Claims 39
From the paper 10.1371/journal.pcbi.1013226 · claims file · checked 2026-10-21: ready
- data-composition
The data are choice behaviour of 5 male mice over 14 full-task sessions each, with dLight photometry from nucleus accumbens core in a left, right, no-recording sequence.
We included n = 5 subjects with 14 sessions each
Methods, Quantification and statistical analysis; Fig 1B legendHow it will be judged
Reproduced when the dataset holds 5 mice with 14 full-task sessions each and photometry in about two of every three sessions; partially reproduced when it holds the 5 mice with fewer sessions or only the recorded sessions. The counts found are reported.
- perf-after-pretraining
Just after pretraining mice chose the high-probability port 72.9% of the time on average (95% CI 0.669 to 0.767).
Just after pre-training, mice chose the high reward probability port 72.9% (95% CI ... of the time on average
Results, Mice rapidly adapt to block switches in the 2ABT, paragraph 1How it will be judged
The fraction of trials with a correct choice. Readings of "just after pre-training" (a) the first full-task session, (b) all 14 full-task sessions (the sentence says "on average" and is followed by the change over sessions); and of the average (i) pooled over trials, (ii) the mean of the mice's fractions. Reproduced when one reading lies inside [0.669, 0.767].
- fig1f-performance-increase
Over the 14 sessions performance rose steadily, with a positive linear regression slope of 0.0026 per session (t = 2.743, p = 0.008).
Over the course of 14 training sessions, performance steadily increased ( Fig 1F , linear regression slope coefficient 0.0026, t statistic
Results, Mice rapidly adapt to block switches in the 2ABT, paragraph 1; Fig 1FHow it will be judged
OLS of the fraction of correct choices on session number (1 to 14). Readings of the unit (a) one value per mouse and session, 70 points (t = 2.743 gives p = 0.008 at 68 degrees of freedom, which fits this reading), (b) one value per trial. Reproduced when, under one reading, the slope lies inside [0.001, 0.005] with p < 0.05; partially reproduced when the slope is positive with p < 0.05 outside that interval.
- trials-to-switch
After a block switch mice took 2.49 trials on average to switch to the new correct port in the first 7 sessions, and 2.36 trials in sessions 8 to 14.
In the first 7 sessions, mice took 2.49 trials on average (95% CI ... to switch to the correct port after a block switch. In the 8-14th session this reduced to 2.36 trials on average
Results, Mice rapidly adapt to block switches in the 2ABT, paragraph 1; S1 FigHow it will be judged
Per block switch, the trial (counted from the block switch) of the first choice of the new correct port (S1 Fig legend, trial2sw). Readings of the count (a) the first trial of the new block is 1, (b) it is 0 (Fig 2F legend calls the switch trial 0); and of the average (i) over block switches pooled, (ii) the mean of the mice's means. Reproduced when, under one reading, both values lie in their intervals ([2.39, 2.59] and [2.28, 2.45]); partially reproduced when the later sessions are lower under that reading but a value lies outside its interval.
- trials-to-commit
Mice switched to and committed to the correct port for 3 or more trials after an average of 3.12 trials (95% CI 3.00 to 3.26), following the first 7 sessions.
Mice switched and committed to the correct port for 3 or more trials (average 3.12 trials
Results, Mice rapidly adapt to block switches in the 2ABT, paragraph 1; S1 Fig legendHow it will be judged
Per block switch, the trial (counted from the block switch) of the first correct choice followed by 2 more correct choices (S1 Fig legend, trial2asymp). The sentence cites S4 Fig, whose legend is about LMER lags; S1 Fig is the one that defines the measure. Readings of "following the first 7 sessions" (a) sessions 8 to 14, (b) sessions 1 to 7; of the count, (i) first trial of the new block is 1, (ii) it is 0; of the average, pooled over blocks or the mean of the mice's means. Reproduced when one reading lies inside [3.00, 3.26].
- figs1-switch-measures-decrease
The number of trials to switch and to commit to the correct port fell over blocks within a session and over sessions.
A downward trend for both measures suggests learning and faster switches across different reward blocks within session. Second column x-axis shows that the switch measures decreased as the number of training sessions increased
S1 Fig legendHow it will be judged
trial2sw and trial2asymp as in trials-to-switch and trials-to-commit, per block switch. The paper reports no test. Reproduced when the OLS slope of each measure on block number within session and on session number is negative; the slopes and their p-values are reported. Partially reproduced when two or three of the four slopes are negative.
- fig1e-stay-after-omission
The probability of repeating the last choice fell after one unrewarded outcome (RR against RU) and fell further after two (RU against UU), with Fisher's exact test statistics of 6.47 and 3.98.
