Synchronized pulse-based evidence accumulation task for humans and rodents.

A) During perceptual decision-making, choices are influenced by perceptual factors, for example sensory noise and sensitivity, as well as cognitive factors, including decision thresholds and working memory. Pulse based evidence accumulation tasks were developed to discriminate the relative contributions of perceptual and cognitive factors on behavioral choice across individuals. B) Schematic of the pulse-based evidence accumulation task for cross-species studies. The goal of the task is to identify the side with the higher probability of visual stimuli. The task structure is the same across species and consists of 4 different phases: Initiation: the subject starts the task by clicking an asteroid (humans) or poking in the center (rodents), Evidence: the visual stimuli will be presented in the right or left side. Choice: during the evidence period, the subject observes the cues and responds by selecting a side. Cues continue until the side is selected. The correct choice is the side with the greater underlying probability of cue presentation. Feedback: the subject will receive a reward (increase of the score for humans and sugar water for the rodents) for a correct choice or a time-out (TO) delay for an incorrect choice. C) Left: The human version consists of an online video game format in which the player controls a spaceship in an asteroid field. The player gets points by destroying asteroids with a laser beam. Right: In the rodent version, mice and rats play the task in a 3-port operant training chamber, in which the center port is used for initiation and the left and right ports are used for making choices. D) Left: For humans, data was collected using remote video calls format. Right: For rodents, behavioral sessions were conducted from multiple animals in parallel in a semi-automated operant facility.

Humans, mice and rats accumulate evidence over time to solve the task.

A) Accuracy varies significantly across species (One-way ANOVA; p < 0.0005, F = 271.84), with mice exhibiting the lowest performance and humans demonstrating the highest. B) Mice exhibit faster response times compared to humans and rats, which display similar response times (One-way ANOVA; p < 0.0005, F = 236.39). C) Normalized bias is comparable between humans and rats but is higher in mice (One-way ANOVA; p < 0.0005, F = 11.13). D) Reward rate was similar for rodents but larger for humans (One-way ANOVA; p < 0.0005, F = 1060.97). E) Histogram of response times across species reveals that rats and humans respond at similar speeds during correct or incorrect trials, whereas mice respond faster in error trials than in correct trials (Paired t-student; p < 0.0005). Note that positive values represent correct trials, while negative values represent incorrect trials. F) Accuracy vs response time derived from a generalized additive mixed model on the three species in which we can observe the evidence accumulation process in each of the species. All of the species peak between 0.7-1 seconds but the humans present a higher peak than the rodents. Each dot represents the averaged accuracy per each 0.2-second bin for each species (Pearson Correlations Accuracy-RT, see methods; Mouse: p < 0.017, R = 0.80, Rat: p < 0.011, R = 0.83, Human: p < 0.033, R = 0.85). All comparisons are unpaired two-sample t-tests, corrected for multiple comparisons with Bonferroni correction. Notations: * = p<0.05, ** = p<0.01, *** = p<0.005, **** = p<0.001.

Drift diffusion models explain the choice and response time data for all three species.

A) Illustration of the diffusion model: two boundaries indicate decision thresholds for left and right choice, greater distance between boundaries would mean more separability between the two choice decisions. Drift rate is the average rate of evidence accumulation, the model assumes drift rate varies from trial to trial, following a normal distribution. Starting point measures the bias in choosing one side over the other. Starting point is also assumed to vary from trial to trial, following a uniform distribution. Non-decision time indexes time spent in processes other than the decision process. Trial to trial variability in non-decision time follows a half-normal distribution. Evidence on each trial is accumulated until a decision boundary is reached, evidence accumulation is governed by non-decision time, starting point bias, drift rate; drift across time is subject to a diffusion noise. B) Fit of the diffusion model to one example subject from each species, response time distributions are plotted separately for correct and error responses. C) Comparison of model performance across species. Rats are better fitted by the model than mice and humans (DDM BIC; Humans: 2772 ± 262, Rats: 1492 ± 403, Mice: 2778 ± 136; One-way ANOVA; p < 0.0005, F = 10.97). D-G) Estimates of model parameters compared across species. Rodents present lower drift (One-way ANOVA; p < 0.0005, F = 77.72) and boundaries (One-way ANOVA; p < 0.0005, F = 88.29) than humans. Rats show the slower non-decision time (One-way ANOVA; p < 0.0005, F = 347.26). All comparisons are unpaired two-sample t-tests, corrected for multiple comparisons with Bonferroni correction. Notations: * = p<0.05, ** = p<0.01, *** = p<0.005, **** = p<0.001.

