Participants were recruited, screened, and assessed at West China Hospital of Sichuan University. The research protocol was approved by the Medical Ethical Committee of West China Hospital of Sichuan University, and parents provided written informed consent after receiving a detailed description of the study. Data were used for research purposes only.

182 individuals with ASD were included in this study, all of whom diagnosed by one professional child psychiatrist based on the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5)(1). Participants were excluded if they met any of the following criteria: (1) neurological disorders, such as epilepsy, encephalitis; (2) intelligence quotient< 70; (3) history of craniocerebral injury; (4) monogenetic diseases, such as fragile X syndrome, tuberous sclerosis, and Rett syndrome; (5) taking psychiatric medications during assessment. The common comorbid disorders such as attention deficit hyperactivity disorder, tic disorders and emotional disorders are not the exclusion criteria. After excluding ineligible subjects, 104 participants remained. Details of participants enrolled in this study and the study process had also been described in our previous published paper (31).

After recruitment, a parent interview and a child assessment were conducted using the autism diagnostic interview-revised (ADI-R) (33) and the autism diagnostic observation schedule (ADOS) (34) respectively to confirm ASD diagnosis. At baseline (time 1), all participants were required to take an intelligence quotient (IQ) test and have an MRI scan after receiving their diagnoses, if available, collecting blood of participants and their family members. In this study, 59 blood samples of ASD were collected. Around 4 years after enrollment (time 2), participants were requested to take part in a follow-up assessment. In this study, 90 children with ASD completed the follow-up assessment, according to ADOS total score, 30 participants achieving optimal outcomes (ADOS total score< 7) (labeled as ASD-), other 60 participants achieving poor outcomes (ADOS total score ≥ 7) (labeled as ASD+). See Figure 1 for the flowchart of the participant recruitment.

ADI-R, developed and revised by Le Couteur et al. (33), is a standardized, structured interview instrument to evaluate autism symptoms. It consists of four subscales: communication, social interactions, restricted and repetitive behaviors (RRB) and evidence of early developmental abnormalities (35). ADOS, revised by Lord et al. (34), is a standardized structured interactive autism symptom rating tool revised. Individuals were assessed by observing the ASD-related symptoms during the game interaction. It consists of 4 subscales: communication, social interaction, imagination/creativity, and RRB. The total score greater than or equal to 7 is the diagnostic threshold. It is the most used “gold standards” for the diagnosis of ASD in clinical and scientific research, which was also used as a criterion for the severity of symptoms in follow-up assessment of ASD children in this study.

The Chinese-Wechsler Intelligence Scale for Children (C-WISC) (36) was used to assess the intellectual development. The full-scale intelligence scale includes 11 sub-tests. In this study we employed 4-in-1 short version recommended in the appendix of the manual, including knowledge, comprehension, picture completion, and block drawing.

Infants-Junior Middle School Students’ Social-Life Abilities Scale (S-M) (37) and overall social outcome (OSO) ratings (6, 7) were used to evaluate the social adaptability of participants. The S-M scale consists of 132 questions in six domains. After the completion of the questionnaire, a standardized score based on age was obtained. Standard scores ranged from 5 to 13, and the lower the scores, the worse the social adaptability. The OSO scoring system was derived from Rutter’s non-specific scoring criteria for the outcome of psychiatric disorders (38), which introduced operational scoring rules and focused on the domains of independent living, friendship and career (6, 7). Finally, the social adaptability of individuals was graded trichotomy as very good/good, average, and poor/very poor.

Peripheral blood of 59 ASD subjects and their parents was collected, stored in the refrigerator at -20°C, and regularly sent to the laboratory for DNA extraction. The extracted DNA was stored in a refrigerator at -80°C and sent for the whole exome sequencing (WES).

The main steps of sequencing were as follows: 1) Quality inspection: The quality of DNA samples was tested to detect whether there was obvious DNA degradation and whether there was RNA and protein contamination; 2) Library construction: DNA samples with content above 0.6ug of Agilent SureSelect Human All Exon V6 liquid capture system were effectively enriched to create sequence libraries; 3) Sequencing: high-throughput and deep sequencing was performed on the Illumina HiSeq 4000 platform (Illumina, Inc., San Diego, CA, USA). The preprocessing steps of WES data included extraction, quality control and typing annotation. Details are shown in Supplementary Materials . We focused on the polygenic risk of rare mutations, so only mutations with minor allele frequency (MAF)< 0.01, according to the east Asians from the 1000 Genomes Project database (N=504) (39) and the Exome Aggregation Consortium (ExAC) database (N ~= 3000) (40), were retained for subsequent analysis.

