Title: Mapping neural contributions of GLP-1 receptor agonism to phenotype using machine learning
Abstract:
GLP-1, an incretin hormone, targets the G protein-coupled receptor GLP-1R, expressed in brain regions crucial for regulating food intake and satiety, particularly in the hypothalamus. Recently, two long-acting peptide GLP-1R agonists (GLP1RAs), liraglutide and semaglutide, were approved for the treatment of obesity in the United States.
Peptide GLP1RAs have low bioavailability, requiring subcutaneous injection. Injectable drugs pose challenges for patient adherence, storage, and production costs. Novel, orally administered small molecule GLP1RAs, like Pfizer's danuglipron, are currently in clinical trials. Danuglipron activates GLP-1R signaling by binding to the receptor's N-terminal extracellular domain at tryptophan residue 33, a primate-specific site unique to small molecule GLP1RAs, highlighting the necessity for a mouse model.
In response, our lab created a humanized GLP-1R (hGLP1R) mouse model and a Cre-dependent viral construct containing hGLP1R. This construct enables the expression of hGLP1R in specific neuronal populations via intracranial injection into Glp1r-Cre mice. This approach allows us to target distinct neural regions expressing the receptor, using a small molecule drug.
Recent advancements in deep learning and machine learning enable thorough analysis of rodent behavior in home cage settings. By automating behavioral classification, subtle behavior variations over hours of time become detectable. In our study, we monitored hGLP1R mice for 4 hours in a home cage after administering danuglipron or a vehicle control. Using Social LEAP Estimates Animal Poses (SLEAP), a deep-learning based framework, we performed animal pose estimation on video recordings. Our SLEAP model was trained on 9,433 labeled images of mice, tracking 9 keypoints. We then employed Keypoint Motion Sequencing (Keypoint-MoSeq), which distinguishes keypoint noise from behavior using a generative model, to identify behavioral syllables. We calculated transitions between syllables, frequency, and total time spent on each syllable for each video. Principal component analysis (PCA) followed by K-means clustering was used to model the results. Additionally, a random forest model was trained to predict group (drug and genotype) from behavioral data, using the scikit-learn Python package.
Our analysis revealed distinct behavioral differences in hGLP1R mice treated with danuglipron compared to vehicle. Furthermore, by analyzing hGP1R virally injected mice on danuglipron, we mapped behavioral phenotype contributions to specific neural regions, underscoring how the individual neural targets of small molecule GLP1RA give rise to phenotype. Understanding the behavioral consequences and neural contributions of GLP1RA could facilitate the development of new or improved drug formulations, crucial given the escalating global rates of obesity and millions of GLP1RA users.
Audience Take Away Notes:
- Use of deep learning-based frameworks for automated analysis of rodent behavior
- Application of automated behavioral analysis to the pharmaceutical industry
- Home cage analysis improves reproducibility of behavioral research and saves human time
- Application of clustering algorithms and machine learning to analyze behavioral data
- The neural mechanisms underlying GLP1RA action and food-seeking behaviors
