Interpretação do Genoma Humano e Doenças: Desafios e Avanços no Aprendizado de Máquina

0123456789 Understanding human disease requires comprehensive interpretation of the genome including characterization of the impact of any variant on gene function and reg ulation Broadly this means that for any letter change in DNA we must precisely identify its effects on bio chemical properties such as protein structure splicing and levels of expression and then interpret these effects in terms of their phenotypic consequences In the past decade genomic sequencing has generated enormous amounts of data profiling both normal genetic variation and disease related mutations13 Concurrently func tional experiments profiling the epigenomic landscape of various cells and tissues45 have provided a window into the regulatory signals controlling where and when genes are expressed These experimental advances have fuelled the devel opment of machine learning algorithms that can pre dict the functional impact of any DNA variant Whereas interpretation of exonic mutations that disrupt protein coding and function is guided by the genetic code for missense mutations of uncertain impact computational approaches evaluating conservation611 secondary and tertiary protein structures1216 and additional bio physical properties for example protein stability1719 preservation of interactions with proteins2022 and other small molecules142324 have made great strides in predicting the molecular impact of any amino acid substitution on gene function However the majority of variants associated with disease lie outside exons in the non coding portion of the genome that makes up more than 98 of human DNA and there is no uniform way to decode their impacts Fig 1a Recently compu tational approaches opened the door to interpretation of the molecular and phenotypic effects of regulatory mutations including those never before observed in populations but that might appear in sequencing addi tional genomes2528 In particular a powerful kind of machine learning deep learning is opening frontiers in modelling complex biology and linking it to genomic sequence changes to give biomedical researchers and clinical geneticists insight into regulatory changes that contribute to human health and disease Deep learn ing approaches which typically require large amounts of training data have become increasingly useful for diverse biological applications2933 as high throughput data sets particularly consortium efforts providing systematically generated data across tissues conditions and diseases and greater computing power have become available3435 Furthermore work to provide easy to use frameworks to specify and share deep learning mod els for genomics3637 improve interpretability of deep learning models3839 and allow automated optimization of network architecture40 has improved the accessibil ity and power of deep learning approaches and made deep learning models the preferred method for many genomic applications41 Quantitative genetics studies have clearly shown that most complex human diseases involve contribu tions from multiple genetic variants but there remains a missing heritability problem only a small fraction of Deep learning A family of machine learning approaches involving multilayer models composed of interconnected nodes where the output of each node in the model is a function of its inputs Decoding disease from genomes to networks to phenotypes Aaron K Wong 14 Rachel S G Sealfon14 Chandra L Theesfeld24 and Olga G Troyanskaya 123 Abstract Interpreting the effects of genetic variants is key to understanding individual susceptibility to disease and designing personalized therapeutic approaches Modern experimental technologies are enabling the generation of massive compendia of human genome sequence data and associated molecular and phenotypic traits together with genome scale expression epigenomics and other functional genomic data Integrative computational models can leverage these data to understand variant impact elucidate the effect of dysregulated genes on biological pathways in specific disease and tissue contexts and interpret disease risk beyond what is feasible with experiments alone In this Review we discuss recent developments in machine learning algorithms for genome interpretation and for integrative molecular level modelling of cells tissues and organs relevant to disease More specifically we highlight existing methods and key challenges and opportunities in identifying specific disease causing genetic variants and linking them to molecular pathways and ultimately to disease phenotypes 1Center for Computational Biology Flatiron Institute Simons Foundation New York NY USA 2LewisSigler Institute for Integrative Genomics Princeton University Princeton NJ USA 3Department of Computer Science Princeton University Princeton NJ USA 4These authors contributed equally Aaron K Wong Rachel S G Sealfon Chandra L Theesfeld e mail ogtcsprincetonedu httpsdoiorg101038 s4157602100389 x REVIEWS Nature reviews Genetics 0123456789 total heritability can be explained by disease associated variants identified by genome wide association studies GWAS even as these studies gain more power to detect associations42 Many studies have suggested that much of the heritability could be explained by the cumulative weak effects of many variants many of which fall below statistical significance for association4346 The question of how so many genomic regions with weak effects can influence disease susceptibility has prompted theories such as the omnigenic model where any gene expressed in a disease relevant tissue can affect core disease genes and thus disease risk through interactions in a complex Noncoding regulatory sequence Coding sequence GWAS SNP arrays 1 million bases detected 1060 million bases imputed from haplotypes PIK3CA BRCA2 BRAF HOXB13 KRAS BRCA1 ERBB2 PTEN Targeting sequencing 1 to hundreds of genes Exome sequencing 30 million bases Genome sequencing 3 billion bases Chromatin state Histone marks Transcription factors Exon 1 Exon 2 Exon 3 Exon 4 Exon 1 Exon 2 Exon 3 Exon 4 Ac Me Target gene TF a b Genetic code Phe Leu Ser Tyr Stop Stop Cys Trp Leu Pro His Arg Gln Ile ThrMet Asn Lys Ser Arg Val Ala AspGlu Gly G U A C G U CA A A A C C C U U U G G G Regulatory code deep learning based sequence models Fig 1 interpreting the other 99 of the genome a The genetic code guides our interpretation of disease associated mutations in coding regions of exon sequences these mutations can disrupt protein coding and function However the majority of disease associated variants lie outside exons in the non coding portion of the human genome that makes up 99 of DNA and there is no uniform way to decode their impacts Recent developments in deep learning based sequence models address this challenge by modelling the relationship between non coding sequences and properties influencing gene regulation such as chromatin modifications DNA accessibility and transcription factor TF binding These models can then be used to predict the effects of variants in the non coding portion of the genome b Detection of common and rare human genome variation Variants are discovered by targeted sequencing of genes for example panels of cancer related genes or sequencing of whole exomes or whole genomes Population wide sampling in genome wide association studies GWAS is done using single nucleotide polymorphism SNP genotyping arrays which are chips with probes for 1 million locations of common variation in the genome With this information sequences for many more positions can be imputed using population reference sequence haplotypes as guides wwwnaturecomnrg Reviews 0123456789 interconnected network4748 Thus understanding how diverse protein products interact in the cell under spe cific contexts such as in specific organs or tissues is crucial to pinpointing genes related to disease and the network effects of their dysregulation In this Review we discuss the challenges and advances in using omics data to interpret genetic vari ants associated with disease We cover the major sources of genetic variation and recent methodological advances that predict the regulatory effects of non coding muta tions including biochemical gene expression and pathogenic impact Finally we discuss methods that integrate omics data into tissue specific systems level models which can then be used to identify the genes and dysregulated biological processes associated with specific diseases Characterizing genetic variation Rapid technological improvements and associated reductions in costs have now made it possible to sequence the genomes of hundreds of thousands of people and projects have launched that aim to sequence the complete genomes of millions of participants4950 This unprecedented growth in available genomic data has resulted in enormous progress in deciphering the genetic underpinnings of human phenotypic variation and disease traits At the same time understanding the functional impact of specific genetic changes remains challenging especially for alterations in non coding regions of the genome Sequence variants can be profiled at several dif ferent levels of resolution Fig 1b Single nucleotide polymorphism SNP arrays can be used to genotype polymorphic positions throughout the genome with more than a million SNPs frequently considered For example the landmark UK Biobank data set includes array based genetic profiling for nearly 500000 indi viduals which can then be investigated in combination with phenotypic information and medical histories to identify disease associated genetic signals2 To pinpoint genetic signatures for particular traits GWAS can com pare the variants observed in individuals who have a particular phenotype such as a disease with control individuals to identify loci that are associated with the trait51 Genotyping arrays incorporate common SNPs to tile along the genome backbone To increase SNP density and add statistical power millions of additional SNPs can be inferred by imputation based on haplo types of populations under consideration5253 However the associated SNPs may have indirect relationships to the trait For instance although the common vari ants on the array may be statistically associated with a disease the identity of the actual allele that is causally linked to the trait can be a rare allele that is not assayed in either the array based or imputation based genotyp ing steps Furthermore GWAS may identify multiple positions in linkage disequilibrium that are associated with the trait in which case further investigation of the associated SNPs through fine mapping is necessary to pinpoint trait associated alleles54 Genetic variants can also be identified by genome sequencing which can cover either limited regions or nearly the entire genome Specific gene panels may be sequenced to test patients for exonic mutations that are known or suspected to be involved in disease Gene panel testing is now routinely used in clinical prac tice for example in prediction of breast cancer risk55 In whole exome sequencing WES coding regions of essentially all genes in the genome are sequenced This approach has the advantage of dramatically decreas ing the fraction of the genome that must be sequenced relative to whole genome sequencing WGS while still profiling the most readily interpretable genetic variants Exome sequencing can also be used to detect somatic copy number variation which is important for characterizing tumour genomes56 However many disease causing mutations are located in non coding portions of the genome and WES will not typically pro vide information on these variants although targeted non coding regions for example non coding regions of known clinical importance can be sequenced in addition to exomes Moreover WES sequencing may fail to capture some exonic portions of the genome in particular regions that are GC rich57 By contrast in WGS nearly the entire genome including both coding and non coding portions is sequenced The inclusion of non coding portions of the genome provides crucial information for understanding the genetic architecture of