Polygenic risk score is a comprehensive assessment of the cumulative effects of multiple variants with weak effects on related diseases to evaluate the genetic risk of developing a disease (41). According to the procedure recommended by PRSice-2, the additive model was used to calculate the exome score (42). We followed the approach developed by Chiara Fabbri (43) to calculate ASD EFRS based on the WES data, and to obtain the load scores of the whole exome rare variants. The EFRS is calculated using the following formula:

where n is the number of genetic variants within the whole exome, v_all is the number of alternative alleles, w_s is the corresponding functional score of the gene variant, and w_f is the frequency weight for that variant. By weighting function and frequency simultaneously, EFRS does not depend on the presence of individual variants which could not be observed in some of the tested samples, and thus keeps the final score stable and reliable (44). Different sources (LRT, Mutation Assessor, Polyphen-2, SIFT and CADD (45–48), see Supplementary Materials ) were tested in this study, to determine the damaging of mutations for functional scores (w_s). The frequency weighting (w_f) was determined based on the mean frequency of east Asian populations in the 1000 Genomes Project (https://www.internationalgenome.org/) and ExAc databases (http://exac.broadinstitute.org). The alternative alleles (v_all) were determined based on the mutation sites of ASD identified in a WES study of 175 trios published in Nature in 2012 (49).

All MRI data were collected on a 3T scanner (Philips, Achieva, TX, Best, The Netherlands) at the Tibet Chengban Branch of Sichuan University West China Hospital. Two modalities [structural MRI (sMRI) and resting state fMRI (rs-fMRI)] images were acquired.

sMRI (T1 weighted) images were scanned using a three-dimensional spoiled gradient recalled echoing planar imaging sequence. Detailed scan parameters are described as follows: repetition time, 8.2 msec; echo time, 3.8 msec; flip angle, 7°; slice thickness, 1 mm; field of view, 256 mm × 256 mm; matrix size, 256 × 256; voxel size, 1 × 1 × 1 mm³. Bold signals from the rs-fMRI modality were acquired using a single-excitation gradient echo planar imaging (EPI) sequence with scanning parameters described as follows: echo time, 30ms; repeat time, 2000ms; flip Angle, 90°; FOV = 240mm×240mm; slice thickness, 4mm; gap, 0mm. Each time point was scanned continuously, and a total of 240 time points were acquired.

Raw data acquired from the MRI scanner (DICOM files) were converted from raw DICOM files into analyzable NIfTI images using dcm2niigui software. Use the SPM software package in MATLAB R2013b platform (https://www.fil.ion.ucl.ac.uk/spm/software/spm12/), FreeSurfer package (https://www.freesurfer.net/) and FSL software (http://surfer.nmr.mgh.harvard.edu/fswiki/Fsl) to analyze imaging data processing.

The whole brain functional connectivity (FC) analysis method based on correlation analysis proposed by Salvador (2005) was used for the construction of FC network (50). Individual differential structural covariance network (IDSCN) was constructed to analyze the common structural changes among the brain regions. Details of data preprocessing and description of the construction of the connectivity networks are provided in the Supplementary Materials .

Firstly, through simple logistic regression analysis, we examined the relationship between EFRS and outcomes situations (autism symptoms severity and social adaptability). Then, based on the brain connectivity networks, we used ML method SVR to identify the brain regions related to the outcomes of ASD. And then, we explored the relationships among EFRS, brain and outcomes, through multivariate logistic regression analysis with post hoc Bonferroni correction, significance level 0.05. Finally, taking the targeted brain regions as a mediator, we constructed a mediation model of “gene-brain-behavioral” to analyze how EFRS affects the behavior of ASD patients through the interaction of brain connectivity. At the same time, imaging genetic biomarkers related to the outcomes of ASD were screened and confirmed.

