Variability in Neural Circuit Formation

Abstract

The study of neural development is usually concerned with the question of how nervous systems get put together. Variation in these processes is usually of interest as a means of revealing these normative mechanisms. However, variation itself can be an object of study and is of interest from multiple angles. First, the nature of variation in both the processes and the outcomes of neural development is relevant to our understanding of how these processes and outcomes are encoded in the genome. Second, variation in the wiring of the brain in humans may underlie variation in all kinds of psychological and behavioral traits, as well as neurodevelopmental disorders. And third, genetic variation that affects circuit development provides the raw material for evolutionary change. Here, I examine these different aspects of variation in circuit development and consider what they may tell us about these larger questions.

THE CODING PROBLEM

If we want to understand how variation can affect the processes of development, we need to understand how those processes are encoded in the genome. A good deal of research in developmental neurobiology is about this question, or at least is relevant to it. The main approach in experimental developmental neurobiology has been to isolate specific processes or developmental events, such as axons crossing the midline or innervating a muscle or topographically projecting across a target region, and then to try and identify specific molecules involved in these processes and investigate their functions.

This approach has been successful in identifying many molecules involved in directing axon guidance and synaptic target selection (Kolodkin and Tessier-Lavigne 2011; de Wit and Ghosh 2016; Mitchell 2018a; Sanes and Zipursky 2020). However, this reductive, controlled approach risks giving the impression that individual guidance or targeting events are controlled by individual molecules, acting in isolation. The reality in vivo is much messier and more complex—all guidance and connectivity decisions are influenced by multiple factors acting at once (e.g., Winberg et al. 1998; Yu et al. 2000; Xu et al. 2020).

A growth cone will encounter multiple secreted or cell-surface molecular cues at any moment, as well as adhesive and anti-adhesive proteins, the general environment of proteins in the extracellular matrix, and the physical forces presented by the landscape over which the axon is extending (Franze 2020; Breau and Trembleau 2023). The influence of these factors will be determined by the repertoire of receptor and adhesion proteins that the growing neuron itself is expressing, but not in a linear, one-at-a-time fashion (Klumpe et al. 2023). Many guidance and connectivity receptors operate in signaling complexes, with context-dependent interactions in cis as well as in trans (e.g., Marquardt et al. 2005; Perez-Branguli et al. 2016; for reviews, see Morales and Kania 2017; Stoeckli 2018; Südhof 2018; Zang et al. 2021).

This array of molecular and physical factors collectively generates what can be thought of as an energy landscape, constraining the growth of the axon along specific routes. Conrad Waddington proposed the “epigenetic landscape” as a visual metaphor of cell fate decisions through development (Waddington 1957). His idea was that this landscape was shaped by the functions and interactions of large numbers of genes (what we would now call a gene regulatory network), which collectively constrain the expression profiles of differentiating cells along defined trajectories toward specific attractor states (mature cell types). This idea can be adapted to describe the processes of axon guidance and synaptic target selection, with the added element that the “landscape” is not in abstract “gene expression space,” but describes the real physical (and molecular) terrain of the developing nervous system (Fig. 1). This terrain presents a different set of possible growth channels to axons depending on their repertoire of receptors.

Figure 1.

The neurodevelopmental landscape. A reimagined version of Waddington's landscape, showing axonal growth cones being channeled down alternate paths (A,B) by virtue of their differential sensitivity to molecular guidance cues expressed across the physical terrain of the developing nervous system.

Rather than a one-to-one functional mapping between specific cue-receptor pairs and specific guidance decisions, this gives a view of a collective set of constraints that tend to channel growing neurons to their appropriate target regions and tend to direct synaptogenesis with the appropriate target cells or even the appropriate subcellular regions (Hiesinger 2021a,b). In turn, this landscape of cues and the respective repertoires of receptors are generated by a developmental program orchestrated by gene regulatory networks and patterning mechanisms involving dozens or hundreds of signaling proteins and transcription factors (Zarin et al. 2014; Herrera and Escalante 2022).

The distributed, indirect nature of this genetic encoding of wiring patterns has implications for how the resultant (really emergent) patterns are affected by different types of genetic variation. It also highlights the inherently statistical nature of these processes, which are probabilistic at the level of single cells, due to noise and randomness at molecular levels.

