Posters - Schedules

Posters Home

Genomic instability is characterized by the accumulation of mutations in a cell's genome and is often found in all cancer types. Recent advances in high throughput sequencing have opened the way for the elaboration of projects such as the Pan-Cancer Analysis of Whole Genomes (PCAWG) where tumors from >23,000 cancer patients were sequenced, allowing for the identification of specific mutational patterns in cancers. These patterns can be broken down into what is called mutational signatures. Some of these signatures, contained in databases like the Catalog of Somatic Mutations in Cancer (COSMIC), have been associated with different aetiologies or causes. UV radiation and tobacco smoking are examples of such causes. In some cases, mutational signatures are also linked to deficiencies of certain DNA Damage Response (DDR) pathways, enabling the use of signatures as biomarkers for tumor deficiencies/sensitivities, or even the use of certain pathways. Well-known signatures, SBS2 and SBS13, often found in cancer, are both associated with activity of the AID/APOBEC family of cytidine deaminases, enzymes capable of DNA editing. It has been shown that the deregulation of the APOBEC3 family enzymes contributes to genomic instability in cancer by generating hypermutated regions called kataegis. Upon further analysis of the mutations found in >900 PCAWG samples, we found that SBS2 and SBS13 were more easily identified when only looking at closely-spaced mutations, meaning single base substitution separated by no more than 100 bp in the genome. This method allowed us to strictly look at mutational processes producing multiple mutations near each other (also called clustered mutations). By applying this method to all samples, we were able to identify a new mutational signature only found in clustered mutation profiles that is heavily correlated to the presence of Mismatch repair (MMR) deficiency signatures present in the complete profiles of tumors. This signature can possibly be used to better understand the mechanisms that are recruited by the cell to cope with an MMR deficiency, outlining new therapeutic targets in such cancers. Our next goal will be to identify cell lines from the Cancer Cell Line Encyclopedia (CCLE) harboring this new signature, enabling studies of mutational processes in vitro.

Neuroimaging information plays a crucial role in the diagnosis and prognosis of Alzheimer's disease (AD). Magnetic Resonance Imaging (MRI) is a non-invasive medical imaging technique that uses radio waves to reveal fine details of brain anatomy and pathology. Radiologists are able to use information in MRI along with other clinical data to determine if a patient has this disease or not. However, efforts are being made by researchers to deploy computer-aided diagnostic tools to aid radiologists in MRI interpretation and reduce human errors. Deep CNNs have become the state-of-the-art technique for medical imaging classification on different imaging modalities for both binary and multiclass problems. Deep CNNs are able to extract spatial features from image data in a hierarchical manner, with deeper layers learning more features that are potentially more relevant to the classification application. This study evaluates an explainable deep CNN-based learning model for the classification of AD severity using MRI. The deep learning models are based on three pre-trained neural network architectures: DenseNet121, DenseNet169, and Inception-ResNet-v2. The framework achieved high sensitivity and specificity on the test sample of subjects with varying levels of AD severity. The deep learning framework shows promise in the classification of MR images from subjects with AD.

The Salmonella enterica serovar Newport red onion outbreak of 2020 was the largest foodborne outbreak of Salmonella in over a decade. The epidemiological investigation suggested two farms as the likely source of contamination. However, single nucleotide polymorphism (SNP) analysis of the whole genome sequencing data showed that none of the Salmonella isolates collected from the farm regions were linked to the clinical isolates—preventing the use of phylogenetics in source identification. Here, we explored an alternative method for analyzing the whole genome sequencing data driven by the hypothesis that if the outbreak strain had come from the farm regions, then the clinical isolates would disproportionately contain plasmids found in isolates from the farm regions due to horizontal transfer. SNP analysis confirmed that the clinical isolates formed a single, nearly-clonal clade with evidence for ancestry in California going back a decade. The clinical clade had a large core genome (4,399 genes) and a large and sparsely distributed accessory genome (2,577 genes, at least 64% on plasmids). At least 20 plasmid types occurred in the clinical clade, more than were found in the literature for Salmonella Newport. A small number of plasmids, 14 from 13 clinical isolates and 17 from 8 farm isolates, were found to be highly similar (>95% identical)—indicating they might be related by horizontal transfer. Phylogenetic analysis was unable to determine the geographic origin, isolation source, or time of transfer of the plasmids, likely due to their promiscuous and transient nature. However, our resampling analysis suggested that observing a similar number and combination of highly similar plasmids in random samples of environmental Salmonella enterica within NCBI Pathogen Detection database was unlikely, supporting a connection between the outbreak strain and the farms implicated by the epidemiological investigation. Horizontally transferred plasmids provided evidence for a connection between clinical isolates and the farms implicated as the source of the outbreak. Our case study suggests that such analyses might add a new dimension to source tracking investigations, but highlights the need for detailed and accurate metadata, more extensive environmental sampling, and a better understanding of plasmid molecular evolution.

