Prefrontal cortex interneurons and their contributions to attention, working memory, and adaptive behavior

Abstract

Inhibitory interneurons play central roles in regulating the input and output of cortical circuits, which in prefrontal cortices (PFC) subserve attention control, working memory and adaptive behavior. Understanding how interneurons support these higher order cognitive functions is a key question in a growing number of studies. Here, we delineate recent progress by surveying molecular, functional and computational motifs of interneurons in the prefrontal cortex of nonhuman primates and rodents. Among multiple transcriptomic and molecular subtypes of neurons several electrophysiologically identified 'eType' interneurons are recruited during attention, learning, and working memory tasks. In nonhuman primate PFC, eType neurons with an inhibitory effect on the local circuit encode behaviorally relevant cues, unexpected outcomes, and tune working memory representations. These response profiles are consistent with the functional specializations proposed for PV+ , SST+ and VIP+ interneurons in rodents, which are recruited during attention and memory-guided tasks. We survey how these functional studies of interneuron types are supported by newly developed molecular and analytical tools and guided by computational studies that suggest unique circuit motifs for distinct types of interneurons to flexibly route synaptic inputs, compute prediction errors, and facilitate information retention in working memory.

1. Introduction

Inhibitory interneurons constitute only a small ~15–34 % fraction of prefrontal cortex (PFC) neurons (Gabbott and Bacon, 1996a; Gabbott et al., 1997; Dombrowski et al., 2001; Rudy et al., 2011; Le Merre et al., 2021) but have particularly important circuit functions. Prefrontal interneurons are implicated in ‘control of information flow’ (Fishell and Kepecs, 2020), determine the tuning sharpness of PFC representations, impose selective filtering of synaptic inputs, and regulate the output-gain of a circuit (Akam and Kullmann, 2010; Wilson et al., 2012; Kepecs and Fishell, 2014; Womelsdorf et al., 2014a; Roux and Buzsáki, 2015; Cardin, 2018). These circuit functions are important for performing multiple cognitive tasks that are causally linked to the PFC, including rule-based attention control (Petrides, 2005; Matsumoto et al., 2007; Womelsdorf and Everling, 2015), working memory maintenance (Constantinidis and Klingberg, 2016), and cognitive flexibility (Passingham and Wise, 2012; Cho et al., 2015; Heilbronner and Hayden, 2016; Banaie Boroujeni et al., 2021; Passingham, 2021a; Preuss and Wise, 2022).

Elucidating the unique contributions of interneuron subtypes to PFC functions faces the challenge of simultaneously assessing higher-order cognitive processes and interneuron profiles. Studies in rodents overcome this challenge by targeting PFC interneurons using their expression of molecular markers such as parvalbumin (PV), somatostatin (SST), and vasoactive intestinal protein (VIP), amongst others (Kepecs and Fishell, 2014; DeFelipe et al., 2013; Tremblay et al., 2016). Despite recent progress, identifying discrete molecular signatures of interneurons has remained challenging in non-human primates (NHP) (El-Shamayleh and Horwitz, 2019; Tremblay et al., 2020). Overcoming these challenges is crucial, since relying on studies of rodent PFC to infer interneuron-specific functions in NHP PFC faces limitations stemming from divergent evolution of the prefrontal cortices in rodent and primate lineages (Preuss and Wise, 2022; Murray et al., 2017). Rodent PFC encompass infralimbic and prelimbic cortical areas (Le Merre et al., 2021), with unique circuitry and laminar organization compared to the prefrontal cortical subfields in humans and nonhuman primates (Passingham and Wise, 2012; Preuss and Wise, 2022; Berger et al., 1991; Seamans et al., 2008; Carlén, 2017; Laubach et al., 2018; Schaeffer et al., 2020; Passingham, 2021b; Cools and Arnsten, 2022). In NHP PFC, at least six separable functional subfields are distinguishable that contribute to higher order cognitive processes such as adaptive behavior, which are not easily compared to infra- and peri-limbic cortices in rodents without simplifications (Passingham, 2021a; Murray et al., 2017; Beran et al., 2016; Primates et al., 2019; Stephan et al., 2019; Boroujeni et al., 2022a; Kolk and Rakic, 2022; Schwartz and Beran, 2022; Watson et al., 2023).

So far, functional insights about primate PFC interneurons have largely been derived from fitting neural activity to simulated interneuron activity in models (e.g., Kim and Sejnowski, 2021), or distinguishing interneurons by their suppressive influence on the local circuit (e.g., Banaie Boroujeni et al., 2021). Inferences about the functional roles of interneurons relied on clustering them based on their electrophysiological profiles, specifically their action potential waveforms and firing patterns (eTyping). This approach has recently gained traction in the field, suggesting an outsized importance of PFC interneurons in attention control, working memory, and flexible learning (Constantinidis and Goldman-Rakic, 2002; Kaufman et al., 2010; Womelsdorf et al., 2014b; Ardid et al., 2015; Trainito et al., 2019; Boroujeni et al., 2022b). These three cognitive domains represent core PFC-dependent functions that have been causally established through lesion, inactivation, and stimulation studies across species (Boroujeni et al., 2022b; Kamigaki and Dan, 2017; Rossi et al., 2007; Cho et al., 2020; Treuting et al., 2025). We focus on these domains as they are selectively impaired in distinct interneuron-related pathologies and allow cross-species comparison (Dienel and Lewis, 2019; Paine et al., 2011). How these eTypes in primate PFC relate to known morphoelectric and transcriptomic types in rodents has remained difficult to discern. Here, we address this question by reviewing the molecular and functional composition of interneurons in the PFC of primates and rodents, how they contribute to performing cognitive tasks, and which computations they support in the neuronal circuit.

2. Interneuron composition within the prefrontal cortex of nonhuman primates

Interneurons in NHP PFC share their transcriptomic origins with those of other species including rodents, but histological studies have identified that the proportion of interneurons in PFC and some interneuron subtypes appear to be unique to the primate lineage (Krienen et al., 2020). Interneurons comprise a larger proportion of the total neuronal population in NHPs (~15–34 %, depending on cortical area) and humans (~30 %) compared to rodents (~15–20 %, Fig. 1A) (Gabbott et al., 1997; Dombrowski et al., 2001; Le Merre et al., 2021; Hendry et al., 1987; Gabbott and Bacon, 1996b; DeFelipe, 2002; Džaja et al., 2014; Saunders et al., 2018; Medalla et al., 2023). The interneuron population is proportionally larger in associative areas compared to sensory and motor areas, particularly in PFC, suggesting that interneurons play a more significant functional role in these circuits (Krienen et al., 2020; Džaja et al., 2014; Medalla et al., 2023; Kooijmans et al., 2020; Torres-Gomez et al., 2020).

Fig. 1.

Diversity of molecularly identified interneurons in prefrontal cortex in primates. (A, B) Proportion of interneurons (A) and their differential developmental origin (B) in primate and mouse prefrontal (mPFC) and visual areas based on transcriptomic profiling. (C) UMAP visualization of all identified GABAergic neuron classes in dorsolateral PFC, across marmosets, macaque monkeys, chimpanzees, and humans. Most neurons express PV, SST, VIP, or LAMP5, but each major transcriptomic type is clearly divisible into multiple subtypes. (D) Cell counts of immunolabeled PV, CB, and CR positive neurons through the depth of mouse FC and macaque lateral PFC reveals almost double the number of CR neurons in macaque lateral PFC compared to mouse mPFC, but similar percentages of CB and PV neurons between species. Additionally, ~10 % of labeled neurons in mouse mPFC co-expressed calcium binding proteins, but co-expressing neurons constituted less than 2 % of the labeled population in macaque lateral PFC. (E) Exemplar connectivity of each interneuron class in nonhuman primate PFC. Darkly colored lines indicate axons and synaptic contacts while lightly colored lines reflect dendritic structures. PV chandelier cells form dense cartridges of vertical synapses along the proximal axon segment of excitatory pyramidal neurons. PV basket cells primarily synapse on proximal dendritic segments while CB cells preferentially target distal dendritic segments of pyramidal neurons. In contrast, CR interneurons almost exclusively target other interneurons and avoid contacts with excitatory pyramids. A small percentage of interneurons, primarily LAMP5-expressing neurogliaform cells, maintain diffuse connections with pyramidal neurons through volumetric GABAergic neurotransmission. Last, but not least, some interneurons send long-range projections to other cortical and subcortical brain regions instead of forming local synaptic contacts with excitatory pyramids. (F) Characteristic transcriptomic, morphological, and electrophysiological features of each interneuron subclass illustrated in E; summarized from (Gabbott and Bacon, 1996a; Dombrowski et al., 2001; DeFelipe et al., 2013; Krienen et al., 2020; Hendry et al., 1987; Gabbott and Bacon, 1996b; DeFelipe, 2002; Džaja et al., 2014; Medalla et al., 2023; Ma et al., 2013, 2022; Micali et al., 2023; Hladnik et al., 2014; Condé et al., 1994; Ascoli et al., 2008; Lewis and Lund, 1990; Somogyi et al., 1982; DeFelipe, 1997). PV: parvalbumin, CB: calbindin, CR: calretinin, SST: somatostatin, CCK: cholecystokinin (CCK), NPY: Neuropeptide Y, 5-HT3R: 5-hydroxytryptamine receptor type 3, VIP: vasoactive intestinal polypeptide, LAMP5: lysosomal associated membrane protein family 5, MGE: medial ganglionic eminence, CGE: caudal ganglionic eminence, Pyr: excitatory pyramidal neuron.

(b) Adapted from (Krienen et al., 2020). (c) Adapted with permission from Ma et al., (2022).

This increased abundance of interneurons in primate PFC likely results from several developmental adaptations, including extended periods of neurogenesis, enhanced survival mechanisms, and expanded migratory pathways that continue for months after birth in the PFC (Arshad et al., 2016; Paredes et al., 2016; Fin et al., 2025; Erickson and Lewis, 2002; Feng et al., 2025). It is also worth noting that postnatal programmed cell death can form another critical developmental process affecting interneuron populations which may serve as another target for evolutionary traits of cortical interneurons (García and Fishell, 2024; Southwell et al., 2012; Priya et al., 2018; Wong et al., 2018).

The preferential localization of interneuron types to PFC in primates originates during developmental migration (Krienen et al., 2020; Ma et al., 2013, 2022; Jorstad et al., 2023; Micali et al., 2023). In rodents, neural progenitor cells that develop into PFC interneurons stem from the medial (MGE) and caudal (CGE) ganglionic eminences in roughly equal proportions between sensory and associative areas, but in primates, compared with sensory areas (V1) a significantly larger proportion of PFC interneurons stem from the CGE (Krienen et al., 2020) (Fig. 1B).

Each ganglionic eminence gives rise to two distinct transcriptomic interneuron classes; CGE-derived interneurons preferentially populate superficial cortical layers and predominantly express VIP or lysosomal associated membrane protein family member 5 (LAMP5) while MGE-derived interneurons preferentially populate deep cortical layers and express PV or SST (Fishell and Kepecs, 2020; Krienen et al., 2020; Jorstad et al., 2023; Butt et al., 2005; Wonders and Anderson, 2006; Harwell et al., 2015; Lim et al., 2018). These developmental studies suggest that NHPs share similar types of interneurons with rodents but primate PFC contains a higher proportion of VIP neurons compared with rodent PFC. This increased abundance of VIP interneurons could also be further enhanced by their activity independent survival from programmed developmental cell death (Priya et al., 2018). VIP is one of four cardinal transcriptomic markers, together with PV, SST, and LAMP5, that identify almost non-overlapping classes of interneurons and encompass nearly all interneurons in PFC (Krienen et al., 2020). These interneurons can be further divided into distinct subtypes based on transcriptomic profiles highlighting features related to their synaptic plasticity and connectivity preferences. For example, among the most prominent interneuron subtypes in NHPs are subtypes of LAMP5 interneurons known as Ivy cells that express LIM homeobox 6 (LHX6) (Krienen et al., 2020; Hodge et al., 2019) and subtypes of PV positive chandelier cells with unique NMDA receptor expression (Ma et al., 2022) (Fig. 1C, see Box 1). These interneuron types may support functions in NHP PFC that derive from unique transcriptomes, such as novel modes of NMDA-dependent plasticity, which future studies will need to clarify (Ma et al., 2022).

