Research

Gene Regulatory Networks

Integration of high-dimensional genomic and epigenomic datasets, such as copy number aberration, DNA methylation, gene, and micro RNA expression datasets to compute gene regulatory interactions in cancer.

Overview:

As part of her PhD dissertation, Dr. Bose has designed a computational tool, miRDriver , which computes gene-micro RNA interaction network integrating multi-omics datasets of tumor samples utilizing a machine learning model and predicted several known and novel cancer-driver genes and micro RNAs in multiple cancer types.

Publication(s):

Description:

Patients with cancer frequently experience copy number aberrations, such as chromosomal amplifications and deletions. Regulators like microRNAs (miRNAs), which control downstream target genes involved in the crucial biological processes of carcinogenesis and proliferation, are frequently found in aberrated copy number areas. In this work, Dr. Bose investigated the networks of miRNA-gene interactions based on copy number and developed a computational pipeline, miRDriver based on the idea that cancer patient copy number data can be used to identify cancer driver miRNAs. miRDriver was applied in 18 different cancer types from The Cancer Genome Atlas (TCGA) database and discovered various miRNA-gene interactions that are common and particular to cancer, as well as interactions that are enriched in experimentally verified miRNA-target interactions. The work highlighted a number of oncogenic and tumor suppressor miRNAs that were both frequent in several cancer types and cancer-specific. In choosing significantly known miRNA-gene interactions across a range of cancer types, this strategy significantly outperformed the state-of-the-art strategies. It was discovered that a number of miRNAs and genes were linked to tumor survival and development. Selected target genes were discovered to have a considerable enrichment in Gene Ontology (GO) terms, cancer-related pathways, and cancer hallmark pathways. Additionally, in numerous cancer types, subtype-specific potential gene signatures have been found.

Flowchart of the miRDriver pipeline.

Use of miRDriver computational pipeline in pan-cancer.

Computed miRNAs were significantly enriched in the experimentally-validated oncogenic miRNAs

Twenty two common miRNAs computed by miRDriver in multiple cancer types

Enriched cancer hallmark terms in pan-cancer for computed target genes

Adjusted Kaplan–Meier plots with adjusted log-rank test p-value for 18 common miRNAs in high and low expression groups, (A) let-7a-3 in OV with OS; (B) let-7b in PAAD with OS; (C) miR-149 in ACC with PFI; (D) miR-210 in BRCA with OS; (E) miR-31 in KIRC with OS; (F) miR-3187 in HNSC with OS; (G) miR-3664 in PRAD with OS; (H) miR-4777 in LUAD with DFI; (I) miR-4786 in LIHC with OS; (J) miR-3136 in BLCA with PFI; (K) miR-34b in ESCA with PFI; (L) miR-3667 in LUSC with PFI; M) miR-4662a in UCEC with PFI; (N) miR-548k in PRAD with PFI; (O) miR-6510 in PAAD with PFI; (P) miR-4762 in LUSC with DFI; (Q) miR-486 in HNSC with DFI; (R) miR-675 in ACC with PFI.

miRDriver in ISMB 2020

Precision Medicine

Designing the computational methods which integrate multi-omics datasets of patient tumor samples with drugs’ chemical properties and side effects to predict optimal drug response in cancer patients, thus connecting the patients with the most effective and less toxic therapies.

Overview:

As part of her PhD dissertation, Dr. Bose has started designing computational tools that connect preclinical and clinical drug response models in cancer research and still continuing to develop an optimal model for patient-specific drug response integrating high-throughput genomic and genetic datasets utilizing artificial intelligence and network science methods.