Kidney cancer is a common disease worldwide and its incidence has increased in recent years. In the development of nanomedicines against cancer, it is crucial to select appropriate biomaterials that can influence the subsequent biological responses. Conventional chemotherapy is not effective for most kidney cancers, so alternative and targeted therapies are needed. While previous studies have highlighted the importance of nanomaterials in therapeutic approaches for other cancers such as breast, lung and liver cancer, little has been done in the case of kidney cancer. A thorough study is needed to determine the importance of carbon nanomaterials as therapeutics in the treatment of kidney cancer. The present study is mainly focused on the use of different types of carbon-based nanoparticles as therapeutic approaches in kidney cancers and to obtain the cancer-specific role of carbon-based nanomaterials. Definition, structure and properties, applications of CNMs as drug delivery carriers and potential biomedical/theranostic applications in renal cancer are elaborately discussed in our research. Furthermore, various methods for transforming the surfaces of carbon-based nanoparticles (shape, size, charge and surface chemistry) to target cancer cells are mentioned. Additionally, physical and chemical properties of carbon-based nanomaterials like enzymatic degradation, and biological interaction plays important role in drug delivery with carbon-based nanomaterials and are discussed in detail in our work. Our research highlights recent advances in carbon-nanoparticle-based target-specific drug delivery via gene delivery or peptide delivery to help in renal cancer therapy. Besides, the interaction of carbon nanomaterials with the tumour microenvironment is also one of the objectives of our work. Drug formulations based on carbon nanoparticles for the treatment of kidney cancer are also one of the key objectives of our work. Moreover, issues with using carbon-based nanoparticles as biomedicine in the treatment of kidney cancer are also discussed (as challenges) in this work. The objectives of our work are to illustrate the contemporary research and clinical applications of carbon nanoparticles in kidney cancer detection and treatment.

For a precise diagnosis and efficient therapy, the molecular heterogeneity of brain cancer poses a considerable challenge. Brain tumors are divided into various groups based on cell type, molecular signature, location, and developmental stage. Glioblastoma multiforme (GBM), an aggressive subtype of gliomas, is the most prevalent type of brain cancer. Furthermore, GBM is categorized into three subgroups: classical, proneural, and mesenchymal. Identification of these subtypes is crucial for clinicians to begin systematic treatment. In the current work, we design the deep learning (DL) frameworks using transcriptome and methylome data to classify the subtypes of GBM. DL-model with genomics data is challenging because of high dimensionality and lack of biological interpretability. Due to their lack of biological interpretability, these models are rarely used in clinical practice. Therefore, we created two distinct frameworks to integrate methylation and gene expression, reduce dimensionality, and finally design a model for subtyping GBM. In the first approach, we mapped and integrated the differentially methylated regions (DMRs) and differentially expressed genes (DEGs) based on the promoter regions. Then we applied the LASSO feature selection to differentially methylated gene (gene expression data) and developed the DL model using LASSO features and CNN; and achieved a classification accuracy of 98.20% (±0.05) on training data with tenfold cross-validation and 94.48% (±0.11) using external data. In the second strategy, we use cox regression analysis to screen the survival-associated DMRs and DEGs in order to create a DL model that is more biologically informed. These survival-related features were mapped, and both gene expression and methylation levels of DEGs and DMRs were integrated using an autoencoder. CNN was used to classify the subtype of GBM, achieving a prediction accuracy of 94.07% (±0.01) on training data with 10-fold cross-validation and 86.41% (±0.24) utilizing external data. We noticed that the model's accuracy decreased when survival-related features were chosen and autoencoder-based integration was used. Although accuracy was decreased, the survival-associated features can be used to develop subtype-specific biomarkers for precision therapy. Additionally, significant associations of features with biological processes and pathways confirm that the biologically relevant DL framework can be used in clinical settings to support the diagnosis of GBM subtypes.

Genome-wide association studies (GWAS) aim to find associations between genotypes and phenotypes to identify specific genetic loci that affect disease risk. GWAS has been successful in discovering many new risk loci for diseases, yet we are unable to fully explain the heritability of many diseases with the risk loci discovered from GWAS. This missing heritability could be explained by genetic interactions. Our lab previously showed that genetic interactions underlying human diseases form structured networks connecting within and between biological pathways, and we developed a method called BridGE for the systematic discovery of genetic interactions. This approach has been applied to discover complex genetic interactions from human genotype cohort data for multiple diseases. However, it has not been applied to improve predictive models for quantifying individuals’ disease risk.

In this work, we created a machine-learning pipeline for predicting case-control phenotypes that leverages combinatorial (pairwise) genotype information and curated gene sets (e.g., pathways). We tested our pipeline on two independent Parkinson's disease cohorts, and our preliminary results suggest that models that include variant pairs as features can improve predictions over models based on only collections of single variants. These results demonstrate the utility of incorporating combinations of loci into disease risk prediction models such as polygenic risk scores.

