Extending Eye2Gene to Quantify the Phenotypic Diversity and Similarity of 63 Inherited Retinal Diseases using an Embedding Approach

Abstract

Purpose : Inherited Retinal Diseases (IRDs) are a genetically and phenotypically heterogeneous group of diseases that are a leading cause of blindness in children and working-age adults. The diversity of these conditions often makes identifying their precise genetic diagnosis a challenge. Here we assessed whether the Eye2Gene AI approach could be leveraged to identify groups and subgroups in a data-driven approach.

Methods : Eye2Gene is a deep-learning classifier trained on 181,565 multi-modal retinal scans to identify the causative gene for 63 of the most common IRDs. The penultimate layer of the Eye2Gene model produces a 768-dimensional vector which represents the class activations. Visualising and clustering of these activations allows data-driven evaluation of within-class diversity and between-class similarity. The Uniform Manifold Approximation and Projection (UMAP) dimensionality reduction algorithm was applied to the activations from the Fundus Autofluorescence imaging module of Eye2Gene, to obtain two-dimensional embeddings. The embeddings in UMAP space were then clustered using hierarchical clustering with Ward linkage, to produce gene grouping dendrograms.

Results : The Eye2Gene hierarchical clustering of 476 individuals, recapitulated, in a data-driven way, some of the broad gene classifications and known phenocopies which we report in Figure 1. Additionally, subphenotypes of PRPH2, BEST1 and ABCA4 were uncovered using this approach, with distinct PRPH2 subphenotypes discovered in UMAP space (Figure 2).

Conclusions : Eye2Gene’s data-driven approach is able to identify phenotypically similar genes, allowing for an effective exploration of complex phenotype variations and their corresponding genetic underpinnings. It provides an objective and quantifiable framework to examine phenotypic similarity between IRDs which can help in the interpretation of variants of unknown significance.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Resulting gene grouping dendrogram from the Eye2Gene embeddings. It recapitulates known phenotypic groupings: achromatopsia genes marked in red, Stargardt phenocopy genes in blue, Retinitis Pigmentosa genes in green, Leber Congenital Amaurosis genes in yellow, and ocular coloboma genes in purple.

 

This figure displays distinct clusters of PRPH2 subphenotypes in the Eye2Gene phenotypic space and shows the location of the associated genetic variants within the PRPH2 protein.