Anterior Insula Activity during Alcohol and Social Reward Self-Administration and Choice in Male and Female Rats

Abstract

Alcohol use disorder (AUD) is characterized by the prioritization of alcohol over healthier nonalcohol rewards, posing significant challenges for addiction treatment. Understanding the neural mechanisms driving this maladaptive preference is crucial for effective intervention. We previously showed that rats will choose alcohol over social reward in a discrete-choice task. Here, we used fiber photometry to investigate how anterior insula cortex (aIC) activity relates to choice and employed Linear Ballistic Accumulator (LBA) modeling to dissect the underlying decision processes. Male and female rats, transfected with calcium indicator jGCaMP7f in aIC, were trained to lever press for either social reward or alcohol (20% ethanol) in alternating sessions, followed by discrete-choice sessions, and then punishment of alcohol choices. Rats developed a preference for alcohol over social reward, which was reversed when alcohol choices were punished. Model output successfully described this behavior with the model-derived “decision bias” tracking preference across all phases. Photometry recordings showed that, as alcohol preference emerged, increased aIC activity during the cue period preceding alcohol choices (relative to social choices) was significantly correlated with decision bias toward alcohol. During punishment, aIC activity bias was no longer related to decision bias, despite the preference shift. These results demonstrate that aIC activity is linked to alcohol reward and choice and suggest that aIC contributes to alcohol preference by encoding a bias in the evidence accumulation process. This highlights a specific role of aIC in the cognitive mechanisms of alcohol-seeking and its potential as a target for therapeutic interventions.

Significance Statement

Understanding the neural basis of prioritizing alcohol over natural rewards is critical for combating alcohol use disorder. This study investigates how anterior insula cortex (aIC) activity in rats relates to choices between alcohol and social reward. Using fiber photometry and cognitive modeling, we found that aIC activity is not only consistently higher for alcohol-related actions but also during the decision period prior to choice. Differential activity for alcohol versus social choices significantly correlates with a model-derived “decision bias,” the speed of evidence accumulation, favoring alcohol, once preference is established. This provides novel mechanistic insight, suggesting aIC contributes to alcohol preference by encoding a bias in the decision-making process, highlighting its role in alcohol-seeking and as a potential therapeutic target.

Introduction

Alcohol use disorder (AUD) continues to impose a significant burden on individuals and society. The effects of chronic alcohol use on the body and brain are well documented (Le et al., 2001; Gilpin and Koob, 2008), but the mechanisms by which such use leads to persistent alcohol use despite negative consequences (Marchant et al., 2018; Domi et al., 2021), or to the exclusion of alternative rewards (Augier et al., 2023; Marchant et al., 2023), remain poorly understood. The role of decision-making and choice in addiction is not without controversy (Hall et al., 2015; Pickard et al., 2015; Hart, 2017), in part because intact choice in addiction can be construed to argue against the brain disease model of addiction (Leshner, 1997; Heilig et al., 2021). However, disordered choice caused by the impact of chronic drug and alcohol use on the associated neurobiological systems has been proposed as a potential to addiction (Jentsch and Taylor, 1999; Goldstein and Volkow, 2011; Heatherton and Wagner, 2011; Bickel et al., 2014; Heilig et al., 2017; Heyman, 2021). Furthermore, treatment strategies incorporating choice, such as community reinforcement and contingency management, can be effective in promoting abstinence (Hunt and Azrin, 1973; Prendergast et al., 2006).

Rodent studies of choice between drug and nondrug rewards have found that rats typically show a choice preference for the nondrug reward (Lenoir et al., 2007; Cantin et al., 2010; Caprioli et al., 2015; Augier et al., 2018). Recent research has recognized the importance of social factors in addiction treatment in humans (Heilig et al., 2016). In rodents, choice for social reward can attenuate drug taking and seeking (Venniro et al., 2018). Recently, however, we found that rats choose alcohol over social reward (Marchant et al., 2023), a finding replicated in a different strain of rats (Augier et al., 2023). Furthermore, food pellets are also chosen over social reward in the discrete-choice procedure (Chow et al., 2022). Thus, while for intravenously administered drugs social choice has protective value, this is not the case for orally consumed drugs like alcohol. Research directed toward understanding the neurobiology underlying choice in addiction has the potential to yield treatments which enhance behavioral allocation toward healthy, nondrug alternatives.

Lesion studies implicate the anterior insula cortex (aIC) in addiction (Naqvi et al., 2007; Joutsa et al., 2022), and fMRI studies show that exposure to alcohol-associated cues is associated with greater activation of aIC in patients (Janes et al., 2017; 2020). Preclinical studies have shown that aIC inactivation decreases relapse (Forget et al., 2010; Pushparaj et al., 2013; Venniro et al., 2017; Campbell et al., 2019; Joshi et al., 2020; Ghareh et al., 2022). Insula cortex activity is also related to loss of control (Rotge et al., 2017; Joshi et al., 2020) and compulsivity (Seif et al., 2013; Belin-Rauscent et al., 2016; Jones et al., 2024) and is important for decision-making during conflict (Naqvi et al., 2014; Daniel et al., 2017). Conflict in particular is a construct which is important for decision-making both in choice (approach-approach conflict) and in punishment (approach-avoidance conflict; Miller, 1971; McNally, 2021). Conflict between competing motivations can lead to ambivalence, which is a cardinal feature of addiction that has received little focus in preclinical research (Heather, 1998; Vandaele and Daeppen, 2022). In this study we used male and female Long–Evans rats to measure activity in aIC using fiber photometry calcium imaging, during alcohol and social reward self-administration, during choice between alcohol and social reward, and during choice between punished-alcohol and unpunished social reward. To describe the decision processes underlying choice, we used cognitive modeling [Linear Ballistic Accumulator (LBA); Brown and Heathcote, 2008] to quantify the latent cognitive processes guiding these choices. We estimated the speed of evidence accumulation during the different phases of the experiment, and from this derived a measure of Decision Bias to explore the relationship between aIC calcium and decision-making in choice between alcohol and social reward.

Materials and Methods

Subjects

We obtained 36 Long–Evans rats (18 male and 18 female), aged 8–10 weeks upon arrival (RjOrl:LE, RGD_151356971, Janvier Labs). All procedures were approved by the Vrije University Animal Welfare Body [Instantie voor Dierenwelzijn VU-VUmc (IvD)] and conducted under the authority of the Dutch Central Commission for Animal Research (CCD, permit #AVD11400202010449) in accordance with European law (Directive 2010/63/EU). We trained the rats in two cohorts, the first cohort had 12 experimental rats and 8 social partners, and the second cohort had 12 experimental rats and 4 new social partners; however, we also used the eight social partners from the first cohort as social partners for the second cohort. Each experimental rat was exclusively trained with a single social partner; however, some social partners served in this role to multiple experimental rats.

Housing conditions

Behavioral tests were conducted during the dark phase of the rat's diurnal cycle (12 h/12 h). Food and water were available ad libitum. Upon arrival, we housed the rats in groups of 4, prior to surgery, typically 2 weeks in duration. We then selected one of those rats to be the “Social Partner” for the remaining 3 “Experimental Rats” for Cohort 1. For Cohort 2, the rats were housed as 2 “Experimental” and 2 “ Social Partner” rats. Throughout the experiment, each experimental rat was paired with same social partner rat. Social partners were then housed together in groups of 4, and the experimental rats were single housed for the remainder of the experiment.

Apparatus

We used Med Associates operant chambers for these experiments (Fig. S1). Each operant chamber consisted of a main compartment (for the experimental rat) and a smaller adjacent compartment (for the social partner), with a grid floor in both compartments. One wall of the main compartment was for social reward. There was a retractable lever, and above the lever a white cue light was the discriminative stimulus and a red cue light was used as the conditioned stimulus. There was also an automated guillotine door separating the main compartment and the adjacent compartment with a grid panel inserted so that the social partner could not enter the main compartment of the operant chamber (Venniro et al., 2019). The other wall of the main compartment was used for alcohol self-administration. This was equipped with another retractable lever, and above the lever a house light was the discriminative stimulus and a red cue light was used as the conditioned stimulus. There was also a receptacle to receive the alcohol infusions for the rat to drink out of. Alcohol was infused into the receptacle via tubing [2806 tygonslang e3603 2, 4(ID) × 4 (mm)] connected to a 20 ml syringe controlled by a Razel pump. All the protocols in the chambers were ran by Med-PC program and the fan remained on for the whole duration. During self-administration and choice, all operant chambers had a custom grid floor with 26 rods; however in order to connect the shock harness, we changed to the grid floor with 19 rods (VFC-005) for the punished-alcohol choice sessions. We were unable to counterbalance which side of the chamber was alcohol or social reward.

