Introduction

Autism is a neurodevelopmental condition that affects approximately 1 in 31 children (Shaw et al., 2025). Autistic individuals display differences in social interaction and communication, repetitive behaviors, intense or special interests, and altered sensory experience as core symptoms (American Psychiatric Association, 2022). Sensory alterations strongly affect the way autistic individuals interact with their environment. According to the DSM-5-TR, these are broadly expressed as “hyper- or hyporeactivity to sensory input or unusual interests in sensory aspects of the environment” and are reported in 90% of autistic individuals, spanning all sensory modalities (Robertson and Baron-Cohen, 2017).

Amongst sensory modalities, touch is the first sense to develop and plays a fundamental role in the exploration of the environment, the definition of “self”, and early social bonding (Bremner and Spence, 2017). Tactile processing alterations have a significant impact on daily functioning and are thought to contribute to repetitive behaviors and social interaction difficulties in autism (Foss-Feig et al., 2012; Green et al., 2018; Thye et al., 2018; He et al., 2021a; Zhai et al., 2023; Carati et al., 2024). However, clinical studies focusing on tactile discrimination have yielded controversial results (Zetler et al., 2019), likely due to the substantial heterogeneity of tactile processing alterations in autism (Mikkelsen et al., 2018) and inconsistent terminology across studies (He et al., 2023). In contrast, numerous studies in the visual and auditory domains (Plaisted et al., 1998a; O’Riordan and Plaisted, 2001; O’Riordan et al., 2001; Bonnel et al., 2003; O’Riordan and Passetti, 2006; Samson et al., 2006; Rotschafer, 2021) have supported enhanced sensory processing in autism, as proposed by several influential models of the condition (Plaisted, 2001; Frith, 2003; Mottron et al., 2006).

Although clinical studies highlight the heterogeneity of tactile responses in autism, variability in experimental protocols and stimulus types has made it difficult to draw cohesive conclusions. Tactile discrimination of vibrotactile stimuli delivered to the glabrous skin of the fingers was reported reduced (Puts et al., 2014; Espenhahn et al., 2023), or intact (He et al., 2021a; Asaridou et al., 2022) in autistic individuals. Similarly, studies of sharp-dull discrimination have reported both reduced (Abu-Dahab et al., 2013) and typical performance (Minshew et al., 1997; Minshew and Hobson, 2008), as have studies of form discrimination (Minshew and Hobson, 2008; Demopoulos et al., 2015; Failla et al., 2017). Translational approaches employing mouse models of autism in combination with psychometric measures from well-controlled behavioral tasks may help clarify whether these inconsistencies reflect genuine inter-individual variability or arise from methodological differences.

While tactile discrimination has been explored in several clinical studies, it remains underexplored in mouse models of autism. Rodent studies leveraging innate responses have reported atypical whisker-dependent texture discrimination (Balasco et al., 2021, 2022; Mattioni et al., 2024) and forepaw-dependent roughness discrimination (Orefice et al., 2016). However, these studies have not employed psychophysical methods to quantify tactile discrimination, restricting the translatability of their findings. Implementing psychophysics in animal models might clarify previous inconsistencies between studies in mice (Orefice et al., 2016) and humans (O’Riordan and Passetti, 2006).

Sensory alterations are thought to impact cognition in autism (Haigh, 2018). However, sensory perception is shaped not only by the physical properties of stimuli but also by higher-order cognitive processes such as attention and categorization—both of which are altered in autism (Allen and Courchesne, 2001; Church et al., 2010; Gastgeb and Strauss, 2012; Cascio et al., 2015; Barbot et al., 2018; Rong et al., 2021a). Moreover, autistic individuals demonstrate differences in how prior beliefs and expectations (priors) are integrated with sensory input during perceptual decision-making. They tend to rely less on contextual priors (Amoruso et al., 2019) and more heavily on prior choices (Feigin et al., 2021), and often form priors that are imprecise or inflexible (Zaidel et al., 2015; Sapey-Triomphe et al., 2021b). Notably, variability in the use of priors has also been linked to individual differences in sensory responsivity (Sapey-Triomphe et al., 2021a). Thus, cognitive differences may compound or interact with primary sensory processing alterations, contributing to the distinct perceptual experiences observed in autism.

In this study, we investigated the interplay between tactile discrimination and cognitive processes during perceptual decision-making. To enhance translational relevance and address questions of inter-individual heterogeneity, we developed a forepaw-based 2-Alternative Choice task for mice, using vibrotactile stimuli analogous to those used in human psychophysical protocols. Leveraging this task, we explored how stimuli of varying salience are discriminated in the Fmr1^-/y mouse model of autism and disentangled stimulus-driven from cognitively modulated-tactile responses.

Our results revealed salience-dependent alterations in both tactile processing and cognitive modulation in Fmr1^-/y mice, with a decreased influence of cognitive processes on tactile perception. During training, Fmr1^-/y mice exhibited heightened choice consistency bias during low-salience trials, which contributed to reduced performance during perceptual learning. In contrast, trained Fmr1^-/y mice displayed enhanced discrimination of low-salience stimuli, but a diminished facilitatory effect of categorization on across-category discrimination. Despite their increased discrimination sensitivity at the low-salience range, Fmr1^-/y mice showed attention deficits specifically for these stimuli under high cognitive load conditions. While a strong choice consistency bias was observed during the testing phase for both WT and Fmr1^-/y mice, integration of recent sensory history was selectively disrupted in Fmr1^-/y mice, such that past sensory information no longer informed ongoing decisions.

These findings uncover distinct, salience-dependent cognitive alterations that shape tactile perception in the Fmr1^-/y model, advancing our understanding of tactile perception in autism. Importantly, our results indicate that the variability observed in clinical studies may arise not solely from primary sensory alterations but also from differences in cognitive context, attentional demands, or task complexity—offering a potential framework to reconcile prior inconsistencies in the field.

Results

Fmr1^-/y mice show intact learning rate and trajectory during perceptual learning

To investigate perceptual and cognitive alterations in a highly translational approach, we developed a forepaw-based 2-Alternative Choice perceptual decision-making task in the Fmr1-KO mouse model of autism. In this task, mice learn to discriminate between vibrotactile stimuli that differ in perceptual salience, which we define as the prominence of a stimulus based on its physical amplitude. To ensure that all mice could reliably detect the applied stimuli, we selected amplitudes above the perceptual thresholds previously established for both wild-type (WT) and Fmr1^-/y male mice (Semelidou et al., 2024). Accordingly, all stimuli used in the present study (40 Hz, ≥12 µm) were well above the detection threshold (10 Hz, 4.46 µm for WT; 10Hz, 7.29 µm for Fmr1^-/y mice), allowing us to specifically probe discrimination, categorization, and decision-making rather than sensory detection.

During each trial, head-fixed, water-controlled mice received a 500 ms vibrotactile stimulus to the forepaw at 40 Hz with either high amplitude (26 μm) or low amplitude (12 μm). Stimulus salience was defined in a relative manner: across animals, the lower-amplitude stimulus elicited consistently higher miss rates, indicating reduced perceptual strength and attentional engagement within the suprathreshold range. We therefore refer to the 12 μm stimulus as “low salience” and the 26 μm stimulus as “high salience.” Mice were trained to report high-salience stimuli by licking the right lick-port, and low-salience stimuli by licking the left port within a 2-s response window following stimulus delivery. Incorrect choices resulted in a 5-s timeout (Fig. 1A).

