Linking systemic metabolic state to long‐term motor memory: Insights from respiratory exchange ratio and glucose manipulation

Abstract

Abstract

Respiration is a crucial metabolic process that converts macronutrients, carbohydrates and fats, and oxygen into energy and carbon dioxide to support motor actions. Moreover, the brain is a significant energy consumer, accounting for large portions of the body's total energy expenditure and relying primarily on carbohydrates for neural activity and plasticity. However, it is not known whether gas composition in breathing can serve as an indicator of neural activity and plasticity as they can for movement intensity. In human reaching movement tasks, we evaluated time‐constants of sensorimotor learning during the recording of gas exchange. We computed the respiratory exchange ratio (RER), indicating whether carbohydrate or fat is used preferentially, and found that the RER was unaffected by the execution and learning of reaching movements and that it was stable within but varied across individuals. Interestingly, using computational modelling to identify short and long‐time constants of sensorimotor learning, individual RER levels correlated with the estimated slow component of learning dynamics, suggesting a link between metabolic state and processes underlying long‐term retention. To probe this further, we used glucose administration, known to increase RER by promoting carbohydrate utilisation, before training. Regression analysis indicated that glucose‐induced RER increases during training were associated with enhanced estimated 24 h retention at the intra‐individual level. Together, RER is associated with processes underlying long‐term memory acquisition and retention, and glucose administration shifted the physiological idling state for the processes. Unravelling the specific neurobiological pathway from these intriguing breathing metrics to brain function emerges as a compelling new research direction.

Key points

• The brain is a major energy consumer (20% of total energy from only 2% of body weight), primarily using carbohydrates for neural activity and plasticity.

• The respiratory exchange ratio (RER) in breath signals the body's balance of fat‐carbohydrate fuel use; this study explored whether RER reflects neural processes in motor memory acquisition and retention.

• Individual RER, stable during reaching tasks but varying across participants, correlated with the computationally estimated slow component of learning dynamics, which is linked to long‐term retention.

• Glucose administration, known to increase RER, was associated with improved estimated 24 h motor memory retention at an individual level.

• The results suggest that RER indicates long‐term motor memory processes and that manipulating RER via glucose may enhance motor memory, offering a new neurobiological pathway from these intriguing breathing metrics to memory function and potential practical implications for a simple but plausible intervention.

Abstract figure legend The brain is a major energy consumer relying primarily on carbohydrates. This exceptional metabolic demand suggests that its functional state, which includes cognitive processes such as motor memory acquisition and retention, should be observable in the body's systemic metabolism. To explore this link, we investigated whether the respiratory exchange ratio (RER), comprising a systemic measure of the body's balance of fat‐carbohydrate fuel use, could serve as a physiological marker for the neural processes. Our findings revealed that an individual's RER was specifically correlated with the computationally estimated slow component of learning dynamics. Furthermore, we demonstrated that this relationship is modifiable; glucose administration, known to increase RER, was associated with an estimated enhancement in 24 h retention at an individual level. Together, these results provide compelling evidence that RER can indicate long‐term motor memory processes and that manipulating this systemic metabolic state may offer a new pathway to enhance cognitive function.

Introduction

Respiration is a crucial process in metabolism responsible for the generation of ATP, an energy currency that supports all motor actions (Fig. 1A ). Metabolism converts macronutrients, primarily carbohydrates and fats, along with oxygen (O₂), into ATP and carbon dioxide (CO₂). These metabolic processes exhibit varying rates and efficiencies. Quick, high‐intensity motor actions require rapid energy production from carbohydrates, whereas prolonged, low‐intensity motor actions require significant energy production from fat. Several studies have shown that the intensity of movement, measured by force or muscle activity, can be approximated with gas composition in breathing and is energetically optimised by motor experience (Abram et al., 2022; Huang et al., 2012; Selinger et al., 2015; Shadmehr et al., 2016). Therefore, respiratory gas composition is recognised as a reliable measure for estimating the energy expenditure associated with movement execution.

Figure 1. Respiratory exchange ratio and its potential role for sensorimotor learning.

A, the metabolic cycle converts oxygen (O₂) and macronutrients, carbohydrates and fats, into carbon dioxide (CO₂) and ATP, the energy currency that supports neural activity, plasticity and muscle activity. Breathing, which involves O₂–CO₂ exchange, plays an important role in this process, but it is unknown whether breathing can predict learning and memory, which are influenced by neural activity and plasticity. B, the respiratory exchange ratio (RER), calculated as the ratio of CO₂ production (VCO₂) to O₂ consumption (VO₂), reflects the metabolic preference for using carbohydrates or fats as fuel. In the cortex, where only carbohydrates are used, the RER approaches 1.0. C, we designed Experiment 1 to test whether the RER, calculated by O₂–CO₂ exchange, represents sensorimotor learning of reaching movements when recording muscle activity. D, visuomotor rotation (VMR), which angularly shifts the position of the visual cursor from the hand position, was introduced to elicit sensorimotor learning. This example shows counterclockwise VMR; the leftward visual error shifts the reaching movement to the right. E, a theory posits that two motor memories with different time constants, fast and slow, are stored in sensorimotor learning. The short‐ or long‐time constants of sensorimotor learning, which presumably correspond to neural activity and plasticity. F, to accurately estimate motor memories with varying time constants, a dynamic VMR comprising five sinusoidal functions with different frequencies ranging from 2 to 32 Hz. It is hypothesised that rapidly changing VMR induces fast motor memory formation, whereas gradually changing VMR elicits the development of slow motor memory. G, the sum of the sinusoidal function is used in the reaching movement task.

In addition to the energy demands for muscle contraction, the brain is a significant energy consumer, utilizing 20% of the total energy expenditure despite comprising only 2% of body weight (Raichle & Gusnard, 2002). In particular, cortical energy is crucial for both neural activity (Oakes et al., 2004; Sokoloff, 1977) and plasticity (Lee et al., 2013; zur Nedden et al., 2011), which must play pivotal roles for sensorimotor learning. Sensorimotor learning not only establishes new patterns of neural activity (Li et al., 2001; Paz et al., 2003; Sun et al., 2022), but also induces novel cortical connections through long‐term potentiation‐like mechanisms (Kida et al., 2016; Landi et al., 2011; Rioult‐Pedotti et al., 2000, 1998). Notably, these processes operate on different time scales: trial‐by‐trial emergence of new neural activity (Perich et al., 2018; Veuthey et al., 2020) contrasts with the slower, structural consolidation of connections, such as dendritic spine growth within 1 h at the earliest (Harms et al., 2008; Xu et al., 2009). This temporal dichotomy may mirror a prevalent computational framework in sensorimotor learning, in which fast and slow processes interact to update motor memory (Smith et al., 2006). The fast process adapts quickly but decays rapidly, whereas the slow process learns gradually and supports longer‐term retention. The computational model incorporating these dynamics successfully captures behavioural features of both short‐ and long‐term memory formation (Joiner & Smith, 2008). Given the energetic demands of both fast and slow learning processes, we hypothesised that breathing metrics might provide a window into the temporal characteristics of sensorimotor learning by reflecting cerebral metabolism.