disease traits as an estimated 90 of disease related genomic intervals are in non coding space58 WGS is increasingly seeing use in direct clinical assessments identifying genetic contributors to both acute illness and disease risk5961 Numerous key consortia have aimed to collect the genomic sequences of large numbers of individuals A pioneering effort in this area was the 1000 Genomes Project active from 2008 to 2015 which studied healthy individuals and aimed to catalogue most genetic variants at 1 frequency or higher in the sampled populations1 The recently launched All of Us consortium aims to sequence the genomes of at least a million American participants with diverse ethnic backgrounds and med ical histories62 In parallel several databases have been developed to make large collections of sequence data accessible to researchers For example the Genome Aggregation Database gnomAD contains more than 125000 exome sequences and 70000 whole genome sequences as of its version 3 release63 Efforts to perform WGS on the UK Biobank cohort are in progress2 Discovery of the functional impact of mutations and attributing these effects to disease causation is a major challenge Large scale sequencing data have been used to identify genomic regions that are likely to be disease associated based on their patterns of observed versus expected variation regardless of coding or non coding space6466 reviewed by Eilbeck et al67 For population based studies although large sample sizes associate increasing numbers of loci with diseases they still do not explain most of the estimated heritability of complex traits For complex highly polygenic diseases enormous sample sizes tens of millions of individuals are needed to unravel causality of most related loci each of which may individually have a small impact on susceptibility68 Furthermore many possible mutations Nature reviews Genetics Reviews 0123456789 are never observed either because they do not occur in a given sample or because they would be lethal before birth Thus methods that do not depend on observing an alteration in a population are needed to complement observation based studies in order to disentangle the relationship between genomic sequences and disease traits Interpretation of coding mutations Computational frameworks for assessing the impact of genetic alterations in the coding portion of the genome although still an area of rapid development are com paratively mature relative to approaches for interpreting non coding variants12156973 Multiple types of evidence such as the type of alteration to the protein sequence mis sense nonsense or frameshift the degree of simi larity between a reference and a substituted amino acid the evolutionary conservation of the altered position and the predicted biophysical effect on protein structure can contribute to understanding the likely effect of a change in the coding portion of the genome A large number of methods have been developed to leverage these factors to predict the impact of coding mutations12156973 We describe here a few commonly used approaches that illus trate different classes of methods for coding variant effect prediction One class of methods relies primarily on sequence conservation for predicting variant effects For exam ple SIFT Sorting Intolerant From Tolerant7475 takes a conservation based approach to variant effect predic tion examining conservation across related proteins in a protein family The premise is that if an amino acid is highly conserved among members of a protein family it is likely to be essential whereas if more variation is tolerated at that position it is less likely that an alteration will have a large effect Another key class of features for predicting the impact of a coding variant involve protein structural predictions One example is the VIPUR Variant Interpretation and Prediction Using Rosetta frame work which incorporates protein structural models to predict variant effects12 Rather than simply determin ing how deleterious a mutation is likely to be VIPUR uses the protein structural models to predict the precise effect of the variant for example whether the active site of the protein is disrupted whether the stability of the protein is altered or whether the protein folding is likely to be disrupted PertInInt explicitly integrates structural information with small molecule binding information to predict structure based impacts of pro tein variants on binding to RNA peptides ions and drugs1314 Finally there are many frameworks that integrate combinations of feature types to generate an aggregate prediction reflecting the impact of the mutation on various sequence and protein properties An example of this type of method is PolyPhen2 Polymorphism Phenotyping v2 which uses a combination of sequence based and structure based features to pre dict the impact of a coding variant16 A naive Bayesian classifier is then used to predict the overall impact of the change from the individual features Combined Annotation Dependent Depletion CADD6976 and Eigen77 are two other widely used machine learning frameworks that integrate a large number of feature types in order to predict the effects of both coding and non coding variants Models for understanding the impact of coding genetic variants have played a key role in identifying alleles that contribute to human disease7881 A strength of these methods is that their output is often readily inter pretable However most approaches used to understand the effect of coding variants are highly tailored to lever age properties of protein coding portions of the genome Fundamentally different computational approaches are needed to elucidate the impact of non coding variants Modelling transcriptional effects As only a small fraction of the human genome is protein coding and most variants are situated in the non coding portion of the genome developing approaches that can address the problem of understand ing the effects of non coding variants is a critical chal lenge Methods that can model the relationship between non coding sequences and properties influencing gene regulation such as chromatin modifications DNA accessibility and transcription factor binding can be used to predict the effects of variants in the non coding portion of the genome2527288284 TAble 1 Key to these efforts are large collections of genome wide profiles for chromatin regulators in a range of cell types and tissues45 enabling methods to predict the impact of variants even in primary tissues2527288385 A strength of models that can predict variant effects based only on sequences is that they enable the predic tion of molecular and disease impact of rare as well as commonly observed genetic alterations and even muta tions that have never or not yet been observed This is an important feature because unique and rare vari ants are rapidly being discovered a recent paper by the Trans Omics for Precision Medicine TOPMed con sortium identified 400 million varying bases in 53000 newly sequenced whole genomes Almost all variants were present at low allele frequency 1 and 46 were present in only one individual86 making any statistical association studies infeasible for linking SNPs to disease due to lack of power Several frameworks use traditional supervised machine learning algorithms or probabilistic model ling to predict the impact of non coding alterations For example gkm SVM uses a support vector machine SVM with gapped k mers as features to predict protein binding to sequences2887 gkm SVM significantly out performed an earlier implementation using ungapped k mer features illustrating the utility of incorporating a longer sequence context for predicting regulatory function Success at such prediction problems has been seen recently using deep learning based frameworks which leverage genome scale data on these attributes to gen erate sequence based predictive models252782848890 A network architecture that is particularly suited for extracting complex patterns from sequence data is the convolutional neural network CNN which was Variant effect The biochemical or phenotypic impact of a genetic variant relative to a reference allele Deleterious The attribute of an allele as it relates to phenotypic impact this can be through decreased organismal fitness that is often associated with increased disease risk Support vector machine SVM A standard supervised machine learning approach that identifies the hyperplane dividing line in high dimensional space that optimally separates positive examples from negative examples k mers Short lengths of nucleic acid used in computational algorithms oligomers of k length in bases Convolutional neural network CNN A class of deep learning models that use structure in the input data for example adjacencies of pixels in an image or of base pairs in a sequence to inform connections between nodes of the model Successive outputs often model features at increasing spatial scales for example for sequence models sequence motifs larger sequence contexts wwwnaturecomnrg Reviews 0123456789 initially widely applied to image analysis problems and uses spatial relationships across features to guide net work connectivity For sequence models CNNs take raw DNA sequences as input and model regulatory and epig enomic elements at increasing genomic scales without the need for manual selection of features Intuitively nodes in successive layers can model individual motif elements groups of motifs regulatory regions and relationships across regulatory regions An early advance in modelling large scale data was the DeepSEA framework a CNN based approach trained on a compendium of genome wide transcription factor binding DNase hypersensitivity and chromatin modification data modelling 2000 chromatin features that impact transcription in 200 cell types2526 Fig 2 The model then predicts the molecular effect of a vari ant with single base resolution given only sequence data The DeepSEA framework can be tailored to specific dis eases by training on disease relevant data91 In another method DeepBind88 individual CNNs more accurately modelled protein binding data compared with 26 other traditional non deep learning and k mer methods A related approach is used in the Basset framework which learns a CNN model to predict cell type specific DNase hypersensitive sites from sequence data27 Basenji an extended and modified version of Bassett can predict the quantitative impact of sequence alterations on reg ulatory modifications as well as considering the impact of sequences in distal genomic regions85 An important challenge is predicting chromatin effects of human variation for any cell type However it is experimentally intractable to generate genome wide chromatin immunoprecipitation followed by sequenc ing ChIPseq profiles for all binding factors in all cell types and tissues Thus generalizable methods that can be used to impute missing data for example binding in cell types unobserved in model training are crucial9293 FactorNet developed from DanQ addressed this through combining recurrent neural network RNN architectures that use information learned previously to inform new training examples8290 FactorNet trains on transcription factor binding in reference cell types and tests on distinct cell types ChromDragoNN addresses prediction across cell contexts through cou pling the learning of cis regulatory sequences with gene expression of trans regulators using multimodal residual neural network architectures that have the flex ibility to integrate distinct data types83 FactorNet and ChromDragoNN approaches demonstrate the promise that modelling the accessibility of a genomic region in available data can be a powerful baseline predictor for biological contexts with no available data Deep neu ral networks have also been applied to epigenomics where sequence modifications are difficult to assay For example DeepCpG predicts missing methylation at CpG sites in single cells using a modular arrangement of neural networks to learn sequence patterns around measured CpG sites using a CNN and capture the Table 1 Methods for transcriptionalbiochemical impact Model