Machine learning reveals prominent spontaneous behavioral changes and treatment efficacy in humanized and transgenic Alzheimer's disease models

Abstract

Computer-vision and machine-learning (ML) approaches are being developed to provide scalable, unbiased, and sensitive methods to assess mouse behavior. Here, we used the ML-based variational animal motion embedding (VAME) segmentation platform to assess spontaneous behavior in humanized App knockin and transgenic APP models of Alzheimer's disease (AD) and to test the role of AD-related neuroinflammation in these behavioral manifestations. We found marked alterations in spontaneous behavior in App^NL-G-F and 5xFAD mice, including age-dependent changes in motif utilization, disorganized behavioral sequences, increased transitions, and randomness. Notably, blocking fibrinogen-microglia interactions in 5xFAD-Fgg^γ390-396A mice largely prevented spontaneous behavioral alterations, indicating a key role for neuroinflammation. Thus, AD-related spontaneous behavioral alterations are prominent in knockin and transgenic models and sensitive to therapeutic interventions. VAME outcomes had higher specificity and sensitivity than conventional behavioral outcomes. We conclude that spontaneous behavior effectively captures age- and sex-dependent disease manifestations and treatment efficacy in AD models.

Keywords: App-KI; CP: Neuroscience; DeepLabCut; Keypoint-MoSeq; amyloid; behavioral segmentation; cognition; naturalistic behavior; open field; pose estimation; preclinical.

Figure 1.. Aged App^NL-G-F mice show robust AD-related pathology and mild impairments in the Morris water maze

Spatial learning and memory of 22-month-old App^NL-G-F mice (n = 16; eight females and eight males) and WT littermate controls (n = 21; 11 females and 10 males) were tested in the Morris water maze. (A) Distance swum (left) and latency (right) in the hidden (spatial) and visible (cued) platform components of the Morris water maze test. App^NL-G-F mice had mild deficits in distances (p = 0.025) and latencies (p = 0.017) in the spatial component of the Morris water maze. p values were determined by repeated two-way ANOVA. (B) Time spent in the target (platform removed) and nontarget quadrants during the 24-h probe trial after 10 sessions of hidden training. App^NL-G-F and WT mice did not perform differently at the target quadrant (p = 0.108). However, WT (p = 0.007), but not App^NL-G-F (p = 0.999), mice had a target quadrant preference. p values were determined by one-way ANOVA and Bonferroni post hoc test. (C) Representative images of hippocampus and cortex stained for 82E1-positve Aβ deposits and Iba1-positive microglia show severe amyloidosis and microgliosis in 22-month-old App^NL-G-F mice. (D) Quantification of hippocampal and cortical area (%) occupied by 82E1-positve Aβ deposits or by Iba1-positive microglia in 22-month-old App^NL-G-F (n = 14) and WT (n = 15) mice. ***p < 0.01 by Student’s t test. Values are mean ± SEM.

Figure 2.. VAME segmentation of behavioral motifs during exploration in an open arena

(A) Workflow. Mice were recorded from below at 25 frames per second (25 Hz) for 25 min (~37,500 frames/mouse) in a circular arena 12 inches in diameter. DeepLabCut (DLC) was used for pose estimation of nine body parts. VAME was used to identify behavioral motifs. Custom scripts were used to define egocentric coordinates, communities, and transition sequences. (B) DLC pose estimation of nine defined body parts (ventral view). (C) Allocentric (relative to arena) and egocentric (relative to mouse center) x and y coordinates (pixels) and allocentric speeds of the nine defined body parts (B) of a representative App^NL-G-F mouse recorded over 2 min at 25 Hz. (D) Egocentric coordinates relative to center (yellow) for the defined body parts (B) over 4 min in a representative App^NL-G-F mouse (1 point per 40 ms; 9,600 frames). Note the left-right and rostral-caudal motion range of all body parts. (E) Motif sequence (5-s mode) for (C). VAME model trained on egocentric coordinates for nose, paws, center, and tail base. (F) UMAP representation of the 30 motifs identified by VAME for all mice (n = 32; 18 App^NL-G-F and 14 WT mice). Motifs occupy distinct locations with smooth transitions as expected for continuous behavioral sequences; 10,000 random frames shown from the entire cohort and duration. (G) Egocentric coordinates in the rostral-caudal (y axis, top) and left-right (x axis, middle) axes of the nine body parts for the 30 identified motifs (n = frames; 32 mice). Points are the average location across all mice for the first 20 frames (800 ms) from motif identification. Allocentric center speed (bottom) across the full cohort (n = frames; 32 mice). Motifs show specific and distinct coordinates and speeds. Values are mean ± SEM.

