Phenotypic analysis of 11,125 trio exomes in neurodevelopmental disorders

Genomic sequencing is widely used to identify causative genetic changes in neurodevelopmental disorders, such as autism, intellectual disability, and epilep sy. Most neurodevelopmental disorders also present with diverse clinical features , and delineating the interaction between causative genetic changes and phenotypic features is a key p rerequisite for developing personalized therapies. However, assessing clinical features at a scale that parallels genomic sequencing remains challenging. Here, w e standardize phenotypic information across 11,125 patient-parent trios with exome sequencing data using biomedical ontologies, analyzing 674,767 phenotypic terms. We find that individuals with de novo variants in 69 out of 261 neurodevelopmental genes exhibit statistically significant clinical similarit ies with distinct phenotypic fingerprints . We also observe that phenotypic relatedness follows a gradient, spanning from highly similar to dissimilar phenotypes, with intra-gene similarities suggesting clinically distinct subgroups for seven neurodevelopmental genes. For most genetic etiologies, only a small subset of highly phenotypically similar individuals carr ied de novo variants in the same gene , highlighting the heterogeneous and complex clinical landscape of neurodevelopmental disorders . Our study provides a large-scale overview of the dynamic relationship between genotypes and phenotypes in neurodevelopmental disorders , underscoring how the inherent complexity of these conditions can be deciphered through approaches that integrate genomic and phenotypic data.

neurodevelopmental disorders, exome sequencing, autism, intellectual disability, epilepsy, genotype-phenotype correlation, biomedical ontology, human phenotype ontology. .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint

Neurodevelopmental disorders affect 1-2% of children worldwide and result from a diverse range of etiologies including brain injuries, structural brain malformations, and genetic variants1. Over the last two decades, the identification of causative genetic etiologies has led to a paradigm shift in understanding these conditions. Up to 20% of individuals with autism, intellectual disability, and epilepsy have causative genetic changes identified through novel genomic techniques such as massive ly parallel sequencing 2. Given the overall frequency of monogenic causes among neurodevelopmental disorders, disease-modifying therapies are becoming increasingly available, including antisense oligonucleotide (ASO) therapies and adeno-associated virus ( AAV) based gene therapies3,4. The development of precision medicine approaches and assessment of their efficacy critically depend on a detailed understanding of the clinical features associated with specific genetic causes5. Historically, small case series have provided critical insights into the clinical spectrum and associated features of genetic neurodevelopmental disorders 6,7. However, with over 250 genetic causes of neurodevelopmental disorders identified, the capacity to perform traditional clinical studies has emerged as a major bottleneck8. Given the growing number of neurodevelopmental genes, it has become impossible to generate sufficient phenotypic detail to characterize each disease entity. This presents challenges both with designing outcome measures for clinical trials and providing accurate medical guidance for individuals affected by genetic neurodevelopmental disorders5. Novel methods for large-scale phenotype analysis offer alternative approaches for understanding the clinical spectrum within a wide range of rare disorders, including neurodevelopmental disorders9-13. Previous studies have introduced novel approaches through analyzing phenotypic data from large patient cohorts using clinical data harmonized through biomedical dictionaries9,10. These approaches have been sufficiently powered to explore the range of clinical features within a broad range of conditions and highlight previously unrecognized disease patterns within broader patient populations12,13. .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint Large-scale phenotyping studies combined with genomic data have historically been constrained by the heterogeneous documentation of clinical features. This roadblock has been addressed by emerging tools for standardizing phenotypic terminology, including biomedical dictionaries and ontologies14. One of the most commonly used dictionaries in clinical genomic testing and research is the Human Phenotype Ontology (HPO), a hierarchical dictionary of more than 13,000 clinical terms designed to facilitate computational approaches15,16. This roadblock has been addressed by emerging tools for standardizing phenotypic terminology, including biomedical dictionaries and ontologies 14. One of the most commonly used dictionaries in clinical genomic testing and research is the Human Phenotype Ontology (HPO), a hierarchical dictionary of more than 13,000 clinical terms designed to facilitate computational approaches15,16,10. HPO annotations combined with statistical methods have previously been used to discover novel genetic etiologies in neurodevelopmental disorders 6, map the phenotypic landscape of known genetic etiologies 11, and address the longitudinal trajectories within a subset of genetic neurodevelopmental disorders 9,13. In addition, biomedical dictionaries such as the HPO lend themselves to machine learning approaches that have successfully predicted genetic etiologies based on early clinical features9,12. Here, we analyze 674,767 clinical phenotypic annotations in 11,125 individuals with genetic neurodevelopmental disorders, integrating exome sequencing (ES) data with phenotypic annotations. We find that this approach allows us to delineate the complex interactions between genetic etiologies and phenotypic features , including measures of clinical resemblance across disorders, disease subgroups, and the pattern of causative genetic alterations in groups of individuals that are highly clinically similar. .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint

Harmonization of multiple cohorts using an HPO-based framework To systematically review the interplay between genetic and clinical features, we aggregated clinical and genomic data from four different cohorts : a cohort from a pediatric tertiary care center obtained through diagnostic sequencing (n=192) with a median of 12 clinical features annotated per individual, the Epilepsy Phenome -Genome Project (EPGP; n=335) with a median of nine clinical features17, the Deciphering Developmental Disorders study (DDD; n= 13,424) with a median of six clinical features18, the EuroEPINOMICS-RES cohort (n=319) with a median of 10 clinical features19, and the Children’s Hospital of Philadelphia (CHOP) Birth Defects Biorepository (n=623) with a median of eight HPO terms. Using the framework of the HPO15, we harmonized clinical data across all four cohorts. After data integration, we retrieved a total of 112,733 clinical annotations in 14, 893 individuals, including 1 1,125 individuals with available trio ES data. The clinical annotations consisted of 4,686 unique HPO terminologies, including global developmental delay (HP:0001263 , f=26.1%), delayed speech and language development (HP:0000750, f=15.6%), and seizures (HP:0001250, f=11.9%) as the most common. Among these, 1,952 out of 4,686 annotations (41.7%) were present in two or less individuals, highlighting a sparsity of clinical annotations that has also been observed in prior studies , as well as the presence of clinical features unique to particular genetic neurodevelopmental disorders 13. The amount of phenotypic information per individual was also variable , with the number of annotations per individual ranging from 1 to 64 terms, with a median of six clinical terms per individual. Leveraging the hierarchal structure of HPO to infer higher-level terms, we expanded the dataset to derive a total of 674,767 clinical annotations, which included 5,707 distinct terms with a median of 41 phenotypic features per individual (range of 3-251 clinical features; Fig. 1). The inclusion of more general phenotypic features is referred to as propagation and has previously been shown to aid the harmonization of clinical information by leveraging the tree - like structure of the underlying biomedical ontology9. .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint Genomic analysis identified 261 distinct genetic etiologies with de novo variants After harmonizing phenotypic data across all four cohort s, we sought to identify potentially disease-causing variants among 11,125 available trio exomes, including proband and parental data. We utilized a standardized pipeline across all samples for variant calling20, annotation, and assessment of de novo variants21, the most common gen etic mechanism observed in neurodevelopmental disorders22,23. We applied standard filters to the analysis of genomic data , including heterozygous variants, allele read depth > 10, allele balance between 0.25 and 0.75, and exclusion of population variants present in gnomAD v4.122,23. Applying these criteria to the existing trio ES data, we identified a total of 1,969 de novo variants, including 261 distinct genetic etiologies with de novo variants in two or more individuals (Extended Data Table 1). The most frequent genetic etiologies observed in the combined cohort included KMT2A (n=22), KCNQ2 (n=20), STXBP1 (n=17), MECP2 (n=16), and PPP2R5D (n=16). A total of 18 genetic etiologies had de novo variants detected in >10 individuals , and 130 genetic etiologies were identified in only two individuals. We also identified a total of 81 recurrent de novo variants in 199 individuals among 64 distinct genetic etiologies present in two or more individuals (Extended Data Table 2). The most frequent recurrent de novo variants included PPP2R5D:p.Glu198Lys (n=12), ADNP:p.Tyr719Ter (n=6), HNRNPH2:p.Arg206Trp (n=6), MECP2:p.Arg145Cys (n=5), and AP2M1:p.Arg710Trp (n=4). Using denovolyzeR24, we identified 155 out of 261 genetic etiologies significantly enriched for de novo variants within the joint cohort, providing statistical evidence supporting the involvement of these 155 genetic etiologies in neurodevelopmental disorders. The framework to identify disease -causing variants by comparing observed versus expected frequency of de novo variants is well established 23. We used this framework to assess the significance for each genetic etiolog y based on genomic features only , referred to here as the genomic significance. Analysis of phenotypic features enables assessment of phenotypic similarity We next evaluated phenotypic similarities for each genetic etiology with the aim to contrast significance based on the frequency of de novo variants (genomic significance) with significance .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint based on clinical resemblance for a given genetic etiology (phenotypic significance) . We performed a phenotypic similarity analysis, an established method to assess clinical relatedness, which utilizes the structure of the HPO to compare the intersection of phenotypic features between individuals6,12,15,25. In brief, a phenotypic similarity analysis of a subset of individuals assesses whether clinical features observed in those individuals are more related than expected by chance. We computed similarity score based on the most informative common ancestor concept, assessing pairwise similarity of individuals using the combinatorial comparison of specific phenotypes shared between individuals. Methods to perform phenotypic similarity analyses based on HPO terms have previously been developed and can be perform ed using various algorithms26 (Extended data Fig. 1). We identified 69/261 genes with nominally significant phenotypic similarity, representing genes where phenotypes in individuals with de novo variants were more phenotypically related than expected by chance (Fig. 2a) . SCN1A (n=14), AP2M1 (n=4), and DNM1 (n=6) were the most phenotypically similar genes. For SCN1A, prominent features driving phenotypic similarity were bilateral tonic-clonic seizure s (HP:0002069; freq = 80%; p -val= 8.8x10 -16), focal clonic seizure s (HP:0002266; freq = 53%; p-val= 2.7x10-15), and febrile seizures (HP:0002373; freq = 67%; p-val= 1.9x10-14). In contrast, for AP2M1, the most prominent features were interictal epileptiform activity (HP:0011182; freq = 100%; p-val= 7.35x10-6), atonic seizures (HP:0010819; freq = 75%; p- val= 1.4x10-5), and EEG with spike -wave complexes (HP:0010850; freq = 75%; p-val= 2.4x10-5). For DNM1, myoclonic seizures (HP:0032794; freq = 83%; p-val= 1.06x10-7), EEG with spike-wave complexes < 2.5 Hz (HP:0010847; freq = 66%; p -val= 1.5x10-7), and atypical absence seizures (HP:0007270; freq = 66%; p-val= 1.89x10-7) were the driving features. Genetic etiologies demonstrate variable patterns of phenotypic and genomic significance We next compared phenotypic significance to genomic significance in all 261 genetic etiologies examined (Fig. 2a), comparing the -log10(p-value) for both analyses . We found that a subset of 14 genes, including MECP2, SYNGAP1, and KCNB1, had an excess of de novo variants but low phenotypic similarity . Accordingly, these gene tic etiologies, while relatively common, show .