We trained mice to perform an AV rule-switching task, in which they made decisions using auditory stimuli while ignoring simultaneously presented visual stimuli or vice versa. Trial presentation was self-paced in a virtual foraging environment wherein a visual track was advanced by forward locomotion on a spherical treadmill (Figure 1A). A task-irrelevant random double sweep (RDS) sound was presented during inter-trial intervals (ITIs) for mapping auditory receptive fields in each attentional state (Figure 1B). Decision stimuli were presented after variable track length, consisting of 1 s auditory tone clouds (TCs; centered at 8 or 17 kHz) and/or visual drifting gratings (horizontal or vertical orientation; Figure 1C). One of the decision stimuli for each modality was a rewarded target (A_R, V_R) and the other an unrewarded distractor (A_U, V_U). Lick responses following targets (hits) and distractors (false alarms [FAs]) produced water rewards and dark timeouts, respectively. Withholding licks for targets (misses) or distractors (correct rejects [CRs]) advanced the next trial. Each session began with a block of unimodal decision stimuli, which cued the attended modality of a subsequent AV block (Figure 1D). A second unimodal block from the other modality was then presented, cueing the rule for a final AV block. Decision stimuli had identical physical properties but different behavioral significance between rules (e.g., licks following A_RV_U were rewarded in A-rule but punished in V-rule). Targets and distractor stimuli remained constant throughout training for each mouse and were approximately counterbalanced across animals. Block sequences (A-rule then V-rule, or vice versa) were also counterbalanced across sessions (Figure 1D.c).

We used two approaches to ensure that animals were engaged during both task rules. First, we restricted analysis to sessions in which discrimination was well above chance (d’>1.5) for both rules, and for which FA rates were below 0.5 for the stimuli with reward valences that conflicted across rules (A_UV_R in the A-rule, A_RV_U in the V-rule; Figure 1F). Second, for a subset of sessions (n=14 sessions, 5 mice) we measured pupil size, a well-established correlate of arousal and behavioral performance (Bradley et al., 2008; McGinley et al., 2015; Reimer et al., 2014). We used a computer vision algorithm to automate measurement of pupil size (pupil diameter/eye diameter) for each frame acquired by a CCD video camera (Figure 2A.a). To isolate pupil fluctuations reflecting general arousal, pupil size was measured during an ITI window designed to avoid pupil responses to decision stimulus onset, dark timeouts, and decreased locomotion events following reward administration (Figure 2.A.b–c, Figure 2B). Previous studies have reported that pupil size increases with task difficulty and engagement in humans, non-human primates, and rodents (Hess and Polt, 1964; Kawaguchi et al., 2018; Schriver et al., 2018). We reasoned that comparison of pupil size across the rules would allow us to establish whether task demands differed between the rules. No difference in pupil size was observed between rules during bimodal blocks (Figure 2C; A-rule bimodal: 0.29±0.05 norm. diameter ± SD, V-rule bimodal: 0.30±0.05; Z=–1.0, p=0.30, paired Wilcoxon signed-rank [WSR], Benjamini-Hochberg false discovery rate [FDR]-adjusted p-values). As expected, pupil diameters were significantly larger during the bimodal portion of the task, when visual stimuli were present and the task had increased in difficulty, compared to the auditory-only unimodal portion of the task (Figure 2—figure supplement 1; A-rule unimodal: 0.28±0.04, A-rule bimodal: 0.29±0.05; Z=–2.6, p=0.009). A trend toward smaller pupil size in the unimodal visual rule compared to bimodal rule was also noted, but the difference did not reach significance after multiple comparisons correction (V-rule unimodal: 0.29±0.04, V-rule bimodal: 0.30±0.05; Z=–2.0, p=0.062). Because pupil size also closely tracks locomotion (Figure 2A.b–c; McGinley et al., 2015), we examined locomotion speed during the same ITI window (Figure 2D). Differences in locomotion speed were also not observed between rules (all p≥0.623; all |Z|≤1.29, paired WSR). Arousal and motor activity were thus comparable between rules, suggesting that differences in neuronal activity may be attributable to modality-selective attention.

After mice learned the AV rule-switching task, a craniotomy was made over right AC, to allow for acute recordings during behavior using multichannel probes spanning the full cortical depth (Figure 3A). In total, we recorded AC activity in 10 mice during 23 behavioral sessions meeting inclusion criteria. The putative cortical depth of each sorted single unit (SU) was assigned by calculating the fractional position of the channel with the largest waveform amplitude within the span of channels in AC, as estimated from spontaneous and tone-evoked recordings following the task (Figure 3B). A separate set of experiments to visualize probe tracks with the fluorescent dye Di-I provided support for this depth estimation technique (Figure 3C; DiCarlo et al., 1996; Morrill and Hasenstaub, 2018). We then divided the fractional depth values into superficial, middle, and deep groups, approximating the supragranular, granular, and infragranular laminae. We further divided SUs into narrow-spiking (NS, putative inhibitory; n=130, 18%) and broad-spiking (BS, predominantly excitatory; n=612, 82%) populations based on trough-peak time (Figure 3D; Bigelow et al., 2019; Cardin et al., 2007; Nandy et al., 2017; Phillips et al., 2017).

