Integrative single-cell and spatial transcriptomics combined with machine learning to discover complement-associated biomarkers in temporal lobe epilepsy with hippocampal sclerosis.

Temporal lobe epilepsy with hippocampal sclerosis (TLE-HS) poses significant challenges in therapeutic management. While studies have demonstrated seizure-induced alterations in peripheral immune molecules, the complement system, a central component of immune function, remains insufficiently characterized at single-cell resolution and spatial distribution in TLE-HS. This study aimed to comprehensively investigate the spatiotemporal dynamics of complement system components and their clinical implications in patients with TLE-HS. We first identified patterns of complement activity changes in epilepsy using bulk RNA sequencing. Then, we employed a TLE-HS mouse model for single-cell RNA sequencing and S1000 high-resolution spatial transcriptomics. We performed integrative bioinformatic analyses on single-cell data to quantify complement-system activity and define microglial heterogeneity, leading to the identification of complement-associated microglial subpopulations. Spatial transcriptomic data then validated the anatomical localization of these identified subpopulations. Finally, we developed and evaluated two machine-learning models based on complement-related gene signatures. Complement activity was elevated in epilepsy. The levels were higher in hippocampal sclerosis (HS) tissue than in normal hippocampus. They were also increased in mesial temporal lobe epilepsy with hippocampal sclerosis compared with mesial temporal lobe epilepsy without hippocampal sclerosis, and in patients with high seizure frequency (HSF) compared with those with low seizure frequency (LSF). Complement-related signatures were further associated with antiseizure medication response. Candidate biomarkers including IRF2, GNB2, EHD1, CTSB, and CFH were identified using statistical modeling and machine learning. Among multiple classifiers, the support vector machine model showed the best predictive performance, and SHAP analyses indicated distinct contribution directions for these candidate genes. In the kainic acid (KA) mouse model, single-cell analyses showed that complement activity was upregulated across cell types in HS. Microglia exhibited the highest complement activity. Re-clustering and trajectory inference defined HS-associated microglial subpopulations that were enriched in terminal differentiation states. hdWGCNA together with differential expression highlighted Ctsb, C1qa, and Fcer1g as core complement-linked genes. Using Ctsb-defined microglial states, eight diagnostic biomarkers were selected, and a multi-layer perceptron model achieved superior classification accuracy in epilepsy diagnosis. Finally, cell-cell communication and spatial transcriptomics consistently implicated an Spp1-related signaling axis associated with Ctsbhigh microglia in hippocampal sclerosis regions. The resulting diagnostic and response-prediction models were deployed as exploratory web-based research tools pending independent validation. Our study systematically characterized the complement system in TLE-HS by integrating multi-level omics data, including bulk RNA sequencing, single-cell sequencing, and spatial transcriptomics. We revealed the potential value of complement system gene signatures in clinical diagnosis and personalized treatment of epilepsy.