After excluding ineligible individuals, there were 104 children with ASD enrolled. The demographic information is shown in Table 1 . The average age is 8.01(SD = 3.25) years. Four years later after enrollment, 90 individuals were followed up for clinical diagnostic assessment (evaluated by ADOS) and social adaptability (S-M and OSO). Among the 90 individuals, 30 of them (33.33%) showed optimal outcome (labeled as ASD-), whose ADOS total scores less than 7, losing the diagnoses of ASD; while the other 60 subjects’ diagnoses of ASD persistent (labeled as ASD+). Independent sample t test was employed to identify the clinical differences between ASD- group and ASD+ group at baseline (Time 1), and the results showed that there is no difference in age, IQ, gender and ADOS scores (all p > 0.05). While at follow-up (Time 2), significant differences are found in IQ, ADOS, S-M and OSO between ASD- group and ASD+ group (all p< 0.001) (see Table 1 ).

We calculated the EFRS of the 59 ASD children according to Chiara Fabbri’s method (43), to explore the genetic liability of the outcomes of ASD children. The EFRS values are shown in Supplementary Table 1 . Among the 59 individuals with WES data, 50 of them achieved follow-up: 15 individuals (30%) had optimal outcome (ASD-), ADOS total score less than 7; the remaining 35 (70%) had ASD diagnoses persistent with poor outcome (ASD+).

Simple regression analysis showed that EFRS could predict social adaptability of ASD children after 4 years later. When taking social adaptability (evaluated by S-M) as the dependent variable and EFRS as the independent variable, the results demonstrated that EFRS could act as an independent predictor of the social adaptability (Y = -0.013*X + 9.29, p = 0.003) (see Figure 2A ). However, EFRS could not be an independent predictor of the severity of autism symptoms (Y = -0.0008*X + 0.95, p = 0.914) (see Figure 2B ). EFRS was not significantly associated with the total score of ADOS neither at baseline (r = -0.108, p = 0.538) nor at follow-up (r = 0.065, p = 0.653) (see Supplementary Figure 1 ).

We built SVR models based on structurally and functionally brain connectivity, IDSCN for structural MRI and FC for functional MRI, to precisely identify the outcome predictors of ASD children. Two candidate indexes were included in the SVR models: 1) symptoms severity: ADOS total score at time 2, and the outcome grouping (ASD+ vs ASD-); 2) social adaptability: S-M standard score and OSO grade at time 2.

SVR models for ADOS total scores shown that FC could predict ADOS total scores (the prediction function: Y = 0.30X + 7.67, R² = 0.13, MSE = 79.25, p = 0.08) (see Supplementary Figure 2 ). There are totally 19 pairs of FC with predictive effects; 10 pairs of those show negative weight, predicting a decrease in ADOS scores; while the other 9 pairs show positive weight, predicting an increase in the ADOS total scores (see Figure 3A ). Details of the 19 pairs of FCs are shown in Table 2 . IDSCN has no predictive effect on ADOS total scores.

SVR models for social adaptability of ASD children showed that both FC (fMRI) and IDSCN (sMRI) have good predictive effects. The predictive function of FC for the S-M standard score is: Y = 0.42*X + 4.94, R² = 0.13, MSE = 2.33, p = 0.03 (see Supplementary Figure 2 ). A total of 10 pairs of FC were identified to have predictive effects, of which 4 pairs of FCs have positive weight values, indicating that increased FC of these brain regions could predict better social adaptability after 4 years; the other 6 pairs of FC have negative weight values, suggesting that increased FC of these brain regions might predict worse social adaptability after 4 years (see Figure 3B , Table 3 ). The predictive function of sMRI for the S-M standard score is: Y = 0.30*X + 5.97, R² = 0.17, MSE = 2.66, p = 0.06 (see Supplementary Figure 2 ). A total of 4 pairs of IDSCN were found to have predictive effects with positive weight values (see Figure 3C , Table 3 ).

We further tested the extent to which EFRS and brain connectivity identified above analysis could predict outcomes of ASD when taking them together as independent variables. The brain connectivity associated with outcomes found by the SVR models (see Tables 2 , 3 ) and EFRS were included in the multiple regression model to analyze the joint predictive value of the baseline genetic risk and brain imaging characteristics on the outcomes of ASD.

In the analysis of the autism symptom severity (outcome grouping based on ADOS scores) at follow-up, 19 pairs of FC identified by SVR based on ADOS total score (see Figure 3A , Table 2 ) and EFRS were included as predictive variables. The analyses of multiple logistic regression did not identify any brain connectivity that could predict the outcome grouping of ASD children at 4-year follow-up (see Figure 3A ); In the analysis of social adaptability (OSO grade) at follow-up, the brain connectivity characteristics determined by SVR based on S-M (see Figure 3B , Table 3 ) and EFRS were included. The results of the predicting effects of both fMRI (FCs) and sMRI (IDSCN) on social adaptability were verified. As shown in Figures 4A, B , 3 of the 4 pairs of IDSCN perform predictive effects, as predictors of poor outcome; 7 of the 10 pairs of FC perform predictive effect, 3 pairs of those as predictors of poor outcome, the other 4 pairs as predictors of good outcome. In conclusion, both FC and IDSCN, the two modalities of brain connectivity could predict the social adaptability of ASD children. The characteristic brain regions mostly located in superior frontal gyrus, occipital cortex, temporal cortex, parietal cortex, paracentral lobule, pallidum, and amygdala for functional brain connectivity, and temporal cortex, thalamus, and hippocampus for structural brain connectivity.

In addition, we employed SVR to identify the characteristic brain regions related to ASD EFRS, to explore the relationship between brain connectivity and EFRS. The SVR models based on FC (fMRI) and IDSCN (sMRI) were constructed taking EFRS as a predictor variable. SVR model construction based on fMRI modality succeeded (regression function: Y = -1.64*X + 154.7, p = 0.05, R = -0.32, MSE = 3168.80), meaning that ASD EFRS has significantly predictive effects on the values of FC; while SVR model based on sMRI failed (regression function: Y = -0.09X + 64.34, p = 0.636) (see Supplementary Figure 3 ). There are 10 pairs of FC identified that are associated with EFRS, as shown in Table 4 .

To further explore the relationship of ASD EFRS and brain regions identified above, and the outcomes of ASD children. We took OSO grade and outcome grouping as dependent variables respectively, and characteristic brain regions related to EFRS (see Table 4 ) as independent variables in the multiple logistic regression analysis, and the results showed that FC of left middle occipital gyrus (MOG.L) and right paracentral lobule (PCL.R) (OR = 7.08, 95%CI = 5.32 - 12.03) is a risk factor for social adaptability of ASD children (see Figure 4C ). The logistic regression model taking outcome grouping as the dependent variable was not fitted successfully.

To further explore the relationships of EFRS, brain connectivity, and symptoms of ASD children, we constructed a mediation model using structural equation modeling. The results show that EFRS mediates the communication (ADIR subscale) through the FC of left middle occipital gyrus (MOG.L) and left pallidum (PAL.L) (CMIN = 80.66, DF = 42, p< 0.001, CFI = 0.867, AIC = 152), as shown in Figure 5 . There is no direct effect between EFRS and the communication (r = -0.10, p = 0.462), but an indirect effect (r = 0.17, p = 0.03).

In this study, 33.33% of children with ASD achieve optimal outcome, losing the diagnosis for ASD. The rate of recovery is in the range of 0–37% mentioned above (9–13), consistent with the results of previous studies. However, we exclude these ASD children with IQ< 70 and comorbid other neuropsychiatric disorders, meaning that the truly rate of recovery is much lower than 33.33%, thus requiring larger samples to replicate. Children with optimal outcome (ASD- group) also perform better social adaptability than those with poor outcome (ASD+ group), which is consistent with the results of Harstad’ study that 37% of the children who were clinically diagnosed with ASD at 12 to 36 months did not continue to meet diagnostic criteria for ASD at 5 to 7 years of age, and emphasized that the most related factor with the nonpersistent ASD was adaptive skills (12).

We found that ASD EFRS could predict social adaptability of children with ASD 4 years later, suggesting that outcome conditions of ASD children are affected by genetic regulation. A large population-based study from the UK Biobank found that social-isolation polygenic risk score (PRS) predicted friendship at 18 years old, demonstrating that the genetic factors are able to predict related social traits (51), which is agreement with our results. Social adaptability is an important prognostic indicator. A previous study has demonstrated that PRS could predict the severity of ASD (27), while in our study, we failed to draw this conclusion; Our results shown that exome risk score could not predict the severity of autism symptoms, which could be due to the small cohort size of this study.

We identified 19 pairs of FC associated with autism symptoms severity at follow-up ( Table 2 ), 10 pairs of FC and 4 pairs of IDSCN associated with social adaptability at follow-up ( Table 3 ), and 10 pairs of FC associated with ASD EFRS ( Table 4 ) by SVR models. Associated brain regions with prognostic predictive effects are mainly distributed in the extensive cortex regions and pallidum and amygdala of subcortical regions for FC, and the temporal cortex, thalamus, and hippocampus for IDSCN. The big differences between FC and IDSCN may stem from the limited cohort size, while in the other hand, which might reflect the intrinsically heterology in brain connectivity revealed by FC and IDSCN. In our study, FC obtained more significant results than IDSCN, which supports the results of Traut who claimed that functional MRI was more important for prediction than structural MRI (52). One of the reason is that FC reflects the organization and inter-relationship of spatially separated brain regions (53), alteration in brain FC is expected to provide potential biomarkers for classifying or predicting brain disorders. Previously, Guo identified two ASD subtypes based on the inter-individual deviation of FC patterns, which could predict the severity of social communication impairments and the severity of restricted and repetitive behaviors in ASD (54). Buch identified three latent dimensions of functional brain network connectivity that predicted individual differences in ASD behaviors (55). Our results provide further evidence that FC could predict social adaptability of ASD children at 4 years’ follow-up, involved brain regions including occipital cortex, paracentral lobule, temporal cortex, amygdala, superior frontal gyrus, parietal cortex ( Figures 4B, C ).

Structural covariance network (SCN) was introduced to explore the network level alterations (56), representing the covariance of morphological characteristics between regions, reflecting anatomical correlations in brain structure between brain regions (56–58). Previous researches about SCN based on gray matter density manifested that ASD showed greater covariance between right posterior cingulate cortex and right temporal compared with typically developmental controls (56, 59). In this study, we introduced individualized differential SCN (IDSCN), aiming to explore the heterogeneity of ASD from a structural perspective using SCN constructed at the individual level. Research mapped IDSCN based on Autism Brain Imaging Data Exchange (ABIDE) database to identify structural covariances, and found that IDSCN of ASD differed significantly from controls mainly involved frontal and subcortical regions (60). However, the research about relationship of the IDSCN and outcomes of ASD is few, our study fills this gap. In this study, we found IDSCN could also predict social adaptability of ASD children at follow-up, mainly involved thalamus, hippocampus, temporal cortex ( Figure 4B ). Further studies are needed to investigate the underlying mechanisms by which the brain predicts social adaptability of ASD.

We explored the relationship of EFRS, brain connectivity and behaviors of ASD children, and found that ASD EFRS could not independently affect autism symptoms, but through the mediation of the FC of middle occipital gyrus (MOG) and pallidum (PAL) (see Figure 5 ), indicating that FC of MOG and PAL may be involved the neuroimaging mechanism of social communication symptoms of ASD caused by polygenic genetic risk. Previous brain imaging studies have also revealed evidence for genetic variations in brain activities underlying behavior, for example, it has been found that both local network metrics of the right hippocampus and its functional connectivity with the basal ganglia and thalamus mediated the relationship between the oxytocin receptor gene and interdependence (61); our results provide additional evidence of mediation role of the FC of MOG and pallidum in the genetics of ASD. Likewise, our findings further suggest a pathway from EFRS to behaviors of ASD, mediated by the FC from MOG to the pallidum. This result indicates that FC of the MOG and pallidum has a genetic basis and that the MOG and pallidum are crucial to the predictive prognosis of children with ASD.

Due to rapid developments in genomics and imaging technologies, neuroimaging genetics studies of ASD have developed in the last few years. Neuroimaging genetics helps to identify ASD-risk genes that contribute to structural and functional variations in brain of ASD patients (55, 62, 63), providing a better understanding of the disorder’s neuropsychiatry, and helping identify targets for therapeutic intervention that could be useful for the clinical management of ASD patients. Approaches of integrating neuroimaging and genetics to link gene pathways to neurobiological and phenotypic heterogeneity can reveal subtype-specific gene-brain-behavior associations (64). Recent research leveraged this method to identify three robust dimensional biomarkers, that was gene, brain and behaviors, to parse heterogeneity in ASD into 4 subgroups (55, 65). Within each subgroup, ASD-related FC was explained by regional differences in the expression of distinct ASD-related gene sets (55). These studies provide evidence that polygenic variation in ASD may manifest as intermediate behavior-related brain circuits that give rise to distinguishable ASD subgroup phenotypes by modulating connectivity in ASD-related networks. However, neuroimaging genetics studies about prognosis of ASD is limited, this motivates future research evaluating the reproducibility, validity, and clinical utility of ASD dimensional and subtype models in the outcomes of ASD. In the long term, there is translational potential for prognosis and targeted pharmacological and circuit-based therapies for ASD.