The evolution of cortical development: the synapsid-diapsid divergence

The human cerebral cortex is probably the most-complex structure elaborated during biological evolution (Rakic, 2009). Essential to our unique cognitive abilities, it is the driver of our civilization, and arguably the sole instrument that may allow us to elaborate solutions to the impending anthropogenic ‘sixth extinction’ (Barnosky and Hadly, 2016). It is, therefore, quite natural to enquire about how our cortex evolved.

With the exception of amphioxus (considered the most primitive chordate), the forebrain is present in all vertebrates (Butler and Hodos, 2005). It is divided into the diencephalon (the caudal part of the forebrain) and telencephalon (see Glossary, Box 1), the dorsal part of which contains the cerebral cortex. The basal telencephalic ‘Bauplan’ is present in evolutionarily more-ancient vertebrates, such as lampreys, sharks and fishes. The main parts of the telencephalon were defined in amphibians by Herrick in his pioneering studies (Herrick, 1948), and consist of a mediobasal septum, a medial pallium, a dorsal pallium and a lateral pallium, and laterobasal areas that include ventral pallial areas, striatum and amygdala; the cortex develops in pallial areas (Laberge and Roth, 2007). However, the forebrain in amphibians does not contain a bona fide cortex, defined as a sheet of neurons that migrate radially from germinative periventricular zones and extend their axons inwards to form a hemispheric white matter. The cerebral cortex, as defined in this way, appears for the first time in amniotes (Butler and Hodos, 2005). In this review, I discuss the evolution of the cerebral cortex in amniotes. Given the interdisciplinary nature of this theme, this cannot be an exhaustive review and I refer the reader to other excellent sources for further information on cortical development (Florio and Huttner, 2014; Lein et al., 2017; Rakic, 2009), on comparative neuroanatomy (Aboitiz and Montiel, 2012; Butler and Hodos, 2005; Butler et al., 2011; Puelles et al., 2017) and on amniote evolution (Benton, 2015a; Carroll, 1988; Colbert et al., 2001; Rowe et al., 2011). Here, I focus primarily on the issue of how comparative studies of cortical development in present-day amniotes can help us understand better the molecular mechanisms that shape our cortex. First, I provide a brief overview of amniote evolution, before summarizing how the cortex is organized in synapsids and diapsids. I then discuss two key aspects of cortical development, neurogenesis and folding, and the similarities and differences in the two amniote groups. These discussions also highlight a number of key questions for future analysis.

It is believed that the closed egg of amniotes (also known as the cleidoic egg), which allowed embryos to develop in a terrestrial rather than an aquatic environment, evolved only once (Benton, 2015a; Ferner and Mess, 2011). Amniotes are monophyletic: all lineages, extinct and present, derive from a single ‘stem amniote’, the ancestor of present mammals, reptiles and birds. Although no fossil stem amniote has been unequivocally identified, they likely appeared circa 314 million years ago (MYA), in the late Carboniferous (Pennsylvanian, see simplified tree in Fig. 1), and presumably derived from advanced amphibians of the labyrinthodont type (Benton, 2015a; Colbert et al., 2001).

Two main lineages, the synapsids and sauropsids, diverged very early from stem amniotes: specimens of both lineages appear in the fossil record around the same time that amniote fossils appear (Benton, 2015a). The phylogenetic relationships of amniotes are largely assessed based on a few simple morphological traits of the skull, namely the localization and number of temporoparietal bony arches – structures that derive from the neural crest of the first branchial arch (Box 2). As yet, we do not understand why these traits are so invariant and consistent. Presumably, they reflect the existence of highly conserved genetic components of evolution that govern key aspects of the body plan.

The synapsid lineage (one arch, see Box 2) gave rise to premammals and mammals. Early synapsids are often referred to as ‘mammal-like reptiles’, a term that I will not use further, for reasons explained in Box 2. Because extinct premammals and all mammals, extinct and present, are synapsids, I refer to their common ancestor as the ‘stem synapsid’. The other lineage is the sauropsid lineage, also referred to as Reptilia (Benton, 2015a). This lineage encompasses anapsids, which have no temporal arch/opening and are represented today by chelonians (turtles), and the diapsids, which have two arches/openings and are represented by other reptiles and birds (Benton, 2015a). Modern diapsids include lepidosaurs and archosaurs. Lepidosaurs comprise the large group of squamates (lizards and snakes), and rhynchocephalia, a group of lizard-like reptiles that includes only two living species of Sphenodon (New Zealand's tuatara). Archosaurs are represented by crocodilians (the modern remnants of ornithischian dinosaurs), chelonians (Benton, 2015a) and also birds – derived from saurischian dinosaurs (Brusatte et al., 2015). It is generally assumed that stem amniotes, like their amphibian ancestors, had anapsid skulls, and anapsid chelonians were for a long time considered to reflect the ancestors of all sauropsids, and even perhaps the elusive stem amniote. However, DNA sequence data and the discovery of new fossils led to the revised view that chelonians are derived from archosaurs and belong to the diapsid lineage (Hedges and Poling, 1999; Schoch and Sues, 2016; Zardoya and Meyer, 2001). All living non-mammalian amniotes are therefore diapsids (at least in evolutionary terms), and I propose to simplify terminology and refer to their common ancestor as ‘stem diapsid’ rather than by the more complex and somewhat equivocal terms of ‘stem sauropsid’ or ‘stem reptile’.