Figure 3.. Middle-age App^NL-G-F mice display robust behavioral alterations in motif use

Spontaneous behavior of 13-month-old App^NL-G-F mice (n = 18; eight females and 10 males) and WT littermate controls (n = 14; eight females and six males) was recorded for 25 min in a circular open arena. (A) Thirty behavioral motifs were identified by VAME. Motif use (relative to sex-matched WT littermates) of 30 identified motifs for all mice. Relative to littermate controls, App^NL-G-F mice displayed usage alterations in eight motifs. **q < 0.01, *q < 0.05 by false discovery rate with Benjamini-Hochberg correction for multiple comparisons (FDR-BH). (B) Kinematic analyses of egocentric coordinates of the defined body parts for the first 10 frames (400 ms) of the eight motifs significantly affected in App^NL-G-F mice (A, q < 0.05 by FDR-BH; n = frames; 32 mice). (C) Schematic of body part positions for the indicated motifs. Data points are the average location for 10 frames after motif identification (n = frames; 32 mice). Motifs show precise and distinct coordinates. (D) Absolute motif use (in seconds) by genotype and sex for the eight motifs altered in App^NL-G-F mice (A). Males and females differed in their use of motifs 0, 2, 5, and 12, but App^NL-G-F expression affected both sexes similarly. p values were determined by two-way (genotype and sex) ANOVA. Values are mean ± SEM.

Figure 4.. Hierarchical clustering of motifs identifies behavioral communities and reveals experience-dependent alterations in App^NL-G-F mice

(A) Cost-function-based hierarchical organization of motifs into 11 communities based on motif transitions and motif use at the cohort level (n = 32; 18 App^NL-G-F and 14 WT mice). Motifs with higher probability of transitions and closer hierarchical distance were clustered into communities on the dendrogram. Node size is proportional to motif use. Red borders and branches indicate significant differences between App^NL-G-F and WT mice by FDR-BH for 30 motifs and 11 communities, respectively. (B) Matrix depicting the probability of forward transitions from ${\mathsf{motif}}_i$ (rows) to ${\mathsf{motif}}_j$ (columns) at the cohort level (n = 32; 18 App^NL-G-F and 14 WT mice). Motifs were organized according to the dendrogram order found in (A). (C) Motif use (relative to sex-matched WT littermates) (top) and speeds (bottom) organized by community. Top: motifs in the same community had reliably similar App^NL-G-F effects, consistent with the notion that each community reflects a tightly associated set of postural units. **q < 0.01, *q < 0.05 by FDR-BH for indicated communities. Bottom: allocentric speed of mouse center in motifs organized by community for female and male App^NL-G-F mice and WT controls. Relative to sex-matched controls, female, but not male, App^NL-G-F mice performed ambulatory motifs with excessive allocentric speed. Red indicates disease-affected motifs. (D) Hierarchical organization of motifs and subjects based on motif use relative to sex-matched WT littermates (n = 32; 18 App^NL-G-F and 14 WT mice). App^NL-G-F and WT mice exhibited different clustering (p = 0.0015 by Mann-Whitney rank-sum test). (E) Percentage of time spent in identified communities in 5-min bins during 25 min of exploration in an open arena. Community usages were strongly modulated by time (experience), and App^NL-G-F mice exhibited impaired time-dependent responses. p values were determined by repeated one-way ANOVA. Values are mean ± SEM.

Figure 5.. Deconstructing full sequences of spontaneous behavior reveals disorganized behavioral sequences, increased motif transitions, and randomness in App^NL-G-F mice