Figure 1.

During early training sessions, performance was low, characterized by a low proportion of correct and a high proportion of incorrect choices, which either remained stable or gradually improved over the course of the session (Fig. S1A-left). Licking patterns were also inconsistent during these early sessions and poorly time-locked to the response window (Fig. S1A-right). In contrast, late training sessions showed consistently high correct choice rates and low incorrect choice rates (Fig. S1B-left), along with stable licking patterns during the response window for both high- and low-salience trials (Fig. S1B-right).

Learning difficulties and altered learning trajectories have been reported in autistic individuals and replicated in several mouse models of autism, including Fmr1^-/y mice (Arnett et al., 2014a; Goel et al., 2018; Mercado et al., 2020; Khachadourian et al., 2023; Shenouda et al., 2023; Mol et al., 2024). To assess whether Fmr1^-/y mice exhibit alterations in learning rate in our task, we compared the number of days required to reach the predefined learning criterion (>70% correct choices for each stimulus salience across three consecutive training days). Our results showed no significant differences between Fmr1^-/y mice and their WT littermates in the total number of training days to reach criterion (Fig. 1B; T test: t = 0.438, p = 0.667) or in the duration spent at each stage of the training protocol (Fig. S1C-D, see Methods; Blocks: T test: t = 1.259, p = 0.224; Mixed: T test: t = -0.185, p = 0.855). Similarly, both genotypes required a comparable number of days to learn the association between stimulus salience and the correct lick-port (Fig. 1C-D; High salience: Mann-Whitney U test: U=45.5, p = 0.788; Low salience: Mann-Whitney U test: U = 38.5, p = 0.422).

To capture potential genotype differences in the learning dynamics during training, we analyzed the performance of all mice that reached the learning criterion during the training phase with mixed low- and high-salience trials. Both WT and Fmr1^-/y mice showed similar learning dynamics, with a slow, gradual improvement across days (main effect of day: β = 0.007, SE = 0.004, z = 1.882, p = 0.060, 95% CI [0.00, 0.01]) (Fig. S2A-top). Consistent with the extended training duration required for the task, learning slopes were shallow and comparable between genotypes (Fig. S2B; Mann-Whitney U test: U=38.0, p=0.536). Importantly, the Genotype × Day interaction was small (β = 0.002), indicating no detectable genotype differences in learning rate across animals.

These similar learning trajectories were accompanied by comparable baseline performance on the first day of training across genotypes (Fig. S2A-top, S2C; genotype effect: β = −0.087, SE = 0.066, z = -1.328, p = 0.184, p = 0.184, 95% CI [-0.22, 0.04]) and similar within-animal day-to-day performance variability (Fig. S2A-bottom; Fig. S2D; T test: t=-0.302, p=0.767).

We next assessed whether performance differed during an intermediate stage of learning, defined for each animal as the middle three days of its training period. Here, both genotypes again showed comparable performance (Fig. S2E, slope; T-test: t = 0.801, p = 0.437) with similar correct-choice rates (Fig. S2F; T-test: t=1.240, p=0.228), indicating no genotype-specific differences during intermediate learning.

Together, these results show that Fmr1^-/y mice display an intact learning rate and trajectories in this 2-AFC perceptual decision-making task.

Fmr1^-/y mice display impaired performance on low-salience trials during perceptual learning

Given that comparable learning trajectories and training duration to reach the criterion do not exclude the presence of performance alterations, we next evaluated additional behavioral metrics during training with mixed high- and low-salience trials. Specifically, we analyzed overall discrimination sensitivity (d’) as well as correct and incorrect choice rates throughout training to determine whether Fmr1^-/y mice exhibited altered task performance despite intact learning progression.

To quantify mice’s sensitivity in discriminating high-versus low-salience stimuli throughout training, we used the signal detection theory and computed the sensitivity index d′. This measure captures stimulus discriminability by comparing normalized hit rates (rightward licks on high salience trials) and false alarm rates (rightward licks on low salience trials). Fmr1^-/y mice showed a strong trend with a large effect size toward reduced overall task sensitivity compared to WT littermates (Fig. 1E; T-test: t=-2.083, p=0.052, Hedges’ g=0.90), indicating diminished perceptual performance during training. These results demonstrate that, despite acquiring the task at a comparable rate (Fig. 1B-D) and with similar learning trajectories (Fig. S2), Fmr1^-/y mice display a trend for diminished performance during salience discrimination training.

To complement our sensitivity (d′) analysis, we also quantified correct and incorrect choice rates, allowing us to assess both rightward and leftward lick responses. Correct choice rates were comparable between Fmr1^-/y and WT mice (Fig. 1F, T-test: t=1.064, p=0.307) with similar Hit rates for both high- and low-salience stimuli (Fig. S1E–F, High salience-T-test: t=0.721, p=0.480; Low salience-T-test: t=1.060, p=0.303). In contrast, Fmr1^-/y mice exhibited a significantly higher rate of incorrect choices (Fig. 1G; T-test: t=-2.099, p=0.050; Hedge’s g=0.90), an effect that was specific to low-salience trials (Fig. 1H; Mann–Whitney U = 23.0, p = 0.048; Hedge’s g = 0.97), while high-salience error rates were similar across genotypes (Fig. 1I; T-test: t=-1.072; p=0.298). After outlier analysis and exclusion of one Fmr1^-/y mouse, the genotype difference was reduced to a trend, though with a large effect size, for incorrect choices (T-test: t=-1.803; p=0.089, Hedge’s g = 0.79) and for incorrect choices during low-salience trials (Mann–Whitney U = 23.0, p = 0.079; Hedge’s g = 0.89).

These results suggest a modest reduction in task performance in Fmr1^-/y mice, primarily driven by increased errors on low-salience trials.

Fmr1^-/y mice show higher choice consistency bias during training

Perceptual learning can further reveal potential alterations in decision strategy, manifested as changes in response bias (decision criterion c) or expectation bias (priors). To examine whether mice exhibited stereotyped licking toward one of the two lickports—a potential indicator of response bias—we calculated the decision criterion. This analysis revealed no significant differences between Fmr1^-/y mice and WT littermates (Fig. 2A; Mann–Whitney U: U=65.000, p=0.254), with both groups showing a bias toward the left lickport (WT: Wilcoxon signed-rank test W=5.000, p=0.039; Fmr1^-/y: Wilcoxon signed-rank test W=10.000, p=0.042).

Figure 2.

Since mice did not exhibit differences in stereotyped licking, we could rule out simple response biases and further assess whether their decisions are influenced not only by the sensory features of stimuli but also by previous experiences that shape expectations (i.e., priors). To investigate the impact of prior experience on current choices, we analyzed behavioral responses on a trial-by-trial basis. We first examined how a rewarded choice in the previous trial influenced the performance in the current trial (Fig. 2B). Fmr1^-/y mice exhibited an increased rate of incorrect responses to low-salience stimuli when these were preceded by a rewarded high-salience stimulus (Fig. 2C; T-test: t=-2.352, p=0.030). In contrast, no genotype difference was observed when a high-salience stimulus followed a rewarded low-salience trial (Fig. 2D; T-test: t=0.603, p=0.554).