We found a significant reduction in the probability of repeating the previous choice after one unrewarded outcome (RR vs. RU, Fisher's exact test ... Moreover, mice reduced their stay rate at the chosen port after two consecutive unrewarded outcomes compared to one unrewarded outcome (RU vs. UU, Fisher's exact test
Results, Mice rapidly adapt to block switches in the 2ABT, paragraph 2; Fig 1EHow it will be judged
Trials t whose choices at t-2 and t-1 were at the same port (Results; the code's two-trial history plot uses Decision(t-1) == Decision(t-2) and Trial > 2), classed by the outcomes at t-2, t-1; stay is the choice at t equal to t-1. The paper lists the counts as "RR, RU, UR, RU"; the fourth is read as UU. Fisher's exact test on the 2 x 2 table of history against stay or switch, its statistic read as the odds ratio (scipy); Cohen's d on the 0/1 stay indicators of the two groups with pooled SD. All 14 sessions of the 5 mice, trials pooled. Reproduced when both tests give p < 1e-4 with stay lower in the second group, and the odds ratios and d values are within tolerance; partially reproduced when the tests hold and some values are off. The counts are reported against the paper's.
- figs6a-decaying-history-weights
Logistic regressions fitted to each mouse show weights of past choice-outcome history that decay for more distant trials, consistently across mice.
we fitted logistic regression models to individual mice behavior data and identified consistent patterns of decaying choice outcome weights for more distant outcome histories
S6 Fig A legend; Results, Mice rapidly adapt to block switches in the 2ABT, paragraph 3How it will be judged
Per mouse, the S6 Fig A model following Beron et al. 2022, choice at t on lagged choice (coded -1/1), reward, and their product; this is the code's LR class (formula Decision ~ C_iback + R_iback + C_iback:R_iback), whose default of 4 lags is used, as the legend gives none. Fitted with a binomial GLM (code). Reproduced when in every mouse the lag-1 choice-by-reward weight is positive and the largest of the four, and the mean over mice falls from lag 1 to lag 4; partially reproduced when this holds in 3 or 4 mice. The S6 Fig B blocking model is described but no result is stated, so it has no claim.
- fig2c-bayesian-beat-rl4p
Both Bayesian models fit behaviour better than the standard RL model RL4p, with relative AIC of -536.57 (BIfp) and -530.59 (BRLfwr).
With RL4p as a baseline, we found that both BRLfwr and BIfp significantly improved the AIC measure compared to the RL4p model (BIfp: Δ AIC = − 536.57 ± 59.48 , BRLfwr: Δ AIC = − 530.59 ± 61.86 )
Results, Cognitive model fitting confirms, paragraphs 4 and 5; Fig 2CHow it will be judged
Models fitted per mouse to all its sessions as in the opening comment; relative AIC per mouse = AIC(model) - AIC(RL4p), averaged over the 5 mice (the ± is read as the stated spread; whether it is an SD or SEM is not said). Test, a t-test of the 5 relative AICs against 0 (equivalent to a paired t-test), Holm-corrected. The family and the sides are not given; readings (a) the six models against RL4p named in the sentence (BRLfwr, BIfp, RFLR, RLFQ3p, RL_meta, PearceHall), (b) those and RLCF; (i) two-sided, (ii) one-sided. Reproduced when both mean relative AICs are within tolerance and both corrected p < 0.05; partially reproduced when both are negative with corrected p < 0.05 but a value is outside tolerance.
- fig2c-complex-rl-beat-rl4p
The complex RL models also fit behaviour better than RL4p, with relative AIC of -652.53 (RFLR), -677.84 (RL_meta), -656.44 (PearceHall) and -646.45 (RLFQ3p).
Additionally, other complex RL models also outperformed the RL4p model ( Δ AIC = − 652.53 ± 49.26 , − 677.84 ± 65.69 , − 656.44 ± 62.16 , − 646.45 ± 53.62 for RFLR, RL_meta, PearceHall, and RLFQ3p respectively
Results, Cognitive model fitting confirms, paragraphs 4 and 5; Fig 2CHow it will be judged
As in fig2c-bayesian-beat-rl4p, same test, family readings and sides. RLFQ3p is run with both forgetting rules (Methods, and Table 1 and code). Reproduced when all four means are within tolerance and all corrected p < 0.05; partially reproduced when all are negative with corrected p < 0.05 but a value is outside tolerance.
- fig2c-rlcf
RLCF fits behaviour better than RL4p (relative AIC -361.92) but worse than BIfp (by 174.64), and BRLfwr is significantly better than RLCF.
