Misclassification in memory modification in AppNL-G-F knock-in mouse model of Alzheimer’s disease
Curation statements for this article:-
Curated by eLife
eLife Assessment
This valuable study proposes using a rigorous computational model to assess memory deficits in Alzheimer's Disease with the goal of developing an early diagnosis tool for the disease. Using an established mouse model of the disease, the authors studied multiple behavioral tasks and ages with the goal of showing similarities in behavioral deficits across tasks. Using the model, the authors indicate specific deficits in memory (overgeneralization and overdifferentiation) in mice with the transgene for the disease. However, the evidence presented is incomplete as certain concerns remain regarding the interpretation of the behavioral results and the validation of the model fit.
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (eLife)
Abstract
Alzheimer’s disease (AD), the leading cause of dementia, could potentially be mitigated through early detection and interventions. However, it remains challenging to assess subtle cognitive changes in the early AD continuum. Computational modeling is a promising approach to explain a generative process underlying subtle behavioral changes with a number of putative variables. Nonetheless, internal models of the patient’s reasoning process remain underexplored in AD. Determining the states of an internal model between measurable pathological states and behavioral phenotypes would advance explanations about the generative process in earlier disease stages beyond assessing behavior alone. In this study, we assumed the latent cause model as an internal model and estimated internal states defined by the model parameters being in conjunction with measurable behavioral phenotypes. The 6- and 12-month-old App NL-G-F knock-in AD model mice and the age-matched control mice underwent memory modification learning, which consisted of classical fear conditioning, extinction, and reinstatement. The results showed that App NL-G-F mice exhibited a lower extent of reinstatement of fear memory. Computational modeling revealed that the deficit in the App NL-G-F mice would be due to their internal states being biased toward overgeneralization or overdifferentiation of their observations, and consequently the competing memories were not retained. This deficit was replicated in another type of memory modification learning in the reversal Barnes maze task. Following reversal learning, App NL-G-F mice, given spatial cues, failed to infer coexisting memories for two goal locations during the trial. We concluded that the altered internal states of App NL-G-F mice illustrated their misclassification in the memory modification process. This novel approach highlights the potential of investigating internal states to precisely assess cognitive changes in early AD and multidimensionally evaluate how early interventions may work.
Article activity feed
-
eLife Assessment
This valuable study proposes using a rigorous computational model to assess memory deficits in Alzheimer's Disease with the goal of developing an early diagnosis tool for the disease. Using an established mouse model of the disease, the authors studied multiple behavioral tasks and ages with the goal of showing similarities in behavioral deficits across tasks. Using the model, the authors indicate specific deficits in memory (overgeneralization and overdifferentiation) in mice with the transgene for the disease. However, the evidence presented is incomplete as certain concerns remain regarding the interpretation of the behavioral results and the validation of the model fit.
-
Reviewer #1 (Public review):
Summary:
The authors show certain memory deficits in a mouse knock-in model of Alzheimer's Disease (AD). They show that the observed memory deficits can be explained by a computational model, the latent cause model of associative memory. The memory tasks used include the fear memory task (CFC) and the 'reverse' Barnes maze. Research on AD is important given its known huge societal burden. Likewise, better characterization of the behavioral phenotypes of genetic mouse models of AD is also imperative to advance our understanding of the disease using these models. In this light, I applaud the authors' efforts.
Strengths:
(1) Combining computational modelling with animal behavior in genetic knock-in mouse lines is a promising approach, which will be beneficial to the field and potentially explain any …
Reviewer #1 (Public review):
Summary:
The authors show certain memory deficits in a mouse knock-in model of Alzheimer's Disease (AD). They show that the observed memory deficits can be explained by a computational model, the latent cause model of associative memory. The memory tasks used include the fear memory task (CFC) and the 'reverse' Barnes maze. Research on AD is important given its known huge societal burden. Likewise, better characterization of the behavioral phenotypes of genetic mouse models of AD is also imperative to advance our understanding of the disease using these models. In this light, I applaud the authors' efforts.
