Aberrant computational mechanisms of social learning and decision-making in schizophrenia and borderline personality disorder.

Psychiatric disorders are ubiquitously characterized by debilitating social impairments. These difficulties are thought to emerge from aberrant social inference. In order to elucidate the underlying computational mechanisms, patients diagnosed with major depressive disorder (N = 29), schizophrenia (N = 31), and borderline personality disorder (N = 31) as well as healthy controls (N = 34) performed a probabilistic reward learning task in which participants could learn from social and non-social information. Patients with schizophrenia and borderline personality disorder performed more poorly on the task than healthy controls and patients with major depressive disorder. Broken down by domain, borderline personality disorder patients performed better in the social compared to the non-social domain. In contrast, controls and major depressive disorder patients showed the opposite pattern and schizophrenia patients showed no difference between domains. In effect, borderline personality disorder patients gave up a possible overall performance advantage by concentrating their learning in the social at the expense of the non-social domain. We used computational modeling to assess learning and decision-making parameters estimated for each participant from their behavior. This enabled additional insights into the underlying learning and decision-making mechanisms. Patients with borderline personality disorder showed slower learning from social and non-social information and an exaggerated sensitivity to changes in environmental volatility, both in the non-social and the social domain, but more so in the latter. Regarding decision-making the modeling revealed that compared to controls and major depression patients, patients with borderline personality disorder and schizophrenia showed a stronger reliance on social relative to non-social information when making choices. Depressed patients did not differ significantly from controls in this respect. Overall, our results are consistent with the notion of a general interpersonal hypersensitivity in borderline personality disorder and schizophrenia based on a shared computational mechanism characterized by an over-reliance on beliefs about others in making decisions and by an exaggerated need to make sense of others during learning specifically in borderline personality disorder.

Introduction

Impairments in social cognition are frequently experienced by people suffering from a psychiatric disorder. For instance, patients with major depressive disorder (MDD) and schizophrenia (SCZ) show a reduction in (social) reward sensitivity and motivation to engage in social interactions [1-5]. Despite high levels of social anhedonia, patients with SCZ show a tendency to over-interpret the meaning of social signals [6]. Individuals with borderline personality disorder (BPD) suffer from rapidly changing beliefs about others that polarise between approach and rejection [7]. Together, these impairments are associated with aberrant inferences/beliefs about oneself and the social environment.

In computational terms, the emergence of aberrant inference can be ascribed to an impaired ability to adjust learning in response to environmental changes [8]. Bayesian learning models allow for a parsimonious algorithmic description of changes in beliefs relevant for accurate inference: belief updates can be written as a surprise signal (prediction error) weighted by a learning rate [9]. The learning rate depends on the ratio between the precision of the sensory data and the precision of the prior belief [10,11]. Whereas healthy participants increase their learning rate more strongly in volatile compared to stable environments [12,13], patients with autism do so less owing to an over-estimation of environmental volatility [8]. Impairments in the estimation of environmental volatility have also been studied as a mechanism for psychosis and SCZ [14-17] as well as MDD [18]. One recent study found that, unlike healthy controls, participants with BPD did not show an increase in learning when social and reward contingencies became volatile [19]. The authors suggested that this might be due to higher expected baseline volatility in participants with BPD. However, the computational model employed in that study did not explicitly model beliefs about volatility.

Adopting previous suggestions of aberrant volatility learning in psychiatric disorders and its role in impaired probability learning, the current study employed Bayesian hierarchical modeling to investigate probabilistic social inference in a volatile context across three major psychiatric disorders, which have previously been associated with social impairments: MDD, SCZ and BPD. Here, the current study investigated whether volatility and probability learning is equally affected when inferring on the hidden states of non-social and social outcomes across the three different disorders. We further asked whether aberrant social learning and decision-making were associated with differences in social anhedonia.

To this end, we adopted a probabilistic reward learning task (introduced in [20]), in which participants could learn from two types of information: non-social and social information. In order to probe the spontaneous rather than explicitly instructed use of social information as in previous social learning studies [12,13,21,22], we did not explicitly tell participants to learn about the social information. We used the hierarchical Gaussian filter (HGF; [10,11]) to obtain a profile of each participant’s particular way of updating beliefs when receiving social and non-social information while making decisions in a volatile context. The HGF is a generic hierarchical Bayesian inference model for volatile environments with parameters that reflect individual variations in cognitive style. We went beyond other recent computational psychiatry studies using the HGF (e.g., [8,23-27]) in that we used two parallel HGF hierarchies for social and non-social aspects of the environment (cf. [20,29]). Our modeling framework was specifically designed also to quantify the relative weight participants afforded their beliefs about the predictive value of social compared to non-social information in decision-making.

Materials and methods

Ethics statement

All participants were naïve to the purpose of the experiment and provided written informed consent to take part in the study after a written and verbal explanation of the study procedure. The study was in line with the Declaration of Helsinki and approval for the experimental protocol was granted by the local ethics committee of the Medical Faculty of the Ludwig-Maximilians University of Munich.

Participants

Patients were recruited for the present study after an independent and experienced clinician diagnosed them using ICD-10 criteria for 1) a depressive episode (F32), schizophrenia (F20.0) and emotionally unstable personality disorder (F60.3). HC and patients with MDD were recruited through the Max Planck Institute of Psychiatry. Patients with SCZ were recruited at the Department of Psychiatry and Psychotherapy at the University Hospital Munich. Patients with BPD were recruited at the kbo-Isar-Amper-Klinikum in Haar, Munich. Participants were chosen prior to analysis such that groups were matched for age (χ2 = 5.302, P = 0.151; Kruskal-Wallis one-way non-parametric ANOVA because of difference in age variance between groups, see S1 Table). Exclusion criteria were a history of neurological disease or injury, reported substance abuse at the time of the investigation, a history of electroconvulsive therapy, and diagnoses of comorbid personality disorder in the case of MDD and SCZ. Furthermore, 9 participants had to be excluded from the analysis due to one of the following reasons: unsaved data due to technical problems (1 HC, 2 BPD). Prior participation in another study which involved the same paradigm (1 HC), always picking the card with the higher reward value (1 HC), either following (1 SCZ) or going against (1 BPD) the gaze on more than 95% of trials (indicating a learning-free strategy), interruption of the task (1 SCZ), change to the diagnosis following study participation (1 MDD). The final sample consisted of 31 HC, 28 MDD, 29 SCZ and 28 BPD. We additionally acquired psychometric data (S1 Table) to further characterize the participants: All patients were asked to fill out questionnaires measuring autistic traits with the autism spectrum quotient (AQ [30]) and social anhedonia symptoms with the Anticipatory and Consummatory Interpersonal Pleasure Scale (ACIPS; [31]). We additionally assessed positive and negative symptoms using the Positive and Negative Syndrome Scale (PANSS [32]) and mood symptoms using the Calgary Depression Scale for Schizophrenia (CDSS [33]) in patients with SCZ. To assess the severity of Borderline Personality Disorder we used the short version of the Borderline Symptom List (BSL-23 [34]). Additional questionnaires were employed but analyzed within the scope of a different study and therefore not presented here. Demographic data as well as details regarding the medication can be seen in S2 Table.

