Selective attention to auditory and visual modalities converges onto noncholinergic basal forebrain neurons

Abstract

Selective attention enables animals and humans to prioritize behaviorally relevant stimuli among competing sensory inputs. Although the basal forebrain (BF) is known to modulate cortical activity and support attention, it remains unclear whether BF activity directly conveys an attention signal. Here, we show that selective attention to auditory and visual stimuli converges onto a shared population of noncholinergic BF neurons. Using a cross-modal task where rats rapidly switched attention between modalities, we found that these neurons responded strongly to attended targets but weakly to the same stimuli when ignored, regardless of modality. These effects closely tracked both task-driven and spontaneous attention shifts on a single-trial basis. Moreover, BF responses reflected the linear summation of attended and ignored inputs, suggesting that sensory streams are filtered in parallel before converging in the BF. These findings suggest that the BF may serve as a subcortical hub integrating attention signals across modalities to guide adaptive behavior.

During selective attention, sensory streams are filtered in parallel and summed linearly in the basal forebrain.

INTRODUCTION

Selective attention is a fundamental cognitive function that enables animals and humans to prioritize behaviorally relevant sensory inputs while filtering out distractions (1–5). At its core, selective attention hinges on the ability to select one stimulus from among competing sensory streams. Most research into the neural mechanisms underlying selective attention has focused on corticothalamic circuits, where attention enhances sensory representations of target stimuli, suppresses competing inputs, and enhances perception and behavioral performance (5–9). A potential source of these attentional modulations lies in subcortical neuromodulatory systems, which project broadly to the cortex and exert powerful influence on network dynamics (10–18). However, whether the activity of these subcortical systems is itself modulated by selective attention, and therefore can directly convey an attention signal, remains largely unknown.

A prime candidate among subcortical neuromodulatory systems is the basal forebrain (BF), which has long been implicated in attention (17–24). BF dysfunction is also linked to cognitive decline in aging (25, 26) and in neurodegenerative disorders such as Parkinson’s (27–29) and Alzheimer’s disease (23, 25, 30, 31). Although BF research has historically focused on its cholinergic neurons (18, 20, 32, 33), recent studies have identified a distinct population of noncholinergic neurons within the BF—referred to as BF bursting neurons (16, 34–39)—that exhibit properties consistent with a neural correlate of attention. These neurons respond robustly to reward-predicting stimuli that animals must attend to (34, 35, 37), and their activation enhances behavioral performance (34–37) and drives strong local field potential (LFP) responses in the frontal cortex (16, 38, 40). Despite these properties, it remains unknown whether BF bursting neuron activity is directly modulated by selective attention. If it is, does selective attention toward different sensory modalities engage distinct subsets of BF bursting neurons, as commonly observed in corticothalamic circuits (4, 5, 41–45) and in BF cholinergic neurons (46, 47)? Or, alternatively, do attention signals across different modalities converge onto a shared population of BF bursting neurons?

To address these questions, we developed a cross-modal selective attention paradigm for rats, allowing us to study rapid shifts of attention between auditory and visual modalities, including both task-driven and spontaneous switches. We found that all BF bursting neurons are strongly modulated by selective attention, responding to the same sensory inputs only when they are attended, whereas their responses were drastically weaker when the same stimuli were ignored. These shifts in BF activity occurred rapidly within single trials and were tightly coupled with behavioral performance. Critically, unlike attention signals in corticothalamic circuits, selective attention signals toward auditory and visual modalities converge onto the same BF bursting neurons, with their responses reflecting the linear summation of inputs from both sensory streams. These findings establish noncholinergic BF neurons as a subcortical hub that integrates selective attention across sensory modalities and suggest that BF bursting neurons may shape corticothalamic dynamics during selective attention, regardless of the attended modality.

RESULTS

Behavioral evidence of selective attention in a cross-modal oddball task

To study selective attention, we designed an audiovisual oddball task for rats that required fast switching of attention between auditory and visual modalities (Fig. 1A). Every 2 s, an auditory (A) and a visual (V) stimulus were presented concurrently for half a second at clearly perceptible intensity levels. Each modality consisted of two possible stimuli and followed an oddball task format, where most stimuli were frequent standard stimuli (std, S_A or S_V), whereas the infrequent oddball stimuli (odd, O_A or O_V) served as potential targets for reward. Stimuli were randomly selected within each modality, resulting in four stimulus configurations: O_AO_V, O_AS_V, S_AO_V, and S_AS_V, which were presented throughout the session. A session consisted of several trial blocks, during which one sensory modality was designated as the rewarding modality (bkA or bkV). Water rewards were available at two ports in the operant chamber (portA and portV), with O_A predicting reward availability at portA only during bkA and O_V predicting reward availability at portV only during bkV. After completing 20 to 40 correct trials, the rewarding sensory modality switched without any overt sensory cues to signal the block transition (see Materials and Methods for details). This task design encouraged the animals to dynamically shift attention between auditory and visual modalities in response to the same stimulus configurations.

Fig. 1. Rapid switching of selective attention between auditory and visual modalities in a cross-modal oddball task.

(A) Task schematic: Auditory (A) and visual (V) stimuli were presented simultaneously from the center every 2 s. Each modality was either a frequent standard (S) or rare oddball (O), yielding four configurations (O_AO_V, O_AS_V, S_AO_V, and S_AS_V). The relevant modality alternated across blocks (auditory-relevant, bkA; visual-relevant, bkV) after 20 to 40 correct trials without explicit cues. In bkA, O_A predicted reward at portA; in bkV, O_V predicted reward at portV. Rewards were delivered only if rats licked the correct port within 2 s. (B) Selective attention to audition implies equivalent portA choices (chA) for O_AO_V and O_AS_V, whereas attention to vision implies equivalent portV choices (chV) for O_AO_V and S_AO_V. (C) Top: Reward contingencies by stimulus, choice, and block. Bottom: Choice probabilities for all four configurations in bkA and bkV (dots, rats; bars, means), with two-way ANOVA results shown for each configuration [not significant (n.s.), P > 0.05; **P < 0.01; ***P < 0.001]. (D) Example session: oddball-trial choices and outcomes across three blocks, with block transitions and spontaneous attention switches indicated. Parentheses denote switches to the nonrelevant modality [e.g., A(V)A]. (E and F) During both block transitions (E) and spontaneous switches (F), p(chA|O_AS_V) closely matched p(chA|O_AO_V) and p(chV|S_AO_V) matched p(chV|O_AO_V). Wilcoxon signed-rank tests with Bonferroni correction are shown above panels (colored line: n.s., P > 0.05; *P < 0.05). Because blocks switch only after a correct response, O_AS_V never immediately precedes V→A transition nor S_AO_V precedes A→V, producing blank entries in (E). Choice probabilities are means ± SEM. Only animals with ≥30 switching episodes were included. The number of animals (N) and total transitions (n) are reported as N(n) in each panel.

The key rationale behind this task design is that, if animals selectively attend to one sensory modality while ignoring the other, their behavioral choices should be driven primarily by the oddball stimulus in the attended modality, with little to no influence from stimuli in the unattended modality. Specifically, if animals selectively attend to the auditory modality while ignoring visual inputs, they should respond similarly to the two stimulus configurations containing O_A—O_AO_V and O_AS_V—by choosing portA (chA) (Fig. 1B). Conversely, if animals selectively attend to the visual modality while ignoring auditory inputs, they should respond similarly to the two stimulus configurations containing O_V—O_AO_V and S_AO_V—by choosing portV (chV) (Fig. 1B). Thus, similar choice patterns between stimulus configurations featuring the same attended oddball stimulus but differing in the unattended modality would provide a strong behavioral signature for selective attention.

We trained adult Long-Evans rats (N = 26, 20 male and 6 female; table S1) to perform this task, collecting behavioral data from over 2.8 million stimulus presentations across 1577 sessions. Behavior performance was highly robust, and the vast majority of sessions (1453/1577; 92%) met the performance criteria (fig. S1). Because behavioral response patterns were qualitatively similar between male and female rats (fig. S2), subsequent experiments and all results in the main text were conducted on male rats (N = 20) to maximize the number of trials performed (fig. S1). Correct response rates to rewarded oddball stimuli averaged 87.2 ± 3.4% in auditory blocks [p(chA|O_A, bkA)] and 82.3 ± 4.5% in visual blocks [p(chV|O_V, bkV)] (means ± SD; N = 20) (fig. S1).

