What are mechanisms? Unpacking the term is key to progress in neuroscience

How is the notion of “mechanism” used and valued in your subfield?

My subfield is very small; we are one of the few labs at the intersection of molecular and cognitive neuroscience, learning how molecular events influence the connectivity of the higher cortical circuits that generate cognitive operations in primate brain. We are the tiny footbridge over the increasing gulf between two expanding arenas—the world of molecular neuroscience, mostly focused on mouse brain, and the world of cognitive neuroscience, mostly focused on human brain imaging. The research in these huge subfields proceeds (relatively) rapidly and has attracted very large numbers of scientists, whereas the research to bridge these domains in the primate is tedious and performed by very few, even though it is vital to their integration. As the subfields of molecular and cognitive science diverge, so does their language, and it becomes increasingly harder for those in one world to understand the underlying concepts and discoveries of the other.

Is the meaning of the term “mechanism” used differently in these two different cultures? The dictionary definition of mechanism is “identifying causal relationships.” From where we stand, this fundamental meaning is retained in the molecular and cognitive subfields, and in our tiny subfield as well. As with the Hindu metaphor of infinite regression, “turtles all the way down,” we see “mechanisms all the way down.” However, the expansion and divergence of the molecular versus cognitive subfields has weakened our ability to understand the details of these mechanistic relationships in spheres beyond our own specialty, and without this knowledge, research is often dismissed by reviewers and editors as “not mechanistic.” The entropy created by the rapid expansion of these diverging arenas has left us ignorant of the true shape of modern neuroscience, and we need the energy and discipline to learn about research beyond our own disciplines to generate the enthalpy necessary to hold our field together.

From our vantage point, we can see how mechanisms in neuroscience are often about alterations in connectivity at all different levels—altered cognition caused by changes in the connectivity of circuits, such as correlations in functional MRI BOLD signals; altered circuit connectivity caused by changes in molecular “connectivity,” such as the opening of potassium channels near a synapse weakening its effects; and alterations in ion channel open state caused by changes in the connectivity of atoms, such as a phosphate group connecting to an ion channel rather than to ATP. By learning the languages of these different levels of inquiry, we can hope to provide a mechanistic understanding of cognition with both breadth and depth.

I work at the interface of systems and cognitive neuroscience, and indeed the notion of mechanism varies across subfields. Working across subfields, this fact can be a barrier to developing a consistent definition of mechanism. In systems neuroscience, I think the definition implied by mechanism is one of a more strictly physical causal chain, whereby lower scales of analysis inform upper levels—ionic currents are the mechanism underlying neurons firing, for example, or neurons firing are the mechanism underlying communication between brain areas.

This definition is perhaps in contrast to one that is common in cognitive neuroscience. In cognitive neuroscience, “mechanism” can be used to imply a link that has explanatory value—correlations between the summed activity in a particular brain area and behavior, for example— but that does not necessarily have a causative relationship. The implication is that such a mechanism could be a plausible cause of a particular cognitive process.

Neither the systems neuroscience definition nor the cognitive neuroscience one is exclusionary though—mechanisms can exist in systems without causation, such as a network model that doesn’t include physical forces such as ions; and mechanism can exist in cognitive neuroscience with causation, with the use of neuromodulation, such as transcranial magnetic stimulation, for example. In many situations, readers may alter the definitions they use when navigating literatures that span multiple subfields. This fact is also likely complicated by the potential use of “mechanism” as shorthand to convey value, particularly in the domain of publishing and grant-writing.

I often feel it would be more productive for the field, including journals and funders, to expand on what they explicitly value. The potential for a causative link or the power to explain an observable phenomenon, for example, are both valuable outcomes. If we give more weight to these kinds of outcomes, then there is less stress on the use of the word “mechanism” as a label. Giving scientists the opportunity to explain the value of their work will add accessibility and drive discovery across fields.

How is the notion of “mechanism” used and valued in your subfield?

I do my best to work across subfields, striving to integrate our understanding of memory across levels of biological understanding. We combine experiments focused on the molecular mechanisms of memory persistence and the neuronal network mechanisms of spatial information processing. Although “mechanism” in each instance refers to “how it works,” what constitutes mechanistic evidence is necessarily very different at distinct levels of biological organization.