The RLCF also explained the mouse behavior better than the RL4p model (RLCF: Δ AIC = − 361.92 ± 70.52 ), with a higher relative AIC on average than the BIfp model ( Δ AIC = 174.64 ± 47.83 )
Results, Cognitive model fitting confirms, paragraphs 4 and 5; Fig 2CHow it will be judged
As in fig2c-bayesian-beat-rl4p. BRLfwr against RLCF is a t-test of the per-mouse AIC differences, Holm-corrected within the family of the five BRLfwr comparisons of the sentence (RL_meta, PearceHall, RFLR, RLFQ3p, RLCF), two-sided or one-sided as readings. Reproduced when both values are within tolerance and the corrected p < 0.05 with BRLfwr lower; partially reproduced when the signs hold and a value is outside tolerance.
- fig2c-no-bayes-vs-complex-rl
Behaviour alone cannot tell whether the Bayesian models or the complex RL models are better; BRLfwr does not differ significantly from RL_meta, PearceHall, RFLR or RLFQ3p.
However, there was not sufficient discriminative power to identify whether Bayesian models or any member of the complex RL models was superior
Results, Cognitive model fitting confirms, paragraph 5How it will be judged
A claim of absence, judged on the paper's test. Two-sided t-tests of the five per-mouse AIC differences of BRLfwr against each of RL_meta, PearceHall, RFLR and RLFQ3p, Holm-corrected in the family of the five comparisons the sentence lists (with RLCF). Reproduced when all four corrected p >= 0.05; not reproduced when any is below. The mean differences are reported either way. The paper does not say whether its t-tests are one- or two-sided; the one-sided tests are shown as a variant.
- fig2b-model-identifiability
In model identification, the Bayesian model and the hybrid BRL are poorly told apart, and so are the complex RL models (RL_meta, RLFQ3p, RFLR), but the families can be told apart and the standard RL model RL4p can be rejected.
This revealed relatively poor identifiability between the Bayesian model and the hybril BRL, as well as poor identifiability between the complex RL models (RL_meta, RLFQ3p, RFLR). These results indicate that the 2ABT cannot be used to discriminate within each family of model on the basis of behavioral data, but can discriminate across families
Results, Cognitive model fitting confirms, paragraph 4; Fig 2BHow it will be judged
The 8 models (RL4p, RLCF, RFLR, BRLfwr, BIfp, RLFQ3p, RL_meta, PearceHall), each simulated from parameters drawn uniformly over the range fitted to the mice, each simulated set fitted with all 8, the best by lowest AIC, and the confusion matrix P(fit j | simulated i) (Methods, Model identification). Simulation readings (a) Methods, 1000 sets per model on the 7 to 23 reward blocks; (b) code, 500 sets of 500 trials on 8 to 15 reward blocks. The paper gives no threshold for "poor" or "can discriminate"; read as (1) BIfp and BRLfwr each send more of their row to the other than any non-Bayesian model takes, and at least one of the two has a diagonal below 0.5; (2) the same within RL_meta, RLFQ3p and RFLR; (3) for every row, the summed probability on its own family (Bayesian = BIfp, BRLfwr; complex RL = RL_meta, RLFQ3p, RFLR, PearceHall; RL4p and RLCF each alone) is above 0.5; (4) the RL4p row's diagonal is its largest entry and no other row gives RL4p more than 0.2. Reproduced when all four hold; partially reproduced when (3) and (4) hold. The full matrix is shown.
- figs2-parameter-recovery
Parameters are recovered well for the Bayesian models such as BRLfwr, while RL_meta, the most complex model with 7 parameters, has poor parameter identifiability.
However, RL_meta was also the most complex model with 7 parameters, with poor parameter identifiability ( S2 Fig ). One might therefore argue that that the Bayesian model BRLfwr was better because it exhibited good parameter recovery
S2 Fig legend; Results, Bayesian models and complex RL models, paragraph 4; Methods, Model fitting and comparisonHow it will be judged
For each model, parameters drawn uniformly between the minimum and maximum fitted over the mice (S2 Fig legend), behaviour simulated and refitted; Pearson r between true and fitted values per parameter, after dropping fitted values more than 6 SD from the mean true value as the legend and code do. Simulation readings as in fig2b-model-identifiability. The Methods' sentence that all three model classes recover parameters with high correlation conflicts with poor identifiability for RL_meta; the claim follows the Results. No threshold is given; read as good when every parameter has r >= 0.7 and poor when at least one has r < 0.5. Reproduced when BRLfwr is good and RL_meta is poor; partially reproduced when one of the two holds. All r values are shown.
- brlfw-alpha-zero
In the unrestricted BRL model (BRLfw) the fitted learning rate alpha was zero for most subjects, making it equivalent to BIfp in this task.