To test this hypothesis, we conducted human behavioural experiments involving reaching movement tasks during the recording of O₂–CO₂ gas exchange. We focused on the respiratory exchange ratio (RER) (Fig. 1B ), which is the volume of CO₂ produced divided by the volume of O₂ consumed (McArdle et al., 2010), because it indicates the metabolic state. Carbohydrate metabolism converts O₂ to CO₂ in a 1:1 ratio (RER = 1.0), whereas fat metabolism converts O₂ to CO₂ in a 1:0.7 ratio (RER = 0.7). Thus, the RER theoretically varies between 0.7 and 1.0 at rest, depending on the macronutrient usage. We observed that the RER was not influenced by the execution and learning of reaching movements; rather, it was stable within individuals but varied across individuals.

The brain utilises only carbohydrates for the energy resource, implying that shifts towards carbohydrate metabolism (i.e. a higher RER level) might indicate increase in the brain's energy expenditure for learning and memory. Interestingly, applying a computational model to decompose learning processes into fast and slow components, we found that individual RER levels showed a correlation with the estimated magnitude of the slow component, often associated with long‐term retention, but not the fast component. Furthermore, to explore whether manipulating this metabolic index during training could influence subsequent memory consolidation, we provided 200 kcal of glucose immediately before training and tested retention 24 h later. Although this intervention increased RER during training, inter‐individual variability in RER appeared to obscure a clear group‐level comparison of 24 h retention. However, regression analysis suggested that glucose‐induced RER increases during training were associated with enhanced estimated 24 h retention. Together, these findings suggest that RER, as an index of baseline metabolic state, shows an intriguing association with processes related to long‐term motor memory acquisition and retention, and that modulating metabolic state with glucose intake might shift towards the physiological context for the processes.

Methods

Participants

In total, 18 participants (18 males, aged between 21 and 52 years) in Experiment 1 and 27 participants (16 males and 11 females, aged between 19 and 39 years) in Experiment 2 were recruited; however, we removed data from three participants in Experiment 2 because of their data quality (See Data analysis). All participants were right‐handed, reported no neurophysiological disorders and provided written informed consent prior to the experiments in accordance with the Declaration of Helsinki. The Committee on Ergonomic Experiments of the National Institute of Advanced Industrial Science and Technology (No. HF2022‐1183) approved the protocol of Experiment 1 and the research ethics committee of the University of Tokyo (No. 23–204) approved the protocol for Experiments 1 and 2, respectively.

Reaching movement task

Participants performed rapid 20 cm, point‐to‐point, arm reaching movements with their right hand when sitting on a chair positioned at the centre of the task space (Fig. 1C ). We used an exoskeleton robotic manipulandum in Experiment 1 (KINARM Exoskeleton Lab; Bkin Technologies, Kingston, ON, Canada), which can prevent unwanted fatigue and maintain a constant posture to reliably record muscle activity with electromyography (EMG), whereas we used a digitizing tablet and pen encased in a light grip in Experiment 2 (Wacom Intuos Pro L; Wacom, Kazo, Japan) to reduce muscle activity that might affect the breathing data. The kinematic data were recorded from the robotic manipulandum (the sampling rate was 1000 Hz) in Experiment 1 and the digitizing tablet (200 Hz) in Experiment 2, and the refresh rates of the visual screen were both 60 Hz.

At the beginning of each trial, the participants were asked to hold a white circle (1.0 cm in diameter) representing their hand position on a red start circle (1.5 cm). One second later, the red target circle (1.5 cm) was displayed 20 cm in front of the start circle. One second later, the target colour changed to green, which was a cue to initiate a forward‐reaching movement. Upon reaching the green target, the target colour changed to blue and the start circle turned green, which was a cue to initiate a backward‐reaching movement. Outback movements were completed, and the next trial started after an inter‐trial interval of 1 s. A preliminary experiment showed that the outback movements took ∼2 s. A trial duration of ∼5 s was chosen because each trial would involve at least one breath, which occurred approximately once every 4 s. After the participants familiarised themselves with the task, we introduced visuomotor rotation (VMR), which involved angularly shifting the cursor position (Fig. 1D ).

To dissociate the fast and slow components of sensorimotor learning in Experiment 1 (Fig. 1E ), we designed the sum of sinusoidal functions (Fig. 1F and G ) (Miyamoto et al., 2020). The frequencies of the functions were [2, 4, 8, 16 and 32] Hz over 900 trials. The two‐state model posits that, in sensorimotor learning, two motor memories are updated concurrently with different time constants (Smith et al., 2006). This unique idea, other from more recent frameworks (Heald et al., 2021), is leveraged to reliably estimate time features of motor memory, where slow motor memory predicts next‐day retention at 99% of variances in the group‐level analysis and at 50% of variances in the individual level (Joiner & Smith, 2008). Accordingly, we predicted that fast motor memory would adapt to VMR changes more quickly (16 and 32 cycles), whereas the slow motor memory would adapt to VMR changes more gradually (two and four cycles). The amplitude of the function was 10°. The maximum amplitudes of the dynamic perturbation were ±30.9° and maximum changes in a trial were ±4.3°. We counterbalanced the signs of the amplitudes of the 18 participants. In total, 900 training trials were divided into 12 blocks, with an ∼1 min break between blocks.

In Experiment 2, to test the 24 h retention with the glucose intervention (See Glucose administration) (Fig. 2A), participants adapted to a constant VMR perturbation on day 1 and were tested for motor memory without visual feedback on day 2 (Fig. 2B ). The constant perturbation was 30°, which was counterbalanced across the 24 participants. There were 220 trials after 20 null trials on day 1 and 40 non‐visual trials on day 2. The 1 min break was given every 10 trials to remove short‐term memory and encourage long‐term memory (Hadjiosif et al., 2023).

Figure 2. Glucose manipulation and its potential influence on long‐term motor memory.

A. Based on the results of Experiment 1, which showed that the RER reflects slow motor memory, we tested the hypothesis that increasing the RER with glucose intake would improve 24‐hour retention. B. Experiment 2 was designed to examine how much of the 30° constant VMR learned on day 1 was retained on day 2 during reaching movements without visual feedback. C. On day 1, the RER was recorded for 20 min before and after an intervention in which one group received 200 kcal glucose (WG, n = 16) while the other group did not (Wa, n = 8). Both then performed the same training task. On day 2, the RER was recorded for 20 minutes, followed by the retention task.