An important, sometimes overlooked point is that early synapsids were evolutionarily highly successful during the Permian period (∼299-252 MYA). They were represented by pelycosaurs and therapsids (see Glossary, Box 1), which are well documented in the fossil record, particularly thanks to the superb strata in North America and Russia that contain pelycosaur fossils, and to the Karoo basin in South Africa and Ural region in Russia, where most therapsid specimens are found (Benton, 2015b; Kemp, 2006). Some pelycosaurs and therapsids were large animals and dominated the Permian terrestrial environment (Chinsamy-Turan, 2012), whereas diapsids were relatively small and less successful (Benton, 2015a; Carroll, 1988).

This situation lasted until about 252 MYA, when the Permian-Triassic (PT) mass extinction, the largest of the major extinction events (Benton, 2015b; Shen et al., 2011), wiped out almost all synapsids and most other species. The PT extinction, during which at least 80% of terrestrial vertebrates disappeared, was more dramatic than the later Cretaceous-Paleogene (K-Pg) extinction that led to the demise of dinosaurs 66 MYA. A few therapsids survived the PT extinction, and stem mammals appeared among them. Early mammals were diminutive, shrew-like insectivores that mostly hid under ground, which led to decreased visual performance and the loss – relative to diapsids – of opsins, leading to dichromatism (with a later duplication leading to trichromatism in Old World monkeys) (Gerkema et al., 2013; Rowe et al., 2011). Throughout the remainder of the Mesozoic era, premammals and mammals were minor members of the Jurassic and Cretaceous faunas, which were dominated by diapsid archosaurs and then specifically by dinosaurs. Although they were ecologically quite insignificant, mammals diversified extensively during this long period, and it is in the Cretaceous period that the main divisions of mammals became established (Bininda-Emonds et al., 2007; dos Reis et al., 2012). After the K-Pg extinction that wiped out dinosaurs, mammals became evolutionarily more successful than diapsids and underwent a great expansion at the beginning of the Cenozoic period, ‘the age of mammals’ (Benton, 2015a; Colbert et al., 2001; O'Leary et al., 2013). They progressively came to dominate almost all habitats, with extreme forms ranging from the diminutive bumblebee bat (Craseonycteris thonglongyai) and etruscan shrew (which weighs just a few grams), to blue whales.

It is worth emphasizing that, far from being evolutionarily recent, the primate line is descended from early placental mammals, when the main early placental lineages formed during late Cretaceous times, about 70 MYA. Primates and rodents diverged about 60 MYA, after the K-Pg extinction, when the variety of mammals exploded (Bininda-Emonds et al., 2007; dos Reis et al., 2012; O'Leary et al., 2013). This early divergence of lineages should be kept in mind when comparing cortical development among different mammals.

In all amniotes, the cerebral cortex develops in the telencephalon in an invariant manner. Therefore, a cerebral cortex was presumably present in stem amniotes and defines an evolutionary homology. There are different views about comparative telencephalic regionalization (Butler et al., 2011). Here, I use a simple version of the ‘tetrapartite pallium model’ (Puelles et al., 2016, 2017), in which the amniote cortex/pallium (see Glossary, Box 1) is divided into three dorsal fields and a ventral sector. The dorsal telencephalic fields include medial, dorsal and lateral components, all of which are present in modern mammals and diapsids (Fig. 2). There is general agreement that the medial cortex in diapsids corresponds to the mammalian hippocampal formation, the dorsal cortex to the neocortex (see Glossary, Box 1) and the lateral cortex to the piriform cortex (Puelles et al., 2017). Excitatory neurons in the medial and dorsal cortex are born in corresponding medial and dorsal ventricular zones (VZs; see Glossary, Box 1). The fate and lineage of neurons born in the ventral VZ is somewhat less clear, partly because they migrate along complex and highly convoluted pathways. In diapsids, neurons destined for the anterior dorsal ventricular ridge (ADVR, see Glossary, Box 1 and below) are generated in the ventral VZ; in mammals, this small ventral VZ may contribute to neurons in the insula, claustrum and amygdala.

Prominent among diapsid pallial/cortical areas is the large ventrolateral ADVR (Fig. 2B,C), which in birds corresponds to the nido- and mesopallium. As yet, the ADVR has not been identified in pre-amniotes. Because it is prominent in all modern diapsids, the ADVR is likely to have been present in stem diapsids. Whether a mammalian equivalent to the ADVR exists has been a matter of considerable debate and remains unresolved (Montiel et al., 2016; Puelles et al., 2017). The VZ of the diapsid ADVR expresses markers of the mammalian cortical VZ (Fernández et al., 1998), especially Pax6. In mammals, Pax6 expression is high in the dorsal cortical ventricular zones, and extends to a narrow zone ventral to the angle of the lateral ventricle, also known as the ‘antihem’ (Subramanian et al., 2009). This narrow zone may be the origin of the mammalian remnant of the ADVR (Fernández et al., 1998; Montiel et al., 2016; Puelles et al., 2016, 2017). From these findings, it is reasonable to propose that the ADVR appeared in stem amniotes (Fig. 2D), then enlarged in diapsids and became less prominent in synapsids.

