RECOMB CCB

Abstract. Online assessment of tumor characteristics during surgery is important and has the potential to establish an intra-operative surgeon feedback mechanism. With the availability of such feedback, surgeons could decide to be more liberal or conservative regarding the resection of the tumor. While there are methods to perform metabolomics-based online tumor pathology prediction, their model complexity and, in turn, the predictive performance is limited by the small dataset sizes. Furthermore,  the information conveyed by the feedback provided on the tumor tissue could be improved both in terms of content and accuracy. In this study, we propose a metabolic pathway-informed deep learning model, PiDeeL, to perform survival analysis and pathology assessment based on metabolite concentrations. We show that incorporating pathway information into the model architecture substantially reduces parameter complexity and achieves better survival analysis and pathological classification performance. With these design decisions, we show that PiDeeL improves tumor pathology prediction performance of the state-of-the-art in terms of the Area Under the ROC Curve (AUC-ROC) by 3.38% and the Area Under the Precision-Recall Curve (AUC-PR) by 4.06%. Similarly, with respect to the time-dependent concordance index (c-index), we observe that PiDeeL achieves better survival analysis performance (improvement up to 4.3%) when compared to the state-of-the-art. Moreover, we show that importance analyses performed on input metabolite features as well as pathway-specific hidden-layer neurons of PiDeel provide insights into tumor metabolism. We foresee that the use of this model in the surgery room will help surgeons adjust the surgery plan on the fly and will result in better prognosis estimates tailored to surgical procedures. The code is released at https://github.com/ciceklab/PiDeeL. The data used in this study is released at https://zenodo.org/record/7228791.

Abstract. Single-cell sequencing technologies are providing unprecedented insights into the molecular biology of individual cells. More recently, multi-omic technologies have emerged which can simultaneously measure gene expression and the epigenomic state of the same cell, holding the promise to unlock our understanding of the epigenetic mechanisms of gene regulation. However, the sparsity and noisy nature of the data poses fundamental statistical challenges which hinder our ability to extract biological knowledge from these complex data sets. Here we propose SHARE-Topic, a Bayesian generative model of multi-omic single cell data which addresses these challenges from the point of view of topic models. SHARE-Topic identifies common patterns of co-variation between different ‘omic layers, providing interpretable explanations for the complexity of the data. Tested on joint ATAC and expression data, SHARE-Topic was able to provide low dimensional representations that recapitulate known biology, and to define in a principled way associations between genes and distal regulators in individual cells. We illustrate SHARE-Topic in a case study of B-cell lymphoma, studying the usage of alternative promoters in the regulation of the FOXP1 transcription factors.

Abstract. Emerging ultra-low coverage single-cell DNA sequencing (scDNA-seq) technologies have enabled high resolution evolutionary studies of copy number aberrations (CNAs) within tumors.  While these sequencing technologies are well suited for identifying CNAs due to the uniformity of sequencing coverage, the sparsity of coverage poses challenges for the study of single-nucleotide variants (SNVs).  In order to maximize the utility of increasingly available ultra-low coverage scDNA-seq data and obtain a comprehensive understanding of tumor evolution, it is important to also analyze the evolution of SNVs from the same set of tumor cells.

We present Phertilizer, a method to infer a clonal tree from ultra-low coverage scDNA-seq data of a tumor.

Based on a probabilistic model, our method recursively partitions the data by identifying key evolutionary events in the history of the tumor.  We demonstrate the performance of Phertilizer on simulated data as well as on two real datasets, finding that Phertilizer effectively utilizes the copy-number signal inherent in the data to more accurately uncover clonal structure and genotypes compared to previous methods.

Availability: https://github.com/elkebir-group/phertilizer 

Abstract. Drug discovery is a challenging task, characterized by a protracted period of time between initial development and market release, with a high rate of attrition at each stage. Computational virtual screening, powered by machine learning algorithms, has emerged as a promising approach for predicting therapeutic efficacy. However, the complex relationships between fea tures learned by these algorithms can be challenging to decipher. We have devised a neural network model for the prediction of drug sensitivity, which employs a biologically-informed visible neural network (VNN), enabling a greater level of interpretability. The trained model can be scrutinized to investigate the biological pathways that play a fundamental role in prediction, as well as the chemical properties of drugs that influence sensitivity. The model leverages multi-omics data obtained from diverse tumor tissue sources and molecular descriptors that encode drug properties. We have extended the model to predict drug synergy, resulting in favorable outcomes while retaining interpretability. Given the often unbalanced nature of publicly available drug screening datasets, our model demonstrates superior performance compared to state-of-the-art visible machine learning algorithms.