Experimental paradigm and procedure

After giving informed consent, participants were seated in front of a computer screen in a quiet room where they received the task instructions. In the same probabilistic learning task as in [20,29], participants were asked to choose between one of two cards (blue or green) in order to maximize their score which was converted into a monetary reward (1–6 €) that was added to participants’ compensation at the end of the task. An animated face was displayed between the cards, which first gazed down, then up towards the participant, before it shifted its gaze towards one of the cards (Fig 1A). The blue and green card appeared randomly on the left and right side from the face and participants responded using ‘a’ or ‘l’ on a German QWERTZ keyboard. When a response was logged within the allowed time (6000 ms), the chosen card was marked for 1000 ms until the outcome (correct: green check mark/wrong: red cross) was displayed for 1000 ms. When the correct card was chosen, the reward value (1–9) displayed on the card was added to the score. Participants were instructed that these values were not associated with the cards’ winning probabilities, but that they might want to choose the card with the higher value if they were completely uncertain about the outcome. When the wrong card was chosen or participants failed to choose a card in the allotted time, the score remained unchanged. Participants were told that the cards had winning probabilities that changed in the course of the experiment but they were not informed about the systematic association between the face animation’s gaze and the trial outcome. Specifically, they were not told that the probability with which the face animation pointed towards the winning card on a given trial varied systematically throughout the task according to the schedule given in Fig 1B. Instead, we simply told participants that the face was integrated into the task to make it more interesting. The probabilistic schedules for social and non-social information were independent from each other in order to estimate participant-specific learning rates separately for both types of information. In the first half of the experiment (trials 1–60), the card winning probabilities were stable, whereas in the second half (trials 61–120) they changed (volatile phase). The social cue had a stable contingency during trials 1–30 and trials 71–120, whereas contingency was volatile during trials 31–70. We used two types of schedules for the social cue which were each presented to half of the participants. In one schedule (depicted in Fig 1B), the probability of the social cue looking towards the winning card was 73% in the first stable phase (trial 1–30) and therefore started as congruent to the winning card (congruent-first). The second probability schedule was flipped, so that the probability of the social cue looking towards the winning card was 27% in the first stable phase (incongruent-first). In total, 15 control participants received the congruent-first schedule, 15 participants with MDD, 14 with SCZ and 15 with BPD. Positions of the cards on the screen (blue left or right) were determined randomly. The task was programmed and presented with PsyToolkit [35].

Fig 1

Task design and computational decision and inference model.

(A) Participants were asked to make a choice between blue and green cards after the gray shading on the colored rectangles (cards) had disappeared (i.e., 750 ms after the face shifted its gaze towards one of the cards). After a delay phase, the outcome was presented (correct/wrong). If the choice was correct, the reward amount (number on the chosen card) was added to a cumulative score. The task consisted of 120 trials. (B) Probability schedules from which outcomes were drawn. Volatile phases are marked in grey. (C) Posteriors are deterministic functions of predictions and outcomes. Predictions in turn are deterministic functions of the posteriors of previous trials. Decisions y(t) are probabilistically determined by predictions and decision model parameters ζ and β. Deterministic quantities are presented as boxes and probabilistic quantities in circles.

Computational modeling

Observing the observer

We modeled behavior in the ‘observing the observer’ (OTO) framework [36,37]. This entails a response model, which probabilistically predicts a participant’s choices based on their inferred beliefs, and a perceptual model, on which the response model depends because it describes the trajectories of participants’ inferred beliefs based on experimental inputs. The OTO framework is conceptually very similar to the idea of inverse reinforcement learning [38].

Perceptual models

We used three different perceptual models in order to make inferences on the most likely mechanisms of learning in our paradigm. We used a Bayesian HGF as well as two non-Bayesian learning models, the Sutton K1 model [39] and a Rescorla Wagner model [40]. All three were implemented in the HGF toolbox, such that they perform parallel learning about the social (predictive value of gaze) and non-social (predictive value of card color) aspects of the task environment.

While the Rescorla Wagner learning model assumes fixed domain-specific learning rates for learning the social and non-social information, the Sutton K1 model assumes variable learning rates that are scaled by recent predicition errors. In contrast, the HGF takes into accout that beliefs have different degrees of uncertainty, scaling the learning rate dynamically as a function of uncertainty. In addition, the HGF assumes hierchical learning of different aspects of the environment. On the lowest level of the hierarchy, agents learn about concrete events (i.e. stimuli), whereas at higher levels of the hierarchy, agents learn about more abstract features of the environment, such as probabilistic associations between stimuli and how these change in time (i.e. volatility). Learning at every level is driven by a ratio between the precision of the input (from the level below) and the precision of prior beliefs.

The HGF is an inference model resulting from the inversion of a generative model in which states of the world are coupled in a three-level hierarchy: At the lowest level of the generative model, and represent the two inputs in a binary form (social cue: 1 = correct, 0 = incorrect; card outcome: 1 = blue wins, 0 = green wins). Level and represent the tendency of the gaze to be correct and the tendency of the blue card to win. State and evolve as first-order autoregressive (AR(1)) processes with a step size determined by the state at the third level. Level and represent the log-volatility of the two tendencies and also evolve as first-order autoregressive (AR(1)) processes. The probabilities of and are the logistic sigmoid transformations of and (Eq 1).

Participants’ responses y were coded with respect to the congruency with the ‘advice’ (1 = follow; 0 = not follow) and were used to invert the model in order to infer the belief trajectories at all three levels i = 1,2,3.

On every trial k, the beliefs (and their precisions ) about the environmental states at the i -th level are updated via prediction errors from the level below weighted by a precision ratio (Eqs 2 and 3). This means that belief updates are larger (due to higher precision weights) when the precision of the posterior belief (π2( or ) is low and the precision of the prediction is high. Consequently, prediction errors are weighted more during phases of high volatility (cf. S1 Fig, panel C, dotted blue trajectory). For the analysis, we used (Eq 4), which is a transformation of (Eq 3) (cf. [41], supplementary material) that corrects for the sigmoid mapping between first and second level, effectively making an uncertainty (inverse precision) measure for first-level beliefs.

Participant-specific parameters ω2 and ω2 represent the learning rates at the second level, i.e. the speed at which association strengths change. Correspondingly, ω3 and ω3 represent the learning rates of the volatilities.

Response models

In the response model a combined belief b( (Eq 6) was mapped onto decisions, which resulted from a combination of both the inferred prediction that the face animation’s gaze will go to the winning card and the inferred prediction that the color of the card that the gaze went to would win (see example in S1 Fig). The inferred prediction and were weighted by and (Eqs 7 and 8), which are functions of the respective precisions ( and , Eqs 9 and 10). The precisions (Eqs 9 and 10) represent the inverse variances of a Bernoulli distribution of and .

The constant parameter ζ represents the weight on the precision of the social prediction compared to the precision of the non-social prediction (Eq 7). In other words, this parameter describes the propensity to weight the social over the non-social information. We investigated the effect of varying the social weighting factor ζ, by simulating the combined belief b( (Eq 6) of agents with same perceptual parameters (fixed at prior values as depicted in S3 Table) but different ζ values.

In the response model, we used the combined belief b( (Eq 6) in a logistic sigmoid (softmax) function to model the probability (Eq 11). In this function, the belief was weighted by the predicted reward of the card when the advice is taken r or not r (Eq 11). We took account of possible subject-specific non-linear distortions in weighting the expected reward by using a weighted average (parameter η) of linear and logarithmic weighting of expected reward.

The mapping of beliefs onto actions varied as a function of the inverse decision temperature γ(, where large γ( implied a high alignment between belief and choice (low decision noise) and a smaller γ( a low alignment between belief and choice (high decision noise). Our four different response models varied in terms of how γ( was defined. In response model 1, γ( was a combination of the log-volatility of the third level for both cues combined with constant participant-specific decision noise β (Eq 12). In response model 2, γ( was a combination of the log-volatility of the third level for the social cue and participant-specific decision noise (Eq 13) and in response model 3, γ( was a combination of the log-volatility of the third level for the non-social cue and participant-specific decision noise (Eq 14). In model 4, γ( only included the participant-specific decision noise (Eq 15).

We used the HGF toolbox, version 4.1, which is part of the software package TAPAS (https://translationalneuromodeling.github.io/tapas) for parameter estimation. We fitted six alternate combinations of perceptual and response models, which were subjected to random-effects Bayesian Model Selection [42,43] (spm_BMS in SPM12; http://www.fil.ion.ucl.uk/spm). The HGF was combined with all four response models. The non-hierarchical models were combined with response model 4 only, owing to the lack of third-level belief trajectories. Details of the prior settings of all models can be seen in S3 Table.

