According to the World Health Organization, suicide is a major public health concern, with more than 700,000 people dying by suicide every year (1), a number that has steadily risen over the past decade (2). Moreover, for each death by suicide, there are more than 20 suicide attempts (1). Individuals suffering from depression are 20-fold more likely to die by suicide compared to the general population (3). In major depressive disorder (MDD), the lifetime prevalence of suicide attempts is estimated to be 31% (4) with completion rates between 5 and 10% (5, 6). However, suicidality is a cross-diagnostic outcome that extends beyond depression alone. Suicidal ideation and attempts can occur in individuals with various psychiatric disorders, as well as in those without a formal psychiatric diagnosis (7). Differentiating between suicidal ideation and suicide attempts is crucial, as the former refers to thoughts and feelings related to suicide, while the latter refers to actual behaviors with potentially lethal consequences. The development of suicidal ideation and the progression from ideation to suicide attempts are distinct phenomena, and researchers have increasingly focused on identifying factors that distinguish between the two (8, 9).

Recent studies have shown that ketamine, an N-methyl-D-aspartate receptor (NMDAR) antagonist, significantly reduces suicidal thoughts and behaviors (STB) (10). Within hours, a single infusion of subanesthetic-dose ketamine relieves depressive symptoms (11, 12), including in treatment-resistant individuals (13, 14), and rapidly and significantly reduces suicidal ideation (10, 15–19). In comparison, treatment with first-line antidepressants, such as selective serotonin reuptake inhibitors (SSRIs), is associated with an improvement in depressive symptoms by the end of the first week, and may take up to 6 weeks to achieve maximum therapeutic effect (20). However, repeated ketamine administrations are required to prolong the drug’s effect, a response which typically abates within a week of a single infusion (21, 22). Notably, ketamine’s anti-suicidal effects may be distinguishable from its antidepressant effects (18, 23, 24). While antidepressant effects are crucial for overall mood improvement, the rapid anti-suicidal action of ketamine can prove critical for those in acute crisis and may hold promise for a broader spectrum of patients with STBs (24). This raises the possibility that ketamine may affect neurobiological pathways associated with STBs independent of depression, emphasizing the unique therapeutic potential of this drug (25).

The replication of ketamine’s rapid antidepressant and anti-suicidal effects has prompted detailed study into the drugs’ pharmacological mechanism of action. Ketamine preferentially acts via NMDAR antagonism giving rise to increased glutamate release from pyramidal cells with subsequent activation of α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid receptors (AMPAR). A cascade of downstream signaling pathways is then triggered leading to enhanced brain-derived neurotrophic factor (BDNF) and mammalian target of rapamycin (mTOR) signaling, increased protein synthesis, and synaptogenesis, which may, in turn, promote adaptive rewiring of pathological neurocircuitry (26) [see (27, 28) for detailed reviews of ketamine’s pharmacological profile]. However, ketamine’s effect is not restricted to glutamatergic pyramidal cells, but extends to other neurotransmitter systems, including serotonin, norepinephrine, and dopamine (29–33) as well non-monoaminergic systems (e.g., opioidergic and inflammatory) (14, 34).

Several clinical and neurobiological predictors of ketamine’s antidepressant and anti-suicidal effect have been identified (35–37), however, these data-driven approaches provide little mechanistic insight into how ketamine reduces STBs. This is partly due to the lack of mechanistic understanding of suicidality itself (38). Suicidality is a complex construct that involves multiple factors, such as emotion regulation, cognitive processing, and social and environmental stressors. Advancements in computational psychiatry, namely generative models, hold great potential in unraveling the complexities of suicidality and the mechanisms underlying ketamine’s anti-suicidal effects. Generative models explicitly describe the mechanisms underlying observed neural or behavioral data using biologically-interpretable variables (39, 40). By comparing model simulations against empirical data, the resulting discrepancies can be used to refine the model. This iterative process leads to the development of more informed theories of suicide and compuational models. By leveraging these computational tools, we can gain insight into the dynamic interactions between suicidality and ketamine’s anti-suicidal effects, potentially identifying biomarkers and treatment targets, and enhancing our understanding of the mechanisms involved.