The genome does not contain enough information, quantitatively speaking, to specify the projection of every axon and the connections of every neuron (Hassan and Hiesinger 2015; Koulakov et al. 2021). It does not encode endpoint information directly at all, in fact, nor does it directly encode algorithmic information (Hiesinger 2021a), at least not in the sense of a set of instructions that isomorphically and separably specify the processes of development in a decomposable manner. Rather, it encodes a generative model—a set of latent variables (protein and regulatory element sequences) that collectively constrain the self-organizing processes of development to reliably produce a new individual of some species type (KJ Mitchell and N Cheney, unpubl.). Changes to the sequence of the genome can alter these latent variables and affect the outcome. But the developmental processes themselves also necessarily introduce variability in the detailed outcome (Vogt 2015; Mitchell 2018b).

SOURCES OF VARIATION IN OUTCOME

The outcome of neural wiring will thus vary across individuals of a species for two reasons: inevitable variation in the genetic program and equally inevitable variation in how any individual “run” of any particular version of that program plays out. In addition, exposure to certain environmental stressors, toxins, or other insults can affect neural development in diverse ways. These can be clinically important (e.g., fetal alcohol spectrum disorders; Popova et al. 2023), but will not be discussed further here. Another major source of variation in neural wiring across individuals within a species is sex. Males and females of many species show systematic variation in connectivity of specific circuits, particularly those underlying sexual and reproductive behaviors (Knoedler and Shah 2018; Meeh et al. 2021). Here, we will concern ourselves with sources and consequences of individual differences more generally.

GENETIC VARIATION

On the genetic front, phenotypic variation can arise due to single mutations of large effect and/or due to the combined, polygenic effects of many genetic variants. Experimental work in model organisms typically involves the former, although polygenic background can be an important modulating factor. In humans, both types of genetic influence are at play.

Single Mutations

As anyone who has spent months or years looking for a phenotype in a mutant animal knows, many mutations of single genes—even homozygous complete null mutations—seem to be well tolerated by the developing organism. Even when biochemical and gene expression evidence suggest the likely involvement of some specific protein in the guidance of some axons or the specification of their synaptic connections, removal of that protein may produce no phenotypic effect, or may cause effects in some anatomical contexts where the gene is expressed, but not in others.

This suggests that either the protein in question is not in fact involved in the processes specifying the circuitry in question or that this function is dispensable, due to redundancy, robustness, or regulation (some reactive compensatory processes). In model organisms, such cryptic functions can often be revealed through analysis of epistatic interactions in genetically sensitized backgrounds. These kinds of screens and analyses have provided the means to identify guidance pathways operating in parallel (e.g., Winberg et al. 1998; Yu et al. 2000; Cate et al. 2016) or to elucidate biochemical pathways connecting cues and receptors to cellular responses (for review, see Zang et al. 2021).

That being said, it is also true that many single mutations do have large effects on neural circuitry development. Forward genetic screens have been highly successful in identifying important molecules that specify neural projections and connectivity. These have been based, for example, on direct anatomical visualization of axonal patterns (e.g., Seeger et al. 1993; van Vactor et al. 1993; Leighton et al. 2001) or on behavioral outcomes (e.g., Hedgecock et al. 1990). This approach is highly powerful in that it lets the system tell you what is important in an unbiased way. But such screens will, as a consequence, identify mutations that affect the phenotypes of interest in possibly quite indirect ways. Indeed, as we will see below, most of the genetic variation affecting any given process probably does so indirectly (i.e., by affecting proteins not directly involved in the cellular process itself) (Boyle et al. 2017).

The opposite approach—reverse genetics—is also a powerful method to test in vivo the importance of some putative guidance or connectivity factor identified through biochemical or molecular means for example (e.g., Mitchell et al. 1996; Serafini et al. 1996; Feldheim et al. 2000). This is a more directed (or biased) approach from the outset, but it is usually not possible to predict in advance how the system will react to disruption of any given gene.

One interesting trend is that, while loss-of-function or removal of a single gene is often reasonably well buffered by the system, gain-of-function mutations or manipulations can often produce far stronger effects. For example, ectopic expression of a cue can often induce much more dramatic effects on the system than its removal (e.g., Nose et al. 1994; Mitchell et al. 1996; Winberg et al. 1998). Similarly, dominant-active or dominant-negative mutations can have wider consequences on biochemical pathways and developmental processes than simple removal of a protein. These kinds of effects are particularly relevant in human genetics.

