INTRODUCTION
Approximately 70 years ago, two papers appeared in the psychiatric literature, describing “autistic disturbances of affective contact” (Kanner, Reference Kanner1943) and “autistic psychopathy” (Asperger, Reference Asperger1944). Asperger’s description was published in German during WWII, and did not receive attention in the English speaking world until Lorna Wing described it in Reference Wing1981 and Uta Frith translated it into English in 1991 (Asperger, Reference Asperger1991). Both syndromes are marked by poor social relationships and pragmatic language, obsession with specific interests and insistence on sameness. A third category in the DSM-IV, Pervasive Developmental Disorder-Not Otherwise Specified (PDD-NOS; APA, 2002) shares social impairments, but lacks the same degree of restricted, repetitive behaviors (RRBs) or shows milder deficits.
In the most recent diagnostic manual of the American Psychiatric Association (DSM-5, APA, 2013), these entities are combined into “autism spectrum disorder” (ASD), defined by three social communication symptoms (poor nonverbal communication, poor social-emotional reciprocity, deficits in developing and maintaining relationships), plus two of four RRBs (stereotyped or repetitive speech or movements, resistance to change, highly restricted interests, and hypo- or hyper-reactivity to sensory input). Severity of the syndrome and intellectual or language impairment are coded separately. With decades of research, including an accelerating pace and a focus on biology in recent years, one might expect some definitive answers about the syndrome’s clinical borders, pathophysiology, and etiology. Instead, enormous amounts of data have accumulated with little clarity emerging, which Rutter (Reference Rutter2014) justifiably finds “odd.”
There seem to be two fundamental obstacles to progress. First is a basic question about the coherence of the behavioral syndrome (explored by Hobson, Reference Hobson2014, and Rutter, Reference Rutter2014, in an issue of Autism [vol. 18] devoted to this topic, and Mandy and Skuse, Reference Mandy and Skuse2008). If there is no coherence, then looking for a unified etiology or pathophysiology is pointless. The second obstacle is the failure of findings at one level (e.g., genetic) to map onto features of another (e.g., phenotype) (Rapin, Reference Rapin2014; Waterhouse, Reference Waterhouse2013; Waterhouse, London, & Gillberg, Reference Waterhouse, London and Gillberg2016). Both questions raise doubts about even retaining the ASD diagnosis (Mueller & Amaral, Reference Mueller and Amaral2017; Waterhouse, London, & Gillberg, Reference Waterhouse, London and Gilberg2017). We will explore these two obstacles in Sections A and B and then review potential avenues for facilitating research progress in Section C.
COHERENCE OF THE SYNDROME
Until recently, no one questioned ASD as a valid diagnosis (Rutter, Reference Rutter2013) although criteria, boundaries, and etiologies were much debated and changed over time. A valid syndrome is “a naturally occurring combination of deficits…there is an underlying [causal] factor” (Benton, Reference Benton1961, p. 76).
Autism as such an entity, with a discoverable etiology and pathophysiology, is under question, as are other psychiatric and neurodevelopmental diagnoses (Cuthbert & Insel, Reference Cuthbert and Insel2013). The syndrome view assumes that social-communication deficits and RRBs co-occur at high rates, and bear some causal relationship. Recent theorists (Mandy and Skuse, Reference Mandy and Skuse2008; Ronald, Happe, and Plomin, Reference Ronald, Happe and Plomin2005) have argued that the two key features are not necessarily causally related and should be studied independently.
Four questions bearing on the coherence of autism will be briefly considered: (1) do the two defining traits co-occur in most cases; (2) is there substantial phenotypic overlap with other disorders; (3) how heterogeneous are autism phenotypes; and (4) how are theories about underlying causes shaped by the assumption of coherence?
Do Most Children Who Meet Social Criteria for Autism (as per DSM-IV, DSM-5, or ICD-10) also Have RRBs at Some Time in Their Development?
Wing and Gould (Reference Wing and Gould1979) reported that in a sample of 74 socially impaired children, all showed some degree of RRBs. However, others have identified PDD-NOS subgroups with low rates of RRBs (Brennan, Barton, Chen, Green, & Fein, Reference Brennan, Barton, Chen, Green and Fein2015; Walker et al., Reference Walker, Thompson, Zwaigenbaum, Goldberg, Bryson, Mahoney and Szatmari2004), although RRBs have variable ages of onset and might have appeared later (Barton, Robins, Jashar, Brennan, & Fein, Reference Barton, Robins, Jashar, Brennan and Fein2013; Stone et al., Reference Stone, Lee, Ashford, Brissie, Hepburn, Coonrod and Weiss1999). Factor analyses (Frazier, Georgiades, Bishop, & Hardan, Reference Frazier, Georgiades, Bishop and Hardan2014; Mandy, Charman, Puura, & Skuse, Reference Mandy, Charman, Puura and Skuse2014) suggest that a two-factor solution is the best fit, suggesting that RRBs and social-communication deficits are separable traits, although Constantino et al. (Reference Constantino, Gruber, Davis, Hayes, Passanante and Przybeck2004) found evidence for a single continuously distributed underlying factor.
Supporting the fractionation approach (studying the traits separately) is population based research suggesting that autism traits are associated with different genes (Happe & Frith, Reference Happé and Frith2006). However, the logic of separating these traits because they lack genetic linkage and phenotypic correlation in a general population may not be sound. For example, there are several genetic disorders associated with wide set eyes and intellectual disability (Wolf-Hirschhorn, Cri-du-Chat, Angelman syndromes, etc.). Presumably, these two traits would show little correlation in the general population but still co-occur in specific syndromes.
Fountain, Winter and Bearman (Reference Fountain, Winter and Bearman2012), studying almost 7000 ASD children, found that those who were high functioning or rapidly improving in one domain tended to be so in other domains, consistent with linked development. However, other studies (see Leekham, Prior, & Uljarevic, Reference Leekam, Prior and Uljarevic2011) suggest that social/communication and RRB symptoms can show different trajectories, consistent with fractionation. Alternatively, the traits may be linked, but diverge because one is more amenable to treatment.