To measure the effects of modality-selective attention on stimulus processing in AC, we began by comparing SU responses to bimodal decision stimuli across task rules. These responses reflected physically identical stimuli and similar levels of arousal and locomotion, as shown in Figure 2. Activity patterns evoked by decision stimuli and modulatory effects of task rule were diverse (Figure 4A). To capture a predominantly sensory-driven response component, we measured mean firing rates (FRs) during the first 300 ms post-stimulus onset (Figure 4B), which preceded most lick responses (lick latency median: 611 ms; 5–95th percentiles: 289–1078 ms; 5.4% of licks <300 ms, n=2852 total lick trials across dataset). Trials with licks earlier than 300 ms were excluded from analysis. We first compared A-rule and V-rule responses to the TC rewarded in the A-rule (A_R*: A_RV_R and A_RV_U responses combined). Averaging across units, responses in the deep layers were suppressed in the A-rule relative to the V-rule for both NS and BS units (Figure 4C; deep BS: p=2.8e-4, Z=4.1, median fold change [FC; A-rule/V-rule]: 0.89, n=333 SUs; deep NS: p=0.011, Z=2.9, median FC: 0.87, n=66; paired WSR, FDR-adjusted p-values; see Figure 4—source data 1A for full stats). No significant group-level change was found in middle or superficial units. Consistent with group-level trends, individual units with significant FR decreases in the A-rule (p<0.01, unpaired t-test) substantially outnumbered units with significant FR increases for all unit populations other than superficial and mid-depth BS units (Figure 4C, right).

A similar pattern of attention-related modulation was observed for unrewarded stimuli in the A-rule (A_U*: A_UV_R and A_UV_U responses combined). At the group level, superficial and middle unit responses did not differ significantly between conditions, whereas deep BS units were suppressed in the A-rule (Figure 4D; deep BS: p=2.0e-06, Z=5.11, median FC: 0.81, n=321; paired WSR, FDR-adjusted p-values; see Figure 4—source data 1B for full stats). Relative fractions of units with significantly modulated FRs to A_U* stimuli were similar to those described above for A_R* stimuli (Figure 4D, right), with the exception of the superficial group, in which slightly more unitshad increased FRs. We further found that most units showed the same direction of modulation for A_R* and A_U* stimuli (Figure 4E), with a similar modulation sign observed for 78% of BS units (50% suppressed for both A_R* and A_U*, 28% enhanced for both) and 99% of NS units (68% suppressed for both, 31% enhanced for both). These findings suggest that modality-selective attention similarly influences FRs evoked by task-relevant target and distractor sounds with different acoustic properties and learned behavioral values.

To determine whether these attentional influences might generalize to task-irrelevant sounds, we examined responses to RDS sounds presented during the ITI. Using the same analysis window (300 ms post-stimulus onset, Figure 4F), we found that attention-related modulation of FR responses evoked by task-irrelevant sounds was highly similar to that observed for both types of decision stimuli: middle- and deep-layer BS and NS populations exhibited group-level FR suppression during the A-rule (Figure 4G), whereas superficial layer units showed no difference (middle BS: p=6.0e-3, Z=3.1, median FC: 0.85, n=112; middle NS: p=0.014, Z=2.7, median FC: 0.65, n=28; deep BS: p=1.2e-6, Z=5.2, median FC: 0.84, n=309; deep NS: p=0.021, Z=2.5, median FC: 0.80, n=64; paired WSR; see Figure 4—source data 2 for full stats). Significantly modulated unit counts were again highly biased toward suppression in the A-rule, with pronounced differences in the middle and deep unit groups (Figure 4G, right). Together, these results show that auditory-selective attention tends to reduce FR responses to sounds, regardless of their behavioral relevance, valence, or spectral content, and that these influences are strongest for deep-layer units.