Abstract. There is a growing understanding that normal tissue samples used as controls in cancer studies do not represent fully healthy tissue but are, instead, intermediate between healthy and cancer. Several factors can contribute to the deviation of solid tumor control samples from healthy state including exposure of the control sample to the same environmental and genetic tumor-promoting factors as the tumor, changes due to tumor related immune response and other aspects of tumor microenvironment. Characterizing the relation between gene expression in control samples and tumor progression is fundamental for understanding regulatory roles of tumor environment and for devising cancer prognostic markers and treatment. We developed and validated TranNet, a computational approach to utilize gene expression in matched tumor and control samples to study the relation between the expression of genes in adjacent normal control and tumor progression. TranNet infers a sparse weighted bipartite graph that defines a transition map from gene expression in control samples to gene expression in tumor together with predictors (potential regulators) of this transition. To our knowledge, TranNet is the first computational method to infer such regulation. We applied TranNet to the The Cancer Genome Atlas (TCGA) data from several solid tumors. Our results demonstrate that many predictors identified by TranNet correspond to known marker genes associated with regulation by the tumor microenvironment and are enriched in G-protein coupled receptor signaling, cell-to-cell communication, immune processes, and cell adhesion. In addition, targets of inferred predictors are enriched in pathways related to tissue remodelling (including the Epithelial-Mesenchymal Transition (EMT)), immune response, and cell proliferation. This indicates that genes identified by TranNet as predictors are makers and potential facilitators of environment-related regulation of tumor progression. The results obtained by TranNet provide a proof of the principle that control samples can be used to identify genes involved in the environment-mediated regulation of tumor progression and offer new insights into the relationship between tumor, normal, and the tumor environment. In addition, the set of predictors identified by TranNet will provide a valuable resource for future investigations. TranNet is freely available at https://github.com/ncbi/TranNet.

Abstract. Motivation: New low-coverage single-cell DNA sequencing technologies enable the measurement of copy number profiles from thousands of individual cells within tumors. From this data, one can infer the evolutionary history of the tumor by modeling transformations of the genome via copy number aberrations. A widely used model to infer such copy number phylogenies is the copy number transformation (CNT) model in which copy number aberrations are represented by events that alter the number of copies of contiguous segments of the genome. While the CNT distance between copy number profiles, i.e. the minimum number of events needed to transform one profile to another, can be computed efficiently, current methods rely on heuristics to compute copy number phylogenies under the CNT model.

Results: We introduce the resurrecting copy number transformation (RCNT) model, a simplification of the CNT model that allows the resurrection of zero copy number regions. We derive a closed form expression for the RCNT distance between two copy number profiles and show that, unlike the CNT distance, the RCNT distance forms a metric. We leverage the closed-form expression for the RCNT distance and a characterization of copy number profiles to derive polynomial time algorithms for two

natural relaxations of the small parsimony problem on copy number profiles. While the resurrection allowed in the RCNT model is not biologically realistic, we show on both simulated and real datasets that the RCNT distance is a close approximation to the CNT distance. Using our polynomial time algorithm for the (relaxed) small parsimony problem, we develop an algorithm, Breaked, for solving the large parsimony problem on copy number profiles. Finally, we demonstrate that Breaked outperforms existing methods for inferring copy number phylogenies on both simulated and real data.

Abstract. Integration of multi-omics data holds great promise for understanding the complex biology of diseases, particularly Alzheimer’s, Parkinson’s, and cancer. However, the integration is challenging due to the high dimensionality and com-plexity of the data. Traditional machine learning methods are not well-suited for handling the complex relationships between different types of omics data. Many models were proposed that utilize graph-based learning models to extract hidden representations and network structures from different omics data to enhance can-cer prediction, patient categorization, etc. The existing graph neural network-based (GNN-based) multi-omics approaches for cancer subtype prediction have three shortcomings: (a) Do not consider all types of omics data, (b) Fail to determine the relative significance of the neighboring nodes (in this case, samples or patients) when it comes to downstream analyses, such as subtype classification, patient stratification, etc., and (c) Use the same approach for generating initial graphs for different omics data. To overcome these shortcomings, we present MOGAT, a novel multi-omics integration approach, leveraging a graph attention network (GAT) model that incorporates graph-based learning with an attention mechanism. MOGAT utilizes a multi-head attention mechanism that can more ef-ficiently extract information for a specific sample by assigning unique attention coefficients to its neighboring samples. To evaluate the performance of MOGAT, we explored its capability via a case study of predicting breast cancer subtypes. Our results showed that MOGAT performs better than the state-of-the-art multi-omics integration approaches. 

Abstract. Combination drug therapies are effective treatments for cancer. However, the genetic heterogeneity of the patients and exponentially large space of drug pairings pose significant challenges for finding the right combination for a specific patient. Current in silico prediction methods can be instrumental in reducing the vast number of candidate drug combinations. However, existing powerful methods are trained with cancer cell line gene expression data, which limits their applicability in clinical settings. While synergy measurements on cell lines models are available at large scale, patient-derived samples are too few to train a complex model. On the other hand, patient-specific single-drug response data are relatively more available. In this work, we propose a deep learning framework, Personalized Deep Synergy Predictor (PDSP), that enables us to use the patient-specific single drug response data for customizing patient drug synergy predictions. PDSP is first trained to learn synergy scores of drug pairs and their single drug presonses for a given cell line using drug structures and large scale cell line gene expression data. Then, the model is fine-tuned for patients with their patient gene expression data and associated single drug response measured on the patient ex vivo samples. In this study, we evaluate PDSP on data from three leukemia patients and observe that it improves the prediction accuracy by 27% compared to models trained on cancer cell line data. PDSP is built and available at https://github.com/hikuru/PDSP.