This debate may be resolved by collecting large amounts of longitudinal data and reporting individual patterns of development. Instituting early treatment, however, may make this “natural history” difficult to observe. In the meantime, given the sparsity of data on this question, it seems unproductive to exclude from clinical or research activities children who show impaired social/communication, but have 0 or 1 RRB instead of the 2 required in DSM-5 (APA, 2013).
Phenotypic Overlap with Other Syndromes
Are RRBs necessary features of ASD, while other frequently comorbid symptoms are ancillary (as in DSM-5)? Comorbidities are the rule rather than the exception in ASD (e.g., Matthews, Reference Matthews2016), including intellectual disability (ID) (Blacher & Kasari, Reference Blacher and Kasari2016), psychiatric and medical disorders (Bauman, Reference Bauman2010; Simonoff et al., Reference Simonoff, Pickles, Charman, Chandler, Loucas and Baird2008), and Specific Language Impairment (Bishop, Reference Bishop2010). Furthermore, these disorders share genetic risk factors with ASD (De Rubeis and Buxbaum, Reference De Rubeis and Buxbaum2015; Elia et al., Reference Elia, Gai, Xie, Perin, Geiger, Glessner and Muganga2010). These and other behavioral comorbidities (e.g., self-injury, sleep problems, aggression) may be related to autism’s defining symptoms in many possible ways (Gillberg & Billstedt, Reference Gillberg and Billstedt2000).
Whether any of these “comorbidities” are as central to ASD as RRBs is obscured by evidence that young children are more likely to get an ASD diagnosis when they display comorbid symptoms (e.g., language disorder, intellectual disability, epilepsy) while “autism only” may be overrepresented in studies of older children (Gillberg & Fernell, Reference Gillberg and Fernell2014). To ascertain whether comorbidity profiles indicate distinct pathophysiology and/or the need for distinct treatment, it will be helpful to routinely characterize participants with respect to comorbidities.
Phenotypic Heterogeneity
Another major obstacle to conceptualizing ASD as a syndrome is its unusually high level of phenotypic variability in symptoms and cognition (Boucher, Reference Boucher2012; Georgiades et al., Reference Georgiades, Szatmari, Boyle, Hanna, Duku, Zwaigenbaum and Smith2013; Howlin, Goode, Hutton, & Rutter, Reference Howlin, Goode, Hutton and Rutter2004; Kjelgard & Tager-Flusberg, Reference Kjelgaard and Tager-Flusberg2001; Rommelse et al., Reference Rommelse, Langerak, van der Meer, de Bruijn, Staal, Oerlemans and Buitelaar2015; Waterhouse et al., Reference Waterhouse, London and Gillberg2016). Onset and course is also quite varied; symptoms may appear during the first year of life (Ozonoff et al., Reference Ozonoff, Iosif, Baguio, Cook, Hill, Hutman and Steinfeld2010; Zwaigenbaum et al., Reference Zwaigenbaum, Bryson, Rogers, Roberts, Brian and Szatmari2005) or as late as the third year (Ozonoff et al., Reference Ozonoff, Young, Landa, Brian, Bryson, Charman and Zwaigenbaum2015; Rogers, Reference Rogers2009); some children show regression in language and/or social attainments in the second year (Hansen et al., Reference Hansen, Ozonoff, Krakowiak, Angkustsiri, Jones, Deprey and Hertz-Picciotto2008). A minority of children will lose their diagnosis (Anderson, Liang, & Lord, Reference Anderson, Liang and Lord2014; Fein et al., Reference Fein, Barton, Eigsti, Kelley, Naigles, Schultz and Troyb2013; Helt et al., Reference Helt, Kelley, Kinsbourne, Pandey, Boorstein, Herbert and Fein2008), as will a higher proportion of children whose autism was due to environmental deprivation such as congenital blindness (Hobson & Lee, Reference Hobson and Lee2010; Hobson, Reference Hobson2014; Jure, Pogonza, & Rapin, Reference Jure, Pogonza and Rapin2016) or severe neglect (Rutter et al., Reference Rutter, Andersen-Wood, Beckett, Bredenkamp, Castle, Groothues and O’Connor1999). In some children, RRBs will remit (Esbensen, Seltzer, Lam, & Bodfish, Reference Esbensen, Seltzer, Lam and Bodfish2009), while other individuals will experience cognitive decline in adolescence or adulthood (Howlin, Reference Howlin2010).
Therefore, ASD may represent dozens, or even hundreds, of disorders with overlapping phenotypes, but distinct biological etiologies (e.g., Abrahams, Reference Abrahams2011; Rapin, Reference Rapin2014). Therefore, case-control studies classifying participants only in regard to whether or not ASD criteria are met, are likely reporting on groups of individuals with many different underlying biological conditions, perhaps as many as there are participants, limiting generalizability and replicability.
Implications for Theory
Mirroring the controversial independence of social and RRB symptoms, there are two classes of theory about underpinnings of autism traits: The first class account for one trait, usually the social impairment; these include poor intersubjectivity in early childhood (Hobson, Reference Hobson2014; Hobson and Meyer, Reference Hobson and Meyer2005), reduced social interest and abnormal salience of social stimuli in early childhood (Fein, Pennington, Markowitz, Braverman, & Waterhouse, Reference Fein, Pennington, Markowitz, Braverman and Waterhouse1986; Uddin & Menon, Reference Uddin and Menon2009; Waterhouse, Fein, and Modahl, Reference Waterhouse, Fein and Modahl1996), and an inability to understand others’ minds (Baron-Cohen, Reference Baron-Cohen1997). Fewer theories attempt to explain the trait of RRBs in terms of executive dysfunction or detail focus (South, Ozonoff, & Mcmahon, Reference South, Ozonoff and Mcmahon2007).
