Balancing ethics and statistics: machine learning facilitates highly accurate classification of mice according to their trait anxiety with reduced sample sizes.

Abstract

Understanding how individual differences influence vulnerability to disease and responses to pharmacological treatments represents one of the main challenges in behavioral neuroscience. Nevertheless, inter-individual variability and sex-specific patterns have been long disregarded in preclinical studies of anxiety and stress disorders. Recently, we established a model of trait anxiety that leverages the heterogeneity of freezing responses following auditory aversive conditioning to cluster female and male mice into sustained and phasic endophenotypes. However, unsupervised clustering required larger sample sizes for robust results which is contradictory to animal welfare principles. Here, we pooled data from 470 animals to train and validate supervised machine learning (ML) models for classifying mice into sustained and phasic responders in a sex-specific manner. We observed high accuracy and generalizability of our predictive models to independent animal batches. In contrast to data-driven clustering, the performance of ML classifiers remained unaffected by sample size and modifications to the conditioning protocol. Therefore, ML-assisted techniques not only enhance robustness and replicability of behavioral phenotyping results but also promote the principle of reducing animal numbers in future studies.