We next tested whether Fmr1^-/y mice were more influenced by their prior choices. To assess this, we compared the rates of repeated prior choices in current low- and high-salience trials, controlling for the total number of correct and incorrect responses. Our analysis revealed higher previous choice repetition during low-salience trials in Fmr1^-/y compared to WT mice (Fig. 2E; T-test: t=-2.618, p=0.017), but a comparable impact of prior choices during high-salience trials (Fig. 2F; T-test: t=1.215, p=0.240).

To verify that the observed differences in performance between WT and Fmr1^-/y mice reflected differential reliance on priors rather than differences in prior strength per se, we quantified the strength of the priors induced by high- and low-salience stimuli. Specifically, we calculated the rate of choices that were repeated following a low- or high-salience trial, correcting for the overall rate of correct and incorrect responses. Our results showed similar prior strength between groups, both for high-salience (Fig. S3A; T-test: t=-1.705, p=0.105) and for low-salience stimuli (Fig. S3B; Mann–Whitney U: U=38.500, p=0.425).

Together, these results indicate an increased choice consistency bias during low-salience trials, which contributes to performance differences during perceptual learning in Fmr1^-/y mice.

Fmr1^-/y mice do not show attention deficits during perceptual learning

Stimulus salience is closely linked to attentional processes (Kerzel and Schönhammer, 2013; Forschack et al., 2022; Bouvier et al., 2023), and attentional alterations are well-documented in autistic individuals (Allen and Courchesne, 2001; Grubb et al., 2013; Barbot et al., 2018; Licznerski et al., 2020; Rong et al., 2021b). Given that Fmr1^-/y mice exhibit salience-dependent performance differences during training (Fig. 1-2), we asked whether their reduced task performance might also reflect underlying attention deficits. To assess attention, we analyzed the trials in which the animal failed to respond to the stimulus (Miss trials) (Fig 1A). Within-genotype analysis confirmed studies linking salience and attention in humans (Kerzel and Schönhammer, 2013), and validated that low amplitude stimuli are less salient, yielding significantly higher Miss rates in both groups (Fig. S3C; WT: Wilcoxon signed-rank: W=0.000, p=0.004; Fmr1^-/y: Paired T-test: t=-4.113, p=0.002). However, there were no genotype differences in Miss rates for either high- or low-salience stimuli (Fig. S3D-E; Miss rate High: Mann–Whitney U: U=51.000, p=0.939; Miss rate Low: Mann–Whitney U: U=50.000, p=1.000). These data demonstrate that low-salience stimuli are less effective at capturing attention in both Fmr1^-/y and WT mice, and that the reduced task performance of Fmr1^-/y mice during perceptual training cannot be attributed to an attentional deficit during low-salience trials.

Trained Fmr1^-/y mice show enhanced discrimination of low-salience stimuli

Altered tactile discrimination has been reported in clinical studies, though findings are inconsistent (reviewed in (Zetler et al., 2019)), and has also been observed in mouse models of autism (Orefice et al., 2016; Balasco et al., 2021, 2022; Mattioni et al., 2024). To assess tactile discrimination in our study, we incorporated the original training stimuli along with 6 intermediate amplitudes, spaced 2 µm apart. Stimuli ranging from 20 to 26 µm (20, 22, 24, 26 µm) were categorized and rewarded as high-salience stimuli, while those from 12 to 18 µm (12, 14, 16, 18 µm) were treated as low-salience stimuli (Fig. 3A). Only animals that successfully acquired the task during the training phase were included in the tactile discrimination test. Fmr1^-/y and WT mice showed similar performance during the last three days of training, with no significant differences in task sensitivity (d’; T-test: t=0.800, p=0.438; Fig. S4A), rates of correct (Fig. S4B-C; High: T-test: t=0.442, p=0.666; Low: T-test: t=0.697, p=0.498) and incorrect responses (Fig. S4D-E; High: T-test: t=-0.680, p=0.509; Low: T-test: t=-1.069, p=0.305), decision bias (Fig. S3F; criterion c; Mann–Whitney U: U=28.000, p=0.955), use of priors (Fig S4G-L; Fig. S4G: T-test: t=0.485, p=0.636; S4H: T-test: t=0.155, p=0.879; S4I: T-test: t=-0.667, p=0.516; S4J: T-test: t=-0.667, p=0.516; S4K: T-test: t=-0.072, p=0.944; S4L: T-test: t=-0.472, p=0.644), or attention (Fig. S4M-N; High: Mann–Whitney U: U=32.500, p=0.555; Low: T-test: t=0.087, p=0.932).

Figure 3.

To evaluate task performance, we quantified the rate of right-lick responses, corresponding to reports of high-salience stimuli, and constructed psychometric curves, which increased with stimulus amplitude in both genotypes. Mixed-effects modeling revealed a strong main effect of amplitude (all amplitude coefficients p < 0.005) but no evidence for a genotype-dependent change across the psychometric function (amplitude x genotype interaction coefficients p > 0.266), indicating comparable performance between WT and KO animals across stimulus intensities (Fig. 3A). A genotype-by-amplitude interaction was observed only at 18 µm amplitude (β = 0.127, SE = 0.06, z = -2.18, p = 0.030, 95% CI [-0.24, -0.01]), where responses were slightly reduced in WT mice. However, this effect did not survive post-hoc analysis (T-test: t = 0.568, p = 0.580) and did not generalize across amplitudes.

We then assessed the animals’ ability to discriminate high-salience stimuli by comparing the proportion of right-lick responses to 26 µm and 24 µm stimuli. Our results showed that neither WT nor Fmr1^-/y mice were able to reliably distinguish between these two amplitudes (Fig. 3B; WT: Wilcoxon signed-rank: statistic=3.000, p=0.156; Fmr1^-/y: Paired t-test: t = 1.447, p = 0.221). However, both groups reliably discriminated between 26 µm and 22 µm stimuli (Fig 3C; WT: Wilcoxon signed-rank: statistic = 0.000, p = 0.031; Fmr1^-/y: Paired t-test: t = -4.929, p = 0.001). These results demonstrate that high-salience stimuli differing by 4 µm in amplitude can be discriminated by both genotypes.

We next investigated low-salience stimulus discrimination by comparing the rate of right-lick responses (reporting high salience) to stimuli of 12 µm and 14 µm amplitude. WT mice responded similarly to both amplitudes (Fig. 3D-left; Paired t-test: statistic = -1.684, p = 0.153), consistent with their performance on high-salience stimuli differing by 2 µm (Fig. 3B-left). In contrast, Fmr1^-/y mice exhibited enhanced discrimination for these low-salience stimuli, with increased high-salience report rates for the 14 µm stimuli compared to those for 12 µm (Fig. 3D-right; Paired t-test: t = -5.007, p = 0.001). Both genotypes efficiently discriminated 12 µm from 16 µm stimuli (Fig. 3E; WT: Paired t-test: t = -3.274, p = 0.022; Fmr1^-/y: Paired t-test: t = -6.772, p < 0.001). This enhanced discrimination in Fmr1^-/y mice was further supported by low-salience reports that showed lower low-salience lick rates for 14 µm than 12 µm stimuli (Fig. S5A-right; Paired t-test: t = 2.394, p = 0.044), whereas WT mice responded similarly (Fig. S5A-left; Paired t-test: t = 1.589, p = 0.173), validating that improved discrimination was not driven by higher miss rates for 12 µm stimuli.

