CIS - "Get to know your neighbors" Seminar series: Dr. Wouter Saelens - Deplancke Lab
![Thumbnail](http://memento.epfl.ch/image/24065/1440x810.jpg)
Event details
Date | 06.03.2023 |
Hour | 15:15 › 16:15 |
Speaker | Dr. Wouter Saelens |
Location | Online |
Category | Conferences - Seminars |
Event Language | English |
Title: “Mountains in our DNA: how combining probabilistic modeling with neural networks deepens our understanding of omics data and genetic variation“
Abstract: The move towards precision medicine will require a detailed and multiscale understanding of the various molecular layers of the human cell. While progress in miniaturization and sequencing has now enabled increasingly sensitive and high-resolution profiling of the cell, these have presented some unique machine-learning and statistical challenges that are far from resolved. It is clear that any molecular or cellular process is incredibly complex, with numerous non-linearities and interactions, which would therefore require (deep) neural networks to properly fit and predict the underlying biology. However, limited sample sizes and lack of interpretability often mean we have to rely on simplified, but statistically convenient, models that tend to trade accuracy with generalizability.
Here, I will show in two examples how these statistically-convenient models can easily hide the true biology. I first highlight how DNA accessibility data contains much more information than previously thought, and how this information can be extracted by combining the expressivity of neural networks and normalizing flows, with the statistical power of Bayesian inference. This approach not only deepens our understanding of cellular behavior but most importantly improves our predictive power for genetic variations that underlie diseases. Next, I show how a similar approach can be applied on the cellular level to study the multifactorial effects of genetic perturbations on the cellular phenotype. We propose that a modular, interpretable but expressive, framework is needed that can convert the various molecular and cellular features of a cell with a flexible probabilistic model.
Bio: Wouter Saelens is a postdoc in the lab of prof. Bart Deplancke, where he studies the interaction within and between cellular processes (differentiation, cell cycle, ) and the several molecular layers (transcriptome, proteome and chromatin accessibility). Molecular biologist by training, Wouter started combining his love for biology with machine learning during his PhD at VIB-Ghent University in Ghent, Belgium.
The CIS seminar will take place In hybrid mode: Room INF 328 and by Zoom (Link TBD)
Please connect to your zoom account using your "@epfl.ch" address, as this live event is only open to the EPFL community
Monday, March 6, 2023 from 3:15 to 4:15 pm
NB: Video recordings of the seminars will be made available on our website and published on our social media pages
Abstract: The move towards precision medicine will require a detailed and multiscale understanding of the various molecular layers of the human cell. While progress in miniaturization and sequencing has now enabled increasingly sensitive and high-resolution profiling of the cell, these have presented some unique machine-learning and statistical challenges that are far from resolved. It is clear that any molecular or cellular process is incredibly complex, with numerous non-linearities and interactions, which would therefore require (deep) neural networks to properly fit and predict the underlying biology. However, limited sample sizes and lack of interpretability often mean we have to rely on simplified, but statistically convenient, models that tend to trade accuracy with generalizability.
Here, I will show in two examples how these statistically-convenient models can easily hide the true biology. I first highlight how DNA accessibility data contains much more information than previously thought, and how this information can be extracted by combining the expressivity of neural networks and normalizing flows, with the statistical power of Bayesian inference. This approach not only deepens our understanding of cellular behavior but most importantly improves our predictive power for genetic variations that underlie diseases. Next, I show how a similar approach can be applied on the cellular level to study the multifactorial effects of genetic perturbations on the cellular phenotype. We propose that a modular, interpretable but expressive, framework is needed that can convert the various molecular and cellular features of a cell with a flexible probabilistic model.
Bio: Wouter Saelens is a postdoc in the lab of prof. Bart Deplancke, where he studies the interaction within and between cellular processes (differentiation, cell cycle, ) and the several molecular layers (transcriptome, proteome and chromatin accessibility). Molecular biologist by training, Wouter started combining his love for biology with machine learning during his PhD at VIB-Ghent University in Ghent, Belgium.
The CIS seminar will take place In hybrid mode: Room INF 328 and by Zoom (Link TBD)
Please connect to your zoom account using your "@epfl.ch" address, as this live event is only open to the EPFL community
Monday, March 6, 2023 from 3:15 to 4:15 pm
NB: Video recordings of the seminars will be made available on our website and published on our social media pages
Practical information
- Informed public
- Free
Organizer
- CIS