The second class are unifying theories that attempt to explain both the social communication and the RRBs; and implicitly accept the syndromic view of autism. These include the “Intense World” hypothesis of Markram and Markram (Reference Markram and Markram2010), in which stimuli experienced as painfully intense lead to selective attention and avoidance of social input. Similarly, Kinsbourne (Reference Kinsbourne2011) suggests that an unstable arousal system leads to avoidance of unpredictable social input, to restricted interests, and to inward focus of attention. This focus is associated with the default mode neural network (Menon & Uddin, Reference Menon and Uddin2010), which may not be normally deactivated by task demands in ASD (Kennedy, Redcay, & Courchesne, Reference Kennedy, Redcay and Courchesne2006). Impaired attention shifting can account for poor development of joint attention, resistance to change and perseverative interests (Harris, Courchesne, Townsend, Carper, & Lord, 1999). Another intriguing hypothesis (Gergely, Reference Gergely2003) suggests that at 0–3 months, the infant attends preferentially to perfectly contingent events, including visual input from his/her own movements, promoting the concept of self. At around 3 months, preference switches to imperfect contingency, attracting attention to social input. A failure to make this genetically programmed switch may lead to reduced social attention, stereotyped behavior, resistance to change, and preference for inanimate objects (e.g., Klin, Lin, Gorrindo, Ramsay, & Jones, Reference Klin, Lin, Gorrindo, Ramsay and Jones2009).
Syndrome coherence has implications for which class of autism theories are most tenable and for which individuals or subtypes.
Our own speculation is that ASD coheres at the behavioral (but not biological) level during a particular developmental window in the preschool years, representing a temporary final common pathway or “bottleneck” through which children with a wide range of genetic or environmental risk factors go (potentially due to missed experience-expectant social input resulting from each child’s vulnerability), only to emerge differently, depending on brain maturation, available treatment, and response to treatment. This view is consistent with higher diagnostic reliability and stability at this age, relative to older and younger ages (e.g., Barbaro & Dissanayake, Reference Barbaro and Dissanayake2017; Brian et al., Reference Brian, Bryson, Smith, Roberts, Roncadin, Szatmari and Zwaigenbaum2016). ASD secondary to environmental deprivation (severe, congenital blindness [Jure, Pogonza, & Rapin, Reference Jure, Pogonza and Rapin2016], or neglect [Rutter et al., Reference Rutter, Andersen-Wood, Beckett, Bredenkamp, Castle, Groothues and O’Connor1999]) shows more improvement in later childhood than idiopathic autism (Hobson & Bishop, Reference Hobson and Bishop2003; Rutter et al., Reference Rutter, Beckett, Castle, Colvert, Kreppner, Mehta and Sonuga- Barke2007), whereas children with Rett syndrome and childhood disintegrative disorder fit ASD criteria early on but show a deteriorating course. This conceptualization of ASD as a developmental bottleneck is similar to the view of Williams and Bowler (Reference Williams and Bowler2014) who argue for coherence at the behavioral level that may wax or wane with development.
Differences in behavioral coherence over development may also be a byproduct of flawed measurement (i.e., weak operational definitions, poor reliability) rather than a true phenomenon (Mandy et al., Reference Mandy, Charman, Puura and Skuse2014). DSM-5 provides little guidance on how to adjust for developmental level. When operational definitions are murky, measurement of behaviors will be unreliable (Lord et al., Reference Lord, Petkova, Hus, Gan, Lu, Martin and Risi2012).
LACK OF CORRESPONDENCE AMONG LEVELS OF EXPLANATION
The second major problem facing autism research is the striking failure of findings at the etiological or physiological level to be closely associated to phenotypic features. In this section, we will briefly consider (1) the relationships between genotype and phenotype in autism, and (2) the contribution of brain imaging data.
Genotype–Phenotype Relationships
Many techniques have been used (see Persico and Napolioni, Reference Persico and Napolioni2013) to identify rare or common autism vulnerability genes and variants, copy number variants (sections of the genome that are repeated, CNVs), linkage sites, or defined genetic syndromes with autistic features (e.g., Fragile X, tuberous sclerosis) (see reviews by Abrahams, Reference Abrahams2011; Bourgeron, Reference Bourgeron2016). Concordance for ASD between monozygotic (MZ) and dizygotic (DZ) twins, and the existence of broad autism phenotype (BAP) in siblings and parents (Bishop et al., Reference Bishop, Mayberry, Maley, Wong, Hill and Hallmayer2004; Piven, Palmer, Jacobi, Childless, & Arndt, Reference Piven, Palmer, Jacobi, Childless and Arndt1997) confirm the high heritability of autism (estimated at 90% by Rutter, Reference Rutter2013, although Bourgeron, Reference Bourgeron2016, finds it closer to 50%, with approximately 50% accounted for by non-shared environment).
However, the actual penetrance of autism-related genes, and the apparent complexity of gene–gene interactions and epigenetic influences, has not proved tractable. The Simons Foundation Autism Research Initiative (SFARI) Web site have large populations of affected families, and attempt to map phenotypic features onto established risk gene variants as well as CNVs. SFARI also lists over 800 genes, on virtually every chromosome, with 50+ ranked as “high confidence” or “strong candidate.” Rare mutations have been identified in synapse-related genes, and genome-wide association (GWA) studies using single nucleotide polymorphisms have highlighted potential ASD risk loci on multiple chromosomes (Chaste et al., Reference Chaste, Klei, Sanders, Hus, Murtha, Lowe and Geschwind2015), accounting for a small proportion of cases. CNVs, both de novo and familial, have been identified in 5–10% of individuals with ASD (Marshall et al., Reference Marshall, Noor and Vincent2008; Sebat et al., Reference Sebat, Lakshmi, Malhotra, Troge and Lese-Martin2007), which is a much higher rate than in the general population (1%), but not higher than in the ID population (Whibley et al., Reference Whibley, Plagnol, Tarpey, Abidi, Fulston, Choma and Raymond2010). Some of the newly discovered CNVs contribute to neuronal migration and synaptic function but their presence in some unaffected family members (approximately 3%) leaves their contribution to autism uncertain (Gilman et al., Reference Gilman, Iossifov, Levy, Ronemus, Wigler and Vitkup2011; Persico and Napolioni, Reference Persico and Napolioni2013).
