Machine learning enables high-throughput, low-replicate screening for novel anti-seizure targets and compounds using combined movement and calcium fluorescence in larval zebrafish.

Identifying new anti-seizure medications (ASMs) is difficult due to limitations in animal-based assays. Zebrafish (Danio rerio) serve as a model for chemical and genetic seizures, but current methods for detecting anti-seizure responses are limited by incomplete detection of anti-seizure responses (locomotor assays) or low-throughput (electrophysiology, fluorescence microscopy). To overcome these challenges, we developed a novel high-throughput method using combined locomotor and calcium fluorescence data from unrestrained larval zebrafish in a 96-well plate reader. Custom software tracked fish movement and fluorescence changes (deltaF/F0) from high-speed time-series, and logistic classifiers trained with elastic net regression distinguished seizure-like activity in response to the GABA(A) receptor antagonist pentylenetetrazole (PTZ). A classifier using combined data ("PTZ M + F"; AUC-ROC: 0.98; F1: 0.912) outperformed movement-only ("PTZ M"; AUC-ROC: 0.9) and fluorescence-only classifiers ("PTZ F"; AUC-ROC 0.96). Seizure-like event rate increased in proportion to PTZ concentration, and was suppressed by valproic acid (VPA). Meanwhile, TGB selectively reduced events defined by the "PTZ M + F″ classifier, paralleling previous reports that TGB reduces electrographic but not locomotor seizures. Using bootstrap simulation, we calculated statistical power and demonstrated reliable detection of ASM effects with as few as N = 4 replicates. In a test screen, 4 out of 5 ASMs were detected. This high-throughput approach combines previously orthogonal assays for zebrafish ASM screening.