All behaving animals must interpret sensory information arriving from the environment and use that information to select future actions. How the nervous system solves this problem has been a longstanding question in neuroscience. Multiple studies across a wide range of behavioral tasks and model systems, including humans (Hunt et al., 2012; Krajbich et al., 2012; Ratcliff et al., 2015), non-human primates (Gold and Shadlen, 2007; Huk and Shadlen, 2005; Shadlen and Newsome, 1996) and rodents (Brunton et al., 2013; Carandini and Churchland, 2013; Erlich et al., 2015; Hanks et al., 2015; Raposo et al., 2012; Sanders and Kepecs, 2012) have proposed a framework through which neural circuits gradually accumulate sensory evidence to guide decisions. Yet, despite the observation of neural correlates of evidence accumulation in several brain regions (Ding and Gold, 2010; Gold and Shadlen, 2007; Hanks et al., 2015; Ratcliff et al., 2007; Shadlen and Newsome, 1996), a major challenge in this line of research has been that the neural circuits that are causally responsible for evidence accumulation have not yet been determined. Two of the cortical regions that are most prominently associated with evidence accumulation, namely the posterior parietal cortex (LIP; Huk and Shadlen, 2005; Kira et al., 2015; Roitman and Shadlen, 2002; Shadlen and Newsome, 1996) and the frontal eye fields in primates (FEF; Ding and Gold, 2012a; Gold and Shadlen, 2000; Mante et al., 2013), together with its probable rodent analogue, the frontal orienting fields (FOF; Erlich et al., 2011), have been the focus of studies recently. Surprisingly, these studies have indicated that neither region is central to the computation of gradually accumulating evidence (Erlich et al., 2015; Hanks et al., 2015; Katz et al., 2016).

The anterior dorsal striatum (ADS) serves as an intriguing alternative candidate (Ding and Gold, 2013), due in part to its unique anatomical positioning as a convergence hub for multiple brain regions (Cheatwood et al., 2003; McGeorge and Faull, 1989) where neural signatures of evidence accumulation have been observed (such as the PPC and FEF/FOF; Gold and Shadlen, 2007; Cheatwood et al., 2003; Ding and Gold, 2013; McGeorge and Faull, 1989). The ADS is thus ideally positioned to participate in evidence accumulation as part of its established role in action selection (Bogacz and Gurney, 2007; Graybiel, 2008; Hikosaka et al., 2014; Jin and Costa, 2010; Nelson and Kreitzer, 2014; Redgrave et al., 2010). Modeling work has also suggested that the ADS may participate in post-accumulation decision commitment (Lo and Wang, 2006).

The auditory input to a different striatal subregion, the posterior ‘auditory’ striatum, has been shown to be critical for auditory discriminations, leading to the suggestion that cortical projections into the striatum may provide a general mechanism for the control of motor decisions (Xiong et al., 2015; Znamenskiy and Zador, 2013). Specifically with regard to evidence accumulation, work in primates found neural correlates of evidence accumulation in the ADS (Ding 2015; Ding and Gold, 2010; Ding and Gold, 2012b; Ding and Gold, 2013; Ding, 2015, Lo and Wang, 2006), and revealed that electrical microstimulation of the ADS impacts behavior that is based on accumulation of evidence (Ding and Gold, 2012a). These data led to the proposal that the ADS may contribute to the computations specifically involved in evidence accumulation. Yet three critical questions to test this proposal have been left unanswered.

First, is the ADS required for unimpaired accumulation-based decision making? To date, there have been no recorded inactivations of the dorsal striatum during the accumulation of evidence. Inactivations are important probes of whether a region plays a central causal role for a cognitive variable of interest (Newsome and Paré, 1988; Erlich et al., 2015; Katz et al., 2016).

Second, do neurons in the dorsal striatum encode sensory information in a way that is sufficient to be involved directly in the graded accumulation process? The correlates of evidence accumulation reported to date in striatum have been of two types: either firing rates that, when averaged over trials, ramp upwards with a slope of the ramp that increases as the evidence strengthens (Ding and Gold, 2010), or estimates of the temporal dynamics of firing rate variance across trials (Ding, 2015). However, the trial averages do not distinguish between graded evidence encodings and other encodings that on a single-trial basis do not represent gradually accumulating evidence, such as sharp coordinated steps in firing for which the timing of the step varies across trials (Hanks et al., 2015; Latimer et al., 2015). Furthermore, the variance estimates have not yet produced clearly definitive conclusions, suggesting that the ADS is only partly involved in graded accumulation (Ding, 2015).

We recently developed a complementary approach, distinct from the two earlier methods, to assess evidence accumulation encoding. This most recent approach estimates ‘tuning curves,’ that is, direct descriptions of the relationship between recorded neural firing rates and the graded value of the evidence accumulator, and can discriminate between different encodings that otherwise appear indistinguishable (Hanks et al., 2015). Here, we apply this approach to electrophysiological recordings from the ADS.

Third, does the dorsal striatum play a causal role throughout the period of accumulation? One of the key aspects of interest in gradual evidence accumulation is its relatively long timescale, as it occurs over a period of hundreds of milliseconds or more (thought to be a potential model of mental deliberation [Gold and Shadlen, 2007]). If the striatum is part of the gradual accumulation process that drives behavior, perturbing it at any timepoint during that accumulation process should affect behavior. This feature is thus an essential prerequisite for a component of the accumulator. However, no temporally specific perturbations of the ADS during the accumulation of evidence have yet been carried out to probe for this feature. Indeed, no brain region studied during an accumulation of evidence behavior has yet been reported to possess this feature.