Response times for rats and mice are well described by collapsing bounds.

A) Illustration of the diffusion model with collapsing boundary, three (free) parameters are used to estimate the boundary: initial boundary separation, semi-saturation constant, and amount of collapse; the model is otherwise the same as the fixed boundary model. B) Illustration of the diffusion model with an urgency signal, three (free) parameters are used to estimate the (linear) urgency signal: slope of the signal, magnitude of the signal, and onset delay for the signal. C) Comparison of model performance for the three diffusion models: fixed boundary model, collapsing boundary model, and the urgency model, for each species. Rats and mice were better fitted by the collapsing boundaries (Repeated measures ANOVA; Mice: p < 0.0005, F = 22.12; Rats: p = 0.0014, F = 11.12). D-F) Comparison of boundary related parameters of the collapsing boundary model, across species. Mice show the lowest semi-saturation constant (One-way ANOVA; p = 0.004, F = 5.69). G-I) Comparison of urgency related parameters of the urgency model across species. Mice present the highest urgency-slope (One-way ANOVA; p = 0.027, F = 3.71) and urgency delay (One-way ANOVA; p = 0.003, F = 6.06). Notations: * = p<0.05, ** = p<0.01, *** = p<0.005, **** = p<0.001.

Rats balance speed and accuracy near optimally to maximize reward rate.

A) The speed-accuracy trade-off plot illustrates the relationship between response time and accuracy across species, each dot represents the mean accuracy and response time of an individual. The dashed line represents the accuracy as a function of different choices of RT for a perfect accumulator. B-C) Representations of accuracy (B) and reward rate (C) as a function of different choice of RT for a perfect accumulator, incorporating the average trial initiation time observed in mice, which was different for correct and error responses. In the simulations, flash ratio was kept at 80% vs 20%, and flash rate was 10 Hz. The red line marks the RT associated with the peak reward rate of the accumulator. D-F) Histograms illustrate the distribution of RT observed in humans (D), mice (E), and rats (F) in comparison to the RT associated with the peak reward rate of the perfect accumulator (black dashed lines).

Mice alternate between different strategies while performing the task.

A) Generalized linear model with 4 parameters: Bias (Intercept), Flash Ratio (Right flashes - Left Flashes) / Total flashes), Win-stay Lose-switch, and Previous choice. B) Goodness of fit of the GLM presents a better fit for the human data compared to mice or rats. Test log-likelihood is in units of bits per trial, and is relative to a ‘null’ Bernoulli coin-flip model (One-way ANOVA; p < 0.0005, F = 38.67). C) Psychometric curve fits, and mean predictive accuracy across species. D) Estimated GLM weights showed that humans relied on the flash ratio more than the rodents, followed by the rats (One-way ANOVA; p < 0.0005, F = 156.28), whereas mice tended to use a previous choice strategy (One-way ANOVA; p < 0.0005, F = 18.2). E) Mean estimated GLM weights across species. F) Model comparisons among the one-state GLM, the classic lapse model (’L’), and multiple GLM-HMMs with different numbers of latent states. To compare across species, we measured the change in test log-likelihood, as a function of the number of states included in the model, relative to a one-state GLM, separately for each species. Mice showed greater increase in LL for a GLM-HMM with 3 latent states, suggesting that mice transitioned through multiple latent states while performing the task. All comparisons are based on unpaired two-sample t-tests, corrected for multiple comparisons with Bonferroni correction. Notations: * = p<0.05, ** = p<0.01, *** = p<0.005, **** = p<0.001.

Learning curricula in the cross-species task.

A) Experiments occurred in multiple phases: Shaping (rodents only), Training and Testing. Prior to training, rodents transitioned through three shaping stages to familiarize them with the operant chamber. In shaping stage 1 they received reward for inserting their noise in an illuminated noise poke. In shaping stage 2 they received reward for first poking in the center port to initiate a trial and then in the illuminated side port with the light. In shaping stage 3 they received reward for first poking in the center port to initiate a trial and then in the flashing side port. During the first stage of training both humans and rodents performed a simple version of the task in which the probability of cue presentation was high on the correct side (90% per time point) and low on the incorrect side (10% per time point). After completion of training, subjects transitioned to the testing phase in which the probability of cue presentation was 80% on the correct side, and 20% on the incorrect side. B) Table showing the number of subjects and training criteria for each species across the learning curricula. C) Evolution of averaged accuracy, response time and number of trials across learning in mice, rats and humans. D) Example mouse, rat and human evolution across stages and sessions. The red dashed lines represent the transitions across stages.

