Abstract

Speaker: Emma Rocheteau

Title: Deep Learning with Electronic Health Records

Abstract:

Deep Learning has the potential to change the future of healthcare. I am particularly interested in projects that bring different aspects of the Electronic Health Record together to make predictions such as length of stay and mortality. In this talk I will cover two of my current research projects. The first is a collaboration with Microsoft Research; we use transformers to predict length of stay and discharge location in the Intensive Care Unit (ICU). In the second we focus on the best way to use the diagnosis data in a mortality predict task (again in the ICU ). We form a graph of hospital patients based on diagnostic similarity using an autoencoder. We propose a “sequential graph neural network”; a graph neural network that uses data from neighbouring patients to co-attend over the time series hidden states. I will briefly discuss some preliminary results for both projects.

—————————————————

Speaker: Michal Rozenwald

Title: Neural Networks for 3D Chromatin folding

Abstract:

Recent technological development has enabled the generation of large amounts of biological data. One area which has seen rapid advances is in the field of chromatin folding. The development of the Hi-C method has unraveled many basic principles of nuclear organization including the subdivision of the genome into chromosome territories, chromatin compartments and Topologically Associating Domains (TADs). In addition, several studies have confirmed a correlation between 3D chromatin organization and a host of epigenetic features relevant to cellular activity and cell fate decisions. However, a full description of the functional interplay between 3D organization and cellular behavior remains elusive. Currently, my main research focus is on applying Machine Learning methods to analyze this 3D chromatin organization and its correlations with epigenetic and transcriptional behavior. In this talk, I will present our work at the lab of Dr. Mikhail Gelfand at HSE University, Moscow on predicting TAD characteristics using ChIP-seq epigenetic data on chromatin markers. We have presented a set of baselines including Linear regression models, Gradient Boosting Trees and Recurrent Neural Networks. Following that project, I am now working with Dr. Pietro Lio (Department of Computer Science and Technology) and Dr. Ernest Laue (Department of Biochemistry) as a visiting student at the University of Cambridge. In this project, we aim to combine population-level and single-cell Hi-C data and use Graph Neural Networks to gain new insights about genome organization and its links (if any) with transcriptional behavior during differentiation of mouse embryonic stem cells. Besides that, I’ll be happy to share some tips on applications for summer internships and preparation for the SWE interviews.

Bio:

Michal (Miki) Rozenwald is a research assistant at the Bioinformatics laboratory of Prof. Mikhail Gelfand at HSE University working on applying Neural Networks to 3D chromatin structure analysis. She finished her BSc in Computer Science and is an MSc in Data Analysis in Biology and Medicine at the National Research University Higher School of Economics (HSE) in Moscow. She is currently a visiting student at Cambridge University working with Prof. Pietro Lio and Prof. Ernest Laue on combining population and single-cell Hi-C methods to explore links between chromatin folding and epigenetics.Michal has interned at Google (Munich and Zurich), Facebook (London and California) working on various machine learning projects. She also interned at the bioinformatics lab of Dr. Noam Shomron at Tel-Aviv University, Israel.