My group and others have uncovered specific molecular mechanisms underlying memory, pinpointing structural and enzymatic changes that are crucial for initiating and maintaining LTP, respectively. (For more about this, see “Persistent protein pairing enables memories to last.”) Elucidating these distinct molecular mechanisms of initiation and maintenance relied on causal manipulations employing pharmacology, pharmacogenetics, optogenetics and genetic manipulations.

Unfortunately, how well causal manipulations work depends on the nature of the system, not merely the ingenuity of the experimenter. Indeed, causal manipulations may be an ineffective approach to elucidating network mechanisms, for example, in large part because biological neuronal networks are complex systems with layers of redundancy and feedback. Take the role of gamma oscillations in memory—we have shown that activity in the medial entorhinal cortex promotes slow gamma over mid-frequency gamma oscillations in part of the hippocampus, triggering recollection. Although these electrophysiological phenomena are part of how recollection works, recollection mechanisms are likely redundant. So when researchers attempt to manipulate slow gamma oscillations to test their role in memory, the network almost certainly compensates in an attempt to prevent a complete collapse and the biophysical origins of gamma oscillations contribute to more than just recollection.

Redundancy, multi-functionality, non-stationarity and multi-stability are fundamental features of any complex system, including most, if not all, biological neuronal networks. These features designed to make the system more robust can also render causal manipulations futile. Of course, even molecular mechanisms can be part of redundant, multi-function, multi-stable and non-stationarity systems, making interpretation of causal manipulations fraught (see below).

 Does the apparent preference for mechanisms in journal guidelines and grant calls capture its status in your subfield?

Although the clear preference is to focus on discovering mechanisms, this goal becomes a problem if we assume that only causal evidence can identify mechanisms and when criteria such as necessity and sufficiency are applied to complex systems. This is obviously problematic at the level of neuronal networks and non-linear systems with redundancy and feedback—where does a circle begin? But it is also evident at the level of molecules, which similarly form networks of interaction. Two prominent examples are CaMKII-alpha and PKMzeta, considered by some to be rival candidate mechanisms of how LTP and memory are maintained. Like a molecular switch, once turned on, both of these enzymes keep themselves activated via autophosphorylation.  CaMKII autophosphorylation is crucial to initiate LTP and memory formation because it enables a CaMKII complex to bind the NMDA glutamate receptor subunit GluN2B to initiate LTP and memory, but further enzymatic activity by CaMKII-alpha is not crucial. In contrast, the enzymatic action of PKMzeta is crucial during the maintenance of LTP and memory hours, days and months after initiation because the persistent kinase function is targeted to memory-activated synapses by kidney and brain expressed protein (KIBRA), a postsynaptic scaffolding protein. Definitively demonstrating PKMzeta’s role was difficult because of redundancy in the system—deleting the gene doesn’t abolish LTP or memory because other molecules can compensate for roles of the missing kinase. Instead of being rivals, the early role of CaMKII-alpha in initiation and the later role of PKMzeta in maintenance are complementary and interacting through a multiplicity of additional components.

Even these bona fide molecular mechanisms reveal a fundamental multiplicity of interactions with redundancy and feedback. Similar to neural networks, these molecular mechanisms depend on interactions of the system components, and those interactions are sensitive to context and time. Contemporary concepts of “population dynamics” for understanding neuronal network functions may also prove to be essential for understanding molecular signaling networks by emphasizing that reliable component interactions define low-dimensional component configurations, so-called manifolds, within which the high-dimensional population of components organize lawfully according to a mixture of correlational and causal interactions. Necessity and sufficiency and other causal evidence cannot be readily established in such systems because redundant interactions define the latent subspaces that are crucial for governing the dynamical organization of the system of interacting components.

Are there other types of causal systems (or evidence) that you think should be given more weight?

 It is valuable to seriously consider whether causality can ever be demonstrated or is even a valuable concept to strictly consider and require in certain domains of neuroscience, because causal manipulations can only be meaningful in a subset of processes in which the manipulated element has a single function. Mammalian brain operations often don’t use elements that have a single function with no feedback or redundancy, making successful causal experiments improbable. The example of a common electrical switch circuit that all homes have at the bottom and top of a staircase makes the point that some perfectly well-understood simple systems cannot be elucidated using causal evidence and reasoning. This is cartooned in the figure below to illustrate that we can be mistaken by applying the causal criteria that can be effective in one type of system to other systems with complex interactions, or even a simple electrical system with redundancy.

