Using Machine Learning to Assess the Pathogenicity of Small In-frame Indels in Inherited Retinal Disease

Abstract

Purpose : Interpretation of pathogenicity for in-frame indels variants is challenging compared to other types of genetic variants. Our goal was to systematically evaluate the performance of the latest in silico machine learning (ML) predictive tools and assess the mutation load in whole genome sequencing (WGS) data for in-frame indel variants in patients affected with inherited retinal diseases (IRDs).

Methods : The performance of four in silico ML pathogenicity prediction tools (CADD, FATHMM-indel, VEST4, and MetaRNN-indel) was benchmarked using a dataset of 3964 in-frame indel variants from ClinVar and a Diagnosing Developmental Disorders (DDD) study deposited in DECIPHER. A subset of this dataset was extracted containing variants across known IRDs genes obtained from the Retinal Information Network, resulting in 220 IRD variants. Tool performance was compared using the area under the receiver operating characteristic curve (AUROC) and precision-recall curve (AUC-PR). For the top-performing model, we defined the upper and lower thresholds to classify variants as likely pathogenic (LP) and likely benign (LB) among variants with known pathogenicity. The best performing model was applied to the in-house WGS data of 1013 IRDs patients.

Results : MetaRNN-indel performed the best with an AUROC of 0.942 and AUC-PR of 0.936 on the complete dataset, and an AUROC of 0.938 and AUC-PR of 0.972 on the IRD-related dataset. In our benchmark dataset, 95% of true benign labels were correctly classified at a MetaRNN-indel score ≤ 0.156, and 95% of true pathogenic labels were correctly classified at a MetaRNN-indel score ≥ 0.66. MetaRNN-indel then reclassified 86 (54.4%) of the 158 in-frame indels on IRD genes from the 1013 unsolved patients with IRDs. Among them, 60 (37%) variants were annotated as LB and 26 (16.5%) were annotated as LP. One patient with a variant in RP2 was confirmed to have X-linked Retinitis Pigmentosa based on phenotype, clinical diagnosis, and segregation analysis.

Conclusions : Our results show that the pathogenicity of in-frame indels can be reliably predicted with existing in silico ML tools. We successfully used MetaRNN-indel with re-calculated thresholds to identify novel pathogenic mutations in unsolved IRD patients, although the frequency of such mutations is low. Confirmation for other patients with candidate variants and other insights of the study will be reported subsequently.

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