Previous studies have found that modulation of ongoing activity in sensory cortex can influence subsequent sensory-evoked responses (Arieli et al., 1996; Haider and McCormick, 2009). Thus, the response suppression during auditory attention reported above may either reflect specific decreases in stimulus responsivity or general decreases in ongoing activity. To address these possibilities, we quantified FRs in a pre-stimulus window spanning 300 ms prior to decision stimulus onset in which no sounds were presented (Figure 5A.a). Although this window may include anticipatory modulation of activity (Cox et al., 2019; Egner et al., 2010; Samuelsen et al., 2012), it nevertheless provides a measure of baseline activity for comparison with evoked responses. We observed significant group-level decreases in pre-stimulus FRs during the A-rule for units in the middle NS and deep BS groups, but no modulation of superficial units (Figure 5B; middle NS: p=0.039, Z=2.48, median FC: 0.71, n=28; deep BS: p=4.2e-05, Z=4.49, median FC: 0.87, n=336; paired WSR, FDR-adjusted p-values; see Figure 5—source data 1 for full stats). To test whether the reduction in pre-stimulus FR was sufficient to account for stimulus-evoked changes reported above, we recalculated FRs evoked by decision stimuli as FC from pre-stimulus FRs (Figure 5A.b). Following this adjustment and after FDR correction, the middle- and deep-layer unit population responses no longer differed between rules (Figure 5C; Figure 5—source data 2). Together, these results suggest that group-level decreases in evoked FRs during A-rule are largely due to generalized suppression of ongoing AC activity.

We next sought to determine whether attention-related changes in stimulus response were related to the tuning preferences of units, a phenomenon termed ‘feature attention’ previously observed in both monkey VC (Maunsell and Treue, 2006; Treue and Martínez Trujillo, 1999) and AC (Da Costa et al., 2013). The RDS mapping stimulus, which we have previously used to efficiently identify auditory response properties and AV interactions (Bigelow et al., 2022), was used to generate spectrotemporal receptive fields (STRFs) through reverse correlation (Figure 6A; Aertsen and Johannesma, 1981; de Boer, 1968; Gourévitch et al., 2015). Tuning for each STRF was measured through a trial-to-trial reliability metric, which we used to divide units into those with activity changes that were reliably evoked by defined spectral or temporal features (tuned STRFs, n=172; Figure 6C.a) and those without feature-evoked changes (untuned, n=409; Figure 6C.b). Spiking activity levels were higher in tuned units compared to untuned (Figure 6D.a). To control for possible activity level-dependent effects, we compared our population of tuned units to a randomly selected subset of untuned units which was matched for both sample size and FR to the tuned population (Figure 6D.b). We then examined attentional modulation of stimulus responses between the tuned and subsampled untuned groups. Responses to the rewarded TC (A_R*), unrewarded TC (A_U*), and the RDS mapping stimuli were significantly modulated by task rule in the untuned group, but not the tuned group (tuned: all p≥0.18, all |Z|≤1.72; untuned: all p≤0.023, all |Z|≥2.27; one-way WSR vs. modulation of 1 [equal across rules], FDR-adjusted p-values; Figure 6—source data 1C, D). Nevertheless, comparisons across these tuned and untuned groups showed that the distributions did not significantly differ after multiple comparisons correction (all p≥0.12, all |Z|≤2.06; WSR).

An important caveat is that the RDS stimuli may not capture all units with some degree of tuning preference. As such, a conservative interpretation would be that group-level suppression during auditory attention is driven by units that do not exhibit strong tuning preferences. Additionally, both tuned and untuned populations contained units with significant evoked responses to the two TCs, although fractions of responsive units were higher in the tuned group (Figure 6D.c). This shows that an absence of STRF tuning does not imply that units were not responsive to the task stimuli.

For the tuned group, does frequency preference determine degree of attentional modulation? We measured the best frequency (BF) of the excitatory field in each tuned STRF (Figure 6F). Consistent with previous work showing that task demands shape frequency representation in AC (Atiani et al., 2009; Fritz et al., 2005; Fritz et al., 2003; Yin et al., 2014), we found a strong BF preference for a 1-octave band around the center frequency of the rewarded TC (Figure 6G). Furthermore, distributions of BFs measured during the A-rule and V-rule were strikingly similar. This suggests that in our task, AC had shifted its frequency representation in a manner that was not rule-dependent. To test whether modulation by rule was dependent on tuning, we next divided units by their BF, as measured from the A-rule STRF, into groups near center frequency of A_R (±0.5 octaves), near A_U or with a BF outside of either band. No difference between the tuning groups was observed for responses to A_R* or A_U* TCs (Kruskal-Wallis non-parametric ANOVA, all p>0.12, all H<5.5, FDR-adjusted p-values; Figure 6—source data 1C, D), suggesting that frequency tuning does not determine suppression or enhancement by attention in this task.