In a landmark study, Le Couteur et al. (Reference Le Couteur, Bailey, Goode, Pickles, Gottesman, Robertson and Rutter1996) studied 28 MZ and 20 DZ twin pairs in whom at least one individual had autism; in the discordant pairs, especially the MZ pairs, the non-autistic twin often showed the broader phenotype. Amazingly, however, there was as much cognitive and language variation within concordant MZ pairs as between pairs, including up to 50 IQ point discrepancy; regression also fails to show concordance in multiplex families (Parr et al., Reference Parr, Le Couteur, Baird, Rutter, Pickles and Fombonne2011). The opposite problem has also been demonstrated: Yuen et al. (Reference Yuen, Thiruvahindrapuram, Merico, Walker, Tammimies, Hoang and Gazzellone2015) did a GWA study of 85 families with two children affected by autism, focusing on genes previously associated with autism. They found that approximately 70% of the sibling pairs did not share an ASD-relevant mutation, but rather had different ones. Since there are presumably many autism-related genes not yet identified, this concordance may be underestimated. Nevertheless, the aggregation of ASD diagnoses in families and the lack of demonstrable genetic concordance within families, suggests a contribution from the shared uterine environment (Hallmayer et al., Reference Hallmayer, Cleveland, Torres, Phillips, Cohen, Torigoe and Risch2011). Another interpretation is that a non-genetic factor or a de novo mutation (e.g., in an older father, or with viral infection, etc.) may push an individual into the clinical manifestation of autism in the presence of familial genetic risk (Gaugler et al., Reference Gaugler, Klei, Sanders, Bodea, Goldberg, Lee and Ripke2014).
One approach to this seemingly intractable problem is to increase phenotypic homogeneity in subtypes in an effort to increase linkage signal or significance of association in GWA studies. The success of these efforts thus far is debated. Subtyping along a single or a few broad features often increases genetic homogeneity (even if only modestly) (Chaste et al., Reference Chaste, Klei, Sanders, Hus, Murtha, Lowe and Geschwind2015; De Rubeis and Buxbaum, Reference De Rubeis and Buxbaum2015; Liu, Paterson, & Szatmari, Reference Liu, Paterson and Szatmari2008; Loviglio et al., Reference Loviglio, Leleu, Männik, Passeggeri, Giannuzzi, van der Werf and Reymond2017; Shao et al., Reference Shao, Cuccaro, Hauser, Raiford, Menold, Wolpert and Abramson2003). Chaste et al. (Reference Chaste, Klei, Sanders, Hus, Murtha, Lowe and Geschwind2015) report that in a GWA study, stratifying a group of 2576 individuals by IQ, insistence on sameness, extent of RRBs, and overall severity of ASD, led to only a modest increase in association significance or estimates of heritability; for both, the most successful subtyping was for severity of RRBs. Heritability in the overall sample was approximately .4 and reached approximately .6 in groups with high RRBs. Note: if RRBs reflect anxiety, especially in individuals who cannot verbalize their emotional state, the heritability might be for anxiety.
Linkage studies have arguably fared somewhat better than GWA studies, with subtyping on specific features improving linkage signal (e.g., Liu et al., Reference Liu, Paterson and Szatmari2008, studied 976 multiplex families from the Autism Genome Project consortium, and LOD scores increased when sorting for delayed onset of phrases; Spence et al., Reference Spence, Cantor, Chung, Kim, Geschwind and Alarcón2006, also found increased linkage for language delay but not on the same chromosomes). Of course, to be confirmed, all of these associations need to be replicated on independent samples. See Chaste et al. (Reference Chaste, Klei, Sanders, Hus, Murtha, Lowe and Geschwind2015) for further discussion of linkage and GWA studies in relation to phenotypic subtyping.
Given this rather modestly successful set of efforts, subtyping along more numerous and fine-grained features may or may not yield more etiologically homogenous subgroups; the potential for success, of course, will require very large samples. GWA studies to date may have lacked sufficient sample size, due to the weak effect of individual common genetic variations, and hence have been underpowered (De Rubeis & Buxbaum, Reference De Rubeis and Buxbaum2015). It is also possible that characteristics not considered central to autism, and therefore, not studied in large-scale genetic studies, such as improved behavior with high fever, might increase linkage; uncovering these relationships would necessitate rich phenotyping of all participants.
As described above, the overlap in phenotype and the shared genetic risk between autism and other neurodevelopmental disorders makes it even more difficult to identify genes specific to autism versus those that cause the common features. One assumes that the phenotype depends on interactions with other genes, epigenetic activation/suppression, imprinting, and the quality of pre and postnatal environmental conditions.
This complexity pervades the study of psychiatric and neurodevelopmental disorders. Indeed, central to their study are equifinality (syndrome coherence at the behavioral level despite multiple biological causes) and multifinality (a single cause can produce highly variable symptoms) (Williams & Bowler, Reference Williams and Bowler2014). This complexity may indicate that there are phenotype–genotype links that are simply out of our grasp with the current state of knowledge, data, and technique. Although difficult and expensive, large samples with rich phenotyping may be the only way forward.