Comparison of accuracy, response time and bias across species, separately for female and male subjects.

Notations: * = p<0.05, ** = p<0.01, *** = p<0.005, **** = p<0.001.

Parameter recovery for the DDM.

Based on the parameters observed with the empirical data, we simulated 50 datasets per species with 600 trials each. After fitting the DDM to each of these simulated datasets, we compared the generative and fitted parameters.

Comparison of drift rate, non-decision time and starting point bias across species, separately for the collapsing boundary DDM and the urgency DDM.

Both models present similar parameter estimates and cross-species trends as the base DDM. Notations: * = p<0.05, ** = p<0.01, *** = p<0.005, **** = p<0.001.

Comparison of DDM parameters that control the trial-to-trial variability of its main parameters, such as drift rate, non-decision time and starting point bias.

We assumed variability of drift rate from trial to trial follows a normal distribution, that for non-decision time follows a half-normal distribution, and for starting point bias, trial-to-trial variability follows a uniform distribution. A) corresponds to the original DDM, B) corresponds to the collapsing boundary model, C) corresponds to the urgency DDM. Notations: * = p<0.05, ** = p<0.01, *** = p<0.005, **** = p<0.001.

Differences in intertrial intervals across species and its effects on the reward rate simulations.

A-C) Distributions of initiation times across all correct and erroneous responses, separately for mice, rats and humans. Initiation time was the period between the start of a trial and the center poke, after the center poke, flash stimuli were presented; for humans this was the time to choose an asteroid. D) Reward rate (rewards/min) as a function of different choices of RT for a perfect accumulator, simulated separately for humans (green line) and rodents (purple line). RT corresponding to the peak reward rate is noted. E) Reward rate of a perfect accumulator taking into account different initiation times. F) Reward rate as a function of different choices of RT for noisy (animal) accumulators. Separate simulations were run to account for average trial initiation times observed in rats and in mice, for correct and error responses. In the simulations, flash ratio was kept at 80% vs 20%, and flash rate was 10 Hz, same as that for the actual experiments. In addition to initiation time, choice accuracy for at different RT levels was predicted by the GAMM, previously fit to animal choice/RT data (see Figure 2F). G-I) Reward rate as a function of different choices of RT for humans (G), mice (H) and rats (I) based on a perfect accumulator, incorporating the average trial initiation time observed in each species, for correct and error responses. The black dashed line represents the response time where the reward rate of the accumulator peaks and the colored line is the average response time of each species.

Mouse data is better fit with a hidden Markov model with 3 latent states.

A) Change in test log-likelihood as a function of the number of states included in the GLM-HMM, relative to a (one-state) GLM, modeled separately for each mouse. The classic lapse model, a restricted form of the two-state model, is labeled as ‘L’. Each trace represents a single mouse and the solid black line indicates the mean across animals. B) Change in predictive accuracy relative to a one-state GLM for each mouse, indicating the percentage improvement in predicting choice. C) Gray dots correspond to individual sessions across all mice, indicating the fraction of trials spent in state 1 (engaged) and state 2 (biased left). Points at the vertices (1,0), (0,1) or (0,0) indicate sessions with no state changes, whereas points along the sides of the triangle indicate sessions that involve only two of the three states. Red dots correspond to the same fractional occupancies for each of the mice, revealing that the engaged state predominated but that all mice spent time in all three states. D) Inferred GLM weights for each mouse, for each of the three states in the three-state model. The solid black curve represents a global fit using pooled data from all mice. E) Psychometric curve for each state, conditioned on previous reward (solid line if right and dashed line if left) and previous choice (darker color if rewarded and dashed line if not rewarded) and all the states combined.

Rats exhibited stability throughout the entire session, whereas mice displayed greater variability.

A-D) Averaged accuracy (A), response time (B), number of trials (C), and reward rate (D) across a session. It is important to note that mouse sessions lasted for 60 minutes and rat sessions for 120 minutes.