The Genetics of Neurodevelopment in Humans

Neurodevelopmental disorders in humans are often divided into rare “Mendelian” conditions (many with identified genetic causes, such as Fragile X syndrome, Rett syndrome, Timothy syndrome, etc.), or common disorders with broad diagnostic labels such as autism, schizophrenia, epilepsy, intellectual disability, attention-deficit hyperactivity disorder, and many others. The latter are highly heritable but are taken to be genetically “complex,” and most cases until recently have remained idiopathic (no clear cause has been identified).

In fact, this dichotomy is largely artificial for two reasons (Mitchell 2015). First, more rare mutations are being discovered that confer high risk for these common diagnostic categories, revealing underlying genetic heterogeneity of these umbrella terms. And second, because even the supposedly “Mendelian” conditions show important modifying effects of other mutations or of the polygenic background more generally.

Mutations in axon guidance and synaptic connectivity genes have been implicated in both specific neurological conditions and in these broader diagnostic categories. There is a small number of specific neurological conditions in humans that result from mutations in canonical “axon guidance” genes. (The scare quotes denote the fact that these proteins often play roles in other processes and even in other tissues). In some cases, the neuroanatomical pathology and neurological symptoms likely reflect quite direct roles of the genes in question in patterning specific circuits. These include, for example, defects in oculomotor circuit formation in patients with ROBO3 mutations, resulting in horizontal gaze palsy (Jen et al. 2004), or defects in midline axonal structures in patients with mutations in the DCC gene, resulting in “mirror movements” (Srour et al. 2010), as observed in mice (Fazeli et al. 1997). A similar condition is observed in patients with dominant mutations in ARHGEF7, which encodes a component of the DCC signal transduction pathway (Schlienger et al. 2023). (Other examples of human disorders arising from mutations in guidance or connectivity genes [including CNTNAP2, L1CAM, NTNG1, and PCDH19, for example] are reviewed in Engle 2010; Blockus and Chedotal 2015; Betancur and Mitchell 2015; and Yuasa-Kawada et al. 2023.)

Most of these conditions are recessive, meaning both copies of the gene must be mutated to cause the phenotype (as is commonly observed for axon guidance genes in model organisms). Such conditions are therefore very rare and have usually been identified in consanguineous populations (e.g., for ELFN1 gene; Dursun et al. 2021). However, while this is true of complete loss-of-function mutations, mutations that result in a truncated protein (especially for transmembrane receptors) can often have dominant effects. Because many guidance proteins interact in complexes with other factors, a truncated protein can have a greater (dominant-negative) impact on developmental processes than simple absence of a protein. For example, single-copy de novo truncating mutations in SEMA6B have been implicated in cases of epilepsy (Hamanaka et al. 2020) and intellectual disability (Cordovado et al. 2022), while similar mutations in PLXNA1 have been implicated in autism cases (Fu et al. 2022). This class of mutation may make an important contribution to the etiology of neurodevelopmental disorders more generally (Torene et al. 2023).

Mutations in some other genes, including some involved in synaptic connectivity such as NRXN1, can present with much more variable clinical manifestations and generally increase risk across many common diagnostic categories (Castronovo et al. 2020). Among rare de novo or inherited high-risk mutations in cases with intellectual disability, developmental delay, epilepsy, autism, or schizophrenia, there are examples of genes involved in axon guidance or neuropil patterning (e.g., DSCAM, PLXNA1, PLXNB1) or synaptogenesis (e.g., NRXN1, NLGN4X) (Satterstrom et al. 2020; Fu et al. 2022; Trost et al. 2022).

Risk genes are generally enriched for genes expressed in fetal brain and involved more broadly in neurodevelopmental processes. These include processes “upstream” of circuit specification, like regulation of gene expression and chromatin function, as well as “downstream” processes like synaptic plasticity or function, which may impact on activity-dependent refinement of neural circuits (e.g., NMDA-receptor genes like GRIN2A or synaptic protein genes like SHANK3 or SYNGAP1; Betancur and Mitchell 2015). However, there are also many genes implicated in these illnesses where neither a direct nor indirect link to the processes of circuit formation or refinement is apparent.

Even for conditions defined clinically by direct neuroanatomical dysconnectivity, such as agenesis of the corpus callosum, most of the implicated genes are not ones that encode specific guidance cues and receptors (Pânzaru et al. 2022). This reflects the fact that the processes of neural circuit formation do not just rely on the genes specifying the “instructions” of which axons project where, but also on those encoding all the proteins required for the more general machinery of regulation of gene expression, signal transduction, cellular motility, and so on. There are, genetically speaking (as observed for all complex traits [Boyle et al. 2017] and in forward genetic screens in model organisms [Mitchell 2018a]), simply many more ways to impair the processes of neural circuit formation indirectly and nonspecifically than the converse.