Previous work has established that FR changes do not necessarily imply changes in the amount of information spikes carry about sensory stimuli. For instance, optogenetic activation of inhibitory interneurons can reduce FRs in AC without changing information, suggesting increased encoding efficiency (Phillips and Hasenstaub, 2016). By contrast, locomotion reduces both FRs and information in AC (Bigelow et al., 2019). To determine whether reduced FRs evoked by decision stimuli were accompanied by changes in information or encoding efficiency, we used a peristimulus time histogram (PSTH)-based neural pattern decoder to compare sound discrimination across attentional states (Foffani and Moxon, 2004; Hoglen et al., 2018; Malone et al., 2007). For each unit, the decoder generates a single-trial test PSTH and then compares these to two or more template PSTHs from different stimulus response conditions, generated sans test trial (Figure 7A). The test trial is assigned to the template that is closest in n-dimensional Euclidean space, reflecting n PSTH bins. This is repeated for all trials, generating new templates for each classifier run. After all trials have been classified, a confusion matrix is generated. From this, we calculated accuracy of classification, MI (bits), and encoding efficiency, a spike-rate-normalized MI (bits/spike). As in previous analyses, a 0–300 ms post-stimulus onset window was used in this method to restrict decoding to a predominantly sensory-driven component of the response. The binwidth for generating PSTHs was 30 ms (Hoglen et al., 2018). Only trials with correct responses (hits and CRs) and units with a minimum stimulus response FR of 1 Hz to both stimuli used in the decoder comparison were included.

We found that task rule could be decoded at greater than chance levels from responses to all four AV stimuli, and at all depth and NS/BS groups, showing that attentional state modulates decision stimulus PSTH responses throughout AC (Figure 7—figure supplement 1; Figure 7—source data 1). These comparisons suggest response modulation by task rule, but do not address how information processing changes across the rules. To test this, we used the decoder to compare accuracy in discriminating between responses to A_R* (rewarded in A-rule) and responses to A_U* (unrewarded in A-rule) bimodal stimuli across A-rule and V-rule conditions. This mimics the TC discrimination required by the mice during the A-rule. In both rules, classification accuracy for the auditory decision stimuli (A_R*, A_U*) was higher than chance for all depth and BS/NS groups (see scatter plots in Figure 7B; all p≤1.4e-05, all |Z|≥4.2, one-way WSR vs. chance [50%]; see Figure 7—source data 2A for stats). Sound classification accuracy (A_R*, A_U*) did not significantly differ across the A-rule and V-rule (Figure 7B, A_R* vs. A_U* comparison across rules: all p≥0.17, all |Z|≤1.39, see Figure 7—source data 3A for full stats; paired WSR on decoder accuracy in A-rule vs. V-rule by depth and NS/BS groups). Despite a reduction in activity levels during auditory attention, there was no loss in decoder accuracy, suggesting a possible change in encoding efficiency.

Through analysis of all decoder runs, we found that classifier accuracy and raw information were indeed correlated with FR (accuracy: r(3001)=0.49, p=2.3e-180; MI: r(3001)=0.41, p=1.5e-123; Pearson’s correlation, all A_R* vs. A_U* decoder runs). Thus, normalizing information by mean joint per-trial spike rate for the two responses in each decode (bits/spike) provides insight into the efficiency with which spikes are used to represent stimuli. We found that this encoding efficiency measure increased by ~20% during the A-rule for deep-layer BS units (Figure 7C, A_R* vs. A_U* comparison across rules: deep BS: p=2.9e-04, Z=–4.06, paired WSR, FDR-adjusted p-value; median FC: 1.19 [fold change: A-rule/V-rule]; V-rule: 0.15±0.13, A-rule: 0.19±0.19, mean bits/spike ± SDs, n=233; all other groups p≥0.40, all |Z|≤1.47; see Figure 7—source data 4A for full stats). No other unit subpopulations showed significant changes. Note that for clarity, the above results are presented as the mean of decoder comparisons A_RV_R vs. A_UV_R and A_RV_U vs. A_UV_U, thus collapsing visual stimulus identity. Analysis of these comparisons separately yields highly similar results (Figure 7—figure supplement 2; Figure 7—source data 2–4), suggesting that visual stimulus identity does not contribute substantially to decoder accuracy or encoding efficiency at the level of group analysis.

The analyses above revealed that auditory attention increased the per-spike encoding efficiency of task decision sounds. Does this effect of cross-modal attention switching generalize to encoding of sounds that were explicitly designed to be task-irrelevant? This helps determine whether attention observed here is specific to features of the auditory stream or broadly alters encoding of incoming auditory information. To address this, we tested whether information between STRFs derived from task-irrelevant ITI sounds and spike trains was modulated by attentional demands of the task. We restricted our analyses to only those units with STRFs passing the reliability criterion shown in Figure 6B. To calculate STRF-spike train MI for each SU, we first calculated probability distributions of STRF-stimulus projection values for all stimulus time points (P(x)) and for those time points preceding a spike (P(x|spike); Figure 7D). Intuitively, these projection values reflect the similarity between a windowed stimulus segment at a given timepoint and the STRF. The divergence of the two projection distributions is captured in a spike-rate-normalized MI measure (bits/spike; encoding efficiency), which describes the reliability with which spikes are determined by stimulus features of the STRF (Figure 7E). No differences in encoding efficiency between conditions were observed in the superficial or middle BS/NS groups, or the deep NS group. Instead, consistent with our earlier findings for decision stimuli, encoding efficiency showed a significant A-rule increase in the deep BS subpopulation (Figure 7F; deep BS: p=0.014, Z=–3.05, median FC: 1.25, n=50; paired WSR, FDR-adjusted p-value; FC: A-rule/V-rule; mean bits/spike ± SDs; all other groups p≥0.24, all |Z|≤1.66; see Figure 7—source data 5 for full stats). This finding shows that during auditory attention, stimulus encoding is better described by a linear STRF filter and thus better tracks physical sound features. Furthermore, it suggests that increased encoding efficiency resulting from decreased spiking is a general effect of auditory attention in deep-layer BS units, regardless of the context-based behavioral relevance or learned valence of the sounds.

