Exploring the Strengths and Limitations of AI-Driven Variant Prioritization Versus Manual Curation in Inborn Errors of Immunity

Laith Ibrahim Moushib^1,2,3Nerea Moreno-Ruiz⁴Andrea Martín-Nalda^2,5Jacques G. Rivière^2,5Blanca Angélica Urban Vargas^2,5Romina Dieli Crimi^2,5Janire Perurena-Prieto^2,5Aina Aguiló-Cucurull^2,5Elena Pérez-Estévez⁶Xavier Solanich⁷Pere Soler-Palacín^2,5Roger Colobran^1,2,8*Laura Batlle-Masó^2,5

• ¹Universitat Autonoma de Barcelona, Barcelona, Spain
• ²Vall d'Hebron Institut de Recerca, Barcelona, Spain
• ³Mustansiriyah University, Baghdad, Iraq
• ⁴Universitat Pompeu Fabra, Barcelona, Spain
• ⁵Hospital Universitari Vall d'Hebron, Barcelona, Spain
• ⁶Hospital Universitario Cruces, Barakaldo, Spain
• ⁷Hospital Universitari de Bellvitge Servicio de Medicina Interna, L'Hospitalet de Llobregat, Spain
• ⁸Vall d'Hebron University Hospital, Barcelona, Spain

Introduction: Next-generation sequencing (NGS) has transformed the genetic diagnosis of human diseases, yet many patients remain unsolved due to the complexity of variant interpretation. Manual curation of candidate variants is effective but time-consuming and requires specialized expertise. Artificial intelligence (AI)-driven platforms have emerged as scalable tools for variant prioritization, yet their performance compared with manual curation remains insufficiently evaluated. The aim of this study was to evaluate the performance of AI-driven platforms for variant prioritization in a cohort of patients with inborn errors of immunity (IEI) and to compare their strengths and limitations with manual curation. Methods: We analyzed 22 unsolved IEI cases that had previously undergone inconclusive NGS studies. Whole-genome sequencing was performed, and variant prioritization was carried out using two AI-driven platforms -AIMMARVEL and AION (Nostos Genomics)- and by manual curation. Selected variants were classified according to clinical relevance (very high, high, medium, or low), integrating both molecular and phenotypic evidence. Results: Across the cohort, AI platforms efficiently prioritized variants with clear pathogenic features, often reaching the same conclusions as manual curation but in a fraction of the time. One patient (5%) received a conclusive diagnosis (FAM111B), and four patients (18%) carried variants of high clinical relevance, including strong disease-causing candidates in CD247 and SH2B3. Additional medium-relevance variants were identified in 36% of cases, although evidence was insufficient for functional validation. Notably, concordance between AIMMARVEL and AION was limited, particularly for variants of uncertain significance (VUS), reflecting differences in algorithmic weighting of variant features versus clinical phenotype. Both platforms also highlighted potentially novel associations in RUNX1 and TRAF7, underscoring their capacity to extend beyond classical IEI genes. Discussion: Our results show that AI-driven tools are powerful for detecting clearly pathogenic variants and can markedly accelerate the diagnostic process. However, their strong reliance on curated databases, limited incorporation of phenotypic data, and challenges in handling VUS may reduce their effectiveness. Enhancing phenotype integration, expanding annotations (including non-coding regions), and incorporating up-to-date literature could improve their performance. Ultimately, AI tools should complement expert curation, with future models evolving toward integrative approaches that better capture the complexity of human disorders.