Posters - Schedules

Posters Home

Myeloproliferative Neoplasms (MPNs) are clonal hematopoietic stem cell disorders that develop due to an abnormal mutation in the bone marrow, causing uncontrolled proliferation of blood cells. Detection and treatment for MPNs are through bone marrow biopsies, which tests for mutations in the JAK2, MPL, or CALR genes; these are mostly invasive and painful processes, carrying a risk for infection and especially for the vulnerable older patients the disease commonly targets. This project focuses on a clinal alternative: peripheral blood (PB) samples, which are blood cells produced in the bone marrow and circulating throughout the body. Using an expression profiling dataset, over 60,000 affymetrix probe IDs were analyzed from 60 MN patients, and both peripheral and bone marrow samples were utilized to develop a novel machine learning model. This model leverages many machine learning techniques, including Multi-Layer Perceptron neural networks, one-hot encode, multi-classification, binary and categorical cross entropy, and Principal Component Analysis to detect these mutations through an input of PB samples, reducing the need for bone marrow sampling. The research also recognizes novel biomarkers of the myeloproliferative disorders through feature engineering, like VCAN, THBS1, BLNK, to name a few, which can promote future research on the disease.

Trigeminal neurons play a crucial role in the perception of pain. Single-cell transcriptomics provides a promising avenue to decipher the underlying mechanisms of pain. In this case, a reference remains missing. To fill the gap, we have collected the most comprehensive datasets and employed a meta-analysis. We collected four public datasets for trigeminal tissues across various technologies. Raw sequencing reads were aligned to the mm10 reference genome using CellRanger . Standardized quality control analysis has been performed to exclude estimated empty cells, ambient RNAs, and doublets using dropkick, SoupX, and DoubletFinder. Processed datasets are integrated to reduce the sample effects by using scVI. We clustered the cells using Leiden algorithm. UMAP was used to generate cell clusters. We annotated clusters using known cell type marker genes. We used 4 public available datasets, 3 dropseq, and 1 from 10x genomics. We achieved 44,770 cells, 16 clusters, consisting of 6 neuron populations, 3 glial, 4 immune, fibroblasts and endothelial cells. To facilitate the public use of the generated atlas, we provided an automated cell type annotation utility using scArches, and the reference is visualized by CELLxGENE. This atlas may serve as a valuable data resource for the mouse trigeminal community.

An aptamer composed of single-stranded DNA/RNA forms a three-dimensional structure and binds to a target molecule. Since this method works similarly to the way antigens and antibodies bind, aptamers are regarded as new biomaterials to substitute the antibodies in modern drug development and cancer treatment research. SELEX is an experiment for discovering representative aptamer candidate sequences, but it takes a lot of rounds to search aptamer sequences that interact with a target protein.
In this study, we have developed a pipeline that generates aptamer candidate sequences for a given target protein. We have designed a machine learning model for determining aptamer-protein interactions (API) using pre-trained transformer-based encoders. The scores obtained by the API model are used to generate candidate aptamer sequences using Monte Carlo tree search and Genetic algorithm. Our aptamer candidate sequences have been evaluated using ZDOCK docking simulation with benchmark data publicly available. This evaluation illustrates that the aptamer candidate sequences generated by our pipeline show higher binding affinity than ones by other exiting tools.

In research a patient is often characterized by multiple data types, commonly clinical and (diverse) omics data. Correct pairing of patient samples is essential. However, practice shows that sample swaps regularly occur. The problem can be addressed by estimating clinical characteristics on the basis of omics data, e.g. 'sex' vs reads covering the y-chromosome.

In a feasibility study, we explored the opportunities and limitations of this approach. As ground truth, we simulated data for two clinical parameters: 'sex' and 'height' (n=100). We assume that both parameters can be predicted on the basis of omics data. Considering different levels of variation, we simulated the predicted values of sex and height. To identify sample swaps, a similarity score is calculated. Low scoring pairs are flagged as non-matching and alternative pairings are suggested that result in overall higher scores.

Dependent on the simulated variation, our pipeline succeeded in correcting 47% (variation=4) up to 83% (variation=0.4) of the swaps. For all simulated scenarios, precision was 86-100%. Accuracy of our approach could be further improved as more clinical parameters are considered.

As a next step, we plan to analyze real data, training a machine learning model for the prediction of clinical parameters.