overview input data Resources Model Method Prediction task Public interactive Web interface Access DeepSEA2526 Deep learning CNN Chromatin TF binding TF HM DHS httpshbflatironinstitute orgdeepsea httpsgithubcom FunctionLabselene gkm SVM delta SVM28 SVM Chromatin TF binding TF HM DHS httpwwwbeerlaborg deltasvm DanQ82 Deep learning CNN BLSTM Chromatin TF binding TF DHS HM httpgithubcom uci cbclDanQ Basset27 Deep learning CNN Chromatin accessibility DHS httpwwwgithubcom davek44Basset DeepCpG89 Deep learning CNN GRU CpG state Bisulfite sequencing httpsgithubcom PMBiodeepcpg ExPecto95 Deep learning CNN linear regression Expression prediction TF HM DHS RNA seq httpshbflatironinstitute orgexpecto httpsgithubcom FunctionLabExPecto Basenji85 Deep learning CNN Expression prediction TF DHS HM CAGE peaks httpswwwgithubcom calicobasenji BPNet DeepLIFT TFModisco84 Deep learning CNN TF binding TF httpsgithubcom kundajelabbpnet ChromDragoNN83 Deep learning CNN ResNet Chromatin accessibility DHS RNA seq httpsgithub comkundajelab ChromDragoNN Xpresso8394 Deep learning CNN Expression prediction CAGE peaks gene annotations httpsxpresso gswashingtonedu AMBER40 Auto machine learning RNN deep learning CNN Chromatin TF binding TF HM DHS httpsgithubcom zj zhangAMBER All methods have the sequence as input and single nucleotide resolution for variant effect predictions All models can indicate cell type specific predictions depending on training data BLSTM bidirectional long short term memory CAGE cap analysis gene expression CNN convolutional neural network DHS DNase hypersensitive site DNA accessibility GRU gated recurrent network HM histone marks ResNet residual neural network RNA seq RNA sequencing RNN recurrent neural network SVM support vector machine TF transcription factor Sequence models A deep learning framework that models the relationship between genetic sequences and properties influencing gene regulation Recurrent neural network RNN A type of neural network in which learning of inputs is influenced by past instances of input examples and so output varies depending on the sequence of inputs For example in speech recognition applying the context of prior words is useful for determining the meaning of a new word Hidden layers pass weights to input information based on previously learned examples Nature reviews Genetics Reviews 0123456789 neighbourhood of CpG site across cells using a gated recurrent network enabling single cell genome wide analysis of methylation states89 The development of interpretable models is crucial because they can lead to biological insight improve user confidence in predictions and reveal potential techni cal biases As deep learning frameworks are composed of many layers of transformations and in effect subject input data to a complex mathematical transformation it is often difficult to trace how input features contribute to the final prediction One approach to interpreting sequence models is to systematically alter bases in the input data and observe how the output predictions change in silico mutagenesis2588 Another approach exemplified by the DeepLIFT framework uses backpropagation through the neural network to elucidate which input features contribute to which aspects of the output38 For sequence models the ability of DeepLIFT to define motifs for transcription factor binding was applied to understand regulatory grammar syntax84 and trans regulatory mechanisms for the usage of transcription start sites94 Output from models that predict proximal molecu lar effects of variants can be incorporated into frame works to predict the wider functional impact of these changes Information on the chromatin modification and transcription factor binding profile of sequences can be used directly to predict effects on gene expression ExPecto uses a modified version of the DeepSEA model and adds a tissue specific regularized linear model of gene expression based on the chromatin mark DNase hypersensitivity and transcription factor binding profile of its upstream sequence95 By comparing the predicted expression of the gene in the presence of a variant versus a reference allele ExPecto can infer the likely effect of a sequence mutation on gene expression in a specific tis sue A recent approach integrated SNP effect predictions from multiple methods used multiple approaches to link genes to SNPs and predicted expression from sequence features in order to infer disease impact of variants96 By leveraging predictions of the direct effects of non coding changes for example alterations in transcription factor binding or epigenetic profile or expression in specific Train deep learning model TGTTGTCAATCGCTACGGAGC Allele C TGTTGTCAATGGCTACGGAGC Allele G Allele C Allele G a Input genomic sequences c Output allelespecific regulatory predictions d b Genomewide chromatin profiles ENCODE Roadmap Epigenomics CLIP Regulatory effects model e Output predicted variant pathogenicity f ASD variant disease impact scores chr6137918071 G C RBP binding Transcription factor binding Histone modification Regulatory event probability 0 5 10 15 20 25 Ac Me Proband Sibling G C G C A A T T G G C A C A T T G G C A T A T T Train pathogenicity model Curated disease regulatory mutations Disease impact model Disease impact score Disease impact score 0 5 10 15 20 chr6137918071 G C Fig 2 Learning the regulatory code to predict variant effects a Sequence based deep learning models predict the biochemical impact of alterations in non coding sequences by comparing predicted effects between two alleles b Models are trained on genome wide profiles of attributes such as chromatin modifications transcription factor binding sites and RNAprotein binding sites They can then predict the profiles of the novel input sequences pinpointing the effect of sequence variants As the models require only sequences as input and are not trained on variant information they can predict the effect of mutations that are rarely or never seen in sequence databases c The model predicts the specific impact of a given variant on attributes such as binding of specific RNA binding proteins RBPs histone modifications and transcription factor binding de These biochemical effects can be used for further analyses For example to assess how biochemical effects relate to disease impact a pathogenicity model was trained on predicted biochemical effects of known regulatory mutations that contribute to human disease part d enabling prediction of pathogenicity for any variant part e f This framework was applied to predict the impact of de novo mutations in probands with autism spectrum disorder ASD and unaffected siblings Mutations in probands had significantly higher predicted disease impact than mutations in unaffected siblings26 CLIP cross linking immunoprecipitation ENCODE Encyclopedia of DNA Elements wwwnaturecomnrg Reviews 0123456789 cell types and tissues to understand the impact on gene expression these frameworks facilitate the understand ing of the disease impact of non coding changes at the biochemical regulatory and phenotypic levels Making deep learning models and resources available to the broader biomedical community is key to accel erating development and adoption by researchers This includes user friendly dynamic web interfaces where researchers can directly generate predictions and model repositories that collate available models36 Resources that empower scientists who are not deep learning experts to develop algorithms and train and test deep learning models include code libraries such as Selene37 Modelling post transcriptional effects Variants that alter post transcriptional properties of genes such as interaction with RNA binding proteins or splicing can also contribute to disease risk97100 Sequence based deep learning models can be used to predict the precise post transcriptional effect of a spe cific variant including regulatory effects of synonymous mutations26101106 TAble 2 One key task is to predict the impact of variants that affect splicing and various innovative methods have taken advantage of the molecular mechanisms of splic ing to detect splicing events For example the DeepSplice and SpliceRover frameworks use CNN models to predict splice junctions from sequences101104 The Deep Splicing Code framework trained multiple deep learning mod els for distinct prediction tasks including identifying whether an exon has alternative 5 or 3 splice sites and determining whether an exon is constitutive or alterna tively spliced102 Deep learning models have also been used to infer intronic RNA splicing branch points107108 Despite the extensive progress in identifying alternative splicing events largely through data from short read next generation sequencing technologies and genome annotations determining full length and alternative RNA isoforms and how variants impact their forma tion is still an open challenge109110 Sequencing technol ogies for example long read direct RNA sequencing and targeted deep sequencing and modelling methods are rapidly rising to find end to end full length RNA species111113 reviewed recently in ReF114 A broader goal is predicting diverse post transcriptional regulatory consequences including not only splicing but also RNA structure stability and trans lation efficiency through interactions with RNA binding proteins115119 including tRNAs and microRNAs120121 Notably sequence based models of RNAprotein inter actions using CNN frameworks have been developed for de novo non coding variant effect prediction2688122 For example Seqweaver was used to predict the post transcriptional regulatory impact of variants from a whole genome analysis of simplex autism families demonstrating significant RNA regulation disruption in de novo variants of children with autism spectrum disorder26 In an even broader application of this model Park et al found that the dysregulation of wide ranging categories of post transcriptional regulation is a pri mary contributor to the inherited risk of psychiatric disorders122 Table 2 Methods for post transcriptional impact Model overview input data Resources Model Method Prediction task RBP profiles chromatin profiles Genome annotations Public interactive Web interface Access DeepBind88 Deep learning CNN Chromatin and RBP binding PBM CLIP SELEXTFs httptoolsgenestoronto edudeepbind DeepSplice188 Deep learning CNN Splice junctions Splice sites httpsgithubcom zhangyimcDeepSplice LaBranchoR107 Deep learning BLSTM Intronic branch point sites Annotated branch points httpbejeranostanfordedu labranchor SpliceRover101 Deep learning CNN Splice donor and acceptor sites Annotated donor and acceptor sites httpbioit2ircugentbe roversplicerover Seqweaver26122 Deep learning CNN RBP binding impact CLIP httpshbflatironinstitute orgseqweaver httpshbflatironinsti tuteorgseqweaver about Deep Splicing Code102 Deep learning CNN Alternative splice site usage ESTs HEXEvents annotated splice events httpsgithubcom louadiDSC SpliceAI109 Deep learning CNN Splice donor and acceptor sites GENCODE RNA annotations httpsgithubcom IlluminaSpliceAI All methods have the sequence as input and have single nucleotide resolution for variant effect predictions All models can indicate cell type specific predictions depending on training data BLSTM bidirectional long short term memory CLIP cross linking immunoprecipitation CNN convolutional neural network EST expressed sequence tag PBM protein binding microarrays RBP RNA binding protein SELEX systematic evolution of ligands by exponential enrichment TF transcription factor Nature reviews Genetics Reviews 0123456789 Deciphering pathogenic variant effects In order to decipher the importance of a variant in contributing to disease processes it is crucial to go beyond biochemical effect and understand the disease impact and ultimately clinical consequence Several data resources aggregate current knowledge on variant impact For example the ClinVar database is a pub lic resource from the US National Institutes of Health NIH that has so far compiled information on 800000 genetic variants annotated to multiple diseases and with various levels of clinical significance using literature