Interoceptive inference and prediction in food-related disorders

In this Outlook, Subramanian and Thaiss discuss how the brain integrates multisensory (interoceptive and exteroceptive) signals to control feeding behaviors and disorders.

Abstract

The brain's capacity to predict and anticipate changes in internal and external environments is fundamental to initiating efficient adaptive responses, behaviors, and reflexes that minimize disruptions to physiology. In the context of feeding control, the brain predicts and anticipates responses to the consumption of dietary substances, thus driving adaptive behaviors in the form of food choices, physiological preparation for meals, and engagement of defensive mechanisms. Here, we provide an integrative perspective on the multisensory computation between exteroceptive and interoceptive cues that guides feeding strategy and may result in food-related disorders.

Multisensory integration of external and internal cues by the brain

The brain functions as a Bayesian estimator of external (Knill and Pouget 2004; Doya 2007; Chater et al. 2020; Yon and Frith 2021) and internal (Khalsa et al. 2009; Berntson and Khalsa 2021) environments. The integration of multiple exteroceptive and interoceptive cues providing nonredundant information improves perceptual accuracy, minimizes sensory uncertainty, and consequently increases the physiological salience of events (Stein and Meredith 1993). As such, multisensory computations can reduce response time, increase response magnitude due to enhanced salience, and improve the adaptiveness of behavioral responses through increased perceptual accuracy (Stein and Meredith 1993; Nozawa et al. 1994; Ernst and Banks 2002; Calvert et al. 2004; Friston et al. 2006; Bar 2007; Friston and Stephan 2007; Gillmeister and Eimer 2007; Bubic et al. 2010).

Beyond augmenting event salience, multisensory computations facilitate learning of associations across sensory modalities, enabling predictive control (Talsma 2015; Tong et al. 2020; Lauzon et al. 2022). Importantly, cross-modality associative learning can occur on timescales ranging from seconds (e.g., Pavlovian conditioning) (Pavlov 1927) to minutes (e.g., flavor–nutrient conditioning) and is partially dependent on dopaminergic subsystems (Grove et al. 2022). The increased efficiency conferred by multisensory integration is crucial for maximizing host fitness during environmental interactions through (1) accurate event detection, (2) faster and/or stronger responses to high-salience events, and (3) predictive control through associative learning.

Multisensory integration in response to nutritional cues

A fundamental requirement for all animals is the acquisition of nutrients through the selection of food from multiple options in their environment (Simpson and Raubenheimer 2012). Animals efficiently attribute relative food quality to inform adaptive behaviors within a food environment in a manner that is tunable based on the host's physiological state. For instance, animals deprived of specific nutrients (both macronutrients and micronutrients) can rapidly select foods supplemented with those nutrients (Berridge et al. 1984; Beauchamp et al. 1990; White et al. 1992; Wallis de Vries and Schippers 1994; Provenza et al. 1998; Coelho et al. 2006; Chaumontet et al. 2018; Pedernera et al. 2021; Malita et al. 2022; Münch et al. 2022). Similarly, allergen-sensitized mice tend to avoid solutions containing food allergens (Florsheim et al. 2023; Plum et al. 2023).

The ability to modulate behaviors based on anticipated postingestive consequences necessitates a set of priors that inform associations between postingestive consequences and preingestive cues. This heuristic has been well described through flavor–nutrient conditioning paradigms, which have provided substantial evidence that associations between flavor and postingestive value of ingested substances, established by pairing oral flavor stimuli with intragastric nutrient infusions, can drive preference for the associated flavor (Capaldi 1992). These cross-modal associations can be formed between any exteroceptive cue (gustation, olfaction, vision, tactioception, or audition) or combination thereof, paired with postingestive consequences (blood glucose levels, inflammatory responses, serum metabolites, serum electrolytes, hormonal changes, and gastric signals to name a few) detected by interoceptive sensory mechanisms. The utility of these associations lies in their capacity to predict impending changes in host physiology during feeding. Anticipatory responses to food consumption, known as cephalic responses, were first described by Pavlov (1927), who observed that sensory cues paired with food intake could elicit secretion of pancreatic enzymes, saliva, and gastric acid. Subsequent research has elucidated anticipatory responses in key metabolic hormones, including insulin, glucagon, gastric inhibitory polypeptide, ghrelin, and pancreatic polypeptide, in response to food-related sensory cues (Schwartz et al. 1979; Simon et al. 1986; Yamazaki and Sakaguchi 1986; Goldschmiedt et al. 1990; Wøjdemann et al. 2000; Simonian et al. 2005; Frecka and Mattes 2008; Monteleone et al. 2008; Teff 2010; Schüssler et al. 2012; Zhu et al. 2014; Rigamonti et al. 2015; Dhillon et al. 2017). Notably, the magnitude of cephalic responses can be modulated based on the nutritional/energy content, the food form, and the sensory modality conveying information about the incoming meal. The remarkably fine predictive control of physiological responses to meals aids in enabling the host to accommodate changes to their internal environment.

