Central nervous system (CNS) toxicity accounts for 25% of safety failures across drug discovery and development, yet it is rarely detected during standard GLP toxicology studies1. Seizure risk remains one of the most difficult CNS adverse effects to predict before human exposure because 2D cell-based assays and animal models cannot adequately measure human neural activity.
Seizures arise from instability in neural networks, driven by disruptions in the balance between excitatory and inhibitory signaling. These effects emerge at the level of interconnected systems, not isolated cells. As a result, reductionist in vitro models and species-dependent animal models often generate signals that fail to align with clinical outcomes.
Functional Neural Activity as a Predictor of Seizure Liability
The 28bio Endurance Study approaches this challenge differently by asking whether functional neural activity in CNS-3D Brain Organoids can serve as a direct predictor of seizure risk.
The study is a retrospective, non-interventional evaluation of 66 small molecules with documented human clinical outcomes from 120,551 patients and known positive controls. By anchoring the analysis in known human outcomes, the study is designed to assess predictive performance in a clinically relevant context. The drug set spans 30 seizure-inducing and 36 clinically safe drugs with diverse mechanisms of action, including ion channel modulators, neurotransmitter pathway agents, and compounds with paradoxical effects, reflecting the complexity of real-world drug development.
Rather than relying on indirect markers, the approach measures drug-induced changes in functional neural activity using human iPSC-derived cortical organoids. These patterns are translated into quantitative features describing oscillatory behavior, burst structure, and network synchronization. To complement the functional data, molecular features derived from chemical structure are incorporated using established fingerprinting methods. Together, these inputs form the basis of an AI neurotoxicology prediction model trained to classify seizure risk, by learning how drugs perturb neural systems and how those perturbations relate to clinical outcomes.
Results from the Endurance Study demonstrate CNS-3D Brain Organoids predict clinical seizure liability with 83% sensitivity and 89% specificity. This performance indicates strong discrimination between seizure-associated and clinically safe drugs, while maintaining a balance between detecting risk and avoiding unnecessary exclusion. In practice, this allows teams to deprioritize compounds with a high likelihood of seizure liability while reducing the risk of eliminating clinically viable candidates.
Beyond binary classification, the model produces continuous scores that stratify drugs by risk level. These scores tend to align with known mechanisms and with the clinical prevalence of seizure events, suggesting that the system captures biologically meaningful gradients rather than arbitrary thresholds.
Published comparisons indicate that CNS-3D Brain Organoids demonstrate 13x higher predictive performance than animal models and outperform 2D cell-based assays.
Seizure Liability as a Measurable Endpoint
Seizure liability is often treated as an emergent risk that cannot be fully predicted because compounds with similar profiles can diverge significantly in clinical outcomes, reflecting the complexity of neural systems.
The 28bio approach addresses this limitation. By incorporating functional human data, molecular features, and AI modeling, researchers can assess drug candidates earlier in development, deprioritizing compounds that destabilize neural networks ahead of in vivo studies and moving forward safe compounds with greater confidence.
For a field defined by late-stage surprises, the ability to measure dysfunction before it manifests clinically represents a meaningful advancement.
References
1Walker, et al. Drug discovery and development: Biomarkers of neurotoxicity and neurodegeneration. Experimental biology and Medicine 2018; 243: 1037-1045.