Fluctuating internal states mediate neural-behavioral covariations in V1

Abstract

Our perception of the world depends on a complex interplay between external sensory inputs and our internal states. How and where in the brain the interactions between internal states and sensory inputs are implemented remain open questions. To study the neural basis of these interactions, we used whole-cell recording to measure membrane potential (Vm) of single V1 neurons in macaque monkeys performing a reaction-time detection task. We find that most V1 neurons gradually depolarize in preparation for target onset. Remarkably, trial-to-trial variations in this preparatory buildup are correlated with the monkey’s reaction times, and these covariations are similar when the animal selects a target in either hemifield. This finding implies that variation in the preparatory buildup are shared between hemispheres and are unrelated to a competitive spatial attention mechanism. Further, we find that while the gradual buildup is largely uncorrelated with the animals’ choices, transient fluctuations in Vm around target-evoked response onset time are correlated with choice. The dependence of these Vm-to-choice covariations on the visual stimulus is consistent with a multiplicative gain, reflecting a variable internal state that interacts with variable sensory signals. Overall, our results reveal surprising covariations between the membrane potential of single V1 neurons and behavior, and that these covariations are mediated by mixture of additive and multiplicative task-related modulations of incoming sensory information.

Visual perception represents the outcome of a complex interplay between sensory processing and fluctuating internal states. Where in the primate brain the interactions between internal states and visual inputs occur, how these interactions are implemented, and the extent to which these fluctuating internal states contribute to previously reported neural-behavioral covariations in early sensory cortex remain open and important questions^1–7. While the effects of internal mechanisms such as attention have been reported previously, these effects tend to be weak or absent in spiking activity of single neurons in early sensory cortex in NHPs, and increase in magnitude as one progresses to higher sensory cortical areas^8–13. Here, we focus on the hypothesis that mechanisms related to fluctuating internal states modulate incoming sensory information in V1, the earliest stage of visual cortical processing, and contribute to moment-by-moment neural-behavioral covariations.

Neural signals related to variable internal state may be prevalent in early sensory cortex but difficult to observe in the suprathreshold spiking activity of single neurons. However, such modulations may be evident in subthreshold membrane potentials, which reflect the pooled inputs from thousands of pre-synaptic neurons¹⁴. In such pooling, the independent neural variability tends to be averaged out, while the weak shared variability becomes dominant^15,16. Similarly, because visual representations in primate V1 are widely distributed^7,17, if some of the fluctuations in subthreshold signals are shared between V1 neurons, these shared fluctuations will be amplified by downstream circuits that pool inputs from large number of V1 neurons. Such pooling could therefore lead to robust covariations between the membrane potential of single V1 neurons and behavior.

To study the impact of internal states on membrane potential (Vm) of single V1 neurons and to measure the relationship between subthreshold Vm fluctuations and behavior, we performed whole-cell patch recording from V1 while three macaque monkeys performed a reaction-time visual detection task in which a low contrast, oriented target appeared in one of two possible locations (Fig. 1a, top left). One target location overlapped the receptive field (RF) of the recorded neuron (“target in” or “Tin”) while the other location was in the opposite hemifield (“target out” or “Tout”). Target contrasts spanned the monkeys’ detection threshold and included zero contrast, where the correct target location was assigned randomly. The monkey responded by shifting gaze to the target location as soon as it was detected, giving us access to behavioral variations in both choice and reaction time. As expected, accuracy increased with target contrast while reaction times decreased with target contrast (Fig. 1b).