Strengths:
(1) Combining computational modelling with animal behavior in genetic knock-in mouse lines is a promising approach, which will be beneficial to the field and potentially explain any discrepancies in results across studies as well as provide new predictions for future work.
(2) The authors' usage of multiple tasks and multiple ages is also important to ensure generalization across memory tasks and 'modelling' of the progression of the disease.
Weaknesses:
(1) I have some concerns regarding the interpretation of the behavioral results. Since the computational model then rests on the authors' interpretation of the behavioral results, it, in turn, makes judging the model's explanatory power difficult as well. For the CFC data, why do knock-in mice have stronger memory in test 1 (Figure 2C)? Does this mean the knock-in mice have better memory at this time point? Is this explained by the latent cause model? Are there some compensatory changes in these mice leading to better memory? The authors use a discrimination index across tests to infer a deficit in re-instatement, but this indicates a relative deficit in re-instatement from memory strength in test 1. The interpretation of these differential DIs is not straightforward. This is evident when test 1 is compared with test 2, i.e., the time point after extinction, which also shows a significant difference across groups, Figure 2F, in the same direction as the re-instatement. A clarification of all these points will help strengthen the authors' case
(2) I have some concerns regarding the interpretation of the Barnes maze data as well, where there already seems to be a deficit in the memory at probe test 1 (Figure 6C). Given that there is already a deficit in memory, would not a more parsimonious explanation of the data be that general memory function in this task is impacted in these mice, rather than the authors' preferred interpretation? How does this memory weakening fit with the CFC data showing stronger memories at test 1? While I applaud the authors for using multiple memory tasks, I am left wondering if the authors tried fitting the latent cause model to the Barnes maze data as well.
(3) Since the authors use the behavioral data for each animal to fit the model, it is important to validate that the fits for the control vs. experimental groups are similar to the model (i.e., no significant differences in residuals). If that is the case, one can compare the differences in model results across groups (Figures 4 and 5). Some further estimates of the performance of the model across groups would help.
(4) Is there an alternative model the authors considered, which was outweighed in terms of prediction by this model? One concern here is also parameter overfitting. Did the authors try leaving out some data (trials/mice) and predicting their responses based on the fit derived from the training data?
-
Reviewer #2 (Public review):
Summary:
This manuscript proposes that the use of a latent cause model for the assessment of memory-based tasks may provide improved early detection of Alzheimer's Disease as well as more differentiated mapping of behavior to underlying causes. To test the validity of this model, the authors use a previously described knock-in mouse model of AD and subject the mice to several behaviors to determine whether the latent cause model may provide informative predictions regarding changes in the observed behaviors. They include a well-established fear learning paradigm in which distinct memories are believed to compete for control of behavior. More specifically, it's been observed that animals undergoing fear learning and subsequent fear extinction develop two separate memories for the acquisition phase and the …
Reviewer #2 (Public review):
Summary:
This manuscript proposes that the use of a latent cause model for the assessment of memory-based tasks may provide improved early detection of Alzheimer's Disease as well as more differentiated mapping of behavior to underlying causes. To test the validity of this model, the authors use a previously described knock-in mouse model of AD and subject the mice to several behaviors to determine whether the latent cause model may provide informative predictions regarding changes in the observed behaviors. They include a well-established fear learning paradigm in which distinct memories are believed to compete for control of behavior. More specifically, it's been observed that animals undergoing fear learning and subsequent fear extinction develop two separate memories for the acquisition phase and the extinction phase, such that the extinction does not simply 'erase' the previously acquired memory. Many models of learning require the addition of a separate context or state to be added during the extinction phase and are typically modeled by assuming the existence of a new state at the time of extinction. The Niv research group, Gershman et al. 2017, have shown that the use of a latent cause model applied to this behavior can elegantly predict the formation of latent states based on a Bayesian approach, and that these latent states can facilitate the persistence of the acquisition and extinction memory independently. The authors of this manuscript leverage this approach to test whether deficits in the production of the internal states, or the inference and learning of those states, may be disrupted in knock-in mice that show both a build-up of amyloid-beta plaques and a deterioration in memory as the mice age.
