Can the "female protective effect" liability threshold model explain sex differences in autism spectrum disorder?

Abstract

Male sex is a strong risk factor for autism spectrum disorder (ASD). The leading theory for a "female protective effect" (FPE) envisions males and females have "differing thresholds" under a "liability threshold model" (DT-LTM). Specifically, this model posits that females require either a greater number or larger magnitude of risk factors (i.e., greater liability) to manifest ASD, which is supported by the finding that a greater proportion of females with ASD have highly penetrant genetic mutations. Herein, we derive testable hypotheses from the DT-LTM for ASD, investigating heritability, familial recurrence, correlation between ASD penetrance and sex ratio, population traits, clinical features, the stability of the sex ratio across diagnostic changes, and highlight other key prerequisites. Our findings reveal that several key predictions of the DT-LTM are not supported by current data, requiring us to establish a different conceptual framework for evaluating alternate models that explain sex differences in ASD.

Figure 1.. The liability threshold model

(A) This normal curve illustrates ASD liability in a population. Individuals exceeding a liability threshold would be diagnosed with ASD (shaded areas). The 4:1 ratio of males to females under the DT-LTM can be observed with the female threshold being shifted to the right. (B) A variation of this model conceptualizes that a single ASD threshold exists for the population. In this version, both males and females would have normal distributions of ASD liability, but the male mean would be shifted slightly toward ASD, such that more males cross the single diagnostic threshold (black line) to yield a 4:1 ratio. (C) The LTM has three explicit prerequisites: risk must be multifactorial and additive, the genetic risk factors and their relative impact must be identical between sexes (perfect genetic correlation), and the variance in liability must be the same for both sexes.

Figure 2.. Relationship between sex ratio and ASD penetrance rates

(A) The DT-LTM predicts more strongly penetrant ASD genes (y axis) should show a lower proportion of males (x axis). We have plotted currently available data for high confidence ASD genes with >25 cases to report ASD rate and >10 cases with sex reported. At current sample numbers, these do not show the predicted relationship (r = 0.58, p = 0.082). (B) In (B), genes with more mutations observed in neurodevelopmental disorders (NDDs) are labeled NDD > ASD and those with more mutations observed in ASD are labeled ASD > NDD. The ASD > NDD genes are those we interpret as having greater specific ASD liability under the DT-LTM. NDD > ASD genes have lower sex ratios than those genes where ASD > NDD for all three association strength cutoffs (FWER < 0.05, FDR < 0.05, or FDR < 0.1). ASD > NDD genes are thus consistently more male biased, contrary to the DT-LTM. Association data for ASD and differences between NDD and ASD from Satterstrom et al. (2020), and data on mutations observed by sex are from Turner et al. (2019). Because ASD gene Ns are lower in Turner et al. and only one gene showed significant heterogeneity (p < 0.05) with more mutations in ASD in Satterstrom et al., we used a lenient p < 0.25 threshold.

Figure 3.. The difference in the means of a QAT does not account for the difference in the extremes when constrained by the “equal variance” prerequisite of the DT-LTM

Population SRS scores show a male bias in the ASD-impaired direction. (A) Parameterizing two normal curves with these population means, but the same variance, as expected under the LTM, will result in a <2:1 ratio of males (blue shading) and females (red shading) exceeding a diagnostic threshold (dashed line) of 70. (B and C) (B) Parameterizing with the same means, but with the observed SRS male and female standard deviations increases the ratio to nearly 4. (C) Predicted M/F ratios in individuals exceeding the thresholds in (A) and (B). A cutoff of 70 was selected here for illustration, but similar results are seen with a variety of diagnostic thresholds.

Figure 4.. Rate of ID does not positively correlate with rate of ASD across monogenic causes of ASD

Literature review of most well-studied ASD-associated genes does not indicate there is a positive correlation between rate of ASD diagnosis and rate of ID diagnosis (r = −0.011, p = 0.07). This suggests a given gene’s prevalence of ID (when mutated) cannot be used as a proxy for that gene’s contribution to genetic liability for ASD.

Figure 5.. ASD prevalence changed over time although M:F ratio did not

A histogram (A) of M:F ratios from a comprehensive review of 200 ASD prevalence studies from 1966 onward (Centers for Disease Control, 2020) indicates a mean M:F ratio of 4.1. The DT-LTM predicts that if the agreed threshold to diagnose ASD is shifting over time to be more lenient (e.g., to go from ~ 1:2,000 in 1999, illustrated in (B), to 1:54 in 2020, as illustrated in (D)), it should have also altered the M:F ratio exceeding that new diagnostic thresholds from 4.1 to 2.6:1. ((C) and (E) are zoomed views of (B) and (D), respectively). Examining the dates of each of the 200 studies shows that although (F) prevalence rates (in cases per 1,000) of ASD have gone up with time (Pearson’s r > 0.5), (G) sex ratios (M:F) have remained flat or slightly increased (Pearson’s r > 0.1). The DT-LTM predicts they should have gone down (gray line: observed linear fit to data, orange line: expected line from 4.1 to 2.6).

Figure 6.. Summary of evidence regarding the LTM in ASD suggests the model should now be rejected

A theory can be disproven by a single counterexample. Although no single study is perfect, across our 7 domains of review and inquiry, we identified multiple lines of evidence that did not support the FPE/DT-LTM (red) or had evidence from several large studies that disagreed in their conclusions (e.g., epidemiology). Although some observations remain consistent with some aspects of the DT-LTM (green), the number of counterexamples here indicates new models are needed that can better fit the entirety of the observations.