In line with paleontological views and with the early divergence of the synapsid and diapsid lineages, comparisons of the telencephalon among mammals, turtles, squamates, Sphenodon, crocodiles and birds show that, apart from the conserved features mentioned above, the telencephalon of all diapsids studied to date is very different to that in mammals. Despite important differences among diapsid cortices, they are clearly more similar to one another than they are to any mammalian cortex (Butler and Hodos, 2005). Comparative studies of embryonic brains have emphasized this even further (Goffinet, 1983; Goffinet et al., 1986). The mammalian piriform cortex is perhaps the only area that resembles its diapsid homolog, both in terms of relative size and connectivity with the olfactory bulb. In monotremes, marsupials and eutherian mammals, the hippocampus is much more developed than it is in any diapsid, in which it is reduced to a simple cortical layer with no definite morphological evidence of CA (‘cornu ammonis’) fields and no anatomically defined dentate gyrus. The diapsid dorsal cortex contains a single layer of pyramidal cells and interneurons, and the large pallium in birds is organized in nuclei rather than in layers (Dugas-Ford et al., 2012). By comparison, even lissencephalic mammals (see below and Glossary, Box 1) have a thick cortex with the classical neuronal layers overlaid by a marginal zone (see Glossary, Box 1). Mammalian cortical layers 6 to 2 develop in sequence, from inside to outside (Rakic, 2009). In contrast, cortical layering is rudimentary in diapsids and, with the possible exception of the turtle dorsal cortex (Xi et al., 2008), the cortical plate (see Glossary, Box 1) develops either from outside to inside or with no preferential gradient at all (Goffinet et al., 1986). The neocortex is also widely expanded tangentially in monotremes, marsupials and eutherian mammals, and is partitioned into many more areas than in diapsids. Conversely, as mentioned above, the large ADVR of diapsids – which also matures from outside to inside – is very diminutive in mammals (Fernández et al., 1998; Montiel et al., 2016; Puelles et al., 2016, 2017).

The comparative data on cortical regionalization summarized above are clearly in line with the fundamental division of amniotes into synapsid/diapsid lineages. In the next sections, I will consider the cellular events and processes that regulate cortical development irrespective of regionalization.

Mammalian cortical development is best documented in mice and primates (Florio and Huttner, 2014; Lui et al., 2011; Molnár and Clowry, 2012; Rakic, 2009). It begins with neurogenesis (Fig. 3A), during which different neurons form from neural progenitor cells (NPCs) located around lateral ventricles, in the VZ (Fig. 3A), and later in the VZ and in subventricular zones (SVZs) (Fig. 3B,C). Early born neurons migrate through the intermediate zone (IZ); they initially form a loose horizontal plexus called the preplate (PP) (Fig. 3A) and later the denser cortical plate (CP). With the appearance of the CP, early PP neurons settle in the marginal zone to form pioneer layer 1 neurons (distinct from to Cajal-Retzius cells; see Glossary, Box 1 for definitions of both) and under the CP, in the subplate (future layer 6b) (Hoerder-Suabedissen and Molnár, 2015) (Fig. 3B). Neurons are deposited in the CP from inside to outside, first forming layers 6 and 5 (populated by deep layer neurons), followed by layers 4, 3 and 2 (upper layer neurons) (Fig. 3C). Most (at least 80%) cortical neurons are excitatory glutamatergic cells generated from cortical NPCs and they migrate radially to the cortex. The others are GABAergic cortical interneurons generated in the ganglionic eminences in the ventral telencephalon that reach the cortex by tangential migration (Anderson et al., 1997; Lavdas et al., 1999; Métin et al., 2007). CP neurons are radially ordered and this requires reelin signaling (see Glossary, Box 1) (Tissir and Goffinet, 2003). Near the end of migration, neurons extend dendrites and axons to initiate wiring; synaptogenesis occurs later, mostly after birth in most mammals.

Below, I will discuss briefly neurogenesis, neuronal migration, formation of PP, and CP and reelin signaling in mammals and diapsids. As will become apparent, some features are similar in all species examined, likely inherited from stem amniotes and regulated by conserved genetic networks, whereas others are absent or poorly defined in diapsids and probably specific to the mammalian brain.

In all mammals, neural progenitor cells (NPCs) are initially present as radial neuroepithelial (NE) cells that span the whole thickness of the neural tube (Fig. 3A). Interphase NE cells are attached together along the lateral ventricle by a junctional belt that surrounds their apical domain, and form end feet that contact the basal lamina at their pial pole. They divide symmetrically to amplify the progenitor pool (Fig. 3A). During the cell cycle, the nuclei of NPCs move such that S phase occurs at a distance from the ventricle, whereas M phase occurs along the ventricle where cells retain their junctional belt. The movement of the nucleus with the mitotic cycle, initially described as ‘to and fro’ (Sauer, 1935), is now referred to as ‘interkinetic nuclear migration’ (Miyata et al., 2014). At the onset of neurogenesis, NPCs modify their transcriptional profile and acquire markers of glia, such as brain lipid binding protein (BLBP). They are therefore labeled apical radial glia (RGs) or ventricular RGs. Like NE progenitors, aRGs are in contact with the ventricle as well as the pia. As they both have apical domains, NE cells and aRGs are often grouped together as apical progenitors (APs). aRGs can divide symmetrically, but also asymmetrically to give rise to an aRG and a neuron or an intermediate progenitor (IP) (Englund et al., 2005; Kowalczyk et al., 2009; Pontious et al., 2008). Early postmitotic neurons are identified morphologically and by their expression of Tbr1 (T-box brain gene 1). IP cells lose all attachment to the ventricle and the pia, and settle in the SVZ (Fig. 4A). They express the transcription factor Eomes1 (also known as Tbr2) and several neuronal genes, albeit at a lower level than do postmitotic neurons, and are committed to a neuronal fate. IP cells are thought to undergo a few (number not known) symmetric divisions, before undergoing terminal division into neurons (Kowalczyk et al., 2009; Pontious et al., 2008). The transition from radial glial NPC to IP is inhibited by FGF (Kang et al., 2009).

