An artificial intelligence based, markerless motion capture approach for quantifying motor deficits of dystrophic zebrafish larvae


Topic:

Pre-Clinical Research

Poster Number: S30

Author(s):

Jeffrey Widrick, PhD, Boston Children's Hospital, Matthias Lambert, PhD, Boston Children's Hospital, Felipe de Souza Leite, PhD, Boston Children's Hospital / Harvard Medical School, Youngsook Jung, PhD, Boston Children's Hospital, Junseok Park, PhD, Boston Children's Hospital, James Conner, BS, Boston Children's Hospital, Alice Lee, PhD, Boston Children's Hospital, Alan Beggs, PhD, Boston Children's Hospital, Louis Kunkel, PhD, Boston Children's Hospital

BACKGROUND: The sapje zebrafish strain is the smallest vertebrate model of Duchenne muscular dystrophy (DMD). Like DMD, sapje larvae have an observable motor deficit. OBJECTIVES: Our goal was to develop a high-temporal and -spatial resolution tool capable of quantifying the extent of these motor deficits, and more broadly, how the underlying biomechanics of locomotion are altered by the absence of dystrophin. RESULTS: We obtained high-temporal resolution video recordings (1000 frames/s) of individual sapje larvae performing escape responses in a shallow arena (123 total larvae: 29 +/+, 45 +/−, 49 −/−). A deep learning neural network was trained to identify seven anatomical keypoints located along the body and tail with human level accuracy. The two-dimensional keypoint coordinates output by the network were used to model escaping larvae as six linked body segments. Escape response distance, displacement, and peak instantaneous speed were highly repeatable across three separate escape response trials. A random forest classifier, using 29 linear and angular kinematic variables derived from the models, distinguished homozygous and heterozygous wild-types from dystrophic larvae with an out-of-bag error of < 2.5%. The two most predictive variables were mean distance and peak instantaneous speed which were both over 3 standard deviations lower in mutants vs. wild-types. The poorer performance of the mutants was attributed mainly to their slower tail beat angular velocity rather than to a reduced tail beat amplitude. In fact, tail kinematic variables were as effective as non-survival in vitro measures of tail muscle contractility in distinguishing mutants from wild-types. We repeated the study (except for contractile measurements) with a second strain of dystrophin deficient larvae, sapje-like (16 +/+, 38 +/−, 52 −/−), and obtained similar results. CONCLUSIONS: The approach described here describes novel, repeatable, high-precision kinematic variables that, 1) provide new insight into how dystrophin deficiency impacts vertebrate locomotion, 2) can serve as non-lethal surrogates of muscle function, and 3) represent statistically powerful outcomes for future studies.