Explainable Machine Learning and Epigenomic Profiling Decipher the Topological Determinants of Lentiviral Integration and Longitudinal Persistence in SCID-X1 Gene Therapy
- 1Department of Cellular and Molecular Biology, Faculty of Basic Sciences, Islamic Azad University, Ardabil Branch, Ardabil, 56131, Iran
Abstract
Lentiviral vectors (LVs) are clinically established for SCID-X1 gene therapy, yet the quantitative epigenomic and spatial determinants governing integration targeting and long-term clonal persistence across chromosomes remain incompletely characterized. Using clinical multi-omics datasets from SCID-X1 trials, we curated 274,959 unique clinical integration sites (VIS) across 276,839 clonal records from 10 patients (hg38). Balanced against length-weighted genomic controls (total N = 549,918), five epigenomic and topological features were modeled. We evaluated Logistic Regression, XGBoost, and a Deep Genomic ResNet using inner 5-fold chromosome-grouped cross-validation and evaluation on held-out test chromosomes (chr19-22, chrX). Statistical uncertainty was quantified via 1-Mb block-bootstrap (5,000 replicates) and 1-Mb block permutations (10,000 replicates). On held-out chromosomes, Deep ResNet achieved an ROC-AUC of 0.7857 [95% CI: 0.7672-0.8029] and PR-AUC of 0.8027 [95% CI: 0.7640-0.8337] over the 0.5755 prevalence baseline, outperforming linear baselines. 1-Mb block permutations confirmed significant proximity to scATAC-seq peaks (Cliff's {delta}=-0.3364, P_perm<0.0001) and TSSs ({delta}=-0.2446, P_perm<0.0001), alongside 82.71% gene body overlap (OR = 3.47). Targeted analysis showed 99.91% of VIS reside distal (>100 kb) from proto-oncogenes without proximal enrichment (OR = 1.05, P = 0.617). Furthermore, longitudinal tracking identified 79,587 persistent clones ([≥]2 time points), which exhibited a significant 31% depletion near proto-oncogenes (OR = 0.69, P = 0.014). Compact topological and epigenomic features robustly predict lentiviral integration without spatial data leakage. The long-term depletion of persistent clones near oncogenes confirms the insertional safety of SIN lentiviral gene therapy.
Related
- Longitudinal persistence of EEG brain fingerprints:phase and complex coherency geometry across sessions andfunctional states
- Longitudinal trajectories and persistence of frequent emergency department use: A two-year individual-level cohort study
- An Enhanced Pipeline for Multi-Omic Integration Based on Topological Data Analysis
- Machine learning prediction of 1-year thrombosis risk in membranous nephropathy: integration of renal pathology and early longitudinal trajectories
- Genome-wide association study reveals putative genetic determinants of longitudinal immune responses to avian influenza and Newcastle disease in chickens
BibTeX
@misc{saber2026explainable,
title = {{Explainable Machine Learning and Epigenomic Profiling Decipher the Topological Determinants of Lentiviral Integration and Longitudinal Persistence in SCID-X1 Gene Therapy}},
author = {H. Saber and Z. Mousavipour},
howpublished = {bioRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.05.749598},
url = {https://www.biorxiv.org/content/10.64898/2026.09.05.749598v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - Explainable Machine Learning and Epigenomic Profiling Decipher the Topological Determinants of Lentiviral Integration and Longitudinal Persistence in SCID-X1 Gene Therapy
AU - H. Saber
AU - Z. Mousavipour
PY - 2026
DA - 2026/09/05
PB - bioRxiv
DO - 10.64898/2026.09.05.749598
UR - https://www.biorxiv.org/content/10.64898/2026.09.05.749598v1
AB - Lentiviral vectors (LVs) are clinically established for SCID-X1 gene therapy, yet the quantitative epigenomic and spatial determinants governing integration targeting and long-term clonal persistence across chromosomes remain incompletely characterized. Using clinical multi-omics datasets from SCID-X1 trials, we curated 274,959 unique clinical integration sites (VIS) across 276,839 clonal records from 10 patients (hg38). Balanced against length-weighted genomic controls (total N = 549,918), five epigenomic and topological features were modeled. We evaluated Logistic Regression, XGBoost, and a Deep Genomic ResNet using inner 5-fold chromosome-grouped cross-validation and evaluation on held-out test chromosomes (chr19-22, chrX). Statistical uncertainty was quantified via 1-Mb block-bootstrap (5,000 replicates) and 1-Mb block permutations (10,000 replicates). On held-out chromosomes, Deep ResNet achieved an ROC-AUC of 0.7857 [95% CI: 0.7672-0.8029] and PR-AUC of 0.8027 [95% CI: 0.7640-0.8337] over the 0.5755 prevalence baseline, outperforming linear baselines. 1-Mb block permutations confirmed significant proximity to scATAC-seq peaks (Cliff's {delta}=-0.3364, P_perm<0.0001) and TSSs ({delta}=-0.2446, P_perm<0.0001), alongside 82.71% gene body overlap (OR = 3.47). Targeted analysis showed 99.91% of VIS reside distal (>100 kb) from proto-oncogenes without proximal enrichment (OR = 1.05, P = 0.617). Furthermore, longitudinal tracking identified 79,587 persistent clones ([≥]2 time points), which exhibited a significant 31% depletion near proto-oncogenes (OR = 0.69, P = 0.014). Compact topological and epigenomic features robustly predict lentiviral integration without spatial data leakage. The long-term depletion of persistent clones near oncogenes confirms the insertional safety of SIN lentiviral gene therapy.
ER -