Gastric Cancer, the fifth most common cancer and the fourth leading cause of cancer-related deaths, is a global disease that continues to plague the world. Recent studies have also shown that microbiome composition and diversity changes could contribute to gastric carcinogenesis; however, little is known about microbial community interactions and keystone species that drive these changes. We present a comprehensive microbial network analysis of the gastric microbiome from 985 gastric samples spanning four gastric conditions, i.e., Healthy control, Gastritis, intestinal metaplasia, and gastric cancer, revealing essential microbial interactions and keystone species that affect gastric carcinogenesis. We employed Semi-Parametric Rank-based approach for Inference in the Graphical model (SPRING) through the NetCoMi package in R for the microbial network analysis. The topological properties of the various networks for gastric conditions revealed that the clustering coefficient, indicative of the strength and complexity of microbial interactions, decreased with increasing carcinogenesis. Higher modularity and lower averaged path was observed in the healthy control group compared to pre-cancerous and cancer groups, reflecting
less divergent and functional groups with increasing carcinogenesis. Pseudoxanthomonas spadix and Sphingobium xenophagum were represented as major keystone species driving the healthy control community. In the metaplasia group, Prevotella salivae identified as keystone species showed interaction with Bacteroides dorei, Dialister invisus, and Prevotella nigrescens, which have been reported to be associated with cancer progression. Microbial interactions in the gastric
cancer group were dominated by Prevetella spp.; however, Filifactor alocis showed significant interaction with pathogenic bacteria Fusobacterium nucleatum in gastric cancer communities. In the network comparison, different keystone species were observed when the interaction of the core microbiome between the metaplasia and gastric group. Three hubs (Anaerococcus nagyae, Bactetiodes dorei, and Pelomonas aquatica) were observed in the intestinal metaplasia group while two hubs (Catonella morbi and Peptostreptococcus stomatis) in the gastric cancer group. In
conclusion, this study provides significant insight into the gastric microbial community interaction and how they could serve as potential biomarkers for understanding gastric carcinogenesis, thereby enhancing treatment and therapeutic strategies for gastric cancer.

Electronic Health Records (EHRs) encompass a diverse array of personal clinical data, and their widespread adoption has given rise to extensive EHR databases. Machine learning techniques, including topic models, have been developed to leverage EHR data in identifying phenotypic correlations and predicting patients' disease risks. However, existing topic modeling methods often face challenges due to the scarcity of reliable disease labels for topic interpretation and the inability to integrate predictive properties and survival-related information. To overcome these limitations, we propose MixEHR-SURG, which is a PheCode-guided topic model that jointly models the patient survival data and diagnostic code data, with the aim of improving both the interpretability and predictive capabilities of the model.

In order to identify phenotype-specific topics related to patient survival outcomes, we first mapped patients' International Classification of Diseases (ICD) codes to well-defined Phenotype Codes. We jointly fit a Cox Proportional Hazards (PH) model and the guided topic inference with the former operates inferred phenotypic topic mixture for each patient and the latter depends on the Cox partial likelihood. For efficient inference, we employ a modified stochastic collapsed variational inference method to estimate model parameters

We evaluated the interpretability and prediction accuracy of our model using both simulation and real-world electronic health record (EHR) datasets. The first EHR dataset was the Quebec congenital heart disease (CHD) database, which includes 84,498 patients and 28 years of follow-up from 1983 to 2010. We used all EHRs before a patient's first heart failure discharge to predict their death time. The second dataset was the MIMIC-III relational database, which contains data from 38,597 adult patients and 7,870 neonates admitted to critical care units between 2001 and 2012. We incorporated a patient's EHRs during their first inpatient period to predict their mortality time. We compared our model with the baseline methods and evaluated the predictive results using the concordance index and dynamic cumulative area under the receiver operating characteristic (AUROC) curve. Our model achieved AUROC of 0.93 on experiment data and 0.63 on real dataset with more interpretable and clinically meaningful topics.

In summary, MixEHR-SURG improves topic modeling techniques by incorporating survival information supervision and trustworthy phenotypes as topic priors, resulting in increased interpretability and predictiveness of mortality outcomes of the recently discharged ICU patients.

Motivation: Multi-protein assemblies play a crucial role in several cellular processes. Studying the functional basis of such complexes begins with the analysis of protein-protein interactions. Several studies have highlighted the significance of interfacial residues in protein-protein complexes and their role in conferring stability and specificity to the complex. However, inter-chain bifurcated interactions which are very significant for protein assemblies remains an unexplored corner.
Results: In this work, the features of inter-protein bifurcated interactions in multi-protein complexes have been investigated. We found that bifurcated inter-protein interactions are present in over 635 out of 682 multi-protein assemblies in our dataset. Arg, Tyr, and Leu display the highest propensity to participate in bifurcated inter-protein interactions. Further, we found that most of these residues are hotspots, and are moderate to highly conserved, with a few exceptions. We explain the biological significance of bifurcated interactions through a few case studies. Overall, this study expands the knowledge on protein-protein interactions paving the way for the learning on multi-protein assemblies.