Prefrontal cortex

In Brief:

Chafee and Heilbronner describe how the complex functions and anatomy of the prefrontal cortex have been untangled over the last century.

The prefrontal cortex is a well-studied, but in terms of understanding what it is for, deeply divisive part of the brain located at the front of the head. Perhaps the least controversial feature of the prefrontal cortex is its complexity. The prefrontal cortex is anatomically, functionally, and computationally complex. It is anatomically complex, containing a number of subregions each sending and receiving projections to a unique set of other cortical and subcortical areas. This interconnectivity presents a serious challenge to efforts to localize function to prefrontal cortex, because it can seem as though information flows everywhere all at once in prefrontal networks. Perhaps as a result, prefrontal cortex is also computationally complex: working memory, abstraction, sensory attention, value-based decision making, planning, and motor control are all functions that have been attributed to the prefrontal cortex. This diversity of functions is likely to reflect the diversity of brain regions that prefrontal cortex communicates with while carrying out the computations it performs to influence behavior.

The task then, if one is to understand prefrontal cortex, is to identify ideas or principles that encompass and explain as much of the complexity of prefrontal cortex as possible. In this Primer, we borrow heavily from influential researchers who have approached prefrontal cortex from different perspectives. Although their ideas have been refined over time, they have not been disproven. Our goal is to synthesize these views and integrate them with recent data.

The motivating thesis of this Primer and our point of departure derives from the work of Patricia Goldman-Rakic and Joaquin Fuster. Beginning their independent work in the 1970s and 1980s, these two investigators sought to understand the functions of prefrontal cortex by characterizing how individual prefrontal neurons in nonhuman primates are activated during behavior. This provided insight into the nature of the computations that are carried out by prefrontal cortex. A combination of their ideas can be summarised in the following hybrid organizing principle: the function of the prefrontal cortex is to generate persistent internal representations (Goldman-Rakic) that organize behavior in time (Fuster).

An internal representation is a pattern of neural activity that encodes the state of the environment or of the self. Internal representations are not representations of current stimuli or actions: rather, they can be sustained independently of the timing of the external sensory and motor events that they represent (we will see examples below). This property frees us from the present moment, giving human behavior its autonomy and self-generated quality. Language exemplifies this: when we express ourselves, we self-generate a sequence of internal representations that can diverge in timing and content from what is happening around us. Notably, damage to prefrontal cortex produces Broca’s aphasia, marked by difficulty in finding words during speech.

Is this all a good thing? One could argue that the temporal independence that internal representations provide has authored much human suffering. Whereas a wellness coach might instruct us to ‘live in the moment’, prefrontal cortex seems to have evolved for the opposite purpose. It allows us represent the future (with all of its pitfalls), as well as ruminate on the past (and its potential for regret). But there are profound benefits. Prefrontal cortex has organized human behavior in time to great effect. Civilization likely owes a good deal to the temporal independence from the here and now afforded to human thinking by the expansion of prefrontal cortex during our evolution.

What and where is prefrontal cortex?

There are multiple perspectives on which brain areas constitute the prefrontal cortex. The simplest definition is that the prefrontal cortex consists of everything in the frontal lobe excluding motor/premotor cortices. Alternatively, it may be just the part of this large region that is cytoarchitectonically granular — has a well-defined layer IV — or that connects primarily with the mediodorsal nucleus of thalamus. Each of these definitions will also have separate implications for the question of which animals have a prefrontal cortex (for example, mice do not have a granular frontal cortex, but do have frontal areas that connect with the mediodorsal thalamus) (see Preuss and Wise, 2022 for a thorough account).

One important takeaway is that most everyone would agree that the prefrontal cortex is not unitary: its subregions have different connections, cytoarchitecture, and function, and some of these regions are unique to particular species. We can divide the primate prefrontal cortex into broad regions: the orbitofrontal cortex; the dorsal prefrontal cortex; the ventrolateral prefrontal cortex; the ventromedial prefrontal cortex; the anterior cingulate cortex; and the frontal pole (Figure 1).