Full behavioral sequences of motif transitions in 13-month-old App^NL-G-F mice (n = 18; eight females and 10 males) and WT littermate controls (n = 14; eight females and six males) were annotated at 25 Hz during 25 min of open-arena exploration (~37,500 potential transitions per mouse) and analyzed with the discrete Markov chain model. (A) Community transition sequences (black) and center speed (blue) of an App^NL-G-F mouse during 60 s of behavioral exploration. (B) Leave-one-out log likelihood estimates of orders 0 and 1 for the Markov chain model of community transitions. App^NL-G-F mice, particularly females, had less predictable behavior. p values were determined by two-way (genotype and sex) ANOVA. (C and D) Cytoscape network visualization of the top 20% most probable motif transitions depicting topological proximity of associated motifs and communities (color code) and significant alterations of motif use (circular borders) and motif transitions (arrows) in App^NL-G-F mice. App^NL-G-F mice had increased transitions (red arrows) and reduced dwell time in motifs (blue dashed arrows) (C). App^NL-G-F mice systematically favored transitions from fast ambulatory to slow exploratory communities (D). (E) Transition difference matrix depicting the delta probability of transitions between App^NL-G-F and WT mice. Blue and red indicate reduced or increased probability of transition, respectively, in App^NL-G-F mice relative to controls. (F) Delta indices (App^NL-G-F WT mice) of motif transition probability, motif use, and motif speed. **p < 0.01, *p < 0.05 by t test. (G) Classifier analyses (logistic regression) of ML outcomes (motif use) versus conventional open field (OF) outcomes (distance, speed, time, and location) obtained from the same videos in App^NL-G-F mice. ML outcomes were more sensitive and specific than conventional outcomes. Values are mean ± SEM.

Figure 6.. Pronounced 5xFAD-dependent alterations in spontaneous behavior are prevented by blocking fibrinogen-microglia interactions in 5xFAD-Fgg^γ390–396A mice

Spontaneous behavior of 8- to 10-month-old 5xFAD mice (n = 13; seven females and six males), 5xFAD-Fgg^γ390–396A mice (n = 12; six females and six males), and controls (n = 22; 10 females and 12 males) was recorded for 60 min in a circular open arena. The control group included WT (n = 13; six females and seven males) and Fgg^γ390–396A (n=9; four females and five males) mice, as their behavior did not differ (Figure S4A). See Figure S4B for sex effects. (A) Motif use of 30 VAME-identified motifs for all mice. Relative to controls, 5xFAD mice displayed motif use alterations in 17 motifs (red asterisks). Relative to 5xFAD mice, 5xFAD-Fgg^γ390–396A showed significant improvements in many motifs (blue asterisks). ***q < 0.001, **q < 0.01, *q < 0.05 by FDR-BH. F, female; M, male. (B) 5xFAD mice performed motifs at higher speeds, which was prevented in 5xFAD-Fgg^γ390–396A mice. Genotype and motif effects were determined by two-way ANOVA (p values). (C) Motif transition subtraction matrix depicting the delta probability of transitions between 5xFAD and control mice. Blue and red indicate reduced or increased probability of transition, respectively, in 5xFAD mice relative to controls. (D) Delta indices of motif transition probability, motif use, and motif speed (n = 47 mice; 13 5xFAD, 12 5xFAD-Fgg^γ390–396A, and 22 control mice). ***p < 0.001, *p < 0.05 by one-way ANOVA and Bonferroni post hoc test for multiple comparisons. (E) Time in motif (seconds) by genotype for the eight 5xFAD-affected motifs restored in 5xFAD-Fgg^γ390–396A mice (A, q < 0.05 by FDR-HB). ***p < 0.001, **p < 0.01, *p < 0.05 by one-way ANOVA and Bonferroni post hoc test for multiple comparisons. (F) Classifier analyses (logistic regression) of ML outcomes (motif use) versus conventional OF outcomes (distance, speed, time, and location) obtained from the same videos. ML outcomes were more sensitive and specific than conventional outcomes. Values are mean ± SEM.

Figure 7.. Pronounced 5xFAD-dependent alterations in community usage and motif transition networks are prevented in 5xFAD-Fgg^γ390–396A mice

(A) Hierarchical organization of motifs into communities based on motif transitions and motif use at the cohort level (n = 47 mice; 13 5xFAD, 12 5xFAD-Fgg^γ390–396A, and 22 control mice). Node size is proportional to motif use. Significant differences (indicated in red) between 5xFAD and control mice were determined by FDR-BH for 30 motifs and nine communities. (B) Time in motif (%) of a representative motif for each disease-affected community (A) during the 60 min of exploration in an open arena. 5xFAD mice exhibited profound deficits, which were largely prevented in 5xFAD-Fgg^γ390–396A mice. p values were determined by repeated one-way ANOVA with Bonferroni post hoc test for multiple comparisons. Values are mean ± SEM. (C) Hierarchical organization of motifs and mice based on motif use relative to sex-matched control littermates. (D) Cytoscape was used to visualize the top 20% most probable motif transitions to depict topological proximity of associated motifs and communities (color code) and significant alterations of motif use (borders) and motif transitions (arrows) for control versus 5xFAD mice (top) and control versus 5xFAD-Fgg^γ390–396A mice (bottom).