Gas exchange

We recorded the O₂–CO₂ gas exchange using a portable metabolic measurement system in the breath‐by‐breath mode (COSMED K5; COSMED Srl, Albano Laziale, Italy). In Experiment 1 (Fig. 1C ), we started recording the RER from the familiarisation phase, which performed reaching movements without any perturbation before the experiment, to the end of the experiment to examine how RER, both at baseline before the experiment and during the task itself, can predict the fast and slow components of sensorimotor learning and how the RER alters during sensorimotor learning. In Experiment 2 (Fig. 2C ), to investigate how glucose intake modulates the RER, we recorded before and after the intervention phase (pre‐ and post‐intervention phases) on day 1 (Fig. 2C ). Both phases took 20 min, as determined by the data analysis in Experiment 1, and the data reliability, calculated by the mean squared error against the mean of the whole data, was saturated at 20 min. The gas exchange was recorded continuously during the training phase. On day 2, we recorded the pre‐experiment baseline gas exchange for 20 min and then continuously recorded it during the retention phase. Before the experiment, the K5 system was calibrated in accordance with the manufacturer's instructions.

Electromyography

In Experiment 1 (Fig. 1C ), muscle activity was recorded using a wireless surface electromyography system (WavePlus EMG system; Cometa, Bareggio, Italy). The wireless system had a very low and consistent delay (14 ms reported by the product specifications), which had little to no impact on the results in reaching movements that took 1980 ± 87 ms. Data were recorded from four muscles: the pectoralis major, posterior deltoid, biceps brachii and triceps brachii, which govern large portions of shoulder extension/flexion and elbow extension/flexion during horizontal reaching movements. The electrode size was 41 × 15.5 × 11.3 mm and the inter‐electrode distance was 20 mm (Pico EMG; Cometa). The electrode sites were carefully determined by visually checking the data quality during the ballistic reaching movements before the experiment. The EMG and kinematic data were recorded using an A/D signal converter (PCI‐6229; National Instruments, Austin, TX, USA) at a sampling rate of 1000 Hz. The EMG signals were amplified (×1000) and band‐pass filtered from 10 to 500 Hz via the hardware.

Glucose administration

In Experiment 2 (Fig. 2C ), we examined the effects of glucose intake on RER and 24 h retention. We instructed the participants in both the ‘Water and Glucose (WG, n = 16)’ and ‘Water alone (Wa, n = 8)’ groups not to consume any food or drink with sweeteners for at least 1 h before the experiment, to maximise the effects of the intervention. The WG group consumed water and gel‐type drinks (Glucorescue, Arkray, Japan), whereas the Wa group consumed only water during the intervention phase; therefore, they had to take off their masks to record gas exchange. The main ingredient of the gel‐type drinks was 50 g of glucose, corresponding to 200 kcal. Because the WG group took 387 ± 40 s, the Wa group also waited for ∼360 s, to match the periods of the intervention phase. Note that there was no significant difference in this 24 h interval between the WG (24.4 ± 0.493 h) and Wa groups (24.5 ± 0.500 h, t ₂₂ = 0.370 and P = 0.715).

Data analysis

We smoothed the handle position using a fourth‐order Butterworth filter with a cutoff frequency of 10 Hz. Subsequently, we found the midpoints of the forward movement, that is, 10 cm away from the start position. This is almost identical to the position at peak velocity, which represents the feedforward predictive component of movement. We computed the movement directions at the midpoints relative to the target direction, because VMR is an angular perturbation. We performed outlier removal in each trial of each experiment if the data were outside the median ±3 IQRs across participants, which was almost 0.05% under the hypothetical standard distribution. We removed one participant from Experiment 2 because the data were very distant from the others, especially in the retention phase (38 out of 40 retention trials were outliers). To compare the spectral amplitudes at trained vs. untrained frequencies in Experiment 1, we defined the mean amplitude of untrained frequencies as the average of the amplitudes at all integer frequencies between 2 and 32 Hz, excluding the five trained frequencies ([2, 4, 8, 16 and 32] Hz). For the analysis of learning and retention in Experiment 2, we defined two key performance metrics. The asymptote of learning on day 1 was calculated as the average of the final 40 trials of the training phase. The level of retention on day 2 was calculated as the average of the final 20 trials of the no‐feedback retention test.

We used a two‐state model to represent short‐ and long‐term memory in sensorimotor learning (Smith et al., 2006). The model considered that the total motor outputs ($x$) were determined by the sum of two distinct memories ($x=x_f+x_s$) and that the memories are parallelly updated with the error ($e=x+p$, $p$ is VMR) as $x_f(n+1)={\alpha }_f{x}_f(n)+{\beta }_{f}e(n)$ and $x_s(n+1)={\alpha }_s{x}_s(n)+{\beta }_{s}e(n)$, where $f$ and $s$ indicate labels of fast and slow components of sensorimotor learning, $n$ indicates the trial number and $\alpha$ and $\beta$ are retention and learning factors. The retention and learning factors are considered to be different between the components: $0<{\alpha }_f<{\alpha }_s<1$ and $0<{\beta }_s<{\beta }_f$. We fitted the model to individual data using expectation maximisation methods. This approach allows for robust parameter estimation from noisy individual data and requires significantly fewer subjects for hypothesis testing compared to traditional least‐squares algorithms (Albert & Shadmehr, 2018). In Experiment 1, we recruited a number of participants comparable to those in previous studies (Albert & Shadmehr, 2018). Basically, we utilised the original code that is available online (http://shadmehrlab.org/tools), although we fixed a few initial parameters for the parameter searching: narrower ranges of initial fast and slow motor memories from ±30° to ±5°. Narrower ranges that emphasised zero learning amounts at baseline were more reflective of the current situation. In addition to Experiment 1, we also applied a similar two‐state model to the behavioural data from Experiment 2 to further investigate the relationship between RER and the different memory components. To account for the decay in memory across the 24 h interval, the model incorporated a parameter that allows for a proportional decay of the memory states, a method implemented in the publicly available code by Albert and Shadmehr (2018). For some analyses where outliers were suspected, we utilised a robust regression to minimise their influence on the correlation results.