Brain Imaging Data
The difficulty in relating genetic variants to symptoms also holds for the intermediate levels of brain imaging and post-mortem data. Autopsy and in vivo studies have indicated differences in the organization of mini columns in the neocortex (Casanova, El-Baz, Vanbogaert, Narahari, & Switala, Reference Casanova, El-Baz, Vanbogaert, Narahari and Switala2010), inflammatory processes (Pardo et al., Reference Pardo, Vargas and Zimmerman2005 but see non-replication in Pardo et al., Reference Pardo, Farmer, Thurm, Shebl, Ilieva, Kalra and Swedo2017), abnormalities in neurotransmitter balance (Lam, Aman, & Arnold, Reference Lam, Aman and Arnold2006), and abnormal connectivity (Belmonte et al., Reference Belmonte, Allen, Beckel-Mitchener, Boulanger, Carper and Webb2004) but details vary substantially. It is unclear whether inconsistent findings are due to methodological differences, developmental stage, or individual variation (Waterhouse et al., Reference Waterhouse, London and Gillberg2016).
One somewhat consistent anatomical finding is accelerated brain growth in the first 2 years of life (e.g., Courchesne, Carper, & Akshoomoff, Reference Courchesne, Carper and Akshoomoff2003; Hazlett et al., Reference Hazlett, Gu, Munsell, Kim, Styner, Wolff and Collins2005). We compared medical records for head growth in a group of high functioning autistic individuals and a group who moved off the spectrum; we found accelerated head growth from 12 to 24 months in both groups, failing to correlate an important outcome variation with a biological marker (Mraz, Dixon, Dumont-Mathieu, & Fein, Reference Mraz, Dixon, Dumont-Mathieu and Fein2009). However, head growth may be too coarse a measure; a study of high risk infant siblings (Hazlett et al., Reference Hazlett, Gu, Munsell, Kim, Styner, Wolff and Collins2017) found that an above-average increase in the surface area of the brain in the first year and in whole brain volume in the second year predicted (with 80%) accuracy which infants went on to develop autism. This positive result, together with the moderately promising results in phenotyping, suggests that more fine-grained measurement in both biology and phenotype may lead to progress.
SUGGESTIONS FOR RESEARCH PROGRESS
As the complexity of autism has become increasingly apparent, some new directions have been considered. Here, we speculate about which approaches might be most fruitful, based in large part on overcoming the obstacles discussed above.
The Initial Pool from Which Participants Can Be Drawn Should Represent the Entire Autism Spectrum
Even after decades of research, we do not know whether RRBs must accompany autistic social deficits. The DSM-5 (APA, 2013) requires two RRBs for an ASD diagnosis. In addition to lack of continuity with the body of DSM-IV based research, the social symptoms in DSM-V are difficult to apply to young or cognitively impaired children. Most importantly, these criteria have prematurely closed the category to individuals with autistic social deficits and no (or one) RRBs. Adding PDD-NOS (APA, 2000) or atypical autism (WHO, 1993) to autistic disorder or childhood autism would capture individuals with ASD’s core social domain, who can then be further characterized. In contrast, starting with ASD as defined in the DSM-5 will guarantee that all individuals enrolled in research have both social impairments and RRB but will not reveal whether they cohere as a syndrome.
All Research Participants Should Be Characterized for Detailed Clinical Features, Both Diagnostic and Non-diagnostic
Given the heterogeneity of autism, and the modest success in mapping diagnosis onto biological findings, it is perhaps time to largely abandon simple case control studies, without thoroughly characterizing participants. Waterhouse and Gillberg (Reference Waterhouse and Gillberg2014) argue that very narrowly defined subgroups, both at the phenotypic and biological levels, will increase the probability of being able to link the two. Studies starting with a single diagnostic group of doubtful validity will be uninformative (Uher & Rutter, Reference Uher and Rutter2012).
Furthermore, detailed phenotypes must not be based solely on the defining features of ASD. As discussed, ASD may be a multifinal endpoint of potentially numerous starting states. Disruption in social communication may come about from lack of experience expectant input, which, in turn, may have been caused by a genetic factor (e.g., decreased proclivity to seek eye contact, reduced ability to perceive biological motion, attention shifting impairment, sensory processing impairment, etc.) or an actual lack of input (as in blindness and severe environmental deprivation). Therefore, classifying by severity in social communication may not increase etiological homogeneity.
Several systems for detailed phenotyping have been proposed: First is Gillberg’s ESSENCE system (Gillberg, Reference Gillberg2010; Pettersson, Anckarsäter, Gillberg, & Lichtenstein, Reference Pettersson, Anckarsäter, Gillberg and Lichtenstein2013), which includes motor abnormalities, general delays, speech/language delay, social functioning, hyperactivity/ inattention, sleep disturbances, mood, and feeding difficulties. Second is a comprehensive list by Lai, Lombardo, Chakrabarti, and Baron-Cohen (Reference Lai, Lombardo, Chakrabarti and Baron-Cohen2013), with over 100 variables including developmental course, gender, comorbidities (medical and psychiatric), cognitive level and profile (including social cognition), as well as genetics (including syndromes, multiplex/simplex family, and gene level variants associated with autism), and environmental risk factors. Third is the Research Domain Criteria (RDOC; Cuthbert & Insel, Reference Cuthbert and Insel2013), in which the targets of study are psychological processes rather than diagnostic syndromes. Individuals with ASD show deficits in many RDOC domains (positive and negative emotions, memory, language, perception, social communication, understanding of self and others, arousal, self-regulation, sleep rhythms, etc.)
However, none of these lists focus explicitly on quality of behavior, especially social behavior. We need to agree upon measures of quality and for changes across development. Foss-Feig, McPartland, Anticevic, & Wolf (Reference Foss-Feig, McPartland, Anticevic and Wolf2016) had the interesting idea of sorting symptoms as negative, positive, or cognitive (as with schizophrenia). Wing and Gould (Reference Wing and Gould1979) suggested typing children by their predominant social attitude (aloof, passive, active but odd). A somewhat more detailed system of qualifying social behavior that we have found very clinically useful is to identify simple social processes and partners that may be present, emerging, or not yet present. In the grid (Figure 1), we expect development to proceed more or less from top to bottom and left to right; this system allows clinicians to track development, to target next developmental steps, and perhaps to identify social subtypes.
