Wednesday, 19 December 2018 Guest lecture: Dr. Marinka Žitnik (Stanford University): Deep Learning for Network Medicine
UP FAMNIT and UP IAM
Departments of Information Sciences and Technologies
are pleased to invite everyone to the lecture
»Deep Learning for Network Medicine«
which will be held by Dr. Marinka Žitnik (Stanford University, USA)
on Thursday, 20 December, at 13:30 in FAMNIT-VP1.
About the lecturer:
Marinka Žitnik is a postdoctoral research fellow in Computer Science at Stanford University where she works with Jure Leskovec and collaborates with biomedical research departments around the world. She is also a Chan Zuckerberg Biohub postdoctoral researcher.
Her research investigates machine learning for biomedical sciences, focusing on large networks of interactions between biomedical entities -- e.g., proteins, drugs, diseases, and patients. She leverages these networks at the scale of billions of interactions among millions of entities and develops new methods blending machine learning with statistical methods and network science.
She uses her methods to answer burning scientific questions, such as how Darwinian evolution changes molecular networks, and how data-driven algorithms accelerate scientific discovery; and she uses the methods to solve high-impact problems, such as what drugs and combinations of drugs are safe for patients, what molecules will treat what diseases, and how newborns are transferred between hospitals and how these transfers influence outcomes.
About the lecture:
Networks pervade medical research and practice. The primary challenge is how to learn on biomedical networks that involve rich interactions, spanning from the molecular scale all the way to the societal scale encompassing all human interactions. However, prevailing deep learning algorithms are designed for data with a regular, grid-like structure and cannot exploit rich interactions, the essence of biomedical networks.
In this talk, I will discuss methods that learn how to embed nodes in rich biomedical networks as points in a low-dimensional embedding space, where the geometry of this space is optimized to reflect the structure of interactions between the nodes. These embeddings methods are at the technical core of Decagon, the first approach for predicting side effects of drug combinations, not merely of individual drugs. Decagon composes a massive network describing how proteins in our bodies interact with each other and how different drugs affect these proteins. Decagon's deep embedding method uses the network to identify patterns in how side effects arise based on how drugs target different proteins. Today, in many cases, it is unknown what side effects might arise from adding another drug to a patient's personal pharmacy, and Decagon has the potential to lead to more effective and safer healthcare.
