Advancing computational psychiatry through a social lens

Social factors have an outsized role in the progression and treatment of psychiatric conditions, but existing computational models often neglect social context. We propose that integrating social features into computational frameworks will shed light into the complexity of mental health, and provide guidelines for model design.

Humans have the most complex social networks among animal species. Evolutionarily, such complexity has had a major influence on the structure and function of our brain, and consequently, on our behavior and mental health. Social factors — including the nature of interpersonal interactions or relationships to broader societal matters such as inequality and discrimination — overwhelmingly affect the onset and trajectory of neuropsychiatric conditions^1,2. Beyond well-known examples such as autism and social anxiety, socially relevant symptoms are seen in many other conditions including major depression (for example, social withdrawal) and schizophrenia. Social factors, including family and peer support, therapeutic alliance between the client and clinician, and interpersonally oriented elements during therapy, are also crucial during recovery^3,4. The importance of social cognition and behaviors in psychiatry is undeniable.

The field of social psychiatry has long recognized the importance of social context in the emergence, progression and treatment of psychiatric conditions^5,6. Despite this recognition, social features are not widely examined in computational psychiatry, leaving an enormous gap underscored in recent work demonstrating that many mainstream reinforcement learning and decision-making tasks have small or inconclusive effects in patient studies⁷. Here, we identify two interrelated sources of this issue: first, a lack of computational theory that links social cognition to psychiatric disorders and symptoms; and second, more notably, an overreliance on popular but insufficient computational models with low ecological or clinical validity. We propose, as have others³, that advancing computational psychiatry through a social lens — one that incorporates how humans interact with others as well as how they navigate complex social environments — will create greater opportunities to bridge the gap between biological mechanisms and clinical phenomenology.

From identifying mechanisms to biomarker discovery, computational psychiatry aims to characterize mental disorders using theoretical and predictive models that span several levels of explanation⁸, including biophysical and cognitive levels. Biophysical models characterize neurophysiological processes, including molecular mechanisms, firing patterns of single neurons, and circuitry-level dynamics. In parallel, cognitive computational models unpack building blocks of behavior, often relying on reinforcement learning and Bayesian approaches. Biophysical and cognitive models are theory-driven, as they typically require careful consideration of hypotheses and experimental design. Furthermore, data-driven models are increasingly popular owing to their utility in generating predictions (for example, about treatments). Reinforcement learning models are particularly influential as they target behaviors impaired across psychiatric conditions, such as rigid and repetitive behaviors in obsessive–compulsive disorders, insensitivity to rewards in anhedonia and depression, maladaptive drug-seeking in addiction, or aversion to uncertainty in anxiety. Despite theoretical predictions, empirical evidence to support learning deficits in human patients remains inconclusive⁷, raising questions about the ecological, clinical and translational relevance.

Social behavior and computational psychiatry

We propose that considering social context is one way to enhance the ecological and clinical validity of tasks and computational models, among other options like the use of more naturalistic measures. The current scope of computational psychiatry is dominated by topics like reinforcement learning, with only around 5–10% of publications focusing on interpersonal dynamics (as identified by a PubMed search in August 2024). Why have most researchers focused on non-social processes so far? On one hand, it is important to operationalize highly complex behaviors and symptoms into plain, replicable models that can probe context-invariant biophysical and cognitive mechanisms⁹. However, corresponding computational models are often highly simplified, aimed at achieving parsimony. Such approaches may fail to capture clinical symptoms or reliably distinguish affected from non-affected individuals. For example, individuals with major depression have been shown to respond to monetary rewards in experimental settings in ways comparable to individuals without depression⁷. This indicates that an oversimplified reinforcement learning approach might not be sufficient to probe specific symptoms, such as feelings of worthlessness or guilt, which are often socially rooted and concern how the self is compared with others (for example, “I am worse than everybody else”). Similarly, individuals with anxiety can perform probabilistic learning tasks reasonably well, yet core clinical symptoms such as excessive social concerns about self-image (for example, “I constantly worry about how others perceive me”) are not examined by these tasks. These clinically relevant experiences are emergent from cognitive ‘maps’ of the self and others and shaped by interpersonal dynamics, societal structures and cultural norms. To fully capture these complexities, we must move beyond basic biophysical and reinforcement learning mechanisms and quantify specific structural features relevant to social cognition. These social features, such as higher uncertainty related to distorted self-image and beliefs about other people, are crucial to understand how social context constrains maps of the world and shapes mental health. In recent years, funders (for example, the National Institutes of Health (NIH) and the Wellcome Trust) have realized this gap and started to curate specific funding mechanisms to study social behaviors relevant to psychiatric conditions.

