Nazlee Zebardast, MD, MSc, in conversation with David Hutton, Managing Editor of Ophthalmology Times®, on research using machine learning techniques to identify image-based, especially OCT-based, glaucoma phenotypes or structural phenotypes We talked about At this year’s ARVO conference.
video transcript
Editor’s Note: This transcript has been edited for clarity.
David Hutton:
This is David Hutton from Ophthalmology times. The Association for Research in Vision and Ophthalmology recently held its annual meeting in New Orleans. At the conference, Dr. Nasley Zebadast announced “Genome-wide discovery by feature space mapping of deep learning-derived clinical OCT phenotypes to the UK Biobank”. Thank you very much for joining us today. Please tell us about the ARVO presentation.
Nazlee Zebardast, MD, MSc:
Thank you so much for having me as David and for giving me the opportunity to talk about my work. Therefore, we all know that glaucoma is the leading cause of irreversible vision loss worldwide. And that primary open-angle glaucoma is a complex and heterogeneous disease. It has long been recognized, including by Dr. Steven Grant, that there are structural and functional variations in glaucoma.
We also know that glaucoma is highly heritable, with over 120 genetic risk loci identified to date. However, these genetic variants as we know them explain only a limited part of the heritability of the disease. Clearly, large gaps remain. The question is how to fill this gap, and there is something called the endo phenotype. It is a quantitative trait that provides a range of risks and increases the ability to detect genetic variants.
Recently, however, machine learning approaches have emerged to make sense of unstructured, high-dimensional data. What we did in this study was to identify the image-based, specifically his OCT-based or structural phenotypes of glaucoma using unsupervised machine learning methods.
The problem is that not all available datasets contain all the information we need. These include genetic data, information about phenotypes such as glaucoma, medical conditions, and genetic information. What we did in this study was use the large clinical data sets available at Mass Eye and Ear. This was her over 18,000 high-quality OCTs from her over 8,000 patients with at least one diagnostic code for glaucoma. And those he uses in two machine learning approaches. One is the unsupervised autoencoder approach. The other is the contrastive learning approach.
We will use these to discover OCT-based phenotypes, specifically for two layers: the ganglion cell complex and the retinal nerve fiber layer. After training these machine learning models, we transferred the trained models to the UK Biobank. There is an OCT of him in the UK Biobank, but this is a population-based study. So we have data from people who don’t have the disease. In most cases, a minority of people have glaucoma. However, we use 80,000 images of her from 40,000 people with genetic information available in the UK Biobank, not all of whom had glaucoma. I found the same phenotype in the UK biobank.
Next, we perform gene-wide discovery using the same phenotypes found in the UK Biobank. We identified a significant number of variants and a significant number of loci. What is really interesting is that we have found that it overlaps very well with what is already known about glaucoma and its phenotype. We found variants already associated with size and even RNFL thickness. We also found genes associated with these different traits.
However, when we examined the expression level data from the community, we also found new genes and new loci. We found that they are highly expressed in retinal ganglion cells. So many new hypotheses have been generated, and I think we have taken a very exciting approach here, even though there is a lot of work to be done.
David Hutton:
What are the next steps for this research?
Nazlee Zebardast, MD, MSc:
Therefore, the next step here is to truly understand what the functional correlations of genes with the discovered loci are, and what are the potential causative genes. It means to perform further genetic analysis such as analysis, pathway analysis. There are many things we can do to really narrow down all these new variants and loci that we discovered with our approach, but which of them are meaningful and truly disease-related?
And you know, whenever you use machine learning there’s a black box where people don’t fully understand what’s going on. It is about being able to properly quantify and how they are clinically relevant. For example, one pattern is associated with visual field loss, various types of glaucoma, intraocular pressure, and need for eye surgery. So we already have all the data from Mount Sinai. We also obtain a lot of phenotypic data from the UK Biobank and the Whole Body Society.
So expressing our patterns in depth will be one of the next major steps.