Some important principles of the genetic architecture of neurodevelopmental disorders are revealed by the study of a particular class of mutations: copy number variants (CNVs) (Mollon et al. 2023). These are deletions or duplications of segments of chromosomes, or, as in Down syndrome, even whole chromosomes. These mutational events can recur at certain sites in the genome, with a very low frequency, but enough to mean that many, many thousands of people carry the effectively identical genetic lesion. CNVs can cause identifiable syndromes, such as 22q11.2 deletion syndrome, 3q29 deletion syndrome, Williams syndrome, and many others (some classically diagnosable based on things like typical facial morphology, for example). But these genetic lesions are also found at much higher frequency in cases with more general neuropsychiatric conditions than in the control population.

A key finding is that such mutations seem to increase risk across these diagnostic categories, reinforcing the view of an overlapping etiology. The same is true for rare, high-risk, single-gene mutations. Thus, many mutations seem to lead to a general developmental brain dysfunction (Moreno-De-Luca et al. 2013), which can manifest in diverse ways. These manifestations include the qualitatively distinct end states that we recognize as “autism,” “schizophrenia,” “bipolar disorder,” and so on, which may be best thought of as representing maladaptive attractor states that the developing brain may end up in. These phenotypes emerge from the way that the developing brain reacts to a very wide variety of possible insults, rather than any direct molecular function of the disrupted genes. This highlights another key principle at play: the ability of the developing system to buffer such insults is also itself affected by genetic variation.

First, it is observed, not surprisingly, that harboring multiple rare mutations (CNVs or single-gene mutations) increases clinical risk substantially (Girirajan et al. 2010; Guo et al. 2019). Such mutations may be inherited (often from clinically unaffected parents) or may arise de novo. Second, risk is also modified by a polygenic background of common genetic variants of the sort that can be identified by genome-wide association studies (GWAS) (Niemi et al. 2018; Bergen et al. 2019; Antaki et al. 2022; Cirnigliaro et al. 2023).

Polygenic Variation

GWAS for conditions like schizophrenia, autism, bipolar disorder, depression, epilepsy, or other neuropsychiatric conditions have identified hundreds of common single-nucleotide polymorphisms (SNPs), which confer increased risk. SNPs are sites in the genome where some percentage of the population has one base (say an “A”), while others have a different base (say a “C”). If one of the versions is significantly more frequent in cases with a particular condition than in controls, then this association implies that those variants increase risk (given that the opposite direction of causation is implausible). Importantly, the increased risk conferred by any single SNP is tiny—almost, but not quite negligible—but the collective risk conferred by the overall polygenic burden of such common risk variants can be substantial.

This kind of polygenic background has been shown to be a contributor to clinical risk in combination with rare mutations in autism and schizophrenia (Bergen et al. 2019; Antaki et al. 2022; Cirnigliaro et al. 2023) and a modifier even of “Mendelian” neurodevelopmental conditions (Niemi et al. 2018). As with the rare variants, most of this polygenic risk seems to be shared across conditions.

For our purposes, what is interesting is that the genes implicated by these common risk variants, as observed for rare, high-risk mutations, are consistently enriched for ones expressed in the embryonic or fetal brain and involved in various neurodevelopmental processes, including synapse organization and synapse assembly, as well as ion channel biology and synaptic transmission (Mallard et al. 2022; Trubetskoy et al. 2022; Als et al. 2023).

In addition to neurodevelopmental disorders, GWAS also implicate neurodevelopmental genes and processes in the genetic variance contributing to general differences in brain morphology and connectivity as well as psychological and behavioral traits in humans and other species. For example, GWAS of structural brain connectivity measures derived from diffusion-weighted neuroimaging found enrichment for genes involved in “neuronal differentiation,” “neural migration,” “neural projection guidance,” and “axon development” (Sha et al. 2023). This suggests, not surprisingly, that variation in such genes in humans can manifest at the macroscopic level of brain-wide structural connectivity.

