MOTIVATION: Although recent advanced sequencing technologies have improved the resolution of genomic and proteomic data to better characterize molecular phenotypes, efficient computational tools to analyze and interpret large-scale omic data are still needed. RESULTS: To address this, we have developed a network-based bioinformatic tool called Ollivier-Ricci curvature for omics (ORCO). ORCO incorporates omics data and a network describing biological relationships between the genes or proteins and computes Ollivier-Ricci curvature (ORC) values for individual interactions. ORC is an edge-based measure that assesses network robustness. It captures functional cooperation in gene signaling using a consistent information-passing measure, which can help investigators identify therapeutic targets and key regulatory modules in biological systems. ORC has identified novel insights in multiple cancer types using genomic data and in neurodevelopmental disorders using brain imaging data. This tool is applicable to any data that can be represented as a network. AVAILABILITY AND IMPLEMENTATION: ORCO is an open-source Python package and is publicly available on GitHub at https://github.com/aksimhal/ORC-Omics.

Directionally sensitive radiomic features including the histogram of oriented gradient (HOG) have been shown to provide objective and quantitative measures for predicting disease outcomes in multiple cancers. However, radiomic features are sensitive to imaging variabilities including acquisition differences, imaging artifacts and noise, making them impractical for using in the clinic to inform patient care. We treat the problem of extracting robust local directionality features by mapping via optimal transport a given local image patch to an iso-intense patch of its mean. We decompose the transport map into sub-work costs each transporting in different directions. To test our approach, we evaluated the ability of the proposed approach to quantify tumor heterogeneity from magnetic resonance imaging (MRI) scans of brain glioblastoma multiforme, computed tomography (CT) scans of head and neck squamous cell carcinoma as well as longitudinal CT scans in lung cancer patients treated with immunotherapy. By considering the entropy difference of the extracted local directionality within tumor regions, we found that patients with higher entropy in their images, had significantly worse overall survival for all three datasets, which indicates that tumors that have images exhibiting flows in many directions may be more malignant. This may seem to reflect high tumor histologic grade or disorganization. Furthermore, by comparing the changes in entropy longitudinally using two imaging time points, we found patients with reduction in entropy from baseline CT are associated with longer overall survival (hazard ratio = 1.95, 95% confidence interval of 1.4-2.8, p = 1.65e-5). The proposed method provides a robust, training free approach to quantify the local directionality contained in images.

BACKGROUND: Compositional systems, represented as parts of some whole, are ubiquitous. They encompass the abundances of proteins in a cell, the distribution of organisms in nature, and the stoichiometry of the most basic chemical reactions. Thus, a central goal is to understand how such processes emerge from the behaviors of their components and their pairwise interactions. Such a study, however, is challenging for two key reasons. Firstly, such systems are complex and depend, often stochastically, on their constituent parts. Secondly, the data lie on a simplex which influences their correlations. RESULTS: To resolve both of these issues, we provide a general and data-driven modeling tool for compositional systems called Compositional Maximum Entropy (CME). By integrating the prior geometric structure of compositions with sample-specific information, CME infers the underlying multivariate relationships between the constituent components. We provide two proofs of principle. First, we measure the relative abundances of different bacteria and infer how they interact. Second, we show that our method outperforms a common alternative for the extraction of gene-gene interactions in triple-negative breast cancer. CONCLUSIONS: CME provides novel and biologically-intuitive insights and is promising as a comprehensive quantitative framework for compositional data.

The remarkable growth of multi-platform genomic profiles has led to the challenge of multiomics data integration. In this study, we present a novel network-based multiomics clustering founded on the Wasserstein distance from optimal mass transport. This distance has many important geometric properties making it a suitable choice for application in machine learning and clustering. Our proposed method of aggregating multiomics and Wasserstein distance clustering (aWCluster) is applied to breast carcinoma as well as bladder carcinoma, colorectal adenocarcinoma, renal carcinoma, lung non-small cell adenocarcinoma, and endometrial carcinoma from The Cancer Genome Atlas project. Subtypes were characterized by the concordant effect of mRNA expression, DNA copy number alteration, and DNA methylation of genes and their neighbors in the interaction network. aWCluster successfully clusters all cancer types into classes with significantly different survival rates. Also, a gene ontology enrichment analysis of significant genes in the low survival subgroup of breast cancer leads to the well-known phenomenon of tumor hypoxia and the transcription factor ETS1 whose expression is induced by hypoxia. We believe aWCluster has the potential to discover novel subtypes and biomarkers by accentuating the genes that have concordant multiomics measurements in their interaction network, which are challenging to find without the network inference or with single omics analysis.