Here, using a combination of behavioral, pharmacological, optogenetic, electrophysiological and computational approaches, we address these three fundamental questions. The results provide evidence supporting a central causal role for the anterior dorsal striatum in evidence accumulation.

We trained rats on a previously developed decision-making task (Brunton et al., 2013) in which subjects accumulate auditory evidence over many hundreds of milliseconds to inform a binary left/right choice (Figure 1a). In each trial, rats kept their nose in a central port during the presentation of two simultaneous trains of randomly timed auditory clicks, one played from a speaker to their left and the other from a speaker to their right. At the end of the auditory stimulus, the rat’s task was to decide which side had played the greater total number of clicks. Consistent with previous studies using this task, analysis of our rats’ behavior indicated that they gradually accumulated auditory evidence over the entire trial, and used that accumulated evidence to drive a categorical choice (Figure 1—figure supplement 1; Supplementary file 1).

Figure 1 with 2 supplements see all

Download asset Open asset

We began to assess the role of the anterior striatum in the accumulation task using reversible pharmacological inactivation with muscimol (Materials and methods). The anterior striatial region targeted in this study receives convergent inputs from the PPC and the FOF, brain regions previously reported to contain neural correlates of evidence accumulation but later shown to not be central to the accumulation process itself (Erlich et al., 2015; Hanks et al., 2015; Katz et al., 2016). Unilateral inactivation of the ADS biased rats to make more ipsilateral choices relative to controls (Figure 1b; bias for right side inactivation = 19.2 ± 4.4%, p<0.01; bias for left side inactivation = 18.6 ± 3.3%, p<0.01). This effect was not a gross motor bias, but was instead specific for accumulation trials, because no significant bias was caused on interleaved motor control trials in which the rats had to make a similar left/right motor response, but were cued by a simple visual stimulus (Figure 1—figure supplement 2; p>0.4 for both left- and right-side trials). Bilateral pharmacological inactivation caused a substantial impairment in performance for accumulation trials (Figure 1c; impairment = 12.6 ± 3.2%, p<0.01). This impairment was again specific for accumulation trials, with no significant impairment in motor control trials where the decision was not based on the accumulation of evidence over time (Figure 1—figure supplement 2; p>0.6 for both left- and right-side trials).

Psychometric curves such as those shown in Figure 1b,c group together trials based on the click difference accrued by the end of the stimulus stream and treat all trials within each group as if they were the same. But in our clicks task, we have far more information available because the precise temporal pattern of each individual trial’s click trains is known. We have previously used this information, together with a model that takes into account those known individual click times, to quantify our subjects’ behavior in terms of multiple parameters governing the dynamics of a drift-diffusion decision process (Ratcliff and McKoon, 2008). We use an enhanced model of the drift-diffusion process so that we can obtain trial-by-trial, moment-by-moment estimates of accumulating evidence (Brunton et al., 2013). This model converts the incoming stream of each trial’s discrete left and right click stimuli into a scalar quantity a(t) that represents the gradually accumulating difference between the two click streams; each right click increases the value of a(t), whereas each left click decreases a(t). Eight parameters, quantifing sensory and accumulator noise, the leakiness or instability of the accumulation process, a sticky accumulation bound, sensory depression or facilitation, side bias (þ), and lapse rate, govern the dynamics of how a(t) evolves in response to the sensory evidence pulses, and how they are then turned into a binary decision. At the end of each trial’s stimulus, the accumulator a(t_end), together with the parameter þ, drives choices: if a(t_end) > þ, the model prescribes ‘choose right’, whereas if a(t_end) < þ, the model prescribes ‘choose left’. All of the parameters are estimated by fitting the model to the rat’s behavior (Materials and methods).

The original model of Brunton et al. (2013) was not constructed to explain different types of side biases, so it had only a single parameter (þ) that could account for such lateralized effects. By adding three more parameters that could cause different types of side biases, fitting the extended model to behavioral data following unilateral inactivations, and asking which parameters are most affected relative to control trials, we can better estimate which particular aspect of the behavior was impacted by unilateral inactivations. The three side bias parameters that we consider, in addition to þ, are: asymmetric sensory input gain, asymmetric sensory input noise, and asymmetric lapse rates (Materials and methods). Considering all four of these side bias parameters in the case of unilateral inactivations of the FOF, we previously concluded that FOF inactivations were consistent with perturbing a process that was not part of evidence accumulation directly, but was instead downstream of the accumulation process and therefore followed it (Erlich et al., 2015; Piet et al., 2017).

Here, we improve upon this analysis and apply it to our striatum inactivation data. At the time of the Erlich et al. (2015) study, the complexity of determining the derivative of the model with respect to all 11 of its parameters precluded us from fitting all 11 parameters simultaneously. We instead performed exhaustive scans in the space of two parameters at a time while the other nine parameters were fixed to their control (no inactivation) values (e.g., Figure 4 in Erlich et al. (2015)). Since that time, however, algorithmic differentiation packages, which greatly facilitate computing the derivative of arbitrary differentiable models embodied in computer code, have become widely available (Abadi et al., 2016; Baydin et al., 2015; Revels et al., 2016; Al-Rfou et al., 2016). Using the ForwardDiff package of the language Julia (Revels et al., 2016) to obtain automatically the derivative with respect to all 11 parameters in the model of Erlich et al. (2015), we constructed a package that can efficiently and simultaneously fit all 11 parameters in the model. We are publishing this package in open source form, as part of the contribution of the current manuscript (code available at https://github.com/misun6312/PBupsModel.jl [Yoon and Brody, 2018]; copy archived at https://github.com/elifesciences/PBupsModel.jl). We validated this approach and our previous FOF analysis by fitting all 11 parameters simultaneously to our previous FOF unilateral inactivation data. This new analysis (Figure 2—figure supplement 2, Supplementary file 4) confirmed the conclusions about the FOF found by Erlich et al. (2015). Following this conclusion, we next turned to performing the same analysis on the inactivation data collected in the current study for the anterior striatum.