Box 1. Transcriptomic subclasses of prefrontal cortex interneurons.

Transcriptomic studies in nonhuman primates and humans have identified distinct subclasses of the major interneuron types: PV, CB, CR and LAMP5. Each of these interneuron types expresses different genes in PFC compared to sensory cortex, which exert different roles in local circuits of each region, e.g., contributing weaker or stronger to plasticity in the circuit (Rao et al., 1999; Reinert et al., 2021). For example, PV-positive chandelier cells can be divided into two transcriptomic subclasses based on differential expression of NMDA receptors and non-clustered protocadherins (DeFelipe, 1997), which suggests distinct ionotropic synaptic transmission and axoaxonic neurotransmission between these groups. Similarly, a subclass of interneurons that express LIM homeobox 6 (LHX6) in addition to LAMP5 is significantly enriched in primates and humans compared to rodents, derives from the MGE instead of the CGE, and preferentially migrates to deep cortical layers (Cho et al., 2020; Ding and Gold, 2012). These interneurons, known as Ivy cells, exhibit a unique morphology and patterns of synaptic activity related to their connections with basal and oblique dendrites of pyramidal cells (Reynolds and Desimone, 2003; Richardson et al., 2022; Rich and Shapiro, 2007). Recent work demonstrates that LAMP5/LHX6+ interneurons can be further split into two transcriptomic types based on positive or negative expression of prospero-related homeobox 1 (PROX1). PROX1 regulates neocortical cell fate and maturation (Rich and Wallis, 2017) and is responsible for the differentiation of CGE-derived neural precursor cells into mature VIP+ and CB+ interneurons, and into a subtype of LAMP5+ interneurons that co-express reelin (RELN) (Rodriguez-Romaguera et al., 2022). However, how PROX1 affects circuit functions are unclear. These examples illustrate that transcriptionally defined interneuron classes in primate PFC are often unique compared to those of rodent PFC and likely support specialized patterns of connectivity and synaptic activity. Yet more research is needed to causally link each interneuron class to its function within local circuits in primate PFC.

While our review specifically focuses on these well-characterized interneuron classes from the rodent literature to better understand their organization and functions in NHP PFC, it is worth noting that recent studies have identified notable evolutionary innovations of the interneuron repertoire across species, with several subtypes appearing to be specific to or enriched in primates. For example, rosehip neurons which are a subtype of LAMP5 + interneurons have been identified predominantly in superficial layers of the human cortex and are notably enriched in associative regions including PFC (Hodge et al., 2019; Boldog et al., 2018) while they appear to be rare or absent in rodents (Földy et al., 2016). Another example is TAC3 + interneurons whose progenitors emerge in early development and form a primate-specific interneuron subtype in the striatum (Krienen et al., 2020, 2023).

3. Distinguishing PFC interneuron types in non-human primates

While developmental and transcriptomic studies specify multiple molecularly distinct interneuron types in NHP, how each type affects circuit functions in PFC remains elusive. Functional classes of NHP interneurons have been identified by distinguishing electrophysiological activation profiles of neurons, known as eTypes (Markram et al., 2015). While the molecular and morphological signatures of NHP eTypes are often unknown, their putative classification may be inferred from a combination of intrinsic features, as demonstrated for rodent sensory cortex (Gouwens et al., 2020). By combining electrophysiological (eType), morphological (mType), and transcriptomic (tType) features, interneurons fall into distinct classes (Gouwens et al., 2020; Cauli et al., 1997; Gouwens et al., 2019). These interneuron classes are known as MET types (Gouwens et al., 2020) and can be identified in most mouse brain regions (Yao et al., 2023). Homologous e, m, and t features have been independently described for primate interneurons and here, we review features that likely identify MET types in primate PFC.

3.1. Molecular and transcriptomic interneuron diversity

The four major transcriptional tTypes in primate PFC expressing PV, SST, VIP or LAMP5 are largely conserved across Old Word and New World primates, apes, and humans (Krienen et al., 2020; Ma et al., 2022; Jorstad et al., 2023). SST+ neurons often co-express calbindin (CB) and VIP neurons often co-express calretinin (CR), and immunolabeling studies demonstrate that non-overlapping protein expressions of CB, CR, and PV distinguishes > 90% of all primate interneurons (Gabbott and Bacon, 1996a; Dombrowski et al., 2001; Hendry et al., 1987; Gabbott and Bacon, 1996b). VIP/CR interneurons constitute the largest group of interneurons in NHP PFC, unlike mouse mPFC (Dombrowski et al., 2001; Krienen et al., 2020; Džzaja et al., 2014; Medalla et al., 2023; Hladnik et al., 2014)(Fig. 1D). On average, within 50 × 50 μm of a primate prefrontal cortical column, there are approximately 6–9 CR neurons, 3–5 PV neurons, and 3–4 CB neurons (Gabbott and Bacon, 1996a; Condé et al., 1994) (for subarea specific differences see Dombrowski et al., 2001). VIP/CR interneurons primarily target PV and SST neurons and induce circuit-wide disinhibition when activated (Kepecs and Fishell, 2014; Pfeffer et al., 2013). Larger populations of VIP/CR neurons in NHP PFC compared to rodent PFC may thus contribute to functionally specialized disinhibitory circuit motifs (see below). It is worth noting that while rodent studies commonly use PV, SST, and VIP to define molecularly distinct interneuron types, NHP studies, specifically earlier neuroanatomical studies, traditionally have been using calbindin (CB) and calretinin (CR) as molecular markers, which show correspondence to rodent SST and VIP populations, respectively (Fig. 1E,F).

3.2. Morphological mTypes of interneurons

Interneuron mTypes are distinguished by their somata (shape and size), axon (length, arborization, and orientation), dendrite (branching, arborization, and thickness), and the relative spine distribution along the dendritic tree (DeFelipe et al., 2013; Medalla et al., 2023; Yuste et al., 2020; Ascoli et al., 2008; Jiang et al., 2015)(Fig. 1E). Typically, PV neurons are classified as multipolar basket cells or chandelier cells; CB neurons are classified as Martinotti, bipolar, or neurogliaform cells; and CR neurons are classified as double bouquet, bipolar, or Cajal-Retzius cells (DeFelipe et al., 2013; Condé et al., 1994; Lewis and Lund, 1990; Kawaguchi, 1997; Markram et al., 2004; Zaitsev et al., 2005, 2009; Kubota, 2014; Kim et al., 2017)(Fig. 1F). The laminar distribution and proportions of mTypes and their mapping to distinct tTypes have been extensively characterized in rodent brains (DeFelipe et al., 2013; Tremblay et al., 2016; Markram et al., 2015; Gouwens et al., 2019; Yuste et al., 2020; Ascoli et al., 2008; Jiang et al., 2015; Markram et al., 2004), but these relationships are less investigated and are more complex in primate brains. NHP mTypes exhibit greater morphological heterogeneity for tTypes in PFC. For example, there is greater variability observed within PV chandelier and basket cells in NHP PFC compared to NHP visual cortex and rat PFC (Gabbott et al., 1997; Gabbott and Bacon, 1996b; Lewis and Lund, 1990). CR interneurons in superficial cortical layers also exhibit increased morphological variability in NHP PFC compared to rodent PFC, with a higher prevalence of vertically oriented double-bouquet cells compared to bipolar cells (Gabbott et al., 1997; Gabbott and Bacon, 1996b; DĎzaja et al., 2014; Medalla et al., 2023; Zaitsev et al., 2005). This increased proportion of double-bouquet (also called “horse-tail”) CR interneurons, characterized by extensive vertical connections across cortical layers and preferential synapsing on other inhibitory interneurons (Gabbott et al., 1997; DeFelipe et al., 2013; Gabbott and Bacon, 1996b; Pfeffer et al., 2013), could contribute to higher-order cognitive functions that require the coordination of neural activity and integration of information across different cortical layers.

3.3. Electrophysiological eType characterization of interneurons

Querying the electrophysiological activation profiles of interneurons enables their classification into distinct eTypes based on action potential (AP) dynamics, firing thresholds, firing rates and variability, and the capacity for neuronal bursting (Box 2) (Schultz, 2006; Bean, 2007; Kole et al., 2008; Kim and Lim, 2015). Differences in eType features can successfully differentiate interneurons with comparable mType or tType features (Povysheva et al., 2007, 2008). While currently established MET types are largely based on in-vitro data and exclusively for rodents (Markram et al., 2015; Gouwens et al., 2019; Ascoli et al., 2008), a growing number of studies describe eTypes in NHP and propose linkages to m and tTypes. One well studied link of molecular and electrophysiological signatures is evident for neurons with a narrow action potential (AP) waveform shape (Torres-Gomez et al., 2020; McCormick et al., 1985; Wilson et al., 1994; Martina et al., 1998; Wei et al., 2023), which we denote as N-eType interneuron. The AP shape is based on the interplay of ion channels (Segev and Korngreen, 2007; Hay et al., 2011; Li et al., 2014; Palacio et al., 2017; Sabater et al., 2021; Richardson et al., 2022). Among those ion channels, a subtype of K+ channel (KV3 family) causes a brief AP duration and underlies fast spiking phenotypes (Martina et al., 1998; Toledo-Rodriguez et al., 2004; McDonald and Mascagni, 2006; Hu et al., 2014; Kaczmarek and Zhang, 2017; Rowan and Christie, 2017; Gu et al., 2018). This fast-spiking phenotype has a non-accommodating, sustained tonic repetitive firing that predominates in PV interneurons in rodents (Cauli et al., 1997; Markram et al., 2004; McDonald and Mascagni, 2006; Hu et al., 2014; Kaczmarek and Zhang, 2017; Rudy and McBain, 2001; Lien and Jonas, 2003). In rodents, ~90% of fast spiking neurons express PV (Kawaguchi, 1997; Hu et al., 2014; Kawaguchi, 1995; Galarreta and Hestrin, 1999) and ~35% of fast spiking neurons express CB and neuropeptide Y (NPY) (Cauli et al., 1997; Kawaguchi, 1995). A similar conclusion may hold for NHP PFC where > 95% of all PV and ~87% of all SST interneurons are of the N-eType, firing AP’s with narrow spike width (Torres-Gomez et al., 2020; Zaitsev et al., 2005; Ghaderi et al., 2018). Yet narrow spikes also occur in non-fast spiking neuron types in NHP, including in ~20% of VIP interneurons (Torres-Gomez et al., 2020) among other GABAergic neurons (Zaitsev et al., 2005; Krimer et al., 2005), and in a subgroup of pyramidal cells in NHP motor cortex (Soares et al., 2017). It is, however, worth noting that these eType characteristics are largely derived from in-vitro studies and intracellular recordings, often conducted in isolation, which makes a direct connection to in-vivo studies less straightforward and requires further control over stimulation protocols in future research (Koch et al., 2025).

Box 2. Distinguishing eTypes.

Conventionally, eTypes have been distinguished by the temporal dynamics of extracellular spike waveforms into an N-eType (narrow) and a B-eType (broad) type (García and Fishell, 2024; Ghaderi et al., 2018). Finer differences of the waveform shape allows more variability to be captured and more subgroups to be identified than just narrow and broad (Challis et al., 2022; Chamberlin et al., 2023). Besides spike waveforms, the temporal structure of spike trains allows clustering eTypes into further subtypes (Banaie Boroujeni et al., 2020; Challis et al., 2022; Hertäg and Sprekeler, 2020; Rossi et al., 2007). Different clustering techniques have been used to distinguish eTypes. This box describes parameters of the waveform shape and spike train temporal dynamics that are used in the field to distinguish eTypes.

(i). eTypes actional potential waveform dynamics

Intrinsic properties of neurons determine key aspects such as action potential (AP) dynamics, firing thresholds, firing rates, and the capacity for neuronal bursting (Fuentealba et al., 2008; Gabbott and Bacon, 1996a, 1996b; Gabbott et al., 1997). These properties are linked to various voltage-gated ion channels. While APs share basic similarities across different neurons, they are kinetically governed by a combination of sodium (Na+), potassium (K+), Calcium (Ca2+), and leakage channels (Goertsen et al., 2022; González-Burgos et al., 2005, 2004; Gouwens et al., 2020, 2019; Gradinaru et al., 2010). A typical AP begins with a Na+-mediated depolarization, followed by a repolarization phase driven by K+ channel activation and Na+ channel inactivation, and terminated through inactivation of K+ channels and often a calcium (Ca2+)-mediated hyperpolarization phase (Gabbott and Bacon, 1996a). Among the ion channels, a subtype of K+ channel (KV3 family) which are highly expressed on PV+ basket cells results in brief AP durations and is associated with the fast spiking phenotypes (Goaillard and Marder, 2021; Gray and Singer, 1989; Gu et al., 2018; Hadaczek et al., 2009; Hahn et al., 2022; Harris et al., 2019). To distinguish eTypes, the repolarization and hyperpolarization phases of APs can indicate the relative contribution of these family subtypes of ion channels and consequently distinguishing of N-eTypes. For in-vivo recordings, spikes show similar phases but with an opposite polarization. In extracellular recordings, spikes show similar phases to those of intracellular APs, but with opposite polarizations (Gabbott and Bacon, 1996a). Consequently, the pre- and post-hyperpolarization periods provide more information about K+ channel subtype contributions and have been used more frequently to distinguish N-eTypes as listed below.

1. Trough-to-peak duration: is the duration between the depolarization trough and the hyperpolarization peak, which relies on Na+ channels closing and K+ channels opening/closing (Cauli et al., 1997; Ceccarelli et al., 2023; Challis et al., 2022; Chamberlin et al., 2023; Ghaderi et al., 2018; Homayoun and Moghaddam, 2007; Lien and Jonas, 2003).

2. Hyperpolarization-ratio: an index based on K+ channel closing that describes pre-hyperpolarization peak kinetics (Banaie Boroujeni et al., 2020; Hladnik et al., 2014).

3. After-hyperpolarization duration: time from the hyperpolarization peak to resting state, which relies on K+ channels closing and Ca+2 channels opening (Banaie Boroujeni et al., 2020; Challis et al., 2022; Chamberlin et al., 2023; Klausberger et al., 2003; Lien and Jonas, 2003).

(ii). eTypes spike train dynamics

Neuron types differ in their composition of ion channels that set the membrane conductance and neuronal excitability which are key factors contributing to the diversity of electrophysiological activity phenotypes (Rotaru et al., 2015; Roussy et al., 2021). These eTypes have distinct firing patterns in-vitro and in isolated cells, including low-firing continuous spiking, stuttering, bursting, low threshold spiking, fast-spiking, regular spiking, irregular spiking (IrS, with or without bursts), adapting, and intrinsic burst generation (Amita et al., 2020; Federer et al., 2024; Fetsch et al., 2018; Kole et al., 2008; Rossi et al., 2007; Roux and Buzsáki, 2015). Capturing these dynamics, especially in-vivo and within the network is not trivial, and there is lack of support whether these eTypes can be observed in the brain. However, several parameters have been proposed that can capture the temporal dynamics of spike trains and consequently distinguishing different eTypes by examining the regularity of the spike train (Banaie Boroujeni et al., 2020; Challis et al., 2022; Rowan and Christie, 2017; Rudy et al., 2011). These parameters can reliably group neuron types into regular spiking, regular fast spiking firing, irregular random firing, and irregular burst firing (Banaie Boroujeni et al., 2020; Challis et al., 2022; Rudy et al., 2011; Rudy and McBain, 2001).

Coefficient of Variance (CV):

This measure reflects the ratio of the standard deviation of the inter-spike intervals (ISIs) to their mean. Although this measure can provide information about the global regularity of the spike trains, it fails to capture the local regularity of the spike trains and is biased by outlier ISIs.

Local Variation (LV):

local variation of spike train is proportional to the difference between consecutive ISIs divided by their mean. When compared to CV, LV can differentiate different types of spike irregularity such as irregular burst, and Poisson process non-burst irregularities (Rudy et al., 2011). In general, a combination of LV and CV can be used to better distinguish eTypes. While a large LV indicates a high degree of local irregularity, for example a burst pattern in irregular spiking neurons, a large CV indicates a high degree of global irregularity such as pauses of firing. Several studies found that the LV explains a greater percentage of variances in extracellular data compare to other metrics on spike train, such as CV, Fano Factor, and firing rate (Banaie Boroujeni et al., 2020; Challis et al., 2022; Hertäg and Sprekeler, 2020).

(iii). eTypes Clustering

Most studies have grouped eTypes into N-eTypes and B-eTypes based on a bimodal Gaussian distribution fit to a single parameter describing spike waveform dynamics (Ceccarelli et al., 2023; Celka, 2007; Homayoun and Moghaddam, 2007; Lien and Jonas, 2003). However, more recent studies have advanced this bimodal clustering by using graph-based clustering or Gaussian Mixture Models (GMMs) to categorize waveforms into multiple clusters (Chamberlin et al., 2023; Sabater et al., 2021). Unlike earlier studies, these studies involved more than one waveform parameter or latent variables that represented the spike waveform's low-dimensional structure. For all clucstering methods, clustering based solely on the waveform is not sufficient to determine inhibitory eTypes and inferring their connections with those molecularly identified neuron types. Based on Lee et al. (2021) findings, 68% of premotor cortex neurons have narrow spike waveforms, while only ~20% of cortical neurons are interneurons (Sabater et al., 2021). N-eType pyramidal neurons are well described as having narrow spikes in ventral and dorsal premotor cortex (Sadakane et al., 2015). The thin spikes of these neurons are linked to the expression of fast potassium (Kv3.1b) channels (Hendry et al., 1987). To address this issue, including variables that capture the dynamics of spike trains can help identify eTypes that more closely resemble the proportion and characteristics of interneurons and pyramidal neurons.

Unlike in rodents, where the correspondence between tTypes and eTypes is well established through optogenetic tagging, with PV neurons showing fast-spiking properties, SST neurons regular-spiking, and VIP neurons irregular-spiking patterns at 80–90% accuracy (Tremblay et al., 2016; Gouwens et al., 2020; Tasic et al., 2018), such correspondence in primates remains challenging due to technical limitations. Consequently, in vivo identification of tType-eType correspondence in NHPs has largely relied on statistically separating N-eTypes neurons based on AP waveform shape and, in some studies, firing statistics, which show that at least three distinct N-eTypes are distinguishable based on firing rate and firing variability (Box 2) (Banaie Boroujeni et al., 2021; Ardid et al., 2015; Trainito et al., 2019; Dasilva et al., 2019; Boroujeni et al., 2020).

3.4. In-vivo characterization of eTypes circuit functions in NHP prefrontal cortex

The differentiation of N-eTypes in primate PFC, along with N-eTypes more likely encompassing inhibitory interneurons promises and justifies further investigations and effort in understanding their unique in-vivo circuit functions and their functional roles during cognitive task performance. The starting point for considering N-eTypes as interneurons is establishing their inhibitory effect on the circuit. Inhibitory circuit effects of N-eTypes in NHP PFC can be indirectly inferred from studies documenting the GABAergic action of interneurons in rodent mPFC (Barthó et al., 2004; Sirota et al., 2008). Alternatively, an early antidromic stimulation study in lateral PFC of NHP distinguished N- and B-eType neurons and found that long-range connections of the PFC to the superior colliculus originated exclusively from B-eType neurons, as would be expected if N-eType neurons are primarily local interneurons (Johnston et al., 2009). Cross-correlation analysis and cell-type specific pharmacological manipulation provide more direct means of documenting inhibitory effects in N-eType neurons (Wilson et al., 1994; Rao et al., 1999; Constantinidis et al., 2002). Within lateral PFC of NHP, in-vitro stimulation of N-eTypes elicits GABA-A mediated inhibitory postsynaptic potentials in connected broad spiking neurons (González-Burgos et al., 2004, 2005). Consistent with in-vitro findings, in-vivo recordings in lateral PFC and ACC confirmed that only about 20% of extracellularly recorded neurons in these areas are N-eTypes (Fig. 2A). In addition, spiking events of N-eType neurons are associated with inhibition of multiunit activity measured on neighboring electrodes (Banaie Boroujeni et al., 2021), and isolated suppression of spiking activity in nearby broad spiking B-eType neurons (Constantinidis and Goldman-Rakic, 2002; Rao et al., 1999; Constantinidis et al., 2002; Diester and Nieder, 2008; Oemisch et al., 2015; Viswanathan and Nieder, 2017a) (Fig. 2B). The inhibitory effects of N-eType neurons on the local circuit peaks within a ~5 ms time lag and can last ~20–50 ms (Banaie Boroujeni et al., 2021; Viswanathan and Nieder, 2017a).

Fig. 2.

Dissociation of N-eType and B-eType neurons in nonhuman primate prefrontal cortex. (A) Visualization of lateral PFC (area 9/46, green) and ACC (area 24, purple) in macaque brain. N-eType neurons (red) constitute about 19 % of the population with a narrow spike waveform. (B) Spiking activity of N-eType neurons (red) measured during task performance triggers a faster drop of multi-unit activity at neighboring recording electrodes than broad spiking neurons (blue). The difference between cell types is portrayed in brown. This pattern is consistent with a net suppressive effect of N-eType activity on MUA, via stronger pre- than post-spike responses for N- versus B-eType neurons in both ACC and lateral PFC. (C) Neurons were recorded in NHP PFC, areas 8/46, during the delay phase of a virtual reality spatial working memory task before and after injection of NMDA antagonist ketamine which impaired performance. A schematic showing at the synapse-level, Ketamine antagonizes NMDA receptor activity which can suppress inhibitory interneurons, reduce their rate and consequently releasing inhibition on excitatory neurons and increasing their rate. Similar reduced activity is shown in N-eType but not B-eType neurons when memorizing the preferred location in the working memory task. (D) Schematics of networks with states dominated by inhibitory (red) and excitatory (blue) cells. Example traces of spikes of N-eType (red) neurons phase locked to the local field potential (LFP) fluctuations and gamma, and B-eType neurons (blue) locked to beta, and theta/alpha oscillations in the PFC of NHPs. Right column: Spike-LFP synchronization of N-eType neurons (red) in the ACC and lateral PFC of NHPs shows peaks in the theta/alpha and beta frequency bands, respectively. N-eTypes in both ACC and lateral PFC show stronger tendencies to synchronize to gamma frequency band (40 Hz) than B-eType neurons. The bottom panel schematically illustrates the different phase-locking preferences of N- and B-eType neurons. (E) In the PFC of NHP, N-eType neurons fire burst spikes during top-down controlled attention states that precede increases of coherent LFP oscillations (illustrated at top in red). In contrast, burst spikes of B-eType neurons on average follow the rise of coherent LFP oscillatory activity (illustrated in blue at bottom). The relative lead and lag of burst spikes to the time of increasing LFP oscillations distinguished the groups of N- versus B-eType neurons. Panels A-C adapted from (Banaie Boroujeni et al., 2021; Womelsdorf et al., 2014b; Roussy et al., 2021).

3.5. Pharmacological modulation of N- versus B-eType neurons

Pharmacological manipulations have documented that N-eType neurons within the PFC show distinct neural response dynamics compared with B-eTypes, which likely derive from unique postsynaptic receptor distributions. N-eType neurons in lateral PFC of NHP are differentially affected by NMDA antagonists during tasks requiring working memory and the encoding of rules (Ma et al., 2015; Roussy et al., 2021). Systemic administration of the NMDA antagonist, ketamine, in primate lateral PFC causes a use-dependent reduction of N-eType neuron activity and concomitant disinhibition of B-eType neurons (Roussy et al., 2021) (Fig. 2C). A similar differential effect is observed for N- and B-eType neurons in rodent PFC after administering selective NMDA antagonists (Homayoun and Moghaddam, 2007; Deng et al., 2020). These findings are consistent with an increased density of NMDA receptors in local cortical interneurons since their blockade reduces the recruitment of inhibition and impairs inhibitory control of nearby pyramidal cells (Ma et al., 2015; Roussy et al., 2021)(Fig. 2C).

N-eTypes neurons in NHP PFC also have distinct electrophysiological profiles following local cholinergic, noradrenergic, and dopaminergic interventions. Cholinergic manipulation using the nonselective agonist carbachol in lateral NHP PFC causes a systematic increase in firing rate in B-eType neurons, but a heterogeneous firing rate modulation with no net change in N-eType neurons (Major et al., 2018). In contrast, selective blockade of the muscarinic M1 sub-receptor primarily suppresses N-eType activity compared to B-eType neuron activity (Vijayraghavan et al., 2018). Thus, N- and B-eType neurons have different sensitivity profiles to nicotinic and muscarinic modulation, consistent with differential neuromodulatory receptor densities in inhibitory and excitatory neurons in NHP (Disney et al., 2006; Tsolias and Medalla, 2022). Interneurons in frontal cortices have also been reported to show stronger expression of adrenergic α2 and β receptors (Wang et al., 2013; Kawaguchi and Shindou, 1998; Lee et al., 2020). Consistent with these expression patterns, a recent study found that α2-A noradrenergic receptor stimulation with Guanfacine selectively improved the encoding of reward prediction errors during reversal learning in N-eType but not in B-eType neurons in the anterior cingulate cortex of NHPs (Hassani et al., 2024). A differential neuronal responses of prefrontal cortex neurons is also evident to local release of dopamine. Iontophoretic dopamine administration systematically suppresses firing rates of N-eType neurons in NHP lateral PFC but increases firing rates of B-eType neurons in the same area (Jacob et al., 2013). In summary, N-eType neurons show distinct response dynamics to local pharmacological modulations affecting GABAergic, glutamatergic, cholinergic, noradrenergic, and dopaminergic systems. These unique response profiles are consistent with considering N-eTypes in primate PFC as being composed mostly of inhibitory interneurons with distinct electrophysiological characteristics.

3.6. Spike timing and synchronization preferences of interneurons in prefrontal cortex

Inhibitory interneurons not only have a suppressive effect on neural activity of their local circuits, but also display unique spike timing and propensities to synchronize to local circuit activity, which suggests that the temporal dynamics of N-eType neurons distinguish them from principal cells (Womelsdorf et al., 2014a; Klausberger and Somogyi, 2008; Cardin et al., 2009; Sohal et al., 2009; Wang, 2010; Buzsáki and Wang, 2012; Hasenstaub et al., 2016). Computational studies also support this notion by showing that inhibitory interneurons play causal roles in resonating with, generating and sustaining rhythmic synchronization in cortical circuits at narrow band frequencies in theta (Stark et al., 2013), alpha (Jensen and Mazaheri, 2010), beta (Spitzer and Haegens, 2017; Sherfey et al., 2020), and gamma band frequencies (Womelsdorf et al., 2014a; Banaie Boroujeni et al., 2021; Wang, 2010; Buzsáki and Wang, 2012; Buia and Tiesinga, 2006). It is therefore not surprising that N-eType neurons in NHP PFC show unique synchronization patterns within the local circuit when compared to B-eType neurons (Banaie Boroujeni et al., 2021; Ardid et al., 2015; Rich and Wallis, 2017; Voloh and Womelsdorf, 2018; Milton et al., 2020; Banaie Boroujeni and Womelsdorf, 2023). N-eType neurons in primate PFC and ACC synchronize significantly more strongly at low gamma band frequencies (35–45 Hz), while B-eType neurons tend to synchronize at beta (15–30 Hz) and theta/alpha (6–14 Hz) band frequencies of their local networks (Fig. 2D) (Banaie Boroujeni et al., 2021). While the gamma band synchronization pattern of N-eTypes is consistent between PFC and ACC, their synchronization with theta/alpha and beta bands is more heterogeneous between areas and may subserve area-specific information routing mechanisms (Banaie Boroujeni et al., 2021; Banaie Boroujeni and Womelsdorf, 2023). For example, in the lateral PFC, N-eTypes synchronize at beta frequencies, while in the ACC they synchronize strongly to theta/alpha frequencies (Fig. 2D) (Banaie Boroujeni et al., 2021). N-eType synchronization at slower theta/alpha frequencies in ACC may result from a slower synaptic activity decay time of interneurons in the ACC compared to the LPFC (Banaie Boroujeni et al., 2021; Banaie Boroujeni and Womelsdorf, 2023; Medalla et al., 2017).

3.7. Burst firing dynamics and network state transitions in N-eType neurons

In addition to local network coordination, recent studies have discovered that synchronized activity of N-eTypes is closely linked to activity in distant brain areas, showing that bursts of interneuron activity coincide with increased inter-areal information flow during states of selective top-down attention (Womelsdorf et al., 2014b; Voloh and Womelsdorf, 2018). Burst firing is traditionally associated with the intrinsic bursting of subsets of pyramidal neurons (Gouwens et al., 2019; Zaitsev et al., 2009; Connors and Gutnick, 1990; Izhikevich et al., 2003; Chang and Luebke, 2007), but is also observed in non-pyramidal and non-fast spiking neurons in rat and primate PFC (Cauli et al., 1997; Kawaguchi, 1997; Rotaru et al., 2015). N-eType neurons more likely exhibit burst spiking during periods of dense barrages of coincident synaptic inputs (Krahe and Gabbiani, 2004), such as those occurring in PFC circuits during high attentional processing demands (Voloh and Womelsdorf, 2018). During these states, burst spikes from N-eTypes synchronize at a 4–10 Hz theta frequency band, preceding increases in LFP theta power (Voloh and Womelsdorf, 2018), which suggests their unique role in transitioning the local network to new activation states (Banaie Boroujeni and Womelsdorf, 2023). While interareal synchronization of PFC and ACC ensues from spikes of both N- and B-eType neurons, N-eType spikes synchronize at slightly earlier phases of the beta band frequency in the LFP of the other area compared to B-eType neurons (Womelsdorf et al., 2014b) (Fig. 2E). In addition to firing at an earlier phase, N-eType neurons also show burst spikes that precede increases of oscillatory power in the beta frequency band in the LFP of the other area, while burst spikes of B-eType on average lag the increase of oscillatory power in the remote area (Womelsdorf et al., 2014b) (Fig. 2E). These temporal dynamics suggest that N-eType neurons may play a leading role in controlling and coordinating interareal communication, potentially serving as gatekeepers for the routing of task-relevant information across prefrontal networks.

In summary, N-eType neurons in primate PFC dynamically synchronize to local population activity, particularly in the gamma band (Banaie Boroujeni et al., 2021), show unique phase relations to theta and beta activity that separates them from B-eType neurons (Womelsdorf et al., 2014b; Voloh and Womelsdorf, 2018), and fire spikes that synchronize to network activity across the PFC, not only prior to the spikes of B-eType neurons, but also before oscillatory power increases in the local network (Womelsdorf et al., 2014b). We use these insights, along with the functionally specific contributions of eTypes, to propose a new paradigm in Box 3. This paradigm describes a multi-level investigation of functional eTypes that aims to establish links between cell type-specific activation dynamics and computational circuit motifs.

Box 3. Multi-level analysis of functional eTypes: from cell type specific activation dynamics to circuit motifs.

Characterizing eTypes in vivo, particularly in primates, allows us to explore their functions at multiple levels of understanding: from identifying their origins as inhibitory or excitatory circuit units to analyzing their interactions at the circuit level, understanding their state-dependent rate variations, and investigating their modulation of network dynamics. Such multi-level eType characterization allows testing hypotheses regarding potential circuit motifs and computational mechanisms that eTypes may carry out to realize different behavioral states and cognitive functions. Here, we outline a step-by-step approach to implementing multi-level investigation of eTypes functional contributions in primate studies.

(i). Identification and classification of eTypes

As explained in Box 2, waveform shape and firing dynamics of neurons can be used to cluster neurons into different eTypes. Characteristics such as narrow spike waveforms (N-eType), high firing rates, and non-burst firing spike train can help narrowing down which types may represent inhibitory interneurons (Banaie Boroujeni et al., 2020; Challis et al., 2022; Fin et al., 2025; García and Fishell, 2024; Sajad et al., 2019). It is worth noting that such classifications can be area-specific, meaning that neurons in different brain regions may exhibit distinct eTypes, as has been shown for neurons in visual and prefrontal areas (Sakamoto et al., 2020).

(ii). Validation of inhibitory or excitatory eTypes

After distinguishing eTypes, we can validate their inhibitory or excitatory effects on nearby neurons or on the local net spiking activity. Cross-correlogram (CCG) analysis between neuron pairs can be particularly useful in high-density recordings to understand possible types of monosynaptic connections. The inhibitory or excitatory origin of neurons can be inferred by characterizing an immediate dip or peak in the spike count between neuron pairs (Holt et al., 1996; Salegio et al., 2012; Saunders et al., 2018; Sawaguchi and Iba, 2001; Sawaguchi et al., 1989; Schaeffer et al., 2020).

Another approach involves examining the relative inhibition/excitation effects of neurons of different eTypes on several simultaneously recorded single neurons (Schultz, 2006), or on the firing of local networks by measuring multiunit activity (MUA) (Banaie Boroujeni et al., 2020; Kobayashi et al., 2019). An eType is likely excitatory if its spike is followed by increased MUA and inhibitory if followed by a decrease in MUA within ~20 ms around the spike times (Banaie Boroujeni et al., 2020; Kobayashi et al., 2019).

(iii). Network dynamics of eTypes

Inhibitory interneurons contribute uniquely to oscillatory networks and are entrained by network dynamics. For example, PV interneurons synchronize with local field potential dynamics, which correlate with behavioral function including selective attention and flexible learning (Armstrong et al., 2012; Kawaguchi et al., 2019; Kawaguchi and Shindou, 1998; Kim et al., 2016a; Mahrach et al., 2020; Ma et al., 2015; Schwartz and Beran, 2022; Seamans et al., 2008; Segev and Korngreen, 2007). As a result, understanding functional eTypes is closely related to their coupling dynamics with the local population of neurons or long-range connected networks (Anastasiades and Carter, 2021; Chernov et al., 2014; McGarry and Carter, 2016; Pi et al., 2013; Seidemann et al., 2016; Shen et al., 2015). This can be done by quantifying their spike timing relative to the slow and fast dynamics of local and distant networks. Commonly used techniques to measure these coupling dynamics rely on the distribution of spike times in relation to the spectral content of LFP (Sherfey et al., 2020; Shinomoto et al., 2005, 2009; Siegle et al., 2014; Sirota et al., 2008).

eTypes can differ in phase locking to different oscillatory dynamics, and fire preferentially during specific phases of LFP in a particular frequency band (Anastasiades and Carter, 2021; Banaie Boroujeni et al., 2020; Kawaguchi, 1995; Pfeffer et al., 2013; Pi et al., 2013; Shinomoto et al., 2009; Soares et al., 2017). Phase locking values describe whether and at what frequency spikes are more strongly coupled to the LFP phases, while phase histograms reveal the phase contents of LFP at which spikes preferentially occur in a particular frequency band. The peak frequency of phase locking indicates the dominant network in which eTypes participate, and the phase preference indicates their activation timing in relation to the activity of the network (Sohal et al., 2009; Somogyi et al., 1982). It is important to consider spike current leakages to the LFP, especially at fast dynamics (gamma bands), which can introduce artificial phase locking. This can be effectively addressed by decoupling spike current leakages from the LFP (Banaie Boroujeni et al., 2020; Southwell et al., 2012).

(iv). State-dependent eType dynamics

Interneurons are believed to shape the temporal dynamics of cortical networks, thereby supporting different behavioral states such as attention and learning (Matsumoto et al., 2007; Ma et al., 2015; Merlin and Vidyasagar, 2023; Spitzer and Haegens, 2017). Such state-specific contribution can be examined by measuring if eTypes display different firing variabilities (e.g., rate or lower dimensional trajectories) and network synchronization patterns (Banaie Boroujeni et al., 2020; Chernov et al., 2014; Mahrach et al., 2020; Martina et al., 1998; Spivak et al., 2022). eTypes, for instance, can show rate enhancement or suppression, and synchronize or desynchronize with learning variables (Banaie Boroujeni et al., 2020). This approach can help determine if an eType covaries with specific functional or behavioral states and if this variation is accompanied by changes in larger network dynamics.

(v). Functional circuit motifs of eTypes

The final step is to propose circuit motifs that explain the observed dynamics and eType features and ideally make predictions for future investigations. These circuit motifs can perform specific functions such as gating different streams of information, error adjustment following the evaluation of an outcome, valence processing, and selection between competing inputs (Almeida et al., 2009; Anastasiades and Carter, 2021; Atallah et al., 2012; Banaie Boroujeni et al., 2020; Kobayashi et al., 2019; Seidemann et al., 2016; Stark et al., 2013; Stephan et al., 2019). In general, there are contradictions regarding the specific functions of different cell classes (Amiez et al., 2006; McCormick et al., 1985; Stergiopoulos et al., 2015), which might be a result of different stimulation protocols or other input output consideration while studying a specific cell class (Stevenson et al., 2012; Szadai et al., 2022; Tasic et al., 2018). Functional circuit motifs, in addition to cell type-specific functions, can further elucidate the different mechanisms that eTypes contribute to network dynamics and realize specific functions (Anastasiades and Carter, 2021; Pi et al., 2013; Stark et al., 2013). For example, the same inhibitory eType can participate in multiple motifs and shape different dynamics. In one circuit motif, it is reciprocally connected to two excitatory eTypes and forms a gamma network through a winner-take-all process to resolve uncertainty. In another motif, it is connected to an excitatory eType and competes with an inhibitory eType with slower dynamics to dominate the network for updating values after high prediction errors (Banaie Boroujeni et al., 2020).

4. Prefrontal N-eType neuron contributions to attention control, learning, and working memory in nonhuman primates

A growing number of studies suggest unique contributions of interneuron firing and synchronization to higher cognitive functions during cognitive task performance. Studies modulating GABA concentrations in PFC and ACC have documented that interneuron activity is instrumental to sustaining working memory (Rao et al., 1999; Sawaguchi et al., 1989; Sawaguchi and Iba, 2001), attention (Paine et al., 2011, 2015), and the flexible adjustment and learning of goal directed behavior (Amiez et al., 2006; Rich and Shapiro, 2007; Enomoto et al., 2011; Urban et al., 2014). Early studies found that functional effects of GABA interventions during working memory performance modulated responses of N-eType neurons, which showed slightly broader spatial tuning during the cue, delay and test periods of the oculomotor delayed response task (Constantinidis and Goldman-Rakic, 2002; Wilson et al., 1994; Wang et al., 2004). These early findings gave rise to an influential computational interneuron model of working memory (Wang et al., 2004) and to multiple studies in NHPs that characterized N-eType neuron responses in the lateral PFC (areas, 9/46, 8, 9), frontal eye field (FEF), ventromedial PFC (area 32), anterior cingulate cortex (area 24), and supplementary eye field (SEF). These studies explore distinct activity profiles of N-eType neurons during rule-guided behavior using attentional cueing and anti-saccade tasks (Banaie Boroujeni et al., 2021; Johnston et al., 2009; Diester and Nieder, 2008; Hussar and Pasternak, 2009; Thiele et al., 2016; Fan et al., 2017; Kawai et al., 2019; Sajad et al., 2019; Ceccarelli et al., 2023), working memory using delayed response tasks (Constantinidis and Goldman-Rakic, 2002; Rao et al., 1999; Diester and Nieder, 2008; Roussy et al., 2021; Qi et al., 2011; Hussar and Pasternak, 2012; Sakamoto et al., 2020; Chung et al., 2023; Wong et al., 2023), and flexible learning using tasks requiring stimulus-reward reversal learning and stimulus-response association learning (Banaie Boroujeni et al., 2021; Hassani et al., 2024; Kawai et al., 2019; Ceccarelli et al., 2023; Shen et al., 2015; Oemisch et al., 2019). Surveying results of these studies (Fig. 3A) suggest that N-eType neurons have functional activation signatures during task performance that are distinct from those of B-eType neurons during attention control, working memory, and learning. While these studies are not proportionally similar across different NHP brain areas and tasks, overall, more studies report modulation of N-eType neurons in attention, rule-guided tasks, and flexible learning, while B-eType neuron activity is more frequently reported to be modulated in working memory tasks (Fig. 3A).

Fig. 3.

Mapping of N-eType neurons to prefrontal cortex functions in Nonhuman Primates. (A) Meta-survey of twenty-five NHP studies distinguishing N- and B-eType neurons from medial and lateral PFC (colored subfields in rendered brain) and reporting their activation during tasks involving flexible learning (e.g., reversal learning and stimulus-response association learning tasks), rule-guided attention tasks (e.g., spatially cued target detection tasks), and working memory (e.g., delayed response tasks). Line lengths reflect relative number of studies in each area and for each type of task. Colors indicate whether a study reported N-eType neurons showed stronger, no, or weaker (red, gray, and blue, respectively) task-relevant firing. (B) N-eType neurons (red) from the lateral PFC encode more likely than B-eType neurons the association of a stimulus and a target response location (upper panel). Encoding was particularly enhanced during learning and remained elevated when learning completed (middle panel). Cross-temporal decoding accuracy (bottom panels) was weak for N-eType neurons during learning (left) and remained relatively weak after learning (right), suggesting a dynamic as opposed to a stable coding of task information. (C) N-eType neurons were distinguished in lateral PFC and anterior cingulate cortex (ACC) while monkeys learned to associate colors with reward in a reversal task (upper panel). In lateral PFC the color cue triggered a low gamma (~40 Hz) spike-LFP synchronization in a fast-spiking N-eType subclass specifically when the cue relevance was uncertain during learning. Uncertainty of the cue corresponds to low choice probability and modulated low gamma activity in lateral PFC (green outlined panel). Uncertainty of the outcome corresponds to periods with large prediction error and modulated low gamma activation of the same fast-spiking N-eType subclass in ACC (purple outlined panel). Shaded areas show significant synchronization peaks. Panel B and C adapted from (Banaie Boroujeni et al., 2021; Ceccarelli et al., 2023).

4.1. Signaling behaviorally relevant stimuli in N-Type neurons of the lateral prefrontal cortex

In lateral PFC N-eType neurons respond more strongly than B-eType neurons to covert attention cues that carry information about target colors (Banaie Boroujeni et al., 2021), to cues indicating the visual target stimulus for a categorical judgement (Diester and Nieder, 2008), to cues indicating the spatial response location in the pro/anti-saccade paradigm (Johnston et al., 2009), to cues indicating the direction of a rewarded target response (Ceccarelli et al., 2023; Qi et al., 2011), and to cues indicating stay or switch decisions (Fan et al., 2017; Ceccarelli et al., 2023) (Fig. 3B). The stronger on-response of N-eType neurons to behaviorally relevant cues differ from those of B-eType neurons. First, the cue onset effect in N-eType neurons is already strong during the learning phase when NHPs acquire the cue’s conditional significance, for example the association of stimulus patterns with a response locations (Ceccarelli et al., 2023)(Fig. 3B), and continues increasing with learning, as demonstrated by a positive correlation of N-eType neuron firing with increases in choice probability during learning (Banaie Boroujeni et al., 2021). The learning related modulation of N-eType neurons is not only evident in firing rate modulations. In a task requiring subjects learning the relevance of visual features, subtypes of N-eType neurons in LPFC and ACC synchronized at gamma band frequencies when the to-be-learned cue or the uncertain outcome was revealed (Banaie Boroujeni et al., 2021)(Fig. 3C). This cue- and outcome- triggered gamma activation was apparent in a fast spiking subclass of N-eType neurons, but not in other N-eTypes and neither in B-eType neurons.

The prominent cue-onset response of N-eType neurons appears to particularly dependent on the behavioral relevance of the stimulus. N-eType neurons show stronger reductions in firing selectivity for task irrelevant, distracting stimulus features (Hussar and Pasternak, 2009), larger rate modulations than B-eType neurons when top-down attention is directed to their receptive field (Dasilva et al., 2019; Thiele et al., 2016), and when actual choices are prepared (Ding and Gold, 2012). One should bear in mind, however, that the effect sizes of these attention related response modulations are small for N-eTypes in FEF and vary across different subgroups of N-eTypes (Dasilva et al., 2021). One reason why N-eType neurons in PFC are less responsive to irrelevant or distracting stimuli may relate to lower spatial sensitivity to these stimuli. In a study of area 8/46, the spatial receptive fields of N-eType neurons to task-irrelevant sweeping bars were reported to be smaller (6.1 °) than those of B-eType neurons (8.7 °) (Viswanathan and Nieder, 2017b). This suggests that without top-down relevance, fewer N-eType neurons respond to stimuli at spatially irrelevant locations than B-eType neurons. Finally, the stimulus onset responses of N-eType neurons are more homogeneous than those of B-Type neurons. One study found that a higher proportion of N-eType neurons (83%, n = 25 of 30) encoded distractor information during the delay period of a delayed response task compared to B-eType neurons, although the average encoding strength was similar across both types (Jacob et al., 2016).

4.2. Working memory and N-eType neurons in lateral prefrontal cortex

Multiple studies have characterized N-eType neurons in lateral PFC in working memory tasks requiring delayed match-to-sample rules, sequential choices, and spatial planning (Fig. 3A). In remarkable agreement these studies show that N-eType neurons and B-eType neurons in lateral PFC similarly likely encode target stimulus variables during working memory delays (Rao et al., 1999; Roussy et al., 2021; Fan et al., 2017; Qi et al., 2011; Chung et al., 2023; Jacob et al., 2016). Only rare reports suggest that a moderately higher proportion of N-eType neurons encode target information during the delay (Diester and Nieder, 2008). However, the selectivity of delay cell firing of N-eType neurons appears somewhat weaker than in B-eType neurons, evident in a broader tuning to the target (sample) stimulus (Constantinidis and Goldman-Rakic, 2002; Diester and Nieder, 2008; Hussar and Pasternak, 2009, 2012; Wong et al., 2023). N-eType neurons are more broadly tuned to the stimulus towards the end of the delay period, which contrasts to subgroups of B-eType neurons that increase their target selectivity over time (Hussar and Pasternak, 2012; Sakamoto et al., 2020). The broader tuning of N-eType neurons during the delay of working memory applies to a variety of stimulus variables including stimulus locations, direction and speed of motion, numerical stimulus category, and reward magnitude tuning, which can involve increased firing to non-preferred stimulus features as well as decreased firing to preferred features (Diester and Nieder, 2008; Hussar and Pasternak, 2009; Fan et al., 2017). Broader tuning to the feature of a target stimulus indicates that N-eType neurons sharpen the tuning of B-eType neurons, because they respond to a wider range of stimulus features than B-eType neurons and thereby suppress activity in pyramidal cells that are weakly activated by a stimulus with features away from their preferred feature (Diester and Nieder, 2008). Consistent with this scenario, pairs of N-eType neurons show stronger firing correlations and are more likely tuned to similar locations when located anatomically adjacent to each other, which differs from pairs of N- and B-eType neurons that frequently show opposite tuning even when recorded from the same electrode location (Rao et al., 1999; Diester and Nieder, 2008; Viswanathan and Nieder, 2017a).

4.3. N-eType neurons in anterior cingulate cortex and supplementary eye field

Medial PFC areas in which N-eType neurons have been studied include ventromedial PFC, ACC, and the SEF. These areas show higher proportions of inhibitory interneurons with stronger inhibitory synaptic currents compared to the lateral PFC (Medalla et al., 2017; Medalla and Barbas, 2009). Interneuron functions in these areas have been inferred from comparing N- and B-eType neurons in tasks involving reversal learning, strategy selection, attention, and response inhibition (Fig. 3A). These studies found that N-eType neurons encode modestly stronger error outcomes (~<10%) than correct outcomes compared with B-eType neurons in choice tasks with deterministic outcomes (Sajad et al., 2019; Shen et al., 2015; Westendorff et al., 2016), and show a more prominent encoding of outcome history during probabilistic reversal learning (Kawai et al., 2019). Stronger encoding of outcome information suggests that N-eType neurons play an important role in learning from outcomes, which is supported by findings from a feature-based reversal learning task (Banaie Boroujeni et al., 2021). This task involved learning which stimulus color is associated with reward by choosing among multiple stimulus features and inferring the target color through feature-specific credit assignment. During learning, the activity of N-eType neurons more strongly correlated with reward prediction error (RPE) activity (Gabbott and Bacon, 1996a), with a significant higher probability of N-eType neurons to encode RPEs for the specific visual feature that gave rise to the prediction error (Oemisch et al., 2019). This feature-specific encoding of RPEs in N-eType neurons indicates that inhibitory neurons in ACC combine information about the expected value of a visual feature with the reward outcome that was received when that specific feature was chosen. Finer grained analysis showed that N-eType neurons in ACC, but not in lateral PFC, engaged in gamma-band synchronization when a choice outcome was unexpected and RPE’s were large, but not when choice coutcomes were expected (Fig. 3C) (Banaie Boroujeni et al., 2021). Enhanced gamma-band synchronization was most evident in a subgroup of N-eType neurons with a fast-spiking firing pattern, regular inter-spike intervals and relatively high firing rates (Banaie Boroujeni et al., 2021). These N-eType specific functional signatures during learning suggest they contribute to improving future behavior. Consistent with this claim, N-eTypes in ACC encode the outcome of a choice stronger than B-eTypes when the subject changes (switches) their choice in subsequent trials, as opposed to continuing (staying) with prior choices (Kawai et al., 2019).

In summary, N-type neurons are associated with stronger activation to behaviorally relevant cues and a broader tuning to behaviorally relevant targets during working memory tasks in lateral PFC. They are also associated with more informative and stronger signaling of prediction errors and outcomes during adaptive behavior in ACC.

5. Unique functional roles of PV, SST and VIP interneurons in rodent prefrontal cortex

In contrast to the eTyping of neurons in studies of primate PFC, studies of mouse mPFC have used optical tagging of molecularly defined interneuron types to infer their specific functions during cognitive tasks. These studies assume that interneurons of the same type should activate whenever their putative functional role is enacted, and that interneuron activity should vary with the animal's performance of tasks requiring that specific function. Support for this assumption comes from many rodent studies involving interneurons expressing PV, SST, or VIP in pre-/infra-limbic areas (mPFC) and anterior cingulate cortex (ACC) (Fig. 4A). Rodent mPFC and ACC are considered prefrontal cortices in the mouse brain because of their high intra-modular connectivity (Harris et al., 2019), stronger input and output connectivity compared to other cortical areas (Le Merre et al., 2021; Anastasiades and Carter, 2021), and selective neuronal activation during higher order cognitive processes, including working memory, attention, and flexible learning (Le Merre et al., 2021; Carlén, 2017; Reinert et al., 2021). Interneuron subtypes supporting these higher order cognitive processes are distinguished either by optically stimulating and recording from neurons expressing light-responsive channels (Kim et al., 2016a, 2016b; Kvitsiani et al., 2013; Pi et al., 2013; Morris et al., 2016); or by observing activity-based fluorescent changes from genetically encoded calcium indicators expressed in specific interneuron subtypes (Box 1) (Kamigaki and Dan, 2017; Pinto and Dan, 2015; Ferguson et al., 2023; Szadai et al., 2022; Allen et al., 2017). These approaches in mPFC and ACC demonstrate that PV, SST, and VIP interneurons modulate their activity during discrete epochs of discrimination-, memory-, and attentional control tasks, summarized in Fig. 4 (Kamigaki and Dan, 2017; Cho et al., 2020; Kim et al., 2016a, 2016b; Kvitsiani et al., 2013; Pi et al., 2013; Morris et al., 2016; Pinto and Dan, 2015; Ferguson et al., 2023; Szadai et al., 2022). It is important to bear in mind that, especially regarding these cognitive domains in rodents, there is no general consensus among neuroscientists. However, our aim is simply to survey these studies using the terminology generally accepted in the original works, in order to provide a roadmap for future studies in NHPs.

Fig. 4.

Mapping of interneuron activation to prefrontal cortex functions in mice. (A) Semi-transparently rendered mouse brain and coronal section with segmented infralimbic (IL) and prelimbic (PL) cortex, and anterior cingulate cortex (ACC). (B) Task structure for sensory discrimination tasks. After a preparation signal, mice receive an auditory cue, and they must determine whether the cue is the target or not (top). A correct detection is rewarded, whereas a false detection is typically punished. The bottom panel shows a similar structure with a working memory component. Mice are first presented with a sample stimulus, then after a delay, they must determine whether the second stimulus is the same as the first. (C) Illustration of response profiles of PV, SST, and VIP interneurons (Magenta, blue, green) in mouse mPFC. The PV interneurons respond prominently to discrete epochs of these task paradigms, including preparation, sensory cue, punishment, and reward (Kamigaki and Dan, 2017; Kvitsiani et al., 2013; Pinto and Dan, 2015; Allen et al., 2017). SST interneurons respond during the choice epoch (goal run time) and to a lesser extent to the sensory cues (Kamigaki and Dan, 2017; Kim et al., 2016b; Kvitsiani et al., 2013; Pinto and Dan, 2015). VIP interneurons respond to reward outcomes as well as punishment (Kamigaki and Dan, 2017; Pi et al., 2013; Pinto and Dan, 2015; Szadai et al., 2022). (D) Compared with PV interneurons, SST interneurons show stronger target selectivity. PV interneurons show an earlier onset and less selective response to target- vs. non-target stimuli (E) Schematic of a delayed non-match to sample spatial working memory task. (F) SST (blue) but not PV (red) interneurons are better able to predict the identity of a sample when trained on their activity during the delay period of correct trials (solid lines). Overall, SST interneurons show stronger spatial selectivity for working memory content than PV interneurons. (G) An example of an attention-demanding task, requiring sustained attention preparing for detecting a target light to collect rewards from a reward port. (H) PV interneurons show elevated activity during attentional periods, and their activity is greater during attention periods in correct compared to incorrect trials (Kim et al., 2016a; Ferguson et al., 2023). (I) Schematic summary of the effects of optical activation (left) and suppression (right) of interneuron types in mPFC on performance accuracy in tasks requiring memory, attention control, and flexible learning. Activating PV interneurons enhances performance accuracy for attention-demanding tasks (Kim et al., 2016a; Ferguson et al., 2023) and tasks requiring flexible learning (Cho et al., 2015, 2020; Murray et al., 2015; Cho et al., 2023), but interferes with memory-required tasks (Kamigaki and Dan, 2017). PV activity suppression negatively affects attentional control (Kim et al., 2016a; Ferguson et al., 2023) and flexible learning (Cho et al., 2015; Murray et al., 2015). Activation or suppression of SST interneurons during the delay period negatively affects performance accuracy in memory-guided and delayed discrimination tasks, respectively. Activating VIP interneurons enhances performance on delayed discrimination tasks (Kamigaki and Dan, 2017), while suppressing a subtype of this interneuron type negatively affects attention (Obermayer et al., 2019). (J) An illustration of a low-dimensional activation state space of interneuron subtypes, mapped across four cognitive domains, based on surveyed studies of functional activation patterns and causal manipulations.

(a) Adapted from (Carlén, 2017). (b) Adapted from (Kamigaki and Dan, 2017; Pinto and Dan, 2015; Szadai et al., 2022). (d) (top and bottom panels adopted from (Pinto and Dan, 2015) and, respectively). (f) (Kim et al., 2016b; Abbas et al., 2018), left panels adopted from (Kim et al., 2016b)).

5.1. Interneurons respond to discrete task events in discrimination tasks

In sensory discrimination tasks, mice distinguish target from non-target stimuli, and lick or nose poke responses reward them for correct choices and penalize them for erroneous ones (Fig. 4B, top). Discrimination tasks unfold in discrete epochs spanning preparatory signals, cues, delays, targets, and outcomes (Fig. 4B, bottom). PV interneurons prominently activate to the onset of these discrete task epochs (Kamigaki and Dan, 2017; Kvitsiani et al., 2013; Pinto and Dan, 2015; Allen et al., 2017) (Fig. 4C). They respond more to preparatory signals, sensory cues, and rewards, and less to punishments (Kamigaki and Dan, 2017; Kvitsiani et al., 2013; Pinto and Dan, 2015; Allen et al., 2017). Unlike PV interneurons, SST and VIP interneurons respond less consistently during discrete task events. SST interneurons selectively respond to cues that indicate a target, increase firing during the decision-making process, and suppress firing after collecting reward (Kim et al., 2016b; Kvitsiani et al., 2013) (Fig. 4C). VIP interneurons on the other hand, are activated most by outcome events, with a stronger response to punishment than reward (Kamigaki and Dan, 2017; Pi et al., 2013; Pinto and Dan, 2015; Szadai et al., 2022), which is a pattern also observed outside of mPFC (Pi et al., 2013; Szadai et al., 2022) (Fig. 4C).

In sensory discrimination tasks, PV interneurons respond non-selectively to exogenous attentional cues, while SST interneurons selectively activate when a target is presented (Pinto and Dan, 2015; Fig. 4D, top) or when the second (test) stimulus after a delay matches the cue stimulus (Kamigaki and Dan, 2017; Fig. 4D, bottom). This greater selectivity of SST neurons relative to PV neurons might be related to decision making processes that compare sensory information with endogenously memorized target information. Retrieval of memorized target information underlies successful performance in spatial delayed non-match to position tasks, where mice need to choose running through an arm of a Y- or T- maze that is different than the one visited during a prior ‘sample’ run (Fig. 4E). SST interneurons in mPFC show spatially selective responses, with ensembles of SST neurons more accurately predicting the sample identity during delay periods in correct trials (Kim et al., 2016b) (Fig. 4F).

The spatially selective responses of SST neurons distinguish them from PV interneurons. This dissociation of response profiles suggests that SST neurons contribute to memory-guided choice tasks where decision variables are endogenously generated, while PV interneurons contribute to attention-demanding tasks involving the detection of exogeneous cues to drive choices, such as five/three-choice serial reaction time (CSRT) tasks (Fig. 4G). In CSRT-like tasks, PV interneurons increase their firing during periods of sustained attention in trials leading to correct responses, compared to omitted or incorrect trials (Kim et al., 2016a; Ferguson et al., 2023) (Fig. 4H). This response profile is consistent with computational studies suggesting a role of PV interneurons in controlling the output gain of the local circuit (Wilson et al., 2012; Tiesinga et al., 2004; Atallah et al., 2012; Ferguson and Cardin, 2020). In contrast, SST interneurons are more closely associated with internally maintaining information about spatial or action plans (Kim et al., 2016b; Abbas et al., 2018; Cummings and Clem, 2020). This short-term memory function of SST neurons is reminiscent of their spatial selectivity in other brain areas (Miao et al., 2017). In comparison to PV and SST interneurons, the functional roles of VIP interneurons have not been extensively studied in mPFC. However, one commonality across existing studies suggests that VIP neurons respond most prominently following the onset of outcomes (Kamigaki and Dan, 2017; Pi et al., 2013; Pinto and Dan, 2015; Szadai et al., 2022), which suggests they are involved in credit assignment and outcome monitoring processes.

5.2. mPFC and ACC interneuron contributions to attention control, working memory and memory

Similar to N-eType neurons in the primate, subtypes of interneurons in mouse mPFC respond prominently during task epochs requiring attention control and flexible learning. Optogenetically suppressing PV interneurons in mPFC during attention-demanding periods impairs performance, suggesting they are necessary for successful attention control (Kim et al., 2016a; Ferguson et al., 2023). Indeed, activating PV interneurons at gamma frequency (40 Hz) can rescue attention deficits in PV knock-down mice (Ferguson et al., 2023), and improve performance in attention demanding tasks (Kim et al., 2016a). Similarly, mice with either a mutation affecting PV interneurons in the mPFC, or with optogenetically suppressed PV activity, are impaired in flexible learning tasks (Cho et al., 2015, 2020; Murray et al., 2015; Cho et al., 2023), while activating PV neurons, especially at the gamma frequency, rescues these behavioral deficits or enhances subjects’ cognitive abilities in tasks that involve flexible switching between different rules (Cho et al., 2015, 2020, 2023; Chamberlin et al., 2023) (Fig. 4I).

Experimental manipulations of SST interneurons in rodent mPFC have focused primarily on memory-guided tasks. Activation or suppression of SST interneurons negatively affects performance on memory-guided tasks, particularly when they require long-term retention of information (Kim et al., 2016b; Abbas et al., 2018; Cummings and Clem, 2020). Activating SST interneurons in the ACC during a delay period decreases target detection accuracy (Kamigaki and Dan, 2017). This deficit in target detection suggests that the regulation of SST+ interneuron activity is essential for optimal performance in tasks involving information retention, spatial memory maintenance, and action planning. Studies causally interfering with VIP interneurons in mouse mPFC also reveal specific functional effects. Activating VIP neurons improves memory maintenance during delay periods in sensory discrimination tasks (Kamigaki and Dan, 2017), whereas inhibiting subgroups of VIP neurons that express choline acetyltransferase (ChAT) impairs attention performance (Obermayer et al., 2019). These findings suggest a role for VIP interneurons for memory retention and attention control, which likely involves disinhibitory circuit mechanisms (Kamigaki and Dan, 2017).

In summary, studies causally manipulating and tracking the activation profiles of interneuron subtypes in rodent mPFC and ACC during different tasks suggest distinguishable functional contributions of PV, SST, and VIP interneurons (Fig. 4J). While avoiding oversimplification, we summarize these results with a tentative low-dimensional structure of neural activity states mapped across four cognitive domains. This schematic should be viewed as a heuristic representation that can reflect the relative contributions of interneuron subtypes across tasks, rather than implying a one-to-one relationship between cell type and function. The low-dimensional state space reflects how the relative weight of each interneuron subtype can shift depending on task demands, including attention control, flexible learning, memory maintenance, and outcome processing. PV interneurons are essential when rapid adjustment or reconfiguration of input-output circuit gain is required, for example, when attentional demands are increased (Kim et al., 2016a; Ferguson et al., 2023) or when flexible rule switching is required (Cho et al., 2015, 2020). SST interneurons play a more prominent role in retaining selective information during spatial working memory and memory-guided action planning (Kim et al., 2016b; Abbas et al., 2018; Cummings and Clem, 2020). Lastly, VIP interneurons play prominent roles in encoding the valence of outcomes, modulating attentional control and maintaining specific information contingent on task outcomes (Kamigaki and Dan, 2017; Pinto and Dan, 2015; Szadai et al., 2022; Obermayer et al., 2019) (Fig. 4J).

We emphasize that what we refer to as cell-type specific functional contributions does not imply a simplistic mapping (e.g., “PV = attention”), but rather represents an abstraction of circuit-level findings derived from specific tasks in rodents. These relationships may not directly translate to higher-order cognitive functions in primates, where task structures and neural representations can vary significantly. Instead, this conceptual framework can serve to guide future NHP and human studies in designing tasks that are sensitive to cell-type specific contributions and to inform computational models that aim to link interneuron dynamics to cognitive functions.

5.3. Cross-species functional convergence across task domains

A direct comparison of similarities and dissimilarities in cell-type specific functions between primates and rodents is not trivial, due to difficulties in mapping both behavioral tasks and neuron types across species. Despite these limitations, across conceptually comparable tasks, prefrontal interneurons in rodents and primates show both convergent and divergent functional motifs. During attention tasks, N-eType neurons in the macaque PFC, particularly subtypes with fast-spiking phenotypes (non-bursting with high firing rates), synchronize with gamma-band activity and respond selectively to attentional cues (Banaie Boroujeni et al., 2021; Thiele et al., 2016), similar to PV interneurons in the rodent medial PFC during attentional control (Kim et al., 2016a; Ferguson et al., 2023; Ferguson and Cardin, 2020). In tasks requiring memory retention, findings in NHPs show that while N-eType neurons encode target information during delay periods, they exhibit broader tuning compared to B-eTypes (Hussar and Pasternak, 2012; Wong et al., 2023). However, the difficulty in detecting SST neurons in NHPs due to their less distinguishable waveform shapes makes assessing the direct contribution of this interneuron type more complex (Torres-Gomez et al., 2020). In tasks involving flexible learning and behavioral adaptation, studies in NHPs demonstrate that N-eType neurons encode reward prediction errors more strongly than B-eTypes (Banaie Boroujeni et al., 2021; Hassani et al., 2024; Oemisch et al., 2019), show enhanced responses to outcome history, (Kawai et al., 2019) and exhibit gamma-band synchronization during unexpected outcomes (Banaie Boroujeni et al., 2021) consistent with the involvement of PV interneurons in behavioral flexibility and rule switching in rodents (Cho et al., 2015, 2020). Together, despite methodological constraints and differences in the structure of the PFC in NHPs and rodents, the basic computational functions of major interneuron subtypes, particularly those related to gain modulation, disinhibition, and oscillatory synchronization, appear to be conserved in the PFC. This motivates future studies to better understand how different neuron types and inhibitory circuit-motifs support cognitive functions across species.

6. Functional computations of prefrontal cortex interneurons

The functional contributions of interneurons in primate PFC (Fig. 3) and mouse mPFC (Fig. 4) are realized through cell-type specific circuit mechanisms. Cell-type-specific circuit motifs have been proposed for cortical computations of PV, SST and VIP neurons (Kepecs and Fishell, 2014; Womelsdorf et al., 2014a; Roux and Buzsáki, 2015; Cardin, 2018). Here, we survey recent progress understanding the role of CR/VIP neurons in disinhibitory circuit motifs that uniquely contribute to PFC computations. CR/VIP neurons are particularly prolific in primate PFC and preferentially synapse onto other interneurons, spanning different layers (Gabbott et al., 1997; Krienen et al., 2020; Gabbott and Bacon, 1996b; Džaja et al., 2014; Medalla et al., 2023; Hladnik et al., 2014). In mouse mPFC, disinhibitory circuit motifs appear to support associative learning, memory maintenance, and selective gating (Kamigaki and Dan, 2017; Pi et al., 2013; Letzkus et al., 2011, 2015; Wang and Yang, 2018; Lee et al., 2019; Kullander and Topolnik, 2021; Zhang et al., 2022). The computational mechanisms underlying these disinhibitory effects involve separable architectures. We survey architectures that use CR/VIP disinhibitory motifs to implement (i) the gating of top-down signals, (ii) selective short-term maintenance of information, (iii) inhibitory plasticity supporting learning, and (iv) the dynamic switching of network states contingent with these computations (Fig. 5). The motifs described below are based on anatomical data that describe local circuit structure (specific dendritic locations and cell types that interneuron subtypes target, Fig. 1), electrophysiological data that predict frequency preferences and spike patterns of cells (Fig. 2) and functional data on the activity patterns (Figs. 3,4). Each motif is motivated by a computational study that integrates these structural and electrophysiological data to predict or reproduce the observed functional activity patterns. We highlight which aspects were inferred and therefore require further experimental validation, as well as possible alternative explanations.

Fig. 5.

Role of interneurons for functional computations. (A) Gating through disinhibition motif: Inputs arriving at different dendritic branches of a pyramidal cells are gated through spatially specific disinhibition of SST interneurons by VIP interneurons (Yang et al., 2016). (B) Working memory motif: Sustaining activity after the driving cue disappears is achieved by two inhibitory pools that inhibit each other (Kim and Sejnowski, 2021). Left: input drives an E and I pool that are not directly coupled; Right: without input, the active I pool stays active because it disinhibits the E pool that in turn provides input to it to sustain the I pool’s activity. The illustrated circuit stores or maintains two different activity patterns, but it can be extended to larger numbers of patterns. (C) Prediction error motif: This interneuron motif ensures that pyramidal cells only activate when the sensory input does match what was predicted. Sensory input activates PV interneurons, while predictions activate SST and VIP interneurons. When predictions and sensory inputs do not match, the PV interneurons are inhibited. This releases inhibition of the pyramidal cell whose activation signals the mismatch of input and prediction. The motif can be trained with various inhibitory plasticity rules (Hertäg and Sprekeler, 2020). (D) State switching motif: Example of a firing rate model that uses three types of interneurons to switch and stay in either of two states characterized by either high E cell activity in the gamma band, or a state with low E cell activity and beta band activity (Hahn et al., 2022). In the disinhibited state, PV and VIP interneurons are active, PV interneurons provide gamma patterned activity to the E cell, whereas VIP inhibit SST interneurons and release E cells from their inhibition. In the inhibited state, SST inhibits both PV/VIP and E cells, while self-synchronizing at the beta band. For additional motifs that dynamically switch states using interneurons, see (ter Wal and Tiesinga, 2021).

6.1. Selective gating of inputs by disinhibitory circuit motifs

N-eType neurons in PFC and ACC of NHPs fire vigorously to cue onsets and correlate their firing with behaviorally relevant stimulus features (Fig. 3A). These functional epochs are characterized by increased input to circuits that require mechanisms to resolve competition between these inputs (Reynolds and Desimone, 2003) or learn the appropriate associations between two specific groups of inputs (Letzkus et al., 2015). When multiple input sources activate a circuit, the concept of gating provides a mechanistic understanding of how synaptic inputs are weighted and certain interneurons activate during specific task epochs. Principal neurons can be viewed as receiving inputs in two broadly separable compartments, the apical dendrite and the basal dendrites/soma; each integrating unique sources of information, for instance, top-down feedback via tuft/apical dendrites and feedforward sensory information via basal dendrites (Murayama et al., 2009; Larkum, 2013). Interneuron subtypes differentially target these compartments; SST neurons synapse onto apical dendrites while PV neurons synapse onto the soma and perisomatic regions, and differentially contribute to the gating of behaviorally relevant inputs as well as modulating their output gain in local processing (Fishell and Kepecs, 2020; Wang et al., 2004; Urban-Ciecko and Barth, 2016). Recent findings in rodents support that a disinhibitory circuit involving VIP, PV, and SST interneurons can realize local gain modulations without influencing feature selectivity and gate relevant inputs through translaminar integration of information. Such selective gating in PFC is critical, as it receives a rich diversity of inputs from distant brain areas (Medalla and Barbas, 2009; Anastasiades and Carter, 2021; Wang and Yang, 2018; McGarry and Carter, 2016; Liu et al., 2020a). One circuit motif supporting this selective gating mechanism involves two types of inhibitory interneurons that implement spatially selective disinhibition of dendritic branches of cortical pyramidal neurons (Fig. 5A) (Yang et al., 2016). According to this motif, gating of one input stream over another is achieved by selectively activating VIP interneurons that inhibit SST cells targeting dendritic regions receiving the relevant input, which disinhibits the local dendritic branch (Fig. 5A). Taken together, this motif is supported by local circuit structure measured in rodents and predicts activity patterns that have been partially observed, but also implies correlations among three cell types that remain to be validated.

6.2. Selective and task specific information maintenance

In addition to gating behaviorally relevant inputs, disinhibitory circuit motifs may also support short-term memory. Primate and rodent studies (Figs. 3, 4) suggest that inhibitory circuits maintain relevant information in working memory during goal-directed behavior (Kamigaki and Dan, 2017; Kim et al., 2016b; Cummings and Clem, 2020). In support, a modeling framework by Kim and Sejnowski (Kim and Sejnowski, 2021) suggests that inhibitory-inhibitory circuit mechanisms enable PFC and other association cortices to efficiently encode and retain information from multiple stimuli over longer timescales. According to this motif, interneurons form separable pools with mutual inhibitory connections when driven by a relevant sensory cue (Fig. 5B). Once the cue disappears, a small pool of inhibitory neurons remain active because they disinhibit principal cells that encode relevant cue information, which subsequently sustains the inhibitory pool’s activity (Kim and Sejnowski, 2021). Neurons in this network show spiking activity that resembles empirically measured firing in the mPFC (Qi et al., 2011; Meyer et al., 2011). Networks optimized within this framework to solve a passive fixation task without requiring working memory, did not display this inhibitory network structure, suggesting that connections among interneurons were essential for learning a working memory task. Simply put, strong mutual inhibition creates a bistable system, where activity in one inhibitory pool is sustained by excitation generated by disinhibition of principal cells, via suppression of the other inhibitory pool. There are also alternative circuit motifs based on recurrent excitatory networks and the slow time course of NMDA receptor–mediated currents that have been proposed as mechanisms for working memory (Durstewitz et al., 2000; Compte et al., 2000).

6.3. Inhibition-specific plasticity rules supporting learning of predictions

Interneuron types in rodent mPFC respond to outcomes (Fig. 4), and subtypes of N-eType neurons in primate ACC index prediction errors in learning tasks (Banaie Boroujeni et al., 2021; Oemisch et al., 2019) (Fig. 3C). Computing prediction errors requires integrating inputs about received and expected outcomes. This integration can be achieved in motifs of interconnected PV, SST, and VIP neuron types that connect to principal cells with unique somatic and dendritic specificity (Hertäg and Sprekeler, 2019, 2020). One well described motif generates negative prediction error responses in pyramidal cells when there is a mismatch between a predicted input and a sensory input (Fig. 5C). Learning is achieved by balancing excitation and inhibition through experience-dependent plasticity, similar to what has been documented for mouse visual cortex (Attinger et al., 2017). In this motif, sensory input activates the soma of pyramidal neurons, as well as PV and SST interneurons, while input carrying outcome predictions activates pyramidal neuron dendrites and VIP interneurons. When sensory input alone does not cause activation and prediction input arrives in the absence of the predicted sensory input, the pyramidal cell is disinhibited, thus signaling the mismatch (Fig. 5C). Balancing different inhibitory pathways is necessary for this mismatch detection, achievable through plasticity rules awaiting biological testing (Hertäg and Sprekeler, 2020). Expanding the above motif with variations of inhibitory connections and plasticity rules suggests specific local circuit mechanisms underlying prediction error signals (Hertäg and Clopath, 2022). These extended motifs support rapid adaptation and learning processes that are well suited to support PFC circuits when they integrate multiple top-down (prediction) and bottom-up (sensory) inputs to update predictions about the expected value of each possible during goal directed behavior (Matsumoto et al., 2007; Banaie Boroujeni et al., 2021; Oemisch et al., 2019; Klein-Flügge et al., 2022).

6.4. Interneurons dynamically switching networks states

The circuit motifs above relied on interneuron activity to implement selective gating, maintaining and updating of working memory content and of predictions during learning. In PFC, these processes are accompanied by distinct oscillatory activity dynamics (Akam and Kullmann, 2010; Womelsdorf et al., 2014a; Banaie Boroujeni et al., 2021; Benchenane et al., 2011; Voloh et al., 2015), suggesting that synchronization at gamma, beta, and theta/alpha frequency bands characterizes different processing states of the neural network (Cho et al., 2015; Banaie Boroujeni et al., 2021; Kim et al., 2016a; Lundqvist et al., 2016; Mendoza-Halliday et al., 2024). A growing literature suggests that establishing and switching these processing states depends on inhibitory circuit motifs (Chen et al., 2017; Domhof and Tiesinga, 2021; ter Wal and Tiesinga, 2021; Hahn et al., 2022). Empirical studies have described a inhibitory motif in superficial and deep layers that implements state transitions based on recurrent inhibition between PV/VIP cells and SST cells, translaminar connections from VIP cells, and bidirectional connectivity between interneurons and pyramidal cells (Pfeffer et al., 2013; Jiang et al., 2015). Modeling this cross-layer motif suggests that each layer supports two distinct states: a disinhibited state where PV/VIP neurons inhibit SST cells, releasing pyramidal cells from inhibition and increasing their activity; and an inhibited state where SST neurons inhibit PV/VIP cells, thereby suppressing pyramidal cell activity (Fig. 5D) (Hahn et al., 2022). The oscillatory power in each state follows the intrinsic time scales of the active interneuron types: beta for the SST dominated state and gamma for the PV dominated state (Hahn et al., 2022), matching findings of other computational studies (Domhof and Tiesinga, 2021; Lee et al., 2018). Activating feedforward and feedback projections to PV/VIP cells induces the disinhibited state while activating lateral projections to SST cells induces the inhibited state. These states reflect the dynamic network computations performed by each cell type across diverse cognitive processes. A full experimental validation of such a switching motif, requiring simultaneous recording and identification of multiple inhibitory subtypes, has not yet been performed in vivo. However, this will likely become feasible in the near future using high-density recordings combined with tagging or all-optical approaches.

7. Limitations and future directions

Methodological constraints in NHPs have limited both precise one-to-one mapping between tTypes and eTypes and cross-species comparisons of interneuron functions in the PFC. Although recent technical and methodological advances have led to a growing number of NHP studies that acknowledge the importance of cell types, a long road still lies ahead. These limitations in studying interneuron functions in the PFC pose challenges for cross-species comparisons, particularly due to task alignment issues, for example, the dissimilarity between attention and working memory tasks used in primates versus rodents. Molecular confirmation in primates is also limited: so far, only indirect in-vivo studies or a few ex-vivo investigations have confirmed N-eTypes as GABAergic, (Banaie Boroujeni et al., 2021; Torres-Gomez et al., 2020) unlike the precise genetic targeting achievable in rodents. Similarly, primate research still lags behind in establishing causal inferences to the same degree as in rodents, largely due to the lack of optogenetic tools targeting molecularly defined types. Instead, it relies more heavily on correlational evidence derived from waveform-based classification (Box 2).

Looking ahead, several emerging and promising directions will need to be integrated and adopted by the NHP neuroscience community to overcome current limitations. One key approach involves the continued development of viral vectors targeting interneuron-specific enhancers, which can refine genetic access in primates without relying on transgenic lines (Vormstein-Schneider et al., 2020; Merlin and Vidyasagar, 2023). In addition, NHP research will greatly benefit from multimodal integration, combining electrophysiology with spatial transcriptomics and large-scale network dynamics, to connect molecular, cellular, and systems levels of interneuron-specific function within a unified framework (Bakken, 2021; Banaie Boroujeni et al., 2021, 2025). Importantly, future work will need to establish clearer links of computational motifs (Fig. 5) with empirical findings, as well as determine the bidirectional relationships of cognitive functions with the connectivity patterns and dynamics of these circuit motifs. Such advances will support computational modeling and theoretical frameworks that link cognitive functions to circuit motifs contributing to gain modulation, input gating, and disinhibition, with explicit mapping to experimentally defined interneuron functions, and further facilitate cross-species comparisons beyond task-specific paradigms (Box 3). Together, progress along these directions will advance our understanding of how circuit dynamics give rise to cognitive functions in NHPs and reveal how functional motifs diverge or converge across species, ultimately helping us understand both conserved and primate-specific principles by which interneurons support cognitive flexibility.

8. Conclusion

We surveyed the diversity of interneurons across PFC (Fig. 1 Box 1), emphasized the importance of electrophysiological signatures when distinguishing interneuron types in NHP PFC (Fig. 2Box 2) and inferred their functional contributions in attention control, working memory, and adaptive behaviors (Fig. 3). Interneurons in NHP PFC are more likely disinhibitory CR/VIP neurons compared with other areas and rodent mPFC, suggesting an increased role of disinhibitory motifs in NHP PFC (Figs. 1-2). Testing interneuron specific functions in primates so far have relied on distinguishing N-eType neurons containing a diverse group of PV, SST and VIP interneurons. Despite this heterogeneity it has become apparent that N-eTypes have a unique functional activation signature, encode behaviorally relevant cues, impose a broader inhibition on local circuits during working memory delays, and signal prediction error and outcome information in the lateral and medial PFC of NHPs (Fig. 3). These functional correlates of N-eType activity in primates are consistent with the functional activation of PV, SST, and VIP interneurons in rodent mPFC and ACC, suggesting neuron-type-specific functional contributions. Based on optogenetic evidence in mPFC, PV neurons are closely associated with attention and learning functions and less with memory-guided processes, SST neurons contribute prominently to memory-guided processes, while VIP neurons are apparently linked to valence-specific outcome processes (Fig. 4). Computational studies provide mechanistic insights on how these functional activation signatures emerge from interneuron specific connectivity in circuit motifs (Fig. 5). Circuit simulations suggest that PV interneurons are essential for balancing network inhibition/excitation level that is crucial for flexible learning and attentional control, SST interneurons for selective gating and retention of information, and VIP interneurons for integrating multiple sources of top-down and bottom-up information; all of which occurs in circuit motifs with complex interconnectivity between interneuron types.

While a direct mapping of t- and e-types in primate PFC is still not clear, our review tried to bring together functional and electrophysiological evidence to propose a roadmap for future research to better draw this connection. Our review finds tentative correspondences between primate eTypes and rodent t-Types which need future investigation to be better understood. Fast-spiking N-eType neurons with regular firing patterns and strong gamma synchronization likely correspond predominantly to PV+ interneurons, as both show similar functional signatures during attention control and learning tasks (Figs. 3, 4) and share narrow spike waveforms and high firing rates (Banaie Boroujeni et al., 2021; Torres-Gomez et al., 2020; McCormick et al., 1985; Hu et al., 2014; Kawaguchi, 1995; Kim et al., 2016a). The subset of N-eType neurons with broader tuning during working memory delays may correspond to SST+ interneurons, as both groups contribute to spatial working memory and stimulus retention (Rao et al., 1999; Diester and Nieder, 2008; Hussar and Pasternak, 2009; Kim et al., 2016b; Abbas et al., 2018). VIP/CR+ interneurons, which constitute the largest proportion of interneurons in primate PFC (Dombrowski et al., 2001; Krienen et al., 2020; Džaja et al., 2014), likely include both N-eType and B-eType neurons given their heterogeneous electrophysiological profiles (Torres-Gomez et al., 2020), which may explain the enhanced valence coding observed for outcome monitoring in both primate N-eTypes (Banaie Boroujeni et al., 2021; Shen et al., 2015) and rodent VIP+ neurons (Kamigaki and Dan, 2017; Pinto and Dan, 2015; Szadai et al., 2022). Future work should develop multimodal characterization approaches that directly link eTypes to molecular identities in primates, particularly through post-hoc immunohistochemistry of recorded neurons or by combining in-vivo recordings with transcriptomic profiling and optical tagging of neuron types (see Supplementary Box 1).

In summary, what has been lacking in our understanding of NHP PFC interneurons is an approach that characterizes eTypes, their temporal activation dynamics and functional recruitment in different processing states, and their mechanistic role in circuit motifs. We believe that such a comprehensive approach will be needed to advance our understanding of interneuron specific functions in the primate PFC and how they compare to those of rodent mPFC (Box 3). Such a multi-level approach to the study of interneurons will become more urgent with the increased use of optogenetics tools to interrogate primate PFC neuron types (see Supplementary Box 1). The rich set of studies we surveyed across species are a versatile starting point for identifying interneuron specific functions in PFC, to ultimately uncover the cell- and circuit-level principles underlying adaptive goal-directed behavior in the human brain.

Appendix A.Supporting information

Supplementary data associated with this article can be found in the online version at doi:10.1016/j.pneurobio.2025.102873.