Equally critical to an animal's food strategy is the avoidance of toxins, pathogens, and other potentially harmful exposures. Adaptive responses to such insults, which rely on the association between exteroceptive cues and postingestive consequences detected by interoceptive sensory machinery, have been extensively documented. For instance, conditioned aversions can develop for food components strongly associated with adverse postingestive effects (Reilly and Schachtman 2008). Furthermore, sensory associations can promote immunological activation, manifesting as mast cell degranulation (Russell et al. 1984; MacQueen et al. 1989), antibody production (Husband et al. 1993; Madden et al. 2001), cytokine release (Cohen et al. 2019), and immune cell infiltration (Cohen et al. 2019) in anticipation of potential insults. Although there is a paucity of literature on anticipatory immune reactions during feeding, the ability to form conditioned associations between exteroceptive sensory cues and peripheral immune responses hints that such responses are viable and underappreciated.

The predictive framework derived from multisensory computations is critical to guiding efficient host food strategy by enabling accurate appraisal of nutrient content and quality and, conversely, the harmful components of available food. The framework is important to efficiently (1) make choices that have the most adaptive value to the host depending on the homeostatic needs, (2) kick-start physiological responses to an incoming meal (cephalic responses) that may or may not be tuned to the predicted nutrient composition, and (3) engage in anticipatory defenses against potential noxious components of food.

Consequences of predictive errors in food cue evaluation

The predictive system described above involves sensing of internal and external environments to inform food choices and mount appropriate physiological responses in anticipation of meals with distinct nutrient compositions and associated costs. Two primary sources of potential errors in this system are (1) conflicting sensory information about an event and (2) prediction errors resulting in inappropriate physiological responses.

The brain must parse a constant stream of sensory information to deconvolute components relevant to events and create a uniform perceptual experience. The reliability of different cues in providing relevant information is based on priors and is context-dependent. Established priors help facilitate navigation of a multidimensional sensory landscape. In the framework described, exteroceptive cues conveying sensory information about food and the postingestive response detected by interoceptive sensory mechanisms collectively form a unified representation of meal/nutrient consumption. We propose three mechanisms to resolve conflicting sensory information about the same event, aligning with established principles of multisensory computation (Fig. 1A).

Figure 1.

Schematic of proposed framework and sensory conflicts. (A) Exteroceptive input and interoceptive consequence together form a unified experience of food consumption. Context-dependent associations (a_c) form over time as the host navigates food environments. Sensory features inform on the salience of events with different reliability (w_e and w_i), which leads to physiological effector responses with particular magnitude and time (p_m,t). Importantly, exteroceptive cues can inhibit responses to interoceptive stimuli and vice versa. (B) Cross-modal illusion created when true sensory information (sweet taste) and sensory conflict (lack of nutritional sugar absorption when consuming artificial sweetener) result in intact effector responses (cephalic response and increased artificial sweetener consumption) as one modality is attributed greater reliability. (C) Sensory distraction (misaligned circadian rhythms) can hamper magnitude or time to effector response. (D) Maladaptive associations formed between two sensory cues (for example, the taste of peanut and mast cell degranulation) can cause maladaptive physiological responses.

First, in some scenarios, the brain can integrate conflicting sensory cues, potentially creating an illusion (Fig. 1B). For example, mice provided with a choice between sugar and artificial sweetener solutions consume them at equal rates for up to 24 h (Tan et al. 2020). Despite the lack of postingestive effects from artificial sweeteners, the sweet taste drives consumption. In humans, saccharin can induce robust cephalic insulin responses without affecting blood glucose levels (Dhillon et al. 2017). This dissociation between perceived sweetness and actual glucose availability in both examples illustrates the capacity of cross-modal sensory illusions to drive physiological responses in the absence of corresponding metabolic stimuli. Sham feeding experiments provide another illustrative example. Dogs with esophageal fistulas that drain consumed food still cease eating after chewing and swallowing despite the absence of nutrient absorption (Janowitz and Grossman 1949). The discrepancy between pregastric cues signaling food intake and the lack of nutrient absorption creates cross-modal illusions, promoting eating cessation. Crucially, both paradigms demonstrate rapid extinction of these illusory effects. The reliability of sensory cues undergoes dynamic updating, leading to adaptive behavioral modifications. In the case of artificial sweeteners, prolonged exposure results in an increased preference for sugar solutions, likely mediated by postingestive sensing mechanisms (Tan et al. 2020). Similarly, fistulated canine models exhibit increased food consumption over time, reflecting an adaptive response to the persistent absence of detected nutrients (Davis and Campbell 1973).

