Volatile working memory representations crystallize with practice

Mice

All of the experiments were conducted according to the National Institute of Health (NIH) guidelines and with the approval of the Chancellor’s Animal Research Committee of the University of California, Los Angeles. Experiments were performed with 8–15-week-old adult male and female C57BL/6J (Jackson Laboratory, 000664), C57BL/6J-Tg (Thy1-GCaMP6s), GP4.12Dkim/J (Jackson Laboratory, 025776) and B6;DBA-Tg(tetO-GCaMP6s)2Niell/J (Jackson Laboratory, 024742) mice crossed with B6.Cg-Tg(Camk2a-tTA)1Mmay/DboJ (Jackson Laboratory, 007004) mice. Mice were kept in the vivarium under a 12 h–12 h light–dark cycle.

Viruses

For optogenetic experiments, CaMKIIa-driven soma-targeted anion-conducting channelrhodopsin fused to FusionRed (pAAV-CKIIa-stGtACR2-FusionRed, Addgene, 105669; titre, 1 × 10¹³ viral genomes per ml) was used to express GtACR2 in the soma of excitatory neurons. For control experiments, we used pAAV-CaMKIIa-mCherry (Addgene, 114469; titre, 7 × 10¹² viral genomes per ml) or pAAV-CaMKIIa-EGFP (Addgene 50469; titre, 7 × 10¹² viral genomes per ml).

Head-bar and cranial window implantation

Adult 8-to-12-week-old male and female C57BL/6J-Tg (Thy1-GCaMP6s) GP4.12Dkim/J mice were anaesthetized with isoflurane (5% for induction, 1–2% (v/v) for maintenance). The depth of anaesthesia was monitored continuously and adjusted when necessary. After induction of anaesthesia, the mice were fitted into a stereotaxic frame (Kopf), with their heads secured by blunt ear bars and their noses placed into an anaesthesia and ventilation system (David Kopf Instruments). Body temperature was kept at 37 °C with a feedback-controlled heating pad (Harvard Apparatus). Mice were administered 0.05 ml lidocaine (2%; Akorn) subcutaneously as a local anaesthetic before surgery. The surgical incision site was cleaned three times with 10% povidone-iodine and 70% ethanol. After removing the scalp and clearing the skull of connective tissues, a custom-made lightweight metal head-bar was fixed onto the skull with cyanoacrylate adhesive and covered with black dental cement (Ortho-Jet). A circular craniotomy (diameter, 5 mm) was performed above the secondary motor cortex (centred at 1.94 mm anterior from bregma or centred at bregma for M2/RSA imaging). A cranial glass window consisting of a 5 mm diameter round #1 coverslip (Warner Instruments) was implanted in the craniotomy, flush with the skull surface and sealed in place using tissue adhesive (Vetbond). The exposed skull surrounding the cranial window was then completely covered with black dental cement to build a small chamber for imaging with a water-immersion objective. After surgery, the mice were injected with carprofen (5 mg per kg of body weight) and allowed to recover overnight in cages placed on a low-voltage heating pad. Carprofen was administered once per day for up to 2 days after surgery. Amoxicillin antibiotic (0.25 mg ml⁻¹) was dispensed in the drinking water for 7 days. Animals were returned to the vivarium for 1–2 weeks for recovery before undergoing imaging experiments.

AAV injection and fibre optic cannula implantation

Adult 8-to-12-week-old male and female C57BL/6J mice were anaesthetized with isoflurane (5% for induction, 1–2% (v/v) for maintenance). Skin incisions were made, followed by craniotomies 1 mm in diameter above the secondary motor cortex (centred at 1.94 mm anterior to bregma and 0.5 mm lateral to the midline) using a small steel burr (Fine Science Tools) powered by a high-speed drill. Saline (0.9%) was applied to the skull to reduce heating caused by drilling. Bilateral viral injections were performed by using stereotaxic apparatus (David Kopf Instruments) to guide the placement of bevelled glass pipettes with a tip diameter of about 50 μm (World Precision Instruments) into the secondary motor cortex (1.94 mm anterior to bregma, 0.5 mm lateral to the midline and 0.3 mm from the pial surface). Using the Nanoject II micro-injector (Drummond Scientific), 300 nl of 1:100 PBS-diluted AAV was bilaterally injected using a syringe pump. Glass pipettes were left in place for at least 10 min after virus injection.