Interestingly, the best fitted parameter for α in the unrestricted BRL version (BRLfw) above was zero for most subjects
Methods, Computational modeling, paragraph on BIfpHow it will be judged
BRLfw as in the code (BRL_fw, both weights updated, Eqs 7 and 8, initial weights 0.75 and 1e-4), fitted per mouse. Zero is read as the lower bound to the optimiser's precision (alpha < 1e-6). Reproduced when this holds in at least 3 of the 5 mice. The fitted values are reported.
- figs8-brlwrp-no-difference
A BRL model with fitted initial reward weight and reward probability (BRLwrp) does not differ from BRLfwr in explaining behaviour or dopamine.
No statistical differences were observed between BRLfwr and BRLwrp in their ability to explain mice behavior and dopamine data
S8 Fig legend; Methods, Computational modeling, paragraph on BRLHow it will be judged
A claim of absence. BRLwrp as in the code (BRL_wrp, wr0 and p_rew fitted). The test is not named; read as the paper's model comparison tests. For behaviour, a two-sided t-test of the 5 per-mouse AIC differences, p >= 0.05; for dopamine, the 95% bootstrapped CI of the Δllk_CV difference (as in fig4f-bayesian-rpe-fit) contains 0. Reproduced when both hold; not reproduced when either finds a difference. Effect sizes are reported. The Methods cite S8 Fig for a different comparison, BRLfwr against the unrestricted BRL (BRLfw, "which we found to perform just as well"); the claim follows the legend, and the same tests of BRLfw against BRLfwr are shown as a variant.
- figs7-no-change-across-sessions
Models fitted to single sessions show no sustained, consistent change in model fitness across the full-task sessions, so no change of strategy across days.
We did not find evidence of any significant and consistent increase in model fitness across multiple sessions
S7 Fig legend; Discussion, paragraph 6How it will be judged
A claim of absence; the paper names no test. Each model fitted per mouse and session (behaviour, Discussion paragraph 6). The legend does not say whether "model fitness" is relative to RL4p or absolute; readings (a) per session relative AIC to RL4p, (b) per session AIC divided by the number of trials with a choice; for BIfp and BRLfwr. Read as an OLS of the 70 per-session values on session number, for each Bayesian model. Reproduced when neither slope is negative (improving) with p < 0.05; not reproduced when either is. The slopes are reported under both readings.
- fig2f-adaptation-after-switch
BRLfwr and RLCF adapt to block switches faster than RL4p, resembling the mice.
we observed that BRLfwr and RLCF predicted faster adaptation to block switches compared to RL4p, resembling the adaptation rate of mouse switching behavior
Results, Bayesian models and complex RL models, paragraph 1; Fig 2FHow it will be judged
P(correct) by trial after block switch, block trials 0 to 10 (the code's behaviour plot uses blockTrial <= 10, the first trial of a new block being 0). Model behaviour readings (a) the code's emulate, the model choosing on each mouse's sessions with its fitted parameters; (b) the code's generate, fresh sessions with each mouse's parameters, on 7 to 23 or 8 to 15 reward blocks. Measure, the mean P(correct) over block trials 0 to 10. Reproduced when BRLfwr and RLCF each exceed RL4p and are each closer to the mice's mean than RL4p is.
- fig3b-sse
Against mouse switch probabilities after each three-trial history, the models' errors are RL4p 0.8270, RLCF 0.2748, BRLfwr 0.1796, RFLR 0.1537, PearceHall 0.1084 and RL_meta 0.0787; RL4p fits worst.
We then compared the average mouse switch probability after each outcome history against the mean predictions of the three different cognitive models ( Fig 3B ; sum of squared errors (SSE)
Results, Bayesian models and complex RL models, paragraph 3; Fig 3A and 3BHow it will be judged
Histories as defined in the opening comment, every history with trials. SSE over histories of (mouse switch probability - model switch probability) squared. Readings of the mouse value (a) pooled over trials, (b) the mean of the mice's values; of the model value (i) the fitted model's predicted probability of switching on each mouse trial, averaged per history, (ii) the switch rate of emulated choices. Reproduced when, under one reading, all six SSEs are within tolerance and RL4p is largest; partially reproduced when RL4p is largest but some values are outside tolerance.
- fig3a-both-port-contexts
After histories with rewards at both ports in the last 3 trials, mice rarely switched (0.2143) and BRLfwr predicted a similar rate (0.1801), while RL4p overestimated it (0.4328); RLCF gave 0.2248 and RFLR 0.2778.