In addition to APs and IPs, some progenitors that resemble aRGs keep their attachment to the pia but lose their ventricular attachment and their apical domain. They are named basal (or outer) radial glia (bRGs) and are particularly abundant in the gyrencephalic cortices (see Glossary, Box 1) of ferret, macaque and human (Fietz et al., 2010; Hansen et al., 2010; Reillo et al., 2011). Cell bodies of bRGs are located in the SVZ, and are especially abundant in its outer component (the OSVZ), a layer that is barely present in mice but prominent in macaque and humans (Fietz et al., 2010; Smart et al., 2002). It is worth noting that the SVZ, OSVZ and bRGs are also present but much less prominent in marmoset monkeys that almost lacks cortical folds (Kelava et al., 2012). Like aRGs, bRGs are able to self-renew, as well as to perform asymmetric neurogenic divisions (Betizeau et al., 2013; Martínez-Cerdeño et al., 2016; Pilz et al., 2013; Shitamukai et al., 2011; Wang et al., 2011), and to generate IP cells that then divide into neurons (LaMonica et al., 2012).

A lingering unresolved ssue concerns the relative importance of the apical and basal domains of APs for the maintenance of NPC properties (Fig. 4). The apical domain of APs (and of adult NPCs) (Han and Alvarez-Buylla, 2010) features a monocilium that is ideally positioned to sense signaling molecules in the cerebrospinal fluid (CSF) (Lehtinen and Walsh, 2011; Lehtinen et al., 2011). However, studies of aRGs upon loss of their centrioles show that the apical domain is dispensable for the preservation of stem cell properties (Insolera et al., 2014). It should also be noted that basal progenitors, which do not contact the CSF, also carry a monocilium – on their basolateral membrane (Paridaen et al., 2013; Wilsch-Bräuninger et al., 2012).

Whereas the apical domain of NPCs may not be essential for their maintenance, all NPCs, including adult NPCs in the SVZ and in the subgranular layer (SGL) of the dentate gyrus (Goldberg and Hirschi, 2009; Mirzadeh et al., 2008) contact a basal lamina that is present at the pial surface and around brain endothelial cells; the latter contribute to the neural stem cell niche (Goldberg and Hirschi, 2009; Lange et al., 2016). Factors from the meninges (see Glossary, Box 1), such as retinoic acid and bone morphogenetic proteins (BMPs) (Choe et al., 2012; Siegenthaler et al., 2009), may regulate NPCs, whereas contact of NPC end-feet with the basal lamina that surrounds the meninges may favor the regulated exchange of signals with mesenchymal tissue (Boucherie et al., 2017). Conversely, the detachment of aRGs from the basal lamina does not suppress their stem cell potential, suggesting that this attachment is not necessary for maintaining stem cell properties (Haubst et al., 2006). This latter study showed that detachment of aRGs from basal lamina is accompanied by discontinuities in the external limiting membrane (composed of the basal lamina plus abutting end-feet of aRGs), a condition that increases communication between neural tissue and meningeal space. Interestingly, this is associated with the increased formation of deep-layer neurons from aRGs, and the decreased formation of upper layer neurons, accompanied with reduced formation of IP cells (Myshrall et al., 2012; Zhou et al., 2006). If either contact to ventricle or attachment to basal lamina is required to maintain NPC properties, then the combined loss of the apical domain and of basal lamina contact – while the limiting membrane remains intact – should deprive NPCs from external signals (from the choroid epithelium via ventricles and from meninges) and thereby trigger their differentiation into IP cells and neurons. This would be reminiscent of the situation during neural induction, when neural tissue formation proceeds by default upon neutralization of ectoderm-promoting factors such as BMPs (Munoz-Sanjuan and Brivanlou, 2002). Perhaps the production of cortical neurons might likewise proceed by default, when NPCs are isolated from other tissues, as shown in vitro (Gaspard et al., 2008).

Compared with mammals, much less is known about cortical neurogenesis in diapsids (Goffinet, 1983; Goffinet et al., 1986; Medina and Abellán, 2009; Montiel and Molnár, 2013; Montiel et al., 2016; Rakic, 2009). As in mammals, NEs and, later, aRGs are the predominant neural progenitor cells, but their number is much lower. In birds, germinative layers in some telencephalic areas are thick and a SVZ is present (Cheung et al., 2007; Montiel and Aboitiz, 2015). A study of the distribution of T-box brain gene 1 (Tbr1)- and Tbr2-expressing cells in the chick shows that Tbr1 labels early neurons and that Tbr2-positive, possibly IP, cells are present in that species (Bulfone et al., 1999). Neurogenesis was recently studied in the gecko and results conform with predictions from data in mammals (Nomura et al., 2013). Specifically, NE cells initially divide symmetrically, with a longer cell cycle than in mammalian NPCs (around 50 h versus 8-18 h in mammals), and then asymmetrically to give rise to neurons. No IP cells were detected directly outside of the VZ. A few Tbr2-positive cells could be detected outside the VZ, but were not labeled by a BrdU pulse; conversely, clonal analysis yielded a few clones with two neurons, indicating that some IP cells were present that had committed to the neuronal lineage (Nomura et al., 2013). Tbr2-positive cells that have delaminated from the VZ have been observed in the dorsal telencephalon in reptiles and birds (Cheung et al., 2007; Martínez-Cerdeño et al., 2016). It should be recognized, however, that Tbr2 might not specifically label IP cells in marsupials (Puzzolo and Mallamaci, 2010), and might be even less specific to IP cells in diapsids. Thus, although several key cellular features of neurogenesis are conserved between synapsids and diapsids, both lineages differ quantitatively and massively in the numbers of NPC, and probably quantitatively by the presence of a mammalian specific IP amplification mechanism.