Fig. 1 Characterization of social interaction.
We also suggest adding variables that have distinguished groups in previous studies, or are candidate endophenotypes: pretend play with peers (Kasari, Chang, & Patterson, Reference Kasari, Chang and Patterson2013), spontaneous imitation (e.g., Toth, Munson, Meltzoff, & Dawson, Reference Toth, Munson, Meltzoff and Dawson2006), temperament (e.g., Garon et al., Reference Garon, Bryson, Zwaigenbaum, Smith, Brian, Roberts and Szatmari2009), psychiatric comorbidities (Hutton, Goode, Murphy, Le Couteur, & Rutter, Reference Hutton, Goode, Murphy, Le Couteur and Rutter2008), medical comorbidities (Bauman, Reference Bauman2010), savant skills (Hermelin, Reference Hermelin2001; Howlin, Reference Howlin2010), and family history of psychiatric/neurodevelopmental disorders, including the BAP (Bailey & Parr, Reference Bailey and Parr2003).
Furthermore, It May Be Productive to Study Characteristics That Are Likely to Be Less Influenced by the Environment
The more a behavioral feature can be influenced by the environment, the less we can expect that stratifying ASD groups along that feature will yield increased genetic homogeneity. Rett’s syndrome was discovered by studying the subgroup of ASD children who are female, show excessive hand wringing, marked regression of hand skills, and head growth deceleration in the second year (Percy, Reference Percy2001), all of which are likely to be minimally influenced by environmental factors. The unusual cry in infancy may have been crucial in identifying Cri-du-Chat as a specific syndrome (in combination with variable dysmorphic features) with the genetic underpinnings elucidated later (Lejeune et al., Reference Lejeune, Lafourcade, Berger, Vialatte, Boeswillwald, Seringe and Turpin1963; German, Lejeune, Macintyre, & De Grouchy, Reference German, Lejeune, Macintyre and De Grouchy1964).
One factor that seems relatively independent of environment is improvement of emotional and social responsiveness with a high fever (Curran et al., Reference Curran, Newschaffer, Lee, Crawford, Johnston and Zimmerman2007), perhaps implicating changes in brain chemistry or the operation of the locus coeruleus in a subset of autism cases (Mehler & Purpura Reference Mehler and Purpura2009). Possible high pain tolerance (although the possibility has been raised that it is the expression rather than the experience of pain that differs) (Allely, Reference Allely2013) and distinctive physical parameters (e.g., head and body growth, dysmorphology) are other features likely to be relatively uninfluenced by the environment. Freedman and Foxe (Reference Freedman and Foxe2017) suggest impaired visual saccade and saccade adaptation as an early autism endophenotype, which may be correlated with cerebellar vermis dysfunction, may be measurable as early as 10 months, and may result in abnormal cortical mapping of space, with over-representation of the peripheral visual field, an intriguing idea since some affected children seem to attend preferentially to the visual periphery.
When Considering the Need to Minimize Environmental Influence, the Issue of Timing Is of Utmost Importance
Enrolling children before any intervention is started is especially important for variables that show increasing sensitivity to environment over time. For example, social attention in infancy is strongly correlated with genetic factors (Constantino et al., Reference Constantino, Kennon-McGill, Weichselbaum, Marrus, Haider, Glowinski and Jones2017); however, environmentally determined factors such as amount of face to face interaction (affected by parenting and intervention) will probably play a larger and larger role in their social attention as development unfolds and so will only be likely to yield more genetically homogenous groups if measured early in development (e.g., Jones and Klin, Reference Jones and Klin2013).
Gathering detailed behavioral data in the first year or 2 of life, given ASD’s prevalence, is a logistical challenge that may be met in part by studying high-risk samples (premature babies, baby sibs of affected children). Some groups are attempting to do structural and even functional imaging with very young children with ASD or risk for ASD during sleep; a recent study reported 80% correct prediction of the ultimate diagnostic status of baby siblings by monitoring early brain growth (Hazlett et al., Reference Hazlett, Gu, Munsell, Kim, Styner, Wolff and Collins2017). Two limitations of this high-risk strategy are that large samples will still be needed, although much smaller than general population studies, and that autism unfolding in these groups may not be wholly representative of autism in other individuals.
Several aspects of clinical course have some specificity for autism (Rutter, Reference Rutter2011), including early regression of language or social milestones in approximately a quarter of cases (Baird et al., Reference Baird, Charman, Pickles, Chandler, Loucas, Meldrum and Simonoff2008; Pickles et al., Reference Pickles, Simonoff, Conti‐Ramsden, Falcaro, Simkin, Charman and Baird2009), savant skills in approximately one third of cases (Howlin, Reference Howlin2010), early accelerated brain growth in a minority of cases (Libero, Nordahl, Li, & Amaral Reference Libero, Nordahl, Li and Amaral2017; Woodhouse et al., Reference Woodhouse, Bailey, Rutter, Bolton, Baird and Couteur1996), onset of epilepsy at distinctive times in approximately one third of cases (Bolton et al., Reference Bolton, Carcani-Rathwell, Hutton, Goode, Howlin and Rutter2011), extreme variability in adult outcome (Howlin et al., Reference Howlin, Goode, Hutton and Rutter2004), with onset of psychiatric comorbidity at puberty in 20% of cases (Hutton et al., Reference Hutton, Goode, Murphy, Le Couteur and Rutter2008), and loss of diagnosis in a significant minority (Fein et al., Reference Fein, Barton, Eigsti, Kelley, Naigles, Schultz and Troyb2013). Therefore, as suggested by Rutter (Reference Rutter2011), and Uher & Rutter (Reference Uher and Rutter2012), adding longitudinal course to characterization of participants is also likely to reduce etiological heterogeneity. Clinical course should include age of onset of symptoms in early childhood, as a third to a half of parents of children with autism recall abnormalities beginning in the first year (Yirmiya & Charman, Reference Yirmiya and Charman2010).