Changes in gene expression identify important genes with key roles in disease diagnosis, prognosis and therapeutics. While causal genes may vary with patient cohorts and experimental approach, often, these diverse set of genes converge to a small set of pathways. We hypothesize that considering both genes and pathways can lead to better understanding of the biology underlying a disease. To evaluate our hypothesis, we use P-NET, a deep learning framework, to identify the important genes and pathways that differentiate between neoplastic and normal cells/tissues in different cancers. This comparison study is done using both scRNAseq and bulk RNAseq data for breast and prostate cancer and two different scRNAseq datasets for Glioblastoma Multiforme (GBM). P-NET yields a satisfactory accuracy in postulating relevant pathways, whether using bulk or scRNAseq data for all three types of cancers. Although the list of important genes appears to vary with different datasets, a list of prominent and almost similar pathways is obtained for different datasets corresponding to a specific cancer type. We validate the pathway lists with the KEGG disease database for each cancer type. Literature survey further validates genes as well as pathways not explicitly mentioned in KEGG.

Serum response factor is a transcription factor present in all types of cells and plays a role in muscle tissue development. SRF is activated by growth factor stimulation and mitosis, leading to the expression of genes that influence growth and the cytoskeleton. Additionally, SRF in gastric cancer is associated with an aggressive phenotype and a poor outcome due to the downregulation of E-cadherin which promotes the epithelial-mesenchymal transition. In colorectal cancer, SRF is overexpressed in metastatic tissues, leading to increased cell motility and invasiveness. Based on this, we decided to look for potential SRF inhibitors. We are currently using chemical similarity algorithms and clustering techniques, like Tanimoto similarity and UMAP, to determine SRF inhibitor candidates based on limited existing inhibitors. Those candidates will then be docked using Autodock Vina. Molecules with high binding affinities will be tested for drug-induced liver injury (DILI) and toxicity in cells (DeepCDR). Furthermore, AutoGrow, an open-source program which uses a genetic algorithm to ‘evolve’ known ligands based on binding affinity to the target, will be used for de-novo drug design. Preliminary results reveal inhibitors with better binding affinities than positive controls from a ChemBL dataset along with preliminary drugs from Autogrow with even better binding affinities.

Serum response factor is a transcription factor present in all types of cells and plays a role in muscle tissue development. SRF is activated by growth factor stimulation and mitosis, leading to the expression of genes that influence growth and the cytoskeleton. Additionally, SRF in gastric cancer is associated with an aggressive phenotype and a poor outcome due to the downregulation of E-cadherin which promotes the epithelial-mesenchymal transition. In colorectal cancer, SRF is overexpressed in metastatic tissues, leading to increased cell motility and invasiveness. Based on this, we decided to look for potential SRF inhibitors. We are currently using chemical similarity algorithms and clustering techniques, like Tanimoto similarity and UMAP, to determine SRF inhibitor candidates based on limited existing inhibitors. Those candidates will then be docked using Autodock Vina. Molecules with high binding affinities will be tested for drug-induced liver injury (DILI) and toxicity in cells (DeepCDR). Furthermore, AutoGrow, an open-source program which uses a genetic algorithm to ‘evolve’ known ligands based on binding affinity to the target, will be used for de-novo drug design. Preliminary results reveal inhibitors with better binding affinities than positive controls from a ChemBL dataset along with preliminary drugs from Autogrow with even better binding affinities.

Lung cancer is one of the most commonly occurring malignant cancer with highest rate of mortality globally. Difference between lung squamous cell carcinoma (LUSC) and lung adenocarcinoma (LUAD) and their treatment strategies according to genetic markers may be helpful to reduce the cancer progression and increase the overall survival (OS) in patients. The mRNA-seq data and clinical information of LUAD and LUSC cohorts from UCSC Xena comprising a total of 437 and 379 patient samples were extracted. Differential expression and weighted network analyses revealed a total of 47 and 18 hub differentially expressed genes (DEGs) corresponding to LUAD and LUSC cohorts. These hub DEGs were further subjected to protein-protein interaction network (PPIN) and OS analyses. Lower mRNA expression levels of both RSP15A and RPS7 worsened the OS of LUSC patients. Additionally, both these prognostic biomarkers were validated via external sources such as UALCAN, cBioPortal, TIMER, and HPA. RPS7 had higher mutation frequency compared to RSP15A and showed significant negative correlations with infiltrating levels of CD4+T cells, CD8+T cells, neutrophils and macrophages. Our findings provided novel insights in the biomarker discovery and a critical role of ribosomal biogenesis especially smaller ribosomal subunit in pathogenesis of LUSC.