In this paper, we review current computational theories of suicidality and ketamine’s mechanism of action, and discuss various computational modeling approaches being used to understand ketamine-induced changes in STBs. We close with a brief discussion on the implications of computational psychiatry for suicide prevention and address open-ended questions in ketamine therapy research.

Normative theories of learning and decision-making in computational neuroscience offer a theoretical framework for understanding optimal decision-making processes and can provide insight into the vulnerabilities associated with STB. While computational accounts of STB are only beginning to emerge, there is already a substantial body of literature on cognitive task performance that can inform such accounts.

Several studies have investigated impaired decision-making in individuals with STBs using the Iowa Gambling Task (IGT) and Cambridge Gambling Task (CGT). A meta-analysis of these studies showed an association between history of suicide attempts and riskier decisions compared to both patient controls and healthy controls (41, 42). However, the complexity of the IGT makes it challenging to attribute variations in task performance to specific cognitive processes. Dombrovski et al. (43) addressed this problem by separating choice processes from learning using reinforcement learning (RL) models and a three-armed bandit task. The authors showed that suicide attempters had impaired value comparison and diminished behavioral sensitivity to reinforcement, suggesting impaired reward learning, a deficit that scaled with suicide attempt lethality and was partially explained by poor cognitive control.

In addition to impairments in decision-making processes, decision-making biases may also play a role in STBs (44). Millner et al. (45) used a drift-diffusion model (DDM) and RL model to isolate two decision-making biases in an Avoid/Escape Go/No-Go task, namely active-escape bias (i.e., a bias to “do something” to escape) and an inhibitory-avoid bias (i.e., a bias to “do nothing” to escape). The authors showed that individuals with STBs exhibited a higher bias for active responses to escape an aversive state compared to controls, even after controlling for clinical variables such as depression and hopelessness. However, the inhibitory-avoid bias was comparable between groups, and no significant differences in bias parameters were found between suicide attempters and ideators (45). A recent study by Myers et al. (46), found that reduced inhibitory control, as measured by the Go/No-Go task, was predictive of a suicide attempt within 90 days, and that the miss rate in the same task was a stronger predictor of near-term suicide attempts compared to other commonly used measures. These preliminary findings suggest that decision-making biases assessed through cognitive tasks may hold clinical predictive value.

One overarching explanation of the above findings is an increase in Pavlovian vs. instrumental control (47, 48). Pavlovian control dictates reflexive behavior that rigidly specifies stimulus–response mappings regardless of outcomes. In contrast, instrumental control allows for the adaptation of behaviors to environmental contingencies to achieve desired outcomes in a goal-directed fashion. Increased Pavlovian control can therefore lead to more impulsive actions and biases that are not optimal in the long run. Karvelis and Diaconescu (49) used computational modeling to formally conceptualize the active-escape and Pavlovian biases in suicidality as a product of perturbations in probabilistic learning. According to the model, the Pavlovian active-escape bias and other suicide risk markers, including hopelessness and reduced cognitive control, may stem from the following four mechanistically distinct parameters that capture components of learning and stress responsiveness: increased stress sensitivity, increased learning from stressors, reduced sense of controllability of stressors, and a reduced ability to unlearn maladaptive beliefs. The model was validated by simulating performance in an Avoid/Escape Go/No-Go task, showing that altering each of the four parameters reproduces the findings of increased active-escape bias reported by Millner et al. (45). The authors proposed that these four mechanisms, which represent different hypotheses about the cognitive mechanisms underlying suicidality, may also correspond to distinct subtypes of suicidality. Although the model remains to be tested empirically, the authors’ proposed hypotheses on the connection between suicide neurobiology and cognition/behavior can inform potential mechanisms of action in treatments aimed at reducing STBs (we discuss this in more detail in Section 4.2).

Pavlovian biases may also explain increased loss aversion in STBs. Liu et al. (50) used the balloon analog risk task (BART) to investigate decision-making biases in MDD patients with and without suicide attempts. The study used Bayesian computational modeling to show that the suicide attempter group demonstrated stronger loss aversion than the non-attempter group and healthy controls. The authors posit that increased loss aversion, as a Pavlovian response, may prompt individuals to focus more on current painful experiences and may motivate suicide as a means of avoiding future psychological pain.