Since these outcome contexts occurred only around block switches, mice usually stayed at the newly rewarded ports (mean switch rate ... BRLfwr showed a similarly low switch rate after these contexts (0.1801), whereas RL4p model predicted a high average switch rate of 0.4328
Results, Bayesian models and complex RL models, paragraph 3; Fig 3AHow it will be judged
The contexts are the histories with at least one A and at least one B (rewards at both ports, e.g. ABb, AaB). Readings of the mean (a) pooled over trials of those contexts, (b) the mean over contexts, (c) the mean of the mice's values; model values as in fig3b-sse. Reproduced when, under one reading, all five are within tolerance; partially reproduced when RL4p is the highest and BRLfwr is within 0.05 of the mice but values are outside tolerance.
- fig3c-aic-around-switch
Around block switches (first 5 trials after a switch) BRLfwr and the dynamic RL models fit similarly better than RL4p, with relative AIC of -163.68 (BRLfwr), -163.92 (RL_meta) and -158.91 (PearceHall).
We found similar Δ AIC results for BRLfwr and dynamic RL models (BRLfwr: −163.6763, RL_meta: −163.9207, PearceHall: −158.9066)
Results, Bayesian models and complex RL models, paragraph 4; Fig 3CHow it will be judged
The models fitted to all trials, the AIC taken over the first 5 trials after each block switch (Results). Readings (a) block trials 0 to 4 or 1 to 5; (b) 2k - 2 log L on those trials, or -2 log L alone; (c) the mean of the per-mouse relative AICs or their sum. Reproduced when, under one reading, all three are within tolerance; partially reproduced when all three are negative and within 10% of each other but outside tolerance.
- rl4p-learning-rates
The best-fitting RL4p model has large learning rates, alpha+ 0.85 and alpha- 0.73 on average.
In our dataset, the best fitting RL4p model had an average α + of 0.85 (95% CI ... and average α − of 0.73
Results, Predictions from BRLfwr and BIfp, paragraph 6How it will be judged
RL4p fitted per mouse as in the opening comment; the mean of the 5 mice's values. The narrow CI for alpha- may come from per-session fits; reading (b) is the mean over per-mouse-and-session fits. Reproduced when, under one reading, both lie inside their intervals; partially reproduced when both are above 0.5 but outside.
- fig4bc-reward-omission
Rewards are followed by an increase of NAc dLight signal (estimate 1.6345) and omissions by a decrease (estimate -1.5840).
After mice received a water reward, we observed a dLight dopamine signal (Z(DA)) increase on average (estimate ... When reward was omitted at the peripheral port, we observed a reduction (OLS estimate
Results, Predictions from BRLfwr and BIfp, paragraphs 1 and 2; Fig 4B and 4CHow it will be judged
The OLS with the four factors of the Results (switch, choice laterality relative to the recorded hemisphere, reward, omission), read as DA ~ 0 + reward + omission + switch + contralateral, the coding under which both a reward and an omission estimate exist. The paper prints the reward estimate once as -1.6345 with the CI [1.613, 1.656]; it is read as positive. Readings of the response (a) DA-PT, (b) the mean de-based Z(DA) over 0 to 1 s; of the trials (i) cleaned trials, (ii) all trials of the 34 sessions with valid DA. Reproduced when, under one reading, both coefficients are inside their CIs with p < 1e-5; partially reproduced when the signs hold with p < 1e-5 but values are outside.
- fig4bc-switch-effect
Outcome dopamine is shifted upward on switch trials (coefficient 0.5030); reward responses are larger when mice have just switched ports, and omission dips are larger when they stayed.
We observed a significant Z(DA) difference (OLS coefficient estimate for switch
Results, Predictions from BRLfwr and BIfp, paragraph 2; Fig 4B and 4CHow it will be judged
The same OLS and readings as fig4bc-reward-omission. Reproduced when, under one reading, the switch coefficient lies in [0.454, 0.552] with p < 1e-5 and the mean response of rewarded switch trials exceeds rewarded stay trials while unrewarded stay trials are more negative than unrewarded switch trials; partially reproduced when the signs hold and the coefficient is outside its interval.
- fig4bc-no-laterality
Choice laterality relative to the recorded hemisphere has no significant effect on outcome dopamine (p = 0.523).
There was no significant effect of hemisphere (laterality, contra or ipsi relative to reward port) on NAc dopamine release (p = 0.523)
Results, Predictions from BRLfwr and BIfp, paragraph 2; Fig 4B and 4CHow it will be judged
A claim of absence, judged on the paper's test, the laterality coefficient of the four-factor OLS of fig4bc-reward-omission. Reproduced when its p >= 0.05 under the readings of that claim; not reproduced when it is below. The coefficient is reported.
- fig4de-port-duration
The duration of the outcome dopamine response follows the time the mouse stays at the reward port, with the trough after omissions reached just after the mouse leaves.
