AI For Health - April Meetup

AI For Health is a monthly gathering of students, researchers, and entrepreneurs working at the intersection of AI and Health/Life Sciences.

When: When: April 29th, Time: 4-6pm
Where: Mila auditorium (6650 Rue Saint Urbain, Montreal, QC H2S 3H1, 1st floor)

Agenda

3:30 PM - Doors open
4:00 PM - Talks starts
6:00 PM - Pizza
7:30 PM - Doors close

Speakers:

• Julien Cohen-Adad.
  Title: Connecting MRI physics and A.I. to advance neuroimaging
  Abstract: Magnetic Resonance Imaging (MRI) can have multiple flavours: T1, T2, proton density, fMRI, diffusion MRI, etc. These so-called "Quantitative MRI" techniques are useful for monitoring pathologies such as multiple sclerosis and Alzheimer's disease. However, quantitative MRI data require complex analysis pipelines that are often executed manually and hence suffer from poor reproducibility. Deep learning (DL) appears to be an ideal candidate to help automatize certain analysis tasks.
  Unfortunately, while dozens of DL papers applied to medical imaging are published every year, most methods have been validated in well-curated single-center datasets only. In the rare case where the code is publicly available, the algorithm usually fails when applied to other centers (a.k.a. Real life data!).
  This happens because images across different centers have slightly different features than those used to train the algorithm (contrast, resolution, etc.), combined with the fact that low amount of data and manual labels are available. Recent DL techniques such as domain adaptation have tackled this issue. However, these techniques are not well adapted to our situation because in MRI, image features not only varies between centers, but also across a large number of acquisition parameters (e.g., repetition time, flip angle).
  The purpose of this presentation is to sensitize the DL community to unmet needs in MRI analysis, and explore possible ways to leverage MRI physics to advance impactful DL applications in the medical domain.

• Mathieu Lavallée-Adam
  Title: Improving the sensitivity of mass spectrometry-based proteomics using supervised learning and protein-protein interaction data.
  Abstract: Mass spectrometry-based proteomics is widely used to identify proteins in complex biological samples. Current proteomics approaches generate hundreds of thousands of mass spectra, yet, on average, only 25% of the mass spectra acquired in a mass spectrometry experiment lead to protein identifications. Increasing protein identification sensitivity is critical to provide a comprehensive understanding of the underlying biology of complex samples. Protein-protein interactions contain information that can improve protein identification rate in mass spectrometry; information that is not used by most current algorithms identifying proteins from mass spectra. We therefore propose a novel supervised learning algorithm, named MS-PROTINI, that assesses the confidence of peptide and protein identifications using mass spectrometry data features and confidence scores along with protein-protein interaction data. Our approach is based on the hypothesis that the confidence of the identification of a given protein P in a sample increases when proteins interacting with P are also observed in the same sample. When benchmarked against the state-of-the-art Percolator algorithm, MS-PROTINI identified more spectra, more peptides and more proteins. We also show that MS-PROTINI improves protein sequence coverage over Percolator. Overall, our machine learning algorithm improves our ability to identify proteins in complex proteomes and will provide a more comprehensive understanding of the molecular mechanisms taking place in biological samples.