Together, these results demonstrate that Fmr1^-/y mice present enhanced tactile fine-discrimination for low-salience stimuli.

Fmr1^-/y mice exhibit intact salience categorization

In addition to assessing stimulus discrimination, our task allowed us to evaluate whether Fmr1^-/y and WT mice form the same categories of low- and high-salience stimuli. Using the psychometric curves of the animals’ high-salience responses (Fig. 3A), we measured the categorization bias and precision. No differences were observed between Fmr1^-/y mice and their WT littermates regarding their categorization bias (Fig. 3F; T-test: t=0.432, p=0.673) or precision (slope, Fig. 3G; T-test: t=-0.283, p=0.782), showing comparable salience categorization between groups. These findings were further confirmed by similar overall task sensitivity (d’, Fig. 3H; T-test: t=0.571, p=0.578). Furthermore, Fmr1^-/y mice showed similar rates of correct (Fig. S5B-C; High: T-test: t=-0.160, p=0.875; Low: T-test: t=1.389, p=0.188) and incorrect choices (Fig. S5D-E; High: T-test: t=0.421, p=0.681; Low: T-test: t=-0.709, p=0.491), as well as comparable overall decision strategy (Fig. S5F, criterion c; T-test: t=0.639, p=0.533) compared to their WT littermates. In conclusion, these results demonstrate similar categorization of high- and low-salience stimuli in Fmr1^-/y and WT mice.

Fmr1^-/y mice show reduced impact of categorization on stimulus discrimination

Higher-order processes such as categorization can amplify perceptual differences between stimuli belonging to different categories (Goldstone, 1994). Alterations in this behavioral process have been reported in autistic individuals, who, despite exhibiting intact visual categorization (in line with our findings in the tactile domain, Fig. 3F-H), show a reduced influence of categorization on perceptual discrimination that leads to diminished facilitation for stimuli belonging to different categories (Soulières et al., 2007). To examine whether salience categorization differentially impacts tactile discrimination in WT and Fmr1^-/y mice, we compared the discrimination accuracy for stimulus pairs that differ by 2 µm and either span across different salience categories (across-category, i.e., 18 µm and 20 µm stimuli) or fall within the same category. Discrimination accuracy was calculated as the difference in rightward choices between each pair of stimuli.

In WT mice, discrimination accuracy was higher for across-category stimulus pairs than for both within-category low- (Fig. 4A-left; Paired t-test, t = 4.452, p=0.007) and high-salience pairs (Fig. 4B-left; Paired t-test, t = 5.453, p = 0.003), indicating a robust facilitatory effect of categorization. Fmr1^-/y mice, in contrast, did not show this facilitation for low-salience stimuli, exhibiting a trend with medium effect size toward improved discrimination for across-category versus within-category low-salience pairs (Fig. 4A-right; Wilcoxon signed-rank: stat = 6.000, p = 0.055; Paired Hedges’ g = 0.640). For high-salience stimuli, however, Fmr1^-/y mice performed similarly to WT, showing enhanced discrimination for across-category pairs relative to within-category high-salience comparisons (Fig. 4B-right; Paired t-test, t = 3.496, p = 0.008).

Figure 4.

To directly compare WT and Fmr1^-/y mice, we quantified the categorical facilitation effect as the difference in discrimination accuracy between across- and within-category (high and low-salience) stimulus pairs. Fmr1^-/y mice showed a trend with a strong effect size toward reduced facilitation (T test: t = 1.81, p = 0.095; Hedge’s g = 0.86), which became significant after exclusion of outliers (Fig. 4C; T test: t = 4.57, p = 0.001; Hedge’s g = 1.94).

Together, these findings suggest that reduced categorization influence in Fmr1^-/y mice diminishes discrimination facilitation for stimuli belonging to different categories, revealing a genotype-specific alteration in how categorical information modulates tactile perception.

High cognitive load conditions reveal salience-dependent attention deficits in Fmr1^-/y mice

Our results during perceptual learning revealed comparable, salience-dependent attention in WT and Fmr1^-/y mice (Fig. S3D-E). During the tactile discrimination/categorization phase of our task, animals were exposed to a high cognitive load, as they needed to discriminate and categorize eight stimulus intensities to obtain a water reward. To assess whether Fmr1^-/y mice exhibit attentional deficits in these high cognitive load conditions, we analyzed the rate of Miss trials as a proxy for attention.

While miss rates for high-salience stimuli did not differ between genotypes (Fig. 5A; all genotype x amplitude interactions p > 0.296), Fmr1^-/y mice showed elevated miss rates for low-salience stimuli during categorization (Fig. 5B). A mixed-effects linear model revealed that Fmr1^-/y mice had significantly higher miss rates than WT controls at the reference amplitude of 12 µm (β = 0.067, SE = 0.024, z = 2.77, p = 0.006, 95% CI [0.019, 0.114]). The genotype differences at 14 µm (β = −0.015, SE = 0.021, z = −0.70, p = 0.483) and 16 µm (β = −0.029, SE = 0.021, z = −1.37, p = 0.171) did not differ significantly from the effect observed at 12 µm. However, a significant genotype × amplitude interaction was observed at the highest amplitude (18 µm; β = −0.045, SE = 0.021, z = −2.16, p = 0.031, 95% CI [−0.087, −0.004]), indicating that genotype differences diminished the highest stimulus intensity of the low-salience range. Post-hoc comparisons further indicated a trend for increased misses at 12 µm with a large effect size (T-test: t = -2.8437, p = 0.058, Hedge’s g = 1.23), but showed no significant differences for the other low-salience stimuli (p > 0.21). Notably, during the training period—when cognitive load was lower—miss rates were comparable across genotypes (Fig. S3D-E).

Figure 5.

Together, these results reveal a salience-specific attentional deficit in Fmr1^-/y mice under high cognitive load, with low-salience stimuli being particularly vulnerable.

Fmr1^-/y mice show preserved choice perseveration but disrupted sensory history integration during tactile categorization

We next examined how trial history—including prior stimulus, choice, and outcome—shapes current decisions during tactile categorization. To do so, we used generalized linear models (GLMs) with a binomial link function to predict high-salience licks (right-lick choices) based on the current stimulus, trial history, genotype, and their interactions. We first fit a main-effects model including the current stimulus, previous stimulus, previous outcome, previous choice, and genotype, and then extended this model to assess genotype-specific modulation of history effects through interactions between genotype and previous choice, outcome, and stimulus (Fig. 6a).

Figure 6.