Collectively, these findings suggest that impaired value-based decision-making and cognitive control are important factors of suicidal vulnerability (51) [see (48) for a review on impaired decision processes in suicidal behavior], particularly in social and aversive contexts (52). Decision making processes in STBs may be influenced by an affective bias, whereby negative outcomes drive learning more than positive outcomes (53, 54). Suicidal ideation has been associated with a processing bias toward negative stimuli (55), while suicide attempts have been associated with blunted positive affective forecasting for future positive events (56). This may suggest that individuals with STBs fixate on negative outcomes and envision a future with predominantly negative events, and even in the presence of unexpected positive events, their predictions about the future are resistant to change. Over time, an excessive focus on and learning from negative outcomes may lead to a general updating bias toward negative information and result in the formation of overly precise negative prior beliefs (e.g., lower self-esteem, pessimistic worldview). These beliefs can exert excessive influence on one’s thoughts leading to a state of hopelessness—i.e., the belief that no actions can improve one’s situation—and in turn, may result in the adoption of maladaptive action policies, which are defined as strategies used by an agent to determine the next action based on what they have learned. For instance, a reduced sense of behavioral control may foster increased Pavlovian learning, leading to increased Pavlovian biases, where an agent continues to take actions that do not lead to the highest expected reward, such as actively escaping aversive events (45, 49), or avoiding loss (50). This bias is problematic as it can prevent the agent from exploring potentially more rewarding actions and thus hinder the learning of an optimal policy. In a suicidal crisis, vulnerable individuals may respond with increasingly stochastic choices due to impaired reward valuation (43), leading to a misestimation of the value of suicide over superior alternatives (48, 57). Utilizing computational models, one can generate novel hypotheses regarding impairments in learning and decision-making, which can be experimentally tested to develop a mechanistic understanding of the neurocognitive processes that underlie the progression of a suicidal crisis.

Ketamine has shown promising therapeutic potential in treating suicidal ideation, but its mechanism of action remains unclear. Marguilho et al. (58) recently proposed a unified model of ketamine’s dissociative and psychedelic properties, suggesting that its therapeutic effects are driven by acute modulation of reward circuits and a sub-acute increase in neuroplasticity. More specifically, ketamine may block NMDAR-dependent bursting activity of neurons in the lateral habenula (LHb), a basal ganglia nucleus known as the “anti-reward” center, resulting in the disinhibition of downstream monoaminergic reward centers. Moreover, the sub-acute increase in neuroplasticity, driven by sustained AMPAR activation in excitatory pyramidal neurons and potentiating BDNF and mTOR signaling, allows for sustained antidepressant effects. Additionally, the authors suggest that ketamine’s dissociative and psychedelic properties are driven by dose-and context-dependent disruption of the salience network (SN) and the default-mode network (DMN). The SN, composed of nodes that include the anterior cingulate cortex (ACC) and anterior insular cortex (AIC), helps prioritize relevant stimuli for decision-making and action, while the DMN, composed of nodes such as the posterior cingulate cortex (PCC), medial prefrontal cortex (mPFC), and inferior parietal lobule (IPL), is associated with self-referential thought. Furthermore, functional connectivity between the DMN and ACC has been shown as a key connection for explaining ketamine’s rapid antidepressant action (59). Computationally, the authors propose that nodes of the SN represent high-level priors about the body, and under low doses of ketamine, disintegration of the SN leads to relaxed priors about bodily self-experience thus accounting for ketamine’s ‘dissociative’ effects. At high ketamine doses, disintegration of the DMN leads to relaxed priors about narrative self-experience, accounting for ketamine’s ‘psychedelic’ effects.