Second, in some instances, sensory cues may be ignored as distractions or interferences, impacting response magnitude or speed (Fig 1C). Circadian cycles provide anticipatory information regarding meal timing, and consumption patterns that deviate from these endogenous rhythms can be conceptualized as a form of sensory conflict. Studies on circadian rhythm misalignment, often observed in jet lag and shift work, have demonstrated impaired glucose tolerance and delayed insulin responses to meals, which contribute to increases in metabolic diseases like diabetes, obesity, and cardiovascular diseases (Spiegel et al. 2009; Donga et al. 2010; Pan et al. 2011; Wang et al. 2011). Similarly, disruptions of the core circadian clock genes Clock or Per2 in mice result in a loss of rhythmic food intake and increased incidence of metabolic diseases (Turek et al. 2005; Yang et al. 2009). Interoceptive sensory systems possess the ability to detect circadian rhythms through the perception of oscillating physiological parameters such as melatonin levels (Zisapel 2018), diurnal intestinal microbiota oscillations (Thaiss et al. 2014), core body temperature (Okamoto-Mizuno and Mizuno 2012), and rest–activity cycles (Brito and Thosar 2023). The integration of such cues, influenced by circadian cycles, can therefore play a role in meal anticipation. Sensory conflict that arises when meals are consumed out of sync with circadian cycles can therefore result in impaired ability to mount appropriate physiological responses to handle meals. Sensory-specific satiety serves as another example of a similar conflict resolution. In this paradigm, despite satiation signals, experience of a new food through exteroceptive senses can promote unregulated overconsumption, particularly amidst the availability of highly palatable foods (Wisniewski et al. 1992).

Third, the formation of maladaptive associations between disparate sensory modalities can lead to aberrant physiological responses, as exemplified by food allergies, which represent a spurious association between dietary nutrients and the engagement of host defense mechanisms (Fig. 1D). Importantly, the involvement of a cephalic component to dietary ingestion in this context is exemplified by the phenomenon of psychosomatic food allergies. In one case report, a patient experienced food allergy symptoms after peanut exposure despite negative skin prick and IgE tests (Kelso et al. 2003). A double-blind food challenge that aimed to block any sensory recognition of a peanut yielded no physical reaction. Upon cognitive reappraisal of this information, the patient's allergic responses to peanuts ceased almost immediately, suggesting a robust psychoneuroimmunological interaction (Kelso et al. 2003). In this instance, sensory cues associated with peanuts were nonspecifically linked to features of an allergic response. Sensing of the allergens therefore can induce anticipatory immune responses that can present clinically as an allergic response. The true prevalence of psychosomatic allergies is difficult to quantify due to the prevalent use of open food challenges—unblinded challenges with allergen—in clinical food allergy diagnosis, which represents a departure from the gold standard test for food allergy diagnosis: a double-blind, placebo-controlled food challenge. In one study, only 60% (28 of 46) of patients with a positive open food challenge remained positive when subjected to a double-blind, placebo-controlled food challenge, suggesting that psychosomatic sources of food allergy may be more prevalent than originally estimated (Venter et al. 2007).

Conclusion

Plasticity in multisensory integration and associative learning paradigms can be elucidated through the lens of predictive coding and Bayesian inference models of brain function. The examples delineated above exemplify the recalibration of sensory cue reliability and the consequent adaptive behavioral modifications aimed at minimizing prediction errors arising from sensory association predictions. The processes of association unlearning and belief updating are fundamental to maintaining cognitive homeostasis in the face of dynamic interoceptive and exteroceptive environmental fluctuations. The stability of associative learning is a function incorporating variables such as the magnitude of postingestive consequences, stimulus novelty, event frequency, evolutionary adaptive significance, and contextual environmental factors. Consequently, the plasticity of belief updating and associative unlearning exhibits significant variability depending on the nature and strength of the initial associative encoding.

The global prevalence of food-related disorders has been largely attributed to a mismatch between the ancestral environments that shaped our evolutionary adaptations and the contemporary nutritional landscape. We hypothesize that intragenerational environmental transitions, as exemplified by migratory phenomena, can precipitate maladaptive responses when previously formed associative networks demonstrate diminished adaptive utility in novel food environments. Epidemiological investigations have revealed a significantly higher incidence and earlier onset of diabetes mellitus in migrant populations compared with indigenous cohorts (Agyemang and van den Born 2019). This observation can be interpreted within a predictive coding framework in which a discordance between the original environment (in which critical associations between preingestive cues and postingestive consequences were established) and the current environment may drive metabolic and immunological dysregulation. This perturbation persists until the associative networks are updated and new adaptive associations are formed. Drawing on discussions that took place at Cold Spring Harbor Laboratory's 88th Symposium, dedicated to the topic of Brain Body Physiology, a critical future direction for the field will be to decipher the biological principles underlying plasticity in exteroceptive and interoceptive inference and prediction, with the goal of understanding the molecular etiology of food-related disorders.