This schematic (right) of a common electrical circuit the “single pole double throw switch” illustrates how so-called causal investigations can be ill-conceived depending on the assumptions about the underlying circuit architecture. A) Observing this position of switch 1, one might hypothesize it turns on the light, and B) confirm this inference by manipulating switch 1 to the other position. However, when switch 2 changes position as in C, D), the opposite conclusions will be reached about switch 1 (and 2). It would be wrong to conclude that switch 1 does not control the light, because it does, and although manipulating a single switch cannot reveal its function, the circuit logic is readily understood by observing the correlations between switches 1 and 2 and the on/off state of the light bulb. Note also that removing either switch would show that it is necessary for the functioning of the circuit, but not reveal the operational logic.

In the field of clinical psychology and developmental neuroscience, the notion of “mechanism” is frequently invoked but often misunderstood. Although mechanisms are framed as causal explanations, most human research remains inherently observational. There are few methods that allow us to ethically manipulate brain structures or neural circuits in living humans to establish direct causality. Instead, we often rely on correlational data, yet the term “mechanism” is frequently applied to findings from studies designed to assess associations rather than true causality. The overuse of the term “mechanism” in observational studies runs the risk of conflating correlation with causation, implicitly suggesting a direct causal relationship that the study design was not equipped to demonstrate.

From a broader perspective, the idea of a mechanistic explanation in human behavior research is both too simplistic and infinitely complex. There is a hierarchy of mechanisms contributing at different levels of analysis: Dysfunction in brain regions or circuits may arise from underlying cellular or molecular processes that are not functioning properly. These cellular and molecular disturbances may be linked to genetic predispositions. Gene expression may be differentially activated due to environmental influences. Thus, if I study brain circuits, it would be easy to say that I’m not tapping into the most fundamental underlying causal mechanism (molecular, cellular, genetic or environmental). However, as we move through these levels, we realize that no single level of analysis captures the full scope of causality. Even when we identify a so-called mechanism at one level, this raises further questions about how that mechanism interacts with broader biological, environmental or contextual factors. Mechanisms are not static or universal but vary across individuals and conditions, complicating the search for definitive causal pathways.

The apparent preference for mechanisms in journal guidelines and funding calls reflects an overly simplistic view of causality that does not fully capture the subtleties inherent in human behavior and neuroscience research. Moving forward, it will be important to recognize that all levels of analysis—whether genetic, cellular/molecular, neural or behavioral—offer valuable insights. Instead of emphasizing a vague call for “mechanistic research,” we should encourage integrative approaches that foster collaboration across disciplines and methodologies to capture the complexity of human behavior and its underlying processes.

Dani Bassett and Lauren Ross propose that it is high time that some clarity and order were brought to bear on the ubiquitous use (abuse) of the term mechanism in neuroscience and beyond. I wholeheartedly agree, but I do not think their piece provides either diagnosis or cure.  The problem is that there has been a lot of thoughtful work, mainly by philosophers, on the concept of mechanism. Neuroscientists, for the most part, alas, do not really care to read this philosophical literature but are nevertheless convinced that proper explanations of behaviors generated by the nervous system must be mechanistic: They must show how physical parts in space and time interact to generate the whole behavior. This cogs-and-springs-in-a-clock view is also expected to be causal: The clock will cease to work in predictable ways if either cog x or spring y are messed with. The favorite way to mess with the intricate springs and cogs of neural circuits these days is optogenetics. I apologize for the blunt simplicity, but this is the stance of mainstream neuroscience. The idea that there can be legitimate causal explanations that are not about physical things that act on each other locally (circuits) or distantly (networks) is not really taken that seriously in mainstream neuroscience.