We additionally included two non-learning models, which assume that participants repeat the actions that lead to reward and switch the strategy immediately after loss (Win-Stay-Lose-Shift) and a model that assumes random responding throughout the task (random-responding). These models were adapted from [44] and implemented in the HGF toolbox to calculate the log model evidence.

Model comparison and validity

The log model evidence (LME) for each participant and each model were subjected to Bayesian Model Selection [43,44] (spm_BMS in SPM12). This procedure estimates the expected posterior probabilities (EXP_P), i.e. the posterior probability of the prevalence of each model in the population, the exceedance probability (XP), i.e. the probability that a given model outperforms all others in the comparison, and the more conservative protected exceedance probability (PXP), which additionally considers the possibility that all models are equally good. We additionally performed within-participant model comparisons to identify the model with the highest LME for each participant (cf. S4 Table). This turned up 16 subjects where random responding had a higher model evidence than the overall winning model. Excluding these participants did not change the pattern of results (cf. S7–S9 Tables).

Posterior predictive validity of model parameters and Parameter Recovery

To test the adequacy of the model, we simulated responses based on the estimated parameters from the winning model for each participant 10 times, resulting in 1160 simulations. As for the real data, we then calculated percentage of high reward probability choices for the two cues and the two phases, subjecting them to the same ANOVA that was performed with the real behavioral data. In this way, we checked whether the simulated responses produce the same group differences in response accuracy. In a second step, simulated responses were again used to invert the winning model to check whether model parameters could be recovered. For each subject, we calculated the average parameter values estimated from the simulated data and correlated (Pearson’s correlations) them with the values estimated from the real data, which served as input for the simulation.

Regression-based choice sequence analysis

We ran a regression based choice sequence analysis in order to investigate adaptation to environmental volatility without needing to fit a learning model. According to the HGF, agents should increase their learning rate in more volatile environments, giving more weight to recent outcomes. In more stable environments, agents should adopt lower learning rates, affording less weight to recent outcomes in order to better filter out noise. To test this in a model-agnostic way, we implemented two general linear models (GLMs) with the responses y as outcome variables for the ‘card’ and ‘gaze’ space respectively (card: 1 = blue taken; 0 = green taken; gaze: 1 = follow; 0 = not follow). For each GLM, we included as predictors the outcomes of the past 5 trials (t-1:t-5: 1 = blue correct; 0 = green correct) and (t-1:t-5: 1 = gaze correct; 0 = gaze incorrect) as well as two predictors for the expected reward (1–9) for either card winning. We ran these GLMs for the stable and volatile phases separately. The slopes of the coefficient estimates of the past outcome predictors were taken as a model-agnostic readout of the ‘learning rate’, indiciating the degree to which more recent information is weighted. A t-test was applied to compare the difference between these slopes during stable and volatile phases of the task.

Statistical analysis

Performance (% correct responses) was subjected to a one-way ANOVA with group (HC vs. MDD vs. SCZ vs. BPD) and schedule (congruent first vs. incongruent first) as between-subject factors.

In order to understand task performance in a domain-specific way, we additionally calculated response accuracy based on the ground truth reward probability (high probability choices) of both cues (non-social and social) and both phases (stable and volatile). The proportion of high probability choices was subjected to a mixed ANOVA with Cue Type (Non-Social vs. Social) and Phase (Stable vs. Volatile) as within-subject factors and Group (HC vs. MDD vs. SCZ vs. BPD) and schedule (congruent first vs. incongruent first) as between-subject factors. Advice taking (advice followed or not on a given trial) was subjected to a mixed ANOVA with social accuracy (high vs. low) and schedule stability (stable vs. volatile) as within-subject factors. Group (HC vs. MDD vs. SCZ vs. BPD) and schedule (congruent first vs. incongruent first) were included as between-subject factors.

Mean precision weights on the second and third level (q(ψ2) and ψ3) separately entered two mixed ANOVAs as dependent variables with schedule stability as a within-subject factor (stable vs.volatile), information type as within participants factor (social vs. non-social). The group (HC vs. MDD vs. SCZ vs. BPD) and schedule (congruent first vs. incongruent first) were between subject factors.

We subjected the posterior estimate for ζ to a one-way ANOVA with group (HC vs. MDD vs. SCZ vs. BPD) as between-subject factor and schedule (congruent first vs. incongruent first) as a covariate.

We hypothesized that social anhedonia (measured by the Anticipatory and Consummatory Interpersonal Pleasure Scale, ACIPS) would be associated with a reduction in learning in the social domain. To test this, we first performed a one-way ANOVA with ACIPS scores as dependent variable and group as the factor (HC vs. MDD vs. SCZ vs. BPD) followed by a multivariate regression with ACIPS as dependent variable and the social learning rates ω2 and the weighting factor ζ as predictors of social learning and decision making. The group factor (HC vs. MDD vs. SCZ vs. BPD) was entered as covariate. This analysis was done for all participants who completed the ACIPS questionnaire (n = 106 of ntotal = 116).

All ANOVA post hoc t tests were Bonferroni-corrected for multiple comparisons. All p-values are two-tailed with a significance threshold of p < .05. Statistical tests were performed using JASP (Version 0.9 2.0; https://jasp-stats.org/) or Matlab (Version 2018b; https://mathworks.com).

Results

Behavior

There was a significant difference between the groups in the overall performance, i.e. % of rewarded responses (F(3,108) = 7.504, p<0.001, η2 = 0.167): Post-hoc comparisons showed that both patients with SCZ and BPD performed significantly worse compared to HC and patients with MDD (SCZ–HC t = 3.781, pbonf = 0.002, d = 0.994, SCZ–MDD t = 2.817, pbonf = 0.035, d = 0.745, BPD–HC t = 3.732, pbonf = 0.002, d = 0.979, BPD–MDD t = 2.78, pbonf = 0.038, d = 0.732). There was no significant difference in performance between patients with BPD and SCZ (t = -0.01, pbonf = 1.000, d = -0.003) nor between HC and patients with MDD (t = 0.88, pbonf = 1.000, d = 0.240). Performance was not significantly affected by the schedule order (congruent first vs. incongruent first; F(1,108) = 0.027, p = 0.870, η2 = 0) or its interaction with the patient groups (F(3,108) = 1.302, p = 0.278, η2 = 0.029).

There was a significant difference between the groups in the overall proportion of trials where the better option was chosen, i.e. proportion of correct choices based on the ground truth reward probability of both cues (F(3,108) = 4.945, p = 0.003, η2 = 0.116, Fig 2 and S6 Table). Post-hoc comparisons showed that both SCZ and BPD patients performed significantly worse compared to HC (SCZ–HC t = 3.373, pbonf = 0.006, d = 0.313, BPD–HC t = 3.227, pbonf = 0.01, d = 0.3) across domains (i.e., cue types) (Fig 2 and S6 Table). Overall, response accuracy did not significantly differ between cue types (F(1,108) = 2.577, p = 0.111, η2 = 0.019), but there was a significant Cue Type × Group interaction F(3,108) = 4.820, p = 0.003, η2 = 0.108), with HC (Post-hoc: t(30) = -2.157, pbonf = 0.039, d = -0.387) and MDD patients (Post-hoc: t(27) = -2.181, pbonf = 0.038, d = -0.412) showing higher response accuracy with regard to the non-social cue, whereas BPD patients showed the opposite pattern (Post-hoc: t(27) = 2.058, pbonf = 0.049, d = 0.389).

Fig 2

Proportion of trials where the better option was chosen.

The better option was defined to be the one that according to the ground truth probability schedule was more likely to be rewarded. Importantly, ‘better’ choice according to color and according to social cue could be different in the same trial. Patients with SCZ and BPD showed poorer response quality in the task compared to HC. A Group x Cue interaction analysis showed that response quality was higher for the non-social cue for HC and MDD patients, whereas BPD patients showed the opposite pattern. Means are plotted with boxes marking 95% confidence intervals and vertical lines showing standard deviations. See also S6 Table.

