Even though recent advances show promise in tackling heterogeneity in ASD, more work remains to be done. For example, while the relationships between ASD symptoms and neuroanatomy are enhanced after disentangling, effect sizes remain small, indicating that a large portion of individual variation is not yet explained. This could be due to multiple reasons: coarse behavioral measures, relevant ASD-specific variation not being reflected in neuroanatomical data, or inability of the CVAEs to capture all the relevant information. Each of these represents an area of potential improvement.

Current measures (such as ADOS, ADI-R, and Vineland scores) are designed for diagnosis using single aggregate scores; finer-grained measures might be needed to capture the multidimensional nature of ASD (Tang et al., 2020). Current measures do contain subdomains that can be used to capture some multidimensional variation (e.g., repetitive behavior or social-communication subdomains in ADOS). However, such experimenter-defined subdomains might not correspond to the dimensions of individual variation within ASD: a single dimension of individual variation might drive correlated changes across multiple subdomains, and vice versa. Therefore identifying dimensions that capture the difference between ASD participants and controls might not be sufficient to characterize all relevant individual variation within ASD. In addition, recent research has developed new computational approaches to measure social cognitive abilities such as the attribution of beliefs, desires, emotions, and traits (Baker et al., 2017; Houlihan et al., 2022). These approaches make it possible to estimate parameters that capture individual differences (“computational phenotyping”; Patzelt et al., 2018; Hauser et al., 2022). The convergence of computational phenotyping with techniques that disentangle disorder-specific variation such as CVAEs has the potential to paint a more detailed picture of individual differences in behavior associated with ASD.

Analyzing jointly multiple kinds of data within each individual can offer a more complete understanding of individual variation. Data acquired with one modality can reveal features that are not visible in the data acquired with other modalities. For example, structural MRI can offer anatomical information that cannot be detected at the lower spatial resolution of fMRI; conversely, fMRI can reveal differences in functional responses that complement anatomical information. Furthermore, the joint analysis of different data modalities can reveal relationships between them. A recent study demonstrated this by examining jointly structural and functional MRI data (Hong et al., 2018). Structural MRI data identified 4 different subtypes of autism, and fMRI revealed different connectivity anomalies associated with each subtype (Hong et al., 2018). Similar strategies can be adopted to integrate other data modalities (e.g., EEG, MEG, and fMRI, see Mash et al., 2018).

Genetics can further complement the characterization of each participant. Joint analysis of genetic and neuroimaging data is of particular interest because it can reveal associations between genotypes and the resulting neurophenotypes. Moving an initial step in this direction, a recent study used fMRI to measure alterations in functional connectivity in ASD, and compared their location to gene expression maps for genes associated with ASD (Benkarim et al., 2021). Another recent study (Lu et al., 2022) found that combining genetics and fMRI data improved ASD classification, suggesting that these data modalities offer complementary information.

Future studies will need to leverage large datasets to link genetics, neurophenotypes and symptoms. In particular, acquiring neuroimaging and behavioral measures as part of newly collected genetic datasets will be essential to understand the impact of individual differences in genetics on the brain and on behavior. In this context, disentangling ASD-specific variation from shared variation in neural and behavioral measures will help guide the identification of genes associated with ASD phenotype.

ASD rarely occurs in isolation—ADHD and anxiety are just some of the most frequently co-occurring conditions (Matson and Goldin, 2013). Individuals with comorbid ASD consistently show higher severity scores (Hayashi et al., 2022), making it difficult to discern based on the ADOS scores alone whether a person has severe or comorbid autism. Extremely high rates of comorbidity (ADHD is present in 50–70 percent of ASD individuals) suggest that perhaps comorbidity might be better conceptualized as another facet of individual variation.

It is therefore pressing to understand the neural bases of comorbidity in order to inform the etiology of ASD. For example: Do individuals with comorbid ASD+ADHD present with a conjunction of neural features seen in individuals who have only ASD and only ADHD, or with a distinct set of features? Research has identified both distinct and overlapping features in comorbid ADHD+ASD but definitive etiology of this condition remains elusive (Hours et al., 2022).

Several other neurological conditions (epilepsy, early onset stroke, Parkinsons; Brainstorm Consortium et al., 2018), gastrointestinal issues (Martínez-González and Andreo-Martínez, 2019), and psychiatric disorders (Obsessive Compulsive Disorder, Schizophrenia, Bipolar disorder, and Major Depressive Disorder) are comorbid with ASD (Brainstorm Consortium et al., 2018). Recent studies are revealing shared genetic, neural and cognitive mechanisms between ASD and comorbid psychiatric disorders (Mizuno et al., 2019; Thorp et al., 2021). CVAEs—with appropriate architecture modifications—can help to identify features specific to the presence of each disorder in isolation as well as features that are uniquely associated with comorbidity.

There are already a variety of computational approaches in psychiatry (see Graham et al., 2019). State-of-the-art deep learning methods have been used to distinguish between ASD participants and controls (Alsaade and Alzahrani, 2022; Chen et al., 2022; Santana et al., 2022), but are just beginning to be applied to investigate ASD subtypes and factors underlying ASD variation (Akter et al., 2021; Aglinskas et al., 2022; Parlett-Pelleriti et al., 2022). In studying individual variation specifically, even the methods that have been applied have not been fully tested in all their possible variants. For example, in deep networks, the dimensionality of the layers and their number are tied to model performance. However, larger models are prone to overfitting, therefore training set size, and diversity become increasingly important (Dinsdale et al., 2022), as well as the use of regularization techniques (e.g., dropout). Increasing model size in CVAEs while taking adequate steps to mitigate overfitting could better capture ASD-specific variation.

Alternatively, breakthroughs could come from utilizing novel network architectures. For example, Generative Adversarial Networks and Diffusion models (Goodfellow et al., 2020; Rombach et al., 2021) are able generate more realistic outputs than VAEs (on which CVAE is based on), often indistinguishable from real inputs (Karras et al., 2018; Yi et al., 2019). This ability could prove advantageous if it captures finer-grained information that might have been lost in CVAEs.

A challenge for these models remains making latent dimensions explicit (conditioning on features like age, diagnosis, or genotype) or at least interpretable. A recent study went in this direction, predicting disease-specific brain aging without longitudinal data conditioning generative models using disorder and age information (Xia et al., 2019). Applied to the case of ASD, these approaches might be used to predict individuals' response to treatment (conditioning on current neuroimaging data) without the need for traditional factorial design, i.e., all participants completing all treatments. Disentangling ASD-specific variation would make it possible to condition the models selectively using neuroimaging features that are relevant to ASD, removing confounding features that could impair the models' performance.

Descriptive research into ASD biomarkers and subtypes must go hand in hand with intervention studies aimed at improving the quality of life in affected individuals. While there are various non-medical interventions available for ASD treatment (Bond et al., 2016), response across individuals is highly variable (Kamio et al., 2015). ASD individuals respond to some interventions, but not others (Zachor and Ben-Itzchak, 2017), suggesting that selecting the right treatment for the right individual is critical. Currently, targeting specific treatments based on behavioral profiles shows only limited success in predicting response to treatment and ASD progression (Hollocks et al., 2022). Recently, fMRI biomarkers have shown promise in predicting response to treatment better than behavioral measurements (Yang et al., 2016)—disentangling disorder-specific variation can potentially bring such neuromarker approaches closer to clinical validity. In order to bridge the gap between descriptive studies and the development of personalized interventions, research findings should increasingly be used to predict future symptom trajectories and the interventions that individuals are most likely to respond to Bzdok et al. (2021).