“Manifold learning using graphs and neural networks to uncover structure in big biomedical data”

Speaker: David van Dijk
Postdoc, Yale University

Talk summary: In this talk I will describe several algorithms that we developed that use graphs, graph signal processing, and deep neural networks to learn manifolds on a wide range of biomedical data. First, I will present a data denoising and imputation algorithm, called MAGIC, that is designed to recover missing values, and generally reveal the underlying structure, of single-cell RNA-sequencing data by diffusing on an underlying data graph. I use MAGIC to characterize the epithelial-to-mesenchymal transition, which is a cellular state transition associated with cancer metastasis. Next, I will present MELD, a method that uses graph filtering for integration of external data to a manifold, and PHATE, a data visualization method that uses a novel informational distance to emphasize both local and global structure in data. I show PHATE on single-cell, gut microbiome and Facebook network data. Finally, I will talk about SAUCIE, which is a neural network autoencoder framework that uses several new regularizations to allow for interpretable encodings that we use to simultaneously visualize, cluster, batch-correct and impute biomedical data of large patient cohorts.

Speaker bio: David completed his PhD at the University of Amsterdam (with Professor Jaap Kaandorp) and the Weizmann Institute of Science (with Professor Eran Segal) in Computational Biology. He used machine learning to understand how gene regulation - i.e. the activity of genes - is encoded in DNA sequence. Currently he’s an Associate Research Scientist in the departments of Genetics and Computer Science in the group of Prof. Smita Krishnaswamy at Yale University. His work focuses on developing new machine learning methods for big biomedical data, with a focus on applying graph signal processing and deep learning to single-cell and other high-throughput biological data. He developed tools that are widely used in the biomedical community, such as MAGIC, a single-cell data imputation method, PHATE, a dimensionality reduction and visualization method, and SAUCIE, a deep learning framework for integrated and scalable biomedical data analysis.

.