For simplicity of presentation, below, we illustrate some of the results of the model fits in terms of psychometric plots (i.e., graphing the probability of a decision to one side as a function of total #R – #L clicks, averaged over trials), but we note again that our model and its fits are sensitive to the detailed timing of the click stimuli in each individual trial, which is information that is obscured in the trial-averaged psychometric plot. As a result, the model and its fits can resolve the effects of different parameters that are indistinguishable in a psychometric plot (see also illustrations of this point in Supplementary Figure S4 in Brunton et al. (2013)). For example, a leaky (i.e., forgetful) accumulator and an increased overall lapse rate both predict an overall performance impairment. But the leaky accumulator impairment will be greater for trials that by chance had their clicks earlier rather than later, whereas the lapse impairment will be independent of the timing of each trial’s clicks. A model that is sensitive to the timing of each trial’s clicks can thus distinguish the two. Similarly, an asymmetric sensory input gain and an asymmetric lapse rate both predict a side bias. But the magnitude of the bias due to an asymmetric input gain will scale with the number of clicks presented on each trial. This contrasts with the bias that would be induced by an asymmetric lapse rate, which would be independent of the number of clicks presented. This again allows the effects of the two parameters to be distinguished. In sum, trial-by-trial and detailed click-timing effects, although not visible in the trial-averaged psychometric plot, impact the likelihood of the data under the model, and thus impact the model fits and the likelihood landscapes (such as those shown in Figure 2b and d below). When two parameters trade off in a manner that impairs our ability to distinguish them, this is revealed in the likelihood landscape as a ridge of high likelihood. The shape of the ridge quantifies the extent and scaling of the parameter trade-off (Materials and methods and for example Figure 2D in Brunton et al. (2013)).

Figure 2 with 4 supplements see all

Download asset Open asset

Simultaneously fitting all parameters of the enhanced 11-parameter model to data from sessions with unilateral muscimol inactivation of the anterior dorsal striatum revealed that two parameters differed enough from their control values to produce substantial changes in behavior (Supplementary file 2). First, the side bias in the lapse rates (the contralateral lapse rate parameter κ_C and the ipsilateral lapse rate parameter κ_I, which are unitless parameters in terms of fraction of trials; Materials and methods) significantly increased in favor of ipsilateral choices (κ_I: from 0.29, 95% C.I. = [0.18 0.43] in control sessions to 0.00, 95% C.I. = [0.00 0.14] for inactivation sessions, κ_C: from 0.20, 95% C.I. = [0.04 0.58] in control sessions to 0.60, 95% C.I. = [0.14 0.93] for inactivation sessions). An effect on lapse rates was also seen after unilateral FOF inactivations, where it was interpreted as an effect on processes subsequent to the accumulator, and not part of it (Erlich et al., 2015). Second, the magnitude of the accumulator and sensory noise parameters, which respectively describe diffusion noise intrinsic to the accumulator and noise associated with the addition of each sensory click, also increased significantly (Figure 2a,b and Figure 2—figure supplement 1). The trade-off between these parameters (Brunton et al., 2013) was large enough that it was impossible to distinguish which of the accumulator noise σ²_a or the ipsilateral and contraletral sensory noise parameters σ²_s,I and σ²_s,C was responsible for the increase. We note that we do not mean to imply that the combination of sensory and accumulator noise is a single, biologically interpretable quantity, but simply that our data cannot distinguish between the different trade-offs between these parameters that fit the data equally well. There was a suggestion that the intrinsic accumulator noise σ²_a specifically increased, with this parameter being significantly greater than zero during inactivation trials whereas it was not distinguishable from zero in control trials (Figure 2b and Figure 2—figure supplement 1), but the difference between control and inactivation σ²_a values was not significant (p<0.15). Lapse rate and noise parameters described distinct effects, and did not trade off with each other (Figure 2—figure supplement 1a, third row, middle column).

For data from bilateral inactivation sessions, the combination of sensory and accumulator noise parameters was again significantly greater than for control sessions (Figure 2c,d; w₁σ²_a + w₂σ²_s: from 31.64 clicks²/sec [95% C.I. = [17.40 55.84]] in control sessions to 117.00 clicks²/sec [95% C.I. = [71.12 156.53]] , where w₁ = 0.92, w₂ = 0.39 during inactivations. We note that noise magnitudes cannot be less than zero, implying that confidence intervals for both σ²_a and σ²_s are bounded by zero).

These fits contrast with those following FOF inactivation (Erlich et al., 2015). In particular, we note that the sensory and accumulator noise parameters were minimally altered after FOF inactivation, whereas ADS inactivation significantly impacted them.