Furthermore, If There Is a True Increase in the Incidence of Autism (not Simply an Increase in Those Diagnosed, and This Issue Is yet to Be Resolved), then Changing Environmental Exposures May Be Serving as Epigenetic Triggers in the Presence of Background Familial Risk
Prospective longitudinal studies, such as the Norwegian mother and baby study of 100,000 individuals hold the greatest promise for shedding light on this question (Rønningen et al., Reference Rønningen, Paltiel, Meltzer, Nordhagen, Lie, Hovengen and Hoppin2006). Other biological studies can be done even earlier, such as studies of placental abnormalities at birth (Anderson et al., Reference Anderson, Jacobs-Stannard, Chawarska, Volkmar and Kliman2007), midpregnancy variables (Kolevzon, Gross, & Reichenberg, Reference Kolevzon, Gross and Reichenberg2001) such as maternal autoantibodies to fetal brain protein (Croen et al., Reference Croen, Braunschweig, Haapanen, Yoshida, Fireman, Grether and Van de Water2008), and bleeding (Gardener, Spiegelman, & Buka, Reference Gardener, Spiegelman and Buka2009). Cases with environmental factors such as toxins (Lyall, Schmidtt, & Hertz-Picciotto, Reference Lyall, Schmidt and Hertz-Picciotto2014) or medication during pregnancy (Christensen et al., Reference Christensen, Grønborg, Sørensen, Schendel, Parner, Pedersen and Vestergaard2013) and early sensory deprivation (Jure et al., Reference Jure, Pogonza and Rapin2016; Rutter et al., Reference Rutter, Andersen‐Wood, Beckett, Bredenkamp, Castle, Groothues and O’Connor1999) are clearly needed for subtyping, and perhaps should be excluded from genetic studies. Although just beginning, recent studies have suggested that the gut microbiome, influenced by mode of delivery, feeding, and mother’s health status, may be abnormal in autism (Li et al., Reference Li, Han, Dy and Hagerman2017) and may represent a future target of treatment.
We Also Need to Distinguish What Individuals with ASD Cannot Do and What They Do not Do.
In many tightly controlled lab or clinic settings, individuals with ASD demonstrate proficiency on a task, but they do not deploy these skills in natural circumstances (Dalton et al., Reference Dalton, Nacewicz, Johnstone, Schaefer, Gernsbacher, Goldsmith and Davidson2005; Kinsbourne and Helt, Reference Kinsbourne and Helt2011; Mueller and Amaral, Reference Mueller and Amaral2017; Schultz et al., Reference Schultz, Gauthier, Klin, Fulbright, Anderson, Volkmar and Gore2000). Although participants with ASD often display typical social behaviors such as mimicry (McIntosh et al., Reference McIntosh, Reichmann-Decker, Winkielman and Wilbarger2006), imitation (Hamilton, Brindley, & Frith, Reference Hamilton, Brindley and Frith2007), and eye contact (Hadjikhani et al., Reference Hadjikhani, Joseph, Snyder, Chabris, Clark, Steele and Harris2004) in experiments where these behaviors are specifically cued, ASD severity consistently correlates with what these individuals typically do when observed in natural settings (e.g., McIntosh, Reichmann-Decker, Winkielman, & Wilbarger, Reference McIntosh, Reichmann-Decker, Winkielman and Wilbarger2006).
For example, fMRI studies of face perception have shown that individuals with ASD fail to look at the eyes of target stimuli if they are adult strangers (thus failing to significantly activate the fusiform face area (FFA), e.g., Schultz et al., Reference Schultz, Gauthier, Klin, Fulbright, Anderson, Volkmar and Gore2000) but will look at the eyes (and activate FFA) if the stimuli involve familiar people (Pierce, Haist, Sedaghat, & Courchesne, Reference Pierce, Haist, Sedaghat and Courchesne2004), or if they are explicitly cued to do so (Zürcher et al., Reference Zürcher, Donnelly, Rogier, Russo, Hippolyte, Hadwin and Hadjikhani2014). In addition, individuals may perform well on a task but use a different process; for instance, individuals with a history of ASD but who have achieved “optimal outcomes,” perform as well on tests of language and cognition as their TD peers, but their brain scans suggest that they are using compensatory strategies, activating more areas of the brain to complete these tasks normally (Eigsti et al., Reference Eigsti, Stevens, Schultz, Barton, Kelley, Naigles and Fein2016). These factors make study design especially delicate in regards to ASD. We must be careful not to conclude that underlying neural systems are “not intact” in individuals with ASD simply because they do not display a specific behavior in specific circumstances.
Since Biological Causes for Autism Are Numerous and Complex, and Narrowly Defined Subgroups Will Also Be Numerous, There Will Be a Need for Very Large Samples, with Data Aggregated across Many Sites to Make Significant Progress
Sophisticated and exploratory big data analyses should be used to produce phenotypic subtypes and correlate them with etiological possibilities, or to identify gene variants and correlate them with phenotypic characteristics (see Campbell, Kohane, & Kong, Reference Campbell, Kohane and Kong2013, for an example of exploratory analysis of biological data). Such attempts at aggregation have been started by the National Database for Autism Research, and by networks of brain imaging researchers (Di Martino et al., Reference Di Martino, Kelly, Grzadzinski, Zuo, Mennes, Mairena and Milham2011). The MSSNG (“Missing”) project is a collaboration of Autism Speaks and Google, which hosts the database in the cloud, and is directed by Toronto’s Hospital for Sick Children. As of March 2017, they had enrolled and analyzed data from over 5000 families with an autistic member, with data contributed by pre-existing databases such as the Autism Genetic Research Exchange, and had identified many more autism-associated gene variants, some of which were also associated with comorbidities such as epilepsy. The Simons Foundation SPARK study aims to collect genetic and medical information from 50,000 affected individuals and their family members, and is collecting data from multiple autism research centers around the country.
These collaborative databases will no doubt accelerate the pace of autism discoveries, particularly in genetics. However, in our view, large scale data aggregation must occur in concert with agreed-upon phenotypic measures with a minimum of key features as described above, and not just a positive diagnosis. As mentioned above, studies of linkage and GWA studies require very large samples and adding longitudinal factors may make these prohibitively expensive. Alternatively, if autism risk genes, or specific CNV’s are used to genetically characterize a group which is then intensively phenotyped, such studies may be productive and more feasible.
An example of such a “genotype first” study with more fine-grained subtyping as well as animal modelling is that of Bernier et al. (Reference Bernier, Golzio, Xiong, Stessman, Coe, Penn and Schuurs-Hoeijmakers2014); they focused on mutations in a gene (CHD8) which previous studies had identified in some autistic individuals. They resequenced the gene in approximately 8000 individuals, identifying additional mutations affecting function (all de novo except one maternally inherited). None of these mutations were found in unaffected siblings or in the general population. Clinical information was available from 15 affected individuals to correlate with the mutations. Features that were present in more than half of the participants, in addition to diagnosed ASD (13/15), were ID or borderline cognition (11/15), macrocephaly (12/15), tall stature (12/14), a characteristic facial feature (8/9), gastrointestinal problems (12/15), and sleep problems (10/15). Finally, the team studied the expression over development of CHD8 (highly conserved across species) and its disruption in zebrafish, which resulted in increased head size and impaired gut motility, both reported in a subset of individuals with autism.
CONCLUSION
There is widespread agreement that ASD represents a complex behavioral endpoint with hundreds of possible biological and environmental vulnerabilities in potentially endless combinations. Truly recognizing the massive heterogeneity within ASD means that case-control studies of heterogeneous groups (i.e., those that only require that ASD is present) have limited utility, and, in many cases, are simply muddying the waters. It is likely that even GWS involving thousands of participants are underpowered due to the sheer number of genetic risk factor combinations for autism.
Given this level of complexity, we propose that autism research going forward must adopt several principles that will allow us to move forward. We should include in ASD samples all individuals who struggle with social communication (of the type described in DSM-IV and 5 and ICD-10) regardless of whether RRBs are present. Genotyping, brain imaging, and environmental contributions should then be studied for groups who are homogeneous on particular phenotypic variables, by characterizing each individual on an expansive number of behavioral and medical variables, with a focus on those most likely to be robust to environmental influence (e.g., high pain threshold, improvement with fever, poor GI motility, impaired visual saccades, decreasing eye contact in the first year). The field as a whole must agree upon a minimal standardized set of history and clinical phenotype measures and begin to relate these measures to genetic and environmental factors.
In “genotype-first” studies, where the participants share a probable genetic contributor, these same variables should be used. Studying these variables before any intervention will increase the independence from the environment. Phenotyping measures should also include longitudinal course, including onset of symptoms and delays, regression and response to treatment. The fact that sensory deprivation such as blindness or neglect can lead to autism suggests that a lack of social input from any of multiple causes disrupts the sense of self and other and validates the idea of environmental causes, which need to be vigorously pursued.
Advances in neuroimaging, genetic/genomic analyses, and exploratory statistical methods (e.g., Campbell et al., Reference Campbell, Kohane and Kong2013) will possibly accelerate the discovery of etiologies and pathophysiology, especially with very large data sets accrued by cross-site aggregation. Even if, in many cases, narrowly defined subgroups fail to link genotype to phenotype, the possibility of linking the two for even one subgroup of ASD is worth pursuing; if we could learn how ASD behaviors result from underlying biology for one subgroup, it may shed light on how behavior and biology are linked in other subgroups. Furthermore, defining phenotypic and biological subgroups will likely lead to better predictions about treatment and prognosis.
We believe that defining subgroups of ASD on behavioral variables that show minimal sensitivity to environmental input is our best chance of connecting ASD symptoms to underlying biology. However, given the disconnect between genotype and phenotype thus far, we acknowledge the possibility that even this type of subtyping will not prove ultimately successful. In this case, the most fruitful way forward is still to begin defining subgroups on phenotypic variables, including symptom emergence, response to medical and behavioral treatments, comorbidities, and longitudinal course. This should allow clinical advancements, with earlier and earlier detection and better and better treatment improving outcomes.
ACKNOWLEDGMENTS
Parts of this study were given as the Herbert Birch Memorial Lecture at the International Neuropsychological Society meeting by DF in 2015. We thank the many individuals with autism, and their families, who have participated in our research and clinical work, and who have informed much of our thinking. We also thank the graduate student and faculty colleagues with whom we have worked over the last decades, and who have stimulated and corrected our thinking (especially UConn colleagues Drs. Barton, Eigsti, and Naigles). We are very grateful to colleagues and mentors who have been tremendously influential on our understanding of neurodevelopmental disorders and brain function (Drs. Yoko Kamio, Edith Kaplan, Marcel Kinsbourne, Harriet Levin, Allan Mirsky, Nahit Motavalli Mukaddes, Isabelle Rapin, and Lynn Waterhouse). Our research has been funded by the National Institute of Mental Health, the Eunice Kennedy Shriver National Institute of Child Health and Development, Autism Speaks, the National Alliance for Autism Research, Maternal and Child Health Bureau, the U.S. Department of Education Student-Initiated Research Grant and the University of Connecticut and Trinity College faculty grants. Dedication: We dedicate this study to Isabelle Rapin, MD, who engaged with us in countless discussions on these topics, about which she was incomparably wise, knowledgeable, and passionate. Her death in May, 2017, is an enormous loss for her colleagues, patients, and family. Disclosures: M.H. has no potential COI. D.F. is part owner of the M-CHAT-R, LLC, a screener for autism in toddlers. The LLC collects royalties when the screener is incorporated into a commercial product, but is available for free to pediatricians.