Figure 1: Membrane potential buildup in anticipation of target onset.

a, Monkeys were trained to fixate and saccade to a Gabor target of variable contrast as soon as it was detected (top left). They were temporally cued 300 ms before target onset by dimming of the fixation point. The target was placed within the receptive field (RF) of the recorded neuron (blue dashed circle) or in the opposite hemifield. On a fraction of the trials, no dimming occurred, no target appeared, and the monkey was rewarded for maintaining fixation (“blank” trials, bottom right). All trial types were randomly interleaved. b, Three monkeys were trained to accurately (top) and quickly (bottom) detect the target. c, Average membrane potentials from an example recorded cell (average correct reaction times (RTs) noted by arrows). Target within the receptive field (“Tin”, right column); target in the opposite hemifield (“Tout”, left column). Contrast and the number of trials (n) are noted in the upper left corner. For some of the cued trials, no target appeared (0%, bottom row, left) and for other trials, the monkey was never cued (“blank”, bottom row, right). d, Examples of the average change in membrane potential (Vm) over time for two neurons from cue onset (cue) to stimulus onset (stim) for all cued trials. The dashed line shows the average response for blank trials without a cue. e, Same as d averaged across all cells (N). The black data points are the differences in membrane potentials computed in 50-ms bins between cued and no cue trials averaged over all cells. f, The distribution of the potential differences for all cells for the first minus last bin (arrows) in e (area of circle represents number of trials). Purple data points are the example cells in d. All shaded regions and error bars for all panels are the standard error of the mean and asterisks note significant data points (P < 0.05).

To assess the impact of task preparation on Vm of V1 neurons, we provided the monkeys with a temporal cue – dimming of the fixation point – 300 ms before target onset (Fig. 1a). If signals related to task preparation operate in V1, we would expect changes in Vm to occur following the temporal cue and before target onset. To identify such preparatory signals, we randomly interleaved with the detection trials reference “blank” trials in which no temporal cue was provided, no visual target appeared and no behavior, beyond fixation, was required (Fig. 1a, bottom right). By comparing Vm in detection and blank trials, we could identify the impact of task preparation on subthreshold signals in V1.

We discovered a robust preparatory depolarization signal that starts to build up following the temporal cue and is absent in blank trials. Fig. 1c shows the membrane potential of an example V1 neuron during the task. There is a clear visual-evoked response following target onset in Tin trials, and this response decreases with decreasing target contrast and is absent at zero contrast, Tout and blank trials. In addition to the visual response, there is a gradual depolarization which is present in detection trials and absent blank trials (e.g., bottom left panel in 1c). The Vm depolarization for the detection (cue) trials relative to the blank (no cue) trials is evident in an additional example cell (Fig. 1d) and in the summary across 31 cells (Fig. 1e–f). This buildup is largely independent of the animal’s choice (Fig. S1). Overall, these results reveal a robust preparatory Vm buildup, reflecting the impact of an internal state at the level of single neurons in macaque V1. Because we see this effect in most recorded neurons irrespective of their stimulus preferences, this effect must be highly prevalent across V1.

If this task-related depolarization reflects the animal’s preparedness or attention, we would expect that on trials with larger buildup the animal’s reaction time would be shorter than in trials with smaller buildup. In other words, we would expect a negative correlation between Vm buildup and reaction time (Fig. 2a, top). To test for this possibility, we first focus on zero contrast trials in which the monkey chose Tin (“choice in” trials). Fig. 2a (bottom) shows Vm traces in two example trials. Reaction time is shorter in the top trial in which Vm was more depolarized at target onset time than in the bottom trial in which Vm was less depolarized at this time. The summary of all zero-contrast choice-in trials from two example neurons shows a weak negative correlation between Vm buildup and reaction time (Fig. 2b). The negative correlation between Vm buildup and RT in zero contrast choice-in trials holds in the summary across all of our recorded neurons (Fig. 2c) and is robust across the exact pre-stimulus temporal interval (Fig. S2), demonstrating a surprising behavioral correlate of fluctuations in preparatory Vm buildup in single V1 neurons. Because reaction times in our task are likely to depend on the activity of a large pool of V1 neurons, this result suggests that variations in Vm buildup are highly correlated across V1 neurons, and could therefore contribute to behavioral variations in reaction times.