This pharmacological demonstration that the striatum is required for unimpaired decisions that are based on the accumulation of evidence, and the model-based suggestion that the striatum affects properties of the accumulator, led us to explore the detailed neural dynamics that may support its potential causal contribution. To do so, we conducted single-unit recordings from freely behaving subjects engaged in the evidence accumulation task. Consistent with previous work (Graybiel, 2008; Jin and Costa, 2010; Kravitz and Kreitzer, 2012), we found that the neural activity of many striatal neurons was modulated by movement initiation (Figure 3—figure supplement 1a–c). However, we also found that over a third of the recorded neurons significantly modulate their activity in a side-selective manner (p<0.05) during the fixation period many hundreds of milliseconds before the movement initiation reporting the decision (64/173 [37%] of the neurons active during the fixation period [Figure 3—figure supplement 1d–f]). This timing suggests that they may have a role in forming the upcoming decision (go-left or go-right). These neurons were termed as ‘side-selective’ and for each we further defined the neuron’s prefered side as that yielding the largest activation, as done previously by Hanks et al. (2015). Consistent with previous work in primate dorsal striatum (Ding and Gold, 2010), we found that the average responses of these rat striatum neurons ramped upwards for stimuli in the preferred direction (Figure 3), and moreover, that after an initial onset latency, the slope of the ramp was proportional to the stimulus strength (Figure 3; Figure 4a). Importantly, however, a gradual ramping profile is not conclusive evidence for the encoding of gradually accumulating evidence, because such a response profile can also be consistent with other encoding schemes (Ditterich, 2006; Hanks et al., 2015; Latimer et al., 2015) such as step changes in firing rate that occur at different times in different trials (Latimer et al., 2015). Thus, we extended our analysis to include a more direct test in which the influence of single quanta of sensory evidence on the responses of the cells is quantitatively assessed.

Figure 3 with 1 supplement see all

Download asset Open asset

Figure 4 with 1 supplement see all

Download asset Open asset

If indeed temporal integration underlies the ramping activity of the striatal cells, then each single quantum of sensory evidence (an auditory click) should result in a fixed-magnitude and a sustained increase in the neuron’s firing rate (Figure 4b, model) (Hanks et al., 2015; Huk and Shadlen, 2005). We thus estimated the effect of each sensory evidence quantum by computing the click-triggered average response of the side-selective striatal neurons. We found that striatal neurons modulated their activity in close agreement with this theoretical prediction (Figure 4b, data), arguing in favor of a role of this anterior striatal subregion in the behavioral accumulation of evidence process.

We also took advantage of a recently developed method, i.e., direct estimates of firing rates as a function of accumulated evidence, to compute neural tuning curves (Hanks et al., 2015). Model-derived estimates of the moment-by-moment value of the accumulating evidence in each trial are collated with simultaneously recorded firing rates to generate tuning curves for accumulated evidence (see Hanks et al. (2015)), Materials and methods, and the illustration of the method in Figure 4—figure supplement 1). When applying this analysis to the striatal data, we found that the side-selective neurons encoded accumulating evidence in a remarkably graded manner throughout the period of evidence accumulation (Figure 4c,d). This graded encoding was consistent across different neurons in the population of recorded striatal cells (Figure 5). Such a graded representation implies that the striatum carries information about the graded value of accumulated evidence, as would be necessary for a brain structure involved in such a process.

Figure 5

Download asset Open asset

Our pharmacological methods address the questions of whether the anterior dorsal striatum is involved in the process of accumulation of evidence, and our electrophysiological and computational methods address how the anterior dorsal striatum represents the accumulation of sensory evidence. However, neither directly addresses the question of when the anterior dorsal striatum is involved. This question is critical and has proven to be pivotal in assessing the involvement of a brain region in the evidence accumulation process. For example, some brain regions can be required for decisions that are based on accumulation of evidence, yet contribute at times suggesting that they are instead required for processes that are subsequent to the gradual accumulation of evidence itself (Erlich et al., 2015; Hanks et al., 2015). No region to date has been reported to be required at points of time that fully coincide with the evidence accumulation period.

To delineate the precise timing of the anterior dorsal striatum’s contribution, we used optogenetic inactivation, mediated by halorhodopsin (eNpHR3.0), to unilaterally and transiently inactivate this region during the Poisson Clicks task. We expressed eNpHR3.0 using viral delivery methods (Figure 6a; Materials and methods). Acute neural recordings in our experimental rats verified that we could indeed transiently silence neural activity in the striatum with fine temporal precision using the delivery of green light (Figure 6b). We began with full-trial unilateral optogenetic inactivation and found, in agreement with the pharmacological inactivation described above, that optogenetic manipulation resulted in more ipsilateral choice biases relative to control trials, which in this case were randomly interleaved with the inactivation trials (Figure 6c; bias = 9.0 ± 2.3%, p<0.01). These effects were consistent across rats (Figure 6d). Control rats whose striatum was injected with the same virus expressing YFP alone did not show a behavioral bias (bias = 0.1 ± 1.8%, p=0.89). Next, to resolve directly when the striatum contributes to the auditory accumulation of evidence task, we transiently inactivated it unilaterally during one of four different 500 ms time periods during the task: (i) the delay period immediately preceding stimulus onset (‘pre-accumulation’), (ii) the first half of a 1 s sensory stimulus (‘first half’), (iii) the second half of a 1 s sensory stimulus (‘second half’), or (iv) the movement period (‘post-choice’). In contrast to similar inactivation assays of the cortical FOF, which have no effect during the early parts of the accumulation period (Hanks et al., 2015), we found that transient optogenetic inactivation of the anterior dorsal striatum during both the first half and second half of the accumulation caused a significant bias for the ipsilateral choices, with a similar magnitude of effect in these two periods (first half bias = 10.4 ± 4.0, p<0.01; second half bias = 12.9 ± 3.7%, p<0.01; difference = 2.5 ± 2.8, p=0.2; Figure 6e; the first-half effect in the striatum is significantly greater than that in the FOF, p<0.01, Figure 7). Remarkably, the effect in striatum was limited to the stimulus presentation period and we found no significant effect of optogenetic inactivation during the pre-accumulation or post-choice periods (pre-accumulation bias = 0.4 ± 5.4%, p=0.42; post-choice bias = 0.9 ± 5.2%, p=0.38; Figure 6e). These results are consistent with the idea that the anterior dorsal striatum plays a direct causal role throughout the entire evidence accumulation process.

