Deep Neural Network BAsed Prediction of Monoubiquifination Sites Using Physicochemical Properties of Protein Sequences / The Causal Relation Between Somatic Genome Alteration and Tumor Immune Micro-Environment / Design Ideas for a Learning EMR
Rathnam Abstract: Ubiquitin is arguable one of the most important molecules involved in post-translational modifications as it is present in all eukaryotic cells and plays a key role in mediating a wide assortment of biological processes, such as cell cycle regulation, endocytosis of cellular proteins, and transcriptional regulation. Due to the highly dynamic and reversible nature of ubiquitination, it is often labor-intensive, expensive, and time-consuming to determine ubiquitination sites in vivo. This created a serious need for more time and cost efficient methods for ubiquitination site prediction, which, in turn, gave rise to the development of various in silico methods. These methods use machine learning techniques in conjunction with experimentally obtained protein sequence and ubiquitination data to predict ubiquitination sites. We describe the use of deep neural networks, which have not been applied yet to this domain, to predict monoubiquitination sites from protein physicochemical properties and compare its performance to existing machine learning-based prediction methods.
Chen Abstract: The immune system may have a protective role in tumor development, but may also simultaneously function to promote or select tumor variants with reduced immunogenicity. Tumors are heterogeneous groups of cells e.g., cancerous cells, stem cells, stromal cells and a wide range of immune cells. These heterogeneous populations secrete multiple signals which may hinder or promote the hallmarks of cancer required for tumor growth, development and progression. Cytokines expressed by tumor cells, immune cells, and other noncancerous cell types in and surrounding tumor tissue can regulate the tumor-promoting and -suppressing processes. In this study we systematically studied the causal relation between tumor genome alterations and its immune micro-environment among a large number of tumors across multiple cancer types, which could help with identifying treatment targets and pathways regarding the genome alterations of individual patients. Expression of signature genes and deconvolution algorithms were used to estimate the infiltration of immune cells in individual tumors. High accuracy was observed with prediction of tumor immune infiltration and immune gene expression pattern using tumor somatic gene alterations. Thus indicating causal structures between the tumor and immune microenvironment.
Calzoni Abstract: To reduce the risk of cognitive overload associated with large amounts of data in Electronic Medical Records (EMRs), we are designing a Learning EMR for use in Intensive Care Units (ICUs) that is able to draw a physician's attention to the right patient data at the right time.
To facilitate the development of an EMR user interface that prioritizes the presentation of high-value data, it is necessary to acquire a thorough understanding of the information practices of ICU providers. Most prior work in the literature focuses on investigating ICU clinicians’ data needs, reasoning strategies, and use of information sources. Our review of relevant literature provided us with useful insights on design aspects that we found applicable to the design of a Learning EMR. We identified four main themes, which we explored by designing a series of wireframes:
1. EMRs should convey and summarize clinical information effectively;
2. EMRs should highlight changes in clinical outcomes;
3. EMRs should facilitate comparisons across patients;
4. EMRs should provide support for analytical reasoning.
To inform the design of our Learning EMR user interface, we intend to combine observations with design activities. We will shadow ICU physicians and collect observational data on their information seeking strategies and interactions with the EMR. Interviews conducted outside of the care setting will help us characterize our observations and organize them into general process models. These models and our review of information visualization literature will inform the creation of a new series of wireframes, exploring possible ways to highlight high-value data in our Learning EMR. We will then use techniques inspired by Participatory Design to actively involve ICU physicians in the design process. In a focus group session, clinicians will be asked to provide feedback about the Learning EMR concept, to generate design ideas for the EMR user interface, and to provide feedback on our wireframes.
Study findings will help us gain a deeper understanding of ICU clinicians’ information needs and practices, and will inform future design choices for our Learning EMR.