Figure 2: Larger buildups in membrane potential predict faster reaction times.

a, If the post-cue buildup in membrane potential of V1 cells reflects the animal’s preparedness and if this preparedness varies from trial to trial, then trials with large buildup should generally have faster reaction times (RTs). Top – schematic. Middle and bottom - two example trials with a Vm buildup consistent with this prediction. b, Two example cells with this inverse relationship between RT and Vm buildup for all 0% contrast Choice In trials. c, The entire distribution of correlations for each cell for no target (0%) and Choice In (area of circle represents number of trials). Purple data points are the example cells in b. d, The predicted correlation between RT and Vm buildup for Choice Out trials (red line) depends on the relationship between the buildup in the two hemispheres (competitive interhemispheric buildup – left; correlated interhemispheric buildup – middle; independent interhemispheric buildup – right). e, Average RT-Vm correlations across cells for 0%, low contrast (≤5%), and high contrast (12.5–100%) targets divided into choice out (red) and choice in (blue) trials. Error bars for all panels are the standard error of the mean and asterisks note significant data points (P < 0.05).

The preparatory Vm buildup in single V1 neurons could reflect general alertness or distributed attention. Alternatively, it could reflect a competitive focal attention mechanism that sequentially alternates between the representations of the two target locations in the two hemispheres. These competing hypotheses make testable predictions with respect to the relationship between the preparatory buildup and reaction times in choice-out trials. If the buildup reflects a competitive selective attention signal, we would expect a negative correlation between the buildup in V1 in the two hemispheres, as stronger attention to one target location in one hemisphere implies weaker attention in the other hemisphere. In this case, a larger Vm buildup in the recorded hemisphere (“Hemi-in”) will imply a weaker buildup in the opposite hemisphere (“Hemi-out”), and therefore, a slower reaction time in choice-out trials (Fig. 2d, left). In other words, we would predict a positive correlation between buildup in Vm in Hemi-in and reaction times in choice-out trials. Alternatively, if the preparatory buildup reflects general alertness or distributed attention¹⁴, Vm buildup would be shared between the two hemispheres, and we would predict a negative correlation between the Vm buildup in Hemi-in and reaction times in both choice-in and choice-out trials (Fig. 2d, middle). Finally, if the buildups were independent in the two hemispheres, we would predict no relation between the Vm buildup in Hemi-in and reaction times in choice-out trials (Fig. 2d, right). Irrespective of their sign, we expect the absolute value of the correlations between Vm buildup and reaction times to decrease with increasing target contrast, because at higher target contrasts reaction times are likely to be dominated by variability in target evoked responses or in motor commands rather than by variability in the preparatory buildup in V1.

Our Vm buildup to reaction times correlations results are consistent with the general alertness or distributed attention hypothesis and inconsistent with the competitive attention and independent hemisphere hypotheses (Fig. 2e). The correlations between Vm buildup and reaction times are negative for both choice-in and choice-out, and the absolute value of these correlations decrease with increasing target contrast. Overall, these results reveal that most single V1 neurons receive a common preparatory signal that is largely shared between the representations of the two target locations in the two hemispheres, and that fluctuations in this preparatory signal are predictive of the speed with which the subject reports the presence of the target.

Our results so far reveal a task-related preparatory buildup in Vm of single V1 neurons that is correlated with the monkeys’ reaction times. Could fluctuations in this buildup also predict trial-to-trial variability in the monkeys’ choices in zero- and low-contrast trials? Covariations between neural responses and choice are typically quantified using choice-probability (CP) metric, which measures the probability with which an ideal observer could predict choice solely based on the pre-choice neural responses to identical stimuli. CP of 0.5 implies no predictive power while CP of 1 implies perfect predictability. If fluctuations in the preparatory buildup are related to variability in choice, Vm prior to target onset time would already be predictive of choice. However, when we compare the average Vm traces in choice-in and choice-out in zero contrast trials, we find that Vm levels at the time of target onset are uncorrelated with choice in two example cells (Fig. 3a; CP values ~0.5) and in our entire recorded population (Fig. 3b). This result implies that the slow preparatory buildup in Vm, which is predictive of reaction time (Fig. 2), is discounted by the downstream circuits that form the perceptual decision in our detection task. Because our Vm-to-reaction time correlation results imply that the buildup is positively correlated between the two hemispheres (Fig. 2e), some of the discounting of the effect of the buildup on choice could be achieved by a downstream mechanism that compares, or subtracts, the responses in the two hemispheres. In addition, downstream decision circuits may discount slow and gradual variations in neural activity and form the perceptual decision based on transient V1 responses that occur around the expected target response onset time¹⁸. Irrespective of the mechanism, our results reveal a surprising decoupling between the effect of the preparatory buildup on reaction time and choice.