Contrary to embryonic cortical neurogenesis, adult neurogenesis is more active in diapsids than in mammals. With the exception of the olfactory bulb and dentate gyrus, few if any neurons are produced in telencephalic areas in adult mammals, at least under physiological conditions (reviewed by Ming and Song, 2011). In contrast, forebrain neurons are generated continuously in adult diapsids, such as crocodiles or lizards, the bodies of which grow during their whole life (Garcia-Verdugo et al., 2002), and also periodically in birds (Alvarez-Buylla and Kirn, 1997).

The importance of NPC proliferation for cortical development is underpinned by the identification of genes that are implicated in human hereditary primary microcephaly (MCPH) and related disorders (see Table 1 and references therein). MCPH is characterized by a small, yet reasonably well formed and folded, cortex, as if built to scale. Several causal genes for MCPH have been identified and the number is regularly increasing. Of particular relevance are the genes that code for the centromeric proteins that regulate chromosome attachment to the mitotic spindle, as well as those that encode centrosomal proteins important for the integrity and function of centrosomes and cilia, and proteins implicated in spindle formation and cell division. Presumably, mutations that impair NPC proliferation and that favor the premature formation of neurons, lead to a reduction in the NPC pool and thus to cortical atrophy. The relative preservation of cortical folding (discussed further below) in MCPH suggests that mutant apical progenitors are still able to produce basal progenitors (bRGs or IP cells, or both) in nearly normal proportion, so that the ratio between apical and basal progenitors remains unaffected, enabling cortical folding to proceed. Human samples of microcephaly are rarely examined pathologically, and studies of mature brains can only yield indirect information about development. If the model above is correct, the inactivation of microcephaly genes in a species with a foliated cortex, such as the ferret or macaque, should mimic the human condition and allow us to understand the contribution of the different NPC types to shaping the pattern and extent of cortical folding. Gene editing technologies such as TALEN and especially CRISPR/Cas have already been applied successfully in ferret (Kou et al., 2015) and monkey (Niu et al., 2014), and should make this possible.

Differential regulation of cortical neurogenesis in synapsids versus diapsids represents only one mechanism that can help to account for the changes in cortical structure across evolution. Next, I address the development of early postmitotic neurons, focusing on neuronal migration, especially radial migration, and on reelin signaling.

In all amniotes where it has been studied, glutamatergic excitatory neurons are generated in cortical ventricular and subventricular zones and migrate radially along the expansions of aRGs to form the cortical plate (Goffinet, 1983; Rakic, 2009). In contrast, GABAergic neurons formed in the ganglionic eminences (LGE and MGE; see Fig. 2A) in the basal telencephalon reach the cortex by tangential migration (Anderson et al., 1997; Cobos et al., 2001; Medina and Abellán, 2009; Métin et al., 2007; Tuorto et al., 2003). There is therefore a likely conservation of mechanisms that regulate migration in all amniotes. Neuronal migration is orchestrated by intrinsic (cell-autonomous), as well as by extrinsic, cues provided by diffusible signals or extracellular matrix molecules (Lambert de Rouvroit and Goffinet, 2001; Manzini and Walsh, 2011; Métin et al., 2006). A detailed account of the cellular and molecular mechanisms of neuronal migration lies beyond the scope of this Review (I refer the reader instead to Tan and Shi, 2013), but I discuss below some key insights that would have been difficult to gain from in vitro cell biological investigations alone and have emerged from studies of human genetic disorders and mutant mouse models.

Like axonal growth cones, the leading process of migrating neurons progresses mainly by actin treadmilling. This is coupled to microtubule dynamics that regulate the progression of the nucleus in the leading process (‘nucleokinesis’). Retraction of the trailing edge at the rear of migrating neurons requires actomyosin contractility, which implicates myosin 2 (reviewed by Vallee et al., 2009).

When actin microfilament dynamics are perturbed, neuronal migration is drastically impaired, sometimes barely initiated, as best exemplified by the presence of periventricular heterotopias (see Glossary, Box 1) in females heterozygous for mutations in the X-linked gene encoding filamin A (Fox et al., 1998) and by the presence of large subcortical band heterotopia (SBH) in mice with inactivation of the small GTPase Rho, an actin regulator (Cappello et al., 2012). Similarly, individuals with Baraitser-Winter syndromes, which are caused by heterozygous mutation in actin genes (Table 1), feature pachygyria (see Glossary, Box 1), lissencephaly or heterotopias. The scarcity of pathogenic mutations in humans of genes implicated in actin dynamics could be due to gene redundancy and/or to the aggressive resulting phenotypes, many of which might be lethal.