Strengths:
I think the authors' proposal to leverage the latent cause model and test whether it can lead to improved assessments in an animal model of AD is a promising approach for bridging the gap between clinical and basic research. The authors use a promising mouse model and apply this to a paradigm in which the behavior and neurobiology are relatively well understood - an ideal situation for assessing how a disease state may impact both the neurobiology and behavior. The latent cause model has the potential to better connect observed behavior to underlying causes and may pave a road for improved mapping of changes in behavior to neurobiological mechanisms in diseases such as AD.
Weaknesses:
I have several substantial concerns which I've detailed below. These include important details on how the behavior was analyzed, how the model was used to assess the behavior, and the interpretations that have been made based on the model.
(1) There is substantial data to suggest that during fear learning in mice separate memories develop for the acquisition and extinction phases, with the acquisition memory becoming more strongly retrieved during spontaneous recovery and reinstatement. The Gershman paper, cited by the authors, shows how the latent causal model can predict this shift in latent states by allowing for the priors to decay over time, thereby increasing the posterior of the acquisition memory at the time of spontaneous recovery. In this manuscript, the authors suggest a similar mechanism of action for reinstatement, yet the model does not appear to return to the acquisition memory state after reinstatement, at least based on the examples shown in Figures 1 and 3. Rather, the model appears to mainly modify the weights in the most recent state, putatively the 'extinction state', during reinstatement. Of course, the authors must rely on how the model fits the data, but this seems problematic based on prior research indicating that reinstatement is most likely due to the reactivation of the acquisition memory. This may call into question whether the model is successfully modeling the underlying processes or states that lead to behavior and whether this is a valid approach for AD.
(2) As stated by the authors in the introduction, the advantage of the fear learning approach is that the memory is modified across the acquisition-extinction-reinstatement phases. Although perhaps not explicitly stated by the authors, the post-reinstatement test (test 3) is the crucial test for whether there is reactivation of a previously stored memory, with the general argument being that the reinvigorated response to the CS can't simply be explained by relearning the CS-US pairing, because re-exposure the US alone leads to increase response to the CS at test. Of course there are several explanations for why this may occur, particularly when also considering the context as a stimulus. This is what I understood to be the justification for the use of a model, such as the latent cause model, that may better capture and compare these possibilities within a single framework. As such, it is critical to look at the level of responding to both the context alone and to the CS. It appears that the authors only look at the percent freezing during the CS, and it is not clear whether this is due to the contextual US learning during the US re-exposure or to increased response to the CS - presumably caused by reactivation of the acquisition memory. For example, the instance of the model shown in Figure 1 indicates that the 'extinction state', or state z6, develops a strong weight for the context during the reinstatement phase of presenting the shock alone. This state then leads to increased freezing during the final CS probe test as shown in the figure. By not comparing the difference in the evoked freezing CR at the test (ITI vs CS period), the purpose of the reinstatement test is lost in the sense of whether a previous memory was reactivated - was the response to the CS restored above and beyond the freezing to the context? I think the authors must somehow incorporate these different phases (CS vs ITI) into their model, particularly since this type of memory retrieval that depends on assessing latent states is specifically why the authors justified using the latent causal model.