For fiber photometry, excitation and emission light was relayed to and from the animal via optical fiber patch cord (0.48 NA, 400 µm flat tip; Doric Lenses). Blue excitation light [490 or 470 nm LED (M490F2 or M470F2, Thorlabs)] was modulated at 211 Hz and passed through a 460–490 nm filter (Doric Lenses), while isosbestic light [405 nm LED (M405F1, Thorlabs)] was modulated at 531 Hz and passed through a filter cube (Doric Lenses). GCaMP7f fluorescence was passed through a 500–550 nm emission filter (Doric Lenses) and onto a photoreceiver (Newport 2151). Light intensity at the tip of the fiber was measured before training sessions two times a week and kept at 40–50 µW. A real-time processor (RZ5P, Tucker-Davis Technologies) controlled excitation lights, demodulated fluorescence signals, and received timestamps of behavioral events. Data was saved at 1,017.25 Hz and analyzed with Matlab scripts.

Alcohol

We prepared 20% ethanol by diluting 70% ethanol (VWR International) with water (v/v). The first cohort of rats (12 experimental rats) had home-cage water that was acidified (pH 2.7), and the operant chamber alcohol was mixed with sterile water. The second cohort of rats (12 experimental rats) had home-cage water that was autoclaved and thus not acidified and the operant chamber alcohol was mixed with autoclaved water.

Viral vectors

We purchased premade viral vectors from the University of Zurich viral vector core: ssAAV-9/2-mCaMKIIα-jGCaMP7f-WPRE-bGHp(A) (jGCaMP7f). The titer injected was as follows: jGCaMP7f, 5.0 × 10¹² gc/ml.

Surgery

Both 1 d and 30 min prior to surgery, we injected rats with the analgesic Rymadil (5 mg/kg, s.c.; Merial); we also administered buprenorphine 30 min prior to surgery (0.01 mg/kg, s.c.). Surgery was performed under isoflurane gas anesthesia (PCH). We placed the anesthetized rat in a stereotactic frame (David Kopf Instruments) and injected lidocaine 10 mg/kg (Fresenius Kabi; RVG 51674) into the incision site prior to the incision. A craniotomy above aIC was performed, and 0.5 µl of AAV solution was injected bilaterally into aIC [AP: +2.8, ML: +4.0 (2° angle), DV: −5.8 mm from bregma] over 5 min. The needle was left in place for an additional 5 min. Two 400 µm optic fibers (Doric Lenses) were then implanted above each aIC [AP: +2.8, ML: +3.8 (0° angle), DV: −5.5 mm from bregma] and secured to the skull using dental cement (Antibiotic Simplex; Stryker, Orthopaedics) and jewelers screws (Jeveka; M1x2 stainless steel screws). Rymadil (5 mg/kg, s.c.) was administered for 2 d after the surgery. Two rats (1 M/1F) were given only unilateral aIC injection of AAV, with bilateral fiber implants, to test the extent to which contralateral projections could be contributing to the signal recorded in each hemisphere.

Behavioral procedure

Phase 1: self-administration

For both alcohol and social reward, we first gave the rats a session where noncontingent rewards were delivered, and no levers were present in the chamber. For alcohol, this 60 min session consisted of an alcohol reward (0.1 ml) being delivered every 3 min (20 in total), and each delivery also coincided with 5 s presentation of the alcohol-associated cue (no lever press). For social, this 60 min session consisted of the guillotine door being opened for 60 s, paired with the red cue light for the first 20 s, every 5 min.

For alcohol self-administration training, we trained the rats to lever press for alcohol reward (FR1) for a total of 10 × 30 min sessions. The alcohol lever was inserted into the chamber, and the alcohol discriminative stimulus was turned on, throughout the whole session (note the social lever was not inserted into the chamber during these sessions). One press on the lever resulted in 0.1 ml of alcohol into the receptacle, the red cue light above this lever being turned on for 5 s. During this time, additional lever presses resulted in no alcohol infusion (i.e., time-out period). Any alcohol remaining in the magazine after session was recorded. We calculated total alcohol consumption in grams per kilogram body weight by multiplying total alcohol rewards by 0.1 (ml), subtracting alcohol remaining, and dividing this by body weight (kg). For social self-administration training, we also trained the rats for a total of 10 × 60 min sessions. Before each session, we first placed the social partner rat into the adjacent social holding component of the operant chamber, and then the experimental rat (of the same sex) was placed into the main component of the operant chamber. The social lever was inserted into the chamber, and the social discriminative stimulus was turned on, throughout the whole session (note the alcohol lever was not inserted into the chamber during these sessions). A single lever press on the social lever resulted in 60 s of social reward signaled by the opening of the guillotine door and the red cue light above this lever turned on for 20 s. Responses on the social lever during this period were recorded but had no programmed consequence.

In this experiment we trained alcohol and social reward self-administration in alternating sessions over successive days (Augier et al., 2023). Alcohol self-administration was the first session, and social self-administration training followed. In the first cohort of rats, one alcohol self-administration session was excluded (Session 7), because this was the first day the rats were tethered to the optic fiber patch cord, and there was a significant reduction in alcohol self-administration on this day (data not shown).

Phase 2: choice tests

Figure 1, E and F, describes the choice session procedure. We tested choice between alcohol and social reward over 6–8 sessions after the last self-administration session. Our choice test procedure was modified slightly from our previous study (Marchant et al., 2023). Each choice test session was composed of 20 trials, each of 4 min duration. Each trial started with presentation of both the alcohol and social reward discriminative cues followed 10 s later by insertion of both levers into the chamber. The levers were programmed to remain inserted into the chamber for 2 min, or until one was pressed. After a lever press, both levers retracted and the corresponding outcome was delivered (either 0.1 ml alcohol infusion and 5 s alcohol-associated CS, or 60 s social interaction and 20 s social-associated CS). Because the choice session is divided into 20 × 4 min blocks, the ITI is variable depending on the response latency of the choice, but it does not vary depending on the outcome (alcohol or social). If no lever was pressed after 2 min, the discriminative cues were turned off and both levers were retracted. Thus, if the rat pressed either lever at the beginning of the trial, then the ITI would be 3 min 50 s, but if they did not press at all (omission), then the cues are turned off and levers retracted for 1 min and 50 s.

Figure 1.

Experimental outline. A, Outline viral targeting strategy. We injected AAV encoding jGCaMP7f into aIC and implanted fiber optics above. B, Experiment design. We trained rats in alternating sessions of alcohol or social reward. During choice, the first cohort (n = 12) received six choice sessions, the second cohort (n = 11) received eight choice sessions. All rats received four punished-alcohol choice sessions, where alcohol lever press resulted in 0.3 mA intensity, 0.5 s duration shock. C, Representative jGCaMP7f and fiber placement in aIC. D, Depiction of fiber placement for all rats, horizontal lines represent the base of the fiber for rats included in the analysis. E, Depiction of the choice session procedure. Each choice session consisted of 20 × 4-minute trials. F, Trial design. A trial begins with onset of the discriminative stimuli (DS) for 10 s and then both levers are inserted. After a response, both levers are retracted, both DS are turned off, and the relevant reward is delivered. If no response is made in 2 min, then both levers are retracted, both DS are turned off. ITI, intertrial interval; DS, discriminative stimulus.

Phase 3: punished-alcohol choice tests

The timing parameters of this phase are identical to the previous choice phase. When an alcohol choice was made, the outcome was 0.1 ml alcohol infusion into the magazine, 5 s presentation of the alcohol-associated CS, and an electric footshock of 0.5 s duration and 0.3 mA intensity. Social choice resulted in the same outcome as before, with no footshock. We gave all rats four sessions in this phase, recording left and right hemisphere each twice. We chose the intensity of 0.3 mA initially because we observe substantial individual variability in single lever operant conditions (McDonald et al., 2024). We were surprised to find that this intensity was sufficient to fully shift preference to social reward for both male and female (except two rats) by the fourth session and as such used this intensity for the second cohort of rats.