Figure 3: Membrane potential predicts choice after target onset.

a, Membrane potential traces of two example cells in 0% target trials separated by choice. The responses begin to differ and are predictive of choice (Choice Probability (CP) ~ 0.6) more than 100 ms after stimulus onset and about 100 ms before saccade onset (right). b, The average membrane potential (in 50-ms bins) across all cells. Both Choice In and Choice Out responses slowly increase together before and after stimulus onset and only diverge 100 ms before saccade onset. c, Choice probability across all cells is only significantly above 0.5 (noted by asterisks, p < 0.05) 50 ms before the choice (saccade onset). d, The latencies of target-evoked responses increase with decreasing contrast, which corresponds to increase in average reaction times (RT) with decreasing contrast. Response to low contrast targets (3–5%) appears at almost 200 ms after stimulus onset. Dashed horizontal line (mean buildup at stimulus onset from Fig. 1f) is larger than the response evoked by near-threshold targets. Shaded regions and error bars for all panels are the standard error of the mean.

To examine the possibility that fluctuations in Vm of single V1 neurons after target onset time predict choice, we compared the average Vm traces in 0% contrast choice-in and choice-out trials time-locked to the monkeys’ reaction time. We find that shortly before the reaction time, Vm tends to be higher in choice-in than in choice-out trials in the two example cells (Fig. 3a, right) and in the summary data (Fig. 3b and 3c, right). Therefore, choice-related signals are present in single V1 neurons in 0% contrast trials, but only shortly before the decision is made.

The observed choice-related Vm signals in 0% contrast trials could reflect random spontaneous transient Vm fluctuations that are correlated across V1 neurons with nearby receptive fields, as observed in fixating monkey V1¹⁹. When such random fluctuations happen to occur in neurons that represent Tin location around the expected time of target onset, downstream decision circuits could confuse these spontaneous transients with a target-evoked response, which would lead to a choice-in perceptual decision. If such random fluctuations were additive with a deterministic target-evoked V1 response (Fig. 4a, top), we would expect positive choice-related Vm signals also in low-contrast Tin and Tout trials, with CP that is stronger than in 0% contrast trials (Fig. 4b, left; see supplementary materials for model derivations and predictions).

Figure 4: Vm choice predictability consistent with a variable gain hypothesis.

a, In one model (Additive, top), stimulus and buildup (‘internal state) are combined by addition before being compared across hemispheres. Choice probability is estimated after adding downstream noise. In the second model (Multiplicative, bottom), stimulus variability is multiplied by internal state variability before being compared across hemispheres. In both cases, internal variability is positively correlated across hemispheres. b, The additive model predicts choice probability above 0.5 for 0% contrast and for low-contrast Tin and Tout trials. The multiplicative model predicts that choice probability decreases below 0.5 for Tout trials. c, The presence of low contrast targets changed the predictive properties of Vm and this choice-related activity coincided with the timing of the weak target-evoked response in Fig. 3d. Top row: there was a larger and earlier predictive difference in Vm between Choice In (correct detection) and Choice Out (error trials) in low contrast Tin compared to 0% contrast target (Fig. 3b). Bottom row: there was a larger response for Choice Out (correct detection) versus Choice In (error trials) in low contrast Tout, which is the opposite relationship compared to when there was no target (Fig. 3b). d, Vm choice probability across all cells is strong and positive in Tin trials, weaker in zero contrast trials, and negative in Tout trials, matching the prediction of the multiplicative model. e, The entire distribution of CPs for all cells for the three time bins noted by the arrows in d (area of circle represents number of trials). Purple data points are the example cells in Fig. 3a. All error bars for all panels are the standard error of the mean and asterisks note significant data points (P < 0.05).