Proteases play a crucial role in various cellular processes, and the precise cleavage of substrates by proteases is essential for these processes to occur correctly. Accurately predicting substrate cleavage sites is a crucial step in understanding protease function and substrate specificity. Many bioinformatics methods have been developed to predict protease-specific substrate cleavage sites. However, with the development of mass spectrometry technology, an enormous amount of protease substrate cleavage data has been generated and will continually grow in the future. Consequently, it is not efficient and practical to train new models based on these rapidly accumulated data every year and update the prediction server for the wider community. In this study, we developed a multi-faceted, versatile bioinformatics tool, termed ProsperousPlus, that enables fast, accurate and high-throughput prediction of substrate cleavage sites for 110 proteases. Benchmarking tests show that ProsperousPlus achieves competitive predictive performance compared with state-of-the-art approaches. Furthermore, ProsperousPlus provides sought-after assistance for non-programming background users to build their customised in-house models and easily meet specific needs. It is anticipated that researchers with little bioinformatics expertise will be able to efficiently use rapidly accumulating substrate cleavage data to train in-house prediction models to meet their specific requirements.

Gene expression profiling is used to study cellular and disease contexts. While large-scale transcriptome data are searched using phenotype metadata, reverse inference of phenotype from a newly generated transcriptome profile called “content” is of interest too. Such profile comparisons and content-based queries are important for drug repositioning and identifying novel drug-disease connections or drug-drug relationships. Using a deep-learning model, Subramanium et al. 2017 showed that landmark (L1000) genes can predict whole transcriptome profiles. We have developed a database of global expression profiles from legacy microarray and RNASeq experiments to investigate the exclusiveness of and alternatives to L1000 genesets in reproducing whole-transcriptome similarity and profile-profile comparisons. Biological pathways, gene ontologies, and functional features critical for selecting such sets were investigated. Our results suggest that many L1000-like subsets are equally powerful in detecting profile-profile similarities. We also identify some genes, gene ontologies, and pathways frequently enriched in the best-performing geneset alternatives, highlighting their criticality for determining whole transcriptome patterns. Best-performing genesets and frequently occurring genes have been further examined for their roles as potential master regulators and ability to allow large-scale imputation of missing gene expressions in transcriptomic data sets.

Glioblastoma (GBM) is a lethal brain cancer characterized by inevitable relapse. Understanding intra-tumoral heterogeneity and the microenvironmental influence on tumor progression is key to developing improved therapies. We employ spatial transcriptomics to study patient derived models of GBM from early engraftment to disease endpoint. This resource enables the study of tumor growth at spatial resolution and allows robust species-specific distinction of tumor and its microenvironment (TME).

We identify gene expression programs specific to the tumor and TME using matrix factorization and observe unique niches and cell types. These include progenitor cell-states along the tumor leading edge, oligodendrocyte-like states at the tumor core, the homing of microglia, macrophages, and reactive astrocytes to the tumor site. Invasion is a major therapeutic obstacle, thus, targeting the leading edge of GBM could improve patient outcomes. In our models, the invasive state exhibits upregulation of diverse ligand/receptor genes including angiopoietin, fibronectin, and macrophage migration pathways, highlighting prospective targets. A subset of these genes constitute hubs in protein-protein interaction networks and their upregulation is significantly associated with reduced survival. Further investigation of these druggable program-specific vulnerabilities will help inform combinatorial therapy strategies and preclinical testing of treatment in xenografts.