Publication(s):

Description:

The functional link between genes and drug response as well as drug response prediction based on genomic and chemical properties have both been investigated using laboratory-cultured cell lines. However, determining how a medicine will affect actual patients is a more crucial and difficult issue. Bose's research combines information from primary tumors and cancer cell lines to uncover connections between cell lines and tumors in order to address this problem. To fill in these gaps, she designed a computational pipeline CTDPathSim, which uses a pathway activity-based method to compare the genetic, genomic, and epigenetic properties of primary tumor samples and cell lines while integrating multi-omics datasets in breast and ovarian cancers. According to her findings, highly comparable sample-cell line pairs respond to pharmaceuticals similarly to lowly similar pairs when using a variety of FDA-approved cancer medications, including Paclitaxel, Vinorelbine, and Mitomycin-c.

Flowchart of the CTDpathSim pipeline

CTDpathSim in ISMB 2020

Drug response concordance with sample-cell line similarity scores. A) Conditional density plot of matching (~ grey) and mismatch (~ black) of Vinorelbine’s response between BRCA samples and CCLE cell lines with computed similarity scores by five different methods: A1) CTDPathSim, A2) TSI method, A3) TC analysis, A4) SRCCM, A5) Celligner. The matching percentage of response of drug vinorelbine increases as the similarity score increases in A1, whereas in A2-A5, this drug does not show concordance between response and similarity scores. B) Boxplots of matching and mismatch response percentage with common drugs between TCGA BRCA cohort and CCLE cell lines in a low similarity threshold (≤ -1.5) for all the methods except TSI threshold (≤ -1) (with a Wilcoxson p-value in each box in red: B1) CTDPathSim shows lower median value of matching drug response in low similarity threshold with significant p-value (<0.05) , whereas all other four methods in B2-B5, there was no difference between medians of matching and mismatch responses ( p-values > 0.05). C) Boxplots of matching and mismatch response percentage with common drugs between TCGA BRCA cohort and CCLE cell lines in a high similarity threshold (≥ 1.5 )for the five methods with a Wilcoxson p-value in each box in red: C1) CTDPathSim shows the higher median value of matching drug response in high similarity threshold with significant p-value (<0.05) , whereas for TSI method and TC analysis in C2 and C3, there was no difference between medians of matching and mismatch responses ( p-values > 0.05); C4)SRCCM and C5)Celligner also show significant higher matching response in high similarity threshold.

Scatter plots of cell lines’ hazard ratios (cell-HR) vs BRCA patients’ hazard ratios (patient-HR) with natural logarithm using 43 breast cancer driver genes. A) shows Spearman rank correlation with a significant p-value in a highly similar sample-cell lines group; B) shows Spearman rank correlation with an insignificant p-value in a lowly similar sample-cell lines group.

Rank of all 1,018 CCLE cell lines based on median similarity scores with all samples predicted by CTDPathSim. A) rank of all cancer cell lines in BRCA cohort with top rank breast cancer cell lines with red dots; B) rank of all cancer cell lines in OV cohort with top rank ovarian cancer cell lines with red dots.

Overlapping KEGG pathways in highly similar sample cell line pairs (aligned group) vs lowly similar sample-cell line pairs (unaligned group) with CTDPathSim’s computed scores between BRCA cohort and CCLE cell lines. A) logarithm of p-values of overlapping pathways in aligned group vs unaligned group. Aligned group shows lower median i.e. significant p-values compared to unaligned group; B) number of significant overlapping pathways in aligned vs unaligned group. Aligned group shows higher median value than unaligned group. Both the violin plots have significant Wilcoxson p-value (< 0.05) (in red).

Elucidating sexual dimorphism in cancer

Exploring the molecular basis of sexual dimorphism within and across cancers at the molecular systems genomics level in tumors and discovering the contribution of germline genetic variation to cancer risk utilizing statistical genetics and machine learning approaches.

Overview:

As part of this project, Dr. Bose is working to identify a group of sex-biased genes that can reveal the disparity between male and female patients in a variety of tumor types using machine learning models. She is also developing a computational tool that applies various models of sex-dependent genetic effects using statistical genetics and genome-wide analyses of complex traits and disease conditions.

Publication(s):

Manuscripts are in progress.