to document evidence and panel review to indicate confi dence in each annotation ClinVar statistics 8 December 2020123126 The commercial Human Gene Mutation Database HGMD is a curated collection of more than 275000 variants HGMD 20194 sourced from pub lished associations between genetic variants and human disease127 The American College of Medical Genetics and Genomics currently maintains a list of 59 genes with mutations that lead to well defined and highly penetrant clinical phenotypes128 that can be systematically profiled in clinical genetic testing and reported MaveDB129 and BioGRID Open Repository of CRISPR Screens ORCS130 collect experimentally determined molecular impacts for thousands of sequence variants systematically profiled as part of biomedical research131 Through synthesizing and making accessible the current state of knowledge on disease causing variants such databases provide an invaluable resource for interpreting genetic data However such clinical information is available for a very small and likely biased fraction of variants especially so in the non coding space Further sys tematic characterization of variants at the pheno typic and endophenotypic level such as through GWAS clinical databases and experimental studies for example in organoids is essential see Integrative network models below for discussion of systematic network based methods deployed for this purpose That said full characterization of the genome must rely on computational approaches especially for the vast non coding genome approximately 294 billion bp where most disease associated mutations map Given the regulatory consequences of variants predicted by the deep learning models described above and in Fig 2 and TAbleS 12 machine learning methods can be trained to predict how deleterious each variant might be on the phenotypic level by training on example variants with associated phenotypic consequences Fig 2d TAble 3 The resulting pathogenicity scores predict which variants of molecular impact will also have potential clinical impact Pathogenicity refers to the impact on organism fitness which is often closely correlated with the appearance of diseases especially those that mani fest before or during reproductive years For example CADD scores76132 reflect the deleteriousness of variants the scores differentiate variants based on the likelihood they are evolutionarily constrained or neutral through integrating dozens of features including genome anno tation such as intron or untranslated region UTR evolutionary GERP710 and SIFT73 and functional information on variants in both coding PolyPhen2 ReF16 and non coding epigenomic profiles from the Encyclopedia of DNA Elements ENCODE133 and Roadmap Epigenomics5 portions of the genome The predicted regulatory features from DeepSEA and Seqweaver for both coding and non coding variants can be leveraged to predict the disease impact score DIS for any mutation The DIS is calculated by training a logistic regression model with predicted regulatory effects for verified disease causing mutations in HGMD posi tives and low frequency non disease variants nega tives This score was used in whole genome analysis to discover increased mutational burden in probands with autism compared with unaffected matched sib lings improving upon traditional count based analy ses by enabling assessment of the functional impact of de novo mutations26 Constraint violation scores predict on a per gene and per tissue basis disease allele versus non disease allele status for variants without knowledge of disease association These scores integrate predicted expression effects from ExPecto with inferred evolution ary constraints in a given tissue95 For example if a vari ant a SNP or small insertion or deletion indel causes a specific highly expressed positively constrained gene to have a significant decrease in expression constraint is violated and the variant is likely to be deleterious and pathogenic This method accurately prioritized puta tive causal SNPs in the GWAS catalogue over SNPs in linkage disequilibrium Another approach LINSIGHT combines modelling the effects of natural selection INSIGHT with functional molecular data to predict the deleteriousness of non coding variants134135 There is great potential in developing more accurate pathogenicity scores both through integrating addi tional information and by leveraging newly character ized disease mutations as such data become available Benchmarking of variant prediction methods against well annotated sets of disease mutations and functional assay data for example in the extensively characterized BRCA1 and BRCA2 genes indicates that variant predic tions are significantly correlated with but do not fully recapitulate experimental data suggesting that although such methods can aid in clinical applications they cannot yet be used in isolation to confidently determine the clin ical consequences of a variant136138 Although the vari ant databases provide samples of disease mutations the comparatively small number and lack of strong evidence for disease association makes for noisy representation of true disease mutations their associated functional genomic profiles and evolutionary signatures Thus a limitation in developing any sequence model is the lack of unbiased experimentally and clinically confirmed var iant effects that can be used as gold standards in evalua tions Despite this an orthogonal analysis demonstrated that the predicted variant effects from sequence models are significantly enriched in disease heritability pro viding additional evidence that they are informative for discovering disease influencing variants139 With regard to evaluations the ClinVar variants reviewed using community agreed standards girded with multiple lines of evidence and reviewed by an expert panel will become an important benchmark in develop ing and evaluating variant impact scores Methods that use conservation scores or allele frequencies are of less Endophenotypic An aspect of a complex trait that may be more experimentally measurable than the entire complex trait and may be closer to an underlying biological process For example educational attainment is an endophenotype examined in the study of autism genetics because it is a readily measurable trait associated with autism and expression levels of insulin receptors are endophenotypes contributing to type 2 diabetes Non coding genome The portion of the genome that does not encode proteins which comprises more than 98 of the total human genome length Probands in a genetic study individuals with the disease typically the particular affected individuals within families who are the starting points for genetic analyses wwwnaturecomnrg Reviews 0123456789 Table 3 Methods for pathogenicity Model overview input data Resources Model Method Prediction task tissue speci ficity RnA expre ssion Protein struc ture Amino acid bioche mistry biophysics Genome annota tion cDs UtR con ser vation chro matin profiles RBP profi les Public interactive Web interface Access PhastCons69 MSA phylo HMM Within conserved elements per base estimate of negative selection Yes httpcompgen cshleduphast https githubcom CshlSiepelLab phast GERP710 MSA MLE Per base and per element score of constraint Yes httpmendel stanfordedu SidowLab downloads gerp PhyloP911 MSA statistical analysis by clade P value conservation acceleration scores per base divergence from neutral rate allowing clade specific selection pressures Yes httpcompgen cshleduphast https githubcom CshlSiepelLab phast CADD6976 Multi data integration classifier SVM Coding and non coding variant impact deleteriousness of SNVs and small indels based on functional and evolutionary information Yes Yes Yes Yes https caddgs washin gtonedu LINS IG HT8135 Multi data integ ration score GLM probabilistic model of constraints Non coding mutation fitness impact based on functional and evolutionary information RNA seq Yes Yes Yes https githubcom CshlSiepelLab LINSIGHT DANN132 Deep learning version of CADD CNN Coding and non coding variant fitness impact for SNVs and small indels based on functional and evolutionary information Yes Yes Yes Yes httpscbcl icsuciedu publicdata DANN Eigen77 Unsupervised meta score Non coding mutation impact Yes Yes Yes Yes TF DHS HM httpwww columbia eduii2135 eigenhtml Constraint violation score ExPecto95 Multi data integration score Non coding variant impact disease risk associated with SNV and small indels based on variant impact violation of cumulative regulatory impacts across genomic interval By gene and tissue RNA seq Yes TF DHS HM Supp Data 2 has directionality scores for calculating CVS Nature reviews Genetics Reviews 0123456789 value for sets of SNPs that contribute to disease but are not directly linked to fitness For example cumulatively common variants contribute significantly to risk for complex diseases but may be under weak selection and disease variants in linkage disequilibrium with benefi cial alleles can be present in the population at higher fre quency than expected Such conservation based methods will underscore these variant sets Scores that accurately predict functional clinical impact that operate at a wide range of variant frequencies in coding and non coding regions of the genome and regardless of genetic archi tecture for example common inherited variants and de novo rare variants will continue to have increasing utility for biomedical researchers studying the aetiology of diseases and clinicians choosing treatment modalities to meet the promise of personalized medicine Integrative network models Although sequence models can predict the molecu lar effects of mutations including on tissue specific gene expression explaining how these alterations give rise to disease phenotypes requires an understanding of the dysregulated pathways and processes Recent work attempting to explain how even common variants with weak effects can still increase disease susceptibil ity theorizes an omnigenic model wherein any gene expressed in a disease relevant tissue can affect core disease genes through interactions between genes that function together in processes pathways and larger net works Thus understanding how the molecular effects of genomic mutations influence disease requires models of cellular networks and pathways that are active in various tissues in healthy or disease states Diverse functional genomic data including gene expression45140141 and proteomics142143 provide a window into the genome scale interactions between biomolecules in various contexts Inferring accurate models from these data is challeng ing because these data are produced on heterogeneous data modalities experimental designs and technological platforms Furthermore the models must capture both shared and unique processes across cell types and physio logical conditions Sophisticated computational tech niques are needed to extract context specific biological relationships from the raw data sets Integrative approaches that combine diverse data sources can identify genome scale functional maps of cells across biological and environmental contexts144149 Whereas individual experiments vary in biological relevance and signal by analysing data collections as a whole faint but recurrent signals can be amplified Model overview input data Resources Model Method Prediction task tissue speci ficity RnA expre ssion Protein struc ture Amino acid bioche mistry biophysics Genome annota tion cDs UtR con ser vation chro matin profiles RBP profi les Public interactive Web interface Access Disease impact score DeepSEA26 Multi data integration classifier SVM Coding and non coding variant impact disease association of SNVs and small indels based on transcriptional regulatory impact in a genome sequence model TF DHS HM Yes httpshb flatironinstitute orgdeepsea https hbflatironi nstituteorg asdbrowser CADD Combined Annotation Dependent Depletion CDS coding sequence CNN convolutional neural network DHS DNase hypersensitive site DNA accessibility GLM generalized linear model HM histone marks indel insertion or deletion MLE maximum likelihood estimation MSA multiple