A ferrule-terminated optical fibre (Thorlabs) was placed above the injected site. The fibre tip was aimed to terminate at the pial surface. The optical fibre was secured to the skull using cyanoacrylate adhesive and black dental cement (Ortho-Jet). After surgery, the mice were left overnight in cages placed on a low-voltage heating pad. Mice were allowed to recover for 2–3 weeks before the experiments. The locations of injections and implanted optical fibres were validated histologically for all experimental mice.

Behavioural training

After recovery from surgery, mice were handled and water-restricted to 85–90% of their original weight. The mice were subsequently habituated to head fixation, airflow and water port for two sessions (one session per day). During the two shaping days, the mice were presented only with the combination of the odours (A, 1-pentanol; B, butyl formate; C, 3-methyl-2-buten-1-ol; and D, ethyl acetate; Sigma Aldrich, 138975, 261521, 162353 and 270989) that led to reward (AC and BD trials) and water was automatically delivered. After 2 days of shaping, the mice were trained to perform the complete delayed-association WM task. The lick port was connected to a touch sensor, and mouse tongues had to touch the lick port at least once to receive a water reward. Each training session consisted of 150 to 250 trials. Odour combinations were presented in a random order. Responses were assessed based on mouse licking during the choice window. If any licks occurred during the choice window, the trial was considered to be a hit for AC and BD trials or false alarm for AD and BC trials. If no licking occurred during the choice window, the trial was considered to be a miss for AC and BD trials or correct rejection for AD and BC trials. Mice were not punished for miss or false alarm trials. A training session was aborted early if a mouse had more than three misses within the most recent ten trials, indicating the animal’s lack of motivation to obtain the water reward. Performance was quantified as the number of hits and correct rejections over the total number of completed trials. The airflow and odour delivery were frequently monitored using an Aurora Scientific photo-ionization detector at the beginning of each training session.

In vivo calcium imaging

Two-photon laser-scanning microscopy was conducted using the Thorlabs multiphoton mesoscope using a 12 kHz resonant scanner with a water-immersion objective with 0.6 excitation NA, 1.0 collection NA and 2.7 mm working distance. The excitation laser was a 920 nm Tiberius Ti:Sapphire Femtosecond Laser, and the laser intensity was 30–80 mW at the sample. Images were acquired using the ScanImage software (Vidrio Technologies). Fully awake mice were mounted in a 2-inch-diameter transparent tube by securing its head bar onto a custom-made head-bar holder under the microscope. 600 px × 1,200 px to 600 px × 2,500 px images were acquired at 8–17 Hz at 150–250 μm depth. To track the mouse movement, a camera mounted underneath the animal acquired the paw location of the animals at 30 Hz. The locomotion data were acquired simultaneously with the calcium imaging data and synchronized with the scanning mirror signals. The microscope and behavioural set-up were encased in a light-tight box, and the mice were kept in darkness during the imaging sessions. We performed online image processing at the beginning of every session to align cells across days. We tried to maximize the correlation between the moving average of frames of the current field of view and the average of frames of the previous sessions.

Two-photon LBM was conducted using a custom-built microscope equipped with a 960 nm, 4.89  MHz repetition rate optical parametric chirped-pulse amplification (OPCPA) pumped by an ytterbium laser at 1,030 nm with 80 W power, delivering a 2 μJ pulse energy and a 90 fs pulse width. The LBM featured a rapid 12 kHz resonant scanner and was paired with a 0.6 excitation NA, 1.0 emission water-immersion objective lens with a 2.7 mm working distance. The LBM technique divided a single pulse into 30 distinct subpulses of varying intensities, targeting 30 separate depths of the specimen separated by 15 μm, yet eliciting a consistent level of fluorescence across these layers²². In our initial LBM experiments, we successfully recorded a region measuring 1,450 × 1,825 × 450 μm³ at a frequency of 7.95 Hz in two mice. Subsequent experiments extended the recorded area to 2,000 × 2,000 × 450 μm³, recorded at 6.45 Hz, in another two animals.