Most of the genes implicated in human and in mouse neuronal migration disorders are linked to microtubule organization and function, as shown by studies of congenital lissencephaly and pachygyria. Lissencephaly type 1 is characterized by a near complete absence of cortical folds, with thickened cortical ribbon and poorly defined lamination (Reiner et al., 1993). A breakthrough in the field came with the realization that some of the genes that cause lissencephaly are orthologous to nuclear distribution (NUD) genes that encode microtubule-interacting proteins required for the regular distribution of nuclei in the syncytium in filamentous fungi (Xiang et al., 1995). Inactivation of the mammalian orthologs of NUD genes in mice preferentially impacts nucleokinesis, the progression of the nucleus in the leading process (Morris et al., 1998; Walsh and Goffinet, 2000). Nucleokinesis is defective in children and/or mice with mutations in genes encoding tubulin α1, the adaptor Nde1, the kinase Cdk5, and its co-factors p25 and p35, the microtubule-associated proteins lissencephaly Lis1 and doublecortin, the microtubule-severing kinesin Kif2a, and the nuclear envelope proteins SYNE and others (see Table 1 for a more complete list and references). Given that the genes mutated in lissencephaly encode proteins that influence microtubule dynamics, it is not surprising that these proteins regulate NPC division and spindle organization as well as nucleokinesis. Perhaps the most surprising finding is that some genes implicated in microtubule dynamics, such as Cdk5 (Gilmore et al., 1998), affect nucleokinesis and neuronal migration with little impact on NPC proliferation, whereas others, such as Wdr62 and Aspm impact NPC proliferation to a greater extent (Jayaraman et al., 2016). For genes such as Cdk5, their expression in neurons and not in NPC explains this specificity, but this trivial explanation is not always valid. These different phenotypes likely reflect different roles for microtubule-associated proteins in NPC division and nucleokinesis, and could provide new avenues of investigation into the underlying mechanisms which, to my knowledge, are not fully understood.

‘Doublecortex’ is a human malformation caused by mutations in the X-linked gene doublecortin, which encodes a microtubule-associated protein (des Portes et al., 1998; Gleeson et al., 1998). In males, mutations generate lissencephaly type I, whereas heterozygous females have a complex phenotype of SBH overlaid with a thin but normally organized cortex. These phenotypes are interpreted in terms of defective neuronal migration. The defect would affect all cortical neurons in males, and of about half of them in females (due to random X inactivation) (des Portes et al., 1998; Gleeson et al., 1998). SBH can also result from inactivation of autosomal genes, in which case the phenotype is attributed to defective migration of late-generated neurons due to a disorganized glial radial scaffold and aRG delamination. As noted above, large SBH were first induced upon inactivation of small GTPAse Rho in the mouse cortex, and were shown to result from gene inactivation in RGs, rather than in migrating neurons (Cappello et al., 2012). SBH were observed in humans and mice with mutations of Eml1 that encodes echinoderm microtubule-associated protein like 1 (Eml1). In mouse NPCs that lack Eml1, spindle orientation is abnormal and mutant NPCs loose attachment and delaminate from the VZ (Kielar et al., 2014). Intriguingly, although Eml1 is not X-linked like doublecortin, the phenotype is very reminiscent of the doublecortex malformation, and the thin cortical plate that overlays the heterotopic band is normally organized. The Eml1 mutant SBH contains more late than early generated neurons, and mutant aRG extensions are disorganized. Presumably, Eml1 inactivation affects primarily external processes of aRGs, which are more important for long-distance migration of late-generated cells than for short-distance migration of early-born neurons (Kielar et al., 2014). Rather similar phenotypes have been described in mice mutant for Mllt4 (Afdn) which encodes afadin, a protein that connects nectin to the actin cytoskeleton at adherens junctions (Gil-Sanz et al., 2014; Yamamoto et al., 2015), for Rap1, a regulator of the actin network (Shah et al., 2016), and for Llgl1 (lethal giant larvae homolog 1), which encodes a polarity protein expressed in NPCs (Jossin et al., 2017). In all those cases, a likely primary mechanism is the delamination of aRGs, leading to neuronal ectopia (Cappello et al., 2012; Gil-Sanz et al., 2014; Shah et al., 2016; Yamamoto et al., 2015). These recent findings hint at new features of neuronal migration that remain to be investigated further.

The mechanisms of neuronal migration in the diapsid cortex have yet to be studied in detail. Nevertheless, available morphological data show that, in turtle, lizard and crocodile embryos, cortical neurons move radially or tangentially, as in mammals (Garcia-Verdugo et al., 2002; Goffinet, 1983; Goffinet et al., 1986; Nomura et al., 2013), suggesting a conservation of mechanisms. However, the diapsid cortex is much thinner and less complex than its mammalian counterpart (Butler and Hodos, 2005; Butler et al., 2011; Puelles et al., 2017). As actin and microtubule dynamics, and radial glial guidance, are particularly important for long-distance migration and for the formation of mammalian upper cortical layers, these cellular mechanisms are likely to prove less crucial to cortical development in diapsids than in mammals.

A key difference between mammalian and diapsid cortices is the number and properties of Cajal-Retzius (C-R) cells (Fig. 4). In mammals, C-R cells are early-born transient neurons that originate from VZ at the frontier between ventral and dorsal telencephalic regions, and then in larger numbers from the dorsal cortical hem (Meyer et al., 2000). From their multiple sites of origin, C-R cells migrate tangentially in the subpial tier of the marginal zone (MZ), over all cortical sectors (Bielle et al., 2005; Meyer and Goffinet, 1998). They disappear progressively by apoptosis during late cortical maturation and only a few remain in the adult marginal zone (Meyer et al., 1999). They are characterized by the secretion of reelin (D'Arcangelo et al., 1995), a large extracellular protein that binds to lipoprotein receptors on the surface of end-migration neurons and activates a signal that orchestrates the radial organization of the cortical plate and cortical folding (Hong et al., 2000; Tissir and Goffinet, 2003). Turtles, lizards, crocodiles and birds have a few reelin-positive cells in the embryonic telencephalic marginal zone. However, those cells are scarce and the reelin signal individually associated with them is very weak compared with that in mammalian C-R cells (Fig. 4) (Bar et al., 2000; Bernier et al., 2000, 1999; Goffinet et al., 1999; Tissir et al., 2003). Lineage-tracing studies in quail indicate that avian reelin-positive cells originate from medial cortical fields, as in mice, but do not migrate from the anti-hem (Nomura et al., 2008). Therefore, the prominence of C-R cells in the embryonic MZ, and the resulting secretion of high amounts of reelin, is a defining feature of the mammalian cortex. It is impossible to know whether C-R cells were present in the early synapsid lineages (pelycosaurs and therapsids), but it should be possible to assess whether they are present in the embryonic cortex of monotremes (echidna and platypus), which are evolutionarily related to early mammals.

Altogether, available evidence suggests that similar cellular mechanisms orchestrate neuronal migration in mammals and diapsids, and that differences are more quantitative than qualitative. On the other hand, the high number of reelin-producing C-R cells and their production of Reelin in the cortical marginal zone are key mammalian specific features and reflect qualitative differences between synapsids and diapsids. Next, I will consider how evolutionary and developmental similarities and differences between mammals and non mammals relate to the complex issue of cortical folding.

The most remarkable difference between the cortex of mammals and diapsids is the propensity of the mammalian cortex to fold. This has been emphasized by others (De Juan Romero and Borrell, 2015; Lewitus et al., 2014; Molnár and Clowry, 2012; Nonaka-Kinoshita et al., 2013; Striedter et al., 2015) and can be appreciated by consulting the superb database ‘Comparative Mammalian Brain Collection’ (www.brainmuseum.org/), on which most of statements below are based. Cortical folding is extreme in humans, apes, cetaceans and elephants, intermediate in many mammals, and absent in lissencephalic mammals, most of which have small body sizes. What is striking is that some cortical folding is present in species from all mammalian branches (Borrell and Reillo, 2012; Fernández et al., 2016; Lewitus et al., 2014; Reillo et al., 2011). For example, although most rodents are lissencephalic, porcupines have some cortical gyri and the cortex of the largest rodent, the capybara, is quite folded. Conversely, even in the primate lineage, where most species have a folded cortex, a small species such as the marmoset has a nearly smooth cortex, with only a shallow sylvian sulcus (Garcia-Moreno et al., 2012). Most marsupials, such as the opossum or koala, have a smooth cortex, but in the kangaroo, it is folded, although not extensively. Even in monotremes, platypus is lissencephalic, whereas echidnas have a highly gyrated cortex (Ashwell and Hardman, 2012). A correlation exists between folding and body size in a given mammalian phylum but not across phyla. In contrast, the brains of all diapsids, including the largest ones, such as crocodiles, monitor lizards and ostriches, are completely smooth.

Thus, the development of a folded cortex occurred repeatedly in all mammalian lineages. Is this a homologous (synapomorphic; see Glossary, Box 1) or a homoplastic (see Glossary, Box 1) feature generated by evolutionary convergence? Did premammals have a smooth or a folded cortex? Some therapsids were large animals with a significant cranial cavity. However, the rare endocasts (see Glossary, Box 1) that have been obtained from therapsid skulls indicate that their telencephalon was sagittally elongated and mostly invested with olfactory projections, most likely lissencephalic (Kemp, 2006). As mentioned above, premammals were diminutive, mouse-sized animals or smaller. Computerized tomography (CT) scan reconstruction of their cranial cavity provides strong evidence that their cortex was smooth, as predicted from their very small size (Rowe et al., 2011). Conversely, some studies using cladistics methods conclude that mammalian ancestors were gyrencephalic (Fernández et al., 2016; Lewitus et al., 2014; O'Leary et al., 2013). Whether stem mammals were lissencephalic or gyrencephalic remains a matter of debate and definitive evidence may never be provided. As cortical folding can occur in all mammalian but none of the diapsid lineages, a reasonable assumption is that the conditions required for cortical gyrus formation were acquired early in the synapsid branch. If cortical folding in several mammalian lineage were homoplastic, due to evolutionary convergence, then I see no reason why some folding could not occur in some diapsids, particularly in birds, which generate large, primate-like numbers of neurons in their forebrain (Olkowicz et al., 2016), yet never develop well defined layers and folds.

I propose that cortical folding is best considered in terms of necessary conditions, bearing in mind that none of them is sufficient. Elegant models have been proposed to simulate cortical folding (Mota and Herculano-Houzel, 2015; Richman et al., 1975; Tallinen et al., 2014). Although they fall short of pointing to cellular mechanisms and experimental tests, they all underscore the importance of physical constraints. A first, quite obvious, condition for cortical folding is the generation of a sufficient number of neurons through the regulated proliferation of NPCs. To a first approximation, proliferation of NE and aRG progenitors is a key feature of tangential cortical spreading, whereas IP cells and bRGs contribute more to radial cortical thickness by forming preferentially upper layer neurons.

A second likely factor of folding is the relative rate of tangential expansion of deep and upper cortical layers. As mentioned above, several primary microcephalies in human with defective progenitor proliferation are associated with an elaborate cortical folding pattern (Friede, 1989), indicating that folding is possible with reduced numbers of neurons and hinting that the ratio of deep and upper cortical layers is relatively preserved in those malformations. Cortical size and folding is increased in mice by the expansion of basal NPCs (Florio et al., 2015; Nonaka-Kinoshita et al., 2013; Stahl et al., 2013) and gyrification can be induced by FGF2 (Rash et al., 2013). Folding of the mouse cortex can also be induced by the double inactivation of the genes Flrt1 and Fltr3, which encode adhesion proteins, indicating that finely tuned adhesion among young neurons is crucial to gyrus formation (Del Toro et al., 2017). Importantly, in these mouse models, gyrification is restricted to the cortical ribbon and thereby differs from the folding induced by the overexpression of β-catenin (Chenn and Walsh, 2002), the inactivation of GSK3 (Kim et al., 2009) or by the downregulation of apoptosis upon inactivation of caspases (Kuida et al., 1998). In the latter models, the whole telencephalic wall undergoes folding, including the VZ and SVZ, in addition to the CP. A note of caution about mouse models with increased gyrus formation concerns the possibility that it could be caused by death of deep layer neurons rather than expansion of NPCs. As in humans with micropolygyria (see Glossary, Box 1) (Friede, 1989), the death of deep layer neurons could result in an imbalance between the upper and lower layers, resulting in folding.

A third condition required for cortical folding is the migration of neurons to the cortical plate. As mentioned above, mutations in several genes that regulate cytoskeletal dynamics lead to defective folding and to lissencephaly (Table 1). A fourth condition for gyrus formation is the concept of intercalation (Reillo et al., 2011; Striedter et al., 2015); upper layer neurons need to migrate across deep cortical layers and to intercalate, and this is necessary for folding. Intercalation at the end of migration is controlled by the reelin pathway and by C-R cells, which are therefore a key feature of cortical folding, as demonstrated by the specific lissencephalic phenotype observed in individuals with defective reelin signaling (Hong et al., 2000).

A fundamental question is which of the variety of mechanisms mentioned above prevent gyrification in the diapsid brain? The lower numbers of cortical neurons in diapsids relative to mammals is certainly a key factor, although the forebrains of some birds contain neuron numbers comparable with those of primates, and this could be enough to trigger some folding (Olkowicz et al., 2016). As mentioned above, although basic features of NPC proliferation appear quite similar in mammals and diapsids, there are obvious quantitative differences in NPC numbers as well as probable qualitative differences in NPC variety, such as bRGs and the IP cell amplification mechanism. What could be the genetic basis for premammals being able to invest a large energy budget during development to foster a vast expansion of cortical NPCs, whereas highly successful diapsids, such as dinosaurs and birds, could not? Could there be a way experimentally to increase cortical neurogenesis in birds or reptiles and assess the consequences? Another key factor is reelin. Compared with mammals, reelin-expressing cells are very few and reelin production very low in the diapsid cortex (Bar et al., 2000; Meyer et al., 2000; Nomura et al., 2008). This might result in the defective intercalation of end-migration neurons into the cortical plate, and thus contribute to an inability to fold, even if high numbers of neurons are generated (Striedter et al., 2015). To increase reelin in the bird marginal zone, reelin-expressing COS7 cells were grafted into the marginal zone of the quail embryonic cortex. Contrary to predictions, this did not modify cortical architecture, but it increased the extensions of aRGs and their attachment to the pia (Nomura et al., 2008). To try and induce the formation of C-R cells in birds, Dbx1 – a gene expressed in mouse C-R cells from the antihem – was ectopically expressed by electroporation in the quail pallium, and this resulted in an increase in the number of reelin positive-cells reminiscent of an induction of C-R cells, and modifications of the morphology of aRGs and neuronal bipolar migration behavior, but not in fold formation (Nomura et al., 2008). These are encouraging results and the hypothesis that increasing C-R cell number and reelin secretion in the embryonic diapsid cortical marginal zone may promote cortical folding should be pursued further.

Over the past 25 years, many genes and proteins that regulate neurogenesis, neuronal migration and architectonic development have been identified in human and mouse, and we are beginning to understand their mechanisms of action. As a result, the time is probably ripe to broaden our scope and to compare in depth the molecular mechanisms of cortical development among mammals and diapsids. Such issues are fascinating and reach into the heart of biology and evolution. They are not new, as most of them were formulated decades ago. Thus far, however, they could be addressed only using descriptive morphological methods. Now, new technologies are emerging to allow studies of cortical development in diapsid embryos, including development of species-specific antibody reagents, single-cell RNA sequencing, gene editing, high-speed and high-resolution video-microscopy, and embryo and stem cell culture and induction. The application of these techniques should help us to address several key unresolved issues, such as the genetic origin of the basic differences in the synapsid and diapsid brain Bauplan, the evolutionary origin of the machinery for cortical folding, and the mechanisms that orchestrated the fast evolution of the human cortex.

I thank colleagues, especially Zoltan Molnar and Fadel Tissir, as well as anonymous reviewers and journal editors whose comments helped me focus some ideas.