The O₂–CO₂ gas exchange ratio was used to compute the RER offline for each breath. The RER was recorded breath‐by‐breath and it was obtained at discrete times over a few seconds. For the RER during tasks, we performed linear interpolation between the RER to match the kinematic data with a sampling rate of 1000 Hz and used the RER at the midpoint of movement as a representative index in the trial. The range of RER was theoretically defined as 0.7–1.0, the outsides of which were regarded as outliers. In particular, we removed two participants in Experiment 2 because ∼35% of trials were determined as outliers, which is a high ratio compared to the others (7.45 ± 6.70%, mean ± SD). For the steady‐state RER, such as the pre‐experiment phase in Experiment 1 and the pre‐ and post‐intervention phases in Experiment 2, we computed the mean of the RER from the individual breathing data after outlier removal in each breathing data.

EMG data were subtracted from the mean and rectified for each trial. Here, we used muscle activity to estimate the intensity of movement execution, as it affects the RER (Goedecke et al., 2000). Therefore, instead of EMG activity during the middle movement, we computed a trial‐wide summation of the EMG of each muscle. We then performed self‐normalisation with the mean across trials for each muscle. This allowed us to examine how muscle activity changes with sensorimotor learning, independent of the different EMG responses of the muscles.

To examine whether the intensity of the reaching movements influenced the subsequent RER in Experiment 1, we performed a lagged linear regression analysis within each of the 12 experimental blocks. For each possible time lag of i trials (where i ranged from 0 to 74), we regressed the muscle activity data from the beginning of a block (trials 1 to 75 – i) against the RER data from later in the same block (trials i + 1 to 75). From each regression, we computed the coefficient of determination (R ²) to quantify the variance in RER explained by the preceding muscle activity. To establish a statistical baseline, we compared these R ² values against a null distribution generated by randomly permuting the order of the 12 blocks 5000 times and recalculating the lagged regressions for each permutation. The P values from comparing the experimental and permuted R ² values at each of the 75 lags were corrected for multiple comparisons using the false discovery rate (FDR) method.

To perform an exploratory estimation of the potential magnitude of intra‐individual enhancement in 24 h retention, we used a regression‐based analysis. This analysis is justified by our finding that RER is a highly stable intra‐individual trait and is intended to quantify the potential effect of the observed deviation from this stable baseline. First, we established a linear regression model ($y=ax+b$) from the data of all 24 participants, relating RER during the day 1 task ($x$) to the measured 24 h retention ($y$). Using the parameters of this model, we then calculated two values for each participant: a hypothetical baseline retention ($y_{pre}$) by applying their pre‐intervention RER ($x_{pre}$) and the retention level associated with their task RER ($y_{task}$). Finally, we calculated the percentage change between these two estimated values as $(y_{task}-y_{pre})/y_{pre}\times 100(\%)$. This value provides an estimation, not a direct proof, of the intra‐individual change in retention that corresponds to the observed RER shift. This approach relies on the assumption (as discussed further in the Discussion) that the cross‐sectional relationship can serve as a plausible model for intra‐individual changes.

We performed one‐/paired‐sample or two‐sample two‐tailed t tests to compare the mean variables and performed Pearson's correlation when testing the covering variables. P < 0.05 was considered statistically significant.

Results

Sensorimotor learning of dynamic perturbation

Experiment 1 focused on the dynamics of sensorimotor adaptation itself, aiming to distinguish potential short‐ and long‐time constant processes within the learning period during the monitoring of O₂–CO₂ gas exchange and muscle activity. Participants were introduced to dynamic VMR perturbation (Fig. 1G ), which allowed for an adequate assessment of learning components with different time constants. Overall movement directions largely followed changes in the dynamic VMR (Fig. 3A ). Fourier transform analysis revealed large amplitudes only at trained frequencies ([2, 4, 8, 16 and 32] Hz), indicating appropriate adaption to the dynamic VMR (Fig. 3B ). Despite not reaching the perturbation levels of 10°, the mean amplitudes at the trained frequencies (6.97 ± 1.24°, mean ± SD) significantly surpassed those at the untrained frequencies between 2 and 32 Hz (0.564 ± 0.421°) (t ₁₇ = 20.0 and P = 3.01 × 10⁻¹³) (Fig. 3C ). These results indicate appropriate participant adjustment to the dynamic VMR.

Figure 3. Few effects of execution and learning of reaching movements on RER.

A, movement directions (purple line and area, mean ± SD across 18 participants) changed in response to the dynamic VMR perturbation. The black line represents the ideal learning required to perfectly counteract the perturbation. B, Fourier transform of the movement directions. The participant's performance (purple line) shows clear peaks at the five trained frequencies, tracking the ideal performance (black line). C, the average amplitude at the trained frequencies was significantly larger than at the untrained frequencies. D and E, self‐normalised muscle activity (D) was high at the beginning of the experiment and decreased rapidly over the first 300 trials. By contrast, RERs (E) were very stable over the course of the experiment. F and G, regression analysis with trial lags revealed that self‐normalised muscle activity cannot explain the changes in RER (F, permutation test, P > 0.05 for all lags with FDR correction). Furthermore, there was no change between task and break (G). H and I, no relationships with possible factors, peak velocity (H) and reaction time (I), were observed. J, no differences between RER in the first and last block. The violin plots illustrate the distribution of the data. Large circles and error bars indicate the mean ± SD, respectively. Small dots represent individual participant data. K and L, No obvious peaks at the trained frequency were found with the Fourier transformation for a series of RER (K), although the mean amplitude was statistically higher at the trained frequency (L). Shaded areas or error bars represent the SD.

Effects of reaching movements on RER

The RER is known to covary with the intensity of movement execution (Goedecke et al., 2000); therefore, changes in muscle activity could potentially confound the relationship between the RER and sensorimotor learning. Thus, we investigated how sensorimotor learning altered muscle activity and whether these changes affected the RER. EMG data from representative muscles including the pectoralis major, posterior deltoid, biceps brachii and triceps brachii were recorded to observe movement execution intensity effects. Muscle activity normalised by the trial average was high at the experiment's outset but quickly decreased, stabilizing at around one‐third of the experiment's duration (Fig. 3D ). This pattern, recognised as the relaxation of co‐contraction into a novel environment (Franklin et al., 2003), contrasted with the stability observed in the mean RER across participants, which did not fluctuate across trials (Fig. 3E ), suggesting a deviation from typical sensorimotor learning patterns (Fig. 3A ).

We examined whether muscle activity could account for RER variance across trials. Performing a linear regression with trial lags, we found that only 3.55% of the variance could be explained without lags, with minimal improvement observed across trials, even when accounting for trial lags (Fig. 3F ). This was in comparison with the permuted data from randomised trial blocks without significance noted for all trial lags after FDR correction (P > 0.05 in all datapoints, mean 0.279, 95% confidence interval = 0.222–0.608). A slight enhancement was detected with 36–46 trial lags (P < 0.05 in most datapoints without FDR correction, mean 0.0340, 95% confidence interval = 0.0171–0.0516). However, these trials elapsed within ∼193–243 s (trial duration was 5.27 ± 0.45 s), which is 1.5–2.0 times longer considering the time constant of the metabolic loop compared with the previous report showing a 120 s lag (Short & Sedlock, 1997).