The same trend emerges for GWAS of cognitive traits (e.g., Davies et al. 2018) and personality traits (e.g., Karlsson Linnér et al. 2019; Belonogova et al. 2021). These consistently show statistical enrichment for genes expressed in fetal brain and involved in a variety of neurodevelopmental processes, such as “neurogenesis,” “regulation of nervous system development,” “regulation of neuron projection development,” and “synapse assembly,” among others. (Note, again, that such enrichments are not exclusive—many other kinds of genes can also contribute to such phenotypic differences, presumably less directly.)

In dogs, a very large study identified genetic drivers of diversification of behavioral traits between breeds. These highlighted variants in noncoding regions of genes enriched for fetal brain expression and for rather broad functional categories of “development” and “neurogenesis.” Post hoc analyses of a sheepdog cluster identified variants in a cluster of genes involved in axon guidance, including members of the Ephrin, Netrin, slit, and semaphorin pathways (Dutrow et al. 2022).

Collectively, these genetic studies show that variation in genes directly involved in processes of neural circuit assembly contributes to the etiology of specific, rare neurological disorders, as well as common neuropsychiatric conditions. In addition, variation in neurodevelopmental genes contributes to heritable differences in brain structural connectivity and in psychological and behavioral traits across the general population. It is important not to overplay these enrichments, however—the majority of genetic variants affecting neural circuit development will likely do so indirectly. Regardless, the upshot is that genetic differences affecting brain wiring (directly or indirectly) can manifest in important ways.

However, genetic differences are not the only source of variance in neural circuit development and brain wiring. Another important source—often overlooked—is stochasticity in the processes of development themselves.

DEVELOPMENTAL VARIATION

As mentioned above, the genome does not contain enough information to encode precisely the numbers and positions of all the different cell types in the nervous system and the myriad connections they each make. Rather, it encodes a generative model that constrains the self-organizing processes of cell differentiation and migration and the subsequent processes of axon guidance and synaptic target selection. These processes have evolved to robustly and reliably direct development toward an outcome that falls within a species-typical range. However, this is achieved statistically through collective cellular interactions rather than deterministically on a cell-by-cell basis.

The processes involved rely on wet, jiggly, jittery components and are consequently noisy on molecular and cellular levels (Symmons and Raj 2016). There is substantial stochasticity at play in gene expression (Raj and van Oudenaarden 2008), protein–protein interactions, diffusion of molecules, and all other cellular processes (Tsimring 2014). The system has, necessarily, evolved to minimize the impact of molecular and cellular noise on the resultant processes and outcome of development (Masel and Siegal 2009; Wagner 2013). However, this robustness has limits—considerable variation in the microscopic details of neural anatomy and connectivity still arises due to this inherent randomness. The processes of development do not play out in exactly the same manner in any individual “run” of the program, even when starting from the identical genome (Vogt 2015; Mitchell 2018b).

This kind of variability is apparent even in wild-type animals but manifests much more obviously in mutants. The robustness of the system is an evolved property that depends on the integrity of the genomic “program” as a whole. In the presence of genetic variation—either large-effect single mutations or a polygenic background of common variants—this property of distributed robustness is degraded. The consequence, well known since the early days of genetics, is that mutations tend to not only shift the mean of some quantitative phenotype or produce a tightly defined qualitatively novel phenotype, but also increase the phenotypic variance (Waddington 1957). This observation is commonplace in research on model organisms, where mutations can result in probabilistic outcomes of neurodevelopmental phenotypes.

To take two arbitrary examples (from my own experience), mutations in the Semaphorin-6A gene in mice result in defasciculation and misprojection of the fibers of the fornix—the output projections from the hippocampus to basal forebrain regions (Rünker et al. 2011). However, this phenotype is only partially penetrant and manifests in an apparently probabilistic way on either side (or both, or neither) across genetically identical animals (Fig. 2). Similar diversity is observed in the projections of the corticospinal tract across individual Sema6A mutant mice (Okada et al. 2019).

Figure 2.

Probabilistic phenotypes. (A,B) PLAP-stained adult mouse brain sections from the Sema6A gene trap line (Leighton et al. 2001) showing, in coronal cross-section of the tract, a reduced and defasciculated fornix (fx) in (B) a homozygous mutant Sema6A^−/− brain (on one side only; arrow), compared to (A) a heterozygous Sema6A^+/− (phenotypically wild-type) animal. (C,D) Photomicrographs of stage 17 Drosophila embryos stained with mAb 1D4 (anti-Fas II) to show motor projections across a number of hemi-segments. Anterior is left, and dorsal is up. (C) Wild-type shows normal projection of the RP3 motor axon, which runs in the ISNb nerve, to innervate the cleft between muscles 6 and 7 (arrow). (D) In embryos deficient for Netrin-B (normally expressed specifically on muscles 6/7), this innervation is absent in about 30% of segments (arrow). (A and B reprinted from Rünker et al. 2011 under the terms of the Creative Commons Attribution License; C and D reprinted, with permission, from Winberg et al. 1998.)

At an even finer level, probabilistic effects can be observed in the projections of individual motor axons from segment to segment in the Drosophila embryo. For example, in embryos mutant for the Netrin-B gene, the RP3 motor neuron sometimes fails to project to and thus innervate the 6/7 body wall muscles (Mitchell et al. 1996). But it does so with an apparently random probability from segment to segment of around 30% (Fig. 2). (This probability can be changed by simultaneously raising or lowering the levels of other guidance cues [Winberg et al. 1998].)

These examples of intra-individual variation strongly argue against the idea that developmental variability observed between genetically identical animals can be ascribed to unknown environmental factors. Rather, they demonstrate both the inherent variability of developmental processes and the fact that this variability is itself a genetic trait.

In some cases, this kind of noise can manifest in a dichotomous manner at the level of macroscopic connectivity. This is due to the highly contingent nature of brain development and the resultant fact that small changes due to noise in some early processes can have cascading effects over later development, resulting in highly divergent trajectories. One well-studied example is the formation of the corpus callosum, the set of axons connecting the two cerebral hemispheres in mammals (Suárez et al. 2014).

The corpus callosum is pioneered by a small set of early-projecting axons (Fig. 3). Their ability to cross the midline depends on the earlier formation of a small astroglial bridge between the two hemispheres (Gobius et al. 2016). If formation of this structure is disrupted, then the pioneer axons cannot cross the midline, and the hundreds of thousands or millions of follower axons that normally comprise the corpus callosum cannot either. Under normal circumstances, this whole process happens highly robustly. But mutations in diverse genes, well studied in both mice and humans, can result in failure of these processes and agenesis of the corpus callosum.

Figure 3.

Corpus callosum (CC) development. (A) (Top) A section through an adult mouse brain shows the cerebral cortex covering the two hemispheres and the CC connecting them. (Bottom left) The stages of normal CC development. (1) Midline cells fuse and form a bridge between the two hemispheres. (2) Pioneer axons cross. (3) Follower axons cross. (Bottom right) When the midline cells fail to fuse, pioneer and follower axons fail to cross, resulting in absence of the CC. (B) In some mouse strains, a proportion of animals end up with no CC, despite every animal being genetically identical. This probabilistic effect is inherited regardless of the phenotype of the parent. (Reprinted, with permission, from Mitchell 2018b.)

The contingent nature of these processes—which make them highly sensitive to noise at an early stage—results in a (more or less) bimodal distribution of phenotypes across genetically identical individuals (e.g., Ruge and Newland 1996). This has been well studied in mice, where different inbred (and thus isogenic) lines have different frequencies of callosal agenesis—this structure either forms fairly normally or is completely absent (Wahlsten et al. 2006). This indicates that what is inherited is a certain risk or probability of this phenotype arising, but that the particular outcome depends on the way stochastic developmental processes play out (Fig. 3). There are clear parallels to the inheritance of risk for neuropsychiatric conditions, such as schizophrenia or epilepsy, where even monozygotic twins may differ in clinical manifestation.

These examples show the ubiquity of developmental variation and its importance as a source of phenotypic variance, directly visible at the level of neuroanatomical phenotypes. It is not surprising that such differences in neural circuits can be correlated with behavioral differences, as demonstrated in flies (Linneweber et al. 2020) and humans (Ruge and Newland 1996).

In humans, it seems highly likely that stochastic developmental variation is the source of much of what is called the “non-shared environment” (Mitchell 2018a; Tikhodeyev and Shcherbakova 2019). This is the component of phenotypic variance identified from twin and family studies that is attributable to neither genetics nor the family environment. (It is most readily identified as the source of residual differences between identical twins reared together, but, of course, is a source of variance that affects everyone.) Researchers in behavioral genetics have often ascribed such variance to nonsystematic environmental exposures or idiosyncratic experiences (Plomin 2011) (without much luck in identifying what those might be). An alternative explanation is that this variance is not “environmental” at all, but inherent to the processes of development themselves.