The model proposed by Marguilho et al. (58) builds upon earlier predictive processing models describing the mechanisms underlying the therapeutic effects of classic serotonergic psychedelics. The Relaxed Beliefs Under Psychedelics (REBUS) model by Carhart-Harris and Friston (60) and self-binding model of psychedelic ego dissolution of Letheby and Gerrans (61) share the idea that serotonergic psychedelics weaken high-level priors, thereby creating an opportunity for belief revision. The REBUS model integrates theories of the entropic brain and free-energy principle within the framework of predictive coding, and proposes that classic psychedelics induce a heightened entropic brain state under high levels of serotonin 2A (5-HT2A) receptor agonism. This results in increased neural plasticity and a relaxation in the precision weighting of high-level priors or beliefs, possibly through reduced connectivity within the DMN. Belief relaxation, in turn, may enable increased sensitization to bottom-up signaling, rendering aberrant beliefs more amenable to revision (60). Letheby’s self-binding model of psychedelic ego dissolution on the other hand, focuses on ego dissolution by weakening the precision of maladaptive self-representations. The authors identify both the SN and DMN as crucial networks to self-representation, proposing that the SN plays a role with a more minimal or embodied sense of self while the DMN is implicated in higher-level narrative self-representation. Letheby posits that psychedelics reduce the brain’s confidence in expectations about reality and the Self, thereby decreasing their influence on phenomenal awareness (i.e., the subjective experience of the world and oneself), and driving therapeutic change through alterations in self-perception possibly through a reduction in the precision-weighting of high-level priors (61).

All three models converge on the idea that the therapeutic effects of these substances involve the relaxation or weakening of high-level priors and beliefs, which subsequently allows for the revision of entrenched maladaptive beliefs or self-representations (see Table 1 for summary of models). Ketamine may achieve this through a different mechanism than serotonergic psychedelics by directly affecting NMDAR and AMPAR signaling. Previous research has shown that glutamatergic NMDARs facilitate the communication of top-down predictive signals, while AMPARs communicate bottom-up prediction error signals (62–64), and the weighting of these prediction errors is driven by neuromodulators such as dopamine and acetylcholine (65–67). Ketamine’s unique mechanism of action, directly influencing NMDAR and AMPAR signaling, offers a distinct pathway for addressing its therapeutic impact on maladaptive cognitive processes associated with STBs (23).

Throughout the remainder of the paper, we attempt to explain ketamine’s anti-suicidal effect using an array of computational modeling approaches. Firstly, we explore its therapeutic potential in the context of the mismatch negativity and predictive coding framework. Secondly, we examine the impact of ketamine on the neurocircuits involved in STBs using Karvelis and Diaconescu’s model of suicidality (49). Finally, we discuss Dynamic Causal Modeling (DCM), a neurobiologically interpretable model that uses neuroimaging data to infer effective connectivity, such as forward and backward connection strengths, among brain regions (68–70). By studying these parameters, we can investigate the altered connectivity strengths and receptor densities implicated in STBs and targeted by pharmacological interventions. The above approaches make use of generative models, which allow for an explicit description of the underlying mechanisms that produce the data. As a result, they provide a detailed disease model and enhance our understanding of the anti-suicidal effects of ketamine.

While theory-driven computational models offer a promising approach for bridging existing knowledge of suicidality and ketamine’s mechanism of action, multiple questions remain.

In conclusion, while the underlying mechanisms of STBs and the anti-suicidal effect of ketamine are still not fully understood, computational models can provide valuable insights into the complex pharmacology of ketamine and its influence on suicidality. However, further studies are needed to optimize modeling approaches and task design, as well as evaluate external factors such as set and setting and the potential therapeutic value of psychedelic-assisted therapy. With continued advancements in computational psychiatry, a more comprehensive understanding of ketamine’s anti-suicidal effects may be achieved, ultimately leading to improved treatment options for individuals at risk of suicide.

CC, PK, and AD developed the theoretical framework. CC drafted a first version of the manuscript. PK, RM, and AD provided edits and suggestions, and assisted with draft finalization. All authors contributed to the article and approved the submitted version.

This work was supported by the Krembil Foundation (to AD).

RM has received research grant support from CIHR/GACD/National Natural Science Foundation of China (NSFC) and the Milken Institute; speaker/consultation fees from Lundbeck, Janssen, Alkermes, Neumora Therapeutics, Boehringer Ingelheim, Sage, Biogen, Mitsubishi Tanabe, Purdue, Pfizer, Otsuka, Takeda, Neurocrine, Sunovion, Bausch Health, Axsome, Novo Nordisk, Kris, Sanofi, Eisai, Intra-Cellular, NewBridge Pharmaceuticals, Viatris, Abbvie, and Atai Life Sciences, and is a CEO of Braxia Scientific Corp.

The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.