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint highly diverse phenotypes that cannot be clearly distingui shed from the larger cohort. Conversely, a subset of 37 genes had a more substantial phenotypic significance than genomic significance, highlighting genetic etiologies in the cohort that had strong phenotypic resemblance, even if they were rarely observed . These genes included SCN1A, AP2M1, and DNM1. We had previously predicted the existence of genetic etiologies identifiable through strong clinical resemblance rather than a burden of de novo variants13. The current analysis emphasizes that these genetic etiologies are hidden in large, heterogeneous datasets and can be identified through phenotypic similarity approaches. No single genetic etiology stood out with simultaneous strong genomic and phenotypic significance; an observation that seems counterintuitive. It is possible that the datasets examined in our study are depleted for genetic disorders that are both common and recognizable, such as tuberous sclerosis complex and Angelman syndrome, as these disorders may be either subject to population-based screening due to their relative frequency or diagnosed via targeted testing due to their unique clinical manifestations27-29. Phenotypic similarity algorithms converge in large datasets Finally, we examined the impact of various similarity algorithms with respect to their ability to identify genetic etiologies with clinical similarity. We had previously reasoned that various algorithms may differentially emphasize selected features of the HPO tree , such as density of sub-branches. However, in our current cohort, we observe that three different algorithms largely generate comparable results, when assessing the significance and ranking of specific genes (Extended data Fig. 1) . Accordingly, p rior observations suggesting a divergence of results of phenotypic similarity algorithms may have been falsely exaggerated by limited sample sizes. Compared to individuals with synonymous variants, de novo protein-disrupting variants displayed a pattern of higher -than-expected similarity across all algorithms, with several genes consistently displaying the highest phenotypic similarity across multiple algorithms (Fig. 2b-2c). .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint Analysis of phenotypes in recurrent variants reveals “diseases within diseases” While clinical homogeneity can be detected at the gene level, the same method can be extended to recurrent de novo variants. Among the 81 recurrent de novo variants detected in our cohort, 26 showed significant phenotypic similarity (Fig. 3). For three genes (KIF1A, SLC6A1, SRCAP), phenotypic similarity of recurrent variants exceeded gene-level median similarity by two-fold, and for five genes (SCN2A, SMARCA2, KCNQ2, ADNP, TCF7L2), phenotypic similarity of recurrent variants surpassed gene -level similarity by three-fold. The most prom inent phenotypic resemblance was seen for SCN2A:p.Arg853Gln, with EEG with focal sharp waves (HP:0011196), infantile spasms (HP:0012469), and developmental regression (HP:0002376) representing the most prominent phenotypes driving clinical relatedness. For SLC6A1, the p.Phe465Ser variant showed a two-fold more prominent phenotypic similarity driven by seizures (HP:0001250), bilateral tonic-clonic seizures with focal onset (HP:0007334), and growth delay (HP:0001510). For KCNQ2:p.Gly281Arg, we observed a three-fold increase in phenotypic similarity resulting from shared clinical features including encephalopathy (HP:0001298), generalized tonic seizure s (HP:0010818), and EEG with generalized epileptiform discharges (HP:0011198). The most common recurrent variant in the cohort was PPP2R5D:p.Glu198Lys (n=12). Although the gene did not reach statistical significance , this recurrent variant exhibited higher similarity and phenotypic significance driven by macrocephaly (HP:0000256), and hypotonia (HP:0001252). Recurrent variants with increased phenotypic similarity compared to gene-level similarity can be conceptualized as “diseases within diseases” , each with distinct features hinting at distinct molecular mechanism s. For example, SCN2A:p.Arg853Gln is known to demonstrate unique electrophysiological properties affecting the Nav1.2 sodium channel30, highlighted by HPO terms such as EEG with focal sharp waves (HP:0011196) and Infantile spasms (HP:0012469). Patterns of phenotypic similarities hint at hidden subgroups within genes To examine whether genes for neurodevelopmental disorders encompass distinct subgroups based on phenotypic characteristics, we analyzed the distribution of pairwise similarity scores of .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint each genetic etiology . We found that similarity score s varied widely across genes . V isual inspection of these distributions suggested possible subgroups within several genes (Fig. 4a). Cluster analysis to formally define phenotypic subgroups revealed seven genes with distinct phenotypic clusters, including STXBP1, ADNP, GRIN2B, DDX3X, PPP2R5D, ANKRD11, and KCNQ2. For example, among individuals with STXBP1-related disorders (n=17), we identified one subgroup of individuals (n=7) all with seizures (HP:0001250), epileptic spasms (HP:0011097), and abnormal EEG findings (HP:0025373), while the second group (n=10) had muscular hypotonia (HP:0001252) and global developmental delay (HP:0001263) without seizures (Fig. 4b-d). Knowledge of phenotypic subgroups is critical for the development of clinical outcome measures in a trial setting. Our results suggests that a large-scale phenotype analysis can aid in identifying hidden phenotypic subgroups, characterized either by increased phenotypic resemblance in individuals with recurrent variants or phenotypic clusters based on the distribution of pairwise comparisons. A phenotypic neighborhood approach identifies recognizable genes and clinical mimics It is well established that individuals with a common genetic etiology may have highly similar features. Conversely, some groups of individuals display highly similar clinical features but have distinct or even unknown disease etiologies . We hypothesized that these “phenocopies”, or individuals with similar phenotypes lacking a common genetic etiology , could be detected by quantitatively identifying individuals who are clinically related to those with a particular genetic etiology. We first analyzed the phenotypic neighborhood associated with individual genes. A phenotypic neighborhood consisted of individuals who shared phenotypic features and had high pairwise similarity scores. These individuals are connected by edges when a given individual is among the k most phenotypically similar to another. Initially, we assessed the 10-neighborhoods (i.e. k=10) of SCN1A, the gene which exhibited the highest overall phenotypic similarity in our dataset. This .