To further explore the impact of movement execution, we compared the average RER during the reaching movement tasks with that during breaks (Fig. 3G ). We found a strong positive correlation between the two conditions (r = +0.908, P = 1.96 × 10⁻⁷), indicating that RER is a highly stable individual trait; participants with a higher RER during the task also had a higher RER during breaks. However, there was no statistically significant difference in the mean RER between the conditions (t ₁₇ = 0.287, P = 0.777), suggesting the low‐intensity reaching task did not cause a systemic shift in the group's average metabolic state. Additionally, RER values during tasks exhibited no correlation with mean peak velocity (r = −0.102 and P = 0.687) (Fig. 3H ) or reaction time (r = −0.261 and P = 0.297) (Fig. 3I ). Together, these results suggest minimal influence of movement execution on the RER as a result of the insufficient intensity of reaching movements.

Effects of sensorimotor learning on RER

Next, we investigated whether there was any indication of sensorimotor learning reflected in the RER (Fig. 3E ). However, we found that the RER remained constant during sensorimotor learning when comparing the RER between the early and late training blocks (t ₁₇ = 1.91 and P = 0.0734) (Fig. 3J ). Nonetheless, these values were highly consistent within participants (r = +0.652 and P = 0.00332).

Given that we used the dynamic VMR as the sum of the sinusoidal function, the amount of VMR to be learned in the early training block was equal to that in the late training block. Consequently, if the RER were to fluctuate with the amount of motor memory, it should remain stable during sensorimotor learning. In such a scenario, the peaks in the frequency domain of the RER are the same as those in the dynamic VMR. Alternatively, the RER might fluctuate with changes in motor memory, particularly following the derivatives of the sum of the sinusoidal functions, which would manifest as cosinusoidal function at the same frequencies. Thus, we would expect to observe peaks at [2, 4, 8, 16 and 32] Hz if sensorimotor learning affected the RER. Figure 3K illustrates the amplitudes of the frequency analysis, revealing no discernible peaks. However, we found a slight increase in the mean amplitudes at the trained frequencies compared to those at the untrained frequencies (t ₁₇ = 4.01 and P = 9.12 × 10⁻⁴) (Fig. 3L ).

The significant difference in the RER at the trained and untrained frequencies may not necessarily signal sensorimotor learning for the following three reasons. First, the absolute difference (0.00149 ± 0.00158) was smaller than the standard deviation of the within‐participant RER fluctuation (0.0446 ± 0.0127). Second, the RER may not behave as a random process similar to white noise, and it may exhibit characteristics of pink noise (or 1/f noise), where the power is inversely proportional to the frequency. Such patterns of pink noise are widely found in human physiology (Pritchard, 1992; Zarahn et al., 1997) and cognition (Gilden et al., 1995) and the tendency for the RER values to display larger amplitudes at low frequencies (Fig. 3K ) could imply a pattern of pink noise. If the RER is 1/f scaled, the mean amplitudes at the trained frequencies will be 2.41 times larger than those at the untrained frequencies. Third, the Fourier analysis could not identify polarity, meaning that we cannot ascertain whether the RER increased or decreased with sensorimotor learning. To further investigate whether the RER changed in tandem with sensorimotor learning, we analysed RER in Experiment 2 (see below).

Individual variability in RER reflects slow motor memory

We designed the dynamic VMR to rigorously estimate both fast and slow components of sensorimotor learning. We fit the two‐state model to each participant's individual data. On average, the model's estimate of overall motor output captured the observed behaviour well (R ² = 0.816 ± 0.172) (Fig. 4A ). As predicted, we observed peaks at the trained frequencies for both fast and slow components, as well as overall (Fig. 4B ). The mean amplitude at the trained frequencies was significantly higher for the overall data (t ₁₇ = 24.2 and P = 1.33 × 10⁻¹⁴) (Fig. 4C ). However, neither the experimental data nor the overall data correlated with the RER during the task (r = 0.0397 and P = 0.876 for experimental data, r = 0.159 and P = 0.530 for overall) (Fig. 4D ), suggesting that the RER does not reflect overall sensorimotor learning.

Figure 4. Individual RER reflects slow components of sensorimotor learning.

A, the two‐state model was fit to individual data from the 18 participants in Experiment 1. The figure shows the group‐averaged model fit. The model's estimate of overall performance (green line and area) and its decomposition into slow (blue) and fast (red) components are shown relative to the ideal learning curve (black), analogous to Fig. 3. B, Fourier transform of the model‐derived components. The model's estimate of overall performance (green) and its decomposition into slow (blue) and fast (red) components show clear peaks at the trained frequencies. The black line indicates the ideal performance, analogous to Fig. 3. C and D, the mean amplitudes of overall performance at the trained frequency were higher than those at the untrained frequency (C). However, the experimental and fitted overall performance did not correlate with the RER during the task (D). E and F, the amplitudes of the fast and slow components were increased and decreased, respectively, with the trained frequency (E). The mean amplitudes at the trained frequency were higher than those at the untrained frequency (F). G and H, The mean amplitudes of the fast (G) and slow (H) components were negatively and positively correlated with the RER during the task, respectively. I and J, for comparison with the model parameters, the RER during the task correlated only with the slow retention factor (I) and the slow learning factor (J). Shaded areas or error bars represent the SD.

Figure 4E illustrates that the peak amplitudes for the fast components gradually increased with frequency (slope > 0, t ₁₇ = 3.39 and P = 0.00351), whereas those for the slow components gradually decreased (slope < 0, t ₁₇ = 8.07 and P = 3.26 × 10⁻⁷). A statistically significant difference between fast and slow components was found at the highest frequency of 32 Hz (t ₁₇ = 7.45 and P = 9.42 × 10⁻⁷), with a marginal difference at the lowest frequency of 2 Hz (t ₁₇ = 1.32 and P = 0.205). The mean amplitudes at the trained frequency were higher than those at the untrained frequency (t ₁₇ = 11.0 and P = 3.85 × 10⁻⁹ for the fast component, t ₁₇ = 6.71 and P = 3.69 × 10⁻⁶ for the slow component) (Fig. 4F ). Interestingly, a statistically significant negative correlation was found between RER during the task and the mean amplitude of the model‐derived fast component (r = −0.581 and P = 0.0114) (Fig. 4G ), whereas a positive correlation was found for the model‐derived slow component (r = 0.668 and P = 0.00244) (Fig. 4H ). A similar, though weaker, positive correlation was observed between the slow component and RER measured before the experiment (r = 0.471 and P = 0.0483; for fast component, r = −0.337 and P = 0.171), suggesting some predictive value of baseline RER for this learning component.

To further explore this relationship, we examined correlations between RER and the model's fitting parameters (retention and learning factors) (Fig. 4I and J ). RER during the task showed a negative correlation with the retention factor (r = −0.526 and P = 0.0251) and a positive correlation with the learning factor (r = 0.627 and P = 0.00535) specifically for the slow component. However, no significant correlations were found for the fast component (r = −0.353 and P = 0.150 for retention factor, r = −0.135 and P = 0.593 for learning factor). These findings could suggest that individuals with a higher RER might exhibit faster initial adaptation but also potentially greater decay within this slow learning process. Given that RER did not appear to be modulated by task execution or learning itself, these results point towards an association between an individual's metabolic state (as indexed by RER) and the characteristics of their slow timescale sensorimotor learning process, as estimated by the two‐state model.

Increase in RER following glucose intake

Experiment 1 suggested a link between individual RER and the slow component identified during sensorimotor learning. To test whether this relationship extends to long‐term memory consolidation and can be modulated by metabolic intervention, Experiment 2 used a two‐day design with glucose administration (see Methods for details) (Fig. 2A ).

We found that the RER increased only after glucose intake. The RER before the intervention was not significantly different between the WG and Wa groups (t ₂₂ = 0.847 and P = 0.406) (Fig. 5A ). The RER immediately after the intervention did not increase but rather slightly decreased from the pre‐ to post‐intervention phases (t ₁₅ = 2.23 and P = 0.0412 for WG, t ₇ = 1.24 and P = 0.252 for Wa) and was not significantly different between these groups (t ₂₂ = 0.422 and P = 0.677). Notably, however, we found a significant increase in the RER from the pre‐intervention to task phases only in the WG group (t ₁₅ = 5.94 and P = 2.72 × 10⁻⁵ for WG, t ₇ = 1.15 and P = 0.288 for Wa), with a significant difference noted between these groups (t ₂₂ = 2.77 and P = 0.0111). The finding that glucose intake did not immediately increase the RER in the post‐intervention phase but rather slowly increased it during the subsequent training phase importantly suggested that glucose intake induced a gradual physiological response. The increase in the RER did not persist for the next day, and no significant differences were found between these groups in either the pre‐retention (t ₂₂ = 0.434, P = 0.669) or retention phases (t ₂₂ = 0.426 and P = 0.674). This result suggests that the RER can be successfully controlled by glucose intervention. Notably, there were no changes in the RER during the training phase in the Wa group, providing important evidence that sensorimotor learning does not affect the RER.

Figure 5. RER predicts glucose‐intake enhancement of long‐term motor memory.

A, No difference in the original RER was found in the pre‐intervention phase. Furthermore, the change in the RER (ΔRER) was not found in the post‐intervention phase but in the training phase. The effects did not persist on day 2; no differences were found in the pre‐retention and retention phases. The violin plots illustrate the distribution of the data. Large circles and error bars indicate the mean and standard deviation (SD), respectively. Small dots represent individual participant data. B,C. We observed the typical learning curve for VMR training on day 1 (left panel in b). On day 2, the amount of learning largely dropped, but we can still see significant retention (right panel in b). Notably, no differences were found between the groups (C). d. The correlation analysis showed a significant positive relationship between retention and intervention RER during training phase, but no relationships against RER in the other phases. Shaded areas or error bars represent the standard deviation (SD).

Association between individual RER responses to glucose intake and retention

We observed basic trends in the VMR learning curves. After short baseline periods, the hand movement directions gradually shifted towards ideal learning (Fig. 5B ). The resulting learning levels were 24.2 ± 2.19° for the WG group and 21.6 ± 7.72° for the Wa group, which were not significantly different (t ₂₂ = 1.30 and P = 0.208) (Fig. 5C ). The amount of retained adaptation on day 2 was obviously reduced from the end of the training period, but were still greater than the baseline at 4.89 ± 2.71° for the WG group but 3.00 ± 2.81° for the Wa group (Fig. 5C ). Consistent with high inter‐individual variability in baseline RER (Fig. 5A ), which probably obscured group‐level effects, the retention amounts on day 2 were not significantly different between the WG and Wa groups (t ₂₂ = 1.59 and P = 0.125).

To identify the metabolic phase most critical for memory consolidation, we examined the relationship between 24‐h retention on day 2 and the RER measured across five distinct phases (Fig. 5D ). Of these, a significant positive correlation was found only with the RER measured during the day 1 training phase (r = 0.552, P = 0.00514). By contrast, no significant correlations were observed between retention and RER during any of the four non‐learning phases: the pre‐intervention rest phase on day 1 (r = 0.161, P = 0.452), the post‐intervention rest phase on day 1 (r = 0.270, P = 0.201), the pre‐retention rest phase on day 2 (r = 0.166, P = 0.437) or during the retention task itself on day 2 (r = 0.307, P = 0.145). These findings strongly suggest that the systemic metabolic state specifically during the initial learning period is critically linked to subsequent long‐term memory consolidation.

Further supporting this, we applied the same two‐state model to the behavioral data from Experiment 2. This model‐based analysis revealed that the relationship between RER and retention was driven specifically by the slow, but not fast, component of memory (Fig. 6). The model captured the learning and retention curves well (R ² = 0.789 ± 0.150) (Fig. 6A and B ). Although the asymptote of learning on day 1 showed a significant positive correlation only with the slow component (r = 0.423, P = 0.0397), the model‐derived retention on day 2 revealed significant positive relationships for both the overall and slow components after robust regression analysis (P = 0.0190 and P = 0.0122, respectively), but not for the fast component (P = 0.589) (Fig. 6C and D ). These results provide model‐based evidence that a higher RER during learning is associated with better long‐term retention, a relationship primarily driven by the slow memory process.

Figure 6. Two‐state model analysis of Experiment 2 reveals a specific link between RER and the slow memory component.

A and B, the two‐state model was fit to the behavioral data from the WG group (A) and Wa group (B). The model's overall performance (green), slow component (blue) and fast component (red) are shown relative to the experimental mean data and the ideal learning curve. Shaded areas or error bars represent the SD. C, scatter plots showing the relationship between RER during the day 1 training phase and the model‐derived asymptote of learning on day 1. D, scatter plots showing the relationship between RER during the day 1 training phase and the model‐derived retention on day 2.