Photometry recording sessions

We recorded either left or right hemisphere aIC for all rats over alternating sessions of alcohol and social self-administration, choice, and punished-alcohol choice. The self-administration sessions were only recorded for the last eight sessions, resulting in four sessions each of alcohol or social recorded (two left and two right for each). All choice and punished-alcohol choice sessions were recorded.

jGCaMP7f expression and fiber placement validation

We deeply anesthetized rats with isoflurane and Euthasol injection (i.p.) and transcardially perfused them with ∼100 ml of normal saline followed by ∼400 ml of 4% paraformaldehyde in 0.1 M sodium phosphate, pH 7.4. The brains were removed and postfixed for 2 h and then submerged in 30% sucrose in 0.1 M PBS for 48 h at 4°C. Brains were then frozen on dry ice, and coronal sections were cut (40 µm) using a Leica Microsystems cryostat and stored in 30% sucrose in 1.0 M PBS stored at −20°C.

Immunohistochemical procedures are based on our previously published work (Marchant et al., 2009; 2010; Campbell et al., 2019; Ghareh et al., 2022). We selected a 1-in-4 series and first rinsed free-floating sections (3 × 10 min) before incubation in PBS containing 0.5% Triton X-100 and 10% Normal Donkey Serum (NDS) and incubated for at least 48 h at 4°C in mouse anti-NeuN primary antibody (1:1,000; Chemicon, MAB377). Sections were then repeatedly washed with PBS and incubated for 2–4 h in PBS + 0.5% Triton X-100 with 2% NDS and donkey anti-mouse secondary antibody DyLight 649 (1:500; Jackson ImmunoResearch, 715-495-150). After another series of washes in PBS, slices were stained with DAPI (0.1 ug/ml) for 10 min prior to mounting onto gelatin-coated glass slides, air-drying, and coverslipping with Mowiol and DABCO.

Slides were all imaged on a Vectra Polaris slide scanner (VUmc imaging core) at 10× magnification, and QuPath was used for image analysis (Bankhead et al., 2017). Images containing aIC, from bregma +4.2 mm to +2.5 mm, were identified and the boundary of expression for each rat was plotted onto the respective Paxinos and Watson atlas (Paxinos and Watson, 2008).

Behavioral data analysis: LBA modeling

To investigate the latent decision processes underlying choice behavior, we fitted reaction time (RT) and choice data from each rat and experimental session (“early”, “late”, “punishment”) to a hierarchical LBA model (Brown and Heathcote, 2008). The LBA is a well-established sequential sampling model that describes decision-making as a race between independent accumulators, one for each response option (in this study, alcohol, and social reward). The first accumulator to reach a decision threshold determines the choice made and the decision time for that trial. By implementing the model in a hierarchical framework, individual subject- and session-level parameters are estimated as being drawn from overarching group-level distributions. This approach allows the model to borrow strength across subjects and sessions, separating true variability from estimation noise and resulting in more stable and reliable parameter estimates.

LBA model parameters

The LBA model characterizes choice and RT through several key parameters, which were estimated for each rat and session.

Mean accumulation rate (v)

The average rate at which evidence accumulates for a specific response option (i.e., v_alcohol, v_social). For each trial, the actual accumulation rate for an accumulator is sampled from a normal distribution with mean (v_option) and standard deviation (s). Higher mean accumulation rates lead to faster and more frequent choices of that option.

Between-trial variability in accumulation rate (s)

This parameter captures trial-to-trial variability in the speed of evidence accumulation. Following common practice for model identifiability and comparison with previous work (Annis et al., 2017; Choi et al., 2022), we fixed s to 1. This allows other parameters, particularly the mean accumulation rates, to scale accordingly and represent the signal-to-noise ratio of the accumulation process.

Start-point variability (A)

The upper bound of a uniform distribution (U[0,A]) from which the starting point of evidence accumulation for each accumulator is randomly sampled on each trial. Larger A indicates more variability in the initial state of evidence.

Decision threshold (b)

The amount of evidence required for an accumulator to trigger a response. In our parameterization, b was derived from two estimated parameters: a relative threshold component (k) and the start-point variability (A), such that b = k + A. This formulation ensures that the decision threshold b is always greater than any possible starting point (k is constrained to be positive).

Nondecision time (tau)

This parameter accounts for time consumed by perceptual and motor processes that are peripheral to the evidence accumulation and decision stage (e.g., stimulus encoding, response execution). The total observed RT on a given trial is the sum of the decision time (determined by the race) and tau.

Hierarchical model specification and comparison

To systematically determine which cognitive mechanisms rats used to adapt their behavior, we employed a “power set” approach. We specified three core mechanisms that could vary across the experimental sessions: (1) Stimulus Bias (S), a change in the relative evidence accumulation rates (v_alcohol vs v_social); (2) Caution (C), a change in the decision threshold, parameterized via start-point variability (A); and (3) Response Bias (R), an asymmetry in the starting point of accumulation, parameterized via the relative threshold (k). We then constructed and fit eight hierarchical LBA models, representing all possible combinations of these mechanisms (i.e., a baseline model with no mechanisms varying, three single-mechanism models, three double-mechanism models, and one triple-mechanism model).

We performed a formal model comparison using Leave-One-Out Cross-Validation (LOO-CV), a robust method for estimating out-of-sample predictive accuracy. The results, assessed via the expected log predictive density (ELPD), indicated that model M_SC, in which Stimulus Bias (S) and Caution (C) were allowed to vary across sessions, provided the best and most parsimonious account of the data (Supplementary Materials). Therefore, all subsequent analyses and interpretations are based on the posterior parameter estimates from this winning model.

Parameter priors for the winning hierarchical model (M_SC)

We fitted the LBA model within a Bayesian framework using PyMC (Salvatier et al., 2016). Priors for the group-level model parameters were chosen to be weakly informative: Group-level means (μ): for parameters that varied across sessions in the winning M_SC model (v_alcohol, v_social, and A), separate group-level means were estimated for each session. Parameters held constant across sessions for each subject (k and τ) were estimated with a single group-level mean. Group-level standard deviations (σ): HalfNormal distributions were used for the standard deviations of all group-level parameters. Subject/session-level offsets: Individual subject- and session-level parameters were modeled as offsets from the group-level mean, drawn from a Normal distribution with a standard deviation determined by the corresponding group-level σ. This noncentered parameterization improves sampling efficiency. Non-decision time (τ): For each subject, the τ parameter was bounded to be less than the minimum observed RT for that subject, ensuring plausibility.

Derived parameters

From the primary posterior distributions of the winning model, we derived two key cognitive metrics: Response caution: Defined as b − A / 2, reflecting the average distance from the mean starting point to the threshold. Decision bias: We define this metric as the accumulation rate difference (v_alcohol − v_social). This value reflects the relative speed and efficiency of evidence accumulation toward one choice over the other, rather than a predecisional bias in the starting point.

Model fitting and convergence

The hierarchical LBA model was fitted to the complete dataset of all trials from all rats. We used the No-U-Turn Sampler (NUTS; Hoffman and Gelman, 2014), as implemented in PyMC. We ran four parallel chains, each with 1,500 tuning (warm-up) steps and 2,000 sampling steps, resulting in 8,000 posterior samples per parameter. Convergence of the MCMC chains was assessed using the R-hat $(\hat{R})$ statistic, with $\hat{R}<1.01$ indicating successful convergence (Gelman and Rubin, 1992), and the effective sample size (ESS), with ESS > 400 considered adequate (Gelman et al., 2013). Trace plots were also visually inspected for chain mixing and stationarity.

