Prediction of Neurodevelopmental Disorders Based on De Novo Coding Variation

Neurodevelopmental disorders (NDDs), such as autism spectrum disorder (ASD), epilepsy, intellectual disability (ID), and developmental disability (DD) are complex disorders characterized by impairment in cognition, learning, and motor skills. From twin and family studies, it has become apparent that NDDs possess a strong genetic component (Freitag, 2007; Gejman et al., 2010). Estimates of heritability for various NDDs have ranged from 0.3 to 0.9, with heritability estimated to be greater than 0.5 for both ASD and ID (Flint, 2001; Kaufman et al., 2010; Tick et al., 2016). The evident contribution of genetic factors to NDD diagnoses has provided reason for routine prenatal whole exome or genome sequencing to identify potentially deleterious genetic variations (Soden et al., 2014; Tărlungeanu & Novarino, 2018). In particular, whole exome sequencing has proved a useful tool to identify, at a low cost, coding variants in genes that are highly intolerant to mutation and play important roles in typical neurodevelopment (Srivastava et al., 2019).

The identification and prioritization of NDD risk genes is important for the discovery of underlying biological mechanisms that are perturbed in NDDs (Cardoso et al., 2019). Previous studies have identified many monogenic forms of NDDs and revealed the multifactorial and polygenic nature of most NDD diagnoses (De Felice et al., 2015; Niemi et al., 2018; Sztainberg & Zoghbi, 2016). In particular, rare de novo mutations that are observed in genes in NDD cases at a significantly higher rate than expected relative to unaffected controls have pinpointed many candidate NDD genes, with more than one thousand genes estimated to be NDD risk genes (De Rubeis et al., 2014; Heyne et al., 2018; Iossifov et al., 2012; Kaplanis et al., 2020; McRae et al., 2017; O’Roak et al., 2012a, b; Sanders et al., 2012; Satterstrom et al., 2020; Wilfert et al., 2017).

De novo mutations are a class of rare genetic variation in which variants, that are not observed in parental genomes, exist in offspring due to mutagenesis in germ cells or errors in replication or recombination (Acuna-Hidalgo et al., 2016). De novo mutations may exist as single nucleotide variants, insertions and deletions (indels), and copy number variants. Because de novo mutations are not inherited, highly penetrant mutations can arise in genes that are critical to neurodevelopment and likely under purifying selection (Iossifov et al., 2012; Uddin et al., 2014). In fact, individuals affected by NDDs experience a greater burden of non-synonymous de novo mutation compared to unaffected controls (Coe et al., 2019; Wilfert et al., 2017). Study of ASD simplex families from the Simons Simplex Collection (SSC) has found that de novo likely gene disruptive (LGD) mutations occur at a nearly twofold increased rate in affected cases (0.21) relative to controls (0.12), as well as displaying an increased rate of missense mutation (Iossifov et al., 2014). Furthermore, the study of genetic modules impacted by these de novo mutations has pinpointed several biological processes relevant to NDD etiology, such as chromatin remodeling, the Wnt pathway, synaptic transmission, and the long-term potentiation pathway (Chow et al., 2019; Kwan et al., 2016; O’Roak et al., 2012a, b; Wilfert et al., 2017).

The benefits associated with successful early prediction of NDDs include the improved ability of parents to make informed decisions about potential early application of treatments (Boivin et al., 2015; Cioni et al., 2016; Corsello, 2005). It is important to note that most NDDs cases cannot be predicted using de novo coding variation alone; the exome constitutes 1–2% of the human genome and the majority of NDD-associated variants are likely to reside in non-coding regions involved in the regulation of gene expression (Short et al., 2018; Turner & Eichler, 2019). Currently, it is estimated that only ~ 10% of ASD cases and ~ 20–30% of ID/DD cases have de novo LGD variants, and the rate of such variants in the general population is significantly lower (Wang et al., 2021). The genetically and phenotypically heterogeneous nature of NDDs indicates that many factors, including common and non-coding genetic variants and non-genetic factors, account for a large fraction of diagnoses, further complicating our ability for the early prediction of these disorders. However, it is possible to confidently predict a subset of individuals who will likely develop NDDs due to de novo coding variation in the form of non-synonymous de novo mutations. Despite the polygenic nature of NDDs and the multitude of potential genetic or environmental causes, focusing specifically on un-inherited, de novo mutations that disrupt protein coding sequence permits early prediction for a small fraction of cases with very low false positive rates.

The early prediction of NDDs requires a very low false positive rate (FPR) due to potential negative consequences, such as the costs associated with early intervention treatments, that may result from false positive prediction. The minimization of the FPR is clinically most relevant in genetic counseling settings for parents with suspected or confirmed familial risk for NDDs and to aid in the decision to begin early intervention treatments in young children. Early diagnosis of NDDs via a combination of behavioral and motor assessments, imaging, and genetic testing followed by early prediction methods can greatly benefit patient outcomes and lead to timely, appropriate treatment (Hadders-Algra, 2021). Previously, a method for the early prediction of complex disorders, Odin, used de novo LGD variants observed in cases and controls and co-expression data to identify cases at very low FPR (Huynh & Hormozdiari, 2018). The shallow neural net (SNN) with novel objective function introduced here incorporates LGD de novo mutation, constraint, and conservation data to achieve a higher (> 0.30129) true positive rate (TPR) at very low FPR (< 0.01) in comparison to traditional classification models such as random forest (RF), support-vector machine (SVM), and logistic regression (LR). Furthermore, the proposed SNN model achieves similar PR-AUC and ROC-AUC to other machine learning approaches. An ensemble model that averages predictions among the SNN, RF, SVM, and LR models is able to achieve a slightly increased TPR at FPR < 0.01 and comparable PR-AUC. Additionally, the SNN is able to rank genes according to their relative importance in NDDs given LGD or missense de novo variation, prioritizing candidate NDD genes.

To identify, at very low FPRs, a subset of affected NDD cases using rare coding variation consisting of de novo LGD and missense variation, LGD- and missense-specific feature matrices indicating the presence of de novo variation within genes were constructed (Fig. 1A). Additional features incorporating constraint and conservation data were used to improve classification of NDD cases using LGD variation. The ability of SNNs (Fig. 1C) to classify NDD cases at very low FPRs were compared to various classifiers, including RF, SVM, and LR (baseline models), in addition to three heuristics.

To distinguish NDD cases from unaffected controls at extremely low FPRs using de novo coding variation and measures of gene constraint and conservation, we developed a SNN with a customized objective function to maximize TP while simultaneously minimizing FP (Fig. 1). Although most cases of NDDs arise from a variety of classes of genetic variation, particularly common, non-coding, and structural variants, focusing specifically on de novo coding variation of relatively large effect size is a tradeoff to obtain significantly reduced FPR on a small but significant subset of samples. Compared to traditional machine learning techniques, such as RF, support vector machine (SVM), and LR (referred to as ‘baseline’ models), the constructed SNN is able to achieve greater TPR at FPR less than 0.01 given LGD-specific variation (Table 1, Fig. 2). The ability of the SNN to capture more than 30% of cases at FPR < 0.01, corresponding to at least 5.29% more cases than any baseline model (Table 1), indicates that the use of a SNN with the custom loss function (Eq. 1) is beneficial in classifying NDD cases at very low FPR. Note that it is estimated that LGD variants have been observed in roughly 10% of ASD cases and up to 30% of DD cases (Wang et al., 2021). Thus, our results indicate that the proposed SNN should be able to identify > 3% of ASD and > 10% of all DD cases while having an FPR of virtually zero simply by considering de novo LGD variants.

In addition, a simple ensemble method that uses the average predicted probability from SNN and baseline model predictions is able to identify NDD cases at greater TPR at FPR < 0.01 and slightly increased precision at lowered recall than any of its constituent models (Table 1, Supplementary Fig. 4). Excluding SNN predictions from the ensemble model reveals that the SNN, compared to baseline models, contributes substantially to the ensemble model’s ability to accurately classify NDD cases at low FPR values. In fact, for LGD-specific variation, an ensemble method that excludes SNN predictions produces decreased TPR at FPR < 0.01 metrics compared to the SNN alone (Table 1).

The ability of SNN and baseline models to use only missense variation to identify NDD cases is relatively poor. However, the incorporation of both missense and LGD-specific predictions during ‘combined’ prediction for samples containing both LGD and missense variation, in which the maximum predicted probability from two separately trained missense- and LGD-specific models are returned, increases average TPR at FPR < 0.01 compared to using only probabilities predicted by an LGD-specific model (Table 1, Fig. 2). The improved performance of combined predictions indicates that certain samples possessing very deleterious missense variation (in addition to LGD variation) are correctly classified as cases when the predicted probability associated with the missense-specific model, rather than the LGD-specific model, is retrieved.

SNNs trained on LGD- and missense-specific feature matrices containing de novo coding variation from NDD cases and controls are able to rank genes according to their relative importance to NDD risk when applied to artificial samples which each contain a single type of de novo variant in a single gene (Supplementary Table 6). An increased enrichment of de novo LGD and missense mutation in NDD cases relative to controls is observed in highly ranked genes (those with higher predicted probability of being a case) using LGD-specific variation (Fig. 3). Significant, strong correlation exists between predicted probability for artificial samples for both the pLI and LOEUF constraint metrics, showing that the ranking via LGD-specific variation can accurately detect most high risk NDD genes. Among the 50 most highly ranked genes using LGD-specific variation, a total of 47 out of 50 genes are classified as high confidence (39 genes with score 1), strong candidate (6 genes with score 2), and suggestive evidence (2 genes with score 3) autism spectrum disorder (ASD) risk genes, including genes relevant to syndromes, according to SFARI Gene and OMIM (Supplementary Table 6). Among genes with predicted probabilities greater than 0.90 (ranks 1–55), four genes (WDR45, CLTC, BRPF1, and GATAD2B) do not possess SFARI annotations, but have been associated with neurodegeneration and intellectual disability according to OMIM annotations. Highly ranked genes lacking both SFARI Gene scores and OMIM annotations suggest candidate NDD genes susceptible to de novo LGD variation. Evidence of association with NDDs [ZFHX3 (Fuller et al., 2018), CHD5 (Parenti et al., 2021), UBR3 (Murcia Pienkowski et al., 2020)) or enrichment of de novo LGD mutation in NDD cases (ANP32A, SKIDA1 (Coe et al., 2019)), neurodegeneration (ANP32A (Podvin et al., 2020), HECTD1 (Schmidt et al., 2021)), gliomas (LARP4B (Koso et al., 2016)), synapses and neuronal formation (LMTK3 (Takahashi et al., 2020), DOT1L (Franz et al., 2019)] have been studied in model organisms, cell lines, and families for these candidate NDD genes.

Weaker correlation is observed for missense-specific rankings with pLI and LOEUF values, and enrichment of de novo non-synonymous mutation is also present in NDD cases relative to controls, although to a lesser extent compared to LGD-specific rankings (Supplementary Fig. 5). The missense-specific rankings are distinct from LGD rankings in their ability to identify genes potentially sensitive to missense variation (Supplementary Table 6). Among highly ranked genes lacking SFARI Gene scores and OMIM annotations, previous studies suggest association with NDDs and schizophrenia [OBSCN (Hashimoto et al., 2016), PLEC (Dincer et al., 2015), RYR2 (Lieve et al., 2019), ZSWIM8 (Tischfield et al., 2017)], cortical formation and thickness [LAMA5 (Omar et al., 2017), GOLGA3 (Kim et al., 2017)), and neurodegenerative diseases (PKHD1 (Santos-Laso et al., 2020), DNAH1 (Thonberg et al., 2017)].

Our results indicate that we can accurately predict a small, yet significant fraction of NDD cases using de novo coding variants. Currently, whole-exome or whole-genome sequencing of trios is not common practice. However, to make the early prediction of these disorders a reality, such sequencing should become common practice. Furthermore, our approach only covers a small fraction of affected patients and additional methods that use other types of biomolecular signatures, such as common variants, rare non-coding variants, and epigenomic markers, are needed to increase the reach of early prediction to a larger fraction of cases.

In summary, the described SNN identifies NDD cases at higher TPR while having very low FPR in comparison to traditional machine learning methods. Several factors contribute to the improved performance of the proposed approach, namely: the use of gene constraint and conservation features in LGD-specific prediction and a custom loss function that specifically seeks to maximize the TPR while minimizing the FPR. An ensemble method, aggregated from SNN and baseline model predictions, is able to correctly classify a greater proportion of cases at FPR < 0.01 compared to any individual model. The SNN itself is a major contributor to increased TPR at FPR < 0.01 observed in the ensemble model. Although de novo missense mutation alone is a poor predictor of case status relative to LGD mutation, missense-specific predictions are useful during combined prediction for identifying additional cases that possess highly deleterious missense mutation in addition to LGD mutation. Fully trained SNNs on LGD- or missense-specific variation are also useful in NDD risk gene prioritization, revealing candidate NDD genes enriched in de novo non-synonymous mutations in NDD cases relative to controls.