Connectome-based prediction of the severity of autism spectrum disorder

Abstract Background Autism spectrum disorder (ASD) is characterized by social and behavioural deficits. Current diagnosis relies on behavioural criteria, but machine learning, particularly connectome-based predictive modelling (CPM), offers the potential to uncover neural biomarkers for ASD. Objective This study aims to predict the severity of ASD traits using CPM and explores differences among ASD subtypes, seeking to enhance diagnosis and understanding of ASD. Methods Resting-state functional magnetic resonance imaging data from 151 ASD patients were used in the model. CPM with leave-one-out cross-validation was conducted to identify intrinsic neural networks that predict Autism Diagnostic Observation Schedule (ADOS) scores. After the model was constructed, it was applied to independent samples to test its replicability (172 ASD patients) and specificity (36 healthy control participants). Furthermore, we examined the predictive model across different aspects of ASD and in subtypes of ASD to understand the potential mechanisms underlying the results. Results The CPM successfully identified negative networks that significantly predicted ADOS total scores [r (df = 150) = 0.19, P = 0.008 in all patients; r (df = 104) = 0.20, P = 0.040 in classic autism] and communication scores [r (df = 150) = 0.22, P = 0.010 in all patients; r (df = 104) = 0.21, P = 0.020 in classic autism]. These results were reproducible across independent databases. The networks were characterized by enhanced inter- and intranetwork connectivity associated with the occipital network (OCC), and the sensorimotor network (SMN) also played important roles. Conclusions A CPM based on whole-brain resting-state functional connectivity can predicted the severity of ASD. Large-scale networks, including the OCC and SMN, played important roles in the predictive model. These findings may provide new directions for the diagnosis and intervention of ASD, and maybe could be the targets in novel interventions.