Figure 6

Download asset Open asset

Figure 7

Download asset Open asset

Studies carried out over more than two decades have attempted to elucidate neural circuits that underlie the accumulation of evidence over time (starting with Shadlen and Newsome (1996); see Gold and Shadlen (2007), Carandini and Churchland (2013), Brody and Hanks (2016), and Hanks and Summerfield (2017) for reviews; see also Carandini and Churchland (2013), Gold and Shadlen (2007), Krajbich et al. (2012), Shadlen and Newsome (1996). Ding and colleagues have shown that microstimulation of the ADS perturbs decisions based on the accumulation of evidence (Ding and Gold, 2010, 2012a; Ding, 2015; but see Histed et al., 2009 and Tehovnik and Slocum, 2013) for discussion as to whether or not microstimulation primarily affects axon terminals, which would add complications for its interpretation when localizing neural function). Despite these many years of studies, no brain region has previously been identified as: first, being required for unimpaired accumulation-based decision-making behavior; second, having the graded neural encoding required for direct involvement in computing the graded, gradually evolving, value of the accumulating evidence; and third, making a causal contribution throughout times that fully coincide with the accumulation process. By demonstrating that the anterior dorsal striatum satisfies all three of these criteria, our work suggests that the anterior dorsal striatum is the first identifiable node in the neural circuit causally responsible for computing evidence accumulation. The anterior dorsal striatum is well positioned anatomically to participate in evidence accumulation as it receives diverse convergent anatomical input from multiple cortical areas (Cheatwood et al., 2003; McGeorge and Faull, 1989) and it is connected via recurrent loops with cortical and subcortical areas that are widely believed to play a role in action selection (Ding and Gold, 2013). Whether the anterior dorsal striatum possesses a unique role in evidence accumulation, or whether it is an important node of a more extended network of brain regions that operate in coordination to mediate evidence accumulation, remains to be resolved. Corticostriatal loops are organized as distinct parallel circuits (Alexander et al., 1986; Kim and Hikosaka, 2015); future studies dissecting the contribution of different loops will be important for resolving this major question.

Our results, together with those of Ding and colleagues (Ding and Gold, 2010; Ding and Gold, 2012a; Ding, 2015), suggest that the striatum may be directly involved in a more expansive set of computations, traditionally considered to be more cognitive in nature, than the already well-established functions of the dorsal striatum in action selection, response initiation, evaluation of reward uncertainty, and habit formation (Ding and Gold, 2010; Graybiel, 2008; Hikosaka et al., 2014; Jin and Costa, 2010; Kravitz and Kreitzer, 2012). It will be important to better understand how the striatal involvement in computing accumulation of evidence, as identified in this study, may contribute to those previously established functions. Extensions to our paradigm (for example, free response protocols or more extended trial durations) are likely to be useful for reconciling the functions indciated by results with the other functions of the striatum. The computations involved in evidence accumulation may perhaps provide an efficient mechanism for extracting important pieces of information from the environment in the service of other roles of the striatum.

By identifying model parameters affected by the inactivations, our model fits suggest specific aspects of the evidence accumulation computation that could be prioritized as potentially particularly strongly related to the ADS’s role in the computation. The model fits to unilateral pharmacological inactivation data found that, similar to unilateral inactivations of the FOF, the side bias in lapse rates was increased by the inactivations. But in contrast to what occurs in the FOF, noise parameters, including the accumulator’s intrinsic noise, were also substantially increased after striatal inactivation (Figure 2a,b). Model fits to bilateral anterior dorsal striatum inactivation data found that the sum of sensory and accumulator noise magnitude parameters was significantly increased by striatal inactivation (Figure 2c,d). A parsimonious account suggests that the main noise parameter that is affected may perhaps be the magnitude of the noise in the evidence accumulator. This would be consistent with the idea, supported by our electrophysiological and optogenetic data, that the striatum plays a role in the accumulation process. The lack of a significant effect on other parameters should be treated with caution: it remains possible that future studies with greater statistical power could discern an effect of striatal inactivation on some of these other parameters. Nevertheless, even while we emphasize that we do not take the modeling results on their own as conclusive, they do suggest that accumulator noise is a principal parameter of interest. It is conceivable that bilateral striatal inactivation increases accumulator noise by destabilizing the accumulator’s representation without biasing it, but a circuit model hypothesis to explain precisely how the striatum might affect the accumulator noise level remains to be developed. Another important direction for future studies will be the development of models with temporally specific parameters that could be used to model the effects of temporally specific optogenetic inactivation appropriately.

Independently of whether the anterior dorsal striatum operates alone or as part of a broader circuit for computing gradual evidence accumulation, and independently of the precise nature of its contribution to the evidence accumulation computation, the data reported here provide a critical foothold towards delineating the relevant causal circuit: for example, the anterior dorsal striatum’s major inputs and outputs become important candidate regions to be examined for a potential role in the process. The possibility that the causal circuit for computing evidence accumulation may be delineated in the near future suggests that we will soon be able to elucidate the circuit and cellular mechanisms that support evidence accumulation, a computation that is crucial for decision-making behavior in a wide range of species, including humans.

All data generated or analysed during this study are included in the manuscript and supporting files.

1. Conference

   1. Abadi M
   2. Barham P
   3. Chen J
   4. Chen Z
   5. Davis A
   6. Dean J

   (2016)

   TensorFlow: A system for large-sale machine learning

   Proceedings of the12th USENIX Symposium on Operating Systems Designand Implementation (OSDI ’16).

