P142 When artificial intelligence meets autoimmunity: benchmarking automated and expert-guided whole exome sequencing of lupus-causing genes

Abstract

Background/Aims Whole-exome sequencing (WES) is increasingly used to investigate patients with suspected monogenic forms of systemic lupus erythematosus and other systemic autoimmune rheumatic diseases (SARDs). With the increasing use of AI-assisted variant interpretation tools, there is a need to evaluate their real-world utility against expert-guided approaches. In a SARD cohort, we compared the performance of AI-assisted and expert-guided WES analysis strategies in identifying disease-relevant variants within lupus-causing genes. Methods We analysed WES data, generated with Illumina NovaSeq6000, from 120 patients with SARD disease and 20 healthy controls. Two analytic strategies were applied using the same platform, Emedgene: (1) an AI-based approach and (2) a semi-automated approach. The platform offers a fully automated phenotype-based variant analysis using AI, resulting in a list of prioritised variants for each case that are more likely to be causative, labelled as “most likely”. The “most likely” variants were extracted in a spreadsheet for further analysis (Table 1). For the expert-guided approach, the Emedgene platform was used for the initial part of the analysis, as it allows the application of a preset of filters followed by the extraction of data for additional analysis (Table 1). Results Before analysing data outside the platform, a total of 2,906 variants likely to be causative were identified with the Emedgene AI tool across all cases (mean 20.8 variants per case) compared to 8,391 variants (mean 59.93 variants per case) with the expert-guided method (Table 1). The majority of variants with both methods were loss-of-function variants of frameshift: 1,623 variants (77.4%) with the AI-guided variant analysis and 8,044 (95.9%) with the semi-automated approach. A comparable number of variants (78 vs 56) in lupus-causing genes were identified at the end of the analysis with both methods (Table 1). Notably, a known pathogenic variant in PEPD (c.819-1G>A, homozygous), previously identified in a patient with monogenic SARD, was detected using both approaches. Conclusion AI-assisted WES analysis offers efficiency and standardisation, yet expert review remains essential to ensure clinically meaningful interpretation in autoimmune genomics. Further studies in well-characterised monogenic cohorts will delineate the optimal integration of AI into autoimmune genomics. Disclosure A. Madenidou: Honoraria; Boehringer Ingelheim. Grants/research support; Janssen. I.N. Bruce: Consultancies; AstraZeneca, Eli Lilly, GlaxoSmithKline, Merck Serono and UCB. Honoraria; speaker for AstraZeneca, GlaxoSmithKline and UCB. Grants/research support; Genzyme/Sanofi, GlaxoSmithKline, Roche, Jansen and UCB. G.I. Rice: Grants/research support; Janssen.