(3) This is related to the second point above. If the question is about the memory processes underlying memory retrieval at the test following reinstatement, then I would argue that the model parameters that are not involved in testing this hypothesis be fixed prior to the test. Unlike the Gershman paper that the authors cited, the authors fit all parameters for each animal. Perhaps the authors should fit certain parameters on the acquisition and extinction phase, and then leave those parameters fixed for the reinstatement phase. To give a more concrete example, if the hypothesis is that AD mice have deficits in differentiating or retrieving latent states during reinstatement which results in the low response to the CS following reinstatement, then perhaps parameters such as the learning rate should be fixed at this point. The authors state that the 12-month-old AD mice have substantially lower learning rate measures (almost a 20-fold reduction!), which can be clearly seen in the very low weights attributed to the AD mouse in Figure 3D. Based on the example in Figure 3D, it seems that the reduced learning rate in these mice is most likely caused by the failure to respond at test. This is based on comparing the behavior in Figures 3C to 3D. The acquisition and extinction curves appear extremely similar across the two groups. It seems that this lower learning rate may indirectly be causing most of the other effects that the authors highlight, such as the low σx, and the changes to the parameters for the CR. It may even explain the extremely high K. Because the weights are so low, this would presumably lead to extremely low likelihoods in the posterior estimation, which I guess would lead to more latent states being considered as the posterior would be more influenced by the prior.
(4) Why didn't the authors use the latent causal model on the Barnes maze task? The authors mention in the discussion that different cognitive processes may be at play across the two tasks, yet reversal tasks have been suggested to be solved using latent states to be able to flip between the two different task states. In this way, it seems very fitting to use the latent cause model. Indeed, it may even be a better way to assess changes in σx as there are presumably 12 observable stimuli/locations.
-
Reviewer #3 (Public review):
Summary:
This paper seeks to identify underlying mechanisms contributing to memory deficits observed in Alzheimer's disease (AD) mouse models. By understanding these mechanisms, they hope to uncover insights into subtle cognitive changes early in AD to inform interventions for early-stage decline.
Strengths:
The paper provides a comprehensive exploration of memory deficits in an AD mouse model, covering the early and late stages of the disease. The experimental design was robust, confirming age-dependent increases in Aβ plaque accumulation in the AD model mice and using multiple behavior tasks that collectively highlighted difficulties in maintaining multiple competing memory cues, with deficits most pronounced in older mice.
In the fear acquisition, extinction, and reinstatement task, AD model mice …
Reviewer #3 (Public review):
Summary:
This paper seeks to identify underlying mechanisms contributing to memory deficits observed in Alzheimer's disease (AD) mouse models. By understanding these mechanisms, they hope to uncover insights into subtle cognitive changes early in AD to inform interventions for early-stage decline.
Strengths:
The paper provides a comprehensive exploration of memory deficits in an AD mouse model, covering the early and late stages of the disease. The experimental design was robust, confirming age-dependent increases in Aβ plaque accumulation in the AD model mice and using multiple behavior tasks that collectively highlighted difficulties in maintaining multiple competing memory cues, with deficits most pronounced in older mice.
In the fear acquisition, extinction, and reinstatement task, AD model mice exhibited a significantly higher fear response after acquisition compared to controls, as well as a greater drop in fear response during reinstatement. These findings suggest that AD mice struggle to retain the fear memory associated with the conditioned stimulus, with the group differences being more pronounced in the older mice.
In the reversal Barnes maze task, the AD model mice displayed a tendency to explore the maze perimeter rather than the two potential target holes, indicating a failure to integrate multiple memory cues into their strategy. This contrasted with the control mice, which used the more confirmatory strategy of focusing on the two target holes. Despite this, the AD mice were quicker to reach the target hole, suggesting that their impairments were specific to memory retrieval rather than basic task performance.
The authors strengthened their findings by analyzing their data with a leading computational model, which describes how animals balance competing memories. They found that AD mice showed somewhat of a contradiction: a tendency to both treat trials as more alike than they are (lower α) and similar stimuli as more distinct than they are (lower σx) compared to controls.
Weaknesses:
While conceptually solid, the model struggles to fit the data and to support the key hypothesis about AD mice's ability to retain competing memories. These issues are evident in Figure 3:
(1) The model misses key trends in the data, including the gradual learning of fear in all groups during acquisition, the absence of a fear response at the start of the experiment, the increase in fear at the start of day 2 of extinction (especially in controls), and the more rapid reinstatement of fear observed in older controls compared to acquisition.
(2) The model attributes the higher fear response in controls during reinstatement to a stronger association with the context from the unsignaled shock phase, rather than to any memory of the conditioned stimulus from acquisition.
These issues lead to potential overinterpretation of the model parameters. The differences in α and σx are being used to make claims about cognitive processes (e.g., overgeneralization vs. overdifferentiation), but the model itself does not appear to capture these processes accurately.
The authors could benefit from a model that better matches the data and that can capture the retention and recollection of a fear memory across phases.
Conclusion:
Overall, the data support the authors' hypothesis that AD model mice struggle to retain competing memories, with the effect becoming more pronounced with age. While I believe the right computational model could highlight these differences, the current model falls short in doing so.
-
Author response:
We appreciate the reviewers’ constructive comments and suggestions. We plan the following revisions to address the public reviews.
Regarding model selection (from Reviewers 1 and 3)
We will test whether the latent cause model has a better explanatory power for the observed reinstatement data compared with at least two other models, including the Rescorla-Wagner model. For each model, the prediction errors across all trials and those in the test 3 trial (reinstatement) will be calculated for individual animals. The explanatory power of the models will be discussed based on these results.
Regarding model validation (from Reviewers 1, 2, and 3)
We acknowledge the reviewers’ concerns about potential parameter overfitting and misinterpretation. First, the simulation in the latent cause model will be run under other possible …
Author response:
We appreciate the reviewers’ constructive comments and suggestions. We plan the following revisions to address the public reviews.
Regarding model selection (from Reviewers 1 and 3)
We will test whether the latent cause model has a better explanatory power for the observed reinstatement data compared with at least two other models, including the Rescorla-Wagner model. For each model, the prediction errors across all trials and those in the test 3 trial (reinstatement) will be calculated for individual animals. The explanatory power of the models will be discussed based on these results.
Regarding model validation (from Reviewers 1, 2, and 3)
We acknowledge the reviewers’ concerns about potential parameter overfitting and misinterpretation. First, the simulation in the latent cause model will be run under other possible conditions to test whether our original condition can be justified, then clarify how certain parameters affect the predicted CR. Second, we will confirm if the prediction errors are comparable between experimental groups, present the correlation between parameters, and discuss this result in the revision.
To evaluate the effect of context in explaining reinstatement in the latent cause model, simulations of CR in test 3 when only context or tone is presented will also be performed and discussed with the behavioral data.
Regarding the interpretation of the behavioral data (from Reviewers 1, 2, and 3) We will clarify our interpretation of the behavioral data by incorporating the additional analyses mentioned above; for example, to clarify the contribution of context in test 3, we will provide data on the CR before the tone presentation in our revision. In addition, how we expected and interpreted the reversal Barnes maze results from the memory modification characteristics estimated in the reinstatement test will be further discussed.
Regarding the application of the latent cause model to the reversal Barnes maze task (from Reviewers 1, 2)
We acknowledge the reviewers’ suggestions to apply the latent cause model to our Barnes maze results to strengthen the link and consistency. To further clarify the reason for including Barnes maze results, we will explicitly discuss how associative learning is involved in spatial learning in the revision. However, we will not be able to directly apply the latent cause model for the Barnes maze data for the following reasons. As we noted in the Results and Discussion, the latent cause model was built on associative learning and cannot be directly applied to the Barnes maze data. The cognitive processes in the Barnes maze task involve maintaining spatial representation of the environment, integrating own position and expected goal, and evaluating potential actions. Importantly, the chosen actions in this task directly affect subsequent observations, while an animal’s response based on an expected outcome typically does not alter future observation in a simple associative learning paradigm.
Thus, although associative learning (e.g., associations between the spatial cue and the location of the escape box) is certainly a critical building block and contributes to performance in the Barnes maze task, this mechanism alone cannot fully explain the animal’s navigation in the maze. We agree that having solid modeling results in the reversal Barnes maze task is an important direction, but extending the latent cause model for this purpose is beyond the scope of this study. We have suggested some possible approaches in the Discussion and will elaborate further on these conceptual distinctions and how latent cause framework assists in the interpretation of results.
-
-
-