When we visualized single trial dopamine traces around "outcome" events, we found that the temporal span of the dopamine response qualitatively aligned with the amount of time mice stayed at the peripheral reward port
Results, Predictions from BRLfwr and BIfp, paragraph 3; Fig 4D and 4EHow it will be judged
The paper shows this only qualitatively, in heatmaps sorted by port_dur. Read as (1) on rewarded trials, the Spearman correlation between port_dur and the time after outcome at which de-based Z(DA) first falls below half its peak after the peak is positive with p < 0.05; (2) on unrewarded trials, the Spearman correlation between port_dur and the time of the minimum in 0 to 6 s after outcome is positive with p < 0.05, and the median of (trough time - port_dur) lies between 0 and 1 s. Cleaned trials of the 34 sessions, pooled. Reproduced when all hold; partially reproduced when (1) and the correlation of (2) hold.
- trial-cleaning-counts
After trial cleaning 20,952 trials (about 68% of the data) remain across 34 sessions with valid dopamine recordings.
After this cleaning, we were left with 20952 trials ( ~ 68 % of the dataset) across 34 sessions with valid dopamine recordings
Methods, Neural data analysis, paragraph 2How it will be judged
Sessions and cleaning as in the opening comment, under both SO_lat limits. Reproduced when, under one reading, the session count is 34 and the trial count within tolerance; partially reproduced when the trial count is within tolerance with another session count.
- fig4f-bayesian-rpe-fit
RPEs of BIfp and BRLfwr explain outcome dopamine better than RL4p RPEs and better than RLCF, RLFQ3p, RL_meta and PearceHall, by cross-validated log-likelihood.
we found that both BIfp and BRLfwr models yielded larger relative log-likelihood ( Δ llk CV ) than the RL4p baseline, evaluated on cross validation sets (70-30) for each animal ... Both Bayesian models also outperformed other complex RL models on this measure
Results, Predictions from BRLfwr and BIfp, paragraph 4; Fig 4FHow it will be judged
Cleaned trials of the 34 sessions. Per model, the Methods' regression outcome_DA_PT ~ rpe:rpe_pos + rpe:rpe_neg + port_dur:rpe_pos + port_dur:rpe_neg + ego_action + session_num + log__MVMT + log__center_dur, by OLS (Gaussian likelihood), with rpe the model RPE of the opening comment, rpe_pos and rpe_neg indicators of rpe >= 0 and < 0, ego_action the choice contralateral or ipsilateral to the recorded hemisphere, session_num numeric, natural logs. Per animal, 70% of trials to fit and 30% to test, test log-likelihood with the training residual variance; Δllk_CV = llk(model) - llk(RL4p) on the same split. Not given ("featurized covariates", the split, the number of splits, the unit of the bootstrap); readings (a) random trial splits, 100 repeats per animal, Δ summed over animals per repeat, bootstrap over repeats; (b) the same with Δ per animal and bootstrap over animals and repeats; (c) splits by session. BIfp RPE and RLFQ3p forgetting with both variants of the opening comment. Reproduced when, under every reading, BRLfwr and BIfp have Δllk_CV > 0 with CI above 0 and each exceeds RLCF, RLFQ3p, RL_meta and PearceHall, and under one reading the six values lie inside the paper's intervals. Partially reproduced when this order holds but no reading puts the values inside the intervals, or when one Bayesian model meets the order.
- figs3-other-metrics
Relative AIC, relative BIC, cross-validated R2 and cross-validated log-likelihood of the dopamine regressions all favour the Bayesian models' RPEs.
All metrics converged on favoring Bayesian model predictions of dopamine at outcome phase
S3 Fig legend; Methods, Neural data analysis, paragraph 2How it will be judged
The regressions of fig4f-bayesian-rpe-fit; AIC and BIC from the fit to all cleaned trials per animal, R2 and log-likelihood from the test sets, each relative to RL4p. Reproduced when, for every metric, BIfp and BRLfwr are better than RL4p and than each of RLCF, RLFQ3p, RL_meta and PearceHall; partially reproduced when this holds for at least two metrics.
- fig4g-history-decrease
On consecutive stay trials, dopamine responses to reward decrease as recent rewards accumulate (from aaa to AAA) but do not flatten to no response.
As mice encountered more rewards in their recent history, dLight dopamine responses to reward gradually decreased (consistent with decreased RPE) but did not completely flatten to no response
Results, Predictions from BRLfwr and BIfp, paragraph 5; Fig 4GHow it will be judged
Rewarded trials t with t-3, t-2, t-1 and t at the same port (Fig 4G legend), by the outcomes at t-3 to t-1 (aaa to AAA). Cleaned trials. The paper reports no test. Read as (1) the Spearman correlation between the number of rewards in the history (0 to 3) and DA-PT is negative with p < 0.05; (2) on AAA trials the mean de-based Z(DA) over 0 to 1 s (DA-PT of a rewarded trial cannot fall below 0, so it cannot show flattening) has a 95% bootstrapped CI above 0. Reproduced when both hold.