This view is consistent with the interpretation mentioned above of the findings from GWAS of psychiatric conditions. The very general enrichment for neurodevelopmental functions among the genes with “hits” from such studies, rather than any convergence on specific functions, is consistent with the idea that an increasing polygenic burden of such variants decreases developmental robustness (Fig. 4). Individuals with a high polygenic burden are thus less able to buffer the effects of new mutations (Mitchell 2015, 2018b), as observed in numerous studies (Niemi et al. 2018; Bergen et al. 2019; Antaki et al. 2022; Cirnigliaro et al. 2023). Thus, not only is developmental robustness a genetic trait, it is a clinically important one.

Figure 4.

Genetic effects on neural development. Possible effects of rare mutations can be buffered by the distributed robustness of developmental systems. However, this robustness is itself a genetic trait and can be degraded by increasing polygenic load (of other rare mutations or many common genetic variants). Robust genotypes may buffer the effects of rare mutations and produce more optimal outcomes with low variance (across genetically identical individuals), while developmentally sensitive genotypes may produce more variable outcomes, including some with clinical consequences.

EVOLUTIONARY DYNAMICS

Understanding the relationship between genotypes and neural circuit phenotypes is essential to understanding how the latter can evolve. Thinking in reductive terms of specific proteins determining specific guidance or connectivity decisions in an isolatable way will naturally lead to the expectation that evolutionary changes in circuit organization should be traceable to specific mutations. There are, in fact, a handful of examples where known mutations affecting protein sequence (e.g., Robo3; Zelina et al. 2014) or expression (e.g., Zic2; Vigouroux et al. 2021 or SATB2; Paolino et al. 2020) are associated with differences in neural connectivity across species (related to midline crossing of various axonal populations).

Whether such mutations were the evolutionary origin of such differences is another matter and much more difficult to test. It is worth examining our expectations about such differences. In general, any new mutation that causes a drastic difference in neural connectivity in some individual is liable to be strongly selected against. It thus does not seem highly likely that single mutations will be solely responsible for large species differences in neural connectivity.

By contrast, the picture outlined above, where neural circuitry and concomitant behavioral phenotypes are affected by thousands of genetic variants across a population, is more conducive to a view of gradual evolutionary change, occurring due to selection or genetic drift as species diverge. Paradoxically, it is the robustness of the system, which clamps down phenotypic deviations in individuals, that is the very thing that allows such change to occur over longer timeframes (Wagner 2013; Hiesinger and Hassan 2018). This distributed robustness allows genetic variations to accumulate in the population while still generating phenotypes within a viable range. This generates a substrate of genetic variance on which evolution can act, gradually pushing traits along separate dimensions, without making any individuals radically different from their parents or their conspecifics.

Importantly, independent selection on distinct phenotypes can occur in an emergently modular fashion, even when all traits are highly polygenic and individual genetic variants are pleiotropic (Clune et al. 2013; Kouvaris et al. 2017; KJ Mitchell and N Cheney, in press). The contingent nature of development, outlined above, may even mean that qualitatively distinct outcomes can arise as a consequence of such polygenic selection. That is, novel phenotypes may be reachable even by gradual genetic change. In addition, in populations undergoing this kind of change, perhaps in a novel environment, new mutations may be much more likely to be beneficial if they reinforce the direction of phenotypic travel.

SUMMARY

The relationship between genotypes and phenotypes reflects the distributed, collective nature of the genomic encoding of neural connectivity and the probabilistic, statistical nature of the developmental processes that realize it. At a genomic level, rare mutations, which can have large individual effects, and the collective effects of thousands of common genetic variants, can influence the phenotype. The magnitude of any such effects is often buffered by the distributed robustness of the whole developmental system, but this robustness is itself a genetic trait and is degraded by increasing mutational load. Developmental variability is thus inevitable and can lead to dichotomous outcomes.

In turn, variation in neural architecture contributes to variation in psychological and behavioral traits across the population and to risk of neuropsychiatric disorders, both rare and common. It is thus of enormous clinical importance to understand both the sources and the consequences of such variation. Over much longer timeframes, the underlying genetic variation also provides the substrate for evolutionary change, with natural selection acting as the critic of any behavioral manifestations. Understanding the nature of the encoding of neural connectivity and the relationships between genotypes and phenotypes is thus essential to understanding how circuits evolve and how this underlies variation in species-typical behaviors.