Model evaluation

To evaluate how well the fitted models could replicate observed behavioral patterns, posterior predictive checks (PPCs) were performed for each converged model. This involved simulating 500 new datasets, each with the same number of trials as the original rat-session dataset, by drawing 500 parameter sets from the joint posterior distribution of the fitted model. We then compared the distributions of these simulated reaction times (separated by choice option) and the simulated choice proportions against the actual observed data. All analyses were performed using Python (version 3.10.15) and the PyMC (version 5.18.2), ArviZ (version 0.20.0), pandas (2.2.3), and NumPy (1.26.4) libraries. The scripts are available at https://github.com/njmarchant/LBA-Alcohol_Social_Choice

Statistics

Behavior

All data was analyzed using IBM SPSS V21. Phases were analyzed separately, and in all tests, we compared male and female rats using Sex as a between-subjects factor. During self-administration the dependent variables were the total number of active lever presses, and total number of alcohol reward deliveries or social reward opportunities (guillotine door open for 60 s). For the choice tests, dependent variables were the total choice of either alcohol or social reward. To analyze we used repeated-measures analysis of variance (ANOVA), using the within-session factor Session where appropriate.

Preference score was calculated using the formula: (Alcohol − Social) / (Alcohol + Social) resulting in values that range from +1 (full alcohol preference) to −1 (full social preference). To calculate the preference scores for “early” and “late” choice sessions, we averaged the preference score of the two sessions for each rat. For the punished-alcohol choice sessions, we averaged the preference score of all four sessions. All comparisons were made using repeated-measures t tests.

Latency data was not normally distributed, and therefore we used Kruskal–Wallis one-way ANOVA to compare latencies for alcohol or social choice in the different sessions in a single analysis. We report the relevant post hoc test output both uncorrected and Bonferroni’s corrected for multiple tests. For “late choice” latencies, we used the last two choice sessions of Phase 2, thus for Cohort 1 this was Sessions 5 and 6, and for Cohort 2 this was Sessions 7 and 8. To calculate mean latencies, omitted trials were simply excluded.

Photometry

Recorded signals were first downsampled by a factor of 64, giving a final sampling rate of 15.89 Hz. The 405 nm isosbestic signal was fit to the 470/490 nm calcium-dependent signal using a first order polynomial regression. A normalized, motion-artifact-corrected ΔF/F was then calculated as follows: ΔF/F = (490 nm signal − fitted 405 nm signal) / fitted 405 nm signal. The resulting ΔF/F was then detrended via a 90 s moving average and low-pass filtered at 3 Hz. Several different epochs were selected for analysis. In self-administration, ΔF/F from 5 s before lever press to 10 s after were collated. To avoid duplicate traces due to overlapping epochs in self-administration, we only included the reinforced active lever presses, and thus all time-out responses were not analyzed. These traces were then baseline corrected using the baseline time period −5 s to −3 s prior to the reinforced lever press. We converted the data into z-scores by subtracting the mean baseline activity and dividing by the standard deviation. In choice, epochs around the lever press (choice) and around the trial start were separately analyzed. For analysis based on the choice response, ΔF/F from 20 s before the lever press to 20 s after were collated. For analysis based on the trial start, ΔF/F from 10 s before the cues turning on to 30 s after were collated. These traces were then baseline corrected and converted into z-scores with the same method described above, using the time period of −10 to −5 s before trial as the baseline for analyses centered on trial start and −20 to −15 s before the response for analyses centered on the response.

In self-administration, traces from the different sessions (alcohol or social) were compared. For analysis of the choice data, we compared choice outcome (response) and decision period (trial start with cues) separately. For the choice we grouped traces by choice outcome: alcohol, social, or omission. For the trial start we grouped traces around the response that was eventually made (alcohol, social, or omission). We used two approaches to analyze the resulting traces: bootstrapped confidence intervals and permutation tests. The rationale for each is described in detail in a previous study (Jean-Richard-Dit-Bressel et al., 2020). The output of every statistical test we conducted is available in the raw data files.

Bootstrapped confidence intervals were used to determine whether calcium activity was significantly different from baseline (ΔF/F = 0). A distribution of bootstrapped means was obtained by randomly sampling from traces with replacement (n traces for that response type; 5,000 iterations). A 95% confidence interval was obtained from the 2.5th and 97.5th percentiles of the bootstrap distribution, which was then expanded by a factor of sqrt(n / (n − 1)) to account for narrowness bias.

Permutation tests were used to compare calcium activity between the different groupings. Observed differences were compared against a distribution of 1,000 random permutations (difference between randomly regrouped traces) to obtain a p-value per time point. Alpha of 0.05 was Bonferroni-corrected based on the number of comparison conditions, resulting in alpha of 0.01 for comparisons between three conditions. For both bootstrap and permutation tests, only periods that were continuously significant for at least 0.5 s were identified as significant.

The output of these tests is displayed in the figures below the traces, where the presence of a significance bar indicates significance during that time period. Not all statistical tests are reported in the figures; however, we have reported the time windows where the statistical tests yielded significant outputs in for every test. The scripts are available at https://github.com/njmarchant/Photom-Alcohol_Social_Choice

Results

Figure 1 shows the experimental outline and targeting strategies for calcium imaging with fiber photometry in aIC. We targeted the anterior insula cortex bilaterally with AAV encoding jGCaMPf driven under the CaMKIIa promoter (Fig. 1A), Figure 1C shows example expression, and Figure 1D shows the fiber placements verified after the experiment. Figure 1B shows the experimental design; we used an alternating training schedule (Augier et al., 2023), where rats were trained in separate days to self-administer on an FR1 schedule either alcohol (20% ethanol in water, 3 s time-out) for 30 min or social reward (60 s open door separating the experimental rat and social partner) for 60 min. Recording sessions were made on the final four self-administration sessions, all choice sessions, and all punishment sessions.

Alcohol and social reward self-administration

Figure 2 shows the self-administration data where only one lever was available for the whole session. Figure 2A shows the total number of rewards that were delivered in each session, and Figure 2B shows the total number of lever presses. For alcohol self-administration, we found a main effect of Session for rewards (F_(9,198) = 5.057; p < 0.001) and for lever presses (F_(9,198) = 4.180; p < 0.001). There was no effect of Sex in these measures (F_(1,22) < 1; p > 0.05). Analysis of alcohol consumed (g/kg; data not shown) revealed a main effect of Session (F_(1,22) = 5.25; p < 0.001) but no effect of Sex (F_(1,22) < 1; p > 0.05). Alcohol consumed (g/kg) in the final session was 0.54 (±0.45) and 0.83 (±0.89) for Male and Female, respectively.

Figure 2.

Activity in aIC during alcohol and social reward self-administration sessions. A, Total rewards and B, Total lever presses, during the alcohol or social reward self-administration sessions. C, Traces depicting aIC activity surrounding the reinforced lever presses (0 s) during the self-administration sessions. Only reinforced lever presses are represented here. D, Mean (±SEM) of z-scored activity during the 3 s outcome period during self-administration. Table 1 shows the time windows where the statistical tests were significant. *p < 0.05. n = 11 males; diamond shapes, n = 12 females, circle shapes.

For social reward self-administration, we found a main effect of Session for rewards (F_(9,198) = 7.139; p < 0.001) and for lever presses (F_(9,198) = 5.317; p < 0.001). There was no effect of Sex in these measures (F_(1,22) < 1; p > 0.05). To compare alcohol and social reward, we converted the data to the rate of responding per minute and found an effect of Reward Type on Total Rewards (F_(1,23) = 4.377; p = 0.048) but not on Total Lever Presses (F_(1,23) = 4.128; p = 0.054). Comparable analysis on the last four sessions of each (the recorded sessions) found no effect of Reward Type on Total Rewards (F_(1,23) < 1; p > 0.05) or Total Lever Presses (F_(1,23) < 1; p > 0.05).

Figure 2C shows the traces depicting calcium transients recorded in aIC centered around the reinforced lever presses. Bootstrapping analysis revealed that aIC activity was higher than baseline both prior to and after the lever press. Comparison between aIC activity for alcohol and social reward using permutation tests revealed a significant difference. Significant time windows are reported in Table 1. Summary statistics using the mean of the z-scored recorded calcium during the 3 s after a reinforced lever press are shown in Figure 2D. There is a significant difference between Alcohol and Social (Alcohol mean 2.43 ± 1.31, Social mean 1.66 ± 1.04; paired t test Alcohol vs Social: t₍₂₁₎ = −2.56, p = 0.019). These data show that while aIC activity is increased relative to baseline for both alcohol and social rewards, the magnitude of activation is higher for alcohol.

Table 1.