Accurate and robust segmentation of lung cancers from CT, even those located close to mediastinum, is needed to more accurately plan and deliver radiotherapy and to measure treatment response. Therefore, we developed a new cross-modality educed distillation (CMEDL) approach, using unpaired CT and MRI scans, whereby an informative teacher MRI network guides a student CT network to extract features that signal the difference between foreground and background. Our contribution eliminates two requirements of distillation methods: (i) paired image sets by using an image to image (I2I) translation and (ii) pre-training of the teacher network with a large training set by using concurrent training of all networks. Our framework uses an end-to-end trained unpaired I2I translation, teacher, and student segmentation networks. Architectural flexibility of our framework is demonstrated using 3 segmentation and 2 I2I networks. Networks were trained with 377 CT and 82 T2w MRI from different sets of patients, with independent validation (N = 209 tumors) and testing (N = 609 tumors) datasets. Network design, methods to combine MRI with CT information, distillation learning under informative (MRI to CT), weak (CT to MRI) and equal teacher (MRI to MRI), and ablation tests were performed. Accuracy was measured using Dice similarity (DSC), surface Dice (sDSC), and Hausdorff distance at the$95^{\textit {th}}$percentile (HD95). The CMEDL approach was significantly (p < 0.001) more accurate (DSC of 0.77 vs. 0.73) than non-CMEDL methods with an informative teacher for CT lung tumor, with a weak teacher (DSC of 0.84 vs. 0.81) for MRI lung tumor, and with equal teacher (DSC of 0.90 vs. 0.88) for MRI multi-organ segmentation. CMEDL also reduced inter-rater lung tumor segmentation variabilities.

MOTIVATION: Convolutional neural networks (CNNs) have achieved great success in the areas of image processing and computer vision, handling grid-structured inputs and efficiently capturing local dependencies through multiple levels of abstraction. However, a lack of interpretability remains a key barrier to the adoption of deep neural networks, particularly in predictive modeling of disease outcomes. Moreover, because biological array data are generally represented in a non-grid structured format, CNNs cannot be applied directly. RESULTS: To address these issues, we propose a novel method, called PathCNN, that constructs an interpretable CNN model on integrated multi-omics data using a newly defined pathway image. PathCNN showed promising predictive performance in differentiating between long-term survival (LTS) and non-LTS when applied to glioblastoma multiforme (GBM). The adoption of a visualization tool coupled with statistical analysis enabled the identification of plausible pathways associated with survival in GBM. In summary, PathCNN demonstrates that CNNs can be effectively applied to multi-omics data in an interpretable manner, resulting in promising predictive power while identifying key biological correlates of disease. AVAILABILITY AND IMPLEMENTATION: The source code is freely available at: https://github.com/mskspi/PathCNN.

We developed a new joint probabilistic segmentation and image distribution matching generative adversarial network (PSIGAN) for unsupervised domain adaptation (UDA) and multi-organ segmentation from magnetic resonance (MRI) images. Our UDA approach models the co-dependency between images and their segmentation as a joint probability distribution using a new structure discriminator. The structure discriminator computes structure of interest focused adversarial loss by combining the generated pseudo MRI with probabilistic segmentations produced by a simultaneously trained segmentation sub-network. The segmentation sub-network is trained using the pseudo MRI produced by the generator sub-network. This leads to a cyclical optimization of both the generator and segmentation sub-networks that are jointly trained as part of an end-to-end network. Extensive experiments and comparisons against multiple state-of-the-art methods were done on four different MRI sequences totalling 257 scans for generating multi-organ and tumor segmentation. The experiments included, (a) 20 T1-weighted (T1w) in-phase mdixon and (b) 20 T2-weighted (T2w) abdominal MRI for segmenting liver, spleen, left and right kidneys, (c) 162 T2-weighted fat suppressed head and neck MRI (T2wFS) for parotid gland segmentation, and (d) 75 T2w MRI for lung tumor segmentation. Our method achieved an overall average DSC of 0.87 on T1w and 0.90 on T2w for the abdominal organs, 0.82 on T2wFS for the parotid glands, and 0.77 on T2w MRI for lung tumors.