   • Google Scholar
2. Preprint

   1. Al-Rfou R
   2. Alain G
   3. Almahairi A
   4. Angermueller C
   5. Bahdanau D
   6. Ballas N
   7. Bastien F
   8. Bayer J
   9. Belikov A
   10. The Theano Development Team

   (2016) Theano: A Python framework for fast computation of mathematical expressions

   arXiv.

   https://arxiv.org/abs/1605.02688

   • Google Scholar
3. 1. Alexander GE
   2. DeLong MR
   3. Strick PL

   (1986) Parallel organization of functionally segregated circuits linking basal ganglia and cortex

   Annual Review of Neuroscience 9:357–381.

   https://doi.org/10.1146/annurev.ne.09.030186.002041

   • PubMed
   • Google Scholar
4. Preprint

   1. Baydin AG
   2. Pearlmutter BA
   3. Radul AA
   4. Siskind JM

   (2015) Automatic differentiation in machine learning: a survey

   Arxiv.

   https://arxiv.org/abs/1502.05767

   • Google Scholar
5. 1. Bogacz R
   2. Gurney K

   (2007) The basal ganglia and cortex implement optimal decision making between alternative actions

   Neural Computation 19:442–477.

   https://doi.org/10.1162/neco.2007.19.2.442

   • PubMed
   • Google Scholar
6. 1. Brody CD
   2. Hanks TD

   (2016) Neural underpinnings of the evidence accumulator

   Current Opinion in Neurobiology 37:149–157.

   https://doi.org/10.1016/j.conb.2016.01.003

   • PubMed
   • Google Scholar
7. 1. Brunton BW
   2. Botvinick MM
   3. Brody CD

   (2013) Rats and humans can optimally accumulate evidence for decision-making

   Science 340:95–98.

   https://doi.org/10.1126/science.1233912

   • PubMed
   • Google Scholar
8. 1. Carandini M
   2. Churchland AK

   (2013) Probing perceptual decisions in rodents

   Nature Neuroscience 16:824–831.

   https://doi.org/10.1038/nn.3410

   • PubMed
   • Google Scholar
9. 1. Cheatwood JL
   2. Reep RL
   3. Corwin JV

   (2003) The associative striatum: cortical and thalamic projections to the dorsocentral striatum in rats

   Brain Research 968:1–14.

   https://doi.org/10.1016/S0006-8993(02)04212-9

   • PubMed
   • Google Scholar
10. 1. Ding L
    2. Gold JI

    (2010) Caudate encodes multiple computations for perceptual decisions

    Journal of Neuroscience 30:15747–15759.

    https://doi.org/10.1523/JNEUROSCI.2894-10.2010

    • PubMed
    • Google Scholar
11. 1. Ding L
    2. Gold JI

    (2012a) Neural correlates of perceptual decision making before, during, and after decision commitment in monkey frontal eye field

    Cerebral Cortex 22:1052–1067.

    https://doi.org/10.1093/cercor/bhr178

    • PubMed
    • Google Scholar
12. 1. Ding L
    2. Gold JI

    (2012b) Separate, causal roles of the caudate in Saccadic choice and execution in a perceptual decision task

    Neuron 75:865–874.

    https://doi.org/10.1016/j.neuron.2012.07.021

    • PubMed
    • Google Scholar
13. 1. Ding L
    2. Gold JI

    (2013) The basal ganglia's contributions to perceptual decision making

    Neuron 79:640–649.

    https://doi.org/10.1016/j.neuron.2013.07.042

    • PubMed
    • Google Scholar
14. 1. Ding L

    (2015) Distinct dynamics of ramping activity in the frontal cortex and caudate nucleus in monkeys

    Journal of Neurophysiology 114:1850–1861.

    https://doi.org/10.1152/jn.00395.2015

    • PubMed
    • Google Scholar
15. 1. Ditterich J

    (2006) Stochastic models of decisions about motion direction: behavior and physiology

    Neural Networks 19:981–1012.

    https://doi.org/10.1016/j.neunet.2006.05.042

    • PubMed
    • Google Scholar
16. 1. Erlich JC
    2. Bialek M
    3. Brody CD

    (2011) A cortical substrate for memory-guided orienting in the rat

    Neuron 72:330–343.

    https://doi.org/10.1016/j.neuron.2011.07.010

    • PubMed
    • Google Scholar
17. 1. Erlich JC
    2. Brunton BW
    3. Duan CA
    4. Hanks TD
    5. Brody CD

    (2015) Distinct effects of prefrontal and parietal cortex inactivations on an accumulation of evidence task in the rat

    eLife 4:e05457.

    https://doi.org/10.7554/eLife.05457

    • Google Scholar
18. 1. Gold JI
    2. Shadlen MN

    (2000) Representation of a perceptual decision in developing oculomotor commands

    Nature 404:390–394.

    https://doi.org/10.1038/35006062

    • PubMed
    • Google Scholar
19. 1. Gold JI
    2. Shadlen MN

    (2007) The neural basis of decision making

    Annual Review of Neuroscience 30:535–574.

    https://doi.org/10.1146/annurev.neuro.29.051605.113038

    • PubMed
    • Google Scholar
20. 1. Graybiel AM

    (2008) Habits, rituals, and the evaluative brain

    Annual Review of Neuroscience 31:359–387.

    https://doi.org/10.1146/annurev.neuro.29.051605.112851

    • PubMed
    • Google Scholar
21. 1. Hanks TD
    2. Kopec CD
    3. Brunton BW
    4. Duan CA
    5. Erlich JC
    6. Brody CD

    (2015) Distinct relationships of parietal and prefrontal cortices to evidence accumulation

    Nature 520:220–223.

    https://doi.org/10.1038/nature14066

    • PubMed
    • Google Scholar
22. 1. Hanks TD
    2. Summerfield C

    (2017) Perceptual decision making in rodents, monkeys, and humans

    Neuron 93:15–31.