Time periods (in seconds) where the photometry statistical tests are significant in Figure 2

Experimental phase	Figure	Statistical comparison	Significance time window (seconds) 0 = Lever press
Self-administration	2C	BCI: Alcohol-reinforced lever press	−2.36→6.61
	2C	BCI: Social-reinforced lever press	−1.92→6.86
	2C	PT: Alcohol v Social	−2.05→5.67

Time 0 refers to reinforced lever press. BCI, Bootstrapped confidence intervals; PT, permutation test.

Discrete choice between alcohol and social reward

Figure 1E shows the choice session procedure. We used a comparable procedure to our previous work (Marchant et al., 2023); however, rather than 15 × 8 min trials, we gave the rats 20 × 4 min trials. Figure 3A shows the choice data from the 8 sessions [Cohort 1 (n = 12) received 6 sessions, Cohort 2 (n = 11) received 8 sessions]. One male rat (from Cohort 2) was removed from the dataset because he made only two choice responses in eight choice sessions. We found that over the course of the training, a choice preference for alcohol emerged. Analysis of the choice responses revealed a significant Choice × Session interaction (F_(5,105) = 5.065; p < 0.001) including both cohorts from c1 to c6. Analysis of Cohort 2 from c1 to c8 also revealed a significant Choice × Session interaction (F_(7,63) = 10.602; p < 0.001). Figure 3B shows preference score for the first two choice sessions (Early: c1–c2 for both cohorts) and the last two choice sessions (Late: c5–c6 for Cohort 1, c7–c8 for Cohort 2). Paired t test revealed a significant difference (t₍₂₂₎ = 3.437; p = 0.0024) reflecting the emergence of choice preference for alcohol. We also analyzed the preference score data with Sex as a between-subjects factor and found no Session × Sex interaction (F_(1,21) = 2.053; p > 0.05) and no main effect of Sex (F_(1,21) < 1; p > 0.05). One-way ANOVA on alcohol consumed (g/kg; data not shown) during the Early sessions revealed no effect of Sex (Male: 0.39 ± 0.23; Female: 0.42 ± 0.32; F_(1,21) < 1; p > 0.05); however, there was an effect of Sex in the Late sessions (Male: 0.55 ± 0.27; Female: 0.91 ± 0.0.39; F_(1,21) = 6.3; p = 0.02).

Figure 3.

Activity in aIC during choice sessions. A, Group data showing alcohol or social choices, and omissions, from the choice sessions. B, Preference score comparing Early and Late choice sessions. C, D, Traces depicting aIC activity centered around the trial start (0 s) during all Early or Late choice trials. E, Mean (±SEM) of z-scored activity during the 10 s cue period during the Early (left) or Late (right) choice sessions. F, G, Traces depicting aIC activity surrounding choice response (0 s). H, Mean (±SEM) of z-scored activity during the 3 s outcome period during the Early (left) or Late (right) choice sessions. Table 2 shows the time windows where the statistical tests were significant. *p < 0.05. n = 11 males; diamond shapes, n = 12 females, circle shapes.

Figure 3, C and D, shows aIC activity centered around trial start during the Early (3C) and Late (3D) choice sessions, with the data separated into whether the trials are those where alcohol or social is chosen. During the Early choice sessions (3C), aIC activity increased relative to baseline for both choice types but was not different from each other (permutation tests). In contrast, during the Late choice sessions (3D), we found that aIC activity during the cue period is significantly higher when alcohol is chosen compared with when social is chosen. Significant time windows are reported in Table 2. This pattern of activity is reflected in the summary statistics, shown in Figure 3E. Within-subjects ANOVA using the factors Session (Early, Late) and Choice (Alcohol, Social) revealed a main effect of Choice (F_(1,19) = 7.9, p = 0.011) and a Choice by Session interaction (F_(1,19) = 9.20, p = 0.007). Post hoc t tests show that aIC activity during the cue period for alcohol choices significantly increased from Early to Late choice (paired t test Early Alcohol vs Late Alcohol: t₍₂₁₎ = −3.21, p = 0.004), and in the Late sessions alcohol is significantly higher than social (paired t test Late Alcohol vs Late Social: t₍₂₀₎ = 3.79, p = 0.001).

Table 2.

Time periods (in seconds) where the photometry statistical tests are significant in Figure 3

Experimental phase	Figure	Statistical comparison	Significance time window (Seconds)
			0 = Trial start
Early Choice (Trial start)	3C	BCI: Alcohol Choice	0.38→3.40; 7.92→10
	3C	BCI: Social Choice	0.38→3.46
	3C	PT: Alcohol v Social	n.s.
Late Choice (Trial start)	3D	BCI: Alcohol Choice	0.38→10
	3D	BCI: Social Choice	0.57→1.89; 3.21→4.28; 6.16→6.92; 7.74→8.93;
	3D	PT: Alcohol v Social	0.44→10
			0 = Lever press
Early Choice (Response)	3F	BCI: Alcohol-reinforced lever press	−13.71→20
	3F	BCI: Social-reinforced lever press	−11.70→−7.55; −6.48→20
	3F	PT: Alcohol v Social	−9.37→−6.54; −5.72→5.03; 7.17→7.92
Late Choice (Response)	3G	BCI: Alcohol-reinforced lever press	−12.64→20
	3G	BCI: Social-reinforced lever press	−12.89→20
	3G	PT: Alcohol v Social	−9.12→3.14; 14.47→20

Time 0 refers to reinforced lever press. BCI, bootstrapped confidence intervals; PT, permutation test; n.s., not significant.

We performed a secondary analysis on aIC activity centered around trial start for the intermediate choice sessions (i.e., “middle” choice sessions are C3 and C4 sessions). These data show that there is a significant increase in aIC activity during the cue period for alcohol choices (Fig. S2B), but not social choices (Fig. S2C), in the middle choice sessions compared with early choice sessions. This is potentially interesting because there are no group differences in preference scores between these sessions (Fig. S2A; paired t test Early vs Middle: t₍₂₂₎ = 0.15, p > 0.05). Comparisons of aIC activity from the middle to late choice sessions show a further increase in aIC calcium toward the end of the cue period (Fig. S2E) and again no change in activity when social is chosen (Fig. S2F). Behaviorally, there was a significant change in preference score (Fig. S2D; paired t test Middle vs Late: t₍₂₂₎ = 4.8, p < 0.001).

Figure 3, F and G, shows aIC activity centered around the choice response (lever press) during the Early (3F) and Late (3G) choice cessions, with the data separated into whether the choices made were alcohol or social reward. Recorded traces of aIC activity reveal comparable patterns of activity between the early and late choice sessions. In both session types, aIC activity was higher for alcohol choices leading up to the lever press and during the outcome period alcohol was higher than social. Figure 3H shows the mean aIC activity during the 3 s period after the choice was made. We found that in both early and late choice sessions, aIC activity was higher for alcohol than social reward and that this did not change between the sessions. Within-subjects ANOVA using the factors Session (Early, Late) and Choice (Alcohol, Social) revealed a main effect of Choice (F_(1,19) = 20.2, p < 0.001) but no Choice by Session interaction (F_(1,19) < 1, p > 0.05). Post hoc t tests show that aIC activity during the outcome period for alcohol is significantly higher than social in both Early and Late choice (paired t test Early Alcohol vs Early Social: t₍₂₁₎ = 4.34, p < 0.001; paired t test Late Alcohol vs Late Social: t₍₂₀₎ = 2.17, p = 0.04) and no differences between Early and Late sessions for Alcohol (Early Alcohol mean 5.27 ± 2.25, Late Alcohol mean 5.02 ± 2.88; paired t test Early vs Late: t₍₂₁₎ < 1, p > 0.05) or Social (Early Social mean 3.58 ± 1.84, Late Social mean 3.68 ± 2.82; paired t test Early vs Late: t₍₂₁₎ < 1, p > 0.05).

Discrete choice between punished alcohol and social reward

Figure 4A shows the total choices made for alcohol or social during the late choice sessions and the punished-alcohol choice sessions. We found that punishment of the alcohol-reinforced choice caused a preference switch to social reward. Analysis of the choice responses comparing response from the “late choice” sessions with the punished-alcohol choice sessions revealed a significant Choice × Session interaction (F_(1,21) = 7.243; p = 0.014). One-way ANOVA on alcohol consumed (g/kg; data not shown) during the punished-alcohol choice sessions revealed no effect of Sex (Male: 0.08 ± 0.15; Female: 0.07 ± 0.03; F_(1,21) < 1; p > 0.05). Figure 4B shows the preference score for the late choice sessions and the punished-alcohol choice sessions. Paired t test revealed a significant difference (t₍₂₂₎ = −9.757; p < 0.001) reflecting the emergence of choice preference for social reward.