sequence alignment phylo HMM phylogenetic hidden Markov model RBP RNA binding protein RNA seq RNA sequencing SNV single nucleotide variant SVM support vector machine TF transcription factor UTR untranslated region Table 3 cont Methods for pathogenicity Fig 3 informing quantitative genetics data with network models Molecular networks can be used to reprioritize the statistical associations from genome wide association studies GWAS using functional information a Methods such as NetWAS Network Wide Association Study first convert GWAS single nucleotide polymorphism SNP P values to gene level P values using various methods189 and then train a machine learning classifier using nominally significant GWAS genes as positive examples blue nodes non significant genes as negative examples grey nodes and features as weights from a gene network The classifier outputs effectively re rank all genes by their connectivity to the top GWAS genes identifying candidates most likely to be associated with the studied diseasephenotype b An example of NetWAS in reprioritizing hypertension HTN systolic blood pressure SBP and diastolic blood pressure DBP GWAS genes148 Manhattan plots and the kidney specific network shown are schematic Gene functional relationships from a kidney specific network were used as features in a support vector machine SVM For each GWAS positive examples were GWAS hits with nominal gene P 001 and non significant genes as negative examples The SVM effectively re ranked all genes by their connectivity to the top GWAS genes enriching for disease relevant signals including for example hypertension associated genes processes and pathways wwwnaturecomnrg Reviews 0123456789 These methods deploy various machine learning tech niques to summarize massive collections of human omics data for example transcriptional profiles proteinprotein interactions and sequence motifs into gene interac tion networks specific to a biological context In these genome wide networks nodes represent genes and edges reflect a predicted confidence that two genes participate in similar biological processes These networks can pro vide both a systems level view and specific experimentally testable hypotheses of gene function and interactions 0 5 10 15 log10P 1 2 3 4 5 6 7 8 9 10 11 12 13 14 16 18 20 22 Chromosome a b Reprioritized gene list Machine learning classifier eg SVM Functional gene network GWAS data Hypertension GWAS Kidneyspecific network Classifiers to rerank HTNassociated candiate genes Reprioritized HTNassociated candidate genes Nominally significant genes as positives 0 5 10 15 log10P Chromosome 0 5 10 15 log10P Chromosome 0 5 10 15 log10P Chromosome Nonsignificant genes as negative examples SBP DBP HTN SBP DBP HTN SBP DBP HTN 0 5 10 15 Enrichment Gene P value Score GWAS Reprioritization AQP2 KCNJ1 XPNPEP2 HNF1B ALDOB PAPPA2 TRIM10 CUBN SLC17A1 HNF4A CYP2A6 SCGB1A1 GIF HFE GC AQP2 KCNJ1 XPNPEP2 HNF1B ALDOB PAPPA2 TRIM10 CUBN SLC17A1 HNF4A CYP2A6 SCGB1A1 GIF HFE GC Nature reviews Genetics Reviews 0123456789 Context specific networks can be constructed directly from correlations of gene expression profiles150152 or by combining global proteinprotein interaction networks with disease or tissue specific gene expression153 However these methods are limited by the availability and quality of tissue expression profiling data which for many human tissues and cell types remains challenging or even infeasi ble to assay experimentally More recent methods address these limitations by applying machine learning to predict tissue specific functional interactions from large com pendia of genomic data148149154157 For example GIANT Genome Wide Integrated Analysis of gene Networks in Tissues148 now available at humanbaseio uses a regu larized Bayesian classifier to predict networks for more than 100 tissues and cell types by assessing and weight ing each data set by relevant tissue signal The method can predict genegene interactions in tissues for which few tissue specific genome scale data exist A systematic study by Huang et al158 evaluated 21 human interaction networks on their ability to predict disease genes and found ConsensusPathDB159 GIANT148 and STRING147 to have the best performance overall although the evaluation was limited to global that is tissue unaware networks Network analysis of quantitative genetics data Network based machine learning approaches can be used to leverage prior experimental knowledge and improve the interpretation of large scale quantitative genetics studies or individual targeted disease specific studies Intuitively these methods use the functional genomic information about pathways encoded in these networks to increase the signal to noise ratio in genetics studies Non coding regulatory variants can be linked to putative target genes and these target genes can then be subjected to network based analyses One approach is to reprioritize GWAS hits based on information in networks that reflect relationships among genes encapsulated in thousands of biologi cal experiments160 Broadly these methods identify disease specific connectivity patterns based on the behaviour of GWAS prioritized genes and reprioritize all genes based on these patterns Fig 3 Importantly these methods are discovery driven and do not depend on prior literature based knowledge in their reprioriti zation instead they use the genome wide information from disease focused GWAS projects For example the NetWAS Network Wide Association Study framework is a machine learning method that takes as input margin ally significant GWAS hits as positives and lower ranked genes as negatives and uses tissue specific network edge weights for a disease relevant tissue as the feature set in a support vector classifier to reprioritize candidate genes148 A recent NetWAS 20 approach also includes a subsampling procedure to weight negative examples proportional to their GWAS P value155 Another method Camoco Coanalysis of molecular components uses co expression networks to identify GWAS hits that share similar expression patterns across experimental assays finding that networks constructed using tissue relevant co expression data sets showed the best performance161 Network based methods can also directly use muta tional data from sequencing patient genomes to identify genes and pathways significantly associated with dis ease For example a network based method was used to evaluate potential disease genes mutated in patients with hereditary plastic paraplegias and to predict addi tional candidate disease genes162 Another approach networkneighbourhood differential enrichment analysis NDEA uses functional networks to boost power in enrichment analysis in application to autism spectrum disorder NDEA identified pathways signif icantly associated with autism based on the proband specific mutational burden in brain specific network neighbourhoods26 Another group of methods aims to identify cancer genes and pathways by overlaying mutational data onto molecular networks163167 reviewed in ReF168 These methods such as NetSig169 are based on the observation that although many genes carry genetic alterations in a particular tumour profile only a small number of these genes are cancer drivers3 Such driver genes are likely to function in shared biological processes and pathways across multiple tumours170 By examining the connec tivity of mutated genes in gene networks it becomes possible to identify subnetworks with greater mutational burden thus increasing power to detect cancer drivers as well as elucidating their functional impact Molecular architecture of disease More broadly a large number of methods aim to iden tify candidate disease genes based on the molecular net works that summarize large omics data compendia146171173 reviewed in ReF174 These methods analyse the con nectivity patterns of known disease genes from lit erature GWAS and so on in a network and identify novel candidates with similar patterns Importantly the molecular network used for diseasegene predic tion is crucial for accurate and relevant predictions particularly because the dysregulated genes underlying diseases are frequently involved in tissue specific and context specific processes175176 For example in a recent study a brain specific network significantly outperformed the global tissue unaware network in predicting autism genes and the top predicted genes from the brain network showed significant enrichment in autism probands rela tive to unaffected siblings177 Networks can further serve to systematically integrate information and data across organisms facilitating the effective use of model organ isms in the study of human disease154178180 In the dis easeQuest method human quantitative genetics data are coupled with tissue specific networks from Caenorhabditis elegans to predict candidate Parkinson disease genes and then to experimentally verify them in the worm154 Networks also provide a powerful approach to analysing and visualizing relationships among genes or gene sets for example to identify groupings that is modules that elucidate functional themes among disease implicated genes Fig 4ab Various approaches have been developed for network based module discovery177181182 with publicly available implementa tions in tools such as Cytoscape183 cytoscapeorg and HumanBase humanbaseio For example applying a Louvain community detection algorithm to genes associ ated with autism spectrum disorder177 in a brain specific Tissue specific network A network that captures relationships between genes that participate in similar biological processes for a particular tissue or cell type wwwnaturecomnrg Reviews 0123456789 Potassium ion transport Neuron development Differentially regulated genes Genes mutated in subjects Population geneticsassociated disease genes Machine learning predicted disease genes Literaturecurated gene set Disease gene list 25825 genes ranked based on autism association using weighted machine learning framework c d a b Top 10 of autismassociated genes used as input gene list for module detection Gene list rank Actin cytoskeleton ADP metabolic process mRNA splicing Neurogenesis Receptor tyrosine kinase signalling Gene of unknown function implicated in biological process by module membership Ion transport and drug metabolism Gproteincoupled receptor pathway Histone modification Chromatin remodelling Gene 1 Gene 2 Gene 3 Gene 4 Gene 5 Gene 6 Gene 7 Gene n 0 5 10 15 log10P 1 2 3 4 5 6 7 8 9 10 12 14 16 18 22 Chromosome Fig 4 systemslevel functional interpretation of disease molecular architecture Functional module discovery identifies groups of genes that are tightly connected in a molecular network and likely to participate in similar biological processes This can distil a large group of genes into biologically coherent sets and can be used to elucidate disease specific processes and infer the function of unknown genes a Gene lists for module clustering may be any group of genes of interest and may be drawn from literature curation sets of genes mutated in subjects with a specific disease genes differentially regulated in a condition of interest genes computationally implicated in a disease for example based on machine learning predictions or genes associated with a disease in population genetics based studies b The molecular network is first filtered by the input gene list red nodes and then partitioned into modules using a community clustering algorithm based on the connectivity of the genes in a functional network The discovered modules can be biologically characterized based on Gene Ontology enrichment of member genes c In an application of module clustering to autism biology 25825 genes were ranked using a weighted machine learning framework based on their association with autism and the top 10 of genes were used for module clustering d Functional modules of 2500 predicted autism associated genes were identified using a brain specific functional network177 The resulting modules captured distinct aspects of autism biology Each node represents a gene and the size of the node reflects connectivity of the gene that is the average edge weight of edges to