Volumetric lung tumor segmentation and accurate longitudinal tracking of tumor volume changes from computed tomography images are essential for monitoring tumor response to therapy. Hence, we developed two multiple resolution residually connected network (MRRN) formulations called incremental-MRRN and dense-MRRN. Our networks simultaneously combine features across multiple image resolution and feature levels through residual connections to detect and segment the lung tumors. We evaluated our method on a total of 1210 non-small cell (NSCLC) lung tumors and nodules from three data sets consisting of 377 tumors from the open-source Cancer Imaging Archive (TCIA), 304 advanced stage NSCLC treated with anti- PD-1 checkpoint immunotherapy from internal institution MSKCC data set, and 529 lung nodules from the Lung Image Database Consortium (LIDC). The algorithm was trained using 377 tumors from the TCIA data set and validated on the MSKCC and tested on LIDC data sets. The segmentation accuracy compared to expert delineations was evaluated by computing the dice similarity coefficient, Hausdorff distances, sensitivity, and precision metrics. Our best performing incremental-MRRN method produced the highest DSC of 0.74 ± 0.13 for TCIA, 0.75±0.12 for MSKCC, and 0.68±0.23 for the LIDC data sets. There was no significant difference in the estimations of volumetric tumor changes computed using the incremental-MRRN method compared with the expert segmentation. In summary, we have developed a multi-scale CNN approach for volumetrically segmenting lung tumors which enables accurate, automated identification of and serial measurement of tumor volumes in the lung.

More than half of all cancer patients receive radiotherapy in their treatment process. However, our understanding of abnormal transcriptional responses to radiation remains poor. In this study, we employ an extended definition of Ollivier-Ricci curvature based on LI-Wasserstein distance to investigate genes and biological processes associated with ionizing radiation (IR) and ultraviolet radiation (UV) exposure using a microarray dataset. Gene expression levels were modeled on a gene interaction topology downloaded from the Human Protein Reference Database (HPRD). This was performed for IR, UV, and mock datasets, separately. The difference curvature value between IR and mock graphs (also between UV and mock) for each gene was used as a metric to estimate the extent to which the gene responds to radiation. We found that in comparison of the top 200 genes identified from IR and UV graphs, about 20~30% genes were overlapping. Through gene ontology enrichment analysis, we found that the metabolic-related biological process was highly associated with both IR and UV radiation exposure.

Chemoresistance is a major obstacle to the successful treatment of many human cancer types. Increasing evidence has revealed that chemoresistance involves many genes and multiple complex biological mechanisms including cancer stem cells, drug efflux mechanism, autophagy and epithelial-mesenchymal transition. Many studies have been conducted to investigate the possible molecular mechanisms of chemoresistance. However, understanding of the biological mechanisms in chemoresistance still remains limited. We surveyed the literature on chemoresistance-related genes and pathways of multiple cancer types. We then used a curated pathway database to investigate significant chemoresistance-related biological pathways. In addition, to investigate the importance of chemoresistance-related markers in protein-protein interaction networks identified using the curated database, we used a gene-ranking algorithm designed based on a graph-based scoring function in our previous study. Our comprehensive survey and analysis provide a systems biology-based overview of the underlying mechanisms of chemoresistance.

BACKGROUND: Inference of gene regulatory networks (GRNs) from gene microarray expression data is of great interest and remains a challenging task in systems biology. Despite many efforts to develop efficient computational methods, the successful modeling of GRNs thus far has been quite limited. To tackle this problem, we propose a novel framework to reconstruct radio-responsive GRNs based on the graphical lasso algorithm. In our attempt to study radiosensitivity, we reviewed the literature and analyzed two publicly available gene microarray datasets. The graphical lasso algorithm was applied to expression measurements for genes commonly found to be significant in these different analyses. RESULTS: Assuming that a protein-protein interaction network obtained from a reliable pathway database is a gold-standard network, a comparison between the networks estimated by the graphical lasso algorithm and the gold-standard network was performed. Statistically significant p-values were achieved when comparing the gold-standard network with networks estimated from one microarray dataset and when comparing the networks estimated from two microarray datasets. CONCLUSION: Our results show the potential to identify new interactions between genes that are not present in a curated database and GRNs using microarray datasets via the graphical lasso algorithm.

Despite various efforts to develop new predictive models for early detection of tumor local failure in locally advanced non-small cell lung cancer (NSCLC), many patients still suffer from a high local failure rate after radiotherapy. Based on recent studies of biomarker proteins' role in predicting tumor response following radiotherapy, we hypothesize that incorporation of physical and biological factors with a suitable framework could improve the overall prediction. To this end, we propose a graphical Bayesian network framework for predicting local failure in lung cancer. The proposed approach was tested using a dataset of locally advanced NSCLC patients treated with radiotherapy. This dataset was collected prospectively, which consisted of physical variables and blood-based biomarkers. Our experimental results demonstrate that the proposed method can be used as an efficient method to develop predictive models of local failure in these patients and to interpret relationships among the different variables. The combined model of physical and biological factors outperformed individual physical and biological models, achieving an accuracy (acc) of 87.78%, Matthew's correlation coefficient (r) of 0.74, and Spearman's rank correlation coefficient (rs) of 0.75 on leave-one-out cross-validation analysis.