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint neighborhood included 61 individuals, including the 14 with known de novo variants in SCN1A. There were 29 individuals who had no identifiable de novo variants, among the 18 other individuals the most common non -SCN1A genes among this neighborhood were DNM1 (n=2), and GABRB3 (n=2). The identification of de novo variants in these genes in individuals who have highly similar clinical presentations to individuals with SCN1A variants suggests that some individuals with these genetic diagnoses may represent potential phenocopies of SCN1A-related disorders in the clinical setting, also highlighting the importance of a broader approach to genetic diagnosis. Next, we explored which shared clinical features were the primary drivers of clinical relatedness within SCN1A and its broader phenotypic neighborhood (k=1000; Fig. 5). We verified that individual phenotypes associated with SCN1A remained statistically significant when compared to the larger neighborhood of individuals, often demonstrating even stronger associations within the 10 -neighborhood and 100 -neighborhood. Clinical features such as bilateral tonic -clonic seizures (HP:0002069; 61%), generalized myoclonic seizures (HP:0002123; 52%), and generalized clonic seizures (HP:0010818; 46%) were prevalent in this neighborhood (Fig. 6a), emphasizing that the other genes in the neighborhood of SCN1A causing generalized epilepsy with multiple seizure types can be phenotypic mimickers of SCN1A-related disorders, an observation that is often noted in the clinical care of individuals with genetic generalized epilepsies31. Moreover, this pattern of stronger associations s uggests that including individuals with high phenotypic similarities to those with SCN1A increases the potential to identify key phenotypes of SCN1A- related disorders. This finding highlights the role of phenotypic neighborhoods in amplifying the essential features of a genetic disorder. We further hypothesized that computational detection of phenocopies through phenotypic neighborhoods could provide insight into the clinical recognizability of a genetic etiology. This was accomplished by capturing the estimated likelihood that two clinically similar individuals share the same genetic etiology , a measure we refer to as the “neighborhood index”. The neighborhood index represents the proportion of individuals within a subgroup who have at least .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint one “genetic twin” , i.e., another individual with the same genetic etiology within the k- neighborhood. As expected, the neighborhood index increase s with increasing values of k, reflecting the fact that including more individuals increases the likelihood of identifying someone with a de novo variant in the same gene. However, genes displayed distinct patterns of neighborhood indices. Notably, we found that SCN1A had a stable, high neighborhood index ranging from 78% to 100% across different k values (k=10,100,1000), whereas SYNGAP1 had a low neighborhood index of 10% at k=50, which increased continuosly to 72% at k=1000 (Fig. 6b). These results suggest that individuals with clinical presentations similar to those with SCN1A- related disorders are highly likely to possess a de novo variant in SCN1A, whereas the confidence for identifying an individual clinically similar to one with a de novo variant in SYNGAP1 is lower. These findings are consistent with the established, consistent clinical presentation of individuals with SCN1A-related disorders, in contrast to the phenotypic heterogeneity among those with SYNGAP1-related disorders. .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint

Both the genetic and phenotypic landscapes of neurodevelopmental disorders are highly complex, comprising hundreds of genetic etiologies and thousands of potential clinical features. Large-scale genomic sequencing has been the key technology in the discovery of genes associated with neurodevelopmental disorders. However, analysis of phenotypic data at a similar magnitude has not yet been performed. With precision therapies under development for a wide range of genetic conditions, understanding the link between genetic causes and linked clinical features is increasingly relevant. Assessing clinical data at the same level as genomic information derived from massive parallel sequencing studies is challenging . In our dataset of 11,125 individuals with genomic and phenotypic data, we used the Human Phenotype Ontology (HPO) to systematically map clinical information onto a joint framework. The HPO represents a useful scaffold for this analysis for several reasons. First, HPO-based coding is already used in a wide range of studies and datasets. For example, the DDD cohort, which represents the largest cohort within our study, has been annotated phenotypically using HPO terms at the point of data collection 18. Second, the HPO represents a simple yet comprehensive dictionary with a defined relationship between categories. Even though the HPO comprises >13,000 phenotypic categories, the relationship between phenotypes is limited to hierarchical relationships, enabling straightforward inference of broader phenotypic categories. In our cohort , phenotypic inference or “propagation” increased the median number of clinical terms per individual from 6 to 41 terms, a more than six- fold increase of clinical information . Finally, methods to assess phenotypic similarity have been extensively tested using HPO-coded phenotypes6,10,12,25. The phenotypic depth in our dataset comprised 5,707 categories, with almost half of all clinical categories only present in two individuals or less. This observation reflects a common feature of clinical information in neurodevelopmental disorders : phenotypic data is sparse and highly dimensional10. However, we find that the size of the dataset compensates for its sparsity. For example, we were able to validate known genotype-phenotype associations, such the association .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint of de novo variants in KMT2A with short stature (HP: 0004322; pval = 6.07 x 10 -4), growth delay (HP:0001510; pval = 3.12 x 10 -4), and hypertrichosis (HP:0000998; pval = 1.19 x 10 -6), reflecting the well-established clinical spectrum of Wiedemann-Steiner syndrome32. Similarly, in individuals with de novo variants in KCNQ2, we found strong associations with g eneralized-onset seizures (HP:0002197; pval= 5.72 x 10 -6), hypotonia (HP:0001252; pval= 1.52x 10-2), and EEG with burst suppression (HP:0010851; pval = 2.15 x 10 -7), known clinical features linked to KCNQ2-related disorders33. In order to assess the phenotypic gestalt in neurodevelopmental disorders more holistically, we focused on phenotypic similarity , a measure that determines clinical relatedness based on all phenotypic features of an individual. We identified 69 phenotypically significant genes, including established neurodevelopmental genes including SCN1A, SCN2A, SCN8A, AP2M1, DNM1, GABRB3, and STX1B19,34. We did not observe higher than expected phenotypic similarity for synonymous de novo variants, emphasizing that phenotypic similarity algorithms are robust and do not artificially inflate statistical significance. This finding suggests that a large number of neurodevelopmental genes have strong phenotypic fingerprints, which may be critical to identify disease-specific outcome measures in clinical trials. We found that 15 phenotypically similar genes did not achieve genomic significance. These genes represent 5.7% of the 261 genes with de novo variants in two or more individuals and account for 33 individuals, 3.3% of the 992 individuals impacted by these 261 genetic etiologies. This increase in individuals with explained genetic causes emphasizes the utility of phenotypic similarity algorithms as a complementary means to identify neurodevelopmental genes. These

may be particularly relevant given that several hundred neurodevelopmental genes are yet to be identified23, and approaches based on genomic information alone are expected to have diminishing returns in large and heterogeneous datasets22. We observed two unexpected findings when we examined subgroups defined by genetic or phenotypic features. First, we found that the phenotypic landscape of recurrent de novo variants .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint is inherently complex. For almost half of all recurrent de novo variants (35/81, 43%), phenotypic similarity exceeds gene-level similarity. Moreover, recurrent variants SMARCA2, PPP2R5D, and TCF7L2 are phenotypically similar despite lack of clinical resemblance when assessing all de novo variants in these genes. This finding suggests that recurrent de novo variants may have unique biological consequences that frequently differ from the disease mechanism of the wider genetic etiology. Second, when analyzing phenotypic subgroups based on pair-wise clinical resemblance, we observed phenotypic clusters in seven neurodevelopmental genes, including several genes previously thought to result in homogeneous phenotypes . Defining meaningful subgroups is a key prerequisite for future clinical trials, and our results suggest that large clinical and genomic datasets may provide an initial insight for the potential existence of novel subgroups in well- known neurodevelopmental genes. Many individuals with genetic neurodevelopmental disorders present similarly to individuals with different or undetectable genetic diagnoses. To explore how neurodevelopmental genes cluster within groups of individuals with related phenotypes , we introduce d the concept of the neighborhood index as a novel concept. This index describes the likelihood that clinically similar individuals share the same neurodevelopmental gene . Accordingly, a high neighborhood index suggests a strong ability to recognize a genetic neurodevelopmental disorder in a clinical setting. We find that neighborhood indices vary widely across genes with strong phenotypic similarity, suggesting that only a small subset of neurodevelopmental gene s including AP2M1, SCN1A, STXBP1, DNM1, and GABRB3 share clinical features that allow these conditions to stand out from the wide phenotypic range seen in individuals with neurodevelopmental disorders. In summary, by jointly assessing genomic and clinical features in neurodevelopmental disorders in 11,125 individuals with genomic and phenotypic data , we discover an unexpectedly complex disease landscape. As genomic and phenotypic data continue s to grow, exploring the interplay of clinical and genomics features may provide novel insights into phenotypic fingerprints linked to neurodevelopmental genes and patterns within subgroups defined by genetic or clinical features. .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint

1 EuroEPINOMICS-RES Consortium; Epilepsy Phenome/Genome Project; Epi4K Consortium. De novo mutations in synaptic transmission genes including DNM1 cause epileptic encephalopathies. Am J Hum Genet 95, 360-370 (2014). https://doi.org/10.1016/j.ajhg.2014.08.013 2 Stefanski, A. et al. Clinical sequencing yield in epilepsy, autism spectrum disorder, and intellectual disability: A systematic review and meta-analysis. Epilepsia 62, 143-151 (2021). https://doi.org/10.1111/epi.16755 3 Lerche, H., Hedrich, U. B. & Wuttke, T. V. Gene-replacement therapy in neurodevelopmental disorders: progress and challenges. J Clin Invest 135 (2025). https://doi.org/10.1172/jci188703 4 Hill, S. F. & Meisler, M. H. Antisense Oligonucleotide Therapy for Neurodevelopmental Disorders. Dev Neurosci 43, 247-252 (2021). https://doi.org/10.1159/000517686 5 Sullivan, J. et al. Adaptive functioning and neurodevelopment in patients with Dravet syndrome: 12-month interim analysis of the BUTTERFLY observational study. Epilepsy Behav 151, 109604 (2024). https://doi.org/10.1016/j.yebeh.2023.109604 6 Helbig, I. et al. A Recurrent Missense Variant in AP2M1 Impairs Clathrin-Mediated Endocytosis and Causes Developmental and Epileptic Encephalopathy. Am J Hum Genet 104, 1060-1072 (2019). https://doi.org/10.1016/j.ajhg.2019.04.001 7 Parthasarathy, S. et al. A recurrent de novo splice site variant involving DNM1 exon 10a causes developmental and epileptic encephalopathy through a dominant-negative mechanism. Am J Hum Genet 109, 2253-2269 (2022). https://doi.org/10.1016/j.ajhg.2022.11.002 8 Ruggiero, S. M., Xian, J. & Helbig, I. The current landscape of epilepsy genetics: where are we, and where are we going? Curr Opin Neurol 36, 86-94 (2023). https://doi.org/10.1097/wco.0000000000001141 9 Ganesan, S. et al. A longitudinal footprint of genetic epilepsies using automated electronic medical record interpretation. Genet Med 22, 2060-2070 (2020). https://doi.org/10.1038/s41436-020-0923-1 10 Lewis-Smith, D. et al. Computational analysis of neurodevelopmental phenotypes: Harmonization empowers clinical discovery. Hum Mutat 43, 1642-1658 (2022). https://doi.org/10.1002/humu.24389 11 Xian, J. et al. Assessing the landscape of STXBP1-related disorders in 534 individuals. Brain 145, 1668-1683 (2022). https://doi.org/10.1093/brain/awab327 12 Galer, P. D. et al. Semantic Similarity Analysis Reveals Robust Gene-Disease Relationships in Developmental and Epileptic Encephalopathies. Am J Hum Genet 107, 683-697 (2020). https://doi.org/10.1016/j.ajhg.2020.08.003 13 Galer, P. D. et al. Clinical signatures of genetic epilepsies precede diagnosis in electronic medical records of 32,000 individuals. Genet Med, 101211 (2024). https://doi.org/10.1016/j.gim.2024.101211 14 Haendel, M. A., Chute, C. G. & Robinson, P. N. Classification, Ontology, and Precision Medicine. N Engl J Med 379, 1452-1462 (2018). https://doi.org/10.1056/NEJMra1615014 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint 15 Kohler, S. et al. The Human Phenotype Ontology in 2021. Nucleic Acids Res 49, D1207- D1217 (2021). https://doi.org/10.1093/nar/gkaa1043 16 Lewis-Smith, D. et al. Modeling seizures in the Human Phenotype Ontology according to contemporary ILAE concepts makes big phenotypic data tractable. Epilepsia 62, 1293- 1305 (2021). https://doi.org/10.1111/epi.16908 17 Abou-Khalil, B. et al. The epilepsy phenome/genome project. Clin Trials 10, 568-586 (2013). https://doi.org/10.1177/1740774513484392 18 Firth, H. V. & Wright, C. F. The Deciphering Developmental Disorders (DDD) study. Dev Med Child Neurol 53, 702-703 (2011). https://doi.org/10.1111/j.1469- 8749.2011.04032.x 19 Heyne, H. O. et al. De novo variants in neurodevelopmental disorders with epilepsy. Nat Genet 50, 1048-1053 (2018). https://doi.org/10.1038/s41588-018-0143-7 20 Helbig, I. et al. Whole-exome and HLA sequencing in Febrile infection-related epilepsy syndrome. Ann Clin Transl Neurol 7, 1429-1435 (2020). https://doi.org/10.1002/acn3.51062 21 Yang, H. & Wang, K. Genomic variant annotation and prioritization with ANNOVAR and wANNOVAR. Nat Protoc 10, 1556-1566 (2015). https://doi.org/10.1038/nprot.2015.105 22 Deciphering Developmental Disorders Study. Prevalence and architecture of de novo mutations in developmental disorders. Nature 542, 433-438 (2017). https://doi.org/10.1038/nature21062 23 Kaplanis, J. et al. Evidence for 28 genetic disorders discovered by combining healthcare and research data. Nature 586, 757-762 (2020). https://doi.org/10.1038/s41586-020- 2832-5 24 Ware, J. S., Samocha, K. E., Homsy, J. & Daly, M. J. Interpreting de novo Variation in Human Disease Using denovolyzeR. Curr Protoc Hum Genet 87, 7 25 21-27 25 15 (2015). https://doi.org/10.1002/0471142905.hg0725s87 25 Akawi, N. et al. Discovery of four recessive developmental disorders using probabilistic genotype and phenotype matching among 4,125 families. Nat Genet 47, 1363-1369 (2015). https://doi.org/10.1038/ng.3410 26 Pesquita, C., Faria, D., Falcão, A. O., Lord, P. & Couto, F. M. Semantic similarity in biomedical ontologies. PLoS Comput Biol 5, e1000443 (2009). https://doi.org/10.1371/journal.pcbi.1000443 27 Dungan, J. S. et al. Noninvasive prenatal screening (NIPS) for fetal chromosome abnormalities in a general-risk population: An evidence-based clinical guideline of the American College of Medical Genetics and Genomics (ACMG). Genet Med 25, 100336 (2023). https://doi.org/10.1016/j.gim.2022.11.004 28 Northrup, H., Wheless, J. W., Bertin, T. K. & Lewis, R. A. Variability of expression in tuberous sclerosis. J Med Genet 30, 41-43 (1993). https://doi.org/10.1136/jmg.30.1.41 29 Williams, C. A. et al. Angelman syndrome 2005: updated consensus for diagnostic criteria. Am J Med Genet A 140, 413-418 (2006). https://doi.org/10.1002/ajmg.a.31074 30 Mason, E. R. et al. Resurgent and Gating Pore Currents Induced by De Novo SCN2A Epilepsy Mutations. eNeuro 6 (2019). https://doi.org/10.1523/eneuro.0141-19.2019 31 Mullen, S. A. & Berkovic, S. F. Genetic generalized epilepsies. Epilepsia 59, 1148-1153 (2018). https://doi.org/10.1111/epi.14042 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint 32 Castiglioni, S. et al. KMT2A: Umbrella Gene for Multiple Diseases. Genes (Basel) 13 (2022). https://doi.org/10.3390/genes13030514 33 Millichap, J. J. et al. KCNQ2 encephalopathy: Features, mutational hot spots, and ezogabine treatment of 11 patients. Neurol Genet 2, e96 (2016). https://doi.org/10.1212/nxg.0000000000000096 34 Lindy, A. S. et al. Diagnostic outcomes for genetic testing of 70 genes in 8565 patients with epilepsy and neurodevelopmental disorders. Epilepsia 59, 1062-1071 (2018). https://doi.org/10.1111/epi.14074 35 Li, H. & Durbin, R. Fast and accurate long-read alignment with Burrows-Wheeler transform. Bioinformatics 26, 589-595 (2010). https://doi.org/10.1093/bioinformatics/btp698 36 McKenna, A. et al. The Genome Analysis Toolkit: a MapReduce framework for analyzing next-generation DNA sequencing data. Genome Res 20, 1297-1303 (2010). https://doi.org/10.1101/gr.107524.110 37 McLaren, W. et al. The Ensembl Variant Effect Predictor. Genome Biol 17, 122 (2016). https://doi.org/10.1186/s13059-016-0974-4 38 Chen, S. et al. A genomic mutational constraint map using variation in 76,156 human genomes. Nature 625, 92-100 (2024). https://doi.org/10.1038/s41586-023-06045-0 39 Rentzsch, P., Witten, D., Cooper, G. M., Shendure, J. & Kircher, M. CADD: predicting the deleteriousness of variants throughout the human genome. Nucleic Acids Res 47, D886- d894 (2019). https://doi.org/10.1093/nar/gky1016 40 OMIM (2025) Online Mendelian Inheritance in Man, OMIM®. Baltimore, MD: McKusick- Nathans Institute of Genetic Medicine, John Hopkins University. 41 Resnik P. Using information content to evaluate semantic similarity in a taxonomy. Proceedings of the 14th international joint conference on Artificial intelligence, Morgan Kaufmann Publishers Inc., San Francisco, CA, USA. 1995; pp. 448–453. .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint Fig. 1: Harmonization of clinical features using HPO. a, Clinical features for 14,893 individuals mapped to 4,686 distinct HPO terms, yielding 5,707 inferred HPO clinical features after harmonization. b, Seizure subontology with 55 reported and 81 inferred terms across 3,324 individuals , with the most frequent clinical annotation terms highlighted. c, The Abnormality of nervous system HPO branch, highlighting the importance of harmonization. The process of harmonization using HPO lead to nearly six-fold increase in the total data points, increasing the median number of terms per individual from 6 to 41. .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint Fig. 2: Comparison of statistical significance of de novo variant occurrence and phenotypic similarity. a, Comparison of statistical significance of 261 genes with de novo variants, integrating both genetic and phenotypic evidence. The x-axis represents the likelihood of the occurrence of de novo variants in the cohort computed using denovolyzeR, while y-axis represents the phenotypic similarity significance calculated through permutation analysis. We identified 69 genes significant based on phenotypic similarity, among which 15 genetic etiologies , including SCN2A and KCNA2, where phenotypic evidence exceeded genetic evidence. b, qqplot comparing the expected and observed phenotypic similarity for genetic etiologies with protein-altering de novo variants. Several genes, including SCN1A, AP2M1, and DNM1 show higher phenotypic significance than expected. c, To assess algorithm performance, we computed phenotypic similarity for genes with synonymous de novo variants. The observed similarity remained low, validating the robustness of our method and are unlikely due to random chance and serve as a negative control. .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint Fig. 3: Phenotypic similarity of recurrent variants compared to genetic etiologies . Recurrent variants associated with specific genetic etiologies indicated higher median phenotypic similarity scores compared to the overall median similarity score of the respective genetic etiology. The dot size represents the number of individuals with the variant, blue dots indicate variants that reached statistical significance based on phenotypic evidence. .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint Fig. 4: Phenotypic similarity patterns reveal subgroups within genetic etiologies. a, Pairwise phenotypic similarity scores for each genetic etiology suggest the presence of distinct subgroups w ithin several genetic etiologies (indicated by the box). b, Focusing on the STXBP1 gene, which exhibited a wide range of pairwise similarity scores, we performed clustering analysis that revealed two distinct subgroups. The heatmap represents hierarchical clustering, with the colors indicating phenotypic similarity patterns based on a weighted similarity score. The purple tile indicates higher phenotypic similarity, while yellow tile s indicate a lower similarity score. c, A r adar plot highlighting hypotonia, ataxia, and tremor as more frequent in one group. d, Seizures, spasms, and abnormal EEG were more frequent in the other group. .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint Fig. 5: Network representation of phenotypic similarity in genetic neurodevelopmental disorders . a, A network visualization of i ndividuals with de novo variants in known e pilepsy genes and the top 1 000 most phenotypically similar individuals. The distance between each individual is based on the phenotypic similarity score, computed using Fruchterman-Reingold algorithm in Gephi. Each node represents an individual, with colors and size of the dots indicating specific genetic etiologies. b, Focusing on the denser cluster reveals that individuals with de novo variants in SCN1A gene exhibit high phenotypic similarity to each other, suggesting a strong shared clinical presentation. .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint Fig. 6: Phenotypic association in the SCN1A neighborhood. a, Volcano plot highlighting the HPO terms in individuals within the SCN1A neighborhood (k=100). The x-axis represents the odds ratio in log scale, indicating the strength of the association, while y -axis represents the statistical signi ficance. Red dots indicate phenotypic terms positively associated with SCN1A such as Bilateral tonic-clonic seizure, Status epilepticus, and Febrile seizures, while blue dots indicate negative association with the phenotypes. b, Distribution of twin index for individuals within k=100 neighborhood, across all genetic etiologies. .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint

Genetic analysis All exome sequencing data from EPGP, EuriEPINOMICS-RES, and our local CHOP trio cohort were returned in FASTQ format. All the data were re -analyzed using a standardized pipeline. Raw sequencing reads were mapped to the human reference genome HS37d5 using Burrows-Wheeler Alignment (BWA) MEM algorithm to produce an aligned BAM file 35. Duplicate reads were removed using Picard, and base quality scores were recalibrated using GenomeAnalysisToolKit (GATK) software36 to account for the systematic bias by the sequencer. Variant calling (SNP and Indel) was performed using HaplotypeCaller and genotyped using GenotypeGVCFs in GATK. For the DDD dataset, we got trio -variant call files also aligned to the HS37d5 human refere nce genome. To make the dataset harmonized, all the variants from these cohorts were liftovered to human genome build 38. The whole genome sequencing dataset we got access through the Birth Defects Biorepository, data processing were performed by the genomics platform at the Broad Institute of MIT and Harvard. DNA libraries were prepared using the Illumina Nextera or Twist exome capture (~38 Mb target) and sequenced with 150 bp paired -end reads, achieving >85% of targets covered at 20x and a mean target coverage of >55x. Sequencing data were processed through a pipeline based on Picard, with read mapping performed using the BW A aligner to the human genome build 38 (GRCh38). Variants were called using the Genome Analysis Toolkit (GATK) HaplotypeCaller package version 3.5, following best practices for variant detection. Variant annotation for all the datasets were done using ANNOVAR and VEP for expected functional sequence21,37. Other annotations also included allele frequencies from gnomAD and CADD (Combined Annotation Dependent Depletion) scores 38,39. Individual genes were further annotated with Residual Variant Intolerance Score (RVIS) percentile, and pLI or probability of loss- of-function intolerance 40. Variants were categorized as de novo , homozygous, or compound heterozygous if the following filtering criteria was fulfilled: genotyping quality in all samples were greater than 20, total read depth greater was than 10, the variant was not present in the gnomAD .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint database, RVIS percentile was less than 55, and the gene displayed dominant inheritance in Online Mendelian Inheritance in Man (OMIM)40. Additionally, we computed the p -value for all genes with de novo variants using denovolyzeR which compares the expected number of de novo variants against the observed number of de novo variants per gene24. Similarity analysis The Human Phenotype Ontology (HPO) provides a standardized biomedical representation of >15,000 phenotypes in human disease15. Every term in HPO represents a distinct clinical feature. All the individuals in our cohort were annotated with HPO terms capturing their clinical presentation. For the final analysis, all HPO terms were propagated using automated harmonization, as previously described, after removing obsolete and duplicate terms based on HPO release v2023 -08-02. The frequency (f) of each HPO term was computed, allowing us to calculate Information Content (IC), defined as -log2(f). We calculated phenotypic similarity between individuals using three different algorithms. First, Resnik similarity uses information content (IC) of a common parent term, also known as most informative common ancestor (MICA), for two individual HPO terms41. Pairwise similarity for two individuals is computed using the following, where P1 and P2 are two individuals in the cohort having m and n annotations respectively, and Sij is the information content of the MICA of HPO terms i and j: 𝑠𝑖𝑚(𝑃!,𝑃") = 1 2,- 𝑚𝑎𝑥!#$#%𝑆&$ ' &( ! + - 𝑚𝑎𝑥!#&#'𝑆$& % $( ! 2 In the next method, we calculated the similarity of two individuals by sum of the IC of the shared or intersection of HPO terms after propagation. We refer to this method as the ‘cube’ algorithm, also known as total overlap: 𝑠𝑖𝑚(𝑃!,𝑃") = - 𝐼𝐶) 5𝐻𝑃𝑂*!)∩*") 8 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint The last similarity method entailed computing the intersection or total overlap of two individuals after filtering out redundant or implied terms from each individual using the function minimal_set from package ontologyIndex in R: 𝑠𝑖𝑚(𝑃!,𝑃") = - 𝐼𝐶) 5𝐻𝑃𝑂'&%&',-_/0)(*!)∩*"))8 For the primary analysis we used the Resnik -mod (Equation 1) algorithm as it is the most well - established method. To test the significance of the phenotypic similarity within a particular gene with n individuals, we computed the distribution of median similarities of n individuals chosen randomly across one million permutations. We then compared this distribution to the median similarity of the respective gene with n individuals. Comparing the expected and observed median similarity resulted in an exact p-value. Clustering analysis We used agglomerative clustering with Ward’s method within the cluster package in R to cluster individuals and clinical features within individual gene subgroups. For HPO features, we weighted terms by their information content prior to clustering. For individuals, we used the inverse of their Resnik phenotypic similarity as a measur e of distance. We determined the number of meaningful clusters by minimizing the ratio of mean height within the cluster to height between clusters for all values between two and ten. Neighborhood analysis We defined an individual’s phenotypic k-neighborhood as the individual along with the k most phenotypically similar individuals in our cohort. The k-neighborhood of an entire genetic etiology was then defined as the union of k-neighborhoods of all individuals within the gene subgroup. For example, the 10-neighborhood of a gene consisted of all individuals with a de novo variant in that gene, along with the 10 most similar individuals to each within the larger cohort. .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint Next, for each genetic cause, we measured the “neighborhood index.” We computed the k- neighborhood for each individual within a given gene genetic etiology and determined how many individuals had at least one k-neighbor with the same genetic cause. We computed neighborhood indices for each genetic etiology, for k values in increments of 10 up to 100, and k values of 500, and 1000 and compared twin indices across all genetic etiologies in our cohort. .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint Extended Data Table 1. Genes with most identified de novo variants across cohorts Gene Ensembl gene ID n Recurrent variant OMIM # KMT2A ENSG00000118058 22 ENSP00000436786.2:p.Asn3383SerfsTer2 (n=2) 159555 KCNQ2 ENSG00000075043 20 ENSP00000352035.2:p.Gly281Arg (n=3) 602235 STXBP1 ENSG00000136854 17 ENSP00000362396.2:p.Arg367Ter (n=2) 602926 MECP2 ENSG00000169057 16 ENSP00000395535.2:p.Arg145Cys (n=5) 300005 PPP2R5D ENSG00000112640 16 ENSP00000417963.1:p.Glu198Lys (n=12) 601646 ADNP ENSG00000101126 15 ENSP00000483881.1:p.Asn832LysfsTer81 (n=2) 611386 CTNNB1 ENSG00000168036 14 ENSP00000344456.5:p.Arg535Ter (n=3) 116806 SCN1A ENSG00000144285 14 - 182389 SCN2A ENSG00000136531 13 ENSP00000364586.2:p.Arg853Gln (n=2) 182390 DDX3X ENSG00000215301 13 ENSP00000494040.1:p.Arg292Ter (n=3) 300160 MED13L ENSG00000123066 12 - 608771 ARID1B ENSG00000049618 11 ENSP00000490491.2:p.Glu1879AlafsTer9 (n=2) 614556 DYRK1A ENSG00000157540 11 ENSP00000494572.1:p.Arg246Ter (n=2) 600855 GRIN2B ENSG00000273079 11 ENSP00000477455.1:p.Gly689Ser (n=3) 138252 SYNGAP1 ENSG00000197283 11 - 603384 ANKRD11 ENSG00000167522 10 - 611192 KAT6A ENSG00000083168 10 - 601408 SATB2 ENSG00000119042 10 ENSP00000401112.1:p.Arg399His (n=2) 608148 ASXL3 ENSG00000141431 9 ENSP00000269197.4:p.Arg1444Ter (n=3) 615115 KCNB1 ENSG00000158445 9 - 600397 CDKL5 ENSG00000008086 8 - 300203 GATAD2B ENSG00000143614 8 ENSP00000357644.4:p.Arg116Ter (n=3) 614998 KAT6B ENSG00000156650 8 - 605880 IQSEC2 ENSG00000124313 7 - 300522 KIF1A ENSG00000130294 7 ENSP00000438388.1:p.Arg307Pro (n=2) 601255 NEXMIF ENSG00000050030 7 - 300524 SCN8A ENSG00000196876 7 - 600702 SETD5 ENSG00000168137 7 ENSP00000385852.2:p.Arg768Ter (n=3) 615743 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint Extended Data Table 2. Comparison of phenotypic similarity among recurrent de novo variants and associated genetic etiologies Gene No. of individuals median similarity of Gene Recurrent variant No. of individuals recurrent median similarity of variant distinct phenotypic features PPP2R5D 16 6.6 p.Glu198Lys 12 8.5 Generalized Hypotonia, Macrocephaly KCNQ2 20 7.8 p.Gly281Arg 3 30.1 Encephalopathy, Generalized tonic seizure, EEG with generalized epileptiform discharges GATAD2B 8 9.5 p.Arg116Ter 3 16.1 Neonatal hypotonia, Hypoplasia of the corpus callosum SRCAP 6 9.6 p.Arg2435Ter 3 26.9 Growth abnormality, short stature, Frontal bossing KCNQ3 5 12.1 p.Arg230Cys 3 13.8 Atonic seizure, EEG with focal sharp waves, Facial asymmetry ADNP 15 7.5 p.Asn832LysfsTer81 2 25.1 Microcephaly, axial hypotonia, coloboma SCN2A 13 16.9 p.Arg853Gln 2 77.7 Febrile seizure (within age range of 3 months to 6 years), atypical absence seizure, EEG with spike- wave complexes (<2.5 Hz) DNM1 6 38.6 p.Ala177Pro 2 26.3 Generalized myoclonic seizure, Interictal epileptiform activity KIF1A 7 15.2 p.Arg307Pro 2 35.6 Intellectual disability, spasticity, Hypertonia KIF1A 7 15.2 p.Arg307Gln 2 30.2 Neonatal hypotonia, Severe global developmental delay KIF1A 7 15.2 p.Pro305Leu 2 22.8 Seizure, moderate global developmental delay SLC6A1 6 11.9 p.Phe465Ser 2 34.3 Growth delay, Bilateral tonic-clonic seizure with focal onset CREBBP 5 17.1 c.3779+1G>A 2 26.3 Growth delay, abnormality of cardiovascular system TCF7L2 4 5.1 c.686-1G>A 2 17.3 Febrile seizure (within age range of 3 months to 6 years), Talipes SMARCA2 4 6.5 p.Arg525His 2 32.1 Wide mouth, prominent nasal bridge .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint Extended Data Fig. 1: Phenotypic similarity across different algorithms. a-c, QQplots showing us the most significant gene based on phenotypic similarity analysis using HPO terms with three different algorithms. d, The observed phenotypic similarity of synonymous de novo variants remained low, validating the findings in our methods. .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint

The authors thank the participants and their family members for taking part in this study. We would like to thank or clinical research coordinators for their extensive support in enrolling research participants and admisitrative assistance. The DDD study presents independent research commissioned by the Health Innovation Challenge Fund [grant number HICF-1009-003]. This study makes use of DECIPHER (http://www.deciphergenomics.org), which is funded by Wellcome [grant number WT223718/Z/21/Z]. S ee Nature PMID: 25533962 or www.ddduk.org/access.html for full acknowledgement. The CHOP Birth Defects Biorepository is supported by the National Center for Advancing Translational Sciences, National Institutes of Health, through Grant UL1TR001878. Author contributions I.H. and S.G. conceived the project. S.G., P.D.G, and S.M.R. were responsible for data curation. S.G, P.D.G, and S.P. developed the methods for the analysis . S.G, P.D.G, S.P. , and S.R.M. conducted the formal analysis. S.G., S.P.,S.R.C., and I.H. drafted the manuscript. All authors reviewed and edited the manuscript. Competing interests The authors declare no competing interests. Funding I.H is supported by the National Institute for Neurological Disorders and Stroke (R01 NS131512, R01 NS127830) and the Hartwell Foundation (Individual Biomedical research Award). I.H. received support through the German Research Foundation (DFG/FNR INTER Research Unit FOR2715 (He5415/7-1 and He5415/7 -2). Y.W received support through the German Research Foundation (DFG WE4896/4-2) and the German Research Ministry (BMBF Treatlon 01GM2210B). .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint Ethics declaration Informed conset for participation in this study was obtained from the parents of all probands in our CHOP EGRP cohort . The study was completed per protocol with local approval by CHOP Institutional Review Board (IRB 15-12226). Data and code availability Datasets in a deidentified format will be available upon request to the corresponding author. The DDD dataset files are publicly available from the European Genome-phenome Archive (study ID: EGAS00001000775), and the EPGP dataset is available from dbGap (accession: phs000653.v1.p1) respectively. Code for the primary analysis is available at https://github.com/helbig-lab/trio_hpo Additional information Correspondence and request for materials should be addressed to Ingo Helbig. Reprints and permissions information is available at http://www.nature.com/reprints. .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 12, 2025. ; https://doi.org/10.1101/2025.03.11.642649doi: bioRxiv preprint