With regard to advice taking behavior during the different phases of the social schedule, we found a main effect of social accuracy (F(1,108) = 227.935, p<0.001) whereby participants followed the gaze more during phases of high accuracy compared to phases of low accuracy (t = 14.94, pbonf <0.001) (Fig 3). Advice taking was not significantly affected by the schedule stability (F(1,108) = 0.503, p = 0.480), indicating that advice taking did not differ between stable and volatile phases. Advice taking was not significantly affected by an interaction between accuracy of the social information and Group (F(3,108) = 2.222, p = 0.09), or by an interaction between social accuracy, schedule stability and Group (F(3,108) = 1.47, p = 0.227).

Fig 3

Behavioral results with regard to advice taking.

All participants followed the advice significantly more during phases of high compared to low accuracy. The descriptive data shows a trend of BPD patients to follow the advice more during volatile phases of low accuracy (right most plot) but the interaction was not significant. Means are plotted with boxes marking 95% confidence intervals and vertical lines showing standard deviations.

Bayesian model comparison & validity

Model comparison showed that the HGF including subject specific decision noise as well as the volatility estimate and outperformed the other HGF models, the Rescorla Wagner and Sutton-K1 models with subject specific decision noise only as well as the WSLS and random responding models (PXP = 0.173; XP = 0.914). See Table 1 for further details and S4 Table for mean posterior parameter estimates. We repeated all modeling-based analyses with a reduced sample where all subjects were excluded for whom the random responding model outperformed the others (S4 Table). This did not lead to any qualitative change in results (S7–S9 Tables). We therefore report the results for the full sample here.

Table 1

Bayesian model comparison results.

BMS	Model 1	Model 2	Model 3	Model 4	Model 5	Model 6	Model 7	Model 8
EXP_R	0.304	0.181	0.073	0.019	0.066	0.210	0.072	0.076
XP	0.914	0.019	0	0	0	0.068	0	0
PXP	0.173	0.119	0.117	0.117	0.117	0.122	0.117	0.117

Posterior model probabilities (EXP_R), Exceedance Probabilities (XP) and Protected Exceedance Probabilities (PXP). Model 1 refers to the HGF combined with response model 1, Model 2 refers to the HGF combined with response model 2, Model 3 refers to the HGF combined with response model 3, Model 4 refers to the HGF combined with response model 4, Model 5 refers to the Sutton K-1 Model combined with response model 4, Model 6 refers to the Rescorla Wagner Model combined with response model 4. Model 7 refers to the WSLS Model and Model 8 to the random Model.

Posterior predictive validity of model parameters and parameter recovery

Subjecting the simulated behavioral readout of % of high probability choices to the same ANOVA as performed with the real behavioral data, we found that the model was able to reproduce the group differences that were observed in the real behavioral data (cf. S6 Table for comparison). The ANOVA of the simulated data showed that for response accuracy, as in the real data, there was a main effect of Group (F(3,108) = 6.755, p = 0.001, η2 = 0.149) with post-hoc t tests revealing significantly lower response accuracy for SCZ and BPD patients compared to HC (SCZ–HC t = 4.04, pbonf<0.001, d = 0.375, BPD–HC t = 3.675, pbonf = 0.002, d = 0.341) and a Group × Cue Type interaction (F(3,108) = 10.981, p<0.001, η2 = 0.219) showing that HC and MDD patients showed higher response accuracy with regard to the non-social cue, whereas BPD patients showed the opposite pattern (cf. S6 Table and S2 Fig for all results).

Parameter recovery showed that all parameters could be recovered well with the exception of ω3 (cf. S4 Fig).

Dynamic learning rates–second level

For the averaged precision weights (i.e., dynamic learning rates) for learning about the social q(ψ2) and non-social q(ψ2), we found a main effect of task phase (F(1,108) = 24.868, p<0.001, η2 = 0.182), showing that q(ψ2) is higher in volatile compared to the stable phases (t = -5.147, pbonf <0.001, d = -0.478) (Fig 4A). In the model-agnostic regression analysis, we observed increased slopes of beta weights over time delays in the volatile vs. stable phase (S5 Fig) indicating higher weighting of most recent trials, which concurs with the increased learning rates in the volatile phase in the HGF. While this effect was observable for both cues, the difference in the model-agnostic analysis was significant only for the card cue, perhaps because this analysis does not consider individual variations in cue use.

Fig 4

Results for mixed ANOVA using precision weights for updating beliefs about social and non-social contingency and volatility.

(A) Precision weights q(ψ2) and (B) precision weights ψ3. Overall, q(ψ2) and ψ3 increase when transitioning from stable to volatile phase. Patients with BPD show reduced overall q(ψ2). At same time, patients with BPD show higher ψ3 compared to the other groups and a more pronounced increase in response to volatility. Bars indicate SEM. See also S1 Fig, S7 Table and S8 Table for all results.

There was no significant interaction between Phase and Information Type, which indicates that q(ψ2) increases similarly during social and non-social volatility (F(1,108) = 0.131, p = 0.718, η2 = 0.001). There was a significant main effect of group (F(3,108) = 3.557, p = 0.017, η2 = 0.088), and the post-hoc t-tests revealed that participants with BPD showed significantly lower precision weights on the second level compared to HC (t = 3.101, pbonf = 0.015, d = 0.288). The difference in q(ψ2) between groups was not affected by Information type (F(3,108) = 1.038, p = 0.379, η2 = 0.027) or its interaction with Phase (F(3,108) = 0.940, p = 0.424, η2 = 0.025) (see S7 Table for all results and for results with the reduced sample).

Dynamic learning rates–third level

We found a main effect of task phase on precision weights at the third level (F(1,108) = 116.206, p<0.001, η2 = 0.462), showing that ψ3 is higher in volatile compared to the stable phases (t = -9.784, pbonf <0.001, d = -0.908) (Fig 4B). There was a significant main effect of group (F(3,108) = 6.530, p<0.001, η2 = 0.141), and post-hoc t-tests showed that participants with BPD showed significantly higher precision weights at the third level compared to all other groups (BPD–HC t = -4.204, pbonf < .001, d = -0.390; BPD–MDD t = -3.199, pbonf = .011, d = -0.297; BPD–SCZ t = -3.055, pbonf = .017, d = -0.284). In addition, there was a significant Phase × Group interaction (F(3,108) = 5.962, p<0.001, η2 = 0.071) showing that participants with BPD increase their precision weights for both modalities significantly more compared to the other groups when volatility increases. There was a trend of BPD patients showing stronger increases in ψ3 in response to social compared to non social volatility (F(3,108) = 2.620, p = 0.055, η2 = 0.065). The analysis also revealed that ψ3 were affected by the order of schedule (F(1,108) = 5.008, p = 0.027, η2 = 0.036), with ψ3 higher for participants receiving the incongruent-first schedule (i.e gaze starts of being highly misleading) compared to the congruent-first schedule (i.e gaze starts of being highly helpful). This effect was not modulated by Group (F(3,108) = 2.067, p = 0.109, η2 = 0.045) (see S8 Table for all results and for results with the reduced sample).