Optogenetics

Optical stimulation was applied through a ferrule-terminated 200 μm core and 0.39 NA optical fibre (Thorlabs) attached to the 200 μm core and 0.39 NA patch cable using a 1.25 mm ceramic mating sleeve (Thorlabs). We used a blue-fibre-coupled light emission diode (λ = 470 nm, Thorlabs, M470F3). The light was delivered at 20 Hz with a 0.4 duty cycle at an irradiance of 10 mW mm⁻² at the output tip of the fibre.

Optogenetic experiments commenced only when the animals achieved a behavioural performance threshold exceeding 90% accuracy for at least three consecutive sessions. This criterion ensured that the animals were well-trained and proficient in reliably executing the behavioural tasks before the introduction of optogenetic manipulations.

Electrophysiology

For in-vivo electrophysiology recordings, expert mice were anaesthetized with isoflurane (5% for induction, 1–2% (v/v) for maintenance). They underwent a 2 mm craniotomy (centred at 1.94 mm anterior to bregma and 0.5 mm lateral to the midline) and silver wire ground (Warner Instruments) implantation surgery over the cerebellum 1 day before recording. The ground wire was fixed in place with dental cement. The exposed skull was covered with Kwik-Sil, and the mouse was allowed to recover overnight. On the day of the recording, the mice were head-fixed into a tube, the Kwik-Sil covering the craniotomy was removed and replaced with buffered artificial cerebrospinal fluid, and the mouse was aligned to the micromanipulator. A 128-channel silicon microprobe³⁷ was slowly lowered using a micromanipulator into M2, and the surface of the exposed brain was covered with mineral oil. The process was monitored using a surgical microscope (Zeiss, STEMI 2000). The microprobe contained 128 channels that were densely distributed (honeycomb layout with 20 μm spacing between nearest-neighbour channels) on two shanks (placed 0.4 mm apart). After insertion, the microprobe was allowed to settle for at least 30 min before the recording began and continued for the entire duration of the session. The electrophysiological and behavioural data acquisitions were synchronously performed using custom MATLAB software while the mouse performed the task. The probe readout was achieved using a detachable head stage module (Intan Technologies RHD 128). Head stages contained commercial integrated electronic circuits (Intan Technologies RHD 2000 USB interface board) providing a multiplexed signal recorded with open source software (Intan Technologies) at 25 kHz per channel.

Histology

At the end of experiments, the mice were deeply anaesthetized under isoflurane and transcardially perfused with 40 ml 1× PBS followed by 40 ml 4% paraformaldehyde in 1× PBS at a rate of approximately 4 ml min⁻¹. After perfusion, the brains were rapidly extracted and post-fixed in 4% paraformaldehyde. Coronal sections (thickness, 100 μm) were collected using a vibratome. The sections were mounted onto glass slides. The slides were then cover-slipped with mounting medium DAPI. Images were acquired using the Leica DM6 B microscope.

Quantification and statistical analysis

Calcium imaging data processing, including motion correction, segmentation, fluorescence signal extraction and deconvolution, was performed using the Python implementation of Suite2P³⁸. Before segmentation, we performed several steps to enhance image quality, including noise reduction, background subtraction and image registration to correct for tissue movement. We validated our segmentation results by comparing the automated segmentation to manually annotated ground truth data. Adjustments to parameters and algorithms were done to achieve optimal results. We used the deconvolved signal for all our analyses. Silicon probe data processing and spike sorting were performed using custom code, KiloSort³⁹ and Phy⁴⁰.

To visualize the calcium activity of individual neurons, we computed a peristimulus time histogram averaged across all trials for all four combinations of odours, smoothed using a moving average over a 400 ms window. To generate response maps for each neuron, we subtracted its mean spontaneous baseline calcium activity across all trials on a given day during the baseline epoch (5 s before the first-odour onset). We divided it by the s.d. of calcium activity during the baseline epoch. Thus, the response maps show changes in calcium activity in units of the s.d. of spontaneous activity. This method was used for visualization purposes only. Unless stated otherwise, all statistical analyses were performed on unsmoothed, deconvolved calcium activity without baseline calcium activity subtracted.

A neuron was considered to have a significant activity field during a specific time epoch if its activity within that epoch significantly differed from the distribution of its 1,000 times circularly shuffled mean activity.