- fig4h-rpe-only-llk
OLS models predicting dopamine from each model's RPE alone fit best for BRLfwr (cross-validated log-likelihood -985.89, against -991.80 to -995.02 for the RL models).
We compared the model fitness across different cognitive models with cross validated log-likelihood (BRLfwr ... Indeed, BRLfwr RPE predictions achieved the highest model fitness to dopamine data
Results, Predictions from BRLfwr and BIfp, paragraph 5; Fig 4HHow it will be judged
Per model, an OLS of DA-PT on the model RPE only, with cross-validation as in fig4f-bayesian-rpe-fit. Not given; readings of the RPE terms (a) rpe:rpe_pos + rpe:rpe_neg, (b) one rpe slope; of the trials (i) the Fig 4G subset (rewarded consecutive stay trials), (ii) all cleaned trials. Reproduced when, under every reading, BRLfwr has the highest llk_CV of the six and, under one reading, the six values are within tolerance (10% is loose here, as the values differ by under 1%); partially reproduced when the order holds and no reading gives the values. BIfp, which the sentence omits, is shown.
- fig4h-nonnegative-rpe-aaa
BRLfwr and BIfp predict a small but non-negative RPE on rewarded stay trials after consecutive rewards, as dopamine shows, while RL_meta and PearceHall drive the RPE toward zero as rewards accumulate.
This also allowed the best fitting BRLfwr and BIfp model to predict a small but non-negative RPE response on rewarded stay trials ( Fig 4H )
Results, Predictions from BRLfwr and BIfp, paragraphs 6 and 7; Fig 4HHow it will be judged
On the Fig 4G trials, mean model RPE per history. Reproduced when on AAA trials the mean RPE of BRLfwr and of BIfp is above 0 and above that of each of RL4p, RLCF, RLFQ3p, RL_meta and PearceHall; partially reproduced when one of the two Bayesian models meets this.
- fig4-consistent-after-reward
Among trials following a reward at t-1, the dopamine response is consistent whatever the more distant history, and only BRLfwr captures this.
Furthermore, we observed that among trials after an immediate past trial reward, the dopamine response was highly consistent, independent of more distant outcome histories. Only BRLfwr captured this phenomenon qualitatively
Results, Predictions from BRLfwr and BIfp, paragraph 7; Fig 4G and 4HHow it will be judged
The Fig 4G trials with A at t-1 (aaA, AaA, aAA, AAA). No test is given. Read as (1) a one-way ANOVA of DA-PT across the four histories has p >= 0.05; (2) the range (max - min) of the four mean model RPEs, divided by the model's range over all eight histories, is smallest for BRLfwr among BRLfwr, BIfp, RL4p, RLCF, RLFQ3p, RL_meta and PearceHall. Reproduced when both hold; partially reproduced when (1) holds alone.
- fig4i-lmer-switch
On switch trials, past rewards at the chosen port reduce outcome dopamine (slope -0.128) and past rewards at the other port increase it (slope 0.126).
LMER fitting revealed a significantly negative effect of R_chosen on dopamine at outcome phase during animal switch trials (slope estimate ... and a significantly positive effect of R_unchosen on dopamine during animal switch trials
Results, Predictions from BRLfwr and BIfp, paragraph 8; Fig 4IHow it will be judged
Cleaned trials. R_chosen is the number of rewarded trials among t-1 to t-4 at the port chosen at t, R_unchosen at the other port; Reward the outcome at t; Switch and Stay indicators of trial t. The Methods' lme4 formula DA ~ Reward:Switch + R_chosen:Switch + Switch + R_unchosen:Switch + 1 + Reward:Stay + R_chosen:Stay + R_unchosen:Stay, each term with a by-Subject random effect. Not given; readings of the random part (a) (term | Subject) for each term, (b) uncorrelated random slopes (0 + term | Subject) plus (1 | Subject); of DA, (i) DA-PT, (ii) mean de-based Z(DA) over 0 to 1 s; Wald CIs. Reproduced when, under one reading, both slopes have the paper's sign with CI excluding 0 and lie in the paper's intervals; partially reproduced when the signs hold with CI excluding 0 but a value lies outside.
- fig4j-lmer-stay
On stay trials, past rewards at the chosen port reduce outcome dopamine (slope -0.09), as all models predict.