As expected, the current stimulus was the dominant driver of choice in both genotypes (Fig. 6A-B; β = 1.94, SE = 0.15, z = 12.96, p < 0.001, 95% CI [1.64, 2.23]). However, previous choice also strongly influenced behavior (Fig. 6A). Both genotypes exhibited perseveration, as psychometric curves conditioned on the previous choice were systematically shifted following both prior right and left choices, indicating a robust tendency to repeat prior responses (Fig. 6A-C; β = 1.16, SE = 0.10, z = 11.85, p < 0.001, 95% CI [0.97, 1.36]). In contrast to the increased choice consistency bias observed during training (Fig. 2C-E), WT and Fmr1^-/y mice exhibited comparable levels of perseveration during categorization (genotype x previous choice interaction; β = 0.27, SE = 0.19, z = 1.47, p = 0.142, 95% CI [-0.09, 0.64]).

In contrast, previous outcome (reward or timeout) had a negligible effect on current choice in either genotype (Fig. 6A, D; β = 0.09, SE = 0.10, z = 0.89, p = 0.373, 95% CI [-0.11, 0.28]). Interestingly, the amplitude of the previous stimulus influenced decision-making (Fig. 6A; β = -0.12, SE = 0.04, z = -2.82, p = 0.005, 95% CI [-0.20, -0.04]) in a genotype-specific manner (Fig. 6A, E; genotype x previous stimulus interaction, β = -0.23, SE = 0.06, z = -4.03, p < 0.001, 95% CI [-0.34, -0.12]). In WT mice, larger-amplitude stimuli on the preceding trial slightly reduced the probability of choosing right on the subsequent trial, whereas lower-amplitude stimuli increased it (Fig. 6E). In contrast, choices of Fmr1^-/y mice were not modulated by the amplitude of the previous stimulus.

Together, these results show that tactile decisions in both genotypes are dominated by current sensory evidence and strongly biased by prior choices. While choice perseveration is preserved in Fmr1^-/y mice, integration of recent sensory history is selectively disrupted, revealing a genotype-specific deficit in how past sensory information informs ongoing decision-making.

Discussion

Here, we developed a decision-making task based on psychophysics to dissociate stimulus-driven from cognitively-modulated tactile responses in the Fmr1^-/y mouse model of autism. Our findings reveal salience-dependent cognitive alterations that shape sensory performance. During perceptual learning, Fmr1^-/y mice exhibited an increased choice consistency bias during low-salience trials, which contributed to reduced task performance. During discrimination, Fmr1^-/y mice displayed an enhanced tactile sensitivity under low-salience conditions alongside decreased facilitation of across-category discrimination. Despite this increased sensitivity, these mice showed low salience-specific attentional impairments under high cognitive load, as well as faster updating of their world model with a decreased impact of sensory history on their current choice. Our findings highlight the interplay between sensory and cognitive alterations in autism, emphasizing the importance of cognitive context in interpreting sensory phenotypes, and advocating for a shift beyond traditional sensory–cognitive dichotomies to better understand autism-related phenotypes.

Learning alterations in autism and reliance on priors

Learning disability and intellectual disability are the second and third most common co-occurring neurodivergencies of autism, observed in 23.5% and 21.7% of autistic individuals, respectively (Khachadourian et al., 2023). Notably, even in the absence of these co-occurring conditions, autistic individuals exhibit differences in learning processes and altered neural mechanisms of learning (Mercado et al., 2020). In our study, we observed reduced performance in salience-based perceptual learning, shaped by a higher consistency bias in choice selection in Fmr1^-/y mice. These results align with findings in autistic individuals showing a stronger influence of previous choices on current perceptual decisions (Feigin et al., 2021). Moreover, decreased flexibility in perceptual decision-making is consistent with evidence of altered auditory and visual learning in autism (Harris et al., 2015; Alispahic et al., 2022), decreased performance under volatile conditions (Goris et al., 2021), as well as increased perseveration and diminished sensitivity to feedback (Crawley et al., 2020). Perceptual category learning was also found to be slower in autistic individuals, potentially due to the use of strategies that prioritize response guessing over rule application (Bott et al., 2006; Soulières et al., 2011). Impairments in visual, tactile, and audio-visual perceptual learning have also been shown in Fmr1^-/y mice (Arnett et al., 2014b; Goel et al., 2018; Mol et al., 2024).

Our findings extend this body of work by identifying a salience-dependent shift in reliance on priors in autism, with lower-salience stimuli more strongly evoking dependence on previous choices and impacting perceptual learning.

Increased sensory discrimination in autism

Several theories of autism, including the weak coherence (Frith, 2003), enhanced perceptual functioning (Mottron et al., 2006), and reduced generalization theories (Plaisted, 2001), support the notion of a superior perception of low-level information in autism. Indeed, a body of work has confirmed superior pure tone discrimination (Bonnel et al., 2003; O’Riordan and Passetti, 2006) and discrimination learning of highly confusable patterns (Plaisted et al., 1998b). Tactile discrimination in autism has been studied using diverse experimental protocols that target different tactile features, which has often led to contradictory results due to methodological variability and participant heterogeneity (reviewed in (Zetler et al., 2019)). Similarly, work on roughness discrimination in rodent models (Orefice et al., 2016; Ahmadi et al., 2023) and human studies (O’Riordan and Passetti, 2006) has shown discrepancies.

To address translational challenges between animal models and clinical research, our approach employed vibrotactile stimuli closely aligned with those used in human studies. Consistent with findings in autistic individuals (He et al., 2021b), Fmr1^-/y mice showed intact amplitude discrimination for high-salience vibrotactile stimuli. Clinical studies have typically employed stimuli with more than 10-fold higher amplitudes compared to stimulus detection thresholds (He et al., 2021b; Asaridou et al., 2022). In contrast, our investigation focused on suprathreshold stimuli with much lower intensities (1.5-fold higher above threshold amplitude, 4-fold higher frequency). By employ such stimuli, our study uncovered enhanced discrimination of low-salience stimuli in Fmr1^-/y mice, an aspect that remains underexplored in autistic individuals. Increased “local” processing could account for enhanced discrimination performance (Frith, 2003). However, in our study, Fmr1^-/y mice exhibited enhanced discrimination only under specific stimulus salience conditions, arguing against a global increase in sensory processing precision. Future clinical studies assessing the discrimination of lower-level stimuli are necessary to explore whether enhanced low-salience discrimination also characterizes autistic individuals.

Intact categorization but decreased facilitation of across-category discrimination in autism

Most perceptual discrimination is influenced by categorization (Goldstone, 1994; Beauny et al., 2020; Micher et al., 2024). As mentioned by Goldstone, “a clear distinction between sensory and cognitive processes is not tenable” (Goldstone, 1994). The reduced generalization hypothesis of autism postulates that, apart from superior performance on a difficult discrimination task, autistic individuals will have inferior performance in stimulus categorization (Plaisted, 2001). However, studies on rule-based and prototype categorization have yielded inconclusive results, reporting both intact (Klinger and Dawson, 2001; Molesworth et al., 2005), slower (Soulières et al., 2011), reduced (Froehlich et al., 2012; Gastgeb and Strauss, 2012) and enhanced performance (Bonnel et al., 2003) in categorization tasks in autistic individuals.

