Zebrafish might be the future of studying neuroactive compounds
Psychedelic compounds have been around for millennia, and we’ve been using them ever since and trying to unsuccessfully understand their mechanisms of action, including widely used compounds such as Ketamine. We lack an understanding of some mechanisms of action due to the complexity of G-Protein Coupled Receptors and ion channel-mediated pathways in the vertebrate nervous system.
Additionally, there are other aspects we must take into consideration, such as the polypharmacology of neuroactive drugs, making a single-target approach to drug discovery nearly an impossible task.
In summary, we use it because we know it works. But we do not understand how or why it works.
One of the solutions presented by science to overcome such issues is to do a phenotypic screening to identify compounds that interact with individual or multiple targets, allowing us to prioritize the desired and often biologically complex readouts of induced phenotypes on higher-level model systems. Phenotypic screening has faced historical limitations until recently when a group of researchers from the University of California, San Francisco, found a way to rapidly profile thousands of compounds using larval zebrafish, a species that shares genetics and CNS anatomy with humans, making scale to-high throughput testing of complex behavioral readouts possible. By combining this method with human-target-based cheminformatics methods like SEA (Similarity Ensemble Approach) and EF (Enrichment Factor) calculations, we can now expand the boundaries of drug discovery and target deconvolution for neuroactive phenotypes in mammals.
Unfortunately, not all is good news. Zebrafish screening data can be both a blessing and a curse for pharmacological exploration due to the challenges in extracting and comparing features in video data. When using larval zebrafish plated on 96-well plates, the team used various compounds and stimuli, including sound and high-intensity light of different colors, generating a broad spectrum of behavioral responses in the fish. The recorded videos include a time series of motion indices to quantify movement over time. Traditionally, the phenotypic distance between these motion index series is determined using correlation distance. Other approaches the team has decided to use include classification and video analysis using machine learning. Correlation distance can differentiate between antipsychotic and anesthetic phenotypes, but it struggles to distinguish more subtle phenotypes. Fish rarely respond to stimuli in a one-to-one video frame, which breaks a basic assumption of how MI distances are traditionally computed. In an experiment with various assays, the strength response to each assay may differ in drug-treated fish. However, correlation distance values all represent equal contributions.
After experimenting with various deep learning and unsupervised learning techniques to classify zebrafish behavior, the team sought an innovative use for Siamese neural networks or twin neural networks, twin-NNs, a model initially developed for biometric fingerprint verification, now adapted to study the behavior of the zebrafish at lower throughput or higher throughput for embryonic development. Allowing the team to screen a library of 650 ligands (from the SCREEN-WELL Neurotransmitter Set, ‘NT-650’, Methods) in high-replicate and train twin-NNs to relate drugs via the phenotypes they induce in larval zebrafish, constructing the screens from the ground up with machine learning, however, these models still exploit unanticipated artifacts in the resulting screen dataset via an undesirable process known more broadly as ‘shortcut learning.’
The team started by studying the effect of retraining the deep metric learning models on randomized datasets designed to test for confounding effects. Thus driving the redesign and collection of a new experimental screen with full physical randomization. The models trained on the revised cluster of diverse neuroactive compounds corresponded well with the known neuroactive biology and phenotypically linked structurally distinct compounds by scaffold hopping.
Learning the distance metric generalizes a screening dataset of diverse drug-like compounds unseen during training, automating the discovery of neuroactive compounds active on human receptors when tested prospectively in vitro.
For now, the research is ongoing and in its early steps, as it was recently approved and published by Nature Communications. However, this could easily be the next step for upcoming neuroactive compounds in the world of pharmacology!:
