Cross-modal plasticity refers to changes in brain function following a sensory loss that allows the spared senses to be used more effectively to guide behavior. There are two main changes: cross-modal recruitment of the brain areas that serve the lost sensory modality and compensatory plasticity of the brain areas that process the spared senses (Figure 1). Cross-modal recruitment has been observed in blind subjects where braille reading activates the deprived visual cortex (Sadato et al., 1996; Buchel et al., 1998; Burton et al., 2002b) and in deaf subjects where visual stimuli activate the deprived auditory cortex (Sandmann et al., 2012). This functional cross-modal recruitment of the deprived cortices is thought to enhance the processing of the remaining senses by increasing the cortical territory. This idea stems from the notion that cortical circuits are functionally equivalent and can process any sensory information presented to them. One of the supporting evidence for the functional equivalence of cortical circuits comes from a study, which rewired visual inputs to the primary auditory cortex (A1) of ferrets during early development causing visually guided behavior to become dependent on A1 (von Melchner et al., 2000). However, primary sensory cortices can be quite specialized in their anatomical organization, for example, barrel cortex in rodents and ocular dominance columns present in primary visual cortex (V1) of some carnivores and primates. Whether such anatomical specializations may affect the functional equivalence of cortical processing is unclear. In addition to the cross-modal recruitment, compensatory plasticity of the spared sensory areas is thought to allow refinement and increase the sensitivity of the spared sensory systems (Pascual-Leone and Torres, 1993; Sterr et al., 1998a,b; Roder et al., 1999; Elbert et al., 2002).

In animal models, early sensory loss results in cross-modal anatomical changes to the primary sensory cortices, especially of thalamocortical inputs (Henschke et al., 2018; Dooley and Krubitzer, 2019), but in adults, most of the cross-modal plasticity likely occurs through functional plasticity of pre-existing circuits (Lee and Whitt, 2015; Ewall et al., 2021). Cross-modal functional plasticity is observed in both the deprived and the spared primary sensory cortices (Figure 1). Evidence from animal studies suggests that these involve experience-dependent plasticity mechanisms based on Hebbian, both long-term potentiation (LTP) and long-term depression (LTD), and homeostatic mechanisms (Lee and Whitt, 2015; Ewall et al., 2021). In this mini-review, I will summarize the synaptic plasticity mechanisms of adult cross-modal plasticity and explain how utilizing the framework of the metaplasticity model can easily account for the global cortical adaptation in adults. To do this, I will first provide a brief introduction to the sliding threshold metaplasticity model.

Metaplasticity refers to the regulation of synaptic plasticity (Abraham and Bear, 1996) and often refers to the sliding threshold model proposed by Bienenstock, Cooper, and Monroe (BCM theory); a theoretical model that can provide stability to Hebbian plasticity (Bienenstock et al., 1982; Bear et al., 1987; Cooper and Bear, 2012). It was recognized that synaptic plasticity solely based on Hebbian mechanisms is limited and is unable to provide network stability that is necessary for neural function on its own. This is due to the fact that correlation based synaptic plasticity mechanisms, such as LTP and LTD, have in-built positive feedback. Strengthening synapse by LTP causes postsynaptic neurons to respond to inputs more strongly, which increases the coincidence of pre- and post-synaptic activity. This in turn promotes further LTP leading to over-excitability when left unchecked. Similar positive feedback occurs upon weakening synapses via LTD, but in the opposite direction that ultimately leads to inactivity. Such bistable property of Hebbian plasticity on neural network activity suggests a need for homeostatic control to provide stability. The sliding threshold model allows such homeostatic control by postulating that the induction threshold for LTP and LTD, referred to as the synaptic modification threshold (θm), slides as a function of the integrated past activity of a neuron (Figure 2). This key property of the sliding threshold allows neurons to dynamically tune their feature selectivity to the dominant input at the moment as a function of the history of overall activity (Bienenstock et al., 1982; Bear et al., 1987; Cooper and Bear, 2012). Here I will discuss the two main properties of the BCM model: the generation of feature selectivity and its dynamic regulation by past activity that endows homeostasis.

One of the main features of the sliding threshold model is the development of feature selectivity by LTP and LTD. Stronger inputs that exceed the θm undergo LTP, while weaker inputs that fail to produce sufficient postsynaptic activity are depressed via LTD. Thus neurons “select” the dominant input, at the expense of other weaker inputs. The ability of sensory cortical neurons to express feature selectivity is critical for sensory processing by enabling discrimination of distinct inputs to generate a percept of a certain feature in the sensory environment. While the initial setup of feature selectivity arises from the developmental organization of axonal projections based on guidance cues (Crowley and Katz, 1999, 2000), the refinement of feature selectivity is known to depend on the activity and in particular sensory experience (Sengpiel and Kind, 2002; Hooks and Chen, 2020). An essential feature of the sliding threshold is its ability to provide network stability, which is endowed by the regulation of θm as a function of the integrated past activity. If the neuron has been under high activity regime for a duration of time, θm increases (i.e., slides up) to make LTP more difficult to induce and promote LTD across the majority of synapses. Weakening the majority of the inputs via LTD causes the overall postsynaptic activity to decrease, hence providing homeostatic control of neural activity. In contrast, if the postsynaptic neuronal activity is low, θm is reduced to promote LTP and decreases the range of activity that produces LTD. While stabilizing the neural activity, the sliding threshold model preserves feature selectivity because stronger inputs will surpass θm to strengthen their connections at the expense of weaker inputs whether the θm is high or low. Implementing the sliding threshold to computational neural networks successfully explains the development of feature selectivity based on LTP/LTD while maintaining network stability (Bienenstock et al., 1982). As will be discussed later sliding threshold also enables dynamic regulation of the neuronal feature selectivity when there are changes to the type of inputs available to the neurons (Figure 2). The sliding threshold model was first experimentally demonstrated in V1, where reducing overall neural activity by dark rearing from birth promotes LTP and reduces LTD (Kirkwood et al., 1996). Since then, it has been confirmed across many brain regions (Lebel et al., 2001; Krucker et al., 2002; Solger et al., 2004; Muller et al., 2007; Narayanan and Johnston, 2010; Hulme et al., 2012) and malfunctioning of the sliding threshold has been reported in various neurological disease models (Hulme et al., 2013; Megill et al., 2015; Jang and Chung, 2016) suggesting its critical role in normal brain function.

At a molecular level, the sliding threshold is implemented by alterations in either the induction or the expression mechanisms of LTP/LTD. The initial postulate of the sliding threshold model stated a horizontal shift of the synaptic modification function (Figure 2), but now there is evidence that the synaptic modification function could shift vertically (Seol et al., 2007; Huang et al., 2012; Hong et al., 2020). Both the horizontal and vertical shift in the synaptic modification function slides the θm in the same manner, but the molecular mechanisms underlying the two are distinct. The horizontal shift in synaptic modification occurs by changes in the induction mechanisms of LTP/D, which in molecular terms has been demonstrated as changes in N-methyl-D-aspartate (NMDA) receptor (NMDAR) subunit composition. Low neuronal activity upregulates the expression of NMDAR GluN2B (or NR2B) subunit (Quinlan et al., 1999; Philpot et al., 2001, 2003), which confers longer current duration compared to GluN2A (or NR2A) containing NMDARs (Chen et al., 1999; Rumbaugh and Vicini, 1999). This in essence allows larger intracellular Ca²⁺ increase through NMDARs to promote LTP. With heightened neuronal activity, GluN2A becomes the dominant subunit for NMDARs, which increases the θm to make LTP induction more stringent. In contrast, the vertical shift in synaptic modification (also referred to as the Pull-Push model) has been shown to occur with neuromodulators acting on LTP/LTD expression mechanisms (Seol et al., 2007; Huang et al., 2012; Hong et al., 2020; Mihalas et al., 2021). Neuromodulators linked to cyclic adenosine 3′,5′-monophosphate (cAMP) signaling promote LTP by increasing phosphorylation of α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid receptors (AMPARs), especially on the protein kinase A target serine-845 (S845) residue of the GluA1 subunit to promote LTP (Seol et al., 2007; Qian et al., 2012). GluA1-S845 phosphorylation has been shown to increase the content of cell surface AMPARs (Oh et al., 2006; He et al., 2011) and prime AMPARs for synaptic recruitment by LTP (Lee and Kirkwood, 2011; Diering and Huganir, 2018). In contrast, neuromodulators linked to phospholipase C (PLC) slide up the θm by promoting LTD and preventing LTP expression (Huang et al., 2012). The main difference between the horizontal and vertical shift in synaptic modification function is that the latter puts synapses into an LTP-only or an LTD-only mode while the former preserves the full range of LTP and LTD (Figure 2).