To estimate the potential magnitude of this intra‐individual effect, we leveraged the regression parameters (slope and offset) from the correlation between RER during the day 1 and 24‐h retention (Fig. 7). We calculated the predicted retention change corresponding to the observed RER increase from the pre‐intervention phase to the task phase for each participant (see Methods). This analysis is predicated on our key finding that RER is an exceptionally stable intra‐individual trait. This stability is supported by strong positive correlations, both within a single experimental session lasting several hours in Experiment 1 (r = 0.652, P = 0.00332) (Fig. 7A ) and across a 24 h period in Experiment 2 (r = +0.683, P = 2.32 × 10⁻⁴) (Fig. 7B ). This regression‐based estimation suggested that the glucose intake was associated with an estimated 21.4 ± 16.3% enhancement in retention for the WG group (t ₁₅ = 5.25 and P = 9.80 × 10⁻⁵) (Fig. 7C ), based on the projection of RER changes onto the observed regression line. By contrast, no such effect was estimated for the Wa group (−5.11 ± 17.4%, t ₇ = 0.829 and P = 0.434; difference between estimates: t22 = 3.67 and P = 0.0134) (Fig. 7C ). Taken together, although lacking a direct group effect, these analyses suggest an association between glucose‐induced RER elevation during training and enhanced long‐term motor memory retention at the individual level.

Figure 7. RER is a highly stable intra‐individual trait, supporting the estimation of glucose‐induced memory enhancement.

A, correlation between the RER measured during the first and last training blocks in Experiment 1 (n = 18). B, correlation between the baseline RER measured on day 1 (pre‐intervention) and day 2 (pre‐retention) in Experiment 2 (n = 24). The solid diagonal line represents the unity line. C, changes in the RER from the pre‐intervention phase to the training phase estimated a 21.4 ± 16.3% increase in retention with the glucose intervention. Error bars represent the SD.

Discussion

Respiration plays a crucial role in energy production and supports neural activity and plasticity (Lee et al., 2013; Oakes et al., 2004; Sokoloff, 1977; zur Nedden et al., 2011), which are essential for sensorimotor learning. Breathing gas composition provides detailed information about whole‐body metabolism including movement intensity (Abram et al., 2022; Huang et al., 2012; Selinger et al., 2015; Shadmehr et al., 2016); however, it is largely unknown whether they are also a prominent index of brain metabolism, that is, whether breathing gas composition predicts sensorimotor learning. According to the two‐state sensorimotor learning model (Smith et al., 2006), the faster component reflects short‐term memory, whereas the slower component reflects long‐term memory. We investigated whether the metabolic state estimated by the RER that shows carbohydrate or fat preferences reflects short‐term or long‐term memory. The results showed that the RER was not influenced by the execution or learning of reaching movements (Fig. 3). Instead, the RER predicted the amount of estimated slow rather than fast motor memory (Fig. 4). We also investigated whether metabolic interventions could improve long‐term memory, that is, next‐day retention (Figs. 5 and 6). To accomplish this, we provided 200 kcal of glucose, which immediately shifted the participants towards more carbohydrate utilisation. Surprisingly, this intervention not only increased the RER, but also improved next‐day retention by up to 21% (Fig. 7). Experiment 1 revealed a correlation between RER and the slow component of learning dynamics. Experiment 2 further showed that RER changes during learning were associated with the strength of memory retention 24 h later. Together, with two separate cohorts of human participants, these findings showed that the metabolic state estimated by the RER can serve as an indicator of long‐term motor memory, particularly at the intra‐individual level, implying a potential avenue for enhancing human motor performance.

Methodological considerations

The metabolic state estimated by the RER varied widely from 0.740 to 0.978 among individuals even at baseline, which is very similar range to the previous findings; for example, from 0.718 to 0.927 in Goedecke et al. (2000). Previous studies have shown that during low‐intensity activities such as unloaded or lightly loaded pedaling, comprising activities probably exerting more load than the reaching movements in our study, RER does not significantly deviate from baseline levels (Wasserman et al., 1973). This is consistent with our findings (Fig. 3E ). We also found no explanation of muscle activity on RER (Fig. 3F ) as well as unchanged RER between at‐task vs. at‐break (Fig. 3G ). Furthermore, the RER is highly stable even during sensorimotor learning (Figs. 3E and 5A , grey) and no correlation against sensorimotor learning amplitudes (Fig. 4D ). The results suggest that the execution and learning of reaching movement do not affect RER accordingly.

Although RER itself remained stable during execution and learning in Experiment 1, it can be altered by factors such as circadian rhythm (Zitting et al., 2018), dietary factors and genetic predispositions (Toubro et al., 1998). Because glucose intake is a simple method known to increase RER by promoting carbohydrate utilisation (Tellez et al., 2013), we used it here as a tool to test the hypothesis that an intervention increasing RER during learning would consequently improve long‐term motor memory. When comparing the presence and absence of glucose intake aimed to directly examine the intervention's effect on long‐term memory, the high baseline RER variability (SD across participants: 0.0461) relative to the average RER increase post‐glucose (0.0466 ± 0.0314) probably contributed to the lack of significant group differences in retention (Fig. 5C ), obscuring potential effects at the group level. This inherent variability underscores the challenge in detecting group‐level effects of interventions such as glucose intake, necessitating careful consideration of intra‐individual analyses at the same time as acknowledging the limitations associated with small sample sizes. The long‐term motor memory would be more enhanced by maximizing the physiological effects of glucose intake, such as the glucose dose or glucose digesting time. Notably, a literature review showed enhancement of memory with glucose intake is limited in the other memory types, such as verbal word memory (Smith et al., 2011). Our results underscore a new type of memory retention improvement at the intra‐individual level and it needs to turn the tide for investigation of physiological basis for sensorimotor learning.

Further methodological considerations include the lack of a non‐caloric, taste‐matched control group for the glucose intervention, which prevents us from definitively isolating the metabolic effects from potential influences of sweetness on learning. For example, glucose intake may induce a positive mood and reduce hunger, perhaps leading to an upregulation of motivation and attention. These cognitive improvements would affect the fast component, reflecting the use of an explicit strategy (McDougle et al., 2015), rather than the slow component. Thus, if glucose intake elicited only cognitive improvements, the slow component would not change. Our findings that the glucose intake affects long‐term motor memory suggest that cognitive improvements from the glucose intake are not the main driver of its enhancement of long‐term motor memory.