Alternatively, we consider a model in which fluctuations in Vm reflect a variable gain that acts multiplicatively on V1 signals (Fig. 4a, bottom). In this case, our model also predicts larger choice-related signals in low-contrast Tin trials than in the zero contrast trials, because in low-contrast Tin trials, the variable gain would be multiplied by a visual evoked response that is absent in zero contrast trials (Fig. 4b, right). Under the variable gain hypothesis, if the gain fluctuations are shared between the two hemispheres, as our Vm-to-reaction time covariations imply for the preparatory buildup (Fig. 2e), a counterintuitive negative choice-related signal emerges for low contrast Tout trials (Fig. 4b, right; see supplementary materials). To see why, consider two scenarios. In Tout trials with high Vm, the gain in both hemispheres is high, the target-evoked responses are large, and the monkey is more likely to make the correct choice (i.e., choice-out). In Tout trials with low Vm and low gain, the target-evoked responses are low, and the monkey is more likely to make an incorrect decision (i.e., choice-in). Therefore, in this case, high Vm in Tout trials is correlated with decreased probability of choice-in and a paradoxical negative CP (Fig. 4b, right; see supplementary materials).

Surprisingly, our results are consistent with the variable gain hypothesis. The choice-related signals are strong and positive in Tin trials, weaker and positive in zero contrast trials, and negative in Tout trials (Fig. 4c–e). The weak but positive CPs in zero contrast trials and the larger absolute value of the CP in Tin than in Tout suggest that our results reflect a mixture of a weak additive choice-related signal and a larger gain effect on choice related activity in V1.

The choice related Vm signals are only observed shortly before the monkey’s reaction time. Could this signal reflect a post-decisional non-causal top-down signal rather than a pre-decisional signal that may be causally related to the choice? Two lines of evidence point against the post-decisional possibility. First, the monkeys performed a reaction time task, and the combination of the long-latency and sluggish V1 responses to low contrast targets (refs^20–22; Fig. 3d) and the fast reaction times (Fig. 1b) reduces the chance of such post-decisional signals appearing in V1 prior to the initiation of the saccade. Second, if the choice-related Vm signals reflect a post-decisional feedback to V1, we would expect a choice-related signal that is independent of the visual stimulus, which is inconsistent with our findings (Fig. 4c–e). In fact, we find such stimulus independent choice-related signals after the saccade (Fig. 4e, right). Therefore, the choice-related Vm signals are unlikely to be a byproduct of a post-decisional top-down signal.

In conclusion, our results reveal robust task-related preparatory signals in membrane potential of single V1 neurons when monkeys prepare to detect a low-contrast visual target. We show that trial-to-trial fluctuations in this buildup are correlated with variability in the monkeys’ reaction times, so that larger buildup is associated with shorter reactions times. This negative correlation between Vm buildup and reaction time holds when the monkey makes saccades to the either hemifield, indicating that the trial-to-trial fluctuations in the buildup are positively correlated between the two hemispheres, thus providing evidence against a competitive interaction between the hemispheres in our task. Even though the preparatory Vm buildup is correlated with reaction time, it is not predictive of the monkey’s choice, suggesting that downstream decision circuits largely discount this slow buildup when forming decisions based on V1 signals. Instead, we find a robust choice-related signal in Vm of single V1 neurons around the expected time of target-response onset. Our results indicate that these choice-related signals are unlikely to be a byproduct of a post-decisional top-down signal. Instead, the choice-related Vm signal reflects a mixture of multiplicative and additive mechanisms that operate in V1 during the critical period between stimulus onset and the formation of the perceptual decision. An important goal for future studies is to identify the source of the multiplicative and additive signals that mediate the effects of internal states on sensory processing and contribute to choice-related variability in V1.

