Machine Learning, Data Topology, Biomedical Datasets - Krishnaswamy Lab at Yale School of Medicine

Transcript

• 00:04Let's say we take a
• 00:05drop of your blood that
• 00:07has, you know, thousands and
• 00:08thousands
• 00:09of cells.
• 00:10If you think about visualizing
• 00:11the cells, they actually exist
• 00:13in some kind of a
• 00:14hundred dimensional space. Right? But
• 00:16they don't take up all
• 00:17this space.
• 00:18So they take up narrow
• 00:20niches within this high dimensional
• 00:22space. It forms a low
• 00:23dimensional manifold. You can think
• 00:25of a two d piece
• 00:26of paper in three d
• 00:27space. It's forming a lower
• 00:29dimensional manifold, and it solves
• 00:31a lot of the challenges
• 00:32in single cell data if
• 00:33you tried to explicitly model
• 00:34that lower dimensional shape or
• 00:35manifold that the cells reside
• 00:37in.
• 00:41My lab focuses on developing
• 00:43novel machine learning, deep learning,
• 00:46and general AI techniques
• 00:48that are a lot of
• 00:49times either motivated by or
• 00:52specifically designed for,
• 00:54analysis and insight from biomedical
• 00:56data, from sequencing machines, from
• 00:59imaging,
• 01:00from other kinds of specialized
• 01:03technology that can measure things
• 01:05like brain activity, cellular activity.
• 01:07So there's been a lot
• 01:08of this kind of data
• 01:09generated in biology that
• 01:11needs to be analyzed that
• 01:13could give us new insights
• 01:14that can accelerate discovery.
• 01:19We develop
• 01:20a class of methods that
• 01:22we call sort of geometric
• 01:24topological
• 01:26deep learning methods. These are
• 01:28kinda like your normal neural
• 01:29networks that are used to
• 01:30build anything else, but we
• 01:32infuse them with a lot
• 01:33of very specific specific kinds
• 01:34of
• 01:35math that's very important in
• 01:37biology. So geometry is something
• 01:39that we use a lot.
• 01:40Geometry happens to be really
• 01:41important in all parts of
• 01:42biology, including
• 01:43shapes of molecules.
• 01:46A molecule has a certain
• 01:47kind of geometry, and if
• 01:48it didn't have that, it
• 01:50couldn't bind to another molecule
• 01:52or perform its proper function.
• 01:54And then topology is a
• 01:56way of numerically characterizing
• 01:59different geometries.
• 02:00We also
• 02:01infused a lot of them
• 02:03with dynamic systems capabilities. So
• 02:05these will be neural networks
• 02:06that can implicitly learn a
• 02:08dynamic system that can generate
• 02:09things like trajectories
• 02:11and forecast what happens to
• 02:13cells and molecules, for example.
• 02:15So these kind of,
• 02:18mathematically enriched neural networks are
• 02:20able to learn more scientifically
• 02:22precise information than the ones
• 02:24that are used out in
• 02:25social media or in other
• 02:27parts of the world.
• 02:31Some prominent areas where we
• 02:33have active grants with a
• 02:34lot of collaborators are cancer,
• 02:36including breast cancer and colorectal
• 02:38cancer,
• 02:39immunology,
• 02:41where we're developing large models
• 02:43that predict immunogenicity,
• 02:46neuroscience,
• 02:47where we're looking both at
• 02:49stem cell development into different
• 02:51neuronal lineages as well as
• 02:53the more neural activity work.
• 02:55Yale is actually unusually well
• 02:57suited for this because you
• 02:59can be at the intersection
• 03:00of all these fields very
• 03:01easily at Yale. So I
• 03:02really love, you know, math
• 03:03and computer science and these
• 03:05kinds of things, and I
• 03:06really wanted to,
• 03:08impact
• 03:08a scientific discovery, biology, medicine,
• 03:11these kinds of things.