Social weighting

The parameter ζ was a measure of the weight given to the social prediction relative to the learned non-social prediction (cf. Fig 5B & 5C for simulation results). Since ζ was restricted to the positive domain, estimate distributions were analyzed log-space, where they were less skewed. We found significant group differences in log(ζ) (F(3,108) = 5.893, p>0.001, η2 = 0.130 (Fig 5A). Both patients with BPD and patients with SCZ showed significantly higher ζ estimates compared to controls (BPD: t = -3.416, pbonf = 0.005; d = -0.847, SCZ: t = -2.855, pbonf = 0.031, d = -0.691) but only patients with BPD differed significantly from participants with MDD (BPD: t = -3.003, pbonf = 0.02, d = -0.818; SCZ: t = -2.451, pbonf = 0.095, d = -0.650). Patients with MDD did not show any significant differences compared to controls (t = -0.335, pbonf = 1, d = -0.095). There was a significant main effect of schedule (F(1,108) = 8.259, p = 0.005, η2 = 0.061), showing that participants receiving the congruency-first schedule had higher ζ compared to participants receiving the incongruency-first schedule (t = -2.874, pbonf = 0.005, d = -0.505). There was no significant interaction between Group and Schedule (F(3,108) = 0.807, p = 0.493, η2 = 0.018). To ensure that the shared mechanism of social over-weighting in BPD and SCZ was not confounded with medication or education status (cf. S1 and S2 Tables), we subjected ζ to an ANCOVA with ζ and Chlorpromazine Equivalence Units and Years of School as covariates. The effects were robust to this (cf. S9 Table for results of full and reduced sample).

Fig 5

Social weighting factor log(ζ).

(A) Patients with BPD gave the social information significantly more weight compared to HC and patients with MDD. Patients with SCZ also had higher ζ compared to HC. Boxes mark 95% confidence intervals and vertical lines standard deviations. (B), Simulation results show the impact of varying weighting factor log(ζ) on combined belief b(t) (see methods Eq 1). The combined belief b(t) was simulated for agents with same perceptual parameters but different ζ values (highest values (log(ζ) = 5) coded in dark blue, lowest values (log(ζ) = -5) in green). (B) shows that the combined belief b(t) of agents with high ζ values is aligned with the social input structure (blue dots) whereas these agents show a stochastic belief structure with regard to the non-social input structure (green dots) in Panel C. Conversely, agents with low ζ values show a belief structure closely aligned to the non-social input structure (C), and a stochastic belief structure with regard to the social input (Panel B). The grey lines represent the ground truth of the respective probability schedules. See also S9 Table for all results.

Social Anhedonia

There was a significant difference in the interpersonal pleasure (ACIPS) ratings between the groups (F(3,103) = 5.719, p < .001) (cf. S1 Table). Post hoc t tests revealed that HC showed significantly higher ACIPS scores compared to patients with MDD (t = 3.088, pbonf = .016) and with BPD (t = 3.802, pbonf = .001) but not with SCZ (t = 1.833, pbonf = .418). No significant differences were observed between patients with MDD and SCZ (t = -1.322, pbonf = 1), patients with MDD and BPD (t = 0.596, pbonf = 1), nor patients with SCZ and BPD (t = 1.978, pbonf = 0.304). The multivariate regression using log(ζ) and social learning rate ω2 as predictors for ACIPS scores did not show any significant results (R2 = 0.009, F(2,106) = 0.477, p = 0.622).

Discussion

This study aimed to improve our understanding of the computational mechanisms that underlie the profound interpersonal difficulties in major psychiatric disorders. To achieve this, we used a probabilistic learning task in conjunction with hierarchical Bayesian modeling transdiagnostically in patients with MDD, SCZ, BPD, and healthy controls. The task required participants to perform association learning about non-social contingencies in the presence of a social cue. This allowed us to characterize and quantify the computational aspects of aberrant social inference and decision-making at an individual level. We found that patients with SCZ and BPD showed significantly poorer performance compared to HC and patients with MDD. Patients with MDD performed comparably well to HC. In addition, patients with BPD showed greater response accuracy in the social compared to the non-social domain during the stable phase, while HC and MDD patients showed the opposite pattern. This is particularly remarkable in light of their overall poorer performance. In effect, BPD patients gave up a possible reward advantage by concentrating their learning efforts disproportionately in the social domain. In addition, we found a tendency in BPD patients to follow the gaze more during volatile phases of low accuracy compared to MDD patients, which however did not reach statistical significance (Fig 3).

These findings raise the question which mechanisms underlie these patterns of behavior. In particular, they call for an investigation of the learning and decision-making mechanisms which give rise to them. Here, computational modeling enabled insights into how beliefs are updated and how these beliefs are translated into decisions: With regard to learning, we found that BPD patients showed increased precision weighting of prediction errors when learning about volatility in both non-social and social information and a tendency for even higher precision weights when learning about social compared to non-social volatility. While volatility learning rates (ψ3) were increased in BPD compared to HC, contingency learning rates (ψ2) were reduced compared to HC both in the social and non-social domain. This accords with a previous finding of blunted social and non-social learning in BPD [19], which was conjectured to result from aberrant volatility beliefs, causing an impairment at detecting contingency changes needed for accurate inference. Because our modeling approach was specifically designed to model beliefs about volatility, it allowed us to address this conjecture. Indeed, our data indicate that impaired contingency learning in BPD is associated with exaggerated learning about environmental volatility. A similar pattern has been observed in autism spectrum disorder (ASD) [8]. Aberrant volatility beliefs in BPD have been suggested to result from unpredictable early-life relationships [19]. However, this is a less likely explanation in ASD which is characterized as a pervasive developmental disorder. This points to a different origin of the mechanistic overlap between our findings and those of [8]. The commonality of aberrant volatility learning may explain the repeated finding of high autism quotient (AQ) values in BPD patients [8], which is confirmed in our sample (cf. S1 Table) and could be taken to suggest a partially shared mechanism of aberrant social inference in these disorders. A previous study from our group found that healthy participants scoring high on AQ showed a similar pattern to the present study’s BPD patients, in that they followed the gaze more than low-AQ participants during periods of low accuracy in volatile phases [20]. In that study, computational modeling showed that high AQ participants failed to use the social information to adapt the precision of their belief about the non-social cue. However, this was not found in any participant group in the present study.

This has been the second study demonstrating that aberrant learning in BPD not only concerns social, but also non-social information (cf. [19]). This suggests that aberrant learning occurs independent of domain, in line with previous findings that precision-weighted prediction errors are computed in similar brain regions, irrespective of domain [22,41].

Unlike previous studies on reward [45-47] or volatility [15] learning in SCZ and healthy subjects at risk for psychosis [16], we did not find significant differences between SCZ and HC in this regard. The same applies to MDD patients, where one possible explanation for this negative finding is the lack of punishment for incorrect choices in our task since recent findings converge on impaired aversive learning in depression (e.g. [48,49]).

With regard to decision-making, computational modeling revealed that SCZ and BPD patients both weighted their social-domain predictions more strongly than HC and MDD. This explains the lower performance of BPD and SCZ patients. Their stronger reliance on social cues compared to HC and MDD patients was detrimental because the social cue was more volatile than the non-social one (5 as opposed to 3 contingency changes).

The commonality of over-weighting social-domain predictions in SCZ and BPD patients suggests itself as the decision-making aspect of a general interpersonal hypersensitivity in both conditions [50]. This is also reflected in excessive, albeit inaccurate, mental state attributions (hypermentalizing) [51-53] that constitute a shared feature of BPD and SCZ [6,51-58].

Hypermentalizing is also a possible explanation for the findings of [59], where similar modeling as in the present study showed that healthy participants at the high end of the paranoia spectrum used similar weighting of social information irrespective of whether incorrect advice was framed to be intentional or not, while low-paranoia participants reduced their social weighting when negative advice was cued to be intentional. Furthermore, a study of healthy participants by our group [29] found that stronger weighting of social over non-social predictions during decision-making was associated with increased activity in the putamen and anterior insula. In future studies, it will be interesting to investigate the involvement of these regions in excessive social-weighting and hypermentalizing in BPD and SCZ. Also, a direct comparison of patients with BPD and autism could help to unveil shared mechanisms of aberrant social inference.