The increase in A-rule encoding efficiency and decrease in average FRs in deep AC led us to further explore the relationship between activity level and information changes. Specifically, we tested whether group-level information efficiency changes are driven by SUs with suppressed responses, and how the minority of units with increased A-rule FRs perform in the decoder. We therefore examined classifier accuracy and encoding efficiency for target and distractor (A_R* vs. A_U*) decoding separately for deep-layer BS units with increased and decreased FRs in the A-rule (Figure 7—figure supplement 3). We found that units with increased FRs (39%; n=96) exhibited a significant increase in A-rule decoding accuracy (Figure 7—figure supplement 3C; p=0.0030, Z=–2.97, med. FC: 1.04, V-rule % correct: 66.4±15.5, A-rule: 69.5±15.5, n=96; paired WSR; FC: A-rule/V-rule; mean ± SDs), but no significant change in encoding efficiency (p=0.84, Z=0.2, V-rule bits/spike: 0.18±0.15, A-rule: 0.18±0.16). By contrast, units with suppressed FRs (60%; n=146) showed no significant change in decoding accuracy (Figure 7—figure supplement 3D; p=0.44, Z=0.77, V-rule: 67.32±15.25, A-rule: 66.53±14.61), but a 44% increase in encoding efficiency (p=1.8e-07, Z=–5.22, median FC: 1.44, V-rule: 0.13±0.12, A-rule: 0.18±0.19; paired WSR). These results suggest that the minority of units that increase FRs in the A-rule perform marginally better at decoding the auditory stimulus, and that the units that decrease FRs drive the shift in encoding efficiency.

To ensure that mice were adequately engaged and attentive in the task, the analyses described above excluded any trials in which the incorrect behavioral response was made. However, an examination of these error trials, which may correlate with lapses in attention, could provide insight into the moment-to-moment behavioral relevance of the attentional effects described above. We have shown that attention to sound is marked by a net suppression of pre-stimulus and evoked FRs. We hypothesized that, if this attentional modulation is behaviorally meaningful, FRs preceding A-rule error trials may be more similar to sound-unattended V-rule trials than to A-rule correct trials. We addressed this possibility by comparing pre-stimulus FRs in error vs. correct trials (300 ms prior to stimulus onset; Figure 8B). Because misses were uncommon (Figure 8A), we restricted our analysis to the comparison of FA and CR trials to allow for adequate sampling of each trial outcome. We included only behavior sessions with at least 10 FA and CR trials (A-rule and V-rule trials considered separately). This decreased unit sample sizes (n=234, 58 across all depth groups for BS, NS; min. group size = 9, 2 for BS, NS). Given the small sample of NS units and the likelihood of insufficient power, NS units were not included in this analysis. When considering BS units with increased FRs in the A-rule, we found no significant group-level difference between A-rule FA and CR trials at any cortical depth (Figure 8C; Figure 8—source data 1A for full stats; all p≥0.29, all |Z|≤1.43, paired WSR, FDR-adjusted p-values). However, deep cortical BS units with A-rule suppression showed significantly higher pre-stimulus FRs prior to A-rule FA trials than CR trials (deep BS [n=98]: mean FR difference between pre-stimulus FA and CR trials = 0.35 Hz, p=0.0098, Z=–3.15; paired WSR; other depth groups: p≥0.28, |Z|≤1.48; all p-values FDR-adjusted). This is unlikely to reflect a motor effect of higher FR before a lick, as it was specific to the A-rule: pre-stimulus FRs in A-rule-suppressed or A-rule-enhanced units did not differ between FA and CR trials in the V-rule (Figure 8C.c; Figure 8—source data 1B; paired WSR: all p≥0.68, all |Z|≤1.59, FDR-adjusted p-values). Together, these findings suggest that FR reductions typical of modality-selective attention directly relate to behavioral outcomes.

