137P Benchmarking in silico tools and a patient-aware machine learning model for pathogenic missense variant prediction on real-world CGP data

Abstract

Results: For IDH prediction, uni-v2 performed best on cross-validation within TCGA, achieving an AUROC (Area Under the Receiver Operating Characteristic curve) of 0.953 and a PRAUC (Area Under the Precision-Recall curve) of 0.960.For external validation on UCL, TITAN led with 0.949 AUROC [CI 0.939-0.957]and 0.965 PRAUC [CI 0.0.957-0.972].For subtyping, uni-v2 again performed best in cross-validation, achieving an AUROCs of 0.947 and a PRAUC of 0.890 on the TCGA dataset.However, in external validation on UCL, both TITAN and uni-v2 achieved similar results (TITAN: AUROC 0.945 [CI 0.937-0.953],PRAUC 0.902 [CI 0.886-0.916],uni-v2: AUROC 0.945 [CI 0.936-0.953],PRAUC 0.903 ).Conclusions: VO models generally outperformed VL, with uni-v2 and TITAN leading their categories.TITAN outperformed its respective patch-level counterpart conch-v1.5,suggesting superior slide-level integration.Findings support brain tumor model development and cross-institutional generalizability for diagnostic workflows in resource-limited settings, potentially improving access to precision neuro-oncology.

Ähnliche Arbeiten