Advances in molecular characterization have reshaped our understanding of low-grade glioma (LGG) subtypes, emphasizing the need for comprehensive classification beyond histology. Lever-aging this, we present a novel approach, network-based Subnetwork Enumeration, and Analysis (nSEA), to identify distinct LGG patient groups based on dysregulated molecular pathways. Using gene expression profiles from 516 patients and a protein-protein interaction network we generated 25 million sub-networks. Through our unsupervised bottom-up approach, we selected 92 subnetworks that categorized LGG patients into five groups. Notably, a new LGG patient group with a lack of mutations in EGFR, NF1, and PTEN emerged as a previously unidentified patient subgroup with unique clinical features and subnetwork states. Validation of the patient groups on an independent dataset demonstrated the robustness of our approach and revealed consistent survival traits across different patient populations. This study offers a comprehensive molecular classification of LGG, providing insights beyond traditional genetic markers. By integrating network analysis with patient clustering, we unveil a previously overlooked patient subgroup with potential implications for prognosis and treatment strategies. Our approach sheds light on the synergistic nature of driver genes and highlights the biological relevance of the identified subnetworks. With broad implications for glioma research, our findings pave the way for further investigations into the mechanistic underpinnings of LGG subtypes and their clinical relevance.Availability: Source code and supplementary data are available at https://github.com/bebeklab/nSEA.
Continuously decreasing cost, speed and efficiency of DNA and RNA sequencing, coupled with advances in real-world sensing, storage of electronic health records, publicly available databases, and new data processing techniques enable precision medicine at unprecedented scale. Machine learning and artificial intelligence emerge naturally as tools for analyzing and summarizing data, supporting clinical decisions with data-driven insights and further unlocking genetically driven mechanisms underlying individualized risk. While these computational tools allow modeling of complex relations in large datasets, they pose new challenges especially because a patient's health is at stake. Due to an often black-box nature and high reliance on the training data, these new tools are prone to biases and most commonly provide correlational rather than causal insights. Results of these analyses have been difficult to validate, interpret, and explain to practitioners, and most genetic studies have struggled to encompass the full spectrum of human diversity. In this work, we summarize recent research trends in addressing these issues with examples from submissions to the Computational Challenges and Artificial Intelligence in Precision Medicine session at Pacific Symposium on Biocomputing 2021. We observe growing research interest in identifying biases, deriving causal and interpretable relations, tuning parameters of models for production, and using artificial intelligence for quality control. We expect further upsurge in work on interpretability and low-risk applications of advanced computational tools.
Physicians’ beliefs and attitudes about COVID-19 are important to ascertain because of their central role in providing care to patients during the pandemic. Identifying topics and sentiments discussed by physicians and other healthcare workers can lead to identification of gaps relating to the COVID-19 pandemic response within the healthcare system. To better understand physicians’ perspectives on the COVID-19 response, we extracted Twitter data from a specific user group that allows physicians to stay anonymous while expressing their perspectives about the COVID-19 pandemic. All tweets were in English. We measured most frequent bigrams and trigrams, compared sentiment analysis methods, and compared our findings to a larger Twitter dataset containing general COVID-19 related discourse. We found significant differences between the two datasets for specific topical phrases. No statistically significant difference was found in sentiments between the two datasets, and both trended slightly more positive than negative. Upon comparison to manual sentiment analysis, it was determined that these sentiment analysis methods should be improved to accurately capture sentiments of anonymous physician data. Anonymous physician social media data is a unique source of information that provides important insights into COVID-19 perspectives.
An early biomarker would transform our ability to screen and treat patients with cancer. The large amount of multi-scale molecular data in public repositories from various cancers provide unprecedented opportunities to find such a biomarker. However, despite identification of numerous molecular biomarkers using these public data, fewer than 1% have proven robust enough to translate into clinical practice1. One of the most important factors affecting the successful translation to clinical practice is lack of real-world patient population heterogeneity in the discovery process. Almost all biomarker studies analyze only a single cohort of patients with the same cancer using a single modality. Recent studies in other diseases have demonstrated the advantage of leveraging biological and technical heterogeneity across multiple independent cohorts to identify robust disease biomarkers. Here we analyzed 17149 samples from patients with one of 23 cancers that were profiled using either DNA methylation, bulk and single-cell gene expression, or protein expression in tumor and serum. First, we analyzed DNA methylation profiles of 9855 samples across 23 cancers from The Cancer Genome Atlas (TCGA). We then examined the gene expression profile of the most significantly hypomethylated gene, KRT8, in 6781 samples from 57 independent microarray datasets from NCBI GEO. KRT8 was significantly over-expressed across cancers except colon cancer (summary effect size=1.05; p < 0.0001). Further, single-cell RNAseq analysis of 7447 single cells from lung tumors showed that genes that significantly correlated with KRT8 (p < 0.05) were involved in p53-related pathways. Immunohistochemistry in tumor biopsies from 294 patients with lung cancer showed that high protein expression of KRT8 is a prognostic marker of poor survival (HR = 1.73, p = 0.01). Finally, detectable KRT8 in serum as measured by ELISA distinguished patients with pancreatic cancer from healthy controls with an AUROC=0.94. In summary, our analysis demonstrates that KRT8 is (1) differentially expressed in several cancers across all molecular modalities and (2) may be useful as a biomarker to identify patients that should be further tested for cancer.
Molecular mechanisms characterizing cancer development and progression are complex and process through thousands of interacting elements in the cell. Understanding the underlying structure of interactions requires the integration of cellular networks with extensive combinations of dysregulation patterns. Recent pan-cancer studies focused on identifying common dysregulation patterns in a confined set of pathways or targeting a manually curated set of genes. However, the complex nature of the disease presents a challenge for finding pathways that would constitute a basis for tumor progression and requires evaluation of subnetworks with functional interactions. Uncovering these relationships is critical for translational medicine and the identification of future therapeutics. We present a frequent subgraph mining algorithm to find functional dysregulation patterns across the cancer spectrum. We mined frequent subgraphs coupled with biased random walks utilizing genomic alterations, gene expression profiles, and protein-protein interaction networks. In this unsupervised approach, we have recovered expert-curated pathways previously reported for explaining the underlying biology of cancer progression in multiple cancer types. Furthermore, we have clustered the genes identified in the frequent subgraphs into highly connected networks using a greedy approach and evaluated biological significance through pathway enrichment analysis. Gene clusters further elaborated on the inherent heterogeneity of cancer samples by both suggesting specific mechanisms for cancer type and common dysregulation patterns across different cancer types. Survival analysis of sample level clusters also revealed significant differences among cancer types (p < 0.001). These results could extend the current understanding of disease etiology by identifying biologically relevant interactions.