Figure 1. Location of the prefrontal cortex and its subregions in the monkey.

ACC, anterior cingulate cortex; dPFC, dorsal prefrontal cortex; FP, frontal pole; M1, primary motor cortex; OFC, orbitofrontal cortex; PM, premotor cortex; vLPFC, ventrolateral prefrontal cortex; vmPFC, ventromedial prefrontal cortex

Working memory

Early research into prefrontal cortex function asked what happens when it is damaged. Investigators focused particularly on the dorsal prefrontal cortex in nonhuman primates. Monkeys deprived of dorsal prefrontal cortex seemed mostly normal in appearance and behavior, but deficits were revealed with rigorous testing. In 1935, Jacobsen noticed that prefrontal damage produced problems in laboratory task performance that depended on recent memory. So began the association between prefrontal cortex, and especially dorsal prefrontal cortex, and working memory. For example, monkeys viewed a piece of food placed in one of two wells, which were subsequently covered (Figure 2A); several seconds later, they were allowed to retrieve the reward, if they could remember where it was. Thus, at the time of the response, there was no stimulus telling them which well to uncover. In this case, monkeys with dorsal prefrontal lesions made random choices, as though they had no memory of which well they recently saw baited.

Figure 2. Working memory and internal representation by primate prefrontal neurons.

(A) Behavioral demonstration of working memory. (1) Monkeys view a bait food item placed in one of two food wells. (2) The wells are covered. (3) Monkeys maintain the spatial location of the bait in working memory. (4) Monkeys demonstrate successful spatial working memory by later choosing one of the two wells to uncover. (B) Individual prefrontal neurons exhibit preferences for different stimuli based on the physical attributes of the stimulus, such as location (top panel), or by increasingly abstract representations of the stimulus, such as category (middle panel), or number (lower panel). (C) Firing rate of a hypothetical prefrontal neuron is elevated when a preferred stimulus is presented in a video display (left). The duration of the visual stimulus (‘Cue’) is indicated by blue shading. The elevated firing rate of the neuron is sustained after the stimulus disappears throughout a subsequent delay period while the stimulus is stored in working memory. Firing rate returns to baseline after the stored information is used to guide the motor response. Firing rate in the same neuron remains at baseline if a nonpreferred stimulus is shown (right).

What could be the underlying mechanism in the prefrontal cortex supporting such a function? In the 1970s and 1980s, Fuster, Goldman-Rakic, and colleagues recorded the action potentials of individual dorsal prefrontal neurons in monkeys performing delayed response tasks (Funahashi et al., 1989; Fuster et al., 1971). Monkeys were trained to remember the location of a brief cue, to wait 3 seconds after its disappearance (the delay period), and then to indicate, based on memory, where the cue had appeared. They made two key observations. First, prefrontal neurons fired action potentials at a persistently elevated rate during the delay period while working memory was engaged. Second, the persistent activation of prefrontal neurons depended on the information stored in working memory, in this case, the spatial location of the remembered stimulus. Presentation of the cue in different locations led to the persistent activation of different subsets of prefrontal neurons with different spatial preferences. The activity of these neurons thus matched working memory both in duration and in information content (Figure 2B, C).

As Goldman-Rakic put it at the time, ‘Out of sight, in mind’. While the persistent activity of prefrontal neurons provides a clear parallel to the operation of working memory, more recently it has become a matter of debate whether this is in fact the neural mechanism. Persistent activity is most evident in data averaged across many trials. Lundqvist, Miller and colleagues analyzed prefrontal activity on single trials and demonstrated that, rather than maintaining a persistently elevated firing rate, prefrontal neurons fired sporadically in bursts of spikes. Masse, Freedman and colleagues have developed artificial neural network models that can store information in working memory by briefly changing the strength of synaptic connections without changing the firing rates of the neurons. Even though the details of the neural mechanisms responsible for working memory continue to be worked out, the pioneering studies of Goldman-Rakic and others provided one of the earliest and most successful links between a cognitive process and a cellular one. This opened a door through which it seemed possible to explain a whole host of cognitive processes in terms of brain activity resolved at the level of individual neurons.

Goldman-Rakic focused on the distinction between internally generated and externally triggered actions. As noted above, her view was that the computational advantage of working memory was that it generated and sustained internal representations, patterns of persistent neural activity that could provide guidance to control behavior in environments even when external stimuli were removed. Goldman-Rakic argued, elegantly, that a wide variety of functional deficits found in humans with prefrontal damage could be explained by the loss of this ability. For example, humans with prefrontal damage are distractible, perseverate, and have difficulty planning. Loss of internal representations to guide behavior would leave one at the mercy of novel stimuli (easily distractible), and cause one to repeat previously rewarded actions after they stopped being rewarding (to perseverate), and prevent one from organizing current actions to achieve future goals.

The schema above derived largely from studies in nonhuman primates. What about the neural basis of working memory in the human brain? Neuroimaging studies have shown that human prefrontal cortex is activated persistently during the delay period of working memory tasks. This provides a strong parallel between human and animal studies. What is more controversial is precisly what is encoded by this persistent activity. Two different schema have been considered (and supported by experimental evidence). In the first formulation, persistent activity in human prefrontal cortex codes the contents of working memory. Indeed, different types of sensory information stored in working memory (such as object identity and location) activate spatially segregated parts of prefrontal cortex in humans.

In the second formulation, referred to as the sensory recruitment hypothesis, persistent activity in prefrontal cortex does not encode information stored in working memory. Instead, enhanced activity in the prefrontal cortex serves as a top-down bias or control signal that recruits persistent activity in posterior sensory cortices, where the information in working memory is actually stored. In this case, sensory information is stored in the same sensory cortical areas where it is initially generated.

In a study supporting this view, Riggall and Postle (2012) decoded stimulus information from patterns of activity measured with functional magnetic resonance imaging (fMRI) in subjects performing a working memory task. The task required that subjects remember either the direction or speed of a circular field of moving dots. Prefrontal cortex was persistently activated during the working memory period, but the pattern of activity did not carry information about the direction of the remembered stimulus. Activity in posterior visual areas, in contrast, carried robust information about the direction of visual motion. This suggested that visual cortex, rather than prefrontal cortex, encoded the information stored in working memory.

This result could, however, reflect differences in the internal structure of prefrontal versus visual cortical areas. To be detectable by fMRI, neurons with different sensory tuning properties (to the direction of motion for example) need to be spatially segregated enough in the brain to recruit distinct patterns of blood flow depending on the stimulus shown. Thus, it is possible for prefrontal neurons to encode sensory features, and for this signal to be obscured in fMRI data, as long as neurons with different tuning are intermingled. Indeed, subsequent studies employing different decoding models have successfully decoded visual feature information stored in working memory from prefrontal fMRI data.

Thus, the role of human prefrontal cortex in working memory remains an area of active research. There is evidence to suggest it operates as a top-down controller of working memory in posterior cortices. There is also evidence to suggest that prefrontal cortex itself is the site of working memory storage where specific items of sensory information are stored.

As is often the case in biological debate, it may prove the two alternative views, each well supported by apparently contradictory data, turn out to not be mutually exclusive. Prefrontal cortex may both store sensory information in the persistent activation of prefrontal neurons, and also induce persistent activity in posterior sensory association cortices to the same task. In this way prefrontal cortex may recruit single neurons in many other cortical areas to join a unitary neural representation that is distributed across many cortical areas, with single neurons in all of them encoding the same information at (nearly) the same time. In fact, at the cellular level, this seems to be a fundamental principle of how distributed cortical networks work. Reciprocal connections rapidly spread and mix neural signals through networks.

Consistent with such an exchange of information at a large scale, it has often been observed that neurons activated to encode a single item of information in working memory exist simultaneously in prefrontal cortex and other cortical areas that are reciprocally connected to it. From this view, the encoding of any single item of information into a pattern of neural activity is a network level phenomenon. Localizing function in such networks becomes a question not of finding localized neural signals, representations, or computations, but rather of understanding how processes taking place in local circuits within each cortical area affect neural activity that is distributed throughout the network, existing everywhere all at once.

Abstraction

Miller and colleagues undertook a pioneering series of experiments in monkeys that broadened the scope of inquiry into the types of information that prefrontal neurons could represent. They trained monkeys to match visual displays based on abstract properties of the displays, rather than the specific stimuli they contained. In one such study (Freedman et al., 2001), monkeys were trained to categorize visual stimuli (Figure 2B,C). A category is an abstraction, in the sense that it generalizes across a potentially infinite set of specific exemplars, capturing the property that they have in common (which defines category membership). In this study, monkeys categorized visual stimuli as being either a cat or dog: the stimuli were volumetric three-dimensional renderings of the animals and the trick was that each stimulus was a mathematical combination (morph) of different cat and dog prototypes in different proportions. For example, a given stimulus might be 20% cat A and 80% dog B. A population of dorsal prefrontal neurons was identified in which firing rate encoded the abstract category to which each stimulus belonged, and not the features of the stimuli themselves. The key attribute of these neurons was that as the features of the stimulus varied continuously from cat to dog, category neurons exhibited a step-like change in firing rate when the morph crossed an implicit category boundary, rapidly shifting the cognitive representation of the stimulus from one to the other in a way that was not closely aligned to the subtle change in its physical attributes.

In other experiments (Nieder et al., 2002), it was shown that prefrontal neurons encode the number of items in a visual display independently of the physical attributes of the stimuli such as their size or location. Such neurons were tuned to numerosity, firing equivalently for any visual display containing, for example, 3 items rather than 2 or 4. It was further demonstrated that prefrontal neurons encode abstract rules independently of the stimuli to which they are applied (Wallis et al. 2001).

Collectively, these studies suggested that abstraction, and potentially thought, in the human brain, may consist of internal representations that can be recognized in the activity patterns of single cortical neurons. That would effectively equate a thought to a cellular event (although admittedly involving the coordinated activation of many millions of individual cells across many communicating areas). This is supported by recent single neuron recording studies in humans demonstrating that abstract mental events, such as reasoning about the beliefs of others when answering questions relating to a narrative passage of text (or not engaging in such a reasoning process), can be decoded from the activity patterns of individual neurons in the human prefrontal cortex with better than 80% accuracy (Jamali et al., 2021). How such signals got there — what learning principles and synaptic mechanisms drive abstraction in hierarchical networks — is a question of deep importance to our understanding of intelligence and ourselves. That question represents the frontier both in human neurobiology and artificial intelligence.

As with the working memory studies, research on abstraction has largely focused on the dorsal prefrontal cortex. There is, however, considerable evidence from human fMRI studies that the premotor cortex, dorsal prefrontal cortex, ventrolateral prefrontal cortex, and frontal pole together are organized along a rostral–caudal gradient in which increasingly anterior regions encode increasingly abstract information. For example, in one seminal experiment (Koechlin et al., 2003), subjects performed a task in which they made a motor response contingent upon increasingly abstract rules. At the lowest level, cues instructed single movements. At the next level, cues instructed whether to ignore the stimulus or perform a single task upon it (categorize a letter as a vowel/consonant, or upper/ lower case). At the highest level, cues indicated which of the two categorization tasks to perform. Functional imaging revealed that increasingly complex levels of cognitive control activated increasingly anterior parts of the frontal lobe. With this scheme, the frontal pole should be responsible for the most abstract decisions. Indeed, there is substantial evidence that the frontal pole is anatomically different in humans relative to other primates.

Decision making

If an overarching principle of prefrontal cortex is that it operates to organize our behavior in time, a decision represents a branch point in this process, where the temporal evolution of behavior proceeds down one path or another (Figure 3). Prefrontal neural circuits exhibit rich temporal dynamics in relation to decision making.

Figure 3. The prefrontal cortex helps to facilitate decision making by organizing behavior in time.

A decision-maker may need to consider options that vary along multiple dimensions, particularly when deciding how to spend their time.

Whereas dorsal prefrontal cortex has been implicated in the representation of control variables reflecting the state of the outside world (see above), ventral prefrontal cortex, especially the orbitofrontal and ventromedial prefrontal cortices, has been implicated in the neural representation of value, suggesting a role in the neural representation of motivated behavior, and potentially contributing to the neural representation of affective states (although that is difficult to pin down). Barbas and colleagues have proposed a structural model of prefrontal cortex relevant to the question of how prefrontal cortex combines affective and environmental information to guide behavior. Their framework draws a distinction between eulaminate (granular) and limbic (agranular or dysgranular) prefrontal areas, based on the extent to which the cortex contains a clearly delineated granular layer IV as well as the pattern of connections.

Eulaminate prefrontal areas (so-called because they have the normal number of laminae), including the dorsal prefrontal cortex, appear most directly involved in the neural representation of external states, or the environment. These areas have strong connections to sensory association cortices, but relatively weak connections to the limbic system. Limbic prefrontal areas, in contrast, appear most directly involved in the neural representation of value to support decision making and guide motivated behavior. Limbic prefrontal areas include large portions of the ventral prefrontal cortex. These areas have strong direct connections to the limbic system, including the nucleus accumbens, the amygdala, hypothalamus, and hippocampus. Lesions of limbic prefrontal areas induce deficits in value-based decision making, whereby patients exhibit difficulty in using the value of actions or outcomes to modify their choices. This suggests a deficit in the ability to represent internal states (as the value of an action, based on reinforcement, is dependent on homeostatic variables such as thirst or hunger). Eulaminate and limbic prefrontal areas are directly connected, and how they interact to integrate information about external and internal states remains an important question.

A large body of evidence has indicated that the ventral prefrontal cortex is strongly concerned with subjective value, or how much a given option is worth to the actor. Consider what would need to happen in your brain if we offered you the opportunity to buy a chocolate-chip cookie from us for $1. You might consider whether you like chocolate-chip cookies, and, if so, how much. But you might also wonder whether the probability of actually receiving the chocolate-chip cookie once you pay us may be quite low, given that we have never met. You may also be concerned about the health qualities of the cookie. You might also want to use that dollar for something else. All of these features (and more) could be combined into a representation of subjective value that is then compared to the value of the $1, and a choice can be made. A great deal of neuroimaging and neurophysiology work has demonstrated that activity in parts of the ventromedial prefrontal and orbitofrontal cortices (which precise parts is the subject of some debate) tracks with subjective value in monkeys and humans, regardless of what type of reward is offered. This suggests that prefrontal cortex combines information about external states and internal states to control behavior.

Summary

Putting all the above together, what picture of prefrontal cortex emerges (Figure 3)? In many ways, the early formulations of prefrontal function appear to have been essentially correct. The function of prefrontal cortex is about organizing behavior in time. It generates neural representations of past events and future outcomes for that purpose. It directs actions in the current moment, based on internal representations, particularly in cases when external stimuli are uninformative, or reflexive actions to them would be unrewarding. It can generate representations of abstract, generalized information derived from experience when behavior is more aptly governed by rules or principles than by specific sensory inputs. It integrates information about the outside world with information about emotions and drives, to impart affective valence to neural representations, and to support selecting the most biologically valuable action. It performs all of these functions by receiving, transforming, and transmitting neural signals encoding behaviorally meaningful information throughout a complex network of cortical and subcortical processing partners. Perhaps more than any other area, prefrontal cortex is responsible for autonomous and internally controlled behavior.