Effective connectivity (EC) is crucial for elucidating causal interactions between brain regions, providing valuable insights into cognitive function. Traditional methods infer EC indirectly through temporal predictions, but unmeasured confounding factors lead to temporal delays, resulting in spurious causal inferences. Compared to the indirect inference of EC using traditional methods, neurostimulation experiments directly infer EC, yet their invasive nature and ethical constraints limit their applicability in assessing whole-brain EC. To address this, we propose a data-theory-driven virtual neurostimulation (VNS) model for directly estimating EC between brain regions. This model constructs a large-scale brain network model as a surrogate for the brain, applying perturbations to specific brain regions based on the physiological mechanism of membrane potential polarization induced by current stimulation. Causal relationships between brain regions are then inferred directly by performing statistical analysis on the responses in blood oxygenation level dependent (BOLD) signals. The model’s accuracy and stability were validated on macaque and human medial temporal lobe (MTL) datasets with known ground-truth EC, demonstrating superior performance over baseline methods. Application to a disease specific dataset further highlights its potential as a personalized biomarker for neurodegenerative diseases, providing a novel pathway for early diagnosis in brain disorders.