Psychology and cognitive science, in contrast, employ componential functional explanations, which critically can also be causal. In an example from Daniel Dennett, which I think I either read or heard, you are walking toward a box, and I say to you, “Don’t go any closer; it has poisonous snakes in it,” and you come to an abrupt stop. The meaning of my utterance has made your muscles contract. This is obviously a causal story but not one composed of a neural account of how sounds in your auditory pathway went through subsequent intermediate processing steps in your brain, culminating in motor commands. In my own work, systematic differences in corrective responses to directional reaching errors can be causally explained by the precise nature of the errors themselves. But neuroscientists consider these mere placeholders or sketches until the associated circuits/networks are identified and characterized. In this view, psychology is a discipline waiting for mechanistic rescue. This, it is sad to say, is the misguided ideology behind the incessant requests by journals for mechanism.

To move forward, there will need to be broader familiarity with the “more is different” framework of complexity science, which has long eschewed privileging a particular level of explanatory granularity or coarse graining, and instead promotes the pluralistic idea of level-dependent effective theories.  If one were to be charitable, one might say that neuroscience is the field, among those interested in the relationship between brain and behavior, that is in fact defined by its requirement for mechanistic causal explanations using structural facts about neurons. Thus, neuroscience is well suited for mechanistic explanations of behaviors such as reflexes, locomotion and eye movements. It can’t, however, accomplish this for all behaviors, but it is nevertheless an invaluable source of confirmatory evidence for non-mechanistic explanations of, for example, linguistic comprehension. Thus, neuroscience is useful because it provides mechanistic explanations of some phenomena and lower-level evidence for non-mechanistic, but nevertheless causal, psychological explanations of other phenomena. The latter will never collapse into the former, which is perfectly OK.

My work spans two fields, cognitive and computational neuroscience, which makes the concept of “mechanism” challenging. In cognitive neuroscience, a mechanism can be how one process affects a higher-order process; selective attention is a mechanism that determines the details that are encoded in memory, for example. This gets translated to regional or network correlates to explain behavior—the level of engagement of the attentional network is a mechanism to explain variation in how well a memory network encodes a memory.

On the computational side, some mechanistic models contain biophysical entities that explain circuit or network behavior. Here, the system of equations may represent excitatory and inhibitory neurons and the connections within and between populations. These mechanisms generate collective behaviors that explain a phenomenon. For example, in development, the increase in inhibitory neuron influence reduces the E/I ratio, which leads to more complicated regional dynamics and, hence, the full network.

In both cases, it’s interesting to note that the mechanism per se is less critical to the explanation than the cascade effect on a network’s state. This is an important point because the brain is a dynamical system, in which there can be many causes for the same effect. This should favor explanations that help constrain the possibilities that link mechanisms to their cascade effect(s).

Unfortunately, for both fields, the emphasis on mechanisms is extreme. A mechanism is supposed to connote essential insight—a “magic bullet.”  But that may not be accurate, especially given the indeterminism between scales in the brain. The action of a mechanism depends on the context. Computational approaches are better for dealing with this contextual uncertainty. Yet I’ve had grant reviews for my computational work that state “ … no clear biological questions, going beyond application of computational models and exploration of parameters,” which sometimes comes from cognitive neuroscience reviewers.

I would assert that, rather than focusing on uncovering the mechanism, it would be more productive to examine how the mechanism reduces the uncertainty of cascade effects. Focusing on cascades would be a more fruitful avenue than making the assumption that a single mechanism can explain a complicated multiscale behavior.

I propose that the concept of mechanism in neuroscience must evolve to match the complexity of brain function, rather than being constrained by traditional reductionist views. The brain is not a machine with neatly separable parts, but a highly entangled system in which function emerges from interactions across multiple spatial and temporal scales. Thus, we have to let go of a Newtonian “billiard-ball” notion of causation and embrace newer frameworks.

In my view, a mechanism in the brain should be characterized as a distributed causal system involving interactions across the neuroaxis—the medulla, brainstem, subcortical forebrain and cortical forebrain—which gives rise to emergent properties and behaviors via non-linear dynamics. This conception goes beyond molecular and cellular interactions to encompass large-scale network and circuit dynamics.

Consider, for example, how we might describe the mechanism of fear processing. A traditional view might focus solely on the amygdala and its molecular cascades. However, this reductive approach misses crucial aspects of how fear actually manifests in the brain. A more comprehensive mechanistic explanation would include how the amygdala interacts with midbrain, subcortical and cortical regions, how these interactions are modulated by neurotransmitter systems, how large-scale network dynamics shift during fear states, and how these processes are constrained and shaped by the brain’s overall anatomical connectivity architecture.

