Neonatal levels of acute phase proteins and risk of autism spectrum disorders

Background Immune signaling pathways influence neurodevelopment and are hypothesized to contribute to the etiology of autism spectrum disorders (ASD). We aimed to assess risk of ASD in relation to levels of neonatal acute phase proteins, key components of innate immune function, measured in neonatal dried blood spots. Method We included 924 ASD cases, 1092 unaffected population-based controls, and 203 unaffected siblings to ASD cases in this case-control study nested within the register-based Stockholm Youth Cohort. Concentrations of nine different acute phase proteins were measured in eluates from neonatal dried blood spots from cases, controls, and siblings using a bead-based multiplex assay. Results C reactive protein was consistently associated with odds of ASD in case-control comparisons, with higher odds associated with the highest quintile compared to the middle quintile (OR 1.50, 95% CI 1.10 – 2.04) in adjusted analyses. In contrast, the lowest quintiles of alpha-2-macroglobulin (3.71, 1.21 – 11.33), ferritin (4.20, 1.40 – 12.65), and Serum Amyloid P (3.05, 1.16 – 8.01) were associated with odds of ASD in the matched sibling comparison. Neonatal acute phase proteins varied with perinatal environmental factors and maternal/fetal phenotypes. Significant interactions in terms of risk for ASD were observed between neonatal acute phase proteins and maternal infection in late pregnancy, maternal anemia, and maternal psychiatric history. Conclusions Indicators of the neonatal innate immune response are associated with risk for ASD, though the nature of these associations varies considerably with factors in the perinatal environment and the genetic background of the comparison group.


Introduction
Perturbations in immune signaling pathways are hypothesized to influence neurodevelopment, as many immune effector molecules have pleiotropic effects in the developing nervous system (1,2).
Peripheral concentrations of cytokines and chemokines involved in the innate and adaptive immune responses vary in individuals already diagnosed with ASD compared to controls (3,4), though the prevalence of multiple disorders of the immune system (e.g., asthma, infections, allergy, autoimmune disorders) is also higher in individuals with ASD (5)(6)(7). Integrative analyses of genetic and transcriptomic data suggest that ASD and several of these commonly co-occurring conditions may be related via shared mechanisms involving innate immunity (8).
In studies employing biological samples collected from etiologically-relevant periods, deviations in neonatal and fetal markers of inflammation and innate immunity exist in some individuals later diagnosed with ASD (9)(10)(11)(12)(13)(14). Such differences may be attributable to genetic background or differential exposure to environmental insults already in utero (15)(16)(17)(18)(19). Acute phase proteins (APP) are key components of the innate immune response. We investigated the associations between neonatal APP and later diagnosis of ASD, and how environmental exposures (e.g., maternal infections) may affect such associations. We also compared individuals with ASD to their unaffected siblings to understand the role of shared familial factors (e.g., genetic background).

Study population
The Stockholm Youth Cohort (SYC) is a register-based cohort of all children aged 0-17 years living in Stockholm County (20, 21). Of the individuals born 1996-2000, we selected all individuals diagnosed with ASD before 12-31-2011, a random sample of individuals in the SYC, and unaffected siblings to ASD cases ( Figure S1). Diagnostic information in the SYC was updated after sample collection to include follow-up until 12-31-2016. Ethical approval was obtained the Stockholm regional review board (DNR 2010/1185-31/5). Individual consent was not required for this anonymized register-based study.

ASD ascertainment Covariates
Covariates were chosen based on previous associations with ASD. Covariate data were extracted from the Medical Birth Register, the Integrated Database for Labor Market Research, and the National Patient Register. We considered maternal age (22), country of origin (23), BMI (24), psychiatric history (25), hospitalization for infection during late pregnancy (3 rd trimester) (15), anemia (26), and supplement use during pregnancy (27); parental income at the time of birth and education level (highest of mother or father) (28); children's birth order, sex, gestational age at birth (29), size for gestational age (30), mode of delivery (31), and Apgar score (32).

Statistical analysis
We log2-transformed the APP values as the distributions were right skewed. Because of differences in total protein and APP distributions according to elution batch (n=3) and analytical plate (n=24; Figure   S2), we created plate-specific standardized z-scores for each APP by subtracting the plate-specific mean from each observation and dividing by the plate-specific standard deviation. We created quintiles of each APP, using the distribution of z-scores among unaffected individuals to set cut-offs.
We examined the association of each covariate with the APP z-scores, separately for ASD cases and controls, estimating mean APP z-score over categories of the covariates with linear regression, followed by a Wald test as a general test of the association between the APP and the covariate.
We used conditional logistic regression models stratified by plate to calculate the odds of ASD associated with each APP. For categorical analysis, the middle quintile was the referent group. For continuous analysis, we used restricted cubic spline models with three knots. Adjusted models included terms for: maternal age, psychiatric history, country of origin, and hospitalization for infection during pregnancy; parental income; children's birth order, sex, gestational age at birth, size for gestational age, and mode of delivery. We tested for interactions between APP levels and the covariates, comparing models including the cross-product terms to those without the cross-product terms using a likelihood ratio test. Matched sibling analysis used conditional logistic regression models, clustered by family, and adjusted for sex, birth order, and total protein concentration.
Overall the patterns of association with APP were similar among the stratified outcomes. The odds associated with the highest quintile of CRP were greatest for ASD only (1.80, 1.15-2.82) compared to ASD with co-occurring ID (1.29, 0.83 -2.01) or with co-occurring ADHD (1.34, 0.85-2.11), with similar results in continuous analyses ( Figures S7-10).

APP and Odds of ASD Among Matched Sibling Pairs
In contrast to the case-control comparison, we found that median levels of all APP except CRP, HAP, and PCT were lower in ASD cases compared to their unaffected siblings (Table S3), with distinct distributions of particularly A2M, SAP, and FER among ASD cases compared to their unaffected siblings ( Figure S5). In matched regression analyses, we observed decreasing odds with increasing levels of A2M, FER, SAA, SAP, and tPA ( Figure 3, Figure S6B). Associations with SAA, SAP, and tPA were attenuated in adjusted models, though the pattern associated with A2M and FER persisted ( Figure 3). Compared to the middle quintile, the lowest quintile of A2M (3.71, 1.21 -11.33), FER (4.20, 1.40 -12.65), and SAP 3.05 (1.16 -8.01) were associated with increased odds of ASD in the matched sibling comparison ( Figure S6B).

Interaction Between APP and Other Risk Factors for ASD
Since many of the covariates we studied (e.g., maternal infection, mode of delivery) represent environmental influences that may influence the innate immune system and are also risk factors for ASD, we hypothesized that variation in the innate immune response to these environmental influences may be associated with risk for ASD. We observed interactions between APP and a number of other risk factors for ASD in our study ( Figure 4, Figure S11, Table S7). We observed interactions between maternal hospitalization for infections during the third trimester and neonatal levels of APP ( Figure 4).
Lower odds of ASD were associated with increasing levels of A2M, FER, FIB, PCT, and tPA among children whose mothers were hospitalized for infection late in pregnancy ( Figure 4A-E), in contrast to those whose mothers were not hospitalized for infection. Similarly, we observed lower odds of ASD associated with increasing levels of FER and SAP among children whose mothers were diagnosed with anemia during pregnancy, in contrast to children whose mothers were not diagnosed with anemia ( Figure 4F-G). The relationship between elevated CRP and ASD was stronger among children whose mother had no history of psychiatric illness compared to those whose mother did have a psychiatric history ( Figure 4H).

Sensitivity analyses
Sensitivity analyses were designed to evaluate whether variation in APP measurements may have influenced results. Estimates from models in sensitivity analyses were similar to the results of the main analysis ( Figure S12).

Discussion
Our results indicate that of the innate immune markers that we studied, only CRP was associated with risk for ASD, with levels of CRP above the population mean associated with increased odds of ASD even after accounting for a range for potential confounders. We observed significant interactions between environmental exposures that could plausibly influence the neonatal innate immune system and levels of APP. With the exception of PCT, higher levels of these markers also tended to be associated with lower risk of ASD when comparing ASD cases to their unaffected siblings.

Strengths and Weaknesses
The setting of the study within a universal healthcare system with developmental screening provided to families free of charge increases the likelihood that cases of ASD are identified, using validated ascertainment methods (21). We included two comparison groups in the study. Unaffected siblings share aspects of the early life environment with ASD-affected siblings as well as on average 50% of their common genetic variants. The large population-based comparison group of unrelated individuals randomly sampled from the larger cohort included in the study is also a strength, allowing us to study the relationship between a large number of potential confounders and neonatal levels of APP. While the use of multiple markers of the innate immune response and information on many putative risk factors for ASD is a strength of this study, there is still a possibility for residual confounding and the analysis of such markers results in multiple statistical comparisons.
The use of APP as markers of the neonatal innate immune response is also a strength of this study.
Therefore, it is difficult to interpret the "cytokine profile" with little information about the environmental factors that may influence such levels, with some studies indicating opposing effects of the same cytokines (10,11,33).

Interpretation
APP are a diverse group of proteins that collectively recognize and opsonize invading pathogens and cellular debris, regulate blood viscosity and clotting, and sequester nutrients (e.g., iron) from pathogens (36). Moreover, many of the APP play key roles in the regulation of the innate and adaptive immune responses and are required for resolution of an acute inflammatory response in experimental models (37-40). Expression of APP occurs primarily in the liver, is influenced by genetic background, and increases by orders of magnitude in response to IL-6 and other inflammatory cytokines (41-44).
Fetal liver produces APP, with expression detectable from early in the second trimester (45). APP are not believed to cross the placenta (46). We consider APP in NDBS as specific indicators of the neonatal innate immune status.
CRP is a pattern recognition molecule, opsonin, and activator of the complement pathway (39). Of the APP studied, CRP had the strongest association with risk of ASD in the case-control comparison, displaying a U-shaped association with odds of ASD. While the shape of the association between CRP and odds of ASD in the sibling comparison was similar, the association was less apparent with considerably wider confidence intervals in the sibling analysis, possibly reflecting shared common genetic variation in innate immune signaling and risk of ASD (8). We also observed a weaker relationship between CRP and risk of ASD among those whose mothers had a previous history of psychiatric illness in the case-control comparison, with a significant interaction detected between maternal psychiatric history and levels of CRP. Environmental stimuli that increase CRP production may be more important to the etiology of ASD among individuals with lower genetic liability for the disorder. Moreover, both positive and negative genetic correlations between psychiatric disorders (e.g., bipolar disorder, schizophrenia) and CRP levels have been reported (47, 48). Thus, levels of CRP in children born to mothers with a high genetic risk for psychiatric illness may exhibit larger variation than children to unaffected mothers. Levels of multiple APP were elevated among neonates whose mothers were hospitalized for infections in late pregnancy, although these markers were not elevated to the same extent among children later diagnosed with ASD. In interaction analyses, higher levels of A2M, FER, FIB, PCT, and tPA were associated with lower risk of ASD among those exposed to maternal infections in late pregnancy. The actions of the APP indicated here are diverse: A2M, FIB, and tPA regulate blood coagulation, while FER binds to iron to limit the growth of pathogens. The physiological function of PCT is poorly defined, but its rapid rise in response to bacterial infections makes PCT useful when diagnosing sepsis, particularly in neonates (49). We previously reported that low levels of APP in combination with maternal exposure to cytomegalovirus and Toxoplasma gondii were associated with greater risk for non-affective psychosis compared to those with higher APP levels (50). Taken together with the results of the current study, this suggests that the degree of protection a neonate's innate immune system offers against an environmental insult may be important in modulating the influence of such insults on neurodevelopmental processes. Of the APP that interacted with infection in terms of risk for ASD, higher levels of all but PCT were also associated with lower odds of ASD in the sibling comparison study. In other words, we observe that, within matched pairs, the sibling who produced higher levels of these APP in response to a presumably similar early life environment was less likely to develop ASD.
While FER concentrations rise to withhold iron from pathogens during the acute phase response (36), in the absence of this response, FER reflects the body's store of iron in both adults and neonates (51).
We recently reported that maternal diagnosis with anemia was associated with ASD, ID, and ADHD in offspring (26). The results of that study indicated that the increased risk associated with maternal anemia may be limited to cases where anemia was severe and long lasting. Here we report evidence that among individuals whose mothers were affected by anemia during pregnancy, higher levels of ferritin were associated with lower odds of ASD. These results suggest that among mothers affected by anemia, an adequate iron store in the fetus may be protective with regard to risk of ASD. We also observed a similar protective association with increasing levels of ferritin when comparing ASD cases to their unaffected siblings, in line with sibling comparisons of exposure to maternal anemia (26).
Numerous studies have posited adverse effects of intrauterine inflammatory signals to explain the associations consistently observed between elevated maternal BMI and risk for ASD (52-55). Here we report that maternal BMI was not associated with APPs in cases or controls measured only a few days after birth, raising the question of whether this particular mechanism to link maternal BMI to risk of ASD is relevant, at least in later pregnancy.

Conclusion
Together, our findings suggest that it is not the levels of inflammatory or regulatory markers per se that may influence ASD risk. Rather, the strength of those signals in the genetic and environmental context of the developing nervous system must be considered. Correspondingly, understanding on a molecular level the individual's ability to resolve environmental insults may help to identify those that are particularly susceptible to such insults.      shown separately for those who were not exposed to maternal hospitalization for infection in late pregnancy (3 rd trimester) compared to those who were (A-E), those who were not exposed to maternal anemia compared to those who were (F-G), and those whose mother had no history of psychiatric diagnoses compared to those whose mother did have a psychiatric history (H). Solid lines represent the OR estimate for each group and dashed lines represent the 95% confidence interval for the fully adjusted model. P-values for interaction (a likelihood ratio test comparing a model with interaction terms to a model without interaction terms) are shown.              Table S1. Diagnostic codes used in the ascertainment of ASD, ADHD, and ID in the Stockholm Youth Cohort. Table S2. Quality control statistics for multiplex assays to analyze acute phase protein concentrations in neonatal dried blood spots. Table S3. Characteristics of unaffected and ASD-affected siblings included in the matched sibling comparison. Table S4. P-values for the association of APP with other covariates among 1092 unaffected individuals. Table S5. P-values for the association of APP with other covariates among 924 ASD affected individuals. Table S6. P-values for non-linearity in restricted cubic spline analyses. Table S7. P-values for interaction using a likelihood ratio test comparing a model without interaction terms nested in model with interaction terms. 2 with total protein value below limit of detection Figure S2. Levels of total protein (A) and one APP (B Ferritin) according to analytic plate and elution batch. Protein elution batches are demarcated by red dotted lines in A. Plate-specific z-scores were calculated to allow comparisons across batches and plates (C).  Figure S5. The distribution of each APP measured in neonatal dried blood spot samples, comparing ASD-unaffected individuals (0) to ASD-affected individuals (1). In (A), 924 ASD cases are compared to 1092 unaffected individuals selected from the cohort. In (B), 203 ASD cases are compared to their unaffected siblings. Figure S6. The relationship between APP and odds of ASD when comparing 924 ASD cases to 1092 unaffected individuals selected from the cohort (A) and when comparing 203 ASD cases to their matched, unaffected siblings (B). Quintiles of each APP were created using the distribution of z-scores among unaffected individuals to set the cut-offs and the middle quintile was used as the referent category. Results are shown for unadjusted estimates (open circles) and fully adjusted estimates (solid circles). Models were adjusted for maternal age, psychiatric history, country of origin, and hospitalization for infection during pregnancy; children's birth order, sex, gestational age at birth, size for gestational age, and mode of delivery. P-values are shown for a Wald test with a null hypothesis that all APP categorical terms were jointly equal to zero, as a test of whether each APP was generally associated with the outcome. Truncated confidence intervals in the matched sibling comparison are indicated with an arrow. Abbreviations: A2M: α-2 macroglobulin; CRP: C reactive protein; FER: ferritin; FIB: fibrinogen; HAP: haptoglobin; PCT: procalcitonin; SAA: serum amyloid A; SAP: serum amyloid P; and tPA: tissue plasminogen activator. Figure S7. The relationship between APP and odds of ASD when the outcome is stratified by cooccurring ID and ADHD. Results are shown for ASD with co-morbid ID, ASD with co-morbid ADHD, and ASD without co-occurring ID or ADHD ("ASD only"), compared to 1092 unaffected individuals selected from the cohort. Quintiles of each APP were created using the distribution of z-scores among unaffected individuals to set the cut-offs and the middle quintile was used as the referent category. Models were adjusted for maternal age, psychiatric history, country of origin, and hospitalization for infection during pregnancy; children's birth order, sex, gestational age at birth, size for gestational age, and mode of delivery. P-values are shown for a Wald test with a null hypothesis that all APP categorical terms were jointly equal to zero, as a test of whether each APP was generally associated with the outcome. Abbreviations: A2M: α-2 macroglobulin; CRP: C reactive protein; FER: ferritin; FIB: fibrinogen; HAP: haptoglobin; PCT: procalcitonin; SAA: serum amyloid A; SAP: serum amyloid P; and tPA: tissue plasminogen activator. Figure S8. The relationship between APP and odds of ASD with co-occurring ID when comparing 321 individuals affected by ASD with co-occurring ID to 1092 unaffected individuals selected from the cohort. Each panel displays the odds of ASD according to APP z-score, flexibly fit using restricted cubic spline models with three knots and a z-score=0 as the referent. The dashed line represents the unadjusted estimate of the relationship between each APP and odds of ASD. The solid line represents the fully adjusted model, adjusted for maternal age, psychiatric history, country of origin, and hospitalization for infection during pregnancy; children's birth order, sex, gestational age at birth, size for gestational age, and mode of delivery. The gray bands represent the 95% confidence interval for the fully adjusted model. P-values are shown for a Wald test with a null hypothesis that all APP spline terms were jointly equal to zero, as a test of whether each APP was generally associated with the outcome. P values are shown for spline terms in both the unadjusted models (p) and the adjusted models (padj). Abbreviations: A2M: α-2 macroglobulin; CRP: C reactive protein; FER: ferritin; FIB: fibrinogen; HAP: haptoglobin; PCT: procalcitonin; SAA: serum amyloid A; SAP: serum amyloid P; and tPA: tissue plasminogen activator. Figure S9. The relationship between APP and odds of ASD with co-occurring ADHD when comparing 322 individuals affected by ASD with co-occurring ADHD to 1092 unaffected individuals selected from the cohort. Each panel displays the odds of ASD according to APP z-score, flexibly fit using restricted cubic spline models with three knots and z-score=0 as the referent. The dashed line represents the unadjusted estimate of the relationship between each APP and odds of ASD. The solid line represents the fully adjusted model, adjusted for maternal age, psychiatric history, country of origin, and hospitalization for infection during pregnancy; children's birth order, sex, gestational age at birth, size for gestational age, and mode of delivery. The gray bands represent the 95% confidence interval for the fully adjusted model. P-values are shown for a Wald test with a null hypothesis that all APP spline terms were jointly equal to zero, as a test of whether each APP was generally associated with the outcome. P values are shown for spline terms in both the unadjusted models (p) and the adjusted models (padj). Abbreviations: A2M: α-2 macroglobulin; CRP: C reactive protein; FER: ferritin; FIB: fibrinogen; HAP: haptoglobin; PCT: procalcitonin; SAA: serum amyloid A; SAP: serum amyloid P; and tPA: tissue plasminogen activator. Figure S10. The relationship between APP and odds of ASD without co-occurring ID or ADHD ("ASD only") when comparing 281 individuals affected by ASD and without a diagnosis of co-occurring ID or ADHD to 1092 unaffected individuals selected from the cohort. Each panel displays the odds of ASD according to APP z-score, flexibly fit using restricted cubic spline models with three knots and a z-score=0 as the referent. The dashed line represents the unadjusted estimate of the relationship between each APP and odds of ASD. The solid line represents the fully adjusted model, adjusted for maternal age, psychiatric history, country of origin, and hospitalization for infection during pregnancy; children's birth order, sex, gestational age at birth, size for gestational age, and mode of delivery. The gray bands represent the 95% confidence interval for the fully adjusted model. P-values are shown for a Wald test with a null hypothesis that all APP spline terms were jointly equal to zero, as a test of whether each APP was generally associated with the outcome. P values are shown for spline terms in both the unadjusted models (p) and the adjusted models (padj). Abbreviations: A2M: α-2 macroglobulin; CRP: C reactive protein; FER: ferritin; FIB: fibrinogen; HAP: haptoglobin; PCT: procalcitonin; SAA: serum amyloid A; SAP: serum amyloid P; and tPA: tissue plasminogen activator. Figure S11. Interaction models for categorical variables. Odds ratios for ASD over the range of APP are shown separately according to strata of parental education level (A), Apgar score at 5 minutes (B), and parental income at birth (C). Solid lines represent the OR estimate for each group and dashed lines represent the 95% confidence interval for the fully adjusted model. P-values for interaction (a likelihood ratio test comparing a model with interaction terms to a model without interaction terms) are shown. Abbreviations: SAA: serum amyloid A; SAP: serum amyloid P; and tPA: tissue plasminogen activator. Figure S12. The relationship between APP and odds of ASD in sensitivity analyses. In (A), 924 ASD cases are compared to 1092 unaffected individuals selected from the cohort, adjusting for birth year, total protein content of the eluate from the neonatal dried blood spot sample, and age (in days) at neonatal blood sample, in addition to the covariates of the fully adjusted model (maternal age, psychiatric history, country of origin, and hospitalization for infection during pregnancy; children's birth order, sex, gestational age at birth, size for gestational age, and mode of delivery). In (B), individuals with a total protein concentration in the lowest quartile were excluded, and results are shown for the fully adjusted model comparing 688 ASD cases to 819 unaffected individuals selected from the cohort. In (C), results are shown for the fully adjusted model using a multi-level modeling approach that accounts for the clustering of APP observations in elution batches and multiplex analysis plate, comparing 924 ASD cases to 1092 unaffected individuals. We used generalized linear and mixed models to fit a 3-level random intercept model with logit link, allowing for the intercepts of both plates and batch to vary, but assuming that the relationship between the log-odds of ASD and each APP does not vary across plates or batches.     Table S6. P-values for non-linearity in restricted cubic spline analyses. Not all relationships we considered are necessarily non-linear. To assess evidence for non-linear relationships, we tested the null-hypothesis that all spline terms that would indicate a change in the slope of the relationship (i.e., all but the first spline term) are equal to zero using a Wald test.