Behavioral chaos in Alzheimer’s disease mice decoded by machine learning

Cutting-edge machine learning tools reveal hidden patterns in Alzheimer’s disease mouse behavior, opening the door to innovative treatments targeting neuroinflammation.

Study: Machine learning reveals prominent spontaneous behavioral changes and treatment efficacy in humanized and transgenic Alzheimer’s disease models. Image Credit: nobeastsofierce / Shutterstock

In a recent study published in the journal Cell Reports, researchers used the machine learning (ML)-based Variational Animal Motion Embedding (VAME) segmentation platform to analyze behavior in Alzheimer’s disease (AD) mouse models and tested the effect of blocking fibrinogen-microglia interactions. They found that AD models showed age-dependent behavioral disruptions, including increased randomness and disrupted habituation, largely prevented by reducing neuroinflammation, with VAME outperforming traditional methods in sensitivity and specificity.

Background

Behavioral alterations, central to neurological disorders, are complex and challenging to measure accurately. Traditional task-based tests provide limited insight into disease-induced changes. However, advances in computer vision and ML tools, such as DeepLabCut, SLEAP, and VAME, now enable the segmentation of spontaneous mouse behavior into postural units (motifs) to uncover sequence and hierarchical structure, offering scalable, unbiased measures of brain dysfunction.

AD, characterized by amyloid accumulation preceding tau pathology and neurodegeneration, often presents subtle behavioral or neuropsychiatric changes like agitation, depression, and loss of motivation decades before dementia onset. These early changes offer a promising window to study AD pathogenesis and therapeutic interventions.

Humanized amyloid precursor protein (App) knockin and transgenic APP mouse models replicate key AD features, such as amyloidosis and neuroinflammation. Despite these advancements, analyzing non-task-oriented spontaneous behavior in AD models has remained technically challenging until the emergence of ML-based behavioral analysis methods. Recent refinements to VAME allow it to incorporate advanced kinematic and network analyses, providing deeper insights into behavioral organization and disease progression.

In the present study, researchers enhanced their VAME ML pipeline to analyze spontaneous behavior, disease progression, and treatment effects in AD mouse models to validate the model’s sensitivity for detecting neuroinflammation-related therapeutic outcomes.

About the study

This study examined Alzheimer’s disease (AD)-related behaviors and pathology using two mouse models: AppNL-G-F and 5xFAD. The AppNL-G-F mice, which express humanized amyloid-β (Aβ) with familial AD mutations, were assessed at young (six months), middle-aged (13 months), and advanced age (22 months).

Behavioral experiments included spontaneous activity in an open arena and spatial memory assessment using the Morris water maze. Histological analysis examined amyloidosis and gliosis. The 5xFAD mice, overexpressing human App and presenilin 1 (PS1) with multiple familial AD mutations, were studied at nine months. To evaluate a potential therapeutic intervention, 5xFAD mice were crossed with Fggγ390–396A mice, a model targeting fibrinogen-microglia interactions.

Behavioral data were captured using DeepLabCut, a video-based pose estimation tool, and analyzed with VAME ML that identifies distinct behavioral motifs and sequences. Motif usage, transitions, and behavioral community structures were examined to identify disease-associated changes.

Spatial learning deficits and increased behavioral randomness were observed in AppNL-G-F mice, while 5xFAD mice exhibited significant motif alterations, with females showing heightened sensitivity to these changes. Many of these abnormalities were restored by the Fggγ390–396A intervention. Classifier analysis was used to compare VAME’s sensitivity and specificity with conventional open-field metrics. A comparison with keypoint-MoSeq was also performed to validate VAME’s outcomes.

Results and discussion

Aged AppNL-G-F mice (22 months) showed mild memory deficits in the Morris water maze, along with severe amyloidosis and gliosis. Using VAME, age- and genotype-related changes in spontaneous behavior were identified. These mice showed significant changes in the use of eight out of 30 identified behavioral motifs, including walking, rearing, and exploration. Higher-order behavioral communities were also disrupted, with impaired habituation, abnormal sensitization, and increased randomness in behavior. Motif transition analysis revealed decreased predictability and premature transitions from active to static behaviors.

In 5xFAD mice, 17 behavioral motifs were found to be significantly affected, with significant abnormalities in transitions and reduced behavioral predictability. Blocking fibrinogen-microglia interactions using the Fggγ390–396A mutation partially or fully rescued these behavioral abnormalities, restoring motif use, speeds, and transitions. Therapeutic effects were particularly evident in fast exploratory and ambulatory behaviors. Importantly, the therapeutic intervention demonstrated disease-specific effects, as fibrinogen blockade did not alter behavior in non-AD controls.

Classifier analysis demonstrated that VAME provided greater sensitivity and specificity than conventional open-field metrics in detecting behavioral differences across genotypes and therapeutic outcomes. Both VAME and keypoint-MoSeq reliably identified disease-associated behavioral alterations, but VAME outcomes were more comprehensive and specific. These findings underscore VAME’s utility in addressing the core disorganization of behavioral sequences observed in AD models.

Together, these results highlight VAME as a robust tool for quantifying complex behaviors and assessing preclinical disease phenotypes and therapeutic interventions with superior specificity and scalability compared to conventional methods.

Additionally, the findings highlight fibrinogen-microglia interactions as a potential therapeutic target. However, the study did not explicitly assess cognitive functions, brain regions, or neural systems. It remains unclear if spontaneous behavioral impairments directly reflect cognitive decline and could potentially serve as biomarkers.

Conclusion

In conclusion, unbiased ML approaches such as VAME enable rigorous quantification of disease-induced behavioral alterations, improving construct and predictive validity assessments in mouse models of neurodegenerative diseases. The incorporation of behavioral community analysis and transition networks provides a scalable and sensitive framework for identifying disease-related disruptions. The method could potentially enhance the translatability of preclinical testing by providing sensitive and scalable tools to evaluate disease progression and therapeutic interventions, addressing a critical gap in neuroscience research.