This broader view of mechanisms aligns with the complex systems nature of the brain. It acknowledges that brain functions cannot be localized to a single region or reduced to a linear sequence of molecular events. Instead, these functions emerge from the dynamic interplay of multiple brain systems operating across different spatiotemporal spatial scales organized in a largely heterarchical manner, meaning they lack a clear-cut hierarchy.

Importantly, this concept of mechanisms doesn’t discard the value of molecular and cellular research. Rather, it places these findings within a larger context, recognizing that understanding the brain requires integrating information across multiple levels of analysis. It calls for new methodological approaches, including network neuroscience, complexity science, dynamical systems modeling and multi-scale imaging methods.

Adopting this view of mechanisms has important implications for how we conduct and evaluate neuroscience research. It suggests that we should value studies that elucidate circuit-level causal dynamics as much as those that uncover molecular pathways. It calls for increased interdisciplinary collaboration to bridge levels of analysis. And it demands that we develop new theoretical frameworks capable of explaining how complex, distributed processes in the brain give rise to cognition and behavior.

As Bassett and Ross elegantly discuss, mechanism is a “suitcase” term in neuroscience that contains multiple, diverse concepts that are not always easy to unpack. In the subfield of cognitive neuroscience, the notion of mechanism is often used to refer to the neurobiological process underlying cognition.

The most widely used tool to noninvasively uncover these neurobiological processes in humans is functional magnetic resonance imaging (fMRI), which gives us correlates of neuronal activity. Bassett and Ross review how mechanisms are often viewed as causal systems and are, for this reason, highly valued. It is somewhat ironic, then, that the field of cognitive neuroscience simultaneously values understanding causal mechanisms of cognition yet relies heavily on a method that can only give us correlates. Of course, there are other tools that let us drive neuronal activity, such as transcranial magnetic stimulation, and can in principle be used to test causal claims. An important goal for cognitive neuroscience will be to more precisely define mechanism at the level of macro-scale neural systems. We might also have to come to terms with the fact that if mechanism necessarily implies causal, then most fMRI studies would not meet the criteria for revealing causal mechanisms underlying cognition.

In the neurobiology of learning and memory, “mechanism” is highly valued because it helps explain how processes such as synaptic plasticity, gene expression and neural activity contribute to memory formation, consolidation and recall. To establish a mechanism, we often look for two key conditions: necessity and sufficiency. An element is considered necessary if the process can’t happen without it, typically shown by such “loss-of-function” experiments as blocking a gene or inhibiting neural activity.

On the other hand, something is sufficient if its presence alone can trigger the process, even without other contributing factors. This is demonstrated through “gain-of-function” methods, such as neural stimulation or gene overexpression. Though showing both necessity and sufficiency suggests a solid understanding of a mechanism, learning and memory processes are often far more complex, involving many interacting components that are difficult to capture solely using these simpler manipulations.

Recent work that treats brain processes as dynamical systems may provide a clearer picture of the processes driving behavior and cognition. In this approach, other types of causal evidence—such as statistical dependence, Granger causality, information flow and oscillatory coupling—reveal relationships that may not fit neatly into traditional mechanistic definitions of necessity and sufficiency. How memory is spread across neural networks, for example, or how information flows between the hippocampus and cortex is better explained through network or attractor models. Broadening what we value beyond strict mechanisms gives us a deeper understanding of brain function on many levels.

“Mechanism” certainly holds a privileged status in the neurobiology of learning and memory, reflected in high-impact journal editorial decisions, as well as peer review. Funding bodies such as the National Institutes of Health also often emphasize mechanistic insights as desirable in grant calls, and this is also reflected in study section evaluations. However, this emphasis can sometimes be limiting, because it encourages a narrow focus on molecular and cellular explanations, potentially underappreciating systems-level, statistical or computational approaches that are equally valuable. Though mechanistic clarity is crucial, broadening the conceptual framework to include other causal explanations would enrich the field and promote interdisciplinary approaches, fostering a more complete understanding of brain function.