    https://doi.org/10.1016/j.neuron.2016.12.003

    • PubMed
    • Google Scholar
23. 1. Hikosaka O
    2. Kim HF
    3. Yasuda M
    4. Yamamoto S

    (2014) Basal ganglia circuits for reward value-guided behavior

    Annual Review of Neuroscience 37:289–306.

    https://doi.org/10.1146/annurev-neuro-071013-013924

    • PubMed
    • Google Scholar
24. 1. Histed MH
    2. Bonin V
    3. Reid RC

    (2009) Direct activation of sparse, distributed populations of cortical neurons by electrical microstimulation

    Neuron 63:508–522.

    https://doi.org/10.1016/j.neuron.2009.07.016

    • PubMed
    • Google Scholar
25. 1. Huk AC
    2. Shadlen MN

    (2005) Neural activity in macaque parietal cortex reflects temporal integration of visual motion signals during perceptual decision making

    Journal of Neuroscience 25:10420–10436.

    https://doi.org/10.1523/JNEUROSCI.4684-04.2005

    • PubMed
    • Google Scholar
26. 1. Hunt LT
    2. Kolling N
    3. Soltani A
    4. Woolrich MW
    5. Rushworth MF
    6. Behrens TE

    (2012) Mechanisms underlying cortical activity during value-guided choice

    Nature Neuroscience 15:470–476.

    https://doi.org/10.1038/nn.3017

    • PubMed
    • Google Scholar
27. 1. Jin X
    2. Costa RM

    (2010) Start/stop signals emerge in nigrostriatal circuits during sequence learning

    Nature 466:457–462.

    https://doi.org/10.1038/nature09263

    • PubMed
    • Google Scholar
28. 1. Jones BL
    2. Nagin DS
    3. Roeder K

    (2001) A SAS procedure based on mixture models for estimating developmental trajectories

    Sociological Methods & Research 29:374–393.

    https://doi.org/10.1177/0049124101029003005

    • Google Scholar
29. 1. Katz LN
    2. Yates JL
    3. Pillow JW
    4. Huk AC

    (2016) Dissociated functional significance of decision-related activity in the primate dorsal stream

    Nature 535:285–288.

    https://doi.org/10.1038/nature18617

    • PubMed
    • Google Scholar
30. 1. Kim HF
    2. Hikosaka O

    (2015) Parallel basal ganglia circuits for voluntary and automatic behaviour to reach rewards

    Brain 138:1776–1800.

    https://doi.org/10.1093/brain/awv134

    • PubMed
    • Google Scholar
31. 1. Kira S
    2. Yang T
    3. Shadlen MN

    (2015) A neural implementation of Wald's sequential probability ratio test

    Neuron 85:861–873.

    https://doi.org/10.1016/j.neuron.2015.01.007

    • PubMed
    • Google Scholar
32. 1. Krajbich I
    2. Lu D
    3. Camerer C
    4. Rangel A

    (2012) The attentional drift-diffusion model extends to simple purchasing decisions

    Frontiers in Psychology 3:193.

    https://doi.org/10.3389/fpsyg.2012.00193

    • PubMed
    • Google Scholar
33. 1. Kravitz AV
    2. Kreitzer AC

    (2012) Striatal mechanisms underlying movement, reinforcement, and punishment

    Physiology 27:167–177.

    https://doi.org/10.1152/physiol.00004.2012

    • PubMed
    • Google Scholar
34. 1. Latimer KW
    2. Yates JL
    3. Meister ML
    4. Huk AC
    5. Pillow JW

    (2015) NEURONAL MODELING. Single-trial spike trains in parietal cortex reveal discrete steps during decision-making

    Science 349:184–187.

    https://doi.org/10.1126/science.aaa4056

    • PubMed
    • Google Scholar
35. 1. Lo CC
    2. Wang XJ

    (2006) Cortico-basal ganglia circuit mechanism for a decision threshold in reaction time tasks

    Nature Neuroscience 9:956–963.

    https://doi.org/10.1038/nn1722

    • PubMed
    • Google Scholar
36. Book

    1. MacKay DJ

    (2003)

    Information Theory, Inference and Learning Algorithms.

    Cambridge University Press.

    • Google Scholar
37. 1. Mante V
    2. Sussillo D
    3. Shenoy KV
    4. Newsome WT

    (2013) Context-dependent computation by recurrent dynamics in prefrontal cortex

    Nature 503:78–84.

    https://doi.org/10.1038/nature12742

    • PubMed
    • Google Scholar
38. 1. McGeorge AJ
    2. Faull RL

    (1989) The organization of the projection from the cerebral cortex to the striatum in the rat

    Neuroscience 29:503–537.

    https://doi.org/10.1016/0306-4522(89)90128-0

    • PubMed
    • Google Scholar
39. 1. Nagin DS

    (1999) Analyzing developmental trajectories: a semiparametric, group-based approach

    Psychological Methods 4:139–157.

    https://doi.org/10.1037/1082-989X.4.2.139

    • Google Scholar
40. 1. Nelson AB
    2. Kreitzer AC

    (2014) Reassessing models of basal ganglia function and dysfunction

    Annual Review of Neuroscience 37:117–135.

    https://doi.org/10.1146/annurev-neuro-071013-013916

    • PubMed
    • Google Scholar
41. 1. Newsome WT
    2. Paré EB

    (1988) A selective impairment of motion perception following lesions of the middle temporal visual area (MT)

    The Journal of Neuroscience 8:2201–2211.

    https://doi.org/10.1523/JNEUROSCI.08-06-02201.1988

    • PubMed
    • Google Scholar
42. 1. Piet AT
    2. Erlich JC
    3. Kopec CD
    4. Brody CD

    (2017) Rat prefrontal cortex inactivations during decision making are explained by bistable attractor dynamics