Figure 4.

Activity in aIC during punished-alcohol choice sessions. A, Group data showing alcohol or social choices, and omissions, from the punished-alcohol choice sessions. B, Preference score comparing Late choice and Punished alcohol choice sessions. C, D, Traces depicting aIC activity centered around the trial start (0 s) during all punished-alcohol choice trials. E, Mean (±SEM) of z-scored activity during the 10 s cue period in trials where alcohol is chosen (left) or social is chosen (right). F, G, Traces depicting aIC activity surrounding choice response (0 s). H, Mean (±SEM) of z-scored activity during the 3 s outcome period in trials where alcohol is chosen (left) or social is chosen (right). Ch, Late Choice; Pun, Punished alcohol choice; Dif, Difference; Alc, Alcohol; Soc, Social. Table 3 shows the time windows where the statistical tests were significant. *p < 0.05. n = 11 males; diamond shapes, n = 12 females, circle shapes.

Figure 4C shows aIC activity centered around trial start for trials where alcohol is chosen comparing Late choice to Punished alcohol choice sessions. Figure 4D shows the same comparison for trials where social was chosen. This analysis shows that aIC activity during the cue period significantly decreases in the punished alcohol choice sessions. In contrast, there is no difference in aIC activity to trials where social is chosen between the two session types. Significant time windows are reported in Table 3. This pattern of activity is reflected in analysis of the mean aIC activity during the cue period shown in Figure 4E. Within-subjects ANOVA using the factors Session (Late, Pun) and Choice (Alcohol, Social) revealed a main effect of Choice (F_(1,19) = 9.5, p = 0.006) and Session (F_(1,19) = 8.3, p = 0.01), as well as a Choice by Session interaction (F_(1,19) = 7.38, p = 0.014). Post hoc t tests show that aIC activity during the cue period for alcohol choices significantly decreased from Late choice to Punished alcohol choice (paired t test Late Alcohol vs Punished Alcohol: t₍₁₉₎ = 3.48, p = 0.003), and in the Late sessions there is no significant difference between (punished) alcohol and social choice trials [paired t test Punished Alcohol vs Social (Punish sessions): t₍₂₀₎ < 1, p > 0.05].

Table 3.

Time periods (in seconds) where the photometry statistical tests are significant in Figure 4

Experimental phase	Figure	Statistical comparison	Significance time window (Seconds)
			0 = Trial start
Alcohol Chosen (Trial start)	4C	BCI: Late Choice	Figure 3D
	4C	BCI: Punished Alcohol Choice	0.44→1.70; 4.47→5.35
	4C	PT: Late v Punish	1.70→10
Social Chosen (Trial start)	4D	BCI: Late Choice	Figure 3D
	4D	BCI: Punished Alcohol Choice	0.38→10
	4D	PT: Late v Punish	n.s.
			0 = Lever press
Alcohol Choice (Response)	4F	BCI: Late Choice	Figure 3F
	4F	BCI: Punished Alcohol Choice	−10.63→10.06
	4F	PT: Late v Punish	−9.06→−8.36; −5.22→0.31; 11.51→20
Social Choice (Response)	4G	BCI: Late Choice	−12.64→20
	4G	BCI: Punished Alcohol Choice	−12.89→20
	4G	PT: Late v Punish	−12.96→−12.52; −11.89→−11.38; −2.26→6.35

Time 0 refers to trial start for C and D and for reinforced lever press for F and G. BCI, bootstrapped confidence intervals; PT, permutation test; n.s., not significant.

Figure 4, F and G, shows aIC activity centered around the choice response during the Late (4F) and Punished alcohol (4G) choice cessions, with the data separated into whether the choices made were alcohol (+shock) or social reward. Recorded traces of aIC activity show that aIC activity leading up to alcohol choices is decreased during the punished sessions compared with the prior choice sessions, but permutation tests revealed no significant difference during the outcome period. For social choices, we found that aIC activity is significantly lower just prior to and for ∼5 s after the response is made. Figure 4H shows mean aIC activity during the 3 s period after the choice was made (outcome period).

Within-subjects ANOVA using the factors Session (Late, Pun) and Choice (Alcohol, Social) revealed main effects of Choice (F_(1,18) = 10.7, p = 0.004) and Session (F_(1,18) = 10.5, p = 0.005), but no Choice by Session interaction (F_(1,19) < 1, p > 0.05). Post hoc t tests show that aIC activity during the outcome period for Alcohol in Late choice is not different to Alcohol + shock in Punished Alcohol choice sessions (Late Alcohol mean 5.25 ± 2.8, Punished Alcohol mean 4.23 ± 3.29; paired t test Late vs Pun: t₍₁₈₎ = 1.32, p > 0.05); in contrast, aIC activity during the outcome period for Social in Late choice is significantly higher than during Punished alcohol choice (Late choice mean 3.79 ± 2.72, Punished Alcohol mean 2.37 ± 1.19; paired t test Late vs Pun: t₍₂₀₎ = 2.92, p = 0.009).

Sex differences in aIC activity

We performed a secondary analysis on aIC activity centered around trial start separating the data into Male or Female subjects (Fig. S3). These data show significant differences in the recorded aIC activity during the Late choice sessions. Specifically, we found that aIC activity during the cue period was significantly higher in female rats than male rats for trials where alcohol was chosen (Fig. S3C) and where social is chosen (Fig. S3D).

Cognitive modeling of choice between alcohol and social reward

Initial analysis of raw response latencies (time from lever insertion to choice) revealed overall patterns of decision-making speed across conditions. While these latencies provide a general measure of decision efficiency (Choi et al., 2022), we subsequently employed LBA modeling to decompose these reaction times into specific cognitive components. Latency distribution for each session is shown in Figure 5A–C; for each condition, there was a log-normal distribution (Early Alcohol R² = 0.86; Late Alcohol R² = 0.98; Punished Alcohol R² = 0.72; Early Social R² = 0.87; Late Social R² = 0.93; Punish Social R² = 0.95). Mean latency per rat is shown in Figure 5D. There was no difference in response latency in Early choice (Alcohol median 11.2 ± 10.4, Social median 11.1 ± 10.6; paired t test Alcohol vs Social: t₍₂₁₎ = −0.4, p > 0.05), but Alcohol latency was significantly faster than Social in Late Choice (Alcohol median 3.0 ± 4.5, Social median 5.9 ± 7.1; paired t test Alcohol vs Social: t₍₂₀₎ = −3.0, p = 0.007). In Punished alcohol choice, there was no difference in response latency between Alcohol and Social (Alcohol median 18.3 ± 19.9, Social median 14.9 ± 17.2; paired t test Alcohol vs Social: t₍₂₀₎ < 1, p > 0.05).

Figure 5.

Cognitive modeling of choice behavior. A–C, Frequency distributions of response latency (from lever insertion) for Alcohol and Social choices across the three phases of the experiment. D, Median response latency for either alcohol or social during the three phases of the experiment. E, Model depicting the LBA model of choice. Salience (V) of alcohol or social accumulate separately to an evidence threshold (b) and thus to action selection. I, Mean and individual subject data for accumulation rates for alcohol and social choice across the experiment. F, Mean and individual subject data response caution across the experiment. *p < 0.05. n = 11 males; diamond shapes, n = 12 females, circle shapes.

To elucidate the cognitive mechanisms underlying the rats’ choices between alcohol and social reward across the experimental phases, we applied we applied a systematic “power set” of eight hierarchical LBA models to the data (see Materials and Methods). We performed a formal model comparison and found that model M_SC, in which Stimulus Bias (accumulation rates) and Caution were allowed to vary across sessions, provided the best and most parsimonious account of the data (see Table 4 and Fig. S5B for comparison details). All subsequent results are based on this winning model. Model convergence was successful for all datasets, as indicated by R-hat (R^) values consistently below 1.05 (mean R^ = 1.000, range = 1.000–1.001) and effective sample sizes (ESS) well above conventional thresholds (mean ESS = 10,340, min = 1,489) for all key parameters (see example trace plots in Fig. S4). However, posterior predictive checks revealed that while the model could successfully account for choice proportions (Fig. S6A), its ability to account for reaction times was limited, succeeding only for the stable, high-preference choices in the late acquisition phase (Fig. S6B,C). The model's failure to capture reaction times in the punished condition, despite being the best-fitting model, suggests that punishment induces a qualitative shift in cognitive strategy that violates the core assumptions of the standard LBA framework.