It is also important to explicitly state the limitations of our regression‐based estimation of the intra‐individual retention enhancement. This analysis does not constitute definitive proof of causality, as is inherent to a between‐subject design. It relies on the assumption that the cross‐sectional relationship between RER and retention observed across participants can serve as a plausible model for intra‐individual changes. We argue this assumption is well‐justified by our findings that RER is an exceptionally stable intra‐individual trait across various conditions (Figs. 3G and 7A and B ). Nevertheless, to definitively establish a causal link between the glucose‐induced RER increase and memory enhancement, a future study employing a within‐subject cross‐over design would be required.

The cortical metabolism and time‐scale dynamics in sensorimotor learning

A key finding of our study is the specific relationship between RER and the slow (but not fast) component of motor memory. This specificity may be understood from converging perspectives involving the energetic costs of structural plasticity, the matching of temporal scales and the distinct metabolic profiles of the underlying neural substrates.

First, the slow process is considered to underpin long‐term memory, which relies on energy‐intensive biological processes such as protein synthesis, gene expression and structural remodelling of synapses (Kandel et al., 2014; Luft et al., 2004). A high RER indicates a systemic metabolic state primed for carbohydrate utilisation, ensuring a rich supply of glucose and ATP to support these costly, long‐term structural changes. The fast process, conversely, is associated with more transient neural activity and may be less constrained by the baseline systemic energy state.

Second, there is a fundamental alignment in the temporal characteristics of RER and the slow learning process. RER is a slow‐moving, tonic physiological state that remains highly stable within individuals. The slow learning process operates over a similarly long timescale, gradually integrating information across many trials. It is plausible that a stable physiological context such as RER would selectively influence a learning mechanism that shares its slow, integrative nature, whereas the rapid, trial‐by‐trial dynamics of the fast process are mismatched with this baseline state.

Finally, this specificity may relate to the different neural circuits underlying each process. Fast and slow learning processes are considered to rely on different brain regions, with the parietal cortex often implicated in the fast process and the cerebellum in the slow process (Kim et al., 2015). Notably, these regions exhibit different metabolic profiles; aerobic glycolysis, for example, is high in the parietal cortex but well below the brain's average in the cerebellum (Vaishnavi et al., 2010). We speculate that the metabolic machinery supporting the slow, cerebellum‐dependent process may be particularly sensitive to systemic glucose availability (as indexed by RER), whereas the mechanisms supporting the fast process in the parietal cortex might be metabolically distinct and thus less coupled to the baseline RER. Future work should aim to integrate these metabolic and neural network models to develop a more comprehensive picture of motor memory formation.

Further insight into the nature of this relationship comes from an interesting, although seemingly counter‐intuitive, finding regarding the specific parameters of the slow process. Although a higher RER was associated with a larger overall contribution of the slow component to behaviour (Fig. 4H ) and a higher learning rate (${\beta }_s$​), it was also associated with a lower retention factor (${\alpha }_s$​), indicating faster forgetting (Fig. 4I and J ). We propose that this is not a contradiction, but rather reflects a higher RER being associated with a more flexible or labile state of the slow motor memory. In our dynamic sinusoidal VMR task, the goal is not to rigidly retain a single motor command but to continuously track the perturbation. In this context, faster forgetting (a lower ${\alpha }_s$​) can be adaptive because it allows the system to more quickly unlearn an adaptation as the perturbation returns towards zero. Therefore, a physiological state associated with a high RER appears to promote a slow process that is both highly sensitive to new information (high ${\beta }_s$) and quick to discard old, irrelevant information (low ${\alpha }_s$). This combination creates a system optimally tuned for flexibility, leading to more effective tracking and a larger overall contribution of the slow process to behaviour.

The RER would indicate physiological idling state for sensorimotor learning

Although the brain accounts for substantial energy expenditure, RER measured via breath reflects the net metabolic state of the entire body. Therefore, the observed correlations probably represent an indirect relationship rather than RER directly tracking task‐specific neural activity. The link between systemic RER and motor memory remains hypothetical, potentially reflecting how overall metabolic context influences the conditions for memory consolidation. Although the brain has a high metabolic rate relative to its mass, its contribution to fluctuations in whole‐body RER measured via breath may be limited. How does RER reflect long‐term motor memory?

It should be noted that the glucose intake did not immediately increase the RER in the post‐intervention phase, but rather slowly increased it during the subsequent training phase (Fig. 5A ). This temporal pattern suggests a gradual physiological response that unfolds over time, that is, insulin secretion. Insulin secretion is not an instantaneous response but occurs gradually in response to increasing levels of glucose in the bloodstream. Insulin secretion is usually accompanied by an increase in RER (Dickson et al., 1924), and presumably regulates glucose levels in the body and neuronal and glial cells (Milstein & Ferris, 2021; Uemura & Greenlee, 2006) for future energy production. Thus, this increase in cerebral glucose levels would meet the cortical energy needs, especially for learning and memory.

Interestingly, insulin modulates memory formation, as evidenced by studies on Alzheimer's disease, a model of impaired declarative memory function. Alzheimer's disease is accompanied by disrupted glucose regulation and utilisation (Kyrtata et al., 2021); however, surprisingly, insulin injection improves memory function in Alzheimer's patients (Craft et al., 1996; Reger et al., 2008). Consistent results have been found in healthy subjects, indicating that the intranasal administration of insulin upregulates memory function (Benedict et al., 2004, 2007). Furthermore, insulin resistance in healthy obese individuals impairs their ability to learn; however, the administration of liraglutide, which temporarily stimulates insulin release, can restore this ability (Hanssen et al., 2023). Although insulin is known to modulate memory, establishing a causal chain from glucose intake to RER changes, insulin response and specific motor memory enhancement requires further investigation with more direct physiological measures.

Competing interests

The authors declare that they have no competing interests.

Author contributions

T.H., N.N., S.W. and A.M. designed the experiments. T.H. and N.N. performed the experiments. T.H. analysed the data. T.H., N.N., S.W. and A.M. wrote the paper.

Funding

This work was supported by KAKENHI grants from JSPS (JP23H03296) and JST/PRESTO (JPMJPR23S8) to TH. This work was also partly supported by JST Moonshot R&D Grant Number JPMJMS2239.

Supporting information

Peer Review History

Biography

Takuji Hayashi is an Assistant Professor at the Graduate School of Education, The University of Tokyo. He specialises in motor learning, motor control and computational modelling, with his primary research areas in neural control of movement. After receiving his PhD in Education from The University of Tokyo in 2017, he served as a postdoctoral fellow at the Institute of Engineering, Tokyo University of Agriculture and Technology. He then worked as a postdoctoral fellow at the School of Engineering and Applied Sciences at Harvard University. He has held his current position since 2022.

Data availability statement

We have prepared an online repository on Zenodo (https://zenodo.org/) that contains the data and codes from the present study.