All experiments were approved by the Institutional Animal Care and Use Committee at the University of California, San Francisco. Twenty-seven C57BL/6 background male mice were surgically implanted with a headpost and began behavioral training, of which 10 completed the training and successfully performed the task during physiology recording sessions. All mice began the experiment between ages P56 and P84. Mice used in this report expressed optogenetic effectors in various subsets of interneurons, which we intended to use for optogenetic identification of cells (Lima et al., 2009; analysis not included here). These mice were generated by crossing an interneuron subpopulation-specific Cre driver line (PV-Cre JAX Stock Nr. 012358; Sst-Cre: JAX Stock Nr. 013044) with either the Ai32 strain (JAX Stock Nr. 012569), expressing Cre-dependent eYFP-tagged channelrhodopsin-2, or the Ai40 strain (JAX Stock Nr. 021188), expressing Cre-dependent eGFP-tagged archaerhodopsin-3. Of the 10 behavior mice included in this report, 6 were Ai32/Sst-Cre, 3 were Ai32/PV-Cre, and 1 was Ai40/Sst-Cre. In most experiments (n=21 recordings), brief, low-level optogenetic pulses during the ITI of the task were used to identify opsin-expressing neurons (<0.3 mW light; 5 light pulses of 10 ms duration, every ~1.5 min); these analyses are outside of the scope of this report. The optogenetic stimulation protocol was consistent through A- and V-rules of the task. Unit stimulus response FRs and behavioral response error rates were not statistically different between trials immediately after optogenetic pulses and stimulus-matched trials preceding the pulses.

All mice were housed in groups of 2–5 for the duration of the behavioral training until the craniotomy. Post-craniotomy and during physiology recordings, mice were housed singly (up to 6 days) to protect the surgical site. Mice were kept in a 12 hr/12 hr reversed dark/light cycle. All training occurred during the dark period, when mice show increased activity and behavioral task performance (Roedel et al., 2006).

Adult mice (>P56) were trained on an AV go/no-go rule-switching behavior task. In this task, mice were positioned on a floating spherical treadmill in front of a monitor and a speaker, and an optical computer mouse recorded treadmill movement. Mice licked to receive a reward depending on auditory, visual, or AV stimulus presentation (‘decision’ stimuli, either ‘target’ or ‘distractor’), but the modality predictive of the reward changed partway through the behavioral session. Each session would start with a unimodal go/no-go block, in which a series of auditory (A_R, A_U; 17 or 8 kHz TC) or visual (V_R, V_U; upward or rightward moving gratings) stimuli was presented. After stimulus presentation, mice signaled choice by either licking a spout in front of the mouth or withholding licking. Licking at the target unimodal stimulus would trigger a water reward, while licking at the distractor would trigger a short dark timeout. After a fixed number of unimodal trials, the stimuli would become AV, but the rule for which stimulus modality predicted reward would carry over from the unimodal block. All four stimulus combinations (A_RV_R, A_RV_U, A_UV_R, A_UV_U) would be presented in the AV block, such that two AV combinations would be target stimuli and two would be distractor. Then, after completing a fixed number of trials in the AV block, the task using the rule of the opposite modality would begin; a unimodal block with the other modality would start, followed by a second AV block using the rule from the preceding unimodal block. For any mouse, the stimuli predictive of the reward in each rule was kept constant across days and training sessions (e.g., a 17 kHz TC would always predict a reward in the A-rule, and a rightward grating would always predict a reward in the V-rule).

The task was self-paced using a virtual foraging approach, in which mouse locomotion (measured through treadmill rotation) would cause a track of randomly placed dots on the screen to move downward. After a randomly varied virtual distance, a decision stimulus would be presented, at which point the mouse would lick or withhold licking to signal choice. For receptive field mapping during physiology experiments, an RDS stimulus was presented in-between decision stimuli, during the inter-trial track portion. Stimuli are detailed below.