the gene in the network Edges not shown for clarity The function of unknown genes can be inferred based on their module context An intuitive interface for researchers to perform functional module discovery is available at HumanBase hbflatironinstituteorgmodule Nature reviews Genetics Reviews 0123456789 functional network identifies modules enriched in syn aptic transmission and plasticity sensory perception and chromatin remodelling Fig 4cd More recently there have been community led efforts to benchmark disease module discovery methods on various network configurations184 The prospects of these methods that predict disease association are strongly dependent on the quality and coverage of available molecular networks Importantly integrated network models can accommo date the growing breadth and scale of new omics data for example metabolomics lipidomics and proteom ics and sequencing in more contexts single cell types within tissues and across diseases Conclusions and perspectives Increasingly rapid and inexpensive sequencing technol ogies are making it possible to initiate projects aiming to sequence ever larger numbers of participants Coupled with rich phenotypic information and improving com putational approaches this wealth of genetic informa tion will lead to dramatic advances in the understanding of the genetic architecture of complex diseases and traits The goal of human genetics is linking genotype to phenotype and this is especially challenging for vari ants in the non coding genome An emerging class of approaches use deep neural networks trained on large compendia of transcription factor binding and chro matin modification data to predict the biochemical impact of non coding variants Such predictions which in effect function as in silico ChIPseq assays can be further linked to properties such as gene expression Crucially sequence based deep learning methods can predict the impact of rare or never seen variants that may be discovered with more sequencing providing information that cannot be readily learned from pheno typic association studies of large genomic databases Improvements in deep learning methods including the development of more interpretable models as well as increasing the quality and coverage of training epigenetic data are likely to result in further advances in the under standing of the non coding genome Broader availabil ity of WGS studies especially with careful controls and improved representation of more diverse populations are also critical to advancement of this field Another key direction for understanding human dis ease processes involves developing network and pathway models in disease relevant biological contexts includ ing specific cell types155157 developmental stages and environ mental conditions Identifying genes modules and pathways most relevant to specific diseases as well as elucidating how network perturbations and dysre gulations contribute to disease represents an important direction for study The accuracy and disease relevance of such models can be improved by accounting for tissue differences in network wiring which can include key dif ferences for understanding disease processes in specific cell types and organ systems Large scale cross tissue data sets for training such models are increasingly availa ble through consortium efforts5133140185187 but additional data especially at single cell resolution will be highly valuable for improving the resolution and accuracy of these models The availability of accurate context specific models of molecular architecture of disease will in turn be key to interpreting large numbers of variants affecting dif ferent genes and processes An understanding of how to combine each variants direct biochemical impact and its effect on downstream regulatory interaction and functional networks will be critical to quantifying its con tribution to phenotypic alterations and disease pro perties across diverse populations Such an understanding can be facilitated by the ability to deeply profile indi viduals at the molecular level including their genomic sequence cell type specific expression patterns and epi genetic profiles and longitudinal phenotypic informa tion Advances in understanding variant impact taken together with multimodal molecular and phenotypic data will bring the field closer to broad implementation of precision medicine in which genetically or molecu larly targeted therapeutic approaches can be designed in a way that is responsive to the disease processes and drug sensitivities in specific patients Published online xx xx xxxx 1 1000 Genomes Project Consortium et al A global reference for human genetic variation Nature 526 6874 2015 2 Bycroft C et al The UK Biobank resource with deep phenotyping and genomic data Nature 562 203209 2018 3 ICGCTCGA Pan Cancer Analysis of Whole Genomes Consortium Pan cancer analysis of whole genomes Nature 578 8293 2020 4 ENCODE Project Consortium An integrated encyclopedia of DNA elements in the human genome Nature 489 5774 2012 5 Bernstein B E et al The NIH Roadmap Epigenomics Mapping Consortium Nat Biotechnol 28 10451048 2010 6 Siepel A et al Evolutionarily conserved elements in vertebrate insect worm and yeast genomes Genome Res 15 10341050 2005 This paper uses the PhastCons method to give per base estimates of negative selection within conserved elements using multiple sequence alignments and hidden Markov models 7 Davydov E V et al Identifying a high fraction of the human genome to be under selective constraint using GERP PLoS Comput Biol 6 e1001025 2010 This paper uses the pathogenicity GERP method to give nucleotide and element level constraint scores from profiling substitution rates in multiple sequence alignments 8 Gulko B Hubisz M J Gronau I Siepel A A method for calculating probabilities of fitness consequences for point mutations across the human genome Nat Genet 47 276283 2015 9 Ramani R Krumholz K Huang YF Siepel A PhastWeb a web interface for evolutionary conservation scoring of multiple sequence alignments using phastCons and phyloP Bioinformatics 35 23202322 2019 10 Cooper G M et al Distribution and intensity of constraint in mammalian genomic sequence Genome Res 15 901913 2005 11 Pollard K S Hubisz M J Rosenbloom K R Siepel A Detection of nonneutral substitution rates on mammalian phylogenies Genome Res 20 110121 2010 This paper discusses the pathogenicity scoring method PhyloP using multiple sequence alignments and gives the per base P value for conservation acceleration scores per clade that reflect divergence from the neutral rate 12 Baugh E H et al Robust classification of protein variation using structural modelling and large scale data integration Nucleic Acids Res 44 25012513 2016 13 Kobren S N Chazelle B Singh M PertInInt an integrative analytical approach to rapidly uncover cancer driver genes with perturbed interactions and functionalities Cell Syst 11 6374e7 2020 This paper uses PertInInt to assess protein variants for cancer relevance based on predicting the functional impact on physical interactions between proteins and other proteins nucleic acids ions drugs and other small molecules 14 Kobren S N Singh M Systematic domain based aggregation of protein structures highlights DNA RNA and other ligand binding positions Nucleic Acids Res 47 582593 2019 15 Ancien F Pucci F Godfroid M Rooman M Prediction and interpretation of deleterious coding variants in terms of protein structural stability Sci Rep 8 4480 2018 16 Adzhubei I A et al A method and server for predicting damaging missense mutations Nat Methods 7 248249 2010 17 Gerasimavicius L Liu X Marsh J A Identification of pathogenic missense mutations using protein stability predictors Sci Rep 10 15387 2020 18 Rodrigues C H Pires D E Ascher D B DynaMut predicting the impact of mutations on protein conformation flexibility and stability Nucleic Acids Res 46 W350W355 2018 wwwnaturecomnrg Reviews 0123456789 19 Pires D E V Ascher D B Blundell T L DUET a server for predicting effects of mutations on protein stability using an integrated computational approach Nucleic Acids Res 42 W314W319 2014 20 Rodrigues C H M Myung Y Pires D E V Ascher D B mCSM PPI2 predicting the effects of mutations on proteinprotein interactions Nucleic Acids Res 47 W338W344 2019 21 Li M Simonetti F L Goncearenco A Panchenko A R MutaBind estimates and interprets the effects of sequence variants on proteinprotein interactions Nucleic Acids Res 44 W494W501 2016 22 Dehouck Y Kwasigroch J M Rooman M Gilis D BeAtMuSiC prediction of changes in proteinprotein binding affinity on mutations Nucleic Acids Res 41 W333W339 2013 23 Pires D E V Blundell T L Ascher D B mCSM lig quantifying the effects of mutations on proteinsmall molecule affinity in genetic disease and emergence of drug resistance Sci Rep 6 29575 2016 24 Ghersi D Singh M Interaction based discovery of functionally important genes in cancers Nucleic Acids Res 42 e18 2014 25 Zhou J Troyanskaya O G Predicting effects of noncoding variants with deep learning based sequence model Nat Methods 12 931934 2015 This paper presents DeepSEA a multitask deep learning model that trains and predicts cell type specific regulatory factor binding to genomic sequence for 900 features and cell types 26 Zhou J et al Whole genome deep learning analysis identifies contribution of noncoding mutations to autism risk Nat Genet 51 973980 2019 This paper for the first time finds a significant contribution of non coding mutations to complex disease risk by demonstrating higher functional impact of de novo mutations from probands with autism compared with siblings using mutational impacts inferred from deep learning sequence models of transcriptional and post transcriptional effects 27 Kelley D R Snoek J Rinn J L Basset learning the regulatory code of the accessible genome with deep convolutional neural networks Genome Res 26 990999 2016 28 Lee D et al A method to predict the impact of regulatory variants from DNA sequence Nat Genet 47 955961 2015 This paper presents among the first approaches to predict the tissue specific impact of changes in the non coding genome without information on evolution or genome annotations gkm svm implementing an SVM classifier that uses only sequence k mers as input 29 Senior A W et al Improved protein structure prediction using potentials from deep learning Nature 577 706710 2020 30 Esteva A et al Dermatologist level classification of skin cancer with deep neural networks Nature 542 115118 2017 31 Poplin R et al A universal SNP and small indel variant caller using deep neural networks Nat Biotechnol 36 983987 2018 32 Miotto R Li L Kidd B A Dudley J T Deep patient an unsupervised representation to predict the future of patients from the electronic health records Sci Rep 6 26094 2016 33 Zhavoronkov A et al Deep learning enables rapid identification of potent DDR1 kinase inhibitors Nat Biotechnol 37 10381040 2019 34 Ching T et al Opportunities and obstacles for deep learning in biology and medicine J R Soc Interface 15 20170387 2018 35 LeCun Y Bengio Y Hinton G Deep learning Nature 521 436444 2015 36 Avsec Ž et al The Kipoi repository accelerates community exchange and reuse of predictive models for genomics Nat Biotechnol 37 592600 2019 37 Chen K M Cofer E M Zhou J Troyanskaya O G Selene a PyTorch based deep learning library for sequence data Nat Methods 16 315318 2019 38 Shrikumar A Greenside P Kundaje A in ICML17 Proc 34th Int Conf Machine Learning eds Precup D Teh Y W 31453153 PMLR 2017 39 Binder A et al Morphological and molecular breast cancer profiling through explainable machine learning Nat Mach Intell 3 355366 2021 40 Zhang Z Park C Y Theesfeld C L Troyanskaya O G An automated framework for efficiently designing deep convolutional neural networks in genomics Nat Mach Intell 3 392400 2021 41 Eraslan G Avsec Ž Gagneur J Theis F J Deep learning new computational modelling techniques for genomics Nat Rev Genet 20 389403 2019 42 Manolio T A et al Finding the missing heritability of complex diseases Nature 461 747753 2009 43 Wainschtein P et al Recovery of trait heritability from whole genome sequence data Preprint at bioRxiv