We assessed for the first time categorization and its impact on discrimination of vibrotactile stimuli in a mouse model of autism. Our results revealed intact tactile categorization but reduced facilitation of across-category discrimination in Fmr1^-/y mice. Although no clinical studies to date have examined this question in the tactile domain, our findings are consistent with previous research in the visual modality. These studies suggest that enhanced discrimination in autistic individuals may stem from reduced generalization (Plaisted et al., 1998b; Plaisted, 2001), and further indicate that discrimination processes may operate with increased independence from the influence of categorization (Soulières et al., 2007). These findings suggest a bidirectional interaction between sensory and cognitive alterations in autism, indicating that not only atypical sensory perception can impact cognition (Haigh, 2018), but also that cognitive differences can, in turn, shape sensory processing.

Attentional alterations in autism

Although attention deficits are not considered a general characteristic of autism (Grubb et al., 2013), reduced attention in the presence of salient distractors (Venker et al., 2021) and weaker accuracy in executive attention (Ridderinkhof et al., 2018) have been reported in autistic individuals. Cognitive load is known to modulate attentional capacity (Lavie and Dalton, 2014; Murphy et al., 2016), and prior work has shown that high cognitive load alters selective attention and distractor filtering (Head and Helton, 2014; Lavie and Dalton, 2014; Brockhoff et al., 2022). However, less is known about attention to task-relevant stimuli in these conditions. Our results support the view that attention deficits emerge under conditions of high cognitive load in autism. Miss rates increased in our task when Fmr^1-/y mice were required to discriminate and categorize eight stimulus intensities within a single session for water reward, but not during training sessions with only two stimulus intensities.

Notably, cognitive load appears to disproportionately affect attention to low-salience stimuli in Fmr1^-/y mice, suggesting a disruption in late-stage attention processes, which take place after the stimulus is perceived and are involved in top-down goal-directed behavior (Calvillo and Jackson, 2014). In contrast, high-salience stimuli, which tend to capture attention automatically (Smallwood, 2013), may engage early attentional processes and thus less affected by cognitive load. This context-dependent attentional modulation aligns with previous findings in autistic individuals, showing stimulus-dependent auditory attention deficits (Čeponiene et al., 2003). These findings suggest that attentional alterations in autism may be conditionally expressed, particularly under tasks requiring increased cognitive resources.

Our results further dissociate attentional and tactile alterations in Fmr1^-/y mice, as low-salience stimuli were more accurately discriminated yet less attended to in Fmr1^-/y mice. These findings align with studies on linguistic processing in autistic children (Järvinen-Pasley et al., 2008; Ploog, 2010) and extend clinical research demonstrating that tactile sensitivity alterations in autism are not linked to attentional difficulties (He et al., 2021b).

Trial history integration during perceptual decision-making in autism

Sensory perception is shaped not only by stimulus characteristics but also by trial history. Our results show that current stimulus intensity was the primary factor influencing the animal’s choice, together with a strong perseveration across genotypes, consistent with prior work on perceptual decision-making in mice (Aguillon-Rodriguez et al., 2021). This effect remained intact in Fmr1^-/y mice during tactile categorization. Interestingly, previous work in rats and humans has revealed that stimulus history strongly influences current perceptual choices under healthy conditions (Hachen et al., 2021). Our results confirmed these results in control mice and further revealed that this sensory history effect was absent in Fmr1^-/y mice. These findings suggest poor sensory history integration and support models of autism proposing faster world model updating, with new sensory information weighting higher than previous experience (Goris et al., 2022).

Motivation and reward valuation during perceptual decision-making

Motivation and reward valuation are two crucial factors impacting performance during perceptual decision-making. In this study, we minimized confounds from fatigue, satiety, or disengagement by restricting analyses to sessions and epochs in which animals were actively performing the task. Under these conditions, Fmr1^-/y and WT mice showed indistinguishable task engagement, learning rates and learning trajectories, which would not be expected if motivation, reward value, or reinforcement efficacy were altered in our mouse model of autism.

During testing, only the sessions where the mice achieved high accuracy on well-learned training stimuli were included in the analysis. Several additional observations further argue against altered reward valuation as a primary driver of behavioral alterations in Fmr1^-/y mice. We observed no genotype differences in categorical response bias (criterion c), nor any effect of previous reward outcome on subsequent choices, indicating that reward history did not differentially influence decision-making in Fmr1^-/y mice. In contrast, we found reduced across-category discrimination facilitation and diminished influence of recent sensory history on choice, pointing to altered perceptual-decision processes rather than global motivational deficits. Moreover, Fmr1^-/y mice exhibited increased miss rates that emerged selectively under high cognitive load and reduced sensory salience conditions. The salience- and context-dependent nature of these effects, together with preserved engagement and reward sensitivity, supports the conclusion that the observed behavioral differences reflect specific disruptions in how sensory information is integrated and utilized during decision-making, rather than nonspecific changes in motivation, fatigue, or stereotyped responding.

Conclusion and Future Perspectives

Numerous hypotheses have sought to provide a unified explanation for the diverse symptoms of autism. Our findings support the view that altered cognitive processes shape perception in autism, influencing how autistic individuals learn and engage with their environment. We demonstrate differences in choice bias during perceptual learning and increased tactile discrimination in trained animals. These changes are accompanied by reduced categorization influence, sensory history integration, and attention. Rather than reflecting a global sensory deficit or enhancement, our results point to context-dependent alterations in how sensory information is integrated, weighted, and used for decision-making.

Future work will be needed to identify the circuit mechanisms driving these effects, and to test whether they reflect intrinsic changes in top-down cognitive networks or altered feedback between cognitive and sensory circuits that reshape sensory representations. In addition, although the present study focused on amplitude discrimination within the flutter range (40 Hz) to align with translational tactile paradigms, extending this approach to a broader range of vibrotactile frequencies will be important to determine whether sensory–cognitive alterations in autism generalize across stimuli that engage different mechanoreceptor populations. Finally, while the present study focused on vibrotactile processing due to its strong translational relevance, previous studies in the visual modality have reported similar cognitive influences on perceptual decision-making, suggesting that these effects may generalize across sensory systems. Future studies directly comparing modalities within the same experimental framework will be important to confirm this. Elucidating how cognition modulates sensory perception may offer valuable insights for clinical practice, help reconcile conflicting findings in the field, and inform the development of more targeted, mechanism-based interventions.

Materials and Methods

Experimental design

To study tactile perception as well as attention and perceptual biases in autism, we developed a novel 2-Alternative Choice task for the categorization and discrimination of flutter-range vibrotactile stimuli. Throughout the text, we use terms that are preferred in the autistic community and are less stigmatizing (Bottema-Beutel et al., 2021).

Ethical statement

All experimental procedures were performed in accordance with the EU directive 2010/63/EU and French law following procedures approved by the Bordeaux Ethics Committee and Ministry for Higher Education and Research. Mice were maintained in reversed light cycle under controlled conditions (temperature 22–24 °C, humidity 40–60%, 12 h/12 h light/dark cycle, light on at 21:00) in a conventional animal facility with ad libitum access to food and ad libitum access to water before the water restriction period. All experiments were performed during the dark cycle, under red light.

