR1C3, EZH2, MMP1, EDNRB, S100B, and SPP1. From over 10,000 genes, BGWO_SA_Ens identified 1404 in the merged dataset (F1 score: 0.981, PR-AUC: 0.998, ROC-AUC: 0.995) and 1710 in the GSE45827 dataset (F1 score: 0.965, PR-AUC: 0.986, ROC-AUC: 0.972). The intersection of DEGs and BGWO_SA_Ens selected genes revealed 35 superior genes that were consistently significant across methods. Enrichment analyses uncovered the involvement of these superior genes in key pathways such as AMPK, Adipocytokine, and PPAR signaling. Protein-protein interaction network analysis highlighted subnetworks and central nodes. Finally, a drug-gene interaction investigation revealed connections between superior genes and anticancer drugs. Collectively, the machine learning workflow identified a robust gene signature for breast cancer, illuminated their biological roles, interactions and therapeutic associations, and underscored the potential of computational approaches in biomarker discovery and precision oncology.

Identification of a Complex Karyotype Signature with Clinical Implications in AML and MDS-EB Using Gene Expression Profiling

Complex karyotype (CK) is associated with a poor prognosis in both acute myeloid leukemia (AML) and myelodysplastic syndrome with excess blasts (MDS-EB). Transcriptomic analyses have improved our understanding of the disease and risk stratification of myeloid neoplasms; however, CK-specific gene expression signatures have been rarely investigated. In this study, we developed and validated a CK-specific gene expression signature. Differential gene expression analysis between the CK and non-CK groups using data from 348 patients with AML and MDS-EB from four cohorts revealed enrichment of the downregulated genes localized on chromosome 5q or 7q, suggesting that haploinsufficiency due to the deletion of these chromosomes possibly underlies CK pathogenesis. We built a robust transcriptional model for CK prediction using LASSO regression for gene subset selection and validated it using the leave-one-out cross-validation method for fitting the logistic regression model. We established a 10-gene CK signature (CKS) predictive of CK with high predictive accuracy (accuracy 94.22%; AUC 0.977). CKS was significantly associated with shorter overall survival in three independent cohorts, and was comparable to that of previously established risk stratification models for AML. Furthermore, we explored of therapeutic targets among the genes comprising CKS and identified the dysregulated expression of superoxide dismutase 1 (SOD1) gene, which is potentially amenable to SOD1 inhibitors.

A multi-platform normalization method for meta-analysis of gene expression data

Transcriptomic profiling is a mainstay of translational cancer research and is often used to identify cancer subtypes, stratify responders vs. non-responders patients, predict survival, and identify potential targets for therapeutic intervention. Analysis of gene expression data gathered by RNA sequencing (RNA-seq) and microarray is generally the first step in identifying and characterizing cancer-associated molecular determinants. The methodological advancements and reduced costs associated with transcriptomic profiling have increased the number of publicly available gene expression profiles for cancer subtypes. Data integration from multiple datasets is routinely done to increase the number of samples, improve statistical power, and provide better insight into the heterogeneity of the biological determinant. However, utilizing raw data from multiple platforms, species, and sources introduces systematic variations due to noise, batch effects, and biases. As such, the integrated data is mathematically adjusted through normalization, which allows direct comparison of expression measures among studies while minimizing technical and systemic variations. This study applied meta-analysis to multiple independent Affymetrix microarray and Illumina RNA-seq datasets available through the Gene Expression Omnibus (GEO) and The Cancer Gene Atlas (TCGA). We have previously identified a tripartite motif containing 37 (TRIM37), a breast cancer oncogene, that drives tumorigenesis and metastasis in triple-negative breast cancer. In this article, we adapted and assessed the validity of Stouffer's z-score normalization method to interrogate TRIM37 expression across different cancer types using multiple large-scale datasets.

A Predictive Model of Adaptive Resistance to BRAF/MEK Inhibitors in Melanoma

The adaptive acquisition of resistance to BRAF and MEK inhibitor-based therapy is a common feature of melanoma cells and contributes to poor patient treatment outcomes. Leveraging insights from a proteomic study and publicly available transcriptomic data, we evaluated the predictive capacity of a gene panel corresponding to proteins differentially abundant between treatment-sensitive and treatment-resistant cell lines, deciphering predictors of treatment resistance and potential resistance mechanisms to BRAF/MEK inhibitor therapy in patient biopsy samples. From our analysis, a 13-gene signature panel, in both test and validation datasets, could identify treatment-resistant or progressed melanoma cases with an accuracy and sensitivity of over 70%. The dysregulation of HMOX1, ICAM, MMP2, and SPARC defined a BRAF/MEK treatment-resistant landscape, with resistant cases showing a >2-fold risk of expression of these genes. Furthermore, we utilized a combination of functional enrichment- and gene expression-derived scores to model and identify pathways, such as HMOX1-mediated mitochondrial stress response, as potential key drivers of the emergence of a BRAF/MEK inhibitor-resistant state in melanoma cells. Overall, our results highlight the utility of these genes in predicting treatment outcomes and the underlying mechanisms that can be targeted to reduce the development of resistance to BRAF/MEK targeted therapy.

Circular RNA expression and the competitive endogenous RNA network in pathological, age-related macular degeneration events: A cross-platform normalization study

Age-related macular degeneration (AMD) causes irreversible blindness in people aged over 50 worldwide. The dysfunction of the retinal pigment epithelium is the primary cause of atrophic AMD. In the current study, we used the ComBat and Training Distribution Matching method to integrate data obtained from the Gene Expression Omnibus database. We analyzed the integrated sequencing data by the Gene Set Enrichment Analysis. Peroxisome and tumor necrosis factor-α (TNF-α) signaling and nuclear factor kappa B (NF-κB) were among the top 10 pathways, and thus we selected them to construct AMD cell models to identify differentially expressed circular RNAs (circRNAs). We then constructed a competing endogenous RNA network, which is related to differentially expressed circRNAs. This network included seven circRNAs, 15 microRNAs, and 82 mRNAs. The Kyoto Encyclopedia of Genes and Genomes analysis of mRNAs in this network showed that the hypoxia-inducible factor-1 (HIF-1) signaling pathway was a common downstream event. The results of the current study may provide insights into the pathological processes of atrophic AMD.