Although fruit flies are very different from humans, they have nonetheless taught us a great deal about how our brains work. For years, scientists have been manipulating fruit fly genes and studying their behavior to link genetic defects to neurodegeneration.
“For the past two decades, to interpret the behavioral consequences of various genetic perturbations in the fly, the field has used the climbing assay, which assesses how well flies are able to climb when placed into a vertical arena,” said Sherry Aw, Group Leader at A*STAR’s Institute of Molecular and Cell Biology (IMCB). However, climbing performance may not capture the full range of behavioral outcomes that may be relevant to neuroscience.
To overcome this challenge, Aw’s team, in collaboration with Li Cheng’s group at A*STAR’s Bioinformatics Institute (BII) and colleagues at the National University of Singapore, developed a machine-learning image analysis program called Feature Learning-based LImb segmentation and Tracking (FLLIT) for fully automated tracking of fruit fly leg movements.
Unlike other machine learning methods that require training on large, manually-labeled image datasets, FLLIT generates its training datasets based on morphological parameters built into the learning algorithm. The program identifies pixels representing the legs of fruit flies via a skeletonization operation for recognizing thin and spindly structures, and an edge operation that traces the outline of the entire fly.
Testing FLLIT on videos of flies, the researchers demonstrated that approximately 97.5% of the computationally identified fly claws were within three pixels of the manually annotated positions. Confident in the accuracy of FLLIT, the researchers then showed that the program could distinguish gait patterns in fly models of Parkinson’s disease (PD) and spinocerebellar ataxia 3 (SCA3).
“In human patients, SCA3 is associated with a lurching, uncoordinated gait, while PD is marked by a stiff, rigid gait,” Aw noted. Similarly, FLLIT detected that fly models of SCA3 and PD walked very differently from one another, also highlighting resemblance between human and fly gaits under disease conditions. Importantly, using FLLIT, the researchers found that two different PD fly models showed a strikingly similar rigid gait despite being of completely different genetic backgrounds, suggesting that alternative molecular mechanisms converge on the same pathological outcome.
“We are also using FLLIT to study the cellular mechanisms that underlie various distinct movement defects such as tremors, which are very prevalent but remain poorly understood. We believe that such detailed phenotyping will not only be useful for basic science research, but will also apply to sensitive behavioral phenotype-based drug screening. We are looking for industrial collaborators to work with on such drug screening projects,” said Aw.
The A*STAR-affiliated researchers contributing to this research are from the Institute of Molecular and Cell Biology (IMCB) and the Bioinformatics Institute (BII).