Next, I will review the synaptic mechanisms of cross-modal plasticity in the adult primary sensory cortices and discuss how the horizontal shift in synaptic modification function could account for the observed findings.

In the adult primary sensory cortices, the superficial layers (L2/3) are likely a major substrate for adult cross-modal plasticity. L2/3 synapses retain their ability to undergo LTP and LTD as well as homeostatic synaptic plasticity into adulthood (Goel and Lee, 2007; Jiang et al., 2007). This contrasts plasticity in L4, where thalamocortical LTP/LTD and homeostatic synaptic plasticity display a short early critical period for plasticity (Crair and Malenka, 1995; Feldman et al., 1998; Desai et al., 2002; Jiang et al., 2007; Barkat et al., 2011; Rodriguez et al., 2018). Moreover, the functional connectivity of the L2/3 neurons is well poised to integrate top-down contextual information with the bottom-up sensory responses from thalamorecipient neurons in L4. Besides feedforward inputs from L4, L2/3 principal neurons receive long-range inputs from multisensory association cortices and higher order sensory cortices, direct connections from other primary sensory cortices, as well as inputs from higher order thalamic nuclei that carry multisensory information (Ewall et al., 2021).

The sliding threshold model is compatible with experimental observations of synaptic plasticity related to both cross-modal recruitment and compensatory plasticity. Viewing cross-modal plasticity in the framework of sliding threshold presents testable predictions. Because the synaptic plasticity is ultimately driven by NMDAR-dependent Hebbian modification in accordance with the newly defined synaptic modification threshold, it suggests that cross-modal plasticity will be dependent on sensory experience in the spared modalities. Such requirement of experience may explain the varied observations of the outcome of cross-modal plasticity in human subjects, especially when a sensory modality is lost later in life (Frasnelli et al., 2011; Lazzouni and Lepore, 2014). Furthermore, observations that pre-existing functional connections across the sensory cortices remain plasticity even in adults suggest that cross-modal plasticity mechanisms could be used for enhancing adult brain function.

A1, primary auditory cortex; AMPAR, α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid receptor; cAMP, cyclic adenosine 3′,5′-monophosphate; dLGN, dorsal lateral geniculate nucleus; GluN2A, glutamate ionotropic receptor NMDA type 2A; GluN2B, glutamate ionotropic receptor NMDA type subunit 2B; L2/3, layer 2/3; L4, layer 4; L5, layer 5; LTD, long-term depression; LTP, long-term potentiation; mEPSCs, miniature excitatory postsynaptic currents; MGBv, ventral portion of medial geniculate body; NMDAR, N-methyl-D-aspartate (NMDA) receptor; NR2A, NMDAR 2A subunit; NR2B, NMDAR 2B subunit; PLC, phospholipase C; TRN, thalamic reticular nucleus; TTX, tetrodotoxin; V1, primary visual cortex; θm, synaptic modification threshold.

H-KL synthesized and conceptualized the ideas and wrote the manuscript.