The first-odour selectivity of a neuron was assessed by comparing the distribution of its mean deconvolved calcium activity over a time epoch for A and B odour trials using the Wilcoxon rank-sum test with a confidence interval of 99%. A neuron was considered to be purely selective if it exhibited selectivity for a specific odour or choice during a specific epoch and did not show selectivity for any other parameter at any other time. Conversely, a neuron was considered to be mixed selective if it showed selectivity for more than one odour or choice at different epochs.

We considered an animal naive, training or expert if its behavioural performance (p) was, respectively, p < 65%, 65% ≤ p < 80% or p ≥ 80%. An animal was considered a novice during the first training day.

To determine whether a neuron’s response was related to the animals’ motor activity, we used DeepLabCut⁴¹ to find the position of the animals’ paws from which we extracted the animals’ movements. We calculated the correlation coefficient between the activity of each neuron and the unshuffled and 1,000 times circularly shuffled locomotion activity. A neuron was considered to be significantly correlated if its correlation coefficient was at least 2 s.d. away from the mean value of the correlation coefficients of shuffled distribution.

We assessed the WM information content in a population activity of neurons by measuring the classification performance using a SVM with a linear kernel. We implemented the SVM binary classification in MATLAB and performed the computations on high-performance computing clusters using thousands of computing nodes. We used the activity of neurons in 500 ms time bins to train a decoder on 90% of randomly chosen trials and tested its accuracy on the 10% of the trials that were withheld. To ensure our model was not biased or overfit to specific data patterns, we repeated the classification measurements at least 32 times with different sets of randomly chosen trials. We then calculated the average of all measurements. This approach introduces randomness and helps to ensure that the decoding results are not a product of a model memorizing specific instances. Decoding accuracy and its standard error were then found by averaging the prediction accuracy of the decoder across all mice.

For across-day classification, we used the activity of the overlapping neurons for our analyses. We trained a model on all trials on one day and tested that model’s predictions on all trials on another day. To assess statistical significance and determine whether decoding performance surpasses chance, we randomized trial types by assigning random labels to each trial.

For the LSTM decoding analyses, we configured the recurrent neural network architecture so that the dimensions of the input layer were aligned with the number of neurons recorded in the pertinent dataset. The network comprised 128 hidden units followed by a linear layer to compute logits for a softmax classifier using cross-entropy loss. The weights and biases within the LSTM layers during the training phase were optimized using the adaptive moment estimation (Adam) optimizer. LSTMs were trained for 100 epochs (passes through the training set). For the decoding process, temporal data granularity was 500 ms. Training involved 90% of randomly selected trials, with the remaining 10% reserved for testing. Analogous to the SVM decoding procedure, the LSTM classification was iterated at least 32 times using distinct randomly chosen trial subsets. The resultant metrics were then averaged over the 32 samples. Furthermore, we performed trial shuffling to mitigate potential biases and ensure the robustness of the LSTM model’s performance assessment.

To find overlapping neurons across sessions, we used co-registration of spatial cell footprints using CellReg⁴². Neurons were modelled with a maximal centroid distance of 10 μm. Final registration used the probabilistic model with a threshold of more than 95% probability of cells being the same for all mice. We used these parameters to ensure the accuracy of matching cells.

We used t-SNE⁴³ to embed the high-dimensional neuronal activity into two dimensions. We calculated the time-averaged calcium activity of neurons during a specific epoch and found the pairwise distances between the high-dimensional points for each trial. For each point, we calculated a s.d. so that the perplexity of each data point matched a predefined value. Starting from an initial set of low-dimensional points, we iteratively updated the points to minimize the Kullback–Leibler divergence between a Gaussian distribution in the high-dimensional space and a t-distribution in the low-dimensional space.

Statistics and reproducibility

All statistical analyses were conducted using Prism (GraphPad), MATLAB (MathWorks) or Python. Statistical tests used in this study include Wilcoxon rank-sum tests and paired t-tests. The significance threshold was held at α = 0.05; NS, not significant (P > 0.05); *P ≤ 0.05, **P ≤ 0.01, ***P ≤ 0.001, ****P ≤ 0.0001. All behavioural, imaging and optogenetics experiments were replicated in multiple animals. Sample sizes were not predetermined using statistical methods.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.