Methods

Surgery

Three macaque monkeys were used in this study. Each animal underwent two surgical procedures. First, we implanted a head-restraining device. After a recovery period, the animal was extensively trained on a visual detection task (Fig. 1a) in which water or juice were used as positive rewards. Once the animal reached a stable level of performance on the task, a custom-designed recording chambers was positioned over the skull above V1, and a second surgery was performed to prepare the monkey for whole-cell recordings. In this surgery, a cranial window was opened, and the dura was resected and replaced by a transparent artificial dura. Within several weeks after this surgical procedure, the animal’s dura healed and formed a tight seal around a silicone ring extending out from the artificial dura, leaving a central region clear for recordings. The artificial dura was temporarily removed prior to intracellular recording (see¹⁹ for additional details).

All surgical procedures were performed under deep anesthesia, using strictly sterile techniques, in a dedicated surgical suite. All procedures were approved by the University of Texas Institutional Animal Care and Use Committee and conformed to National Institutes of Health standards.

Behavior/Eye tracking

Monkeys were trained to detect a small oriented Gabor target, indicating target presence by making a saccadic eye movement to the target location as soon as it was detected (Fig. 1a, top-left). Each trial began when a fixation spot was displayed at the center of a monitor in front of the monkey. The size of the fixation spot was 0.1°, and the monkey had to shift gaze to the fixation point. After the animal maintained fixation within a small window (less than 2° full width) for 1300–1600 ms, the fixation point dimmed as a temporal cue for 300 ms, and then a small oriented Gabor target (σ = 1/6–1/3°) at 4–5 contrast levels was presented either in the receptive field (“target in”, or “Tin”) or in the opposite side of the visual field (“target out”, or “Tout”). To receive the liquid reward, the monkey had to make a saccadic eye movement to the target location within 600 ms of its onset (but at least 75 ms after target onset) and maintain gaze on the target location for at least 100 ms. For 0% contrast, correct target location was selected randomly. Reaction time is defined as the time from target onset until the animal’s eye reached the saccade target window. Within a block of trials, target-present trials at different contrasts were randomly interleaved with a blank condition in which there was no dimming of the fixation point and the animal was required to maintain fixation until the end of the trial to receive the reward (Fig. 1a, bottom right).

Visual stimuli were presented on a gamma-corrected high-end 21-inch color display (Sony Trinitron GDM-F520) at a fixed mean luminance of 30 cd/m². The display subtended 20.5° × 15.4° at a viewing distance of 108 cm and had a pixel resolution of 1024 × 768, 30-bit color depth, and a refresh rate of 100 Hz. Visual stimuli were generated using a high-end graphics card on a dedicated PC, using custom-designed software. Behavioral measurements and data acquisition were controlled by a PC running a custom-designed software. Eye positions were monitored with an infrared eye-tracker (EyeLink, SR Research, Ottawa, Canada).