Despite identical stimulus configurations across trial blocks, behavioral response patterns toward each stimulus configuration differed significantly between block types [two-way analysis of variance (ANOVA), significant main effects of block type and/or a block × choice interaction, P < 0.01] (Fig. 1C). In O_AS_V and S_AO_V trials, rats responded when the oddball was in the rewarding modality (O_AS_V in bkA; S_AO_V in bkV) by preferentially choosing the corresponding port (chA; chV, respectively) but refrained from responding when the same stimuli were not rewarded (O_AS_V in bkV; S_AO_V in bkA) (Fig. 1C). In O_AO_V trials, rats responded at high rates in both block types, but their choice of port depended on the rewarding modality (chA in bkA; chV in bkV) (Fig. 1C). These results indicate that the attended sensory modality shifted across blocks.

To evaluate whether animals displayed behavioral signatures of selective attention, we first analyzed choice patterns during block transitions, where the rewarded modality switched without explicit cues. Despite the abrupt changes in reward contingencies, rats rapidly adjusted their choices to favor the newly rewarded modality (Fig. 1, D and E). Critically, choice probabilities were nearly identical between p(chA|O_AS_V) and p(chA|O_AO_V), as well as between p(chV|S_AO_V) and p(chV|O_AO_V) (Fig. 1E and fig. S3) (N = 19, Wilcoxon signed-rank test with Bonferroni correction, P > 0.05). These results match the predicted behavioral signature of selective attention (Fig. 1B), indicating that behavioral choices were driven predominantly by the oddball stimulus in the corresponding modality (chA by O_A; chV by O_V), with minimal influence from the unattended modality—even when the unattended stimulus was also an oddball.

Although these findings provide strong evidence for cross-modal selective attention during task-driven shifts at block transitions, one potential concern is that these behavioral changes could be explained entirely by updates in reward contingencies. If rats were genuinely deploying selective attention, they should also be able to shift attention independently of reward-driven task structure. To test this, we identified episodes of “spontaneous switches” of attention—brief stretches of trials in which rats consistently chose the nonrewarded oddball stimulus with high probability, despite the fact that these choices were not rewarded (fig. S4). These episodes occurred during periods when reward contingencies remained unchanged, either within a stable trial block or at the beginning of a session before any rewards were delivered (Fig. 1D and fig. S4),

These spontaneous attention shifts exhibited the same behavioral signatures of selective attention observed during block transitions. Choice probabilities during these unrewarded switches remained nearly identical between p(chA|O_AS_V) and p(chA|O_AO_V), as well as between p(chV|S_AO_V) and p(chV|O_AO_V), despite their rapid and transient nature (Fig. 1F and fig. S2) (N = 9 to 15, Wilcoxon signed-rank test with Bonferroni correction, P > 0.05). Furthermore, response latency patterns during spontaneous switches closely matched those in rewarded trials (fig. S5), supporting the interpretation that rats were actively attending to the nonrewarded modality.

Together, these findings demonstrate that rats flexibly and spontaneously deploy cross-modal selective attention in this task—even in the absence of explicit cues or reward. These results also establish the behavioral signature of selective attention: Choosing portA in the presence of O_A (chA|O_A) reflects attention toward the auditory modality, whereas choosing portV in the presence of O_V (chV|O_V) reflects attention toward the visual modality, consistent with the predictions outlined in Fig. 1B.

BF bursting neuron responses are strongly modulated by selective attention

To investigate the role of the BF in selective attention, we recorded BF neuronal activity while rats performed the cross-modal oddball task (Fig. 2A). Of the recorded neurons, 40% (582/1454) were classified as BF bursting neurons, defined by their strong phasic excitatory responses to reward-predicting stimuli and low baseline firing rates (<10 spikes/s) (Fig. 2B and fig. S6), according to established criteria (16, 34–37). This phasic response is locked to stimulus onset but not to movement events such as port entry or licking (fig. S7). Typical response patterns of BF bursting neurons during the task are illustrated by an example neuron (Fig. 2, C and D) and summarized at the population level (Fig. 2E).

Fig. 2. Selective attention to auditory and visual targets strongly modulates BF activity and converges onto the same neurons.

(A) Bilateral BF recording sites across coronal sections (N = 8 rats); colors denote animals; coordinates relative to bregma. Atlas images adapted with permission from (68). (B) BF bursting neurons (red, n = 582) were defined by low baseline firing and large stimulus-evoked responses in rewarded trials; other BF neurons are gray (n = 872). (C) Example bursting neuron: behavior and spike rasters aligned to oddball onset show block-dependent responses. [s] marks a spontaneous A(V)A switch. (D) Same neuron, rasters and peristimulus time histograms (PSTHs) are shown for the four stimulus configurations across bkA and bkV. (E) Population heatmaps of individual BF bursting neuron PSTHs for each configuration in bkA and bkV, sorted identically by responses to O_AS_V in bkA. Population PSTHs (means ± SEM) below. Significant bkA versus bkV differences are marked (*P < 0.05, Wilcoxon, Bonferroni corrected). (F) Selective attention modulates responses to O_AS_V [F(a)] and S_AO_V [F(b)]. Left: Baseline-subtracted population PSTHs across four conditions: attended in rewarded block, rewarded block average, nonrewarded block average, and miss in the nonrewarded block. The 0.05- to 0.25-s window quantifies amplitude. Middle: Normalized amplitudes (means ± SEM) relative to the attended-rewarded condition with pairwise tests (**P < 0.01; ***P < 0.001, Wilcoxon). Right: Scatterplots compare each neuron’s attended versus miss response. (G) Competing models of modality specificity: divergent (distinct subpopulations) versus convergent (shared neurons). (H) Convergence is supported by a strong across-neuron correlation between normalized responses in O_AS_V-bkA-chA and S_AO_V-bkV-chV trials (see Materials and Methods for normalization). Neurons sorted into quartiles (Q1 to Q4) based on O_AS_V-bkA-chA responses show similarly graded responses across conditions, despite notable latency differences. (I) Correlation matrix across attended oddball conditions. The boxed cell corresponds to (H).

Despite encountering the same stimulus configurations across trial blocks, BF bursting neurons exhibited distinct response patterns to stimulus onset across the two block types (Fig. 2C). These cross-block modulations arose from different response profiles to each stimulus configuration between block types (Fig. 2, D and E) (Wilcoxon signed-rank test with Bonferroni correction, P < 0.05).

A particularly prominent feature is that BF responses to the same stimulus were significantly enhanced when the stimulus was attended and substantially diminished when ignored. This pattern was most evident in O_AS_V and S_AO_V trials: BF bursting neurons exhibited robust excitation when the oddball stimulus predicted reward (O_AS_V in bkA; S_AO_V in bkV) but showed markedly reduced responses when the same stimuli did not predict reward (O_AS_V in bkV; S_AO_V in bkA) (Fig. 2F) (Wilcoxon signed-rank test, P < 10⁻⁹⁵ for O_AS_V: bkA all versus bkV all; P < 10⁻⁹⁶ for S_AO_V: bkA all versus bkV all). Within the rewarded block, BF responses were even stronger when animals attended to the oddball stimulus by choosing the corresponding port (Fig. 2F) (P < 10⁻⁷⁶ for O_AS_V-bkA: chA versus all; P < 10⁻⁹² for S_AO_V-bkV: chV versus all). When these rewarded trials were used as the basis for normalization (100%) and baseline firing rate was set to 0%, BF responses to the same stimulus in the unrewarded block fell to 26.2% (O_AS_V) and 8.9% (S_AO_V), which further dropped to 15.9% (O_AS_V) and 0.1% (S_AO_V) in miss trials (P < 10⁻⁸¹ for O_AS_V-bkV: all versus miss; P < 10⁻⁹³ for S_AO_V-bkA: all versus miss) (Fig. 2F). These findings underscore the strong modulation of BF bursting neuron activity in response to identical audiovisual sensory inputs, reflecting the powerful influence of selective attention.

Auditory and visual selective attention converge onto the same BF neurons

Given the strong attentional modulation observed in BF bursting neurons, a key question is how selective attention signals toward different sensory modalities are organized within the BF. In corticothalamic circuits, selective attention typically recruits distinct neural populations based on sensory modality (4, 5, 41–45). Whether this principle also applies to subcortical systems like the BF, however, remains unclear (Fig. 2G). Do separate subsets of BF bursting neurons mediate attention to auditory and visual modalities, as in corticothalamic circuits? Or do attention signals from different modalities converge onto the same BF neurons, such that individual neurons respond similarly to attended targets regardless of sensory modality?

To address this question, we first compared BF bursting neuron activity under two fully nonoverlapping conditions: O_AS_V-bkA-chA and S_AO_V-bkV-chV trials. These trial types differed in stimulus configuration (O_AS_V versus S_AO_V), block type (bkA versus bkV), and behavioral choice (chA versus chV) (Fig. 2H). Despite these differences—as well as notable differences in BF response latencies—the response amplitudes of individual BF bursting neurons were highly correlated across conditions (r = 0.81, P < 10⁻¹³², Pearson correlation) (Fig. 2H; see fig. S8 for additional analysis). Furthermore, when BF bursting neurons were grouped into quadrants based on their response amplitudes in O_AS_V-bkA-chA trials, their responses in S_AO_V-bkV-chV trials followed a similarly graded pattern (Fig. 2H). This pattern of convergence was not limited to this specific pair of conditions. Similar cross-condition correlations in response strength were observed across all attended oddball configurations (O_AS_V, S_AO_V, and O_AO_V), such that neurons that responded strongly under one attended condition also responded strongly in others (Pearson r = 0.78 to 0.95) (Fig. 2I). These findings demonstrate that, unlike corticothalamic circuits where attention is implemented via modality-specific subpopulations, selective attention signals from both auditory and visual modalities converge onto the same population of BF bursting neurons.

Parallel processing of attended and ignored stimuli before converging in the BF

If BF bursting neuron activity reflects the convergence of selective attention signals for attended auditory and visual oddball stimuli, does this convergence also occur within the same trial when both auditory and visual oddball stimuli are presented (O_AO_V), but only one is attended while the other is ignored? And how does a simultaneously presented, ignored oddball influence the response to the attended one? We considered three competing scenarios: (i) divisive normalization, a well-established computation when concurrent stimuli compete for processing (48, 49), predicting a subadditive response (O_AO_V < attended + ignored); (ii) synergy, classically observed in superior colliculus during multisensory integration (50), predicting a superadditive response (O_AO_V > attended + ignored); and (iii) parallel processing with near-linear summation, predicting an approximately additive response (OAOV ≈ attended + ignored).

To test these alternatives, we introduced 20% unimodal trials in a subset of sessions while keeping reward contingencies unchanged (Fig. 3B). In these trials, either the auditory or visual stimulus was omitted (denoted as Ø_A or Ø_V), allowing us to directly compare BF responses to identical attended oddball stimuli under unimodal versus bimodal conditions. Behaviorally, rats’ choice patterns were highly similar between unimodal and bimodal trials when the attended oddball was in the rewarded modality, regardless of whether a competing stimulus was present in the unattended modality (Wilcoxon signed-rank test, P > 0.05 between O_AO_V, O_AS_V, and O_AØ_V in bkA; P > 0.05 between O_AO_V, S_AO_V, and Ø_AO_V in bkV; N = 10 rats) (Fig. 3C). This finding indicates that the presence or absence of a stimulus in the unattended modality did not meaningfully influence behavior toward the attended target.

Fig. 3. Converging inputs from attended and ignored stimuli are summed linearly in BF bursting neurons.

(A) Three competing models for how the ignored oddball affects BF responses to the attended oddball: divisive normalization (subadditive; O_AO_V < attended + ignored), synergy (superadditive), and parallel processing with near-linear summation (approximately additive; O_AO_V ≈ attended + ignored). (B) Cross-modal oddball task with 20% unimodal trials (arrows) created four extra configurations: O_AØ_V, Ø_AO_V, S_AØ_V, and Ø_AS_V. Reward contingencies in unimodal trials were identical to those in bimodal trials (Fig. 1C). (C) Mean choice probabilities for bimodal and unimodal trial types (N = 10 rats, 53 sessions; dots, rats; bars, means) were similar when the same oddball was rewarded (Wilcoxon signed-rank, P > 0.05), indicating minimal influence from the unattended stream. (D) BF responses in Attend-A (O_A-chA) conditions: O_AO_V, O_AS_V, and O_AØ_V. [D(a)] Heatmaps of single BF bursting neurons (n = 94), sorted by O_AS_V-chA amplitude. [D(b)] Baseline-subtracted population PSTHs (means ± SEM). Statistical comparisons (Wilcoxon signed-rank test with Bonferroni correction) are indicated above (n.s., P > 0.05 shown in colored lines; *P < 0.05). BF responses to the attended O_A were largely unaffected by the ignored O_V. [D(c)] Response amplitudes in O_AO_V-chA strongly correlated with O_AØ_V-chA (Pearson r = 0.96). Each dot represents one neuron. [D(d)] Residual differences between O_AO_V and O_AS_V trials closely resembled miss responses in S_AO_V and Ø_AO_V, consistent with a small additive O_V term. [D(e)] Scatterplot comparing response amplitudes in S_AO_V-miss and the residual (O_AO_V − O_AS_V) (both chA). Each dot represents one neuron. [D(f)] A schematic illustrates parallel filtering of auditory and visual inputs with attention enhancing auditory inputs (and/or suppressing visual inputs), followed by linear summation in BF bursting neurons. (E) Same analyses as in (D) but for Attend-V (O_V-chV) conditions.

At the neuronal level, we first asked whether an ignored oddball alters the response to the attended oddball. Mirroring behavior, BF bursting neuron activity was nearly identical across unimodal and bimodal conditions when the auditory oddball was attended [O_AO_V, O_AS_V, O_AØ_V, and chA; Fig. 3, D(a) to D(c)] and likewise when the visual oddball was attended [O_AO_V, S_AO_V, Ø_AO_V, and chV; Fig. 3, E(a) to E(c)]. At the population level, both the amplitude and timing of the peak response were comparable across conditions (Wilcoxon signed-rank test with Bonferroni correction, P > 0.05 at peak response) [Fig. 3, D(b) and E(b)]. At the single-neuron level, response amplitudes were highly correlated between unimodal and bimodal trial types (r = 0.96 for both Attend-A and Attend-V conditions) [Fig. 3, D(c) and E(c)]. These findings demonstrate that BF bursting neuron responses are dominated by the attended stimulus and show little evidence for divisive normalization or synergy.

If attended and ignored oddballs were processed in parallel and summed linearly, the residual (O_AO_V − attended) should match the ignored oddball response. The ignored oddball exerted a small but measurable influence on BF activity [Fig. 3, D(b) and E(b)]. When the ignored modality was visual, the difference in BF activity between O_AO_V-chA and O_AS_V-chA trials was minor and emerged later [Wilcoxon signed-rank test with Bonferroni correction, P < 0.05; Fig. 3D(b)], closely resembling BF responses to ignored visual oddballs in S_AO_V and Ø_AO_V miss trials (Wilcoxon signed-rank test with Bonferroni correction, P > 0.05 between trial types across all lags) [Fig. 3D(d)]. In contrast, when the ignored modality was auditory, the response difference between O_AO_V-chV and S_AO_V-chV trials was more pronounced and occurred earlier [Wilcoxon signed-rank test with Bonferroni correction, P < 0.05; Fig. 3E(b)], closely resembling BF responses to ignored auditory oddballs in O_AS_V and O_AØ_V miss trials (Wilcoxon signed-rank test with Bonferroni correction, P > 0.05 between trial types across all lags) [Fig. 3E(d)]. These effects were also reflected at the single-neuron level, with a significant correlation in response amplitudes when the ignored stimulus was the auditory oddball (r = 0.38, P < 0.001, Pearson correlation) [Fig. 3E(e)] but not when it was the visual oddball (r = 0.15, P = 0.136) [Fig. 3D(e)], consistent with its negligible impact on BF activity. These results are consistent with the parallel processing scenario where the ignored stream contributes a small additive term.

Together, trials containing both oddballs do not show super- or subadditive interactions. Instead, BF bursting neuron responses reflect the linear summation of two parallel inputs: a strong excitatory response to the attended stimulus and a weaker response to the ignored one, which are processed independently before converging onto BF bursting neurons [Fig. 3, D(f) and E(f)]. Similar response patterns were also observed in sessions without unimodal trial types (fig. S9), reinforcing the conclusion that attended and ignored sensory inputs converge onto the same BF bursting neurons within individual trials.

Binary BF response modes dynamically track rapid attention shifts

Having established that BF bursting neurons are strongly modulated by selective attention across trial blocks (Fig. 2F), we next asked whether this BF attention signal could track rapid shifts of attention. If BF activity reflects selective attention to the target stimulus, it should track such shifts even when they are spontaneous, nonrewarded, and occur infrequently.

We began by examining BF bursting neuron activity in O_AS_V and S_AO_V trials and found that BF responses fell into two distinct modes based on the animal’s behavioral choice. When rats selected the port corresponding to the oddball modality—i.e., when they attended to the oddball stimulus (e.g., O_AS_V with chA; S_AO_V with chV)—BF bursting neurons exhibited strong phasic responses regardless of whether the choice was rewarded (Fig. 4A and fig. S10). In contrast, when rats chose the noncorresponding port (e.g., O_AS_V with chV; S_AO_V with chA), BF responses were weak and resembled those seen in miss trials, where the stimulus was ignored (Fig. 4A). By normalizing BF response amplitudes as in Fig. 2F, we revealed a bimodal distribution of single-trial response amplitudes, reflecting a categorical distinction between trials in which the stimulus was attended (high BF response) and those in which it was ignored (low BF response near baseline).

Fig. 4. Binary BF response amplitudes track rapid shifts of attention between attended and ignored stimuli.

(A) BF bursting neurons show two response modes in O_AS_V [A(a)] and S_AO_V [A(b)] trials depending on whether the oddball was attended or ignored. Left: Baseline-subtracted population PSTHs for each condition (means ± SEM, n = 582). Shaded regions denote the windows for quantifying response amplitudes. Right: Single-trial amplitude distributions reveal a clear bimodality consistent with “attend” (high) and “ignore” (low) modes. Response amplitudes were calculated by pooling BF bursting neurons within each session and normalized per session—separately for O_AS_V and S_AO_V trials—with 100% as the mean rewarded-trial response and 0% as baseline firing. Each distribution was normalized to have unit area. (B) BF responses closely track rapid attention shifts during block transitions. [B(a)] Behavioral choice probabilities (top) and corresponding normalized BF activity (bottom) for O_AS_V (blue) and S_AO_V (red) aligned to the transition (means ± SEM; number of transitions shown in panels). [B(b)] Choice probabilities and BF amplitudes are strongly correlated across lags; each dot is one trial lag from [B(a)]. (C) Same analysis as in (B) for spontaneous switches. (D and E) BF bursting neurons alternate between attended and ignored response modes across all trial positions, regardless of how frequently a choice type occurred. Normalized BF amplitudes from [B(a)] and [C(a)] are replotted after splitting trials by whether animals attended to the oddball (correct port) or ignored it (incorrect or miss). Data are means ± SEM using five-trial bins to improve robustness when certain choice types were rare. Significant differences between attended and ignored trials are marked with * (Wilcoxon rank sum test with Bonferroni correction, P < 0.05). BF activity is consistently elevated when the oddball is attended and suppressed when ignored, even when those choices are infrequent.

To test whether BF bursting neuron activity could track attention shifts on rapid timescales, we analyzed BF responses in O_AS_V and S_AO_V trials during two types of transitions: (i) block transitions, in which reward contingencies changed, and (ii) spontaneous switches, in which attention shifted despite stable contingencies and the resulting choices were not rewarded (Fig. 4, B and C). In both scenarios, BF response amplitudes closely mirrored behavioral choice patterns [p(chA|O_AS_V) and p(chV|S_AO_V)], increasing when the oddball stimulus was attended and decreasing when it was ignored. Notably, this tight coupling persisted even during brief, unrewarded spontaneous switches (Fig. 4C; see also examples in Fig. 2, C and D). Across both transition types, the correlation between behavioral choice patterns and BF responses was exceptionally high (Pearson r = 0.93 to 0.99) [Fig. 4, B(b) and C(b)].

Crucially, this binary pattern of BF activity remained robust across nearly all trial positions, regardless of the frequency of specific choice types (Wilcoxon signed-rank test with Bonferroni correction, P < 0.05 between attend and ignore conditions across all trial lags) (Fig. 4, D and E). Even when certain choice types were rare, attended choices (e.g., O_AS_V-chA in bkV or S_AO_V-chV in bkA) consistently evoked strong BF responses, whereas ignored choices (e.g., O_AS_V-ignore in bkA or S_AO_V-ignore in bkV) consistently evoked weak responses. These findings demonstrate that BF bursting neurons rapidly and reliably alternate between two distinct response modes—attend and ignore—on a single-trial basis, reflecting the animal’s real-time attention toward the stimulus.

BF response patterns reliably decode the attended modality in single trials

Last, we investigated whether BF bursting neuron activity reflects not only whether animals are attending to a stimulus but also which sensory modality they are attending to. Previous analyses showed that BF responses to auditory and visual oddballs differ in their temporal dynamics (Fig. 3, D and E). We therefore asked whether BF temporal response patterns could be used to decode the attended modality at the single-trial level.

To test this, we focused on O_AO_V trials, in which both auditory and visual oddballs were presented simultaneously, but animals selectively attended to one modality. In these trials, the temporal pattern of BF bursting neuron activity reliably aligned with the animals’ choice, regardless of whether the choice was rewarded (Fig. 5A). To quantitatively analyze these patterns, we applied principal components analysis (PCA) to single-trial BF activity across sessions. The first two principal components (PCs) captured key features of the response: PC1 reflected overall activity magnitude, whereas PC2 encoded temporal dynamics that distinguished auditory-attended versus visual-attended trials (Fig. 5A).

Fig. 5. BF temporal response patterns track rapid shifts of the attended modality.

(A) Top left: Trial-by-trial BF bursting neuron activity in O_AO_V trials from one session (#1), grouped by choice (chA or chV) and outcome (rewarded or nonrewarded). BF activity patterns showed distinct temporal dynamics depending on the attended modality. Top right: First two PCs (PC1 and PC2) derived from single-trial population activity within 200 ms of stimulus onset. Bottom left: O_AO_V trials projected onto PC1-PC2 form three clusters—attend-A (chA), attend-V (chV), and miss. Colored dots represent individual trials from the example session; black outlines indicate nonrewarded choices. Bottom right: PC2 score distributions for the four O_AO_V trial types show a clear bimodal separation by attended modality, independent of reward outcome. (B) BF responses from all oddball trial types (O_AS_V, S_AO_V, and O_AO_V) across all sessions projected onto PC1-PC2. Three trial clusters emerge: auditory-attended (O_A-chA), visual-attended (O_V-chV), and stimulus-ignored (miss, O_A-chV, and O_V-chA). PC1 discriminated attended from ignored trials (AUC = 0.94), whereas PC2 distinguished attend-A versus attend-V (AUC = 0.96). (C) AUC values for PC1 and PC2 in pooled data (black squares) and individual sessions (box plots). All three trial clusters were reliably discriminated in individual sessions. (D) Selective attention index (top) and PC2 values (middle) during block transitions, pooled across all oddball trials. The selective attention index was computed as p(chA|O_A) − p(chV|O_V). PC2 values computed from O_AO_V trials (dotted line) closely tracked those from all oddball trials (solid line). Bottom: PC2 values were strongly correlated with the selective attention index during block transition (Pearson r = 0.99). (E) Same as (D) but for spontaneous switches of attention. PC2 reliably tracked dynamic shifts in selective attention between the two sensory modalities, even during brief, spontaneous attention switches (Pearson r = 0.96).

This PCA framework allowed us to analyze all oddball trial types in a unified space. BF responses formed three distinct clusters: auditory-attended trials (O_A-chA, including O_AS_V-chA and O_AO_V-chA), visual-attended trials (O_V-chV, including S_AO_V-chV and O_AO_V-chV), and stimulus-ignored trials (including misses, O_AS_V-chV, and S_AO_V-chA) (Fig. 5B). PC1 reliably distinguished attended (including attend-A and attend-V) from ignored trials [area under the curve (AUC) = 0.94, receiver operating characteristic (ROC) analysis] (Fig. 5B). PC2, on the other hand, discriminated between auditory-attended and visual-attended trials (AUC = 0.96, ROC analysis) (Fig. 5B). This cluster topography was highly consistent across animals and sessions (fig. S11), and different trial types could be reliably distinguished in individual sessions (Fig. 5C). PC1 separates attended from ignored trials because individual BF bursting neurons themselves discriminate these states, with decoding strength scaling with their bursting amplitude (fig. S12), whereas PC2 captures the robust auditory-earlier-than-visual latency shift underlying the Attend-A versus Attend-V separation (fig. S13).

To test whether BF activity could track dynamic changes in the attended modality, we examined how PC2 values varied during dynamic attention shifts. We defined a selective attention index as p(chA|O_A) − p(chV|O_V), capturing shifts between auditory and visual attention. PC2 values tracked this index with high fidelity during both block transitions (Pearson r = 0.99; Fig. 5D) and spontaneous attention switches (Pearson r = 0.96; Fig. 5E).

Together, these findings demonstrate that BF activity patterns not only track whether animals are attending to a stimulus (PC1) but also reliably indicate which sensory modality is attended (PC2). These results establish BF bursting neuron activity as a robust neural correlate of selective attention at the single-trial level across sessions and animals.

BF outcome responses are inversely related to stimulus-evoked responses

Prior works show that BF bursting neurons exhibit reward prediction error (RPE)–like coding, with responses to reward-predicting cues and to reward outcomes varying negatively with each other (34, 35, 37, 51, 52). Although our task was not designed to test RPE explicitly, we likewise observed an inverse relationship between BF activity at stimulus onset and at trial outcome (fig. S14). Thus, while conveying a selective-attention signal, the activity of BF bursting neurons continues to express an RPE-like component.

DISCUSSION

In this study, we establish that noncholinergic BF bursting neurons serve as a subcortical hub integrating selective attention signals across auditory and visual modalities. By using a cross-modal oddball task (Fig. 1), we show that BF bursting neurons selectively respond to target stimuli only when they are attended, regardless of whether their sensory modality is auditory or visual (Fig. 2). Moreover, information from attended and ignored stimuli converges onto BF bursting neurons and is combined linearly (Fig. 3). This BF attention signal closely tracks rapid attention shifts and reliably decodes both whether the stimulus is attended (Fig. 4) and which modality is attended (Fig. 5), positioning this group of noncholinergic BF neurons as a key subcortical hub for integrating selective attention across modalities to promote reward-seeking behavior.

The insights gained in this study were made possible by the cross-modal selective attention task we developed, which presents competing auditory and visual stimuli simultaneously while requiring rapid shifts of attention across sensory modalities. Although many rodent paradigms exist for studying distinct aspects of attention—such as sustained attention, signal detection, and set-shifting (53–58)—few are designed to probe stimulus competition and the selective prioritization of one input over another. Notably, some recent paradigms have used trial-by-trial cueing to precisely control attentional allocation among competing stimuli, offering powerful insights into the prefrontal-thalamic mechanisms that enable flexible sensory gating (5, 43, 44). Our task complements this approach by adopting a block-based design, allowing for the characterization of both instructed and spontaneous shifts of attention in the absence of explicit cues (Fig. 1, E and F). Furthermore, by demonstrating that behavioral choices are unaffected by stimuli in the unattended modality, our task provides robust behavioral evidence for the deployment of cross-modal selective attention (Fig. 1, E and F).

Using this paradigm, we showed that BF bursting neuron activity is strongly modulated by selective attention. Compared to trials in which the same O_AS_V or S_AO_V stimuli were attended, BF bursting neuron responses were reduced to just 15.9% (O_AS_V) and 0.1% (S_AO_V) when the stimuli were ignored (Fig. 2F). The fact that BF activity could switch almost instantaneously between two distinct response modes (Fig. 4) is particularly notable, given that both auditory and visual stimuli were perceptually salient. A similarly strong attentional modulation was observed in O_AO_V trials, where BF responses were dominated by the attended stimulus, whereas the stimulus in the unattended modality contributed minimally (Fig. 3, D and E). These findings highlight the powerful and rapid modulation of BF activity by selective attention.

A central finding of our study is that selective attention signals from auditory and visual modalities converge onto the same BF bursting neurons (Fig. 2, G to I), whose activity reflects the linear summation of inputs from attended and ignored sensory channels (Fig. 3, D and E). Given that BF bursting neurons showed little response to standard stimuli in either modality (Fig. 2E) and that BF responses in O_AS_V and S_AO_V trials were indistinguishable from those in unimodal O_AØ_V and Ø_AO_V trials (Fig. 3, D and E), the principle of linear summation observed in O_AO_V trials (Fig. 3, D and E) can be extended to all stimulus configurations—for example, the response to O_AS_V can be understood as the sum of responses to O_A and S_V stimuli. These findings impose important constraints on upstream circuit organization, indicating that auditory and visual inputs are processed in parallel, each shaped by a strong and dynamic attention filter, before converging onto the same BF bursting neurons [Fig. 3, D(f) and E(f)]. Thus, BF bursting neurons likely do not implement the selection stage itself; rather, their activity aligns with a late-selection/target-detection stage of selective attention (2, 9, 59, 60), providing a modality-common, post–selection amplification that promotes reward-seeking actions. Clarifying this division within selective attention—early selection versus late target detection—and placing BF bursting neurons in the latter is a key contribution of this study.

A key feature of the convergence is that BF bursting neurons exhibiting stronger responses to one modality also show stronger responses to the other (Fig. 2, H and I), suggesting a common circuit mechanism that governs their responses across sensory modalities. This modality-common attention signal contrasts sharply with selective attention mechanisms described in corticothalamic circuits, where neurons are typically tuned to specific modalities or features (4, 5, 41–45). The convergence of auditory and visual attention signals onto the same BF neurons suggests that BF activity alone does not specify which modality is attended. Instead, we were able to decode the attended modality from BF activity only when aligned to stimulus onset, by leveraging differences in the temporal response profiles of BF neurons (Fig. 5 and fig. S13). In this sense, the distinct temporal dynamics of BF bursting neurons serve as a readout of selective attention, reflecting which sensory input has gained priority and access to the BF. This interpretation resolves the apparent paradox: Although BF bursting neurons respond to attended targets regardless of sensory modality, their responses still convey information about the attended modality through their timing. A key implication of this temporal difference is that BF modulation of cortical processing would also occur with different latencies depending on the attended modality. Without recognizing the role of a modality-common attention signal originating in the BF, such latency-dependent cortical modulations could easily be misattributed to modality-specific cortical mechanisms—thereby overlooking a potentially critical subcortical contribution to attentional processing.

Our observation that bimodal and unimodal trials produced similar behavioral performance and BF activity patterns (Fig. 3) suggests that the relationship between BF activity and behavioral output remains consistent, regardless of whether competing sensory inputs are present. This continuity allows us to bridge the current findings with previous studies of BF bursting neurons that used unimodal stimulus presentations (16, 34–38). In those studies, stronger BF bursting neuron activity was associated with increased reward-seeking behavior and faster reaction times, both in well-trained behavioral tasks (34, 35) and during new learning (34, 37). Conversely, inhibition of BF bursting neuron activity was linked to nogo behavior (34) and rapid behavioral stopping (36). Although the present work is correlational, earlier electrical stimulations of the BF produced specific changes in BF bursting neurons and support their causal role in behavior (35, 36). These findings are consistent with the idea that BF bursting neurons function as a bidirectional gain control mechanism—facilitating reward-seeking actions when activated and promoting behavioral suppression when inhibited (61). The current study extends this framework to selective attention, suggesting that BF bursting neurons serve as a crucial link between attended stimuli and reward-seeking behavior, independent of sensory modality.

Although our task was not explicitly designed to test whether BF bursting neurons encode RPE, we nevertheless observed an inverse relationship between BF responses to stimulus onset and trial outcomes (fig. S14)—a hallmark of RPE coding and consistent with prior reports (34, 35, 37, 51, 52). This implies that BF bursting neurons carry an RPE-like component while simultaneously conveying a selective-attention signal. We propose that the two perspectives are not mutually exclusive because reward-predicting stimuli inherently capture attention, blurring the distinction between these constructs especially in reward-based decision-making tasks. Reconciling these interpretations requires recognizing the dynamic evolution of RPE during learning. Initially, RPE signals are tied to reward delivery, but as learning progresses, they shift to the onset of reward-predicting stimuli. This progression aligns with well-established distinctions in learning theory (62) between “attention for learning” and “attention for performance.” The latter—attention guided by well-learned, reward-predicting stimuli—is less emphasized in the RPE literature but is the focus of our study. Our data suggest that BF bursting neurons may transform RPE signals at stimulus onset into attention signals that enhance cortical processing of reward-predicting stimuli. This mechanism would support faster and more accurate behavioral responses by aligning cortical dynamics with task-relevant priorities.

Converging anatomical and physiological evidence points to the prefrontal cortex (PFC) as a principal downstream target through which BF bursting neurons may influence adaptive behavior. Previous studies (16, 40) show that phasic activation of BF bursting neurons was tightly coupled, on a trial-by-trial basis, to a positive LFP deflection in deep PFC layers—the primary recipient zone of BF inputs—with closely matched timing and amplitude. Moreover, brief BF electrical stimulation evokes a layer-specific PFC LFP response that mirrors this BF-linked signal. Together, these observations indicate that BF bursting neurons deliver a fast, phasic drive to deep-layer PFC, positioning the PFC as a key node through which BF signals modulate cortical processing and guide behavior.

Last, although most studies of the BF have focused on cholinergic neurons (21, 63), BF bursting neurons represent a distinct and functionally specialized population. Optogenetically identified BF cholinergic neurons respond to reward and punishment at shorter latencies (33, 64) (10 to 30 ms) compared to BF bursting neurons (34–37) (~50 ms) and exhibit faster instantaneous bursting rates (33, 64) (>200 spikes/s) compared to BF bursting neurons (34, 39) (<100 spikes/s). In addition, BF cholinergic neurons display prominent state-dependent firing rate modulations, increasing during wakefulness and decreasing during slow-wave sleep (33, 65, 66), whereas BF bursting neurons show relatively stable activity across brain states (34). Our current findings further highlight their distinction: Whereas BF cholinergic neurons are tuned to specific sensory features (46, 47), BF bursting neurons are not and instead convey a modality-common attention signal. Together, these physiological and functional differences support the idea that BF bursting neurons are noncholinergic neurons (24, 34, 38). BF bursting neurons likely correspond to cortically projecting GABAergic cells that may enhance cortical activity through a disinhibitory mechanism (61). Identifying molecular markers for BF bursting neurons using single-cell transcriptomics will be critical for enabling cell type–specific optogenetic perturbation to directly test their causal role in selective attention.

MATERIALS AND METHODS

Subjects

Twenty male and six female adult Long-Evans rats (National Laboratory Animal Center, Taiwan), aged 3 to 6 months and weighing 250 to 500 g, were used for behavioral testing. Eight of the male rats were further used for electrophysiological recordings. See table S1 for animal contributions to each analysis. Rats were housed under a 12-hour/12-hour light/dark cycle and provided with 18 g of food/day, with unrestricted access to water. During training and recording procedures, rats were water restricted to maintain 85 to 90% of their baseline body weight and trained in daily 60-min sessions during the light cycle. Water-restricted rats received 15 min of water access at the end of each training day, with unrestricted access on weekends. All experimental procedures were approved by the Institutional Animal Care and Use Committee at the National Yang Ming Chiao Tung University, Taiwan (NYCU, IACUC #1080510).

Apparatus

Behavioral experiments were conducted in custom-built Plexiglass operant chambers [29.4 cm (L) by 24 cm (W) by 31.3 cm (H)], custom-built by LabTalk Ltd. (Tainan, Taiwan) housed within light- and sound-attenuating cubicles [70 cm (L) by 43.5 cm (W) by 54 cm (H)]. Each cubicle was equipped with an exhaust fan (#9S1224M401, Sanyo Denki America Inc.) that also served to mask external noise. The front panel of each chamber contained three modules: a central fixation port and two flanking reward ports. The fixation port was equipped with an infrared (IR) sensor module (IR emitter #SFH 4851, ams OSRAM; IR receiver #BPW76A, Vishay) to detect animals’ snout entry. The reward port was equipped with two IR sensor modules, positioned to detect reward-port entry and sipper-tube licking, respectively. Water rewards were delivered via sipper tubes located within the reward ports, controlled by solenoid valves (#003-0111-900, Parker Hannifin Corp., Hollis, NH), calibrated to dispense 10 μl per drop.

A speaker (#2240, Visaton) was mounted above the central fixation port and controlled by a programmable audio generator (ANL-926, Med Associates Inc.). A white light-emitting diode (LED) lamp (#VC1860245W3D, VCC) positioned above the speaker served as the visual oddball stimulus (O_V). Two warm white incandescent light bulbs (#CM1819, VCC) located immediately above the central fixation port served as the visual standard stimulus (S_V). In addition, an incandescent light bulb was positioned above each reward port to serve as a block-type hint signal during behavioral shaping. Behavioral training was controlled by Med-PC software (Version V, Med Associates Inc.), which recorded all event timestamps with 1-ms resolution and sent TTL signals to the neurophysiology recording systems.

Cross-modal oddball task

Rats were trained to perform a cross-modal oddball task in which an auditory (A) and a visual (V) stimulus were presented concurrently for 0.5 s every 2 s. Stimuli in both modalities were delivered at clearly perceptible intensities, were easily distinguishable from each other and were presented above the central fixation port without any lateral bias. Within each sensory modality, one stimulus was designated as an infrequent oddball, whereas the other served as a frequent standard. The auditory oddball (O_A) was a 6-kHz pure tone (70 dB), whereas the auditory standard (S_A) was white noise (65 dB) for most animals, except in two cases where the S_A was a 10-kHz pure tone (70 dB). Both O_A and S_A were delivered from the same speaker. The visual oddball (O_V) was a diffuse white LED lamp positioned at the top of the central fixation panel, whereas the visual standard (S_V) consisted of two warm white incandescent bulbs located immediately above the central fixation port.

Oddball stimuli within each modality appeared approximately once every five stimulus presentations, with inter-oddball intervals (IOIs) of 6, 8, 10, 12, or 14 s, selected pseudorandomly. A new IOI was drawn after each oddball presentation, and IOIs were determined independently within each modality. This design resulted in four possible stimulus configurations: O_AS_V, S_AO_V, O_AO_V, and S_AS_V (Fig. 1A). One of the two reward ports was designated for auditory oddball rewards (portA; the left port was assigned as portA in 19/26 rats), whereas the other was designated for visual oddball rewards (portV).

To receive a water reward, rats had to lick the correct reward port within a 2-s response window following the onset of an oddball stimulus in the rewarded modality. The rewarded modality alternated between trial blocks: In auditory-relevant blocks (bkA), only O_A was rewarded, requiring rats to respond to O_AS_V or O_AO_V by choosing portA (chA). In visual-relevant blocks (bkV), only O_V was rewarded, requiring responses to S_AO_V or O_AO_V at portV (chV) (Fig. 1, B and C). Correct responses led to the delivery of three drops of water, dispensed at the third, fourth, and fifth licks. The timing of the third lick was considered the trial outcome delivery time.

During correct licking that led to reward delivery, IOI timers for both modalities were temporarily suspended, and S_AS_V was presented until the rat withdrew from the reward port and refrained from reentering for at least 0.15 s, at which point the IOI timers for both modalities resumed. Incorrect licking responses—including licking at the incorrect port or responding to an incorrect stimulus configuration—were not rewarded and reset the IOI timers for both modalities. During incorrect licking, S_AS_V was presented until licking ceased, at which point IOI timers resumed. Each block consisted of 20 to 40 correct trials, determined pseudorandomly. Block transitions occurred without any explicit sensory cues, requiring animals to dynamically track changes in reward contingencies and shift attention accordingly.

In addition to the main task (Task 1), in which oddball stimuli in the rewarded modality were always rewarded (100% probability), we implemented two task variations. In Task 2, the reward probability for the oddball stimulus in the rewarded modality was reduced to 90%, introducing occasional unrewarded trials. In Task 3, 20% of trials were randomly selected as unimodal trials, where stimuli were presented in only one sensory modality while keeping reward contingencies unchanged (Fig. 3). Behavioral performance for these variations are detailed in fig. S1. See table S1 for how each task variant contributed to the analyses.

Although the overall behavioral response patterns were qualitatively similar between male and female rats (figs. S1 and S2), female rats exhibited lower correct response probabilities and completed fewer trials than male rats (fig. S1). To maximize the number of trials performed, recording experiments included only male rats. Given the similar behavioral performance across the three task variants (fig. S1) and the comparable BF neuronal responses between bimodal and unimodal trials (Fig. 3), recording data from all three task variants were pooled for BF activity analyses. The only exception is Fig. 3, which exclusively used data from Task 3.

Throughout this study, the terms “target,” “oddball,” and “trial” are used interchangeably to refer to stimulus configurations containing oddball stimuli that could serve as targets in this task (O_AS_V, S_AO_V, and O_AO_V). Given that standard stimulus presentations (S_AS_V) elicited minimal behavioral and neural responses in both block types, S_AS_V trials were excluded from all analyses unless otherwise stated.

Behavioral shaping procedure

Rats underwent a structured behavioral shaping protocol consisting of four sequential training stages over 2 to 3 months before advancing to the full cross-modal oddball task. These training stages gradually introduced key task elements while maintaining the overall reward contingencies of the final task. Beginning in stage 2, the overall task structure and parameters were identical to the final cross-modal oddball task, except for the specific modifications described in each stage.

Stage 1: Reward port training

In the initial stage, rats learned to associate the reward ports with water delivery. In each trial, one of the two reward ports was randomly assigned to dispense water, signaled by the illumination of a hint light (an incandescent bulb) above the corresponding port. To receive a reward, rats were required to lick the illuminated port, triggering the delivery of nine drops of water, starting with the first lick. The trial ended when the rat withdrew from the port and refrained from reentering for 2 s, at which point an intertrial interval (ITI) was initiated, pseudorandomly set to 6, 8, or 10 s. Each session lasted 40 min or until the rat completed 112 rewarded trials. Rats were required to reach the maximum number of rewarded trials in two consecutive sessions before advancing to the next stage. The hint lights introduced in this step were later used to indicate the rewarding modality in subsequent training stages.

Stage 2: Oddball-reward association in individual modalities

At this stage, rats learned to associate each oddball stimulus (O_A and O_V) with its corresponding reward port (portA and portV). The standard stimuli (S_A and S_V) were not yet introduced at this stage. Only one sensory modality was presented per trial block, which also defined the rewarding modality.

To facilitate learning, hint lights were initially used to help rats associate oddball stimuli with the correct reward port. In early training sessions, the hint light remained illuminated throughout the block to indicate the rewarded port. As rats improved, the hint lights were turned off after the third rewarded trial in each block. In the final phase of training, the hint lights were removed entirely, requiring the rats to memorize the correct reward port for each modality.

Stage 3: Ignoring standard stimuli in individual modalities

At this stage, standard stimuli (S_A and S_V) were introduced for the first time. The training format remained the same, with each trial block restricted to a single sensory modality. This stage reinforced the requirement that rats ignore standard stimuli and respond only to oddball stimuli.

Stage 4: Bimodal stimuli with hint lights

In the final shaping stage, auditory and visual stimuli were presented concurrently, as in the full cross-modal oddball task. Initially, the hint light remained illuminated throughout the block, similar to stage 2. After rats reached a performance threshold, the hint lights were turned off after the third rewarded trial in each block. Once rats performed consistently, they were transitioned to the full cross-modal oddball task without the hint lights.

Stereotaxic surgery and electrode implantation

After reaching stable performance in the cross-modal oddball task, animals underwent stereotaxic surgery for chronic electrode implantation targeting the bilateral BF. Surgery was conducted under isoflurane anesthesia (1 to 5%) following protocols similar to those used in our previous studies. Multiple skull screws were inserted to anchor the implant, with one screw positioned over the cerebellum serving as the common electrical reference and another over the opposite cerebellar hemisphere serving as the electrical ground. Craniotomies were made at Anterior-Posterior (AP): −0.6 mm, Medial-Lateral (ML): ±2.25 mm relative to bregma to target the BF bilaterally.

The electrode assembly consisted of two bundles, each containing 14 to 16 polyimide-insulated tungsten wires (38 μm in diameter; California Fine Wire). Each bundle was enclosed in a 28-gauge stainless steel cannula and controlled by a precision microdrive, allowing gradual electrode advancement postsurgery. Electrode impedances were ~0.1 MΩ, measured at 1 kHz using an Open-Ephys Acquisition Board (open-ephys.org).

During implantation, electrode bundles were retracted inside the cannula, and the tips of the cannulae were lowered to an initial depth of 6.25 mm below the cortical surface using a micropositioner (Robot Stereotaxic, Neurostar GmbH) at controlled speeds of 2 to 50 μm/s. Once the target depth was reached, the electrode assembly and skull screws were secured with dental cement (Hygenic Denture Resin). Electrodes were then gradually advanced to a final depth of 7.25 mm below the cortical surface.

Neural data acquisition and preprocessing

Following surgery, rats received ibuprofen (0.3 to 0.5 mg/ml in drinking water) for pain management and topical antibiotics to prevent infection. They were allowed 1 week to recover with ad libitum access to food and water before resuming behavioral training. Rats were retrained in the cross-modal oddball task before neuronal recordings commenced. BF neuronal activity was recorded daily during task performance for at least 2 weeks at the same electrode depth. Electrodes were advanced in 62.5-μm increments, followed by a 1-week recovery period to allow for tissue stabilization before further recordings. Only sessions recorded at distinct depths or under different task variations were included for analysis.

Neural signals were referenced to a skull screw placed over the cerebellum, band-pass filtered (0.3 Hz to 7.5 kHz), amplified, and digitized at 30 kHz using the Open-Ephys Acquisition Board (open-ephys.org). Spike waveforms were sorted offline using the KlustaKwik sorting algorithm, followed by manual curation in a custom Python GUI. Single units were identified on the basis of clear separation from the noise cluster and minimal spike collisions (<0.1% of spikes occurring within a 1.5-ms interspike interval). In addition, cross-correlation analyses were performed to eliminate duplicate units recorded simultaneously across multiple electrodes. Units with average firing rates of less than 0.1 spikes/s were also removed.

At the end of the experiment, animals were deeply anesthetized with sodium pentobarbital (90 to 150 mg/kg, ip) and perfused transcardially with normal saline followed by 10% formalin. Implants were carefully removed, brains were extracted from the skull and postfixed for 24 hours in formalin at +4°C. Frozen fixed brains were cut using a sliding microtome (SM2010R, Leica) into 50-μm-thick coronal sections. Electrode placements were verified via dark-field microscopy (M205 FCA, Leica) on unstained histological sections (67). All electrodes were confirmed to be within the expected target range: AP [0, −1.2] mm, ML [1.8, 3] mm relative to bregma, and DV [7.2, 8.5] mm relative to the cortical surface (Fig. 2A).

Data analysis

Data were analyzed using custom Matlab (R2019b, MATLAB The MathWorks Inc., Natick, MA) scripts.

Session selection criteria

Sessions must meet the following criteria to be included for analyses: (i) p(reward) [defined as p(chA|O_A, bkA) and p(chV|O_V, bkV)] must be higher than 50% in both blocks; (ii) >100 rewarded trials; (iii) at least four trial blocks (fig. S1).

Define behavioral response types

Behavioral response types were defined on the basis of licking behavior. Therefore, we first identified licking clusters, defined as groups of consecutive licks at the same reward port with interlick intervals shorter than 0.45 s and no intervening withdrawal from the port (minimum of one lick). The first lick within each cluster was designated as the behavioral response timestamp and used to determine the chosen direction. Stimulus presentations in which the first lick occurred within the 2-s response window were classified as chA or chV trials, depending on the chosen reward port. Stimulus presentations where no lick occurred within this period were classified as miss trials. These definitions formed the basis for calculating choice probabilities across different stimulus configurations and trial types. Response latency was defined as the time from stimulus onset to the first lick (fig. S5).

During the cross-modal oddball task, any licking temporarily suspended the IOI timers, ensuring that only standard stimuli (S_AS_V) were presented during the licking period. Consequently, any S_AS_V presentations that occurred while animals were licking were excluded from all analyses involving S_AS_V (e.g., Fig. 2, D and E).

Identify spontaneous switches of attention

To identify periods when rats spontaneously switched their attention to the nonrewarding modality, we first defined a choice block as a sequence of trials in which the animal consistently licked the same port. Each choice block began and ended with a lick to the same port, had a minimum length of one trial, and included intervening no-response trials (fig. S4). Only choice blocks that contained at least one oddball stimulus presentation and no rewarded trials were considered in the analysis. By focusing exclusively on nonrewarded switching behavior, we ensured that animals made choices based on attentional shifts rather than reward contingencies.

To identify spontaneous switching episodes, we applied the following criteria based on the characteristics of nonrewarded choice blocks (fig. S4): (i) no trials within the choice block were rewarded; (ii) the choice direction differed from the currently rewarded modality; (iii) the choice block contained at least two attended choice trials (O_A-chA or O_V-chV); (iv) the choice probability in oddball trials within the block [p(chA|O_A) or p(chV|O_V)] was at least 50%. These criteria allowed us to isolate sustained switching episodes where animals consistently chose the nonrewarded port in response to the nonrewarded oddball across multiple trials with a high probability (≥50%), ruling out random exploratory behavior. Using these criteria, we identified 3909 such episodes across 1315 sessions (fig. S4, D to F).

Episodes were further classified on the basis of whether the animal had received any reward earlier in the session. Episodes that occurred before the first reward were categorized as “spontaneous switches at session start” (fig. S2B). For episodes occurring after at least one reward had been obtained, we applied two additional screening criteria: (I) the start and end of the choice block occurred at least five trials away from a task block transition and (ii) the switch could not occur during the task block transition period and must have occurred after the animal received the first reward in the new block. Episodes meeting these additional criteria were classified as “spontaneous switching of attention within a block” (Fig. 1F and fig. S2A). All subsequent analyses focused exclusively on trials containing oddball stimuli, whereas S_AS_V were excluded because they elicited little behavioral or neural responses.

For clarity, we refer to these spontaneous switches of attention using notation that reflects the sequence of attended modalities, with parentheses indicating nonrewarded choices. For example, A(V)A represents a switch to the nonrewarded visual modality during an auditory block. Spontaneous switching of attention within a block included A(V)A and V(A)V switching types, whereas spontaneous switching at session start included (A)V and (V)A switching types.

Dynamics around rapid behavioral switching

To analyze behavioral dynamics surrounding rapid switching of attention, we focused exclusively on oddball trials (O_AO_V, O_AS_V, and S_AO_V) and excluded standard stimuli (S_AS_V), which elicited minimal behavioral and neural responses. Our goal was to evaluate choice behavior in oddball trials as a function of their temporal position relative to a switching episode.

Each switching episode was aligned to its transition trial. For block transitions (A→V or V→A), alignment was based on the first oddball trial in the post–transition block. Because rats were required to make a correct choice to trigger a block transition, the final trial before a transition was always an O_V-chV trial (S_AO_V-chV or O_AO_V-chV) in V→A transitions or an O_A-chA trial (O_AS_V-chA or O_AO_V-chA) in A→V transitions (see Fig. 1E). Consequently, some oddball types (e.g., O_AS_V in the final trial before a V→A transition, or S_AO_V before an A→V transition) were never observed at trial lag −1, precluding statistical comparison at these positions. These gaps were indicated by blank spaces in the statistical comparison bars in Fig. 1E and fig. S2A1.

For spontaneous switching episodes within a block [A(V)A or V(A)V], trials were aligned to the first oddball trial in the choice block corresponding to the nonrewarded modality—(V) and (A), respectively. For spontaneous switches at session start [(A)V or (V)A], alignment was based on the first oddball trial in the V or A choice block, respectively.

We analyzed up to 20 oddball trials before and after each transition trial, assigning each trial a relative trial lag. If a task block transition occurred during the flanking period of a spontaneous switch episode, oddball trials from the other block were marked as NaN and excluded from analysis at those lags.

Choice probability for each stimulus type (O_AO_V, O_AS_V, and S_AO_V) was computed at each trial lag by averaging across all switching episodes for the relevant trial type [e.g., for p(chA|O_AS_V), only O_AS_V trials were included], whereas other trial types were treated as NaN. Choice probabilities surrounding block transitions and spontaneous attention switches (Fig. 1, E and F, and fig. S2A1) were first averaged across episodes within each animal and then across animals.

Because of the unequal frequencies of oddball trial types (O_AS_V:S_AO_V:O_AO_V ≈ 4:4:1), we required that each animal contribute at least 30 switching episodes (either block transitions or spontaneous attention switches) to be included in the analysis. This criterion ensured that even the least frequent trial type (O_AO_V) yielded approximately three observations per trial lag per animal. The number of animals (N) and total switching events (n) are reported in the relevant figure panels. Statistical comparisons between trial types were performed using the Wilcoxon signed-rank test (signrank.m).

For analyses of spontaneous switching in female rats (fig. S2A2) and at session start in male rats (fig. S2B), due to the lower number of transition episodes (n = 126 to 195), choice probabilities were averaged across all episodes pooled from all animals. Statistical comparisons between trial types were performed using the 1-degree-of-freedom chi-square test (crosstab.m). Trial lags with fewer than five O_AO_V trials were excluded from the analysis.

Identification of BF bursting neurons

BF bursting neurons were defined as BF single units exhibiting a stimulus-evoked increase in firing rate of more than 2 spikes/s during the [0.05, 0.2]-s window after stimulus onset in both rewarded O_AS_V and S_AO_V trials, compared to their baseline firing rate measured in the [−1, 0]-s window relative to all stimulus onsets. In addition, BF bursting neurons were required to have a baseline firing rate between 0.1 and 10 spikes/s based on previously established criteria (16, 34–37). Across 56 recording sessions (N = 8 rats), a total of 1454 BF single units were recorded, of which 40% (582/1454) were classified as BF bursting neurons (fig. S6).

BF response amplitudes and their normalization

Response amplitudes of BF bursting neurons were calculated as the average firing rate within a specified post–event time window, minus the baseline firing rate. This approach captures the change in neural activity relative to baseline. The time windows used for these calculations are indicated in the relevant figures. For stimulus-evoked responses, the analysis window was set to 0.05 to 0.25 s after stimulus onset. For responses to rewarded outcomes, the window was 0.05 to 0.25 s after the first drop of water delivery. For unrewarded trials, the window was 0.05 to 0.5 s following the expected time of outcome, defined as the time of the third lick. To compare responses across modalities and trial types (Fig. 2, H and I), each bursting neuron’s response amplitude was divided by the mean amplitude of all simultaneously recorded BF bursting neurons within the same session and trial condition. This normalization emphasizes each neuron’s response relative to its peers while reducing the variability from trial type–specific performance differences.

To compute BF responses in single trials, we pooled activity from all BF bursting neurons recorded in a given session. When pooling BF responses across sessions, normalization was applied to account for variability between sessions. For stimulus responses to O_AS_V and S_AO_V trials (Fig. 4), BF response amplitudes were normalized within each trial type and session by setting the average response in rewarded trials to 100% and the baseline firing rate to 0%.

For analyses requiring pooled data across all trial types—such as the PCA (Fig. 5) and outcome response analysis (fig. S14)—we applied a more general normalization procedure to account for variabilities in stimulus-evoked response latencies across sessions, especially in responses to O_V stimuli. Specifically, we first identified the peak response latency to stimulus onset in rewarded O_V trials for each session. We then calculated the average firing rate in a time window spanning −0.1 to +0.06 s relative to this peak latency across all rewarded trials (including all oddball trial types). This window reliably captured peak activity for both attended O_A and O_V trials. The resulting response amplitude was defined as 100%, and the baseline firing rate was set to 0%.

PCA of BF response patterns

PCA was applied to single-trial BF bursting neuron responses in O_AO_V trials to examine BF temporal response patterns distinguishing attention to auditory versus visual modalities. For each session, we first identified the peak response latency to stimulus onset in rewarded O_V trials. Then, for every O_AO_V trial, we extracted the BF population activity in a time window spanning −0.12 to +0.08 s relative to this peak latency. These temporal response vectors were baseline subtracted and amplitude normalized as described in the previous section. Normalized single-trial response vectors from all sessions were then subjected to PCA (pca.m). PCA was performed without mean-centering to ensure that the origin of the PCA space corresponded to baseline (zero activity). The first two PCs (PC1 and PC2) accounted for 48.5 and 12.7% of the total variance, respectively. To project other trial types into the same PCA space, we applied the same normalization procedure to single-trial BF activity in O_AS_V, S_AO_V, and O_AO_V trials and projected them onto the PC axes derived from the O_AO_V data.

ROC analysis and AUC

To quantify the separability of trial types based on PC1 and PC2, we performed ROC analysis and computed the AUC using the perfcurve.m function in Matlab.

Statistics

All statistical analyses were performed using the Statistics and Machine Learning Toolbox (version 11.6) in Matlab R2019b. The Wilcoxon signed-rank test (signrank.m) was used to compare paired trial types (e.g., behavioral performance between conditions), with Bonferroni correction applied for multiple comparisons. The Wilcoxon rank sum test (ranksum.m) was used for comparisons between independent groups. In cases where the number of switching episodes was too low to generate reliable per-animal estimates (e.g., fig. S2), a 1-degree-of-freedom chi-square test (crosstab.m) was used to compare choice probabilities. Pearson correlation coefficients were computed using corrcoef.m to quantify relationships between variables. Two-way ANOVA (anova2.m) was used to assess the main effects and interactions of choice type, block type, and their interaction (Fig. 1E and fig. S1).

Supplementary Materials

This PDF file includes:

Figs. S1 to S14

Table S1

References