Patients undergoing thoracic radiation therapy can develop radiation pneumonitis (RP), a potentially fatal inflammation of the lungs. Support vector machines (SVMs), a statistical machine learning method, have recently been used to build binary-outcome RP prediction models with promising results. In this work, we (1) introduce a feature-ranking selection step to limit complexity in ensemble SVM models (2) show that ensembles of SVMs provide a statistically significant performance improvement in the area under the cross-validated receiver operating curve and (3) apply Platt's tuning to generate probability estimates from the component SVMs in order to augment clinical relevance.

Combining different machine learning models (decision fusion) has been shown to be an effective method for estimating the underlying physical mechanism by allowing the models to reinforce each other when consensus exists, or, conversely, negate each other when there is no consensus. To be effective, decision fusion requires that the different models provide some degree of complementary information. In this work, we fuse the results of four different machine learning models (Boosted Decision Trees, Neural Networks, Support Vector Machines, Self Organizing Maps) to predict the risk of lung pneumonitis in patients undergoing thoracic radiotherapy. Fusion was achieved by simple averaging of the 10-fold cross validated predictions for each patient from all four models. To reduce prediction dependence on the manner in which the data set was split, 10-fold cross-validation was repeated 100 times for random data splitting. The area under the receiver operating characteristics curve for the fused cross-validated results was 0.79, higher than the individual models and with (generally) lower variance. The fusion extracted three important features as the consensus among all four models in predicting radiation pneumonitis risk: chemotherapy prior to radiotherapy, equivalent Uniform Dose (EUD) for exponent a = 1.2 to 3, and female gender. The results show great promise for machine learning in radiotherapy outcomes modeling.

Since the early demonstration of the curative potential of radiation therapy for tumor sterilization, normal tissue toxicity continues to be dose limiting. Accurate prediction of patient¿s complication risk would allow personalization of treatment planning decisions. Nonlinear kernel methods can provide a robust framework for learning complex interactions between observed toxicities and treatment, anatomical, and patient-related variables. However, proper application of these powerful methods would require better understanding of a high-dimensional feature space that is spanned by all these variables. In this work, we investigate methods for visualization of this high-dimensional space and compare different approaches for extracting discriminant features. Our preliminary results demonstrate that principle component analysis is a valuable tool for visualizing high dimensional data and for determining proper kernel type. In addition, variable selection based on resampling methods within the logistic regression framework seemed to yield improved prediction performance compared to the recursive-feature elimination method.

Building surfaces from cross-section curves has wide applications including bio-medical modeling. Previous work in this area has mostly focused on connecting simple closed curves on parallel cross-sections. Here we consider the more general problem where input data may lie on non-parallel cross-sections and consist of curve networks that represent the segmentation of the underlying object by different material or tissue types (e.g., skin, muscle, bone, etc.) on each cross-section. The desired output is a surface network that models both the exterior surface and the internal partitioning of the object. We introduce an algorithm that is capable of handling curve networks of arbitrary shape and topology on cross-section planes with arbitrary orientations. Our algorithm is simple to implement and is guaranteed to produce a closed surface network that interpolates the curve network on each cross-section. Our method is demonstrated on both synthetic and bio-medical examples.

In radiotherapy treatment planning of cancer patients, the collection of multiple images of different and yet complementary information is rapidly becoming the norm. Beside CT data sets, PET and/or MRI or MRS images are also being used to aid in the definition of the target volume for treatment optimization. We are investigating methods to integrate available information for joint target registration and segmentation of multi-modality images as perceived by the human observer. Towards this goal, we are exploring multi-valued level set deformable models in conjunction with human perception models for simultaneous delineation of multi-modality images consisting of combinations of PET, CT, or MR datasets. Information from multimodality image sets is integrated based on a logical model to define the final target volume. The methods were demonstrated qualitatively on patient cases of lung cancer with PET/CT and a prostate patient case with CT and MR. We used a series of phantom data of CT, PET, and MR for quantification analysis. Phantom studies suggest 90% segmentation accuracy and less than 2% volume error when integrating all of the three modalities. This is compared with 74% accuracy and 4.4% volume error when using CT-based systems. These results indicate that this semi-automated multimodality-based definition of the biophysical target would provide a feasible and accurate framework for integrating complementary imaging information from different modalities and potentially a useful tool for optimizing of cancer patients radiotherapy plans.