Bromodomain and extra-terminal domain (BET) proteins have been linked to increases in oncogene expression and tumor progression in cancer. BET inhibitors (BETi) and other drugs in combination can moderately reduce colorectal cancer cell proliferation. Limited treatments exist for colorectal cancer due to its malignant nature and existing treatments are often costly or ineffective. Our research centers around determining potential BETi in colorectal cancer through in-silico research and testing identified drug candidates in an in-vitro setting. While previous research has been conducted on BETi, few studies examine the effects of BETi in colorectal cancer. We have created a list of one hundred possible BETi drugs. By utilizing a deep learning Cancer Drug Response (Deep CDR) prediction algorithm, we will rank our list of potential drug candidates. We are also working on identifying additional targets in HCT116 cells. In order to do this, we are analyzing gene expression datasets using R and ranking candidates related to the BET protein pathway. Once the in-silico analysis is complete, the drugs will be ordered/synthesized and tested on HCT116 colorectal cancer cells. They will be tested through MTT Assays, Western Blot and qPCR . We hope to increase the number of cancer treatment options for patients.

Background: Limited prognostic information from single modal data and high cost of acquiring multimodal data make precision treatment of colorectal cancer (CRC) challenging. We aimed to develop and validate a clinically applicable multimodal fusion model for survival prediction (CAMP).

Methods: Using the largest fully matched multimodal CRC cohort (Clinical Omics study of Colorectal Cancer in China, n = 587), we developed a hypergraph convolutional network-based multimodal fusion model to integrate digital pathology, gene expression, and clinical information for survival prediction. The CAMP model was further tested when several modalities were missing and validated in external multimodal cohorts (n = 517). Comprehensive visualization and analyses were conducted for model interpretation.

Results: The CAMP model achieved the highest performance (C-index 0.735) compared to other state-of-the-art models. The CAMP model trained by multimodal data remained stable performance even without certain modal data. Moreover, our model significantly distinguished prognosis of CRC patients in external cohorts without fine-tuning. Interestingly, compared to the single modal model, the multimodal model had higher attention scores at the tumor invasion front and stromal regions.

Conclusion: By effectively integrating multimodal data, our model can provide accurate and flexible survival prediction of CRC to facilitate personalized treatment and follow-up strategies.

Background: Heterogeneity of tumor immune microenvironment accounts for differential prognosis and immunotherapy response among colorectal cancer (CRC) patients. Here, we developed novel immune subtypes through integrative multi-omics analysis to characterize CRC heterogeneity.

Methods: Immune-related gene expression, mutation, and methylation profiles were collected from TCGA (n = 627) to perform multi-omics factor analysis (MOFA) and establish the Multi-Omics Tumor Immune Features Clustering of CRC (MotifCC). Transcriptomic, genomic, and epigenetic landscapes were analyzed to characterize differences among MotifCC clusters. Independent validation of MotifCC was performed in our large-scale, in-house COCC cohort (Clinical Omics Study of Colorectal Cancer in China, n = 1001).

Results: The three MotifCC clusters showed distinct characteristics. Cluster1 was a high immune and stroma infiltration subtype with the worst prognosis. Cluster2 was characterized by the low immune infiltration, high metabolic intensity and the best survival. With medium immune infiltration and intermediate prognosis, Cluster3 was distinguished by the highest methylation and stemness status. Besides, Cluster3 was associated with intermediate prognosis but better response to immunotherapy.

Conclusion: We established the MotifCC, novel immune subtypes capturing the multi-omics heterogeneity of CRC and facilitating patient stratification for immunotherapy.

Background: Gene expression-based molecular subtypes can effectively dissect the tumor heterogeneity. However, the prognostic value of molecular subtypes is limited, and their clinical translation remains challenging. This study aims to establish the consensus molecular subtypes (CMS)-related radiogenomic signatures for survival prediction of colorectal cancer (CRC).

Methods: This study was performed in three steps. Firstly, we extracted feature representations of CMS subtypes by utilizing a supervised deep learning framework. Secondly, we established the mapping relationship between radiogenomic features and deep features and constructed prognostic signatures. Finally, we validated the prognostic value of radiogenomic signatures and explore their biological interpretation.

Results: We enrolled seven genomic datasets(n=2519) including our private cohort from ICGC-ARGO project, one radiogenomic (n = 233) and one radiomic (n = 535) datasets. Our genomic signatures can classify patients into high risk and low risk groups with significant differences. Similarly, radiogenomic signatures exhibited comparable performance in prognostic prediction. Multivariate analysis confirmed the independent prognostic value of radiogenomic signatures. Functional annotation revealed the association between radiogenomic features and CMS subtypes.

Conclusions: We established a general framework for translating molecular subtypes into low-cost radiogenomic biomarkers, providing more accurate prognosis prediction and treatment guidance.