Even for existing studies considering social context, there are still insufficient computational frameworks. For example, social stressor paradigms (such as social defeat or isolation) are frequently used in animal research to study aspects of depression or schizophrenia^10,11. Although these studies are often descriptive, there is a current movement towards incorporating social context into mechanistic preclinical animal research¹¹. Computational theories can link these empirical findings in principled ways. In human research, social deficits are often probed by paradigms such as face perception, making it challenging to rely on ‘off-the-shelf’ computational models to delineate the mechanisms that contribute to psychiatric symptoms. Although some research into autism and personality disorders has incorporated social observational learning paradigms, a key remaining issue is that existing computational models often do not directly account for or integrate social-specific features in model architecture. For example, one may use the same reinforcement learning model framework to account for decision-making under social and non-social conditions and examine differences in parameter estimates (such as learning rates). Although this approach is important, it largely assesses shallow representations of others and contextual effects instead of social-specific cognitive mechanisms such as theory of mind or inferring others’ aversion to unfairness. In other words, standard computational models are necessary, but not sufficient, to explain behaviors in humans with or without psychiatric symptoms.

Modeling social features in computational psychiatry

To improve inferences in computational psychiatry, paradigms and models should match the level of explanation more closely. Recent research has included these ideas by modeling computational mechanisms that underlie specific social cognitive constructs — such as social controllability¹² and theory of mind¹³ — and their dysfunction in psychiatric conditions. Importantly, these models address computation at the level of social representation¹³ — that is, models that directly generate predictions about features that characterize our complex social environment, including predictions about others’ future states. These frameworks integrate work from social psychology, behavioral economics, and computational neuroscience to quantify beliefs held by the self — and how the self makes inference about others — to understand both healthy and disordered interactions, such as when and why individuals experience delusional ideations of social partners^12,14 and how interpersonal perceptions distort self-other beliefs¹⁵. Although these advancements introduce increased complexity of paradigm and model design, and potential challenges in operationalizing or interpreting abstract features, they effectively address current gaps by integrating complex social dynamics and broader context to characterize human interactions, while retaining core computational principles such as learning and uncertainty estimation. Considering this view presents a ripe opportunity for exploring and evaluating new psychiatric theories inspired by methods from other disciplines that have made notable headway in quantifying social cognition and behavior. For example, we can evaluate theories that involve how people generate maps of internal perceptions, emotions or beliefs of others using models inspired by behavioral economics; we can evaluate theories inspired from computational sociology and computer science involving how people map social networks, groups or societal structures, or how people map collective beliefs or norms using models inspired by evolutionary biology, ecology and sociology. The recent application of natural language models in clinical work also shows promise in quantifying emergent social features from real-life conversations, including inferring speakers’ intentions or reasoning about speakers’ language usage, which can help in analyzing and treating social communication difficulties in conditions such as autism and schizophrenia. While not necessarily providing mechanistic insights, language models can complement theory-driven models to discover markers relevant to psychiatric diagnosis and treatment. Together, these approaches would build on foundational models of cognitive and behavioral processes while firmly rooting both the self and other within formal psychiatric theory.

The mind is not an island, whether in health or disorder⁵. Experimental paradigms and computational models of social cognition could shed insights on the complex deficits in neuropsychiatric conditions. We call for the computational psychiatry field to incorporate this view by designing and testing models that characterize the social features rooted in mental health and disorders. We provide guidelines and considerations for accomplishing this goal (Fig. 1), which will enable computational psychiatry to advance toward a more complete suite of theories that do not neglect social context, returning to its roots and embracing the core of what it is to be human^5,6.

Fig. 1: Guidelines and considerations when modeling social features rooted in mental health and disorder.

This schematic provides steps and considerations when designing a study to probe socially relevant symptoms in psychiatric conditions using a computational psychiatry approach. First, identify a well-defined socially relevant problem or symptom observed in a psychiatric disorder. Consider whether intentional mental content (for example, beliefs or values) are core or peripheral to the problem, whether the problem follows a straightforward time course or is characterized by divergent trajectories, and whether the problem can be characterized by a specific neurobiological mechanism⁹. Second, choose an experimental paradigm that captures core social aspects of the identified problem. Consider whether behaviors produce single or joint outcomes, include any necessary control experiments such as a non-social condition that controls for the presence and interactions with social agents, and think about the ecological validity of your paradigm. Third, select and validate computational models that generate specific predictions about healthy and disordered social cognition. Consider what the model architecture captures predictions about social cognitive and behavioral processes. Finally, apply the models to characterize empirical social behavior and cognition in the experimental paradigm.