Mice

Second-generation Fmr1 knockout (Fmr1^−/y) and wild-type littermate mice 5-16 weeks old were used in our study. Mice were maintained in a C57Bl/6 J background (Mientjes et al., 2006). Male wild-type and Fmr1^-/y littermates were generated by crossing Fmr1^+/− females with Fmr1^+/y male mice from the same production, and the resulting progeny used for our experiments was either Fmr1^+/y (wild type) or Fmr1^-/y (KO). Mice were maintained in collective cages following weaning (2-4 litter males per cage). Cages were balanced for genotype and supplemented with cotton nestlets and carton tubes.

The perceptual decision-making data were collected from 5 different cohorts of mice at different time-points during their active phase of the day. Mice of both genotypes were littermates and represented in each cohort. The number of mice is provided in the figure captions. The experimenter was blind to the animals’ genotypes throughout the experiment. The genotype of experimental animals was re-confirmed post hoc by tail-PCR.

2-Alternative Choice task Setup

The vibrotactile decision-making setup was positioned in an isolation cubicle to minimize interference during the experiment. Mice were placed in a body tube and were head-fixed with their forepaws resting on two steel bars (6 mm diameter, Thorlabs). The right bar was mounted to a Preloaded Piezo Actuator (P-841.6, Physik Instrumente) equipped with a strain gauge feedback sensor and controlled (E-501, Physik Instrumente) in a closed loop, as described before (Prsa et al., 2019; Semelidou et al., 2024). A 12.7 mm stainless steel post (ThorLabs) was mounted on the actuator vertically and a 0.6 mm stainless steel rod (ThorLabs) was clamped horizontally onto this post. The horizontal rod served as the contact bar on which the animal rested its right forepaw. Water reward was delivered through either of the two metal feeding needles (20G, 1,9mm tip, Agntho’s AB), placed left and right of the mouse’s mouth, each connected to a lickport interface with a solenoid valve (Sanworks) equipped with a capacitive sensor (https://github.com/poulet-lab/Bpod_CapacitivePortInterface). The perceptual decision-making setup was controlled by Bpod (Sanworks) through scripts in Python (PyBpod, https://pybpod.readthedocs.io/en/latest/). The lickport interface (Sanworks) was equipped with a capacitive sensor (https://github.com/poulet-lab/Bpod_CapacitivePortInterface).

Habituation to head-fixation and water restriction

Mice (P40-P50) were handled using carton tubes and the cupping technique until they were comfortable in the experimenter’s hands, attested by eating while handled. Mice were gradually habituated to the experimental setup and head fixation for 5 days. The third day of habituation, a water-restriction protocol was implemented, where mice had access to liquid water in the setup and to a solid water supplement (Hydrogel, BioServices) in their home cage. The water supplement was divided into small, individual portions, and each mouse received its allotment after the daily training/testing session. Daily body weight measurements were used to monitor hydration and ensure that all animals maintained stable body weight. If necessary, supplemental water was adjusted to maintain animals within the approved weight range. In total, the animals received 1.5-2 ml of water per day, which corresponds to 50-65% of their ad libitum consumption, while ensuring that they did not lose more than 10% of their weight. Each mouse received 6-8 g of Hydrogel (ad libitum) during the weekend. This water restriction protocol was maintained throughout behavioral training and until the end of behavioral testing.

2-Alternative Choice task training and testing Vibrotactile Stimuli

Stimuli were sinusoidal vibrations at 40 Hz with peak-to-peak displacements of 12 μm amplitude (low salience) or 26 μm amplitude (high salience), well above the detection threshold of both groups (10 Hz, 4.46 µm for WT; 10Hz, 7.29 µm for Fmr1^-/y mice) (Semelidou et al., 2024).

Animals were positioned in the setup to ensure stable and consistent forepaw contact with the rod delivering the vibration. Pilot experiments with a sensor to monitor forepaw placement confirmed that the mice did not remove their forepaws from the bar before stimulus delivery.

Analysis of trials were no response followed stimulus delivery (Miss trials) showed that the 12 μm stimulus consistently elicited a higher proportion of missed responses compared to the 26 μm stimulus across animals, indicating lower behavioral performance for the lower-amplitude stimulus. We therefore referred to the 12 μm stimulus as “low salience” and the 26 μm stimulus as “high salience” to denote relative differences in perceptual strength and attentional engagement within the suprathreshold range.

Habituated mice (8 weeks old) were trained to associate high- (26 μm) and low-salience (12 μm) vibrotactile stimuli (pure sinusoid, 500 ms duration, 40 Hz frequency) with a water reward (8 µl) at the right- or left-placed lickport of the setup, respectively. All trials consisted of stimulus delivery followed by a 2 s response window during which the mice could lick to receive the reward. Inter-trial intervals were variable (5-10 s). Training was subdivided in 4 phases: (a) automatic water delivery at the beginning of the response window at the corresponding lickport (left for 12 μm stimuli, right for 26 μm stimuli). (b) Training in blocks: lick-triggered water delivery during blocks of 20 trials with the same vibrotactile stimulus, of either 12 μm or 26 μm amplitude. Licking at the wrong port resulted in 5 s timeout. (c) Training with pseudorandomly delivered high- and low-salience trials: lick-triggered water delivery following pseudorandom delivery of 12 μm or 26 μm amplitude stimuli. Licking at the wrong port resulted in 5 s timeout.

All sessions consisted of 200-300 trials with a 1:1 ratio of high and low-salience stimuli. During training, this ratio was modified when an animal showed consistent bias for one of the two lickports. Pilot experiments with an extra sensor to monitor forepaw placement confirmed that the mice did not remove their forepaws from the bar before stimulus delivery. To complete training in blocks, mice needed to perform with 70% correct choices and 30% incorrect choices for both stimulus amplitudes. To complete training with pseudorandomly delivered high- and low-salience trials, mice needed to reach the criterion of more than 70% correct choices and less than 30% incorrect choices as an average for 3 consecutive days. All mice that fulfilled this criterion were tested for the categorization/discrimination of vibrotactile stimuli. During testing, stimuli (pure sinusoid, 500 ms duration, 40 Hz frequency) were delivered in a pseudorandom manner with a 50% High-salience: 50% Low-salience ratio. Amplitudes varied on a range between 12 μm and 26 μm. Stimuli of 20, 22, 24 and 26 μm were considered high-salience stimuli and rewarded at the right lickport while stimuli of 12, 14, 16 and 18 μm were considered as low-salience stimuli and rewarded at the left lickport.

2-Alternative choice vibrotactile task analysis

All analysis was performed with custom-made Python scripts that can be available upon request. Behavior was quantified based on the lick events and three main outcomes were measured for each stimulus salience: Correct choice rate (number of correct licks divided by the total number of high- or low-salience trials), Incorrect choice rate (number of wrong licks divided by the total number of high- or low-salience

trials), and Missed stimuli rate (number of trials in which the animal did not lick, divided by the total number of high- or low-salience trials). Sessions in which animals disengaged were analyzed only during epochs in which the animal was actively performing the task. Training duration was calculated based on the total number of days each animal passed in training in blocks and with pseudorandom stimulus delivery.