Introduction
Autism spectrum disorder (ASD) is a mosaic of de v elopmental conditions c har acterized by earl y deficits in two behavioural domains: (i) difficulties in social communication and interaction and (ii) highl y r estricted, ster eotypic, and r e petiti ve behaviours (Yang, et al., 2023a(Yang, et al., , 2023 b) b).The diagnosis of ASD is typically made through a comprehensive evaluation conducted by an experienced clinician.Curr entl y, widel y used observ ational measur es for ASD diagnosis include the Autism Diagnostic Observation Schedule (ADOS) and the Autism Diagnostic Interview-Revised, which ar e consider ed the 'gold standard' in ASD-focused assessments.Additionall y, a standard e v aluation pr otocol often involv es assessing IQ and ada ptiv e behaviour (Braconnier & Siper, 2021 ).
Despite ongoing efforts, the diagnosis of ASD is still based on behavioural criteria.Ho w ever, significant advances in identifying br ain ima ging biomarkers and de v eloping tools to aid in the diagnosis of ASD have been made.Machine learning, a pattern recognition technique developed from artificial intelligence studies, has played a central role in this endea vour.T he central tenet of machine learning is to automate inductive reasoning to create new knowledge (i.e.learning), a process achieved by extracting general rules and patterns from large datasets (Tai et al., 2019 ).Many studies have used similar machine-learning techniques with connectivity data in a classification fr ame work to distinguish healthy control (HC) participants from patients, including those with ASD (Anderson et al., 2011 ;Plitt et al., 2015 ).Due to the high pr e v alence rate and heterogeneous nature of ASD, numerous machinelearning a ppr oac hes hav e been de v eloped ov er the last decade to investigate potential differences between ASD patients and neur otypical contr ols using functional ma gnetic r esonance ima ging (fMRI) data.Ho w e v er, a compr ehensiv e understanding of the network basis underlying ASD and the intricate interactions between and within different brain networks is still lacking from the perspective of brain networks.Identifying brain-based predictive models for ASD will not only enhance our current biological understanding of ASD pathophysiology, which can further refine existing interventions, but may also have potential for direct application in real-world clinical practice.
Connectome-based pr edictiv e modelling (CPM) is a machinelearning, data-driv en pr otocol that is used to de v elop pr edictiv e models of brain-behaviour relationships using whole-brain functional connectivity data ('connectomes') through cross-validation.Unlike traditional regression or correlation analyses, CPM does not necessitate a priori selection of networks, and it can identify specific regional connections associated with behaviours (Shen et al. , 2017 ;Zhou et al. , 2021 ).To ensure rigor and r epr oducibility, CPM incor por ates a built-in cr oss-v alidation a ppr oac h to avoid overfitting by testing model replication in a novel sample (Shen et al., 2017 ).This study used CPM to explore the brain network characteristics of participants with ASD for the following reasons: (i) CPM can r e v eal associations between br ain connectivity patterns and behavioural outcomes or clinical dia gnoses, whic h can be used to identify potential neural biomarkers and ASD-specific brain connectivity patterns .T his facilitates ASD diagnosis , allowing it to move beyond reliance on behavioural criteria.(ii) CPM focuses on anal ysing the entir e br ain network r ather than isolated br ain r egions .T his allows r esearc hers to understand the potential network basis and inter actions underl ying ASD fr om a lar ge-scale brain network perspective.(iii) As a machine learning method, CPM may be used to discover previously unknown or unexpected brain-beha viour associations .
Some scholars have already applied CPM to research in the field of ASD.Lake et al.'s study on tr ansdia gnostic social abilities r elated to ASD and attention-deficit/hyperactivity disorder (ADHD) a pplied CPM anal ysis to pr edict ADOS scor es of ASD patients in the ABIDE-I/II datasets (Lake et al., 2019 ).Rohr et al. used CPM to pr edict scor es on the Behavior Rating Inventory of Executive Function in participants from two sites in the Autism Brain Imaging Data Exchange II (ABIDE-II) and observed that the default mode netw ork (DMN) play ed a significant r ole in alter ations in inhibition and shifting (Rohr et al., 2020 ).Dufford et al. used CPM to predict the Social Responsiveness Scale scores of 144 participants in the Healthy Brain Network dataset, including 34 individuals with ASD (Dufford, 2022 ).They identified a common tr ansdia gnostic social impairment network that was rooted in social areas of the brain, the subcortex, and the salience network.
Corresponding to the definition of ASD, the ADOS-G (generic) is a standardized, semistructured assessment that consists of four modules designed to observe social inter action, comm unication, pla y, language , and the ima ginativ e use of materials by c hildr en, youth, and adults who may have ASD.The assessment includes 10-15 activities with 31 accompanying ratings, with questions related to social-emotional aspects and interview items about activities of daily living and additional tasks.Participants were tested at centres in the ABIDE using Modules 3 and 4 of the ADOS-G.Module 3 is intended for v erball y fluent c hildr en in whom playing with toys is a ge-a ppr opriate, while Module 4 is designed for verbally fluent adults and adolescents (usuall y ov er 12-16 years old) who do not have an interest in toys (Lord et al., 2000 ).The assessment provides a total score and subaspect scores, including social, comm unication, cr eativity, and ster eotyped beha viour.T his study aimed to explore the ability of CPM to predict ADOS total scores and subaspect scores in ASD patients.
In addition, ASD is part of a continuum of c har acteristics on a spectrum resulting from multiple nonlinear causative factors ( Araujo , n.d. ).T herefore , individuals with ASD exhibit a variety of clinical pr esentations, whic h makes the design of tests and subsequent inter pr etation of r esults c hallenging.According to the Diagnostic and Statistical Manual of Mental Disorders, fourth edi-tion (DSM-IV), ther e ar e thr ee subtypes of ASD: (i) classic autism (CA), (ii) Asper ger's syndr ome (AS), and (iii) perv asiv e de v elopmental disorder not otherwise specified (PDD-NOS) (Lord, 2012 ).In the current study, we were also interested in whether the different subtypes of ASD show significant differences in CPM c har acteristics.
In the current study, we used dimensional CPM to identify neural networks that predicted the severity of ASD traits (measured by the ADOS) using whole-brain network data.Next, we conducted migr ation v alidation with an independent dataset to investigate the stability of the model and to identify any significant differences between ASD patients and HCs to test the specificity of the CPM prediction network.As many previous studies have suggested that social communication impairment is the most typical phenotype of ASD patients, we hypothesized that the brain networks associated with social cognition and communication would predict the ADOS scores of ASD patients.

Participants
The dataset in our study was obtained from the ABIDE-I/II.The scanning parameters at different sites in the ABIDE dataset vary; for details, please refer to the official website ( http://fcon _ 1000.pr ojects.nitrc.org/ indi/ abide/ ).To shar e r esting-state functional MRI (rs-fMRI), anatomical, and phenotypic datasets with the br oader scientific comm unity, the ABIDE I involv ed 17 international sites and 1112 datasets, including 539 from individuals with ASD and 573 from HCs (aged 7-64 years, median age of 14.7 years acr oss gr oups).
Our exclusion criteria for participants were as follows: (i) without functional or structural images; (ii) with mixed handedness or without handedness information; (ii) without full-scale intelligence quotient (FIQ) information or had an FIQ < 70 (Floris et al., 2021 ); (iv) use of differ ent par ameters, suc h as eye status, repetition time, slice number, or data matrix size, from others used at that site; (v) use of time points different from others used at that site (time points varied between participants at the Stanford site, but all participants had at least 180 time points; consequently, 180 time points were ultimately used for all Stanford participants); (vi) signal losses r e v ealed during visual inspection of functional ima ges; (vii) scan dur ation < 100 time points (Van Dijk et al., 2010 ); (viii) head motion exceeding 2 mm or 2 • ; (ix) bad spatial normalization r e v ealed during visual inspection of functional ima ges; (x) scan cov er a ge of < 91% of the whole br ain; (xi) fr om sites with < 20 individual datasets at each step (UCLA_2 after poor spatial normalization, Caltech after inadequate cover); (xii) a DSM-IV score of −9999 (an invalid score); (xiii) without ADOS scores; or (xiv) mean fr ame wise displacement (mFD) is > 0.2 mm.Ultimately, a total of 359 participants, including 323 ASD patients (151 in model sample and 172 in validation sample) and 36 HCs, were included in our in vestigation.T he demogr a phic information of the included participants is shown in Table 1 .

MRI data preprocessing
Figure 1 depicts the research process.MRI data were preprocessed with Statistical P ar ametric Ma pping (SPM) v .12 ( http://www .fil.ion.ucl.ac.uk/ spm/ software/ spm12/ ) and DPABI V5.1 ( http://rfmri.or g/dpabi ).The pr epr ocessing steps included the following: (i) discarding the first 10 time points, (ii) slice timing and head motion correction, (iii) spatial normalization with the forw ar ds transformation field into the standard Montr eal Neur ological Institute (MNI) space (resampling voxel size was 3 × 3 × 3 mm 3 ), (iv) spatial smoothing with a 6 mm full-width at half-maximum Gaussian kernel, (v) r emov al of the linear trends of time courses, (vi) adjustment for nuisance covariates (i.e. head motion covariates with the Friston 24-parameter model as well as white matter and cer ebr ospinal fluid signals), (vii) eac h time point wher e the fr ame wise displacement (FD) exceeded 0.2 mm was used as a separ ate r egr essor for scrubbing, and (viii) a pplication of a bandpass temporal filter (0.01-0.08 Hz) to the time series (Fig. 1 A).

Functional network construction
Whole-br ain r esting-state functional connectivity (rsFC) was calculated using the GRETNA Toolbox ( https://www.nitrc.org/pr ojects/gr etna/) with the Dosenbach atlas, which consists of 160 r egions of inter est (ROI).Ev ery ROI in the atlas was a spherical region with a 5 mm radius, and all regions were assigned into six predefined networks: the DMN, frontoparietal network, cingulateoper cular netw ork, sensorimotor netw ork (SMN), occipital network (OCC), and cerebellum network (Dosenbach et al., 2010 ).By av er a ging the time courses of all voxels within a node, we extracted the time courses of all nodes and calculated Pearson's correlation coefficients between them.The results were then transformed using Fisher's z -tr ansformation, r esulting in a 160 × 160 correlation matrix for each participant in which each edge represented the rsFC strength between two nodes (Fig. 1 B).

Multisite effect correction
By using the ComBat function available in MATLAB ( https:// github.com/Jfortin1/ ComBatHarmonization ), we r emov ed the site effects to account for site, collection time, and data acquisition par ameter v ariability acr oss eac h of the collection centres in the ABIDE I.A pr e vious study sho w ed that this a ppr oac h can effectiv el y account for scanner-r elated v ariance in m ultisite rfMRI datasets .T he default ComBat function uses a nonparametric empirical Bayes pr ocedur e with a prior and tr eats the dia gnosis, a ge, sex, FIQ, and mFD as biological variables of interest.

Connectome-based predicti v e modelling
As Fig. 1 C shows, all analyses used CPM ( https://www.nitrc.org/pr ojects/bioima gesuite ) to estimate pr edictiv e models based on whole-brain rsFC.We applied the Dosenbach 160 ROI functional template and calculated the R OI-R OI functional connectivity.First, after controlling for confounding variables (i.e .age , sex, and head motion), correlation coefficients between each edge and a vector of behavioural values (here, ADOS scores) were calculated.
A threshold ( P < 0.01) was applied to the matrix to retain edges that were important for the subsequent processing.The positive and negative networks consisted of edges that were significantl y positiv el y (negativ el y) corr elated with ADOS scor es.Next, the sum of edge weights (connectivity strength) in positive and negative netw orks w as calculated for each participant to form single-subject positive and negative network strengths .T he positive and negative network strengths were then included as regression factors to construct linear r egr ession pr edictiv e models for pr edicting ADOS scor es.Finall y, linear r egr ession pr edictiv e models were used on data from new participants to generate predicted values.

Cross-v alida tion
To ensure independence between feature selection and prediction, leav e-one-out cr oss-v alidation was a pplied to the model.In this a ppr oac h, eac h participant's pr edicted v alue (i.e. the 'left-out' subject) was set as the testing set, while all other participants' predicted v alues wer e set as the tr aining set so that all participants had a predicted value according to this iterative manner.Edges that meet the predefined threshold (in this study, P < 0.01) are treated as k e y edges during eac h iter ation.Those edges that ar e consistentl y r etained acr oss all iter ations will be marked in the final network (consensus featur es), whic h is a matrix composed of 0, 1, and −1.Here, 1 r epr esents k e y edges in the positi ve network, and −1 r epr esents k e y edges in the negative network.In addition, we also computed the partial correlation coefficient ( r ) between the model's predicted values and the actual values, taking sex, age , and mFD as co variates .T he significance of r was assessed based on null distributions, which were generated by randomly shuffling the correspondence between behavioural variables and connectivity matrices 1000 times and repeating the CPM analysis with the shuffled data.We used the av er a ge of the resulting 1000 correlation coefficients to represent the average model performance (Fig. 1 D).

Replication of the predicti v e networks across datasets
CPM anal yses wer e used to calculate connectomes fr om ne w ASD datasets to predict ADOS scores.Edge weights corresponding to ASD netw orks w er e then extr acted fr om connectomes (Fig. 1 E).To assess the replicability and specificity of the networks across different datasets, we validated the outcome networks into a inde-pendent clinical sample (172 ASD patients from ABIDE-II dataset) and a healthy sample (36 HCs with ADOS scores).We used partial corr elation anal ysis to examine the r elationship between pr edicted values and actual values (Fig. 1 F).

CPM for predicting ADOS total scores at the o ver all le vel
Among all participants, the CPM r esults demonstr ated that negative network significantly predicted ADOS total scores [ r (df = 150) = 0.19, P = 0.008] (Fig. 2 A1), whereas positive network did not [ r (df = 150) = −0.004,P = 0.540].After establishing the models, we applied their consensus feature to perform another CPM analysis of independent clinical data and HCs for verification.The negative network successfully predicted the ADOS total score in the validation sample [ r (df = 171) = 0.18; P FDR = 0.006] (Fig. 2 A2).T hus , the ability of the identified networks to predict ADOS total scores in a separ ate, heter ogeneous r eplication sample was acceptable.
To verify the specificity of the netw orks, w e applied the models to pr edict ADOS scor es in HCs.P artial corr elation anal yses indicated no significant associations between predicted values and actual values of HCs [ r (df = 35) = 0.10, P FDR = 0.552] (Fig. 2 A3).The results suggest that the identified networks significantly predicted ADOS total scores among participants with ASD.

CPM for predicting ADOS scores in different ASD subtypes
Among participants with CA, the CPM results demonstrated that negative network significantly predicted ADOS total scores [ r (df = 104) = 0.20, P = 0.040] (Fig. 3 A1), ADOS social interaction scores ( r [df = 104] = 0.16, P = 0.030) and ADOS communication scores [ r (df = 104) = 0.21, P = 0.020] (Fig. 3 B1) but not ADOS stereotyped beha viour scores .By contr ast, positiv e network did not reliabl y pr edict an y of these scor es .In the 30 AS participants , CPM did not yield an y effectiv e pr edictiv e networks.Notabl y, we did not calculate the CPM among participants with PDD-NOS because the sample size was too small.The negative networks for predicting ADOS total [ r (df = 171) = 0.17, P FDR = 0.013] (Fig. 3

Netw ork ana tom y and o verlap with canonical neur al netw orks
Finall y, four negativ e networks for predicting ADOS total and comm unication scor es, both for the entir e sample and specificall y for C A patients , were found to be effective after validation in independent samples and HCs.For the convenience of discussions, we will refer to these four network models as T_C A, C_C A, T_C A, and C_CA r espectiv el y.To further mitigate the potential influence of head motion, we excluded participants with mFD > 0.15 mm and obtained similar results .T hese findings are included in the supplementary materials .Across all folds of cr oss-v alidation, 110 negativ e edges wer e r etained as consensus functional connections of T_CA and included connections within and between multiple macroscale brain regions (e.g.frontal, parietal, occipital, temporal, limbic, and cerebellar regions) (Fig. 4 A1).The nodes with the most connections in the negative network included a postoccipital node (Node no.134, 17 • ), a temporal node (Node no.123, 13 • ), an anterior cingulate cortex node (Node no.137, 10 • ), and an inferior parietal lobe node (Node no.117, 9 • ) (Fig. 4 A2).According to the results, the connections between the OCC and SMN and intra-network connections of OCC contributed most to T_CA (Fig. 4 A3).Regarding the C_CA, 123 negative edges wer e r etained as consensus functional connections (Fig. 4 B1).The nodes with the most connections in the negative network are similar to T_CA, including the temporal node (Node no.123, 17 • ), the postoccipital node (Node no.134, 13 • ), and a pr ecentr al gyrus node (Node no.101, 10 • ) (Fig. 4 B2).At the network le v el, the connections between the OCC and SMN and intra-network connections of OCC contributed most to C_CA (Fig. 4 B3).
For participants with CA, the CPM results reveal a more extensiv e ov erla p of neur al networks; 342 negativ e edges wer e r etained as consensus functional connections of T_CA (Fig. 4 C1).The nodes with the most connections in the negative network included a post-parietal node (Node no.117, 26 • ), an occipital node (Node no.123, 24 • ), a mid-insula node (Node no.69, 20 • ), and a post-cingulate node (Node no.24, 20 • ) (Fig. 4 C2).On the basis of the results, OCC and SMN exhibit the most intra-network connections, whereas OCC with SMN and OCC with CON exhibit the most inter-network connections .T he DMN has more connections with networks other than the cerebellum network, especially with the pr e viousl y mentioned OCC, SMN, and CON (Fig. 4 C3).Next, 217 negative edges were retained as consensus functional connections of C_C A (Fig. 4 D1).T he nodes with the most connections in the negative network included a mid-insula node (Node no.69, 27 • and two occipital nodes (Node nos 122 and 123, 20 • ) (Fig. 4 D2).At the network le v el, some intr a-network connections (OCC-SMN, OCC-CON, and SMN-CNN) and intra-network connections of OCC contributed most to C_CA (Fig. 4 D3).

Discussion
In this study, we used a data-driven, connectome-based machine learning protocol to predict ADOS scores.Our research findings indicate that CPM successfully predicted the ADOS total and comm unication scor es for ASD patients, both in all patients and in the C A subtype .Moreo ver, the validation across independent clinical datasets and in HCs provides strong evidence for the replicability of our identified networks .T hese results will deepen our understanding of how functional connectivity coalesces to give rise to the complex autism syndrome and may assist in clinical interventions and treatments (Horien et al., 2022 ).

Critical networks in predicting ADOS total scores
Our study highlighted the importance of connections related to the OCC in predicting ADOS total scores in ASD patients.Numerous studies have demonstrated that patients with ASD exhibit atypical sensory processing, including both hypersensitivity and hyposensitivity (Baron-Cohen et al., 2009 ;Bast, 2007 ;Baum et al., 2015 ).In se v er al sensory domains, individuals often complain of visual symptoms (Simmons et al., 2009 ), and these symptoms increase the burden of depression (Bitsika et al., 2016 ).A pr e vious study r e v ealed that patients with ASD exhibit abnormal activation in the dorsolateral prefrontal cortex, insula, posterior medial pr efr ontal cortex, and occipital cortex compared to typicall y de v eloping individuals (Di Martino et al., 2014 ), suggesting the presence of visual processing deficits in ASD patients .T her efor e, our r esults suggest that both the intranetwork connectivity of the OCC and specific interactions between the OCC and other networks may impair the integration of visual information, which in turn relates to atypical social communication in individuals with ASD.Ho w e v er, this speculativ e conclusion r equir es v erification in a longitudinal sample to establish its validity.
Additionally, the SMN played a crucial role in predicting ADOS total scores in our model.Sensorimotor disturbances are considered a core symptom of ASD in the DSM-5 and persist throughout the lifespan.These disturbances include sensory disturbances, clumsiness, postural instability, and impaired visuomotor coordination (Fournier et al., 2010 ).Mor eov er, motor impairments in gross and fine motor skills, as well as in socially embedded motor skills such as imitation and praxis , ha ve been observed in individuals with ASD (Amonkar et al., 2021 ).These motor dysfunctions have been associated with the clinical symptomology (Sutera et al., 2007 ) and sociobehavioural traits (MacDonald et al., 2013 ) of ASD patients.Se v er al r esearc hers hav e examined the association between motor impairment and the se v erity of ASD core symptoms in the areas of social communication, re petiti ve beha viours , and cognition (Amonkar et al., 2021 ;MacDonald et al., 2013 ).Brian et al. used factor anal ysis and r e v ealed significant correlations between restricted behaviours and hyperresponsive sensory symptoms in c hildr en with ASD (Boyd et al., 2010 ).Penelope et al. also r e v ealed significant corr elations between sensory profiles and Autism Diagnostic Interview-Revised restricted and re petiti ve behaviour (RRB) scores (Hannant et al., 2016 ).We speculate that the patterns of connectivity between the SMN and other networks in the ASD network may impair sensorimotor information integration and are related to atypical social communication in ASD patients .T her efor e, to some extent, the increased connectivity of the SMN may be a neural signature of ASD that can be used for prediction and diagnosis.
Our study r e v ealed numer ous connections both between networks and within networks, pr ominentl y involving the OCC and SMN, which is consistent with previous research (Rohr et al., 2020 ).These findings raise questions about the involvement of the parietal and occipital lobes in the psychopathology of autism.Rohr et al. found that functional connectivity abnormalities in the SMN and visual netw orks w er e r elated to impair ed inhibitory contr ol in c hildr en with ASD (Rohr et al., 2020 ).Heng Chen et al. also found increased connectivity of the SMN and visual networks in ASD, and the insular cortex and occipital cortex were the most connected regions (Chen et al., 2018 ).Applying functional connectivity analyses of intra-and internetwork connectivity, Bosi Chen et al. r eported incr eased network connectivity between regions in the visual network and SMN in the ASD group.This visual-sensorymotor hyperconnectivity is particularly noteworthy in light of the sensory processing abnormalities and multisensory integration impairments in ASD.Ad ditionally, this hyperconnecti vity was correlated with heightened ASD symptomatology (Chen et al., 2021 ).These findings may indicate inadequate integration of visual and somatosensory input into the socioaffective circuits, which in turn affects social interactions.
Ov er all, the OCC, r esponsible for visual pr ocessing, plays a fundamental role in recognizing and interpreting facial expressions, body language, and other nonverbal cues, which are essential for social interaction and understanding emotions (Peelen et al., 2010 ).Similarl y, the SMN, whic h is involv ed in motor planning and execution, is also closely linked to social cognition, as it underlies the ability to imitate and mirror others' actions, fostering social learning and empathy (Hamilton, 2013 ).The heightened connectivity in these networks might indicate an altered neural mechanism for social information processing in ASD, contributing to the difficulties observed in social communication and interaction.Further studies are needed to examine the relationships between activity in these networks and symptom se v erity to support the hypothe-sis that visual-sensory-motor hyperconnectivity may underlie the atypical social interaction and verbal communication observed in ASD.

The predicti v e role of comm unica tion fea tures in ASD diagnosis
When predicting the communication scores for all ASD patients, we obtained a network structure that closely resembled the negative network model for predicting total scores, and this was also the case for C A patients .Furthermore , while social interaction scor es wer e also pr edictible, they could not be v alidated in an independent clinical sample, and this model r etained onl y a fe w k e y edges.Impair ed comm unication is one of the cor e featur es of ASD patients.ASDs coexist with other communication difficulties such as a language disorder, apraxia of speech, speech sound disorders, and/or other neur ode v elopmental disorders .F eatures and clinical markers associated with these coexisting conditions can be used for differ ential dia gnosis and v arious interv entions (Vogindr oukas et al., 2022 ).Our findings may provide neur ological e vidence for that impair ed comm unication might be a k e y featur e in pr edicting the se v erity of ASD.

Differences in predicting ASD subtypes
In terms of subtypes , C A patients were able to successfully predict both total and comm unication scor es and v alidate them in an independent sample, whereas AS patients were unable to derive a pr edictiv e model.Due to the insufficient number of PDD-NOS patients in the sample, this study did not include a discussion on it.It should also be noted that CA is perhaps the broadest and most predominant form of ASD subtypes, and in our study, CA patients constituted the majority of the entire sample (115/157, 73%).
Differing from T_CA and C_CA, in T_CA and C_CA, the role of network connections involved in DMN and CON are emphasized.In ASD, altered functional connectivity involving regions in the DMN has been confirmed to be associated with inhibition (Voorhies et al., 2018 ) and Social Responsiveness Scale scores (Jann et al., 2015 ).The DMN is known to exhibit high activity during the resting state and suppressed activity during cognitiv el y demanding tasks .Furthermore , since the DMN is also associated with a state of readiness for environmental changes, it may be responsible for the poor awareness of and response to social environments in ASD patients .T he CON (also r eferr ed to as the salience or v entr al attention network) is commonl y r eferr ed to as 'cognitiv e contr ol' networks, and is crucial for an indi vidual's executi ve functions (Hausman et al., 2021 ;Sadaghiani & D'esposito, 2015 ).Disruptions in the CON have been linked to se v er al neur opsyc hiatric conditions.Raichle and Snyder discussed the involvement of this network in disorders such as ADHD, emphasizing its role in the regulation of attention (Raichle & Sn yder, 2007 ).K ennedy and Courchesne et al. suggest that alterations in the CON might have potential contributions to the communication and social difficulties observed in individuals with ASD (Kennedy & Courchesne , 2008 ).Furthermore , T_C A and C_C A exhibit a broader overla p of neur al networks.Considering that all models ar e ultimatel y negative networks, meaning that the lower se v erity of the participants, the less close the functional connections involved in these networks .T her efor e, we belie v e that the reduction in brain functional network connectivity may lead to the attention and executive function deficits and motor coor dination issues, as w ell as social and emotional regulation deficits commonly observed in CA patients.

Figure 1 :
Figure 1: Dia gr ams of the anal ysis pr ocess of this study.(A) The fMRI data wer e obtained fr om ABIDE-I and ABIDE-II, and all the data wer e pr epr ocessed using SPM 12 and DPABI.(B) The subplot shows the calculation of task-related ROI-ROI functional connectivity based on the Dosenbach 160 ROI functional template.(C) Schematic of CPM.(D) Permutation testing was used 1000 times to examine the performance of the models.(E) The consensus features of models with significant predictive effects were extracted.(F) To evaluate the replicability and specificity of the models, the netw orks w er e migr ated to ne w independent datasets for the v erification pr ocess.

Figure 2 :
Figure 2: Model performance among all patients.Abbr e viations: T_CA, the negative network model for predicting ADOS total scores of all patients; C_CA, the negative network model for predicting ADOS communication scores of all patients.(A1) and (B1) are the partial correlation scatterplots of the predicted values and the actual values in the model sample, demonstrating that both T_CA and C_CA exhibit great predictive po w er.(A2) and (B2) are the partial correlation scatterplots of the predicted values and the actual values in the validation sample, demonstrating that both T_CA and C_CA ar e r e plicable in inde pendent clinical samples.(A3) and (B3) ar e the partial corr elation scatter plots of the pr edicted v alues and the actual v alues in HCs , pro ving the relative specificity of T_CA and C_CA.

Figure 3 :
Figure 3: Model performance among patients with CA.Abbr e viations: T_CA, the negative network model for predicting ADOS total scores of patients with CA; C_CA, the negative network model for predicting ADOS communication scores of patients with CA. (A1) and (B1) are the partial correlation scatterplots of the predicted values and the actual values in the model sample, demonstrating that both T_CA and C_CA exhibit great predictive po w er.(A2) and (B2) are the partial correlation scatterplots of the predicted values and the actual values in the validation sample, demonstrating that both T_CA and C_CA are replicable in independent clinical samples.(A3) and (B3) are the partial correlation scatterplots of the predicted values and the actual values in HCs , pro ving the relative specificity of T_CA and C_CA.

Figure 4 :
Figure 4: Key nodes and networks of each model.(A), (B), (C), and (D), respectively, illustrate the network anatomy and overlap of T_ALL, C_ALL, T_CA, and C_CA.(1) The circle plot shows the negative network connections among 160 nodes, which were divided into six brain networks according to the Dosenbach 160 ROI functional template and colour coded.(2) The glass brain plot shows the k e y nodes (circles) and edges (lines, r epr esenting functional connections) in the negative network, and the size of the nodes reflects the number of connections related to the node.(3) The matrix plot shows intra-and internetwork connectivity in the negative network.