All models predicted a negative relationship between R_chosen and RPE signal, matching the dopamine observation (slope estimate
Results, Predictions from BRLfwr and BIfp, paragraph 8; Fig 4JHow it will be judged
The same model and readings as fig4i-lmer-switch, the R_chosen:Stay term; model predictions as in fig4i-model-runchosen. Reproduced when, under one reading, the slope is negative with CI excluding 0 inside [-0.11, -0.07] and every model's R_chosen:Stay coefficient is negative; partially reproduced when the data slope holds outside its interval.
- fig4i-model-runchosen
On switch trials BIfp, BRLfwr and RLCF predict a positive effect of past rewards at the other port on the RPE, while the other RL models predict little or none.
On animal switch trials, both BIfp, BRLfwr, and RLCF predicted a positive effect of R_unchosen, but other RL models predicted little to no effect
Results, Predictions from BRLfwr and BIfp, paragraph 8; Fig 4I rightHow it will be judged
Model RPEs regressed with the fixed part of the fig4i-lmer-switch formula by OLS, without subject effects (Methods, Linear mixed effects modeling). The Methods speak of model simulated data with a uniform sample size per simulated session; readings (a) the code's generate with each mouse's fitted parameters, one simulated session per real session of the same length, on 7 to 23 or 8 to 15 reward blocks; (b) RPEs of the models run on the mice's own trials. "Little to no effect" is read as not significantly positive at p < 0.05, or below half the smallest of the three positive coefficients. Reproduced when, under every reading, the R_unchosen:Switch coefficient is positive with p < 0.05 for BIfp, BRLfwr and RLCF and "little to no" for RL4p, RLFQ3p, RL_meta and PearceHall.
- figs5-per-animal
Dopamine results are qualitatively consistent across animals; per-animal regressions on switch trials resemble the Bayesian model predictions in all but one animal, and switch-stay differences in outcome responses are consistent across animals.
All but one animal showed a strong qualitative resemblance to Bayesian model predictions for Switch trials
S5 Fig legendHow it will be judged
Per animal, OLS with the fixed part of the fig4i-lmer-switch formula. Read as resemblance when the R_chosen:Switch coefficient is negative and R_unchosen:Switch is positive, the sign pattern BRLfwr and BIfp predict. Reproduced when 4 of the 5 animals show it and, in at least 4 animals, mean DA-PT of rewarded switch trials exceeds rewarded stay trials and unrewarded stay trials are below unrewarded switch trials (S5 Fig A); partially reproduced when 3 animals show the regression pattern.
- reward-switch-reward-count
Only 94 trials start a reward, switch, reward sequence, out of 30,707 trials.
Yet in our dataset, there are only 94 trials initiating such sequences out of 30,707 trials in total
Discussion, paragraph 5How it will be judged
A trial rewarded at one port followed, in the same session, by a choice of the other port that is rewarded. Readings of the trial set (a) the 34 sessions with valid recordings (20,952 is about 68% of 30,707, which fits this), (b) all recorded sessions, (c) all 70 sessions. Reproduced when, under one reading, both counts are within tolerance.
- figs4-lmer-feature-sets
In cross-validated R2 of dopamine regressions, the 3R+Sw feature set (Reward, R_chosen and R_unchosen, each split by switch and stay trials) captures a similar level of variance to L0:R, the standard set of trial-back choice and reward interactions; 4 lags were used at the expense of about 1% of variance explained.
Together we found that 3R+Sw was enough to capture similar levels of variance compared to L0:R. We used 4 lags because it allows us to capture a relatively high amount of past outcome levels, with only the expense of 1% variance explained
S4 Fig legend; Methods, Linear mixed effects modelingHow it will be judged
Cleaned trials, DA-PT as the output (the legend keeps dopamine as the output variable). Feature sets from the S4 Fig legend, L0:R (choice, reward and their interaction for each trial back) and 3R+Sw (the fixed part of the fig4i-lmer-switch formula). Not given (the regression type, the cross-validation scheme, the lags compared, and what the 1% is measured against); readings (a) OLS with the 70-30 splits of fig4f-bayesian-rpe-fit, (b) the LMER of fig4i-lmer-switch with the same splits; lags 1 to 8. Read "similar" as a cross-validated R2 of 3R+Sw at 4 lags within 0.01 of L0:R at 4 lags, and the 1% as the R2 of 3R+Sw at 4 lags within 0.01 of its best over lags, or of L0:R's best. Reproduced when both hold under every reading; partially reproduced when the first holds. All R2 values are shown.
Pipeline
- Paper
- done (1 of 1)
- Claims
- done
- Claims check
- done (ready)
- Attempt
- todo
- Questions
- none
- Verified
- none