For testing, only sessions with more than 70% correct choices for the training stimuli (12 μm and 26 μm) were analyzed. Psychometric curves were fitted on the rightwards lick rate for each stimulus amplitude using a general linear model. An average of 84 repetitions for each amplitude was used to calculate rightward lick rates.

Based on the signal detection theory (Hautus et al., 2021), sensitivity d’ was calculated as:

Sensitivity d’ for high- and low-salience stimuli was calculated based on the Correct and Incorrect choice rate for high- and low-salience stimuli, respectively.

The strategy of the animals was assessed through their criterion, which was calculated as:

Categorization bias was calculated as the inflection point (i.e., the midpoint parameter μ) of the fitted logistic psychometric function and categorization precision as the slope of the curve.

Prior choice impact was calculated separately for each type of the current trial (high- or low-salience). For low-salience trials, the prior choice impact was calculated as the rate of correct low and incorrect-low trials, divided by the rate of all correct and all incorrect trials. Similarly, for high-salience trials, the prior choice impact was calculated as the rate of correct high and incorrect-high trials, divided by the rate of all correct and all incorrect trials.

Learning trajectory analysis

Learning curves were quantified from the training phase with mixed high- and low-salience trials by calculating, for each mouse, the proportion of correct choices on each training day. For group-level comparisons, performance was plotted across days for WT and Fmr1^-/y mice, and mean learning trajectories with 95% confidence intervals were computed for each genotype.

To statistically compare learning trajectories across genotypes, we used a linear mixed-effects model with fixed effects for Genotype, Day, and their interaction, and a random intercept and random slope for each animal. This model tested whether performance differed between genotypes across days and whether the rate of learning (performance change per day) differed between WT and Fmr1^-/y mice.

To quantify individual learning rates, we computed the slope of performance as a function of training day for each animal. For each mouse, performance values (proportion of correct choices) were regressed against each Day using a simple linear regression model:

The slope (β₁) provides an estimate of the rate of improvement per training day.

Intermediate-stage slopes (middle 3 days of each mouse’s training) were calculated in the same manner. Categorization bias was calculated based on the psychometric curves, as the stimulus amplitude at the inflection point of the sigmoid fitting curve. Categorization precision was calculated based on the slope of the psychometric curve.

Delta discrimination accuracy was computed as the difference in the rate of high-salience (rightward) licks between pairs of stimuli that differed by 2 µm. For analyses of low-salience and high-salience discrimination, delta discrimination accuracy values were calculated for each within-category stimulus pair and then averaged within the low-salience or high-salience category, respectively.

Statistics

All values are presented as mean ± s.e.m. Box plots show the median, interquartile, range, mean and individual values. The numbers of animals used are indicated in the figure legends. Sample size was determined based on previous studies using similar behavioral paradigms. For experiments assessing tactile responses of trained animals, only animals that reached the training criterion were tested. For the discrimination & categorization task, all the testing sessions where the animals showed less than 70% correct responses to the training stimuli (12µm and 26µm) were excluded from tactile discrimination/categorization analysis. No samples or data points were excluded from the analysis.

All statistical analysis was done using Python (SciPy, Pingouin). All datasets were tested for normality using Shapiro–Wilk tests, and potential outliers were identified using a median absolute deviation–based modified Z-score (|Z| > 3.5). When an outlier was detected, statistical analyses were performed both including and excluding the outlier. In such cases, results from both analyses are reported when the inclusion or exclusion of the outlier altered the statistical outcome, to ensure transparency and robustness of the findings. For normally distributed data, two-tailed paired t-tests were used for within-subject comparisons, and Welch’s independent-samples t-tests were used for between-group comparisons; otherwise, the Mann–Whitney U test (between groups) or the Wilcoxon signed-rank test (within subjects) was applied. Bonferroni correction was applied where multiple comparisons were performed. Effect sizes were calculated using Hedge’s g for between-genotype comparisons and paired Hedge’s g for within-genotype comparisons, particularly when statistical analyses showed trend-level differences (p ≈ 0.05–0.1), to provide a measure of the magnitude of the effect independent of sample size. Linear mixed-effects models were used to assess statistical differences in learning trajectories, psychometric curves, and miss rates.

Statistical modeling during categorization taking into account the trial history. Generalized linear models (GLMs) with a binomial link function were used to quantify the effects of current stimulus, previous trial history, genotype, and their interactions on the probability of choosing the right port. Trials without a defined history and Miss trials were excluded. Stimulus amplitudes were z-scored per mouse.

Two models were fit:

1. Main effects GLM – included current stimulus, previous stimulus, previous outcome, previous choice, and genotype.

2. Interaction GLM – extended the main effects model by including genotype × previous outcome, genotype × previous choice, and genotype × previous stimulus interactions.

Cluster-robust standard errors were calculated to account for repeated measurements within each mouse.

Marginal effects and psychometric curves

Predicted probabilities were computed from the interaction GLM to visualize the effect of stimulus amplitude, previous choice, and previous outcome on behavior, separately for WT and KO mice.

Supplementary figures

Figure S1.

Figure S2.

Figure S3.

Figure S4.

Figure S5.

Data availability

The raw data of behavioral experiments generated in this study have been deposited on the figshare database: 10.6084/m9.figshare.29459771. Source data for each figure and the Python code used for all analyses are provided with the paper.

Acknowledgements

We would like to thank the animal housing and genotyping facilities of INSERM U1215 Neurocenter Magendie. We thank Dr. James Poulet (Max-Delbrück-Centrum für Molekulare Medizin, Berlin) for providing the capacitive sensors for the behavioral task. We would like to thank Dr. Melanie Ginger for providing feedback on the manuscript. Mouse schemas were made using Biorender. This project was funded by the Foundation for Medical Research Postdoc Fellowship (O.S.), INSERM (A.F.), Marcel Dassault-Fondation FondaMental Award 2019 (A.F.), Simons Foundation Autism Research Initiative (A.F.), GIS Autisme & TND (O.S.), 2024 NARSAD Young Investigator Grant (O.S.), Fondation FondaMental (A.F.).

Additional information

Contributions

A.F. and O.S. conceived the project. O.S. designed the experiments. O.S., M.T-M.F. and A.W. performed the experiments. O.S. and A.W. wrote the Python code for the behavioral experiments. O.S. analyzed the data and wrote the necessary Python scripts. O.S. interpreted the data. O.S. prepared the figures. O.S. wrote the manuscript, and A.F. and M.T-M.F. provided feedback.

Inclusion and diversity statement

We support inclusive, diverse, and equitable conduct of research. We tried to use inclusive language as much as possible.

Funding

Fondation pour la Recherche Médicale (FRM) (PostDoc France SPF202005011992)

• Ourania Semelidou

INSERM

• Andreas A Frick

Marcel Dassault-Fondation FondaMental Award 2019

• Andreas A Frick

Simons Foundation Autism Research Initiative (SFARI)

• Andreas A Frick

GIS Autisme et TND

• Ourania Semelidou

Brain and Behavior Research Foundation (BBRF) (2024 NARSAD Young Investigator Grant)

• Ourania Semelidou

Additional files

Supplementary file 1. Mixed models statistics.

Soruce data 1.

Soruce code 1.