Whole-cell recording

The whole-cell recording procedures were described in details in our previous work^19,23. Recording chambers were located on the dorsal portion of V1, with the anterior portion of the chamber reaching close to the lunate sulcus and the border between V1 and V2. We verified the retinotopic organization by voltage-sensitive dye imaging²⁴. The cortex in our cranial windows represents stimuli that are approximately 2.5–5° away from the fovea in the lower quadrant of the contralateral hemifield. To increase recording stability, we filled the recording chamber with 2%–4% agarose in artificial cerebrospinal fluid (CSF). To minimize the damage to the brain, we stopped advancing the electrodes when it reached about 1300 mm in the brain, therefore our intracellular recordings were primarily from layer 2/3 in V1. A silver chloride wire was inserted into the agarose as a reference electrode. Pipettes (6–10 MΩ) were pulled from 1.2 mm outer diameter, 0.70 mm inner diameter KG-33 borosilicate glass capillaries (King Precision Glass, Claremont, CA, USA) on a P-2000 micropipette puller (Sutter Instruments, Novato, CA, USA). Patch pipettes were filled with (in mM) 135 K-gluconate, 4 NaCl, 0.5 EGTA, 2 MgATP, 10 phosphocreatine disodium, and 10 HEPES, pH adjusted to 7.3 with KOH (Sigma–Aldrich, St. Louis, MO, USA). Whole cell current-clamp recordings were performed with an Axoclamp 2B Microelectrode Amplifier (Molecular Devices, Sunnyvale, CA, USA). Membrane potential signals was detrended by high pass filtering at 0.1 Hz with MATLAB (Mathworks, Natick, MA, USA).

Membrane potential analysis

For each trial, the mean potential from 950 to 50 ms preceding the cue was subtracted from the entire membrane potential trace and this value was labeled as ΔVm. To determine changes in membrane potential from cue onset to stimulus onset, all cued trials were combined, since there was no statistically significant difference in results between target conditions or choice (Kruskal-Wallis test, P = 0.52) and all detection conditions exhibited a similar rise in potential relative to blank trials (see Fig. S1). The changes in potential from cue onset to stimulus onset were averaged across all recorded cells in 50-ms bins for both cued and no cue trials. Statistical comparisons between cued and no cue trials for each cell were done on the binned data using bootstrap analysis.

Using the same binned data, we computed Pearson’s correlation (r) between RT and ΔVm for all repeated trials of a single condition for each recorded cell. We excluded cells for any condition with fewer than three repeats. Since correlation estimates can be highly variable with small sample sizes²⁵, we used weighted averaging of r-values across cells based on the square root of the number of trials for each cell. Statistical significance of weighted average correlation of the binned data was determined with bootstrap analysis. We found similar results without weighting, but values varied more across bins and conditions. We chose the bin at stimulus onset, but the three bins before stimulus onset also exhibited significant negative correlation between RT and ΔVm that did not depend on choice (Fig. S2). Slope estimates between RT and ΔVm also produce similar, but less stable results with an average slope of 10 ms faster reaction times for every 1 mV increase in membrane potential. In all analyses, we reached the same conclusion that increases in Vm buildup predict faster reaction times at or before stimulus onset. For each cell, “low” contrast was the lowest contrast tested ranging from 3–5% and “high” contrast was the highest contrast tested ranging from 12.5–100%.

Choice probability

Using the same 50-ms bins, we computed the CP for a higher ΔVm predicting Choice In using standard receiver operating characteristic (ROC) analysis¹ separately for 0%, low contrast Tin and low contrast Tout. A CP above 0.5 means that ΔVm is higher in Choice In than in Choice Out and a CP below 0.5 means the opposite relationship. We excluded cells for any condition with fewer than two Choice In or Choice Out conditions. Statistical significance of CP above or below chance for the binned data was determined using bootstrap analysis.

Model data

Details of the additive and multiplicative models’ derivations and predictions are included in the supplementary materials.

Statistical analysis

Experiments and analysis were not performed blind to the conditions of the experiments. Statistical methods were not used to predetermine sample sizes, but our sample sizes are similar to those reported in previous studies. All statistical tests were non-parametric making no assumption about the underlying distribution using the Kruskal-Wallis test or bootstrap analysis of the mean²⁶. For bootstrapping, we resampled each set of data 1,000 times, allowing repeats, to produce surrogate datasets of the same size. Sorted estimates from these datasets were then used to determine P values relative to a null hypothesis (ΔVm = 0, r = 0, CP = 0.5). If the entire surrogate dataset was above or below the null hypothesis, we described the result as P < 0.001.

Supplementary Material

Supplementary Files

This is a list of supplementary files associated with this preprint. Click to download.

• DetectionSupplementary.pdf