Table 4.

LBA model comparison results

Rank	Model name	ELPD(LOO)	SE of Diff	Weight
0	M_SC	−7,125.7	24.1	1.00
1	M_SCR	−7,131.66	24.1	0.00
2	M_S	−7,143.98	23.5	0.00
3	M_SR	−7,149.87	23.5	0.00
4	M_C	−7,465.22	16.3	0.00
5	M_CR	−7,471.31	16.3	0.00
6	M_base	−7,477.25	16.6	0.00
7	M_R	−7,482.81	16.6	0.00

ELPD (LOO) is the expected log predictive density; higher is better. SE of Diff is the standard error of the ELPD difference relative to the top-ranked model. Weight is the Akaike weight.

LBA model parameters across experimental phases

The LBA model (Fig. 5E) allowed us to extract estimates for key decision parameters: accumulation rates for alcohol (v_alcohol) and social (v_social) choices, start-point variability (A), threshold (b), and nondecision time (τ). Figure 5F shows the accumulation rate (v) for alcohol or social choices during the three phases. Based on the posterior distributions from the hierarchical model, we found no difference between the rates in early choice sessions (v_alcohol mean 0.537 ± 0.109, v_social mean 0.547 ± 0.107; paired t test Alcohol versus Social: t₍₂₂₎ = −0.218, p = 0.8293 but in the late choice sessions v_alcohol was significantly higher than v_social (v_alcohol mean 0.678 ± 0.140 v_social mean 0.463 ± 0.106; paired t test Alcohol vs Social: t₍₂₂₎ = 4.296, p < 0.001). During the punished-alcohol sessions, v_social was significantly higher than v_alcohol (v_alcohol mean 0.218 ± 0.092 v_social mean 0.807 ± 0.099; paired t test Alcohol vs Social: t₍₂₂₎ = −14.8, p < 0.001). This demonstrates that the model's accumulation rate parameter successfully tracks the observed shifts in choice preference.

In stark contrast to the behavioral slowing observed in the punishment phase, the model revealed a paradoxical effect on decision caution. Caution is shown in Figure 5G and was found to significantly decrease from the Early choice sessions to the Late choice sessions (Early mean 0.607 ± 0.063, Late mean 0.295 ± 0.029; paired t test Early vs Late: t₍₂₂₎ = 21.5, p < 0.001) and then decreased further in the Punished sessions (Punish mean 0.223 ± 0.020; paired t test Late vs Pun: t₍₂₂₎ = 10.2, p < 0.001). This counterintuitive result, where the model requires less caution to best explain behavior that is demonstrably slower and more considered, indicates that the animals’ adaptation to punishment is not a simple modulation of a standard decision parameter, but a more complex shift in cognitive state that the model cannot capture.

Relationship between choice performance and model-derived decision bias

To quantify decision bias, we used a within-subjects estimation of the relative strength of evidence accumulation toward alcohol versus social reward (accumulation rate difference: v_alcohol − v_social; hereafter Decision Bias). To provide a comparable metric in recorded aIC activity, for each rat we generated an “aIC difference” score (aIC_alcohol − aIC_social). This measure reflects a within-subjects description of the relative magnitude of aIC activity during the cue period toward alcohol or social choices.

Figure 6A shows the Preference Score for the three phases of the experiment. We found a significant increase in preference score from Early to Late (Early mean −0.03 ± 0.53, Late mean 0.39 ± 0.57; paired t test Early vs Late: t₍₂₂₎ = −3.36, p = 0.003) and a significant decrease in preference from Late to Pun (Pun mean −0.77 ± 0.36; paired t test Late vs Pun: t₍₂₂₎ = 9.4, p < 0.001). Figure 6B shows Decision Bias across the experimental phases. There was a significant increase in the Decision Bias toward positive values during the Late choice sessions (Early mean −0.01 ± 0.21, Late mean 0.21 ± 0.24; paired t test Early vs Late: t₍₂₂₎ = −4.4, p < 0.001), indicating a faster evidence accumulation for alcohol choices in the Late choice sessions. In the punished-alcohol sessions, Decision Bias was negative, and significantly lower than Late Choice (Punish mean −0.589 ± 0.19; paired t test Late vs Pun: t₍₂₂₎ = 14.7, p < 0.001). Figure 6C shows the aIC Difference Score across the three phases. There is a significant difference between Early and Late choice sessions (Early mean 0.026 ± 1.13, Late mean 1.30 ± 1.57; paired t test Early vs Late: t₍₂₂₎ = −3.5, p = 0.002) and Late choice and punished-alcohol sessions (Punish mean 0.21 ± 1.25; paired t test Late vs Pun: t₍₂₂₎ = 3.28, p = 0.003) but no difference between Early and Punished (paired t test Early vs Pun: t₍₂₂₎ = −0.5, p > 0.05). This metric reveals a bias in aIC activity during the cue period for alcohol choices in the Late choice sessions. However, these findings also show that aIC activity does not increase for social choices during the Punished alcohol choice sessions.

Figure 6.

Relationship between decision-making and aIC calcium dynamics. A, Behavior: Preference score is a within-subjects value depicting the proportion of alcohol or social reward chosen in each phase of the experiment. B, Decision Bias: Within-subjects value depicting the difference in accumulation rates for alcohol versus social (v_alcohol − v_social). C, aIC activity bias: Within-subject value depicting the difference in aIC activity during the cue period for trials where alcohol is chosen versus social (aIC_alcohol − aIC_social). D–F, Relationship between Preference and Decision bias. Correlations between Preference Score and Decision Bias for Early (D), Late (E), and Punished alcohol (F) sessions. G–I, Relationship between aIC Activity Bias and Decision Bias. Correlations between aIC and Decision Bias for Early (G), Late (G), and Punished alcohol (I) sessions. *p < 0.05. n = 11 males; diamond shapes, n = 12 females, circle shapes.

We next examined the relationship between the behavioral observation, Preference Score, and the LBA-derived Decision Bias. We found a strong positive correlation between Preference Score and Decision Bias across each of the three phases shown in Figure 6D–F: Early: r = 0.979, p < 0.001; Late: r = 0.987, p < 0.001; Pun: r = 0.950, p < 0.001. These findings validate the Decision Bias parameter as a reliable metric capturing decision-making parameters underlying the expressed choice preferences.

Relationship between Decision Bias and aIC activity during the cue period

Finally, we investigated the extent to which aIC calcium dynamics map onto these decision processes. We found that aIC Activity Bias was correlated with Decision Bias only in the Late choice sessions (Fig. 6H; r = 0.766, p < 0.001). In contrast, this relationship was not significant in the Early choice sessions (Fig. 6G; r = 0.065, p = 0.767) or in the Punished alcohol choice session (Fig. 6I; r = 0.223, p = 0.307). In the model that we used for analysis, both accumulation rate and response caution varied between the experimental phases. As such, we conducted additional analysis testing whether response caution and Activity Bias were correlated in any phase. These results show that Activity Bias was not correlated with response caution in either Early (r = −0.379, p = 0.074), Late (r = −0.259, p = 0.233), or Punished alcohol (r = 0.207, p = 0.344) choice sessions, providing evidence that aIC activity during the cue period is related to faster choices for alcohol through a process of faster accumulation rate, rather than reduced decision threshold.

Correlations between aIC activity and individual accumulation rates or latencies are shown in Figure S7. These data show that aIC calcium during the cue is significantly correlated with mean accumulation rate only for alcohol choices in the Late choice phase (Fig. S7B), and no other experimental phase, and not for social choices. These results provide further evidence for the relationship between alcohol preference and aIC activity during decision-making.