Prior to any training, mice were surgically implanted with a stainless steel headplate, used both for head fixation during the task and for physiology recordings after the task was learned (surgical methods described below). Three days post-implant, mice began a water restriction protocol based on previously published guidelines (Guo et al., 2014). Throughout the course of training, mice received a minimum water amount of 25 mL/kg/day, based on weight at time of surgical implant. After recovery from surgery, mice were given ~7 days to adjust to water restriction. Then, mice were head fixed and habituated to the floating treadmill for 15–30 min daily sessions with no stimulus presentation for 2–3 days. After mice appeared comfortable on the treadmill, a phased behavioral task training regimen began. Mice were trained once daily for ~6 days per week. On day 1, mice were introduced to an auditory-only (A-only) stimulus training version of the task in which A_R (‘target’/‘rewarded’) or A_U (‘distractor’/‘unrewarded’) stimuli were presented, and a reward would be automatically administered shortly after the onset of A_R. Next, the mice were put on an operant version of the A-only task, which required licking any time after the onset of A_R to receive a reward and withholding of licking during A_U to avoid a dark timeout punishment. Mice achieved proficiency, defined as 2 or more consecutive days of sensitivity index d’>1.5 (see Data analysis for calculation), on the A-only task after 11.0±4.7 days after start of training (median ± SD, n=10 successful mice). Then, a similar training structure was repeated for the visual task: V-only stimulus training with automatic rewards for V_R, but not V_U, followed by an operant version of the visual task requiring licks for rewards (median time to proficiency: 26.0±7.2 days after start). After learning the tasks for each modality separately, mice were introduced to an auditory-AV (A-AV) version, in which the rule from the auditory stimulus carried over to the AV block. This was intermixed with training days on a visual-AV (V-AV) version of the task. Number of training days on A-AV or V-AV were decided based on prior performance, with extra training given as needed. Mice were considered proficient at this stage after performing with d’>1.5 on each rule (A-AV; V-AV) on 2 consecutive days (median time to proficiency: 40.0±15.8 days after start). Finally, the full rule-switching task was introduced (Figure 1D), generally alternating between days of V-rule-first and the A-rule-first task sequences but allocating more training days to task orders as needed. Because physiology recordings were acute and strictly limited to 6 days after craniotomy, we set a greater threshold for expert-level performance on the full task before advancing to physiology: 3 consecutive days of d’>2.5 (median time to expertise: 90.5±31.8 days). Care was taken to train each mouse at a roughly consistent time of day (no more than ~1–2 hr day-to-day variation). During expert-level task performance, mice typically completed 260–300 trials in a daily session (30 A-only; 100–120 A-AV; 30 V-only; 100–120 V-AV).

The behavior training setup was controlled by two computers: a behavior monitoring and reward control PC (OptiPlex 7040 MT, Dell) and a dedicated stimulus presentation machine running Mac OS X (Mac Mini, Apple). Stimulus presentation was controlled with MATLAB using custom software (https://github.com/HasenstaubLab/AVtrainer-stim/tree/main/demo; copy archived at swh:1:rev:737720f41fd5302b90fd5e60a10822270381818c;path=/demo/; Morrill et al., 2022), and inter-machine communication used the ZeroMQ protocol. Auditory and visual stimuli were generated and presented using the Psychophysics Toolbox Version 3 (Kleiner et al., 2007). Water rewards were administered using a programmable syringe pump (NE-500, New Era Pump Systems, Farmingdale, NY), positioned outside of the sound-attenuating recording chamber. Early in training, water reward volume was set at 0.01 mL per correct response, but over training the reward volume was gradually decreased to 0.006 mL to achieve greater trial counts. Licking events were recorded using a custom photobeam-based lickometer circuit based on plans provided by Evan Remington (Xiaoqin Wang Lab, Johns Hopkins University). Licks were registered when an IR photobeam positioned in front of the lick tube was broken, queried at a sample rate of 100 Hz by an Arduino Uno microcontroller (Arduino, LLC).

Animals in this experiment underwent two surgeries: first, before training a surgery to implant a custom steel headplate over the temporal skull using dental cement was conducted. The animal was anesthetized using isoflurane and a headplate was implanted over AC, ~2.5 mm posterior to bregma and under the squamosal ridge, to allow for physiology recordings after achieving task expertise. When mice completed the training regimen outlined above, a craniotomy surgery was performed. The animal was again anesthetized using isoflurane and an elliptical opening (0.75 mm wide × 1.5 mm long) was made in the skull over AC using a dental drill. This opening was promptly covered with silicone elastomer (Kwik-Cast, World Precision Instruments), and the animal was allowed to recover overnight. The following day, the animal was affixed by its headplate over the treadmill inside of a sound-attenuating recording chamber, the silicone plug over the craniotomy was removed, and the craniotomy was flushed with saline. A silver chloride ground wire was placed into the craniotomy well at a safe distance from the exposed brain. A 64-channel linear probe (20 µm site spacing; Cambridge Neurotech, Cambridge, UK) was slowly inserted in the brain using a motorized microdrive (FHC, Bowdoin, ME) at an approximate rate of ~1 μm/s (Fiáth et al., 2019). After reaching the desired depth, the brain was allowed to settle for 10 min, after which the water spout, lickometer, visual stimulus delivery monitor, and speaker were positioned in front of the mouse, and the behavior session commenced. Behavior sessions were sometimes stopped early and restarted due to poor performance. In approximately half of behavior-physiology sessions (13 of 23 successful recordings), the task was stopped due to low performance after the rule transition and restarted at the beginning (unimodal block) of the second rule. To control for possible effects of task order, attempts were made to counterbalance recordings from A-rule first (15) and V-rule first (8) behavior sessions.