httpsdoiorg101101588020 2019 44 Wood A R et al Defining the role of common variation in the genomic and biological architecture of adult human height Nat Genet 46 11731186 2014 45 Yang J et al Common SNPs explain a large proportion of the heritability for human height Nat Genet 42 565569 2010 46 Yang J et al Genome partitioning of genetic variation for complex traits using common SNPs Nat Genet 43 519525 2011 47 Boyle E A Li Y I Pritchard J K An expanded view of complex traits from polygenic to omnigenic Cell 169 11771186 2017 48 Liu X Li Y I Pritchard J K Trans effects on gene expression can drive omnigenic inheritance Cell 177 10221034e6 2019 49 Lappalainen T Scott A J Brandt M Hall I M Genomic analysis in the age of human genome sequencing Cell 177 7084 2019 50 Shendure J Findlay G M Snyder M W Genomic medicineprogress pitfalls and promise Cell 177 4557 2019 51 Buniello A et al The NHGRI EBI GWAS Catalog of published genome wide association studies targeted arrays and summary statistics 2019 Nucleic Acids Res 47 D1005D1012 2019 52 Pasaniuc B et al Fast and accurate imputation of summary statistics enhances evidence of functional enrichment Bioinformatics 30 29062914 2014 53 Schurz H et al Evaluating the accuracy of imputation methods in a five way admixed population Front Genet 10 34 2019 54 Schaid D J Chen W Larson N B From genome wide associations to candidate causal variants by statistical fine mapping Nat Rev Genet 19 491504 2018 55 Easton D F et al Gene panel sequencing and the prediction of breast cancer risk N Engl J Med 372 22432257 2015 56 Robbins C M et al Copy number and targeted mutational analysis reveals novel somatic events in metastatic prostate tumors Genome Res 21 4755 2011 57 Meienberg J Bruggmann R Oexle K Matyas G Clinical sequencing is WGS the better WES Hum Genet 135 359362 2016 58 Maurano M T et al Systematic localization of common disease associated variation in regulatory DNA Science 337 11901195 2012 59 French C E et al Whole genome sequencing reveals that genetic conditions are frequent in intensively ill children Intensive Care Med 45 627636 2019 60 Hou YC C et al Precision medicine integrating whole genome sequencing comprehensive metabolomics and advanced imaging Proc Natl Acad Sci USA 117 30533062 2020 61 Cassini T A et al Whole genome sequencing reveals novel IGHMBP2 variant leading to unique cryptic splice site and CharcotMarieTooth phenotype with early onset symptoms Mol Genet Genom Med 7 e00676 2019 62 All of Us Research Program Investigators et al The All of Us research program N Engl J Med 381 668676 2019 63 Karczewski K J et al The mutational constraint spectrum quantified from variation in 141456 humans Nature 581 434443 2020 64 Petrovski S Wang Q Heinzen E L Allen A S Goldstein D B Genic intolerance to functional variation and the interpretation of personal genomes PLoS Genet 9 e1003709 2013 65 Lek M et al Analysis of protein coding genetic variation in 60706 humans Nature 536 285291 2016 66 Havrilla J M Pedersen B S Layer R M Quinlan A R A map of constrained coding regions in the human genome Nat Genet 51 8895 2019 67 Eilbeck K Quinlan A Yandell M Settling the score variant prioritization and Mendelian disease Nat Rev Genet 18 599612 2017 68 Watanabe K et al A global overview of pleiotropy and genetic architecture in complex traits Nat Genet 51 13391348 2019 69 Kircher M et al A general framework for estimating the relative pathogenicity of human genetic variants Nat Genet 46 310315 2014 70 Carter H Douville C Stenson P D Cooper D N Karchin R Identifying Mendelian disease genes with the variant effect scoring tool BMC Genomics 14 S3 2013 71 Shihab H A et al Ranking non synonymous single nucleotide polymorphisms based on disease concepts Hum Genomics 8 11 2014 72 Ioannidis N M et al REVEL an ensemble method for predicting the pathogenicity of rare missense variants Am J Hum Genet 99 877885 2016 73 Sim NL et al SIFT web server predicting effects of amino acid substitutions on proteins Nucleic Acids Res 40 W452W457 2012 74 Ng P C Henikoff S Predicting deleterious amino acid substitutions Genome Res 11 863874 2001 75 Vaser R Adusumalli S Leng S N Sikic M Ng P C SIFT missense predictions for genomes Nat Protoc 11 19 2016 76 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 D886D894 2019 This paper uses the CADD pathogenicity score as an SVM classifier that integrates multiple functional genomic and evolutionary data to predict coding and non coding variant impacts 77 Ionita Laza I McCallum K Xu B Buxbaum J D A spectral approach integrating functional genomic annotations for coding and noncoding variants Nat Genet 48 214220 2016 This paper presents the Eigen pathogenicity score as an unsupervised meta score of non coding variant fitness impact 78 Park J S et al Brain somatic mutations observed in Alzheimers disease associated with aging and dysregulation of tau phosphorylation Nat Commun 10 112 2019 79 Bailey M H et al Comprehensive characterization of cancer driver genes and mutations Cell 173 371385e18 2018 80 Buja A et al Damaging de novo mutations diminish motor skills in children on the autism spectrum Proc Natl Acad Sci USA 115 E1859E1866 2018 81 Wright C F et al Genetic diagnosis of developmental disorders in the DDD study a scalable analysis of genome wide research data Lancet 385 13051314 2015 82 Quang D Xie X DanQ a hybrid convolutional and recurrent deep neural network for quantifying the function of DNA sequences Nucleic Acids Res 44 e107e107 2016 83 Nair S Kim D S Perricone J Kundaje A Integrating regulatory DNA sequence and gene expression to predict genome wide chromatin accessibility across cellular contexts Bioinformatics 35 i108i116 2019 84 Avsec Ž et al Base resolution models of transcription factor binding reveal soft motif syntax Nat Genet 53 354366 2021 This paper describes a novel deep learning framework for functional genomics sequence modelling which combines neural network models with model interpretation tools to discover highresolution motif syntax 85 Kelley D R et al Sequential regulatory activity prediction across chromosomes with convolutional neural networks Genome Res 28 739750 2018 This paper uses Basenji as a CNN sequence model that predicts regulatory factor binding and expression based on cap analysis gene expression CAGE peak data 86 Taliun D et al Sequencing of 53831 diverse genomes from the NHLBI TOPMed Program Nature 590 290299 2021 87 Ghandi M Lee D Mohammad Noori M Beer M A Enhanced regulatory sequence prediction using gapped k mer features PLoS Comput Biol 10 e1003711 2014 88 Alipanahi B Delong A Weirauch M T Frey B J Predicting the sequence specificities of DNA and Nature reviews Genetics Reviews 0123456789 RNA binding proteins by deep learning Nat Biotechnol 33 831838 2015 This paper uses DeepBind as a framework of individual CNNs that train on and predict regulatory factor binding to DNA and RNA 89 Angermueller C Lee H J Reik W Stegle O DeepCpG accurate prediction of single cell DNA methylation states using deep learning Genome Biol 18 67 2017 90 Quang D Xie X FactorNet a deep learning framework for predicting cell type specific transcription factor binding from nucleotide resolution sequential data Methods 166 4047 2019 91 Richter F et al Genomic analyses implicate noncoding de novo variants in congenital heart disease Nat Genet 52 769777 2020 This paper presentes whole genome sequence analysis of genetic aetiology of congenital heart disease wherein HeartENN a deep CNN sequence genomic sequence model is applied to functional impact prediction of de novo non coding mutations and an excess burden of high impact mutations is observed in individuals who are affected compared with controls 92 Qin Q Feng J Imputation for transcription factor binding predictions based on deep learning PLoS Comput Biol 13 e1005403 2017 93 Li H Quang D Guan Y Anchor trans cell type prediction of transcription factor binding sites Genome Res 29 281292 2019 94 Agarwal V Shendure J Predicting mRNA abundance directly from genomic sequence using deep convolutional neural networks Cell Rep 31 107663 2020 95 Zhou J et al Deep learning sequence based ab initio prediction of variant effects on expression and disease risk Nat Genet 50 11711179 2018 This paper presents tissue specific gene expression prediction from sequence using a deep CNN and a linear model and application to derive constraint violation score pathogenicity based on cumulative predicted regulatory impacts in genomic intervals 96 Dey K K et al Integrative approaches to improve the informativeness of deep learning models for human complex diseases Preprint at bioRxiv httpsdoi org10110120200908288563 2020 97 Law A J Kleinman J E Weinberger D R Weickert C S Disease associated intronic variants in the ErbB4 gene are related to altered ErbB4 splice variant expression in the brain in schizophrenia Hum Mol Genet 16 129141 2006 98 Sangermano R et al ABCA4 midigenes reveal the full splice spectrum of all reported noncanonical splice site variants in Stargardt disease Genome Res 28 100110 2018 99 de Jong V M et al Post transcriptional control of candidate risk genes for type 1 diabetes by rare genetic variants Genes Immun 14 5861 2012 100 Cardo L F et al A Search for SNCA 3 UTR variants Identified SNP rs356165 as a determinant of disease risk and onset age in Parkinsons disease J Mol Neurosci 47 425430 2011 101 Zuallaert J et al SpliceRover interpretable convolutional neural networks for improved splice site prediction Bioinformatics 34 41804188 2018 102 Louadi Z Oubounyt M Tayara H Chong K T Deep splicing code classifying alternative splicing events using deep learning Genes 10 587 2019 103 Leung M K K Xiong H Y Lee L J Frey B J Deep learning of the tissue regulated splicing code Bioinformatics 30 i121i129 2014 104 Zhang Y Liu X MacLeod J Liu J Discerning novel splice junctions derived from RNA seq alignment a deep learning approach BMC Genomics 19 971 2018 105 Zeng Z Bromberg Y Predicting functional effects of synonymous variants a systematic review and perspectives Front Genet 10 914 2019 106 Barash Y et al Deciphering the splicing code Nature 465 5359 2010 107 Paggi J M Bejerano G A sequence based deep learning model accurately predicts RNA splicing branchpoints RNA 24 16471658 2018 108 Li Y I et al Annotation free quantification of RNA splicing using LeafCutter Nat Genet 50 151158 2018 109 Jaganathan K et al Predicting splicing from primary sequence with deep learning Cell 176 535548e24 2019 110 Cummings B B et al Improving genetic diagnosis in Mendelian disease with transcriptome sequencing Sci Transl Med 9 eaal5209 2017 111 Ray T A et al Comprehensive identification of mRNA isoforms reveals the diversity of neural cell surface molecules with roles in retinal development and disease Nat Commun 11 3328 2020 112 Lagarde J et al High throughput annotation of full length long noncoding RNAs with capture long read sequencing Nat Genet 49 17311740 2017 113 Gupta I et al Single cell isoform RNA sequencing characterizes isoforms in thousands of cerebellar cells Nat Biotechnol 36 11971202 2018 114 Hardwick S A Joglekar A Flicek P Frankish A Tilgner H U Getting the entire message progress in isoform sequencing Front Genet 10 709 2019 115 Pan X Shen HB RNA protein binding motifs mining with a new hybrid deep learning based cross domain knowledge integration approach BMC Bioinforma 18 114 2017 116 Pan X Rijnbeek P Yan J Shen HB Prediction of RNA protein sequence and structure binding preferences using deep convolutional and recurrent neural networks BMC Genomics 19 111 2018 117 Pan X Shen HB Predicting RNAprotein binding sites and motifs through combining local and global deep convolutional neural networks Bioinformatics 34 34273436 2018 118 Yu H Wang J Sheng Q Liu Q Shyr Y beRBP binding estimation for human RNA binding proteins Nucleic Acids Res 47 e26e26 2018 119 Zhang Z et al Deep learning augmented RNA seq analysis of transcript splicing Nat Methods 16 307310 2019 120 Wen M Cong P Zhang Z Lu H Li T DeepMirTar a deep learning approach for predicting human miRNA targets Bioinformatics 34 37813787 2018 121 Kang Q Meng J Cui J Luan Y Chen M PmliPred a method based on hybrid model and fuzzy decision for plant miRNAlncRNA interaction prediction Bioinformatics 36 29862992 2020 122 Park C Y et al Genome wide landscape of RNA binding protein target site dysregulation reveals a major impact on psychiatric disorder risk Nat Genet 53 166173 2021 123 Landrum M J et al ClinVar public archive of relationships among sequence variation and human phenotype Nucleic Acids Res 42 D980D985 2014 124 Landrum M J et al ClinVar improving access to variant interpretations and supporting evidence Nucleic Acids Res 46 D1062D1067 2018 125 Rehm H L et al ClinGenthe clinical genome resource N Engl J Med 372 22352242 2015 126 Harrison S M et al Clinical laboratories collaborate to resolve differences in variant interpretations submitted to ClinVar Genet Med 19 10961104 2017 127 Stenson P D et al The human gene mutation database HGMD optimizing its use in a clinical diagnostic or research setting Hum Genet 139 11971207 2020 128 Kalia S S et al Recommendations for reporting of secondary findings in clinical exome and genome sequencing 2016 update ACMG SF v20 a policy statement of the American College of Medical Genetics and Genomics Genet Med 19 249255 2017 129 Esposito D et al MaveDB an open source platform to distribute and interpret data from multiplexed assays of variant effect Genome Biol 20 223 2019 130 Oughtred R et al The BioGRID database a comprehensive biomedical resource of curated protein genetic and chemical interactions Protein Sci 30 187200 2021 131 Gelman H et al Recommendations for the collection and use of multiplexed functional data for clinical variant interpretation Genome Med 11 85 2019 132 Quang D Chen Y Xie X DANN a deep learning approach for annotating the pathogenicity of genetic variants Bioinformatics 31 761763 2015 This paper presents a pathogenicity scoring method which is a deep learning CNN version of CADD for coding and non coding variant fitness impact 133 Davis C A et al The Encyclopedia of DNA Elements ENCODE data portal update Nucleic Acids Res 46 D794D801 2018 134 Gronau I Arbiza L Mohammed J Siepel A Inference of natural selection from interspersed genomic elements based on polymorphism and divergence Mol Biol Evol 30 11591171 2013 135 Huang YF Gulko B Siepel A Fast scalable prediction of deleterious noncoding variants from functional and population genomic data Nat Genet 49 618624 2017 This paper uses a pathogenicity scoring method LINSIGHT to predict fitness consequences of non coding human variation using linear modelling of functional genomic data with a probabilistic model of molecular evolution 136 Ernst C et al Performance of in silico prediction tools for the classification of rare BRCA12 missense variants in clinical diagnostics BMC Med Genomics 11 35 2018 137 Hart S N et al Comprehensive annotation of BRCA1 and BRCA2 missense variants by functionally validated sequence based computational prediction models Genet Med 21 7180 2019 138 Findlay G M et al Accurate classification of BRCA1 variants with saturation genome editing Nature 562 217222 2018 139 Kim S S et al Improving the informativeness of Mendelian disease derived pathogenicity scores for common disease Nat Commun 11 6258 2020 140 GTEx Consortium The GTEx Consortium atlas of genetic regulatory effects across human tissues Science 369 13181330 2020 141 Consortium G The Genotype Tissue Expression GTEx project Nat Genet 45 580585 2013 142 Deutsch E W et al The ProteomeXchange consortium in 2017 supporting the cultural change in proteomics public data deposition Nucleic Acids Res 45 D1100D1106 2017 143 Schwenk J M et al The human plasma proteome draft of 2017 building on the human plasma peptideatlas from mass spectrometry and complementary assays J Proteome Res 16 42994310 2017 144 Huttenhower C et al Exploring the human genome with functional maps Genome Res 19 10931106 2009 145 Troyanskaya O G Dolinski K Owen A B Altman R B Botstein D A Bayesian framework for combining heterogeneous data sources for gene function prediction in Saccharomyces cerevisiae Proc Natl Acad Sci USA 100 83488353 2003 146 Mostafavi S Ray D Warde Farley D Grouios C Morris Q GeneMANIA a real time multiple association network integration algorithm for predicting gene function Genome Biol 9 S4 2008 147 Snel B Lehmann G Bork P Huynen M A STRING a web server to retrieve and display the repeatedly occurring neighbourhood of a gene Nucleic Acids Res 28 34423444 2000 148 Greene C S et al Understanding multicellular function and disease with human tissue specific networks Nat Genet 47 569576 2015 149 Wong A K Krishnan A Troyanskaya O G GIANT 20 genome scale integrated analysis of gene networks in tissues Nucleic Acids Res 46 W65W70 2018 150 Pierson E et al Sharing and specificity of co expression networks across 35 human tissues PLoS Comput Biol 11 e1004220 2015 151 Keller M P et al A gene expression network model of type 2 diabetes links cell cycle regulation in islets with diabetes susceptibility Genome Res 18 706716 2008 152 Dobrin R et al Multi tissue coexpression networks reveal unexpected subnetworks associated with disease Genome Biol 10 R55 2009 153 Magger O Waldman Y Y Ruppin E Sharan R Enhancing the prioritization of disease causing genes through tissue specific protein interaction networks PLoS Comput Biol 8 e1002690 2012 154 Yao V et al An integrative tissue network approach to identify and test human disease genes Nat Biotechnol 36 10911099 2018 155 Roussarie JP et al Selective neuronal vulnerability in Alzheimers disease a network based analysis Neuron 107 821835e12 2020 156 Goya J et al FNTM a server for predicting functional networks of tissues in mouse Nucleic Acids Res 43 W182W187 2015 157 Ledo J H et al Lack of a site specific phosphorylation of Presenilin 1 disrupts microglial gene networks and progenitors during development PLoS ONE 15 e0237773 2020 158 Huang J K et al Systematic evaluation of molecular networks for discovery of disease genes Cell Syst 6 484495e5 2018 159 Kamburov A Wierling C Lehrach H Herwig R ConsensusPathDBa database for integrating human functional interaction networks Nucleic Acids Res 37 D623D628 2009 wwwnaturecomnrg Reviews 0123456789 160 Califano A Butte A J Friend S Ideker T Schadt E Leveraging models of cell regulation and GWAS data in integrative network based association studies Nat Genet 44 841847 2012 161 Schaefer R J et al Integrating coexpression networks with GWAS to prioritize causal genes in maize Plant Cell 30 29222942 2018 162 Novarino G et al Exome sequencing links corticospinal motor neuron disease to common neurodegenerative disorders Science 343 506511 2014 163 Leiserson M D M et al Pan cancer network analysis identifies combinations of rare somatic mutations across pathways and protein complexes Nat Genet 47 106114 2015 164 Ruffalo M Koyutürk M Sharan R Network based integration of disparate omic data to identify silent players in cancer PLoS Comput Biol 11 e1004595 2015 165 Vandin F Upfal E Raphael B J Algorithms for detecting significantly mutated pathways in cancer J Comput Biol 18 507522 2011 166 Reyna M A Leiserson M D M Raphael B J Hierarchical HotNet identifying hierarchies of altered subnetworks Bioinformatics 34 i972i980 2018 167 Creixell P et al Kinome wide decoding of network attacking mutations rewiring cancer signaling Cell 163 202217 2015 168 Creixell P et al Pathway and network analysis of cancer genomes Nat Methods 12 615621 2015 169 Horn H et al NetSig network based discovery from cancer genomes Nat Methods 15 6166 2018 170 Vogelstein B et al Cancer genome landscapes Science 339 15461558 2013 171 Vanunu O Magger O Ruppin E Shlomi T Sharan R Associating genes and protein complexes with disease via network propagation PLoS Comput Biol 6 e1000641 2010 172 Aerts S et al Gene prioritization through genomic data fusion Nat Biotechnol 24 537544 2006 173 Köhler S Bauer S Horn D Robinson P N Walking the interactome for prioritization of candidate disease genes Am J Hum Genet 82 949958 2008 174 Barabási AL Gulbahce N Loscalzo J Network medicine a network based approach to human disease Nat Rev Genet 12 5668 2011 175 Lage K et al A large scale analysis of tissue specific pathology and gene expression of human disease genes and complexes Proc Natl Acad Sci USA 105 2087020875 2008 176 Winter E E Goodstadt L Ponting C P Elevated rates of protein secretion evolution and disease among tissue specific genes Genome Res 14 5461 2004 177 Krishnan A et al Genome wide prediction and functional characterization of the genetic basis of autism spectrum disorder Nat Neurosci 19 14541462 2016 178 Chikina M D Troyanskaya O G Accurate quantification of functional analogy among close homologs PLoS Comput Biol 7 e1001074 2011 179 Guan Y Ackert Bicknell C L Kell B Troyanskaya O G Hibbs M A Functional genomics complements quantitative genetics in identifying disease gene associations PLoS Comput Biol 6 e1000991 2010 180 Swarup V et al Identification of evolutionarily conserved gene networks mediating neurodegenerative dementia Nat Med 25 152164 2019 181 Zhang B Horvath S A general framework for weighted gene co expression network analysis Stat Appl Genet Mol Biol 4 7 2005 182 Parikshak N N Gandal M J Geschwind D H Systems biology and gene networks in neurodevelopmental and neurodegenerative disorders Nat Rev Genet 16 441458 2015 183 Shannon P et al Cytoscape a software environment for integrated models of biomolecular interaction networks Genome Res 13 24982504 2003 184 Choobdar S et al Assessment of network module identification across complex diseases Nat Methods 16 843852 2019 185 Lizio M et al Update of the FANTOM web resource expansion to provide additional transcriptome atlases Nucleic Acids Res 47 D752D758 2019 186 Regev A et al The human cell atlas eLife 6 e27041 2017 187 Lindsay S J et al HDBR expression a unique resource for global and individual gene expression studies during early human brain development Front Neuroanat 10 86 2016 188 Zhang Y et al Discerning novel splice junctions derived from RNA seq alignment a deep learning approach BMC Genomics 19 971 2018 189 Mishra A Macgregor S VEGAS2 software for more flexible gene based testing Twin Res Hum Genet 18 8691 2015 Acknowledgements This work is supported by National Institutes of Health NIH National Institute of Diabetes and Digestive and Kidney Diseases NIDDK grants U24DK100845 UGDK114907 and U2CDK114886 and NIH grant UH3TR002158 to OGT The authors thank C Park for helpful discussions Author contributions The authors contributed to all aspects of the article Competing interests The authors declare no competing interests Publishers note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations Springer Nature Limited 2021 Nature reviews Genetics Reviews Resumo do artigo Decoding disease from genomes to networks to phenotypes Segue abaixo os principais pontos elencados em números 1 O título do mesmo traduzido para o português é Decodificação de doença de genomas a redes e fenótipos do ano de 2021 2 Sobre o resumo O mesmo aborda que o artigo busca interpretar os efeitos das variantes genéticas sendo essa a chave para entender suscetibilidade à doença e concepção de abordagens terapêuticas personalizadas Modernas tecnologias experimentais estão permitindo a geração de maior entendimento do genoma humano Modelos computacionais integrativos podem aproveitar dados para entender o impacto da variante elucidar o efeito de genes desregulados sobre vias biológicas em contextos específicos de doenças e tecidos e interpretar o risco de doenças além sendo considerado como viável apenas com experimentos 3 O que o trabalho discute Os desenvolvimentos recentes em algoritmos de aprendizado de máquina para interpretação de genoma e para integração de nível molecular modelagem de células tecidos e órgãos relevantes para a doença 4 Ponto específico Destacar os existentes métodos e principais desafios e oportunidades na identificação de doenças genéticas específicas variantes e ligá los a vias moleculares e em última análise aos fenótipos da doença 5 Ponto chave sobre os avanços Os avanços experimentais alimentaram o desenvolvimento de algoritmos de aprendizado de máