Abstract
Age-related hearing loss (presbycusis) affects one-third of the world's population. One hallmark of presbycusis is difficulty hearing in noisy environments. Presbycusis can be separated into two components: the aging ear and the aging brain. To date, the role of the aging brain in presbycusis is not well understood. Activity in the primary auditory cortex (A1) during a behavioral task is because of a combination of responses representing the acoustic stimuli, attentional gain, and behavioral choice. Disruptions in any of these aspects can lead to decreased auditory processing. To investigate how these distinct components are disrupted in aging, we performed in vivo 2-photon Ca2+ imaging in both male and female mice (Thy1-GCaMP6s × CBA/CaJ mice) that retain peripheral hearing into old age. We imaged A1 neurons of young adult (2-6 months) and old mice (16-24 months) during a tone detection task in broadband noise. While young mice performed well, old mice performed worse at low signal-to-noise ratios. Calcium imaging showed that old animals have increased prestimulus activity, reduced attentional gain, and increased noise correlations. Increased correlations in old animals exist regardless of cell tuning and behavioral outcome, and these correlated networks exist over a much larger portion of cortical space. Neural decoding techniques suggest that this prestimulus activity is predictive of old animals making early responses. Together, our results suggest a model in which old animals have higher and more correlated prestimulus activity and cannot fully suppress this activity, leading to the decreased representation of targets among distracting stimuli.
SIGNIFICANCE STATEMENT Aging inhibits the ability to hear clearly in noisy environments. We show that the aging auditory cortex is unable to fully suppress its responses to background noise. During an auditory behavior, fewer neurons were suppressed in the old relative to young animals, which leads to higher prestimulus activity and more false alarms. We show that this excess activity additionally leads to increased correlations between neurons, reducing the amount of relevant stimulus information in the auditory cortex. Future work identifying the lost circuits that are responsible for proper background suppression could provide new targets for therapeutic strategies to preserve auditory processing ability into old age.
Introduction
The inability to comprehend speech in noisy backgrounds is a common challenge for those with age-related hearing loss. Factors that contribute to hearing difficulty in healthy aging can broadly be grouped into two main categories: peripheral degeneration of the auditory transduction mechanisms in the ear (Gopinath et al., 2009; Lin et al., 2011) and changes in the central auditory pathway that further process those peripheral signals (Caspary et al., 2008; Cisneros-Franco et al., 2018). While peripheral changes leading to hearing loss have been well categorized, less is known about how age-related central auditory changes affect auditory processing. Human studies have shown that aging subjects with an intact peripheral auditory system still perform worse on gap detection, sound source localization, and temporal perception (Fitzgibbons and Gordon-Salant, 1998; Gordon-Salant and Fitzgibbons, 1999; Lister et al., 2002; Wambacq et al., 2009).
To hear in natural environments, it is necessary to separate foreground and background sounds. This is most challenging in noisy environments, as both the foreground and background contain complex temporal and spectral features that must be separated (Nelken et al., 1999; Rodriguez et al., 2010). In healthy animals, central auditory areas, such as the primary auditory cortex (A1), encode target sounds in a complex background through a process called stimulus-specific adaptation (Ulanovsky et al., 2004; Rabinowitz et al., 2013; Natan et al., 2015). In this process, A1 neurons adjust the dynamic range of their responses to the dynamic range of the background stimulus to optimally encode deviations from the background stimulus by the foreground stimulus. It has been shown that aging mice passively listening to tones in noise are no longer able to fully shift their dynamic range, leading to increased activity correlations and stronger responses to the background stimuli (Shilling-Scrivo et al., 2021). It is believed that this reduction in dynamic range is because of the reduced inhibitory function in A1 of aging animals (Willott et al., 1993; Shi et al., 2004; Stanley and Shetty, 2004; Burianova et al., 2009; Stanley et al., 2012; Gold and Bajo, 2014; Liguz-Lecznar et al., 2015; Kumar et al., 2019).
Detection of relevant stimuli depends not only on the spectrotemporal relationships between foreground and background but also on the behavioral context of the sounds to the animal. During passive listening, the responses of A1 neurons are robust across a wide variety of stimulus parameters (Blackwell et al., 2016). Moreover, during behavior, the responses of A1 change to more optimally encode the behaviorally relevant stimulus (Fritz et al., 2003; Atiani et al., 2009; David et al., 2012; Kato et al., 2015; Francis et al., 2018; Kuchibhotla and Bathellier, 2018). However, it is unknown how aging affects the detection of behaviorally relevant stimuli in complex backgrounds.
To investigate how A1 detects tones in noise, and how A1 activity is disrupted in the aging brain, we recorded activity from 8078 neurons in young adult (2-6 months, N =10) and old (16-24 months, N =12) mice. We show that neurons in old animals have reduced behavioral gain, decreased offset responses, and increased activity correlations, especially in the prestimulus period compared with those in young animals. Old mice show an inability to decrease the activity correlations during task performance, and we show using a Bayesian decoder that these increased correlations lead to decreased decoding performance.
Materials and Methods
Animals
All procedures were approved by the University of Maryland Institutional Animal Care and Use Committee. Mice were housed in a 12 h reverse light/dark cycle room. We used 22 mice total with 10 young animals (6 male, 4 female, age 2-6 months, mean 4 months) and 12 old animals (8 male, 4 female, 16-24 months, mean 19 months). Thy1-GCaMP6s × CBA mice were used for 2-photon (2P) calcium imaging of cortical neurons. These mice were created by taking the offspring of a cross between Thy1-GCaMP6s mice (JAX: 024275) and CBA/CaJ mice (JAX: 000654), the latter of which are known for their exceptional hearing (Willott et al., 1988, 1991; Spongr et al., 1997; Bowen et al., 2020; Shilling-Scrivo et al., 2021).
Surgery
Mice were prophylactically injected with dexamethasone (5 mg/kg) 2 h before surgery to prevent infection and cortical edema. Mice were anesthetized with isoflurane (3%-4% for induction; 1.5%-2% for maintenance). Internal body temperature was maintained at 38°C using a heating pad with a closed loop homeothermic monitoring system. At the time of surgery, mice were injected again with dexamethasone and atropine (0.1 mg/kg). Hair was removed via plucking and hair removal agent (Nair). The scalp was then disinfected with three alternating swabs of betadine and 70% ethanol. The skin above the skull and temporal muscle was then removed, and the temporal muscle was resected to expose the temporal bone. The headpost was then attached to the skull with a combination of cyanoacrylate (Vetbond) and dental acrylic (C&B Metabond). A 3 mm circular section of bone above A1 was then removed, and the cranial window was implanted. The cranial window consisted of two 3 mm circular glass coverslips affixed to one 5 mm circular coverslip, the edges of which were filled with a clear silicone elastomer (Kwik-Sil). The window was then affixed in place with the same dental acrylic. The dental acrylic and headpost were then coated in iron oxide to prevent optical reflections. Mice were postoperatively given injections of meloxicam (0.5 mg/kg) and were allowed to recover for at least 1 week before experiments.
In vivo 2P imaging
Imaging was performed in animals as described previously (Liu et al., 2019; Shilling-Scrivo et al., 2021). Experiments were performed on a rotatable microscope (Bergamo II series, B248, Thorlabs) using a pulsed femtosecond Ti:Sapphire 2P laser (Vision S, Coherent) and ThorImage and ThorSync software. Imaging was performed at 940 nm excitation wavelength. The imaging field size was ∼370 µm × 370 µm and was imaged at 30 frames per second. Animals were first acclimatized to the microscope for a period of 5-10 min before experiments. During this time, white noise was also played to allow adaptation to the white noise stimulus. Each imaging session contained one block of sound stimuli presented with the animal in the passive condition followed by a block of sound stimuli in the active tone detection condition (Francis et al., 2018). Imaging during the passive set of stimuli and during tone detection task was done of the same FOV such that we obtained both passive and active response from each neuron.
Imaging data analysis
Neuron fluorescence traces were extracted using custom MATLAB code (The MathWorks, version 2017B). Images were motion-corrected to subpixel precision using discrete-time Fourier transforms. Cell bodies were then manually selected with cell bodies and neuropil divisions created automatically. Any pixels that overlapped multiple cells were excluded from analysis. For each neuron, all pixels within the cell body were averaged to create the baseline fluorescence value. To calculate the neuropil fluorescence, all pixels in a ring surrounding the labeled neuron were averaged, excluding pixels that corresponded to other neurons. The corrected neuropil fluorescence was calculated as follows: FCell_Corrected = FCell – 0.7 * FNeuropil. ΔF/F was calculated by dividing fluorescence from each trial by the average F of the preceding silent baseline frames. To test whether neurons were responsive to the tone stimulus, we performed an ANOVA of frames before and after tone stimulus presentation. Neurons were considered stimulus-responsive if their fluorescence after tone presentation was statistically different from baseline fluorescence.
Sound stimuli
All sound stimuli were presented with a free-field electrostatic speaker 10 cm away from the mouse's right ear (ES1 speaker with ED1 speaker driver, Tucker-Davis Technologies). The speaker was calibrated by first recording a 70 dB SPL, 4-64 kHz white noise with a calibrated microphone to find the speaker's natural transfer function. We then calculated the inverse of the function, which is the function that, when added to the speaker's natural transfer function, will equalize the output of the speaker, giving a flat frequency/dB curve. We then tested this calibration by recording pure tones at 70 dB SPL and ensured that the recorded sound level was <5 dB from the target for all tones played. During passive 2P imaging, 1 s sinusoidally amplitude-modulated tones (4-48 kHz ½ octave spacing, 5 Hz full depth modulation) at 70, 60, and 50 dB SPL were presented together with a constant (4-48 kHz, 50 dB SPL) broadband white noise to obtain signal-to-noise ratios (SNRs) of 20, 10, and 0 dB, respectively. Each trial consisted of a 1 s prestimulus silence, followed by 1 s of white noise, 1 s of tone presented in white noise, and then 1 s of white noise alone, and 1 s post-stimulus silence. Finally, to prevent timing strategies, we added a variable 6-10 s intertrial interval between trials as well as a rolling 2 s period in which the animal had to refrain from licking before the next trial would begin. Each unique frequency/SNR combination was presented for five repeats.
Behavioral setup
Mice were first habituated to the sound-attenuating boxes as described previously (Francis et al., 2018). Once habituated, we trained mice on the tone-in-noise detection task daily. Mice would be head-fixed and placed into the training chamber that contained a speaker playing constant 50 dB white noise. The animal was then allowed to habituate to the chamber for 5-10 min to allow the auditory system to compensate for the white noise background. After habituation, the experiment was initiated by providing free water until the mouse licked the waterspout, indicating it was both attending to and motivated by the water reward. Each trial would begin after the mouse refrained from licking for 2 s.
On correct trials, water was dispensed in precise volume (20 µl) via a solenoid that was placed outside the behavioral area to prevent any solenoid sounds from becoming a cue for water delivery.
Training
All animals were first trained for 1 week on a tone detection task in quiet. Then, 50 dB of white noise was added, and once again mice were trained on the task in noise. Then animals were given five sessions of training at two SNR levels (20, 10 dB SNR) before training a minimum of five sessions on the full task at three SNR levels (20, 10, 0 dB SNR). Mice were then trained for 1-2 weeks on the imaging scope to habituate animals to training. Mice trained in the imaging microscope until they were able to reach a d′ ≥ 1 criterion or they reached 1 month of total training time. Mice that were unable to reach this criterion were not used in this study. Mice were then imaged while behaving for 1-4 behavioral sessions.
Statistics
All significance testing was performed using built-in MATLAB functions unless otherwise noted. All one-way ANOVAs were calculated using the anova1 function. Higher-order ANOVAs (indicated by VARIABLE_1 × VARIABLE_2 notation in text) were calculated using the anovan function. Multiple comparisons were corrected for using the multcompare function on the results of the ANOVA. Multiple comparisons were performed using Tukey's Honest Significant Difference unless otherwise noted. Student's t tests between group means were performed using the ttest2 function.
Behavioral detection performance
To determine the behavioral performance of each animal, we used the commonly used signal detection method of inferring the behavioral d′ of each animal. To determine the signed d′, we calculated the ideal observer performance using the animal's hit and false alarm (FA) rates as follows:
Where z is the z score (norminv in MATLAB) and hits and FA are the hit rate and FA rate, respectively. Hit rates and FA rates were adjusted using the log-linear rule (Hautus, 1995) to prevent hit and FA rates near 0 and 1 to bias results.
Signal/noise correlations
Signal/noise correlations were calculated as in Winkowski and Kanold (2013). The signal correlation is defined as how similar (and thus correlated) the tuning curves are for two neurons. We obtain the tuning curve for a single neuron by calculating the average response to each tone/level combination across trials. Tuning curves were calculated for each neuron to create a N × M matrix where N corresponds to the number of neurons and M is the unique combinations of tone/level combinations. To obtain the correlations between two neurons (i and j), we obtained the correlation between the tuning curves we used the Pearson correlation equation below.
Neurons with direct connections to each other or that share common inputs will show correlated trial-to-trial activity. Noise correlations are defined as any trial-to-trial correlated activity that is not explained by signal correlations. To obtain noise correlations, we first found the average fluorescence response during each trial. To remove the effect of signal correlations, we subtracted the cell's average response to that stimulus, leaving only the trial-to-trial variance. Calculating this for each neuron, we created an N × M matrix where N is the number of neurons and M is the neuron's deviation from its average response on each trial. From this noise matrix, we calculated the trial-to-trial correlation between each pair of neurons with the same Pearson correlation equation above. We analyzed significance by creating a two-way age × SNR ANOVA with corrections for multiple comparisons. To look at the effect of attention, for each neuron we calculated the following:
We then further separated this into Δr+ and Δr–, based on whether the change in correlations from active to passive was positive or negative, respectively (Francis et al., 2018). To address the effect of tuning, Δr+ and Δr– were further subdivided into BFin and BFout based on whether the best frequency (BF) of the neuron was within a half-octave of the target frequency. Significance was analyzed using a two-way BF × SNR ANOVA for Δr+ and Δr–, respectively.
Naive Bayes modeling
The Naive Bayes decoder uses Bayes rule to determine which tone was played given the neural activity of a subset of neurons. Data were in the form of a time × trial × neuron matrix. For each time point, a random subset of neurons was selected for the model and used to create a naive Bayes model using the MATLAB function, fitcnb. CIs were determined by running 10 such models and plotting the mean and 95% CIs. See below for a detailed description of the mathematics behind naive Bayes.
The probability that the animal behaved correctly given a certain neural population activity can be written as follow:
The model takes the neural activity of each trial, and calculates the probability of whether the animal performed correctly based only on the neural activity of that trial:
To ensure the validity of the model, it is run with fivefold cross-validation. The model is rerun 20 times using different random subsets of training data to ensure accurate model performance.
Temporal analysis
We began with a time × trial × neuron matrix, where each index represents the calcium activity of a neuron on that time point and trial. We first obtained the average response for each neuron by averaging each neuron's response across trials to create a time × neuron matrix. To show how the response of the neuron changed over time, we then took the approximate derivative of each neuron's response with the MATLAB diff function. Then, for each neuron, we found the time point for which the maximum change in response occurred. We binned these response times for greater statistical power and used a one-way ANOVA with multiple comparisons to determine time points that had a significant number of cells compared with silence (the first two time bins).
Results
To investigate how age affects the processing of behaviorally relevant sounds in noisy backgrounds, we recorded the activity of 8078 neurons via 2P imaging of young adult (2-6 months) and old (16-24 months) Thy1-GCaMP6s × CBA mice performing a tone detection task in noise. Mice were head-fixed in the behavioral box (Fig. 1A) while a speaker played continuous broadband white noise (4-48 kHz 50 dB SPL). Mice were allowed to acclimate to the white noise stimulus for 5-10 min before the start of any behavioral experiments. Once acclimated, each imaging session consisted of a passive block to map the best frequencies of the imaged neurons, followed by an active tone detection block. In the passive block, animals passively listened to a series of 1 s tones at multiple frequencies (4-48 kHz, 70 dB SPL). Following the passive block, the lick spout would be raised and a small amount of water would be provided to the animal to signal the beginning of the active detection block. During active detection, each trial began with a 1 s prestimulus period, followed by a 1 s target tone with 5 Hz amplitude modulation at one of three levels (70, 60, 50 dB SPL), giving an SNR of 20, 10, and 0 dB. Intertrial intervals were randomized between 6 and 10 s. Additionally, to prevent habitual responding, the subsequent trial only began after the animal refrained from responding for 2 consecutive seconds after the ITI. Animals were rewarded with a 0.3 s water reward if the first lick of a trial occurred 0.1-1.5 s after tone onset (Fig. 1A, green bar). If the mouse licked before this window (Fig. 1A, FAs, purple bar), no water was delivered for subsequent licks; and additional 8 s timeout punishment was added to the ITI. All animals were first trained in ideal conditions with no background noise to ensure that they understood the task parameters. Animals were then trained on the full task with background noise added for a minimum of 2 weeks and were imaged once they reached stable behavioral performance. To characterize the performance of animals, we computed the detection sensitivity (d′) over all SNR levels. We find that, while there is variability between animals, overall d′ is high and decreases as a function of age (Fig. 1B, age × level ANOVA main effect of age, F(1) = 6.09, p = 0.016).
Old mice have worse performance during tone-in-noise detection task. A, Mouse A1 was imaged using 2P microscopy while either listening to tones-in-noise passively or detecting a target 12 kHz tone at one of three different SNRs. Top, Task structure used for the active experiment. Bottom, Example 2P imaging field. B, Detection sensitivity of animals to detect the target tone decreases as a function of age (ANOVA main effect of age, F(1) = 6.09, p = 0.016). C, Young and old animals in the study performed significantly better than those that did not learn the task. Left, Young animal stats (age × level ANOVA main effect of performance, F(1) = 72, p = 2.8 e-12, post hoc tests: 0 dB, p = 1.1e-3; 10 dB, 9.53e-5; 20 dB, p = 5.96e-6). Right, Old animal stats (performance × level ANOVA main effect of performance, F(1) = 29, p = 2.4e-6, post hoc tests: 0 dB, p = 0.22; 10 dB, p = 2.4e-3; 20 dB, p = 0.03). D, Example neurons responses in passive and determination of its BF. E, The average percentage of neurons that had BFs at each presented frequency. At the active target frequency (12 kHz), there is a <1% difference in BF representation between young and old animals. F, Lick-triggered average of all behavioral responses that occurred before tone onset.
To show that animals understood the task, we separated young (2-6 months) and old (16-24 months) animals into good and poor-performing groups. Animals that had at least one level in which they performed at a d′ ≥ 1 were labeled as good performers and those that did not reach that level of performance after 1 month of training were labeled poor performers. We then plotted their behavioral performance as a function of SNR. Young good performing animals showed significantly better performance than poor animals across all levels (Fig. 1C, left, performance × level ANOVA main effect of performance, F(1) = 72, p = 2.8 e-12, post hoc tests: 0 dB, p = 1.1e-3; 10 dB, 9.53e-5; 20 dB, p = 5.96e-6). Old good animals had higher variance and only performed significantly better than old poor animals at the 10 and 20 dB conditions (Fig. 1C, right, performance × level ANOVA main effect of performance, F(1) = 29, p = 2.4e-6, post hoc tests: 0 dB, p = 0.22; 10 dB, p = 2.4e-3; 20 dB, p = 0.03). These results indicate that old animals learned the task and performed better than the poorly behaving animals at all but the lowest sound level. All further analysis was performed on the subset of animals that learned the task, and we therefore will use old and young to refer to the old and young animals able to learn the task.
We next investigated the stimulus-evoked activity of neurons in A1 with in vivo 2P imaging (Fig. 1D). Previous work (Francis et al., 2018) has shown that three factors influence the neural response of the auditory cortex during a behavioral task: (1) the responses to the auditory stimulus, (2) attentional gain (if the animal was engaged in a task), and (3) behavioral outcome (hits vs miss and FAs). We investigated separately how aging affects these three factors to identify which changes are responsible for the decrease in performance.
Old and young mice have similar responses at low frequencies
Old mice on the CBA/CaJ strain have reduced hearing loss with age and have been shown to have good hearing during the ages tested (Willott et al., 1988, 1991; Spongr et al., 1997; Bowen et al., 2020; Shilling-Scrivo et al., 2021). To confirm that changes in behavioral performance in the mice studied were not because of peripheral hearing loss, we measured the fraction of cells responding to the 12 kHz target tone in both young and old mice. We presented tones (4-48 kHz, 70 dB SPL) in the passive condition (with the waterspout removed) and obtained fluorescence traces for each neuron (Fig. 1D). We then calculated the average neuronal response for each tone and defined the BF as the frequency that elicited the largest average response (Fig. 1D, arrow). Neurons that showed a significant change between baseline frames and stimulus frames were designated as stimulus-responsive. When plotting the overall BF distributions (Fig. 1E), we found that, while there is a slight reduction in the representation of the highest frequencies in old animals, there is <1% difference between young and old neurons at the target frequency (Fig. 1E, % neurons with 12 kHz BF: young, 13.1%; old, 12.5%). Therefore, differences in detection performance are not because of the reduction of frequency representation at the target frequency used in this experiment. We also verified that the neural activity seen was not because of the licking activity that occurred during the trial. To do so, we calculated the lick-triggered average of all licks that occurred before the trial began. As this licking is separate from the behavioral task, we can isolate the effect of licking on A1 neuron activity. The lick-triggered average (Fig. 1F) showed no activity in either group, indicating that licking did not cause fluorescence increases in A1, consistent with prior reports (Francis et al., 2018).
Old animals do not respond to tone offset
We first investigated the dynamics of the behavioral responses to determine how quickly animals responded to the tone stimulus and whether there are any different strategies between the groups of animals. Moreover, our paradigm requires animals to withhold responding for 100 ms after tone onset (Fig. 2A, green horizontal bar). Young animals responded soon after tone onset, aligning the majority of behavioral responses to the earliest possible reward period (100 ms after tone onset) (Fig. 2B). In addition to licking at tone onset, we observed licking at tone offset in young animals (Fig. 2B, arrow). Old animals also showed increased licking after the end of the waiting period but showed fewer offset responses than young animals (Fig. 2B). In addition, old animals showed a higher level of licking in the prestimulus period (FAs). This indicates that both young and old animals with good performance understood the task parameters. Moreover, young animals with good behavior were the only group that frequently responded after stimulus offset (Fig. 2B).
Temporal dynamics of behavioral response match temporal dynamics of the neural response in A1. A, Diagram of task parameters. B, Histograms of behavioral response timing for young and old animals, respectively. Horizontal blue bar and associated dashed lines represent the timing of the tone onset and offset. Young animals show a licking response at tone onset and offset (arrow). C, To calculate the Δ response of each neuron, the derivative of the neurons' average activity is taken and the time point with the largest change in response magnitude. Resulting peak time points are grouped across animals, such that each circle represents the percentage of neurons that had the peak Δ response at that time bin. D, Left, A1 neurons in young animals have peak responses at tone onset and offset transitions (ANOVA F(8) = 16.7, post hoc tests: onset, p = 8.9e-8; offset, p = 1.67 e-7). Right, Old animals only had significant responses to tone onset (Fig. 2D, right, ANOVA F(8) = 3.78, post hoc tests: onset, p = 1.81e-2; offset, p = 0.657).
Old animals lack offset responses
We next explored whether these differences in response timing between age groups were mirrored by altered temporal response dynamics of A1 neurons. Neurons in A1 generally have the strongest response to large transitions in sound amplitude, such as the onset or offset of a sound (Liu et al., 2019). Since old animals show reduced tone offset responses during passive sound presentation (Shilling-Scrivo et al., 2021), we speculated that a lack of offset responses in the old animals could contribute to the lack of behavioral responses after tone offset. We therefore investigated the time course of the neural responses during the behavioral episodes.
To determine the temporal response characteristics of each neuron, we first calculated the average responses across all trials. We then identified the time point with the greatest Δ response by computing the derivative of those average responses. For each animal, we then plotted a histogram of the resulting time points at 300 ms time bins (Fig. 2C). We then defined significant population response bins as time bins that have a significantly higher percentage of peak responsive cells than in the prestimulus silence. All statistical measures are in comparison to the first time bin, but general significance patterns still apply if the second time bin is used for comparison. We defined onset and offset responses as the first bin that showed response after stimulus onset/offset. As expected (Liu et al., 2019), A1 neurons in young animals show a large fraction of neurons with peak responses at tone onset and offset and are thus particularly tuned to sound transitions (Fig. 2D, left ANOVA, F(8) = 16.7, post hoc tests: onset, p = 8.9e-8; offset, p = 1.67 e-7). In contrast, in old animals, most neurons responded to tone onset and not offset (Fig. 2D, right, ANOVA, F(8) = 3.78, post hoc tests: onset, p = 1.81e-2; offset, p = 0.657). These results show that old animals lack neuronal offset responses, and this lack of offset responses might underlie the failure of old animals to respond to classes of sounds containing sharp declines in amplitude.
Suppressive neurons in old A1 have less attentional gain
Neurons in A1 show task-dependent changes of neuronal activity termed attentional gain (Fritz et al., 2003; Atiani et al., 2009; David et al., 2012; Kato et al., 2015; Francis et al., 2018; Kuchibhotla and Bathellier, 2018). We next investigated whether aging altered these task-dependent changes. Figure 3A shows the average responses of each stimulus-responsive neuron in young (3427 neurons) and old animals (900 neurons) while animals were either actively correctly detecting the target tone (hits, left columns) or passively listening to the target tone (middle column). We first separated neurons into facilitative or suppressive based on whether they had an increase or decrease in fluorescence relative to baseline (Student's t, p < 0.001). We noted that, for both young and old animals, ∼80% of stimulus-responsive neurons had facilitative responses (young = 2626/3427; old = 690/900) and 20% had suppressive responses (young = 601/3427; old = 210/900). We then calculated the attentional gain observed for each neuron by subtracting the passive from the active response (Fig. 3A, right columns).
Old animals had reduced attentional gain for suppressive responses. A, For each neuron, the attentional gain was calculated by subtracting the average response of the neuron during tone detection from the average response of the neuron during passive listening. Each line indicates the activity of one neuron. B, Gain responses were separated into suppressive gain and facilitative gain, based on whether the sign of attentional gain during the tone was positive or negative. C, Stats for facilitative (age × level ANOVA, main effect of age: F(1) = 0.089, p = 0.76, main effect of level: F(2) = 2.99, p = 0.051). D, Stats for suppressive (age × level ANOVA, main effect of age: F(1) = 90.1, p = 5.53e-21, main effect of level: F(2) = 1.09, p = 0.335). E–H, Analysis by sex. Males showed similar patterns for facilitative (age × level ANOVA, main effect of age: F(1) = 0.55, p = 0.45, main effect of level: F(2) = 6.85, p = 1.1e-3) (E) and suppressive neurons (F) (age × level ANOVA, main effect of age: F(1) = 81.4, p = 4.64 e-19, main effect of level: F(2) = 0.366 p = 0.693). Females showed similar patterns for facilitative neurons (G) (age × level ANOVA, main effect of age: F(1) = 2.26, p = 0.13, main effect of level: F(2) = 21.1 p = 7.79e-10). However, both young and old females had similar levels of suppressive gain (age × level ANOVA, main effect of age: F(1) = 81.4, p = 4.64 e-19, main effect of level: F(2) = 0.366 p = 0.693).
We then plotted the average population gain for young and old animals separately for facilitated and suppressed cells (Fig. 3B). Facilitated cells from young and old animals showed a positive gain. The magnitude of the gain for facilitative responses was similar for cells from young and old animals (age × level ANOVA, main effect of age, F(1) = 0.089, p = 0.765). However, in cells from old animals, we observed a positive gain before stimulus onset (arrow), indicating an increase in activity during the prestimulus period in old animals. The increase in prestimulus activity in old animals is consistent with the increase in licking in the prestimulus period in old mice (Fig. 2). Neurons showing suppressive responses in young and old animals showed attentional gain, in that suppressive responses had an increased magnitude (more negative) (Fig. 3B). Cells from old animals had smaller gain of suppressive responses (age × level ANOVA, main effect of age, F(1) = 90.1.4, p = 5.5 e-21). Together, these data show that suppressive neurons in old animals have reduced attentional gain. We presented the target tone at various SNR levels, which reflects varying task difficulty. In old animals, this resulted in differences in task performance. We therefore next investigated whether gain varies with SNR and whether gain deficits with age might underlie the decreased performance of old animals at low SNR. In both facilitative and suppressive responses, we found no significant effect of SNR (Fig. 3C,D, age × SNR ANOVA main effect of SNR, facilitative, F(2) = 2.99, p = 0.051; suppressive, F(2) = 1.09, p = 0.33).
In previous work, we found that old male animals showed larger deficits in suppression when passively listening to tones in noise (Shilling-Scrivo et al., 2021). We therefore would predict that, when comparing by sex, we would similarly see a deficit in suppressive gain. While both male and female mice had deficits in suppressive gain, the deficit was most pronounced in male mice (Fig. 3E–H). Thus, we show that old animals have deficits in suppressive gain during behavior.
Old mice fail to suppress activity at low SNR
Suppression of task-irrelevant neurons is critical for the detection of stimuli in noise. In the healthy auditory cortex, neural responses to irrelevant stimuli are reduced, allowing the optimal encoding of attended stimuli in a noise-invariant manner (Rabinowitz et al., 2011, 2013; Natan et al., 2015; Cooke et al., 2018). With the reduction in suppressive gain in old animals, we would predict that the responses to the background noise may not be fully suppressed, leading to more FAs because of this spurious neural activity. To test this prediction, we looked at the average neural responses of animals when they made a response to identify differences in prestimulus activity depending on the response outcome.
In young animals at both SNR levels (Fig. 4A,B, left), all behavioral choices were associated with an increase in fluorescence in facilitative cells and a decrease in fluorescence in suppressive cells (Fig. 4C,D, left). We also observed that there was more prestimulus activity for both suppressed and facilitated cells for early FA trials than for hit trials. These data show that young animals' behavioral choice is associated with distinct changes in the activity of A1 neurons.
Old animals have increased prestimulus activity and reduced behavioral gain for suppressed responses. Plots represent the average neural response when the animal either made a hit or FA. Neurons were separated into facilitated (A,B) or suppressed (C,D) by their average activity after tone onset. A, B, For facilitated neurons, both groups showed increased prestimulus activity during FA across both SNR conditions. C, D, Old animals show reduced disruptions in suppressive activity. C, At high SNR, old animals show reduced suppression compared with young (arrow). D, At low SNR, normally suppressed neurons become active during FA (arrow).
In old animals (Fig. 4A, right), the choice-related activity in facilitated cells at 20 dB SNR, where old animals showed good performance, largely mirrored the changes in young animals. However, the magnitude of the changes in suppressive cells was markedly reduced (Fig. 4C, right). Together, these data show that, in old animals under easy listening conditions, the neural choice-related signals in A1 largely mirror young animals. In contrast, at 0 dB SNR, where old animals do not perform well, distinct differences in the choice-related activity were evident in facilitated and suppressive cells (Fig. 4B,D, right). Old animals showed reduced suppression and increased facilitation during hit trials. During FA trials, suppressive cells show prestimulus excitation, suggesting a complete breakdown in the suppression of these cells in difficult listening conditions. Separating these results by sex (Fig. 5), we found that this breakdown in suppression is predominately found in the male mice (Fig. 5D). These data reveal that reduced suppression leads to increased prestimulus activity in normally suppressed neurons during FA trials at low SNR.
Increased prestimulus activity and reduced behavioral gain for suppressed responses are greater in old male animals. Plots are similar to Figure 4 but now represent changes between aging males and females. For facilitated neurons, both groups showed increased prestimulus activity during FA during high SNR conditions (A). At low SNR conditions (B), old males showed little change in activity for either hits or FAs. C, D, Old females additionally have more suppressive responses than old males. C, At high SNR, both male and female old animals show reduced suppression. D, However, at low SNR, only the aging male activity becomes excitatory during low SNR.
Increased activity correlations between neurons in old mice
So far, we have shown that multiple measures of single-neuron activity are impaired in old animals. However, behaviorally relevant stimuli are better encoded by distributed networks of neurons in A1 (Francis et al., 2018). Thus, we next investigated how aging affects population activity in A1. Given that in passive listening old mice correlated neuronal activity is increased (Shilling-Scrivo et al., 2021), we speculated that altered population activity might be related to the behavioral deficits.
Noise correlations are the measure of remaining trial-to-trial covariance between neurons after tuning similarity is accounted for; thus, these are correlations independent of the stimulus. These correlations are important, as even weak noise correlations can have a large impact on how well neural populations can encode stimuli (Abbott and Dayan, 1999; Nirenberg and Latham, 2003) with reduced noise correlations leading to better performance (Cohen and Maunsell, 2009; Mitchell et al., 2009). We have previously shown that old animals have higher noise correlations during passive listening (Shilling-Scrivo et al., 2021). If noise correlations are similarly higher during behavior, it could impair detection efficiency in older animals.
We calculated noise correlations during tone presentation in both young and old mice during the passive and active conditions. Thus, the noise correlations reflect activity correlations independent of the tonal stimulus. We find that, for both groups, on average, noise correlations are small and positive in both the 20 and 0 dB SNR conditions (Fig. 6A), similar to prior results (Francis et al., 2018). Moreover, correlations in old animals in the active condition were significantly larger than those of young animals (Fig. 6A: age × level ANOVA: post hoc tests: 20 dB, p = 4.62e-9; 0 dB, p = 3.76e-9). Therefore, old animals showed increased noise correlations during behavior.
Noise correlation increased in old animals across frequency and distance. A, CDF of noise correlation during tone detection. Inset, Average correlation values showing significantly larger correlation values for aging animals at all levels (age × level ANOVA: post hoc tests: 20 dB, p = 4.62e-9; 0 dB, p = 3.76e-9). B, To isolate the magnitude of the attentional gain, we created Δr– and Δr+ populations by taking the difference between active and passive correlation values for each neuron. C, Correlations were separated into those that decreased correlation during behavior and those that increased, which are labeled r– and r+, respectively. Average correlation values were then plotted for each group in the passive and active conditions. Comparing the r– populations in the 20 dB condition between young and old, the old r– neurons have increased correlations in both conditions (age × behavior ANOVA post hoc tests: passive, p = 0.51, active, 3.7e-9). The r+, however, only have significantly increased correlations in the active condition (age × behavior ANOVA, post hoc tests: passive, p = 0.51; active, 3.7e-9). D, Left, Δr+ neurons have increased attentional gain in aging (age × level ANOVA, main effect of age, F(1) = 61.7, p = 5.11e-15, post hoc tests: 20 dB, p = 3.76e-9; 0 dB, p = 1.8-e3). Right, Δr– neurons show no differences in attentional gain (age × level ANOVA, main effect of age, F(1) = 1.78, p = 0.18). E, The results of looking at Δr+ and Δr– by sex. For Δr+ (left), males show the same patterns as the total group (age × level ANOVA, main effect of age, F(1) = 34.5, p = 4.76e-9, post hoc tests: 20 dB, p = 1.61e-8; 0 dB, p = 0.03). Females also show similar patterns (age × level ANOVA, main effect of age, F(1) = 7.3, p = 6.3e-3, post hoc tests: 20 dB, p = 0.89; 0 dB, p = 0.008). For Δr– (right), males and females have opposing results. Old males have an increased reduction in correlations compared with young males (age × level ANOVA, main effect of age, F(1) = 227, p = 9.95e-50, post hoc tests: 20 dB, p = 3.7e-9; 0 dB, p = 3.76e-9). Old females, however, have significantly less reduction in correlation in r– neurons than young females (age × level ANOVA, main effect of age, F(1) = 51.1, p = 2.28e-14, post hoc tests: 20 dB, p = 1.27e-4; 0 dB, p = 3.95e-9).
Since attention alters single neuron responses (Fig. 3), we next wanted to investigate the effect of attention on noise correlations. Attending to a stimulus alters the magnitude of noise correlations in behaving animals, compared with the passive condition (Francis et al., 2018). However, this effect is not uniform: some neurons increase their correlations (Δr+), while others decrease their correlations (Δr–). If this differential attentional effect is impaired in aging, it would also lead to poor performance. To investigate the change in correlations, we first separated the populations of neurons into two separate populations: neurons whose noise correlations decreased during behavior (Δr–) and those whose noise correlations increased during behavior (Δr+) (Francis et al., 2018) (Fig. 5B). Comparing these two populations across age revealed an asymmetry in the effect of behavior on noise correlations. In Δr+ neurons (Fig. 6C), noise correlations were similar between young and old animals in the passive condition, while old Δr+ neurons had significantly higher noise correlations during behavior than young Δr+ neurons (age × behavior ANOVA post hoc tests: passive, p = 0.51; active, 3.7e-9). However, in Δr– neurons, old animals had higher noise correlations than young in both the active and passive conditions (age × behavior ANOVA post hoc tests: passive, p = 3.76e-9; active, 3.76e-9). Thus, Δr+ and Δr– neurons in young and old animals seemed to show different behaviorally related changes in correlations (Fig. 6C). Comparing Δr across ages showed that the magnitude of noise correlations in old animals was larger for Δr+ neurons (Fig. 6D, left, age × level ANOVA main effect of age, F(1) = 61.7, p = 5.11e-15, post hoc tests: 20 dB, p = 3.76e-9; 0 dB, p = 1.8-e3) but not Δr– neurons (Fig. 6D, right, age × level ANOVA main effect of age, F(1) = 1.78, p = 0.18). Therefore, the increased noise correlations in old animals compared with young animals during behavior result from both the increased baseline correlation of Δr– neurons and the increased correlation of Δr+ neurons during behavior. Thus, our results point to an age-dependent decrease in the suppression of prestimulus correlation and a larger attentional increase in correlations in specific groups of cells.
We next separated these results by sex. The results for the Δr+ followed the same pattern, as old male mice had increased correlation at all levels (Fig. 6E, left, age × level ANOVA, main effect of age, F(1) = 34.5, p =4.76e-9, post hoc tests: 20 dB, p = 1.61e-8; 0 dB, p = 0.03), while old female mice only had significant increased correlations at lower SNRs (age × level ANOVA, main effect of age, F(1) = 7.3, p = 6.3e-3, post hoc tests: 20 dB, p = 0.89; 0 dB, p = 0.008). For Δr– neurons (Fig. 6E, right), males and females had opposing results. Old males had an increased reduction in correlations compared with young males (age × level ANOVA, main effect of age, F(1) = 227, p = 9.95e-50, post hoc tests: 20 dB, p = 3.7e-9; 0 dB, p = 3.76e-9). Old females, however, had less reduction in correlation in Δr– neurons than young females (age × level ANOVA, main effect of age, F(1) = 51.1, p = 2.28e-14, post hoc tests: 20 dB, p = 1.27e-4; 0 dB, p = 3.95e-9).
Given that correlated neurons have certain stimulus selectivity, we next wanted to assess how the tuning of neurons relative to the target may affect noise correlations. We separated neurons by whether or not the BF of the neurons was within half an octave of the target frequency and labeled them BFin and BFout, respectively (Fig. 7A). For Δr+ neurons, the frequency distance to the target frequency did not affect the patterns of correlations (Fig. 7B). The same is true for Δr– neurons (Fig. 7C). Therefore, the relative tuning of individual neurons does not impact the network correlation structure.
Correlations higher across long distances and behavioral outcomes in old animals. A-C, Relationship between noise correlations and distance between the neurons' BF and the target. A, Neurons were separated into BF-in and BF-out based on whether their BF was within a half octave of the target frequency or not. BF does not affect the correlation patterns. B, Δr+ neurons show no effect of BF on young (BF × level ANOVA: post hoc tests: 20 dB SNR, p = 1; 0 dB SNR, p = 0.79) or old neurons (BF × level ANOVA: post hoc tests: 20 dB SNR, p = 0.99; 0 dB SNR, p = 0.93). C, Δr– neurons, likewise, show no difference across BF for young (BF × level ANOVA: post hoc tests: 20 dB SNR, p = 0.98; 0 dB SNR, p = 0.93) or old (BF × level ANOVA: post hoc tests: 20 dB SNR, p = 0.99; 0 dB SNR, p = 0.062). D–F, Effect of behavioral outcome on SI correlations. D, Trials where the animal responded were separated by behavioral outcome into hits and FAs. E, In young animals, correlation values were significantly lower in hit trials than in FA for Δr+ (top, performance × level ANOVA, post hoc tests: 20 dB, p = 4.0e-4; 0 dB, p = 0.043) and Δr– correlations (bottom, performance × level ANOVA, post hoc tests: 20 dB, p = 3.58e-9; 0 dB, p = 2.7e-3). F, In old animals, correlations were high in both hit and FA trials. G, H, Noise correlation value of a function of distance. G, Correlation values were binned at 10 µm intervals from 0 to 300 µm. H, In young animals, correlation values decrease as a function of distance; and correlations are, on average, lower during active engagement (hits) than passive. I, In old animals, both passive and active correlations are elevated relative to young animals.
Next, we investigated whether there was any interaction between behavioral outcome and correlation magnitude. As reduced noise correlations can lead to better performance (Cohen and Maunsell, 2009; Mitchell et al., 2009), we would predict that noise correlations may be higher in trials where the animal performed incorrectly. To test this, we calculated the noise correlations for hit trials and FA trials. As predicted, young animals had increased correlations during FA trials compared with hit trials for both Δr+ correlations (Fig. 7E, top, performance × level ANOVA, post hoc tests: 20 dB, p = 4.0e-4; 0 dB, p = 0.043) and Δr– correlations (Fig. 7E, bottom, performance × level ANOVA, post hoc tests: 20 dB, p = 3.58e-9; 0 dB, p = 2.7e-3). Old animals, however, showed high noise correlations regardless of behavioral outcome (Fig. 7F). Therefore, the modulation of activity correlations in A1 by behavioral outcome is impaired in old animals.
Finally, we investigated how changes in correlation changed as a function of distance from the pair of neurons (Fig. 7G–I). Pairwise correlation are highest between nearby cells and show spatial dependence (Winkowski and Kanold, 2013; Bowen et al., 2020; Rupasinghe et al., 2021), likely because of the spatial pattern of intracortical connections (Watkins et al., 2014; Meng et al., 2017). We thus investigated the effect of spatial separation on correlations in the active and passive conditions (Fig. 7G–I). We found that, in young animals during behavior, correlations between distant neurons were reduced compared with the passive in both the 20 and 0 dB conditions (Fig. 7H). Correlations for neighboring cells (<20 µm) were unchanged. Similarly in old animals at 20 dB SNR, where animals performed well, correlations were reduced for pairs from ∼20 to 300 µm apart (Fig. 7I). In contrast, at 0 dB, where animals performed poorly, only correlations for pairs located >200 µm apart were reduced. Correlations for close-by cells (<20 µm) were unchanged under both SNR conditions. These results indicate that aging reduces the ability to control mesoscale (<200 µm) correlated activity and that this reduced ability to control correlations is associated with behavioral performance.
Together, these data indicate that, in old animals, local networks of neurons are more correlated to each other; these increased correlations exist regardless of cell tuning and behavioral outcome; and these networks remain correlated over a much larger portion of cortical space.
Old A1 activity is less predictive of correct responses and more predictive of FAs
We have so far shown that neural responses in aging animals are degraded. Responses contain less temporal information, there is increased activity of normally suppressed neurons, and neurons are more correlated with each other. Data from human studies have shown that older adults show higher cognitive effort to detect tones in noise, regardless of overall accuracy (Pichora-Fuller et al., 1995; Wingfield and Grossman, 2006). Therefore, we speculated that old animals may have poorer stimulus encoding in A1, which leads to worse performance at all but the easiest listening conditions (20 dB SNR). To investigate this, we trained a Naive Bayes decoder model to determine how well populations of A1 neurons encode behavioral performance (Maron, 1961; Shilling-Scrivo et al., 2021) (Fig. 8A).
Neural decoding is impaired in older animals. A, Ideal observer analysis. The diagram represents a Naive Bayes decoder trained on two neurons to discriminate the behavioral outcome of the trial. Based on the labeled training data (colored circles) of how the two neurons respond for each trial, the encoder can create boundaries (dotted lines) to predict future trials. We trained the Naive Bayes classifier to decode tone frequency from neural activity. B, Classifier accuracy improves as more neurons are added to the decoding model for both groups. In the young group, asymptotic performance was reached at ∼10% accuracy at 60 neurons. The old group showed no improvement with an increasing number of neurons. C, Left, Using a model trained to detect whether the animal performed correctly using the activity of n = 60 neurons. Before tone onset, all models show no increase in performance, indicating that the models are well calibrated. Once tone onset occurs, all models increase in accuracy above chance, and performance increases until tone offset. Right, Bar plots represent the effect of SNR on the detection accuracy of the models. Models trained on old neurons were less effective at decoding at all SNR (age × SNR ANOVA, post hoc tests: 0 dB, p = 1.47e-6; 10 dB, p = 1.21e-4; 20 dB, p = 1). D, A second model was trained to detect whether the animal made an early response in the prestimulus period (FAs). Left, The model trained on young animals shows no ability to detect FAs, while the model trained on old increases in predictive accuracy close to tone onset. Right, The model trained on old animals shows significantly better predictive power across all SNRs (age × SNR ANOVA, post hoc tests: 0 dB, p = 1.81e-3; 10 dB, p = 1.43e-3; 20 dB, p = 2.41e-3). E–H, Analysis by sex. For decoding hits, males (E) showed similar patterns with old males showing the worst detection accuracy at the lowest SNR (age × SNR ANOVA, post hoc tests: 0 dB, p = 2.e-4; 10 dB, p = 0.051; 20 dB, p = 0.09). F, Females, however, showed the opposite result, as young females only had better decoding performance at the highest SNR (age × SNR ANOVA, post hoc tests: 0 dB, p = 0.23; 10 dB, p = 0.97; 20 dB, p = 3.0e-3). For decoding FAs, old males (G) showed increased predictive power across SNR (age × SNR ANOVA, post hoc tests: 0 dB, p = 9.280e-6; 10 dB, p = 6.58e-6.97; 20 dB, p = 3.12e-7). H, Old females, however, only had higher FA prediction at the highest SNR (age × SNR ANOVA, post hoc tests: 0 dB, p = 0.17; 10 dB, p = 0.278; 20 dB, p = 5.32e-7).
We first varied the numbers of neurons included to identify whether additional neurons increased classification accuracy. Figure 8B shows the relative increase in decoding accuracy from the single neuron model as the number of neurons used in the model increases. Already, an increase in neuron number to five significantly improved performance (Fig. 8B) consistent with the idea that small networks of neurons can represent stimuli (Francis et al., 2018). When the model was trained on young animals with good performance, model accuracy increased steadily until plateauing at ∼60 neurons (Fig. 8B). In contrast, increasing the number of neurons did not affect model performance for old animals. This is consistent with our observation that old animals have increased correlations (Fig. 6A), given that adding highly correlated neurons to the model will add little new information about the stimulus.
We next investigated the temporal encoding of these populations by plotting the accuracy of n = 60 neuron models over time (Fig. 8C, left). Models trained on young animals showed ∼10% increase in accuracy after tone onset, in line with our previous results in decoding behavioral performance with no background noise (Francis et al., 2018). Old animals, however, showed a much smaller increase in accuracy with time. To quantify these changes, we looked at the maximum accuracy each model obtained during the tone period and plotted it as a function of SNR (Fig. 8C, middle). The model trained on young animals performs significantly better than the model trained on old animals at all but the highest SNR (Fig. 8C, middle age × SNR ANOVA, post hoc tests: 0 dB, p = 1.47e-6; 10 dB, p = 1.21e-4; 20 dB, p = 1). Additionally, the model elucidated a sex dependence in stimuli encoding in young, but not old, animals (Fig. 8C, right). These results indicate that neuronal activity in old animals contained less information about behavioral performance, especially at challenging listening conditions.
We have earlier hypothesized that the increased prestimulus activity causes the old animals to have more FAs. If this were the case, we should be able to predict whether the animal made an FA based only on the prestimulus activity. We thus tested this hypothesis using the same naive Bayes encoder model, trained to detect FAs using only data from the prestimulus period. Plotting the normalized accuracy over time (Fig. 8D, left) showed that the model trained on old animals increased in accuracy over time, while the model trained on young animals showed no change. Testing the effect of SNR revealed that the model trained on old animals has significantly higher FA prediction at all SNR (Fig. 8D, right, age × SNR ANOVA, post hoc tests: 0 dB, p = 1.81e-3; 10 dB, p = 1.43e-3; 20 dB, p = 2.41e-3).
We had earlier shown that there is a greater loss of suppression in male mice. As such, we would expect that, if this loss of suppression is causing the increased prestimulus activity, and given increased FAs, our model would be more predictive for male mice than for female mice. Indeed, we found that, for old male mice, there was increased prediction of FAs at all sound levels (Fig. 8G, age × SNR ANOVA, post hoc tests: 0 dB, p = 9.280e-6; 10 dB, p = 6.58e-6.97; 20 dB, p = 3.12e-7). However, for old female mice, there was only significantly higher decoding performance at the highest SNR (Fig. 8H, age × SNR ANOVA, post hoc tests: 0 dB, p = 0.17; 10 dB, p = 0.278; 20 dB, p = 5.32e-7). Together, these data show that the amount of prestimulus activity is predictive of FAs for old animals.
We speculated that increased activity correlations in old animals contributed to the decreased decoding performance. We thus next quantified the effect that increased correlations had on the decoding performance in the old animals. To do this, we artificially increased the amount of correlation found in the young animal data to produce an artificially “aged” model (Fig. 9A). First, as increasing numbers of neurons did not improve performance in the old model (Fig. 8B), we used only the number of neurons found in young before the increase in neuron number significantly improved performance (5 neurons; Fig. 8B). Then, to mimic the effect of increased correlations, we created noisy copies of these neurons by multiplying each trial of the neuron by a random weight between 50% and 150% (Fig. 9A). This process of creating noisy copies was repeated until the model had 60 total neurons. We then measured model performance, and found that the young-aged model no longer performed better than the old model at low SNR (Fig. 9B, age × SNR ANOVA, post hoc tests: 0 dB, p = 1; 10 dB, p = 0.93; 20 dB, p = 0.04). Additionally, there was no longer any difference in performance at predicting FAs between the old model and the young-aged model (Fig. 9C). Together, these results show that increasing the correlation of the network decreases the behavioral information contained within the network.
Neural decoding impairment in older animals consistent with increased correlation. A, Young-aged data were created by noisy copies of neurons by multiplying each trial of the neuron by a random weight between 50% and 150%. This process of creating noisy copies was repeated until the model had 60 total neurons from five original neurons. B, Young-aged model does not perform better than the model trained on old animals (age × SNR ANOVA, post hoc tests: 0 dB, p = 1; 10 dB, p = 0.93; 20 dB, p = 0.04). C, The young-aged now predicts FAs at a rate that is not significantly different from the old model (age × SNR ANOVA, post hoc tests: 0 dB, p = 0.15; 10 dB, p = 0.43; 20 dB, p = 0.66).
Discussion
We investigated how central changes with aging in A1 lead to decreased performance while detecting tones in noisy backgrounds. We found that, similar to humans, old mice have worse behavior on a tone detection task in noisy backgrounds at lower SNRs. We found that this decrease in behavioral performance was not because of differences in tuning of neurons at the target frequency, but because of an increase in FAs because of increased prestimulus activity and increased activity correlations. Old mice also did not respond to tone offset further decreasing performance because of the decreased temporal offset responses and reduced neuronal suppression.
Neural responses of the auditory cortex during a behavioral task are shaped by the responses to the auditory stimulus, attentional gain, and behavioral outcome (Francis et al., 2018). We show that aging affects all these factors, but that behavioral performance is most closely related to changes in attentional gain and that population activity is related to behavioral outcome. Thus, our results suggest that changes in intrinsic A1 processing underlie behavioral changes in aging.
We find that aging mice have reduced suppressive gain during task performance (Fig. 3). This reduced gain can lead to increased prestimulus activity in otherwise suppressed neurons causing FAs. Suppressive gain can be because of a reduction in excitation or an increase in inhibition. While decreased inhibition is associated with healthy aging (Caspary et al., 1995, 2008; Milbrandt et al., 1996; Llano et al., 2012; Richardson et al., 2021), only a minority of neurons in A1 show suppressive gain indicating that a deficit in inhibition would have to be specific. Other sources of suppressive gain could be a decrease in the strength of excitatory inputs. Layer 2/3 neurons receive excitatory inputs from most cortical layers (Meng et al., 2017); and in particular, inputs from deep cortical layers are thought to play a role in gain control (Bortone et al., 2014; Guo et al., 2017; Clayton et al., 2021). We speculate that aging alters such deep layer inputs to layer 2/3.
On a network level, we found increased noise correlations in old animals, which we determined was caused by both higher correlations in passive, an inability to suppress these increased correlations during behavior, as well as an increase in behaviorally related correlations. This indicated that multiple mechanisms are affected by aging. First, mechanisms controlling baseline correlations have reduced efficacy; second, mechanisms increasing the correlations of subsets of neurons have increased effectiveness. Decreased inhibition could underlie the increase of baseline correlations and also contribute to the increase in correlations during behavior in Δr+ neurons. In addition, changes in excitatory circuits (e.g., more shared input) could be present and boost correlations in Δr+ neurons. However, while Δr+ neurons show an effect of aging, Δr– neurons in old and young animals show similar changes with behavior, suggesting that specific circuits control Δr+ and Δr– neurons during behavior.
We find decreased ability of mice to detect stimuli in a constant noise background. In healthy A1, stimulus-specific adaptation allows A1 neurons to reduce their sensitivity to redundant or irrelevant stimuli and create noise invariant representations of attended stimuli (Rabinowitz et al., 2011, 2013; Natan et al., 2015; Cooke et al., 2018). This suppression of activity relies on complex networks of inhibitory neurons to shape the receptive fields of cortical neurons (Chen et al., 2015; Natan et al., 2015; Kuchibhotla and Bathellier, 2018). Our results suggest that aging affects these inhibitory networks. This inability to fully suppress the neural response to the white noise background stimulus would also explain the increased stimulus-independent correlation as well as the increase in prestimulus activity.
Old animals show reduced neuronal offset responses (Shilling-Scrivo et al., 2021), and we here find reduced behavioral offset responses. This lack of offset responses can be detrimental to tone detection in noise as they are critical for auditory scene analysis (Bregman, 1994). This supports the notion that the loss of offset responses impairs proper stimulus detection in noisy environments as offset cues are not available. Offset responses in A1 are thought to be because of nonlemniscal inputs to A1 as well as inhibitory processing (Liu et al., 2019). Given that alterations in inhibition are associated with aging, we speculate that abnormal inhibitory processing underlies the deficits in offset responses. This lack of offset responses might also underlie the inability to discriminate speech in noise in older adults that exhibit otherwise good hearing in quiet (Dubno et al., 1984; Working Group on Speech Understanding and Aging, 1988; Gordon-Salant and Fitzgibbons, 1995, 1999; Fitzgibbons and Gordon-Salant, 1996; Lister et al., 2002; Lee, 2015).
Noise correlations are an indirect measure of functional connectivity between neurons. Several studies have shown that attention decreases overall correlations, with performance correlating to the level of reduction (Cohen and Maunsell, 2009; Downer et al., 2017; Nandy et al., 2017). While we did see reduced noise correlations in young animals with good performance, the old animals had high correlations regardless of performance. It is possible that the aging animals were unable to reduce their correlations to the level seen in the young animals, leading to their impaired task performance. However, aging animals were able to perform at 20 dB SPL despite the increased baseline correlations and decreased suppressive gain. Since the changes in noise correlations are predictive of the behavioral outcome, we speculate that aging animals have adopted different processing strategies and use different higher-order auditory and prefrontal areas (Winkowski et al., 2013, 2018) to perform properly. Indeed, recent work in humans showed stronger synchrony in activity between prefrontal and sensory areas in aging (Alain et al., 2022). Moreover, we find that old A1 prestimulus activity was predictive of when the animal would make a FA, but activity during the trial was less predictive of whether the animal would correctly detect the stimulus. Finally, we show that artificially increasing the noise correlation in the young animal model is sufficient to reduce decoding accuracy to the old animal level. This indicates that non–stimulus-related factors, and thereby other cortical areas, contribute to the performance of aged animals.
Our analysis also found a significant effect of sex. Old male animals had significantly less suppressive gain (Figs. 3, 5) and increased prestimulus activity leading to FAs (Fig. 8). This is consistent with prior work showing decreased suppression in old males while passively listening to tones in noise (Shilling-Scrivo et al., 2021). More work is needed to map the relationship between decreased neural suppression and the difficulty of hearing in noise, the circuits involved, and whether there is some neuroprotective effect of sex.
In conclusion, our results show that old animals are unable to fully suppress their responses to background noise. We speculate that impaired gain control cannot overcome increased activity correlations present in the aged brain, leading to the inability to separate targets from background noise. Our work also establishes increased noise correlations as a potential neural signature of central auditory deficits.
Footnotes
This work was supported by National Institutes of Health P01AG055365 to P.O.K., RO1DC009607 to P.O.K., and R01DC017785 to P.O.K. We thank Nik Francis, Zac Bowen, Travis Babola, and Ji Liu for technical help.
The authors declare no competing financial interests.
- Correspondence should be addressed to Patrick O. Kanold at pkanold{at}jhu.edu