After completion of the behavior task, the water spout and lickometer were removed, and a series of auditory and/or visual passive experiments were conducted in order to characterize the response properties of the recording site. All stimuli were presented with the auditory and visual stimulation apparatus described above. Following completion of these experiments, the probe was slowly removed, and the brain was covered with a thin layer of freshly mixed 2.5% agarose in saline, followed by a layer of silicone elastomer. The animal was returned to its home cage, and the following day the physiological recording process was repeated. Recordings were made for up to 6 days after the craniotomy. The neural signal acquisition system consisted of an Intan RHD2000 recording board and an RHD2164 amplifier (Intan Technologies), sampling at 30 kHz.

In-task auditory decision stimuli were 1 s TCs, consisting of 50 ms tone pips overlapping by 25 ms, with frequencies in a 1-octave band around either 17 or 8 kHz. TCs were frozen for the duration of the task, so that each mouse always heard the same pip sequences, allowing for direct comparisons of sound-evoked neural responses across rules without concern that stimulus peculiarities may be driving observed differences. TCs were presented at 60 dB SPL. Visual decision stimuli consisted of a circular moving grating stimulus (33° diameter subtended visual space), which appeared at the center of the screen for 1 s (coincident with TC stimulus during bimodal presentation). Gratings moved either upward or rightward with a 4 Hz temporal frequency, 0.09 cycles/degree spatial frequency at 50% contrast. In-between decision stimulus presentations, an RDS stimulus was presented for receptive field mapping (Bigelow et al., 2022; Gourévitch et al., 2015). The RDS comprised two uncorrelated random sweeps that varied continuously and smoothly between 4 and 64 kHz, with a maximum sweep modulation frequency of 20 Hz. RDS stimuli were presented at 50 dB SPL.

After the behavior task, passive auditory search stimuli (pure tones, click trains) were presented to characterize response properties of the electrode channel. Click trains consisted of broadband 5 ms white noise pulses, presented at 20 Hz for 500 ms duration. Pure tone stimuli consisted of 100 ms tones of varied frequencies (4–64 kHz, 0.2 octave spacing) and sound attenuation levels (30–60 dB in 5 dB linear steps), with an interstimulus interval of 500 ms.

Auditory stimuli were presented from a free-field electrostatic speaker (ES1, Tucker-Davis Technologies) driven by an external soundcard (Quad-Capture or Octa-Capture, Roland) sampling at 192 kHz. Sound levels were calibrated using a Brüel & Kjær model 2209 meter and a model 4939 microphone. Visual stimuli were presented on a 19-inch LCD monitor with a 60 Hz refresh rate (Asus VW199), positioned 25 cm in front of the mouse and centered horizontally and vertically on the eyes of the mouse. Monitor luminance was calibrated to 25 cd/m² for a gray screen, measured at approximate eye level for the mouse.

All statistical calculations were performed in MATLAB r2019a and its Statistics and Machine Learning Toolbox, V11.5. For group comparisons of SU responses across task rules, paired WSR tests were used, unless otherwise noted. Because tests were performed separately on each depth and spike waveform subpopulation, the Benjamini-Hochberg FDR procedure was used to correct for multiple comparisons, typically across n=6 comparisons (three depth groups, two spike waveform groups; Benjamini and Hochberg, 1995). This method relies on controlling the Type I error rate (here, q=0.05), providing increased power over typical family-wise error rate controls. To determine if individual SUs were significantly modulated by rule, an unpaired Student’s t-test on FR was used with a threshold of p<0.01. Descriptive statistics reported in text are mean ± standard deviation (SD), unless otherwise noted. Fractional change values between task rules are reported as the median of the A-rule/V-rule. All other statistical tests are described in Results. Sample sizes (n) are indicated for each comparison in Results or source data files.

To test the accuracy of our depth estimation method based on physiological responses (Figure 3), we presented the pure tone search stimuli described above to a separate set of untrained control mice while performing extracellular recordings (n=11 recordings from four mice; Ai32/Sst-Cre). Before insertion, the probe was painted with the fluorescent lipophilic dye Di-I (DiCarlo et al., 1996; Morrill and Hasenstaub, 2018). The depth measurement procedure based on physiological signals was carried out as described above, and then probe tracks from each recording were visualized as described previously (Morrill and Hasenstaub, 2018). Briefly, after recordings, the animal was euthanized, and the brain was removed and placed into a solution of 4% PFA in PBS (0.1 m, pH 7.4) for 12 hr, followed by 30% sucrose in PBS solution for several days. The brain was then frozen and sliced using a sliding microtome (SM2000R, Leica Biosystems) and slices were imaged with a fluorescence microscope (BZ-X810, Keyence). Di-I probe markings showing cortical depth were consistent with physiological activity-based depth measurements described above (Figure 3B–C).