In addition to looking for learning and decision-making differences between the different diagnostic groups that were defined by traditional ICD-10 criteria, we adopted a transdiagnostic perspective to investigate the relation between computational mechanisms of social learning and decision-making with ACIPS, a self-report measure of social anhedonia. Previous studies have adopted such a dimensional approach in the general population and found that patterns in aversive learning mapped onto distinct symptoms of depression, social anxiety, and compulsivity [48,59,60]. We, however, could not find any transdiagnostic association of social anhedonia with computational parameters of social learning and decision-making in our data. A possible explanation for this negative finding is that we used scores from a single questionnaire whereas previous studies applied factor analysis (e.g. [60]) on all items from several questionnaires. In addition, these studies investigated larger samples compared to our study (>400 vs. 116) and therefore had more statistical power to detect signficant effects.

Limitations

We did not use a non-social cue (such as an arrow pointing to a card) as a control condition and therefore cannot fully rule out the possibility that the increased weighting of our social cue observed in BPD and SCZ reflects a more general rather than specifically social peculiarity in information processing. However, eye gaze is a very salient cue and in the paradigm, we aimed to accentuate the social quality of our cue by a clear period of eye contact with the participant before providing the cue.

A further limitation concerns the fact that most patients were in psychopharmacological treatment during data acquisition and had different degrees of disorder severity and chronicity. Furthermore, different patient groups were assessed in different clinical centers, and there was a gender imbalance in the SCZ and BPD groups.

Conclusion

By adopting a computational psychiatry approach [61-65] to data from an inference task with a social component, we show that BPD patients exhibit an aberrant pattern of learning rate adjustment when the environment becomes more volatile. Instead of quickly relearning changed contingencies, they show exaggerated volatility learning. While SCZ and MDD patients showed a tendency to the same pattern, they did not significantly differ from controls in this respect. We also show that BPD and SCZ patients rely more strongly than controls on social-domain beliefs relative to non-social-domain beliefs when making decisions even in a task where doing the opposite would have given them an advantage. Taken together, this shows that there are computational commonalities as well as differences between patient groups, which suggests some underlying mechanisms that may be shared across diagnoses. Since this approach allows for individually quantifying the severity of impairment at a mechanistic level, it has the potential to lead to diagnostic and prognostic advances. Furthermore, it points the way to possible targets for novel interventions which transcend traditional diagnostic boundaries.

Psychometric data of the participants.

All quantities given as Mean ± SD.

(DOCX)

Click here for additional data file.

Demographic data of the participants.

All quantities given as Mean ± SD.

(DOCX)

Click here for additional data file.

Prior configurations of perceptual and response model parameters.

Means and variances of Gaussian priors are given in the space in which the parameter was estimated (native, log, or logit).

(DOCX)

Click here for additional data file.

Within-subjects model comparison.

(DOCX)

Click here for additional data file.

Mean posterior estimates of learning model and decision model parameters estimated from winning model.

(DOCX)

Click here for additional data file.

Statistics for mixed ANOVA with response accuracy (% High probability choices) from real and simulated data for stable and volatile phases (Factor Phase) of social and non-social cue (Factor Cue Type) for all groups (Factor Group) and schedules (Factor Schedule).

The table shows the results for the real and simulated behavior.

(DOCX)

Click here for additional data file.

Statistics for mixed ANOVA with averaged q(ψ2) during stable and volatile phases (Factor Phase) of social and non-social cue (Factor Cue Type) for all groups (Factor Group) and schedules (Factor Schedule).

The table shows the results for the full and reduced sample.

(DOCX)

Click here for additional data file.

Statistics for mixed ANOVA with averaged ψ3 during stable and volatile phases (Factor Phase) of social and non-social cue (Factor Cue Type) for all groups (Factor Group) and schedules (Factor Schedule).

The table shows the results for the full and reduced sample.

(DOCX)

Click here for additional data file.

Statistics for ANCOVA with ζ and Chlorpromazine Equivalence Units and Years of School as covariate for full and reduced sample.

(DOCX)

Click here for additional data file.

Learning trajectories for one example participant.

A, Precisions ψ3 (red) and ψ3 (blue) that modulate the weight on B, prediction errors δ2 (red) and δ2 (blue). C, Precision weights ψ2 in red trajectory and q(ψ2) in red dotted trajectory. Precision weights ψ2 in blue trajectory and q(ψ2) in blue dotted trajectory. Precision weights modulate weight on D) prediction error δ1 (red) and δ1 (blue) signals. E, Dark red dots mark the input structure of the non-social information (blue correct = 1; green correct = 0) and the dotted red line represents the ground truth of this input structure. Light red dots mark the choices (blue card = 1; green card = 0). The red trajectory is the participant specific belief trajectory about the blue card to be correct that was estimated on the basis of the choices. E, The same logic applies to the social input and response structure in blue. The posterior parameter estimates for this particular participant were ω2 = -1.460, ω2 = -3.576, ω3 = -6.021, ω2 = -6.074, log(ζ) = -2.0243, log(β) = 2.207, logit(η) = 0.211.

(TIF)

Click here for additional data file.

Posterior Predictive Validity.

Simulated Behavior from the posterior estimates of all participants revealed the same effects as real behavior. Boxes mark 95% confidence intervals and vertical lines standard deviations.

(TIF)

Click here for additional data file.

Grouped individual data points showing precision weights for updating beliefs about social and non-social contingency and volatility.

A, precision weights q(ψ2). B, precision weights ψ3. Overall, q(ψ2) and ψ3 increase when transitioning from stable to volatile phase.

(TIF)

Click here for additional data file.

Parameter recovery.

We simulated behavioral responses based on the posterior estimates of all participants 10 times, resulting in 1160 simulations. For each subject, we calculated the average posteriors estimated from the simulated data (x-axis) and correlated (Pearson’s correlations) them with the original posterior parameters estimates (y-axis), which is shown in plots a-f. The estimated parameters could be recovered well, however, the third level evolution rate (ω) could not be recoverd well.

(TIF)

Click here for additional data file.

Regression model of participants choices for both cue types.

Top: Regression model for choices in card space (1 = taking blue, 0 = taking green) with predictors of card accuracy (1 = blue correct, 0 = green correct) for the past 5 trials, and reward values (reward value if blue taken, reward value if green taken) for the whole task (left), stable (middle) and volatile (right) phase. The slope between predictors were calculated as a model-agnostic readout of the ‘learning rate’, indiciating the degree to which more recent information is weighted. A t-test was applied to compare the difference between these slopes during stable and volatile phases of the task. Slopes are increasing during volatile compared to stable phases. Below: Regression model for choices in gaze space (1 = taking advice, 0 = not taking advice) with predictors of gaze accuracy (1 = gaze correct, 0 = gaze incorrect) for the past 5 trials, and card reward values (reward value if advice taken) for the whole task (left), stable (middle) and volatile (right) phase.

(TIF)

Click here for additional data file.

24 Apr 2020

Dear Ms Henco,

Thank you very much for submitting your manuscript "Aberrant computational mechanisms of social learning and decision-making in schizophrenia and borderline personality disorder" for consideration at PLOS Computational Biology.

As with all papers reviewed by the journal, your manuscript was reviewed by members of the editorial board and by several independent reviewers. In light of the reviews (below this email), we would like to invite the resubmission of a significantly-revised version that takes into account the reviewers' comments.

As you will see, the reviewers were generally positive about your manuscript. However, they had a few serious concerns that need to be addressed. In particular, we would like to call your attention to reviewer 2's comment regarding the comparability of social vs. non-social learning. Even though this point was already mentioned under Limitations, any conclusions regarding social vs. non-social learning should probably made with more caution. Another point, which was raised by several reviewers, is that demonstration of parameter recovery and model simulations would considerably strengthen the paper. We cannot make any decision about publication until we have seen the revised manuscript and your response to the reviewers' comments. Your revised manuscript is also likely to be sent to reviewers for further evaluation.