    Neural Computation 29:2861–2886.

    https://doi.org/10.1162/neco_a_01005

    • PubMed
    • Google Scholar
43. 1. Raposo D
    2. Sheppard JP
    3. Schrater PR
    4. Churchland AK

    (2012) Multisensory decision-making in rats and humans

    Journal of Neuroscience 32:3726–3735.

    https://doi.org/10.1523/JNEUROSCI.4998-11.2012

    • PubMed
    • Google Scholar
44. 1. Ratcliff R
    2. Hasegawa YT
    3. Hasegawa RP
    4. Smith PL
    5. Segraves MA

    (2007) Dual diffusion model for single-cell recording data from the superior colliculus in a brightness-discrimination task

    Journal of Neurophysiology 97:1756–1774.

    https://doi.org/10.1152/jn.00393.2006

    • PubMed
    • Google Scholar
45. 1. Ratcliff R
    2. McKoon G

    (2008) The diffusion decision model: theory and data for two-choice decision tasks

    Neural Computation 20:873–922.

    https://doi.org/10.1162/neco.2008.12-06-420

    • PubMed
    • Google Scholar
46. 1. Ratcliff R
    2. Smith PL

    (2015)

    Oxford Handbook of Computational and Mathematical Psychology

    Modeling simple decisions and applications using a diffusion model, Oxford Handbook of Computational and Mathematical Psychology, New York, Oxford University Press.

    • Google Scholar
47. 1. Redgrave P
    2. Rodriguez M
    3. Smith Y
    4. Rodriguez-Oroz MC
    5. Lehericy S
    6. Bergman H
    7. Agid Y
    8. DeLong MR
    9. Obeso JA

    (2010) Goal-directed and habitual control in the basal ganglia: implications for parkinson's disease

    Nature Reviews Neuroscience 11:760–772.

    https://doi.org/10.1038/nrn2915

    • PubMed
    • Google Scholar
48. Preprint

    1. Revels J
    2. Lubin M
    3. Papamarkou T

    (2016) Forward-Mode automatic differentiation in julia

    Arxiv.

    https://arxiv.org/abs/1607.07892

    • Google Scholar
49. 1. Roitman JD
    2. Shadlen MN

    (2002) Response of neurons in the lateral intraparietal area during a combined visual discrimination reaction time task

    The Journal of Neuroscience 22:9475–9489.

    https://doi.org/10.1523/JNEUROSCI.22-21-09475.2002

    • PubMed
    • Google Scholar
50. 1. Sanders JI
    2. Kepecs A

    (2012) Choice ball: a response interface for two-choice psychometric discrimination in head-fixed mice

    Journal of Neurophysiology 108:3416–3423.

    https://doi.org/10.1152/jn.00669.2012

    • PubMed
    • Google Scholar
51. 1. Shadlen MN
    2. Newsome WT

    (1996) Motion perception: seeing and deciding

    PNAS 93:628–633.

    https://doi.org/10.1073/pnas.93.2.628

    • PubMed
    • Google Scholar
52. 1. Tehovnik EJ
    2. Slocum WM

    (2013) Two-photon imaging and the activation of cortical neurons

    Neuroscience 245:12–25.

    https://doi.org/10.1016/j.neuroscience.2013.04.022

    • PubMed
    • Google Scholar
53. 1. Xiong Q
    2. Znamenskiy P
    3. Zador AM

    (2015) Selective corticostriatal plasticity during acquisition of an auditory discrimination task

    Nature 521:348–351.

    https://doi.org/10.1038/nature14225

    • PubMed
    • Google Scholar
54. Software

    1. Yoon AM
    2. Brody CD

    (2018) PBupsModel.jl, version 3411019

    Github.

    https://github.com/misun6312/PBupsModel.jl
55. 1. Znamenskiy P
    2. Zador AM

    (2013) Corticostriatal neurons in auditory cortex drive decisions during auditory discrimination

    Nature 497:482–485.

    https://doi.org/10.1038/nature12077

    • PubMed
    • Google Scholar

Author details

1. Michael M Yartsev

   1. Princeton Neuroscience Institute, Princeton, United States
   2. Department of Bioengineering, Helen Wills Neuroscience Institute, Berkeley, United States

   Contribution

   Conceptualization, Resources, Data curation, Software, Formal analysis, Supervision, Funding acquisition, Validation, Investigation, Visualization, Methodology, Writing—original draft, Project administration, Writing—review and editing

   Contributed equally with

   Timothy D Hanks

   For correspondence

   myartsev@berkeley.edu

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0003-0952-2801

2. Timothy D Hanks

   1. Princeton Neuroscience Institute, Princeton, United States
   2. Department of Neurology, University of California, Davis, Sacramento, United States
   3. Center for Neuroscience, University of California, Davis, Davis, United States

   Contribution

   Conceptualization, Resources, Data curation, Software, Formal analysis, Supervision, Funding acquisition, Validation, Investigation, Visualization, Methodology, Writing—original draft, Project administration, Writing—review and editing

   Contributed equally with

   Michael M Yartsev

   For correspondence

   thanks@ucdavis.edu

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0003-4147-4475

3. Alice Misun Yoon

   Princeton Neuroscience Institute, Princeton, United States

   Contribution

   Software, Formal analysis, Visualization, Writing—original draft

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0001-7832-2796

4. Carlos D Brody

   1. Princeton Neuroscience Institute, Princeton, United States
   2. Howard Hughes Medical Institute, Maryland, United States

   Contribution

   Conceptualization, Supervision, Funding acquisition, Writing—review and editing

   For correspondence

   brody@princeton.edu

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-4201-561X

• 5,801

  views

• 940

  downloads

• 119

  citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.