To test whether aIC activity in the 10 s period leading up to the choice response is better related to the accumulation rate than activity during the cue period, we also correlated these values (Fig. S7G–L). These data show no relationship between these measures, indicating that aIC activity during the cue period is a better predictor of accumulation rate toward alcohol choices, than an equivalent period of aIC activity preceding the actual choice response. This dissociation provides strong evidence that aIC activity during the cue sets a “Decision Bias” that influences the subsequent evidence accumulation process, rather than tracking the accumulation as it unfolds.

Discussion

In this study we measured aIC activity using calcium imaging with fiber photometry during choice between alcohol and social reward in a discrete-choice model (Lenoir et al., 2007). Replicating our (Marchant et al., 2023) and others (Augier et al., 2023) previous work, we show that both male and female rats show a choice preference for alcohol over social reward. We report four main findings. First, we show that aIC activity is higher for alcohol compared with social reward during both self-administration and choice. Second, emergence of an alcohol choice preference is associated with increased aIC activity during the cue period for trials where alcohol is chosen. Third, we show that punishment of the alcohol-reinforced response results in choice for social reward, which decreased cue-period aIC activity for alcohol choices but had no change for social choices. Finally, cognitive modeling allowed us to describe the latent cognitive mechanisms underlying alcohol over social choices. Correlating the model-derived bias with activity revealed that aIC cue-period activity is related to Decision Bias for alcohol choices when preference is well established. Overall, these findings highlight the specific role of aIC in the decision-making in choice for alcohol.

Consistent higher aIC activity toward alcohol reward compared with social reward

We found that aIC activity is more strongly associated with alcohol than social reward. We show that response-contingent alcohol elicits greater aIC calcium than social reward in both self-administration and choice. Anterior insula functions include interoception (Craig, 2009), emotional processing (Gu et al., 2013; Zych and Gogolla, 2021), and attentional modulation (Menon and Uddin, 2010; Zych and Gogolla, 2021). Greater aIC activity for alcohol compared with social outcomes may be a function of taste and visceral sensations. Insula cortex encodes visceral sensations, as well as both taste sensation and the stimuli which predict these outcomes (Samuelsen et al., 2012; Gardner and Fontanini, 2014). These observations typically come from IC regions more posterior than in this study, as visceral information from the body arrives in the posterior insula cortex (Allen et al., 1991). The rodent pIC typically shows higher activity for aversive tastes (Gehrlach et al., 2019). While aIC is also involved in aversive learning related to taste (Kayyal et al., 2019), neuroimaging studies show in humans that aIC is responsive to taste stimuli (Small, 2010; Rolls, 2016). One additional consideration is that social interaction involving touch also activates IC (Morrison et al., 2011; Gordon et al., 2013; Suvilehto et al., 2021), although again typically posterior IC. Regardless, one potential interpretational issue related to the observed aIC signal during social reward self-administration and choice is that the barrier prevents substantial touch, which may limit IC activation.

Role of aIC in decision-making and choice

Several lines of research have identified a role for aIC in decision-making (Naqvi and Bechara, 2010; Daniel et al., 2017). Here we describe how decision processes change across the experimental phases. The LBA model specifically conceptualizes choice as a race between multiple independent accumulators which each gather evidence for a specific response (Forstmann et al., 2008; Donkin et al., 2011b). A key finding of our study is the correlation between aIC activity during the cue period and Decision Bias. This presents an important question regarding the temporal alignment of these events, as our model measures response latency from lever insertion, while the decision process likely begins at cue onset. We propose that cue-evoked aIC acts as a causal antecedent to the evidence accumulation process. This activity sets the initial parameters and motivational state governing subsequent race-to-threshold, effectively establishing decision bias before choice is executed. The importance of the cue period activity to this process is further exemplified by the observation that aIC activity in the 10 s preceding the choice action (i.e., lever press) does not correlate with the accumulation rate for any choice type.

The emergence of alcohol preference was characterized by a significant increase in decision bias toward alcohol. Conversely, punishment reversed Decision Bias but also revealed a critical limitation of the model, for which the best fit required a paradoxical decrease in response caution, a finding at odds with observation. This likely reflects a fundamental shift in decision-making strategy, which may involve a switch to a different cognitive state not well captured by the standard race model, which assumes a single cognitive strategy. One possibility is that, especially in punishment, the rats alternate between a task-engaged state and a hesitant state, resulting in a mix of fast and slow reaction times, as observed.

In the Late phase, a strong and stable alcohol preference may be representative of automatic choices where the preferred outcome reduces the need for prolonged deliberation before reaching a decision threshold (Donkin et al., 2011a; Shadlen and Shohamy, 2016; Redish et al., 2022). This form of decision-making would be associated with faster response times, which is observed for both alcohol and social choices. The observation that aIC activity tracks Decision Bias for alcohol aligns with its proposed role in representing the motivational salience or expected value of outcomes (Menon and Uddin, 2010; Parkes and Balleine, 2013). In the Late phase, heightened and biased aIC activity might reflect an amplified salience attributed to alcohol cues, directly translating into a faster evidence accumulation for an alcohol choice. However, in the punished phase, the lack of correlation between aIC activity and Decision Bias, combined with the model's failure to explain the extended latency with respect to increased caution, suggests that a different decision-making process is employed, which is not related to aIC activity. Consistent with this, aIC activity during the cue period is decoupled from Decision Bias. While we did not measure freezing in this experiment, previous work has shown that punishment is unlike fear conditioning and produces very little freezing particularly after learning (Bolles et al., 1980; Jean-Richard-Dit-Bressel and McNally, 2015; Jean-Richard-Dit-Bressel et al., 2018). An additional argument against freezing is that behavioral responses were reallocated toward the social lever by the fourth session. The modeling data indicates that aIC is more related to the expression of preference-driven responses rather than the decision-making processes engaged when the alcohol choice is punished. Some studies have proposed that IC activity is related to decision-making during conflict (Naqvi et al., 2014; Daniel et al., 2017). However, these data show that aIC activity was more closely related to faster choices after preference is established.

The role of choice in addiction is not without controversy. The extent to which individuals with AUD seemingly maintain the ability to voluntarily continue alcohol use has been used to argue against the brain disease model of addiction (Heyman, 2021). However, the presence of choice does not exclude the possibility that the choices are influenced by the neurobiological consequences of alcohol use (Heilig et al., 2021; Pickard, 2022), and there are many plausible dysfunctions in decision-making in AUD (Verdejo-Garcia et al., 2018). Value-based decision-making can be a valuable framework to describe a pathway toward recovery from substance use disorders (Field et al., 2020) and potential decision processes which result in persistent alcohol use (Redish et al., 2008). The use of LBA modeling in this study aligns with the pursuit of computational validity (Redish et al., 2022), which emphasizes that translating behavioral findings across different contexts or species should be grounded in similarity of the underlying information processing and computational algorithms, rather than task resemblances or general theoretical constructs. By decomposing behavior into quantifiable model-based parameters such as decision bias and response caution, we move beyond a descriptive account of choice to probe the latent computational mechanisms driving those choices (Donkin et al., 2011b). It will be of interest in future studies to determine whether comparable metrics are observed in human laboratory tasks (Li et al., 2020; Karlsson et al., 2025).

Previous work has shown that activity in the rat IC, more posterior to recorded here, responds both to consumed rewards as well as cues which predict their arrival (Samuelsen et al., 2012; Gardner and Fontanini, 2014). This study expands on these observations to show that activity in anterior IC tracks cue-induced response anticipation and decision-making during choice. Activity in aIC contributes toward choice between goal-directed actions based on aIC-dependent representation of value (Parkes and Balleine, 2013; Parkes et al., 2015). Communication between aIC and the nucleus accumbens (NAc) core is critical for retrieving outcome values and directing choice (Parkes et al., 2015). Inhibition of the aIC to NAc pathway decreases aversion-resistant alcohol use (Seif et al., 2013), and chemogenetic stimulation increases alcohol consumption (Haaranen et al., 2020). In rodents, alcohol dependence is associated with altered function connectivity between IC and NAc (Scuppa et al., 2020), and increased connectivity between aIC and NAc is associated with AUD in humans (Grodin et al., 2018; Le et al., 2022).

Concluding remarks

In this study we show how activity in aIC relates to choice between alcohol and social reward. We found that aIC activity is consistently higher for alcohol reward both during self-administration and choice. Our modeling of the decision-making process showed that differential activity during the decision period, particularly when preference is established, reflects a stronger evidence accumulation bias toward alcohol.