Document resource
Introduction Brain arteriovenous malformations (bAVMs) remain a significant cause of hemorrhagic stroke in young adults. Current anatomical grading systems fail to capture the molecular heterogeneity driving individual AVM behavior. Somatic KRAS/BRAF mutations are identified in 55-87% of sporadic bAVMs, yet no framework integrates molecular profiling with hemodynamic risk assessment for personalized treatment. We present initial results from a novel pipeline integrating single-cell multi-omics profiling of bAVM tissues and liquid biopsies with AI-driven computational fluid dynamics (CFD) modeling for individualized rupture risk prediction and genotype-targeted therapy selection.Methods bAVM tissue was obtained from 18 patients undergoing surgical resection (n=14) or endoluminal biopsy (n=4), with multi-region sampling of the nidus, arterialized veins, feeding arteries, and perinidal parenchyma. Concurrent liquid biopsies from the efferent draining vein were analyzed for cfDNA and cfRNA. Single-cell RNA sequencing (scRNA-seq), scATAC-seq, and whole-exome sequencing (WES) were performed on 126,450 cells. Patient-specific CFD models were generated from 4D Flow MRI and quantitative DSA data. A gradient-boosted ensemble machine learning model integrated multi-omics signatures with hemodynamic parameters for rupture risk stratification.Results scRNA-seq revealed 14 distinct cell clusters, including 4 novel endothelial subtypes with differential MAPK-ERK activation, pro-inflammatory macrophages with elevated MMP-9/IL-6, and pericytes with deficient PDGFB signaling. scATAC-seq identified 847 differentially accessible chromatin regions with aberrant transcription factor motifs (ETS, SOX, KLF) regulating pathological angiogenesis. Somatic KRAS mutations (G12D/G12V) were detected in 12/18 patients (66.7%), enriched in endothelial cells (VAF 1.2-4.8%). Efferent vein liquid biopsy detected concordant KRAS mutations in 9/12 positive patients (75% sensitivity) via ddPCR, with cfRNA showing VEGFA and ANGPT2 upregulation. The integrated AI model (42 molecular and 8 hemodynamic features) achieved an AUC of 0.89 (95% CI: 0.82-0.95), outperforming hemodynamics-only (AUC 0.74) and morphology-only (AUC 0.68) models. SHAP analysis identified wall shear stress gradient, KRAS VAF, MMP-9 expression, and endothelial inflammatory gene score as top predictive features. In vitro drug screening on KRAS G12D iPSC-endothelial cells demonstrated significant ERK phosphorylation reduction with trametinib (IC 50=12.3 nM) and sotorasib (IC50=38.7 nM).Conclusions This study provides the first integrated single-cell multi-omics and AI-hemodynamic framework for personalized bAVM management. Our results demonstrate that bAVMs harbor significant cellular heterogeneity resolvable by single-cell profiling, efferent vein liquid biopsy can detect actionable somatic mutations non-invasively, integrated AI modeling significantly improves rupture prediction, and patient-derived models enable genotype-targeted drug screening. These findings support a paradigm shift toward precision medicine in bAVM management.Disclosures O. Mansour: 1; C; Alexandria University Research inistitute. A. Hasawy: None. N. Nasr: None. A. Gomaa: None. M. Ashraf: None.Abstract O-038 Figure 1Initial results: single-cell landscape, mutation profiling, AI rupture prediction, and drug response. (A) UMAP visualization of 126,450 single cells from 18 bAVM patients, revealing 14 distinct clusters including 4 novel endothelial subtypes (EC-1 through EC-4) with differential MAPK-ERK activation, pro-inflammatory macrophages (Mac-Infl), and PDGFB-deficient pericytes. (B) Somatic mutation frequencies across the cohort: KRAS mutations detected in 66.7% of patients by tissue WES; efferent vein liquid biopsy (LB) achieved 75% concordance for KRAS detection. (C) Receiver operating characteristic (ROC) curves comparing integrated multi-omics +hemodynamic AI model (AUC = 0.89) with hemodynamics-only (AUC = 0.74) and morphology-only (AUC = 0.68) models. (D) Dose-response curves for trametinib (ICsn = 12.3 nM) and sotorasib (IC;n = 38.7 nM) in patient-derived KRAS G12D iPSC-endothelial cells measuring ERK phosphorylation inhibition. Error bars = SEM. * p < 0.05, ** p < 0.01.