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, more efficacious anti-seizure medications (ASMs) is challenging, partly due to limitations in animal-based assays. Zebrafish (Danio rerio) can serve as a model of chemical and genetic seizures, but methods for detecting seizure-like activity in zebrafish, though powerful, have been hampered by low sensitivity (locomotor/behavioral assays) or low-throughput (tectal electrophysiology or calcium fluorescence microscopy). To address these issues, we developed a novel approach to assay seizure-like activity using combined locomotor and calcium fluorescence features, measured simultaneously from unrestrained larval zebrafish using a 96-well fluorescent plate reader. Using custom software to track fish movement and changes in fluorescence (deltaF/F0) from high-speed time-series (12.6Hz), we trained logistic classifiers using elastic net regression to distinguish seizure-like activity from non-seizure related changes based on event-specific and subject-specific features in response to the GABAAR antagonist, pentylenetetrazole (PTZ). We demonstrate that a classifier trained on combined movement and fluorescence data achieves high accuracy ("PTZ M+F"; area-under-curve receiver-operator characteristic (AUC-ROC): 0.98; F1 score: 0.912) and out-performs classifiers trained on movement ("PTZ M"; AUC-ROC: 0.9, F1: 0.9) or fluorescence features alone ("PTZ F"; AUC-ROC 0.96; F1: 0.87). The rate of classified seizure-like events increases as a dose-response to PTZ (serial dose escalation, 0, 2.5mM, 15mM) and is strongly suppressed by ASM treatment (valproic acid, VPA; tiagabine, TGB). At high-dose PTZ, we show that VPA reduces seizure-like activity defined by either "PTZ M+F" or "PTZ M" classifiers. Meanwhile, TGB selectively reduces events defined by the "PTZ M+F" classifier, paralleling previous reports that TGB reduces electrographic but not locomotor seizures and highlighting the potential for our approach to combine features of previously orthogonal assays. Using ASM benchmark data, we employ bootstrap simulation to calculate the expected statistical power of our method as a function of sample size. We demonstrate that anti-seizure responses (robust strictly standardized mean difference, RSSMD, versus control) with magnitudes similar to those associated with VPA or TGB can be reliably detected (true positive rate (TPR) > 90%) with as few as N=4 biological replicates per group, while maintaining a 5% false positive rate. In a prospective test screen with 3-6 replicates per group and on-plate controls, the anti-seizure effect of 4 out of 5 tested ASMs (CBZ, LEV, LZP, TGB) was detected. In summary, we demonstrate a simple high-throughput approach to whole organism anti-seizure phenotyping combining two previously reported metrics to facilitate screens for novel anti-seizure interventions in zebrafish.