When you are ready to resubmit, please upload the following:

[1] A letter containing a detailed list of your responses to the review comments and a description of the changes you have made in the manuscript. Please note while forming your response, if your article is accepted, you may have the opportunity to make the peer review history publicly available. The record will include editor decision letters (with reviews) and your responses to reviewer comments. If eligible, we will contact you to opt in or out.

[2] Two versions of the revised manuscript: one with either highlights or tracked changes denoting where the text has been changed; the other a clean version (uploaded as the manuscript file).

Important additional instructions are given below your reviewer comments.

Please prepare and submit your revised manuscript within 60 days. If you anticipate any delay, please let us know the expected resubmission date by replying to this email. Please note that revised manuscripts received after the 60-day due date may require evaluation and peer review similar to newly submitted manuscripts.

Thank you again for your submission. We hope that our editorial process has been constructive so far, and we welcome your feedback at any time. Please don't hesitate to contact us if you have any questions or comments.

Sincerely,

Philipp Sterzer

Guest Editor

PLOS Computational Biology

Samuel Gershman

Deputy Editor

PLOS Computational Biology

***********************

Reviewer's Responses to Questions

Comments to the Authors:

Please note here if the review is uploaded as an attachment.

Reviewer #1: In this paper, Henco et al. evaluate alterations in hierarchical evidence in social and non-social contexts across a transdiagnostic sample comprising individuals with schizophrenia, major depressive disorder, borderline personality disorder and healthy controls. They use a well-designed advice taking paradigm that manipulates of volatility and model the data using a variant of the 3-level hierarchical Gaussian filter, a widely used approach developed by this group. The experimental design and computational modeling is rigorous and sound and the transdiagnostic evaluation is a relevant and novel aspect of this work.

While this work has some relevant limitations, the strengths seem to outweigh them. I am generally enthusiastic about this work and only have relatively minor comments that I would like the authors to address.

1. One potential concern has to do with low-level confounds related to poor accuracy or use of heuristics. On Fig 3a, it seems like response accuracy is around chance level in some subjects. Can the authors provide evidence that these subjects understand task instructions and engage with the task as intended? Any basic manipulation checks would be useful. Furthermore, can the authors show that the group differences in fitted model parameters are not driven by low-level issues such as low overall poor accuracy or use of heuristics? Looking for strategies such as win-stay/lose-shift and the percentage of participants using them in each sample would also be advisable. Perhaps a within-subject model comparison could help exclude subjects for whom there is no evidence to support they are learning during the task.

2. One important aspect of this study is the transdiagnostic evaluation of social learning impairments in relation to the ACIPS measure of social anhedonia, which the authors report as a negative finding. This result should be further emphasized in the discussion for balance. Previous work (e.g., Gillan et al.) suggests that transdiagnostic measures are more related to computational phenotypes than DSM categories, but this work seems to suggest otherwise. While this is a nuanced point, the authors should discuss the implications of this negative result along the lines of relevant previous work.

3. The authors make an effort to explain the partially convergent findings in schizophrenia and borderline personality disorders in terms of their shared symptoms, a relevant point that enriches the discussion. Obviously, these disorders also have many aspects that are not shared. Can the authors conduct further analyses to support which specific shared symptom domains may be driving the convergent findings across these two disorders? Can these findings be instead due to other confounds that may be shared between these two disorders or the specific samples included here, for instance similar medication or lower IQ or differences in sociodemographic status?

4. Related to the previous point, one aspect I would like the authors to emphasize further is the integration of the current results with their previous results in autism and paranoia in the general population.

5. It is unclear why the authors decided to simulate the data for a subset of participants in multiple iterations. Can’t the authors simulate data using the fitted parameter for each subject and take average responses for several iterations in a given subject as the responses in that subject so as to avoid differences in sample size and statistical power between the real and the simulated data? This needs clarification.

6. I would like for the authors add a parameter recovery analysis supporting the identifiability of parameters in their winning model, which tends to be a concern with highly complex models such as the one described here.

Minor points:

7. In Fig. 3 it would be preferable to show scattered data points for each subject (at least for the stable versus volatile contrast estimates).

8. Typo on line 149: I believe they mean ‘afforded’ instead of ‘accorded’

9. To clarify, I am assuming these are all new samples, or have any of them been previously reported on? If so, please indicate which subsamples or parts of samples have been included in previous manuscripts.

Reviewer #2: This study aims at assessing probability learning and volatility in schizophrenia (SCZ), borderline personality disorder (BPD) and major depressive disorder (MDD) as well as learning from social and non social cues.

The authors test patients diagnosed with MDD (N=29), SCZ (N=31), and BPD (N=31), and healthy controls (N=34) performing a probabilistic reward learning task in which participants could learn from social (a cue given by the gaze of a person in a photo) and nonsocial information regarding probabilities of two cards (a blue card and a green card) to give a reward. The probabilities of the pairing between each card and rewards could be stable or volatile. The social cue could be congruent or incongruent with the correct answer.

The individual behaviour data was fit using a HGF model.

Participant also filled an AQ questionnaire and a questionnaire regarding social anhedonia.

The raw data showed that BPD and SCZ had worse performance than MDD and HC with patients with BPD following the social advice significantly more compared to patients with MDD during volatile phases of low accuracy.

The modelling suggests impaired learning from social and non-social information in BPD characterised by an exaggerated sensitivity to changes in environmental volatility. Compared to controls, patients with BPD and SCZ showed an over-reliance on their beliefs about the predictive value of social relative to non-social information during decision-making.

The topic is very timely and interesting. The paper is clear and well written (though fairly technical due to being centred on the use of the HGF model). I think using this sort of task and modelling with patients having different types of ICD10 diagnostics and looking at commonalities or differences is very interesting and what needs to be done to progress in the field.

However, I have 2 major concerns about the study:

1) As the authors point out in Discussion, although the authors want to assess “social learning” (as per the title of the article), they didn’t use a control condition to assess whether the differences observed with the “social cue” had anything to do with the “social” nature of the cue (i.e. that it is presented as an eye-gaze), or whether it is more related to having to integrate two types of cues, that is whether they are looking at a difference in cue integration rather than in social learning, where the two cues might also rely on working memory in different ways. 
They could have used a control condition with an arrow for example instead of a gaze, and I don’t really understand why the authors didn’t do that. 
In my mind, it would have greatly enhanced the study and justified framing it as a study about “social learning”. In the absence of such a control condition, I find that all the arguments related to social learning to be weak and speculative.

2) I think because the study relies on the complex HGF model, it would require showing how good the model is at parameter and model recovery for this task/data (cf. Wilson and Collins, Ten simple rules for the computational modelling of behavioural data, eLife 2019). 
The model comparison with simpler models in particular the Rescorla-Wagner model (with fixed learning rate) doesn’t sound very fair: how does the RW model include the social cue?

It is important to justify the use of such complex models (rather than starting from them), and in my mind the way to do this is both to show that parameter and model recovery works, and that no simpler model can account for the data.

Minor comments:

- Table 1 - explain how the reader is supposed to make sense of those numbers

- Abstract and main text could be clearer in terms of what the behaviour shows, what questions it raises and what the modelling reveals in addition.

Reviewer #3: The present study applied a probabilistic reinforcement learning task to three different groups of patients, major depressive disorder (MDD), schizophrenia (SZ), and borderline personality disorder (BPD), and healthy controls. The task allowed directly learning from the outcome of choices and about the reliability of a social cue (the gaze of a face presented centrally between the two choice options). Hierarchical Gaussian filter, a variant of Bayesian inference was applied to the behavioural data to extract different learning parameters. The main finding is that BPD show diminished learning from either source of information (outcomes and gaze information), whilst at the same displaying aberrantly high learning about the volatility of the environment (both the outcome and gaze, which orthogonally varied). Furthermore, choices of both BPD and SZ were guided more strongly by gaze compared to outcome information relative to healthy controls.

Overall, this is a very interesting study and I really appreciate the attempt at linking different clinical diagnoses with underlying computational phenotypes. However, at present, I am not fully confident about the findings presented. Specifically:

1) The group differences are exclusively observed in the model-derived parameters, not in any of the behavioural parameters. It is therefore very important to be able to follow exactly on what the HGF is doing. At present, I find the description of the HGF not very intuitive to follow for readers not familiar with Bayesian inference. Thus, one would have to take the authors' word for it. Rewriting this paragraph in a way that makes it more amenable to non-modellers could provide a lot more clarity - and make the work of greater interest to a broader readership.

2) Relatedly, I would find it very assuring if the authors could present a behavioural readout that does not rely on high-level Bayesian modelling. One option that comes to mind is to use a regression-based approach. Entering the (blue) outcomes and the gaze direction, each for the past n trials, reward available on the current trial (and possibly the interaction of both) could be used as predictors. This could be run separately for the stationary and volatile environments. In my view, the slope of the (social and non-social) weights should reflect the social and non-social learning rates, and the difference between blocks would reflect their adjustments in response to changes in volatility.

3) It is currently hard to judge how well the best fitting model described participants' choices. Could you please provide some more intuitive measure, such as the model's average probability of choosing the option selected by the participant? Is the winning model the same in all patient groups/controls? In how many of the participants was it the best one?

4) Only two non-Bayesian models are presented, and these are not explained, the reader is referred to two references. While I know what an RW model is, I had not heard about the Sutton K1 model before, and even a simple RW model can come in many flavours. What was it that motivated the choice of these specific two non-Bayesian models? Why not others, e.g. some that also feature dynamic learning rates (e.g Pearce-Hall associability), or separate learning rates for social and non-social information? There are various other alternative models that could be compared here, and I am not sure why the authors restricted it to those two.

5) I am also missing details on the task. What was the range of possible outcomes? I am asking because (as described by prospect theory) humans typically do not weight reward amount linearly, tending to underweight higher amounts. Does it improve model fits to account for such subject-specific distortions?

6) In the results, it says: "We simulated responses using the posterior mean parameter values of 60 randomly chosen participants from the best fitting model to demonstrate that this model was capable of reproducing the group differences that were observed in the real behavioral data..."

This seems arbitrary. Why only a subset (of more than 50% of all participants) rather than all participants? Furthermore, can the fitted parameters be recovered from the simulated data?

7) Overall performance (% correct) is impaired in BDP and SCZ, and overall task performance is quite low. In particular in the patient group, there appears to be a significant fraction of people that performs just around chance level (some slightly above, some even below). This leaves me wondering whether there was a substantial proportion of participants (patients) that failed to properly perform the task?

Minor:

1) How did you define "% correct"? Does this refer to choices of the option with higher probability or with higher expected value (probability x points)?

2) Figure 2: please indicate in the legend what values are shown. It looks like medians in 2A+B, but mean ± SEM in 2C? Please clarifiy.

3) Parts of the abstract appear to be redundant, unless I'm mistaken, lines 72-76 merely repeat the findings already stated above?

**********

Have all data underlying the figures and results presented in the manuscript been provided?

Large-scale datasets should be made available via a public repository as described in the PLOS Computational Biology data availability policy, and numerical data that underlies graphs or summary statistics should be provided in spreadsheet form as supporting information.

Reviewer #1: Yes

Reviewer #2: No: At the moment they state "All XXX files are available from the XXX database (accession number(s) XXX, XXX.)." (haven't replaced the XXX).

Reviewer #3: None

**********

PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

Reviewer #3: No

Figure Files:

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, . PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email us at .

Data Requirements:

Please note that, as a condition of publication, PLOS' data policy requires that you make available all data used to draw the conclusions outlined in your manuscript. Data must be deposited in an appropriate repository, included within the body of the manuscript, or uploaded as supporting information. This includes all numerical values that were used to generate graphs, histograms etc.. For an example in PLOS Biology see here: http://www.plosbiology.org/article/info%3Adoi%2F10.1371%2Fjournal.pbio.1001908#s5.

Reproducibility:

To enhance the reproducibility of your results, PLOS recommends that you deposit laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. For instructions, please see

24 Jun 2020

Submitted filename: ResponsetoReviewers.docx

Click here for additional data file.

19 Jul 2020

Dear Ms. Henco,

We are pleased to inform you that your manuscript 'Aberrant computational mechanisms of social learning and decision-making in schizophrenia and borderline personality disorder' has been provisionally accepted for publication in PLOS Computational Biology.

Before your manuscript can be formally accepted you will need to complete some formatting changes, which you will receive in a follow up email. A member of our team will be in touch with a set of requests.

Please note that your manuscript will not be scheduled for publication until you have made the required changes, so a swift response is appreciated.

IMPORTANT: The editorial review process is now complete. PLOS will only permit corrections to spelling, formatting or significant scientific errors from this point onwards. Requests for major changes, or any which affect the scientific understanding of your work, will cause delays to the publication date of your manuscript.

Should you, your institution's press office or the journal office choose to press release your paper, you will automatically be opted out of early publication. We ask that you notify us now if you or your institution is planning to press release the article. All press must be co-ordinated with PLOS.

Thank you again for supporting Open Access publishing; we are looking forward to publishing your work in PLOS Computational Biology.

Best regards,

Philipp Sterzer

Guest Editor

PLOS Computational Biology

Samuel Gershman

Deputy Editor

PLOS Computational Biology

***********************************************************

Reviewer's Responses to Questions

Comments to the Authors:

Please note here if the review is uploaded as an attachment.

Reviewer #1: The authors have been very responsive and adequately addressed all my comments.

Reviewer #2: The authors have answered a lot of my concerns. I am still not completely convinced about the lack of a control condition to test whether the social nature of the cue is really the determining factor in how it is integrated. However, the authors have now expanded their discussion of this aspect. Overall, I think the paper is now greatly improved and will be very interesting to the community.

Reviewer #3: Thank you for the thorough revisions and the detailed additional analyses. One of my concerns had been that a significant fraction of participants displayed random responding. One of the new analyses shows that a model with random responding indeed is the winning model in n = 16 volunteers. Importantly, however, their main results appear immune to this (excluding these volunteers does not change the main pattern of results) and the state of things is clearly and transparently presented to the reader.

The model-agnostic analyses, in my view, provide important support for the modeling-based claims.

Overall, I think the manuscript now is in good shape for publication.

**********

Have all data underlying the figures and results presented in the manuscript been provided?

Large-scale datasets should be made available via a public repository as described in the PLOS Computational Biology data availability policy, and numerical data that underlies graphs or summary statistics should be provided in spreadsheet form as supporting information.

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #3: None

**********

PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: Yes: Guillermo Horga

Reviewer #2: No

Reviewer #3: No

24 Sep 2020

PCOMPBIOL-D-20-00401R1

Aberrant computational mechanisms of social learning and decision-making in schizophrenia and borderline personality disorder

Dear Dr Henco,

I am pleased to inform you that your manuscript has been formally accepted for publication in PLOS Computational Biology. Your manuscript is now with our production department and you will be notified of the publication date in due course.

The corresponding author will soon be receiving a typeset proof for review, to ensure errors have not been introduced during production. Please review the PDF proof of your manuscript carefully, as this is the last chance to correct any errors. Please note that major changes, or those which affect the scientific understanding of the work, will likely cause delays to the publication date of your manuscript.

Soon after your final files are uploaded, unless you have opted out, the early version of your manuscript will be published online. The date of the early version will be your article's publication date. The final article will be published to the same URL, and all versions of the paper will be accessible to readers.

Thank you again for supporting PLOS Computational Biology and open-access publishing. We are looking forward to publishing your work!

With kind regards,

Matt Lyles

PLOS Computational Biology | Carlyle House, Carlyle Road, Cambridge CB4 3DN | United Kingdom ploscompbiol@plos.org | Phone +44 (0) 1223-442824 | ploscompbiol.org | @PLOSCompBiol