Olfactory CAPTCHA
In order to test responses to novel background odors, we used the olfactory equivalent of the visual CAPTCHA(Von Ahn et al. 2003) employed for human verification tasks, which also serves as a benchmark for testing image recognition algorithms(George et al. 2017). CAPTCHA requires a user to identify letters (known targets) superimposed on distracting stimuli (novel backgrounds). Animals exposed to a novel background as part of an olfactory CAPTCHA task have to identify target odors present permitting an unambiguous evaluation of olfactory function. In contrast, it is harder to evaluate olfactory function of animals freely exploring novel background odors(Brunner et al. 2015; Gordon et al. 2016) because there is no task assigned. We hypothesize that WT mice will be able to identify known odors in the presence of novel background odors, since this situation mimics what they experience in the wild and that mouse models of autism will not, because CAPTCHA requires generalization and robust responses to novelty, which are affected in ASD.
Mixtures of target and background odors were divided into a training set and a test set. Animals were trained with the training set and their performance was evaluated with the test set. The training set consisted of 16 mixtures of 3 odors, 1 target odor that solely determined reward availability (4 possible odors), 1 contextual background odor (4 possible odors), and 1 fixed background, (s)-(-) limonene. Both (s)-(-) limonene and the contextual background odors were presented at a higher concentration (0.1% of saturated vapor) than the target odors (0.025%). The contextual background odors and the (s)-(-) limonene were the known background odors. The test set also consisted of mixtures of 3 odors that included the same target odors and contextual background odors, but the (s)-(-) limonene was replaced by one of 11 possible novel background odors, which were also presented at a higher concentration (0.1%). We used a go/no-go design, with two of the target odors indicating the presence of a water reward upon licking the water tube. We chose a relatively complex training set (eight go mixtures and eight no-go mixtures) to promote generalization. We chose a go/no-go licking task because it is easy for WT mice to learn(Komiyama et al. 2010) and it can also be learned by the Cntnap2−/− mouse model of autism(Rosario et al. 2020).
Olfactory stimulus consisted of weak target odor embedded in strong background odors
We built a triple air-dilution odor machine to deliver mixtures of three odors (see Methods). We used intrinsic optical imaging of WT mice to quantify the pattern of glomerular activation of the target and background odors at the same concentrations used in the task (see Fig. 1A and Sup. Figure 1). The four target odors had lower levels of average glomerular activation compared to the four contextual background odors (see Fig. 1B,C). The 11 novel background odors had intermediate levels of average glomerular activation between the target and the contextual background odors. Glomerular patterns of mixtures of odors were dominated by the background odors (see Fig. 1D) resulting in go and no-go mixtures with known odors that were quite similar (correlation = 0.85 ± 0.07) but within range of previous stimuli that rodents were able to discriminate (Gschwend et al. 2015; Uchida and Mainen 2003; Abraham et al. 2004).
WT mice could identify target odors among novel background odors
Four head-fixed water-deprived WT mice were trained to detect the presence of target odors and were rewarded with water if they licked in response to the target go-odors (see Fig. 1E,F). If they licked before the target go-odors appeared or if they licked in response to no-go target odors, odor delivery was stopped and the mice were given a time-out. Mice were trained for ~ 9 days (see training protocol) to detect targets among known backgrounds. Once animals displayed consistent performance at the final concentration (> 80% correct for more than 50 trials), trials with novel background odors were introduced. Presentation of the novel background odor and the contextual background odor preceded the onset of the target odor by 0.75 s and 1.5 s respectively (see Sup. Figure 2). This asynchronous odor presentation emulated a situation where the animal is in a novel environment when the target appears and should be easier for the animals because of sensory adaptation to the novel background(Verhagen et al. 2007). We used a long inter-trial interval to avoid receptor adaptation effects and to permit the airflow to clean the odor delivery system: the separation between trials was 30.4 ± 3.3 seconds(1885 trials).
The trials with different novel background odors were separated by five or six trials with known background odors. These trials with known background odors were included to maintain the animal’s motivation with an easier task and to reinforce the identity of the target odors. Each presentation of a novel background odor was separated by other presentations with the same background odor by ~ 25 minutes. Each novel background odor was presented, at most, four times per day on two separate days, giving a maximum of eight presentations per novel background odor per animal to prevent learning of the novel background odors.
WT mice performance for known background odors was high (87.4%, n = 1563 trials), and significantly higher than the 50% chance level (p = 1.24 e-215, binomial test) (see Fig. 2A). WT mice almost never failed to detect the targets (0.7% of the trials were false rejections), as with other go/no-go behavioral tasks(Rokni et al. 2014; Kuchibhotla et al. 2019). WT mice were able to solve the olfactory CAPTCHA and successfully identify targets in the presence of novel background odors at higher than chance levels (76.9%, n = 334 trials for the novel background trials, p = 3.08 e-24, binomial test). Moreover, WT mice were able to identify target odors among novel background odors, even on the first presentation on day 1 (see Fig. 2B), with WT mice performing at 79% (p = 6.4e-5, binomial test, n = 44 trials, 4 animals). There was no systematic increase in performance as the animals familiarized themselves with a novel background odor, and there was no significant positive linear correlation between a background odor presentation number and performance (r = -0.48, p = 0.22). Interestingly, when the novel background odors were used, animals had a significant increase in the number of false rejections (5% from 0.7%, p = 6.91e-7, Fisher exact test) suggesting increased difficulty in detecting the target odor in novel backgrounds compared to known backgrounds (see Fig. 2C, D).
To further test whether familiarity with the novel background odors increased performance, we trained a different cohort of five WT mice in the same task (see Supplementary material 1). This new cohort had been exposed to 5 of the 11 novel background odors in their home cages for 30 minutes over 6 days before behavioral training started. The performance of the new cohort of animals in response to the previously exposed 5 odors was 79.6% (211 trials, see Supplementary Material), which was not significantly different (p > 0.1, Fisher exact test) from the original group performance on these odors (78.8%, 132 trials, 4 animals). Both exposure to the same novel background odor during the task and longer exposure times outside the task context did not produce systematic performance increases, indicating that WT mice employed an algorithm that did not require previous knowledge of the background odor to detect target odors among novel background odors.
WT recognized the novel background odors after a single presentation
We wanted to determine how quickly the mice familiarized themselves with a novel background odor by measuring the sniffing response. Sniffing rate increases as an animal explores novel odors(Wesson et al. 2008; Verhagen et al. 2007). We monitored non-invasively animal sniffing during the olfactory CAPTCHA task using a flow sensor attached to the odor delivery tube(Bolding and Franks 2017) (see Fig. 3A,B). On the first presentation of the novel background odors, the sniff rate increased compared to the trials with the known background odors (1.37 ± 0.36 extra sniffs per second, p = 0.00067 binary test) indicating that the novel odors were perceived as novel. The sniff rate was also elevated on the first presentation on the second day (day 2, 0.95 ± 0.34, p = 0.007, binary test). Interestingly, average sniff rates dropped to rates similar to known background trials (p > 0.1) on further exposures in a session, indicating that the animals recognized the novel background odor (see Fig. 3C). This was quite remarkable, because the novel background odors were presented with different targets and contextual odors, and subsequent presentations of the same novel background odor were separated by at least 25 minutes.
We also quantified the animal familiarity to the novel odors using the number of sniffs that it took from target onset to lick response. During the known background trials, the animals took 2.93 ± 0.06 sniffs before licking (842 responses). Animals took a larger number of sniffs (5.8 ± 1.11, 20 responses) on the first exposure of a novel background and it was significantly higher than the number of sniffs for known backgrounds (Wilcoxon rank-sum test, p = 0.0023). Consistently with increasing familiarity after the first exposure and reduction of exploratory behavior, the number of sniffs before licking fell after the first presentation of a novel odor to a value that was similar to the number of sniffs in the presence of known backgrounds. Animals increased their sniff rate and number for novel odor but this increase disappears after a few exposures as they recognized the novel background odor. Surprisingly, there was no increase in performance even when WT mice recognized the novel background odor, suggesting that WT mice used an algorithm that required no knowledge of the background odor.
WT mice did not require background adaptation to detect odors in novel environments
In the first set of experiments, the target odor appeared 0.75 s after the onset of the novel background odor. Due to adaptation to the early-appearing novel background odor, the target could become more salient, enhancing its detectability. In order to directly test whether sensory adaptation was necessary for identifying odors in novel environments, we trained a different group of five animals in a condition where the target and background odors were presented synchronously (synchronous case), rather than the target odor appearing after the novel background odor (asynchronous case) (see Fig. 4A). This synchronous odor presentation emulated a situation where the animal is suddenly presented with a mixture that includes a novel odor. In the synchronous case, both the target and the novel background odors endured the same process of adaptation, reducing the impact of differential adaptation to the background odor. For the known background odors, animals in the synchronous case performed at a high level (88.5%, n = 1,191 trials, p = 4e-176, binomial test), similar to that in the asynchronous case.
WT mice were able to identify the target odors with the synchronously presented novel background odors at higher than chance levels (p = 7.4e-21, binomial test, see Fig. 4B). Although the mice had an increase in false rejections (see Fig. 4C), the total performance for the synchronous presentation (77.2%, n = 277 trials, 5 animals) was almost identical (Fisher exact test, p > 0.9) to the asynchronous case performance for the previous group of 4 animals (76.9%). Sensory adaptation did not make a significant contribution to odor identification in novel environments. Animals identified the target odors among novel backgrounds (74.3%, 39 trials) at higher than chance levels (p = 0.001, binomial test), even on the first presentation of a novel background odor (see Fig. 4D). There was no significant correlation between the odor presentation number and the performance of the mice (r = -0.02, p = 0.95); hence, as with the asynchronous case, there was no improvement of performance with further exposures to a novel background odor.
Animals performing the synchronous task also reacted to the first presentation of a novel background odor by increasing their sniffing rate (see Fig. 4E), and this response was also adapted after two presentations, but the number of extra sniffs before a response did not increase for novel background odors (see Fig. 4F, Supplementary Material 3, and Sup. Figure 4).
Fast sniff rates correlated with increased performance on novel background odors
We wondered whether rapid sniff rates would correlate with increased performance as described for other olfactory tasks (Kepecs, Uchida, and Mainen 2007). When animals increased their sniff rates, inhalations become shorter(Jordan et al. 2018). Brief inhalations correlated with higher performance only on novel background odors, suggesting that rapid inhalation induced adaptation(Verhagen et al. 2007) might contribute to odor identification in novel environments (see Supplementary material 2 and Sup. Figure 3); however, the lack of difference in performance between the synchronous and asynchronous case showed that adaptation is not the only mechanism used in odor identification in novel olfactory environments, and other circuit mechanism might play a role.
Linear classifier matched WT mice’s performance in novel background odors
WT mice performance in odor identification for mixtures of known odors can be matched by linear classifiers, trained using calcium imaging of glomerular activation patterns of the dorsal surface of the olfactory bulb(Mathis et al. 2016). The very simple circuit implementation of a linear classifier is less likely to be affected in mouse models of autism. We wanted to determine whether linear classifiers, trained with glomerular activation images from mixtures from the training set, were sufficient to generalize and match WT mice performance in response to novel background odors. Could a linear classifier predict the WT mice’s performance in response to individual novel background odors?
We compared the behavioral performance of nine WT mice for individual novel background odors with the performance of a linear classifier trained using intrinsic imaging (see Fig. 5A). We measured the glomerular activation patterns in response to individual odors in WT mice and created virtual odor mixtures, including different levels of Gaussian noise and an experimentally measured saturation nonlinearity(Mathis et al. 2016) (see Sup. Figure 5). The linear classifier was trained using the 16 mixtures of the training set to determine if the target odor in the training mixture was a go or a no-go. To evaluate the performance for each of the 11 individual novel background odors, the linear classifier was tested using the 16 mixtures where (s)-(-)limonene was replaced by an individual novel background odor. The linear classifier trained with intrinsic images was able to identify odors at a level that matched or exceeded the performance of the animals in novel backgrounds similar to measures using glomerular calcium imaging(Mathis et al. 2016). However, the performance of the linear classifier for a given novel background odor was not a good predictor of the WT mice’s performance for that novel background odor (see Fig. 5B,E,F).
Nearest neighbor classifier (NNC) predicted WT mice performance in response to novel background odors
We wondered if WT mice performance for individual novel background odors could be predicted using a non-linear classifier. The NNC is a non-linear classifier that determined which of the 16 training mixtures that included (s)-(-) limonene was the best match to the observed mixture with a novel background odor. If both the best match and the observed mixture contained the go (or no-go) target odors, the classification was a success (see Fig. 5C). We trained the NNC using the same imaging data from the dorsal surface as used for the linear classifier. The NNC correlated significantly with the mice behavior in response to individual novel background odors (see Fig. 5D,E,F) demonstrating that it was possible to predict WT mice performance using dorsal glomerular data.
The virtual odor mixtures used did not capture other types of nonlinear receptor activation in response to odor mixtures (Zak et al. 2020; Inagaki et al. 2020) that might affect an algorithm performance. To control for these potential confounders, we used imaging data from real mixtures to train the linear classifier and the NNC. The comparison with the behavioral responses of nine animals yielded similar results as the responses of the classifiers calculated using the virtual odor mixtures: the NNC predicted WT mice performance in response to individual novel background odors, whereas the linear classifier was not a good predictor (see Fig. 5G,H). Animals used an algorithm that was better approximated by the NNC than the linear classifier. The NNC requires a recurrent inhibitory circuit in olfactory cortex to perform the winner-take-all computation, and its performance might be affected by reduced inhibition in mouse models of autism (Coghlan et al. 2012).
WT mice responses were delayed in novel backgrounds
We wondered if mice might be performing a different computation when challenged with novel background odors and whether this extra computation might be reflected in increased reaction times, as shown for other behavioral tasks(Abraham et al. 2004). Previous work (Rokni et al. 2014) had shown very little increase in response latency with task difficulty for animals identifying odors among known background odors.
WT mice had slower reaction times for novel background odors compared to known background odors for both the synchronous and the asynchronous task. The lick reaction time for the asynchronous case for novel background odors was 651.3 ± 37 ms and it was significantly slower (n = 151 lick responses, 4 animals, p = 0.013, Wilcoxon rank-sum test) than the response for the known background odors (571.4 ± 11.9 ms, 812 lick responses, see Fig. 6A). The reaction time for novel background odors for the synchronous case was 844.8 ± 46.3 ms, and it was also significantly slower (110 lick responses, 5 animals, p = 0.026, Wilcoxon rank-sum test) than the reaction time for known background odors(739.8 ± 16.5 ms, n = 577 trials). This increase in response time in the CAPTCHA task suggested that identifying odors among novel background odors might further involve recurrent computations compared to odor identification in known background odors.
WT mice performance was less sensitive to amount of training data than the NNC
The NNC was a good predictor of WT mice performance in novel background odors. The NNC requires a training set that includes an appropriate match to a new mixture; otherwise, the NNC performance might decrease (see Fig. 6B). To confirm this, we reduced the number of training examples, from 16 training mixtures (4 contextual background odors x 4 target odors x (s)-(-)limonene) to 8 mixtures (see Fig. 6C). The task became harder because not only the test mixture included a novel background odor, but also the test mixture was a novel combination of known contextual background and target. As expected, the performance of the NNC decreased significantly (p = 0.017, single tailed fisher exact test) from 79.3% when trained with the full set to 70.4% when trained with the restricted set. The performance of a linear classifier also significantly decreased (p = 0.017, single tailed fisher exact test) from 76.7% for the full set to 64.7% for the restricted set. If animals employed these algorithms, their performance would be affected by reducing the number of training examples.
In order to test this, we trained a new batch of three WT mice using the restricted set with asynchronous presentation of the odors (see Fig. 6D). The WT mice performance on novel backgrounds was 74.12% (228 trials) and not significant lower (p = 0.25, single tailed fisher exact test) than the previous group of animals trained with the full set of training mixtures (76.9%, 334 trials, 4 animals, asynchronous task). WT mice used the training data more efficiently than the linear classifier or the NNC and are able to generalize to novel combinations of known odors in the presence of a novel background odor.
Sparse deconvolution algorithms(Koulakov and Rinberg 2011; Grabska-Barwińska et al. 2017; Li and Hertz 2000; Otazu and Leibold 2011) use the training data more efficiently than the NNC and might be a plausible type of algorithm used by WT mice. Sparse deconvolution algorithms decompose the signal into contributions selected from a very large dictionary of known odors, while minimizing the number of odors used, permitting generalization to any combination of known odors. We could not directly test the performance of these algorithms for individual novel background odors because the performance in response to a novel background odor depends on the glomerular representation of all odors that an animal knows, which is hard to estimate. However, we performed simulations (see Supplementary material 4) to determine whether a standard sparse representation algorithm, the Lasso(Tibshirani 1996), could identify the target odors in the presence of novel background odors that are not part of a simulated dictionary. The Lasso solved the behavioral task at a performance of > 90% suggesting (see Sup. Figure 6) that deconvolution algorithms are a plausible hypothesis as the algorithms used by WT mice. The recurrent computations required to generalize, given the reduced size of the training set, might make these algorithms more sensitive to the diverse circuit defects observed in mouse models of autism.
Cntnap2 −/− mouse model of autism odor detection was selectively affected in novel background odors
We tested the Cntnap2−/− mouse model of autism(Poliak et al. 2003) (JAX Stock No: 017482) because Cntnap2−/− mice have either equal(Levy et al. 2019) or better(Peñagarikano et al. 2011) performance than WT mice in simple olfactory tasks, suggesting a functional olfactory system. Nevertheless, Cntnap2 is expressed in olfactory receptor neurons(Gordon et al. 2016) and its absence might reduce neural excitability, affecting glomerular representations. However, the glomerular activation patterns of WT mice and Cntnap2−/− mice were very similar (see Fig. 7A). We compared odor responses across 190 odor pairs and found a significant correlation in similarity between the odor pairs calculated using WT and Cntnap2−/− mice imaging data (R = 0.46, p = 1.774e-11, Pearson linear correlation) (see Fig. 7B). We also compared the performance of the linear classifier trained using the 16 mixtures with known background odors, using WT and Cntnap2−/− imaging data. The performance of the classifiers in mixtures that included novel backgrounds (176 mixtures) was almost identical for the WT and Cntnap2−/− data (see Fig. 7C), indicating that Cntnap2−/− and WT mice had similar responses at the sensory periphery.
We trained four Cntnap2−/− mice using the restricted training set with asynchronous presentation of the odors. Animal training proceeded similarly to that for the WT mice, reaching similar performance levels on well-known background odor after ~ 9 days of training and indicating that learning was not affected in the context of our task. The Cntnap2−/− mice performance on the standard background odors, just before the presentation of the novel odors started, was 85.5% (220 trials) and not significantly different (p = 0.75, Fisher exact test) from the performance of the WT mice (87.1%, 140 trials, 3 animals) (see Fig. 7D). However, for novel background odors, the Cntnap2−/− mice performance dropped to 61.13% (283 trials), which was significantly lower (p = 0.00242, Fisher exact test) than the performance of WT mice (74.1%, 228 trials) (see Fig. 7E).
The Cntnap2−/− mice poor performance might have been due to their hyperactivity(Peñagarikano et al. 2011), possibly causing uncontrolled licks in response to the novel background odors. However, Cntnap2−/− mice did not make many early licks nor false detections; in fact, the main type of error in the presence of novel background odors were false rejections; that is, the mice failed to lick in response to the target go-odor (see Fig. 7F).
Lack of odor exploration in Cntnap2−/− mice (Brunner et al. 2015; Gordon et al. 2016) might explain their poor performance for novel background odors; however, Cntnap2−/− mice strongly increased their sniffing rate in response to novel background odors, similar to WT mice, suggesting that Cntnap2−/− mice explored the novel background odor(see Fig. 7G ). Although Cntnap2−/− mice were reluctant to lick for target odors among novel background odors, the lick reaction time for the asynchronous case for the novel background odors was 664.8 ± 68 ms and not significantly slower (n = 49 lick responses, 4 animals, p = 0.85, Wilcoxon rank-sum test) than the response time for the known background odors (632.5 ± 17.5 ms, 462 lick responses). Cntnap2−/− mice did not delay their responses for novel background odors as WT mice did, suggesting a deficit in recruiting recurrent computations necessary to perform the task (see Fig. 7H). The circuitry required for robust sensory identification in novel environments might be selectively affected in this mouse model of autism, while performance of well-rehearsed mixtures was spared.