CiCi Zheng, 09/09/24, SWDB Workshop Report
This project is built on the work by Piet et al., where the authors characterize the behavioral strategies used by mice during the visual change detection task in the Allen Visual Behavior 2P dataset. By fitting the initiation of lick bouts via a dynamic logistic regression model, two strategies were found to contribute substantially to the model-fit, namely the visual strategy (lick on detected change) and the timing strategy (lick with a fixed probability from the last lick). Moreover, individual mice use a mixture of the two strategies, and the preference of strategies has not only performance correlates (Fig 2G, Fig S5), but the employment of the visual strategy can be mapped to increased activation of the Vip-Sst disinhibitory circuit, which facilitates task- appropriate responses (Fig. 4- 6).
One particularly interesting aspect is how the strategy preference evolves during training. There could be a few plausible and not mutually exclusive trends that one can observe in the data. For example, that the animals might be consistent and maintained their dominant strategy over training. Animal could also show increased preference to the optimal strategy for reward (here, the visual strategy) as training progresses. For the Visual Behavior 2P data, Piet et al. found both trend to be true: that when they classified mice by their dominant strategy (visual or timing) during imaging, 1) the “visual mice” always have more preference for visual strategy during training and 2) the preference for visual strategy (indicated by “visual index”) increases over different training stages in particular for the “visual mice” (Fig. S7 A-C).
The rich training design and the well-documented behavior data allow a deeper dive into the training dynamics and how the strategy preferences on the imaging/recording days are shaped. In particular, in the early training sessions, mice progressed through each training stage by meeting stage-specific criteria. Based on the findings of Piet et al that the mice have a consistent strategy preference, one would hypothesize that the “visual mice” that have a consistent preference for the optimal strategy would progress faster in the training, and they would have experienced less training sessions. In this project we focused on the Allen Visual Behavior Neuropixel dataset, that used the same task as Visual Behavior 2P, but the animals had to meet for additional training criteria and undergo more training sessions after passing all training criteria before they were recorded. I am interested in explore how those different training dynamics might contribute to the final performance in different ways.
Fig 1. Ephys sessions model fit.
Given the limited time, I focused on the wt animals (n = 27). I replicate the same inference framework and applied it to the lick bouts data in ephys sessions, training sessions with the natural image stimuli (Stage 3-5) and the habituation sessions. I directly used the parameters in Piet et al for the timing regressor, but for a more rigorous study one should fit a sigmoid regression from scratch. For pre- ephys sessions, omission related strategies was not included. Model fit were evaluated using area under the ROC curve (AUROC) of 10 fold cross-validated model predictions. For AUROC, .5 means chance level prediction and 1 means perfect classification. Average AUC value is 0.816 for ephys sessions, 0.827 for training sessions, and 0.795 for habituation sessions.
Following Piet et al, we use the absolute value of the change in model evidence after removing either the visual or timing strategies as ‘the ‘visual index’’ and “timing index’’, and deem the difference between the visual index and timing index to be the “strategy index” (Fig 1 a-b).
Strategy index during recording positively correlates with total session number. I first fit the model on the 56 ephys sessions and obtain their strategy indices. As concluded in the 2P dataset, higher values of the strategy index tend to result in a higher number of earned rewards, and is positively correlated with a higher D-prime (Fig 1 c-d). The strategy index also shows a negative correlation with number of abortion trials and false alarms, suggesting that in those sessions the mice demonstrate good task performance (Fig 1 e-f). Before looking into the training sessions, We tested whether mice that have higher strategy index during ephys sessions needed less training session before they became hand-off ready, but only found little correlation (Fig 2 a). Interestingly, when we plot the sessions’ strategy index against the total session number of sessions animal had experienced, we found a positive correlation, suggesting that more training contributes to the preference for the optimal strategy in recording sessions (Fig 2 b). Taken both findings together, it suggests that the strategy index in the recording sessions is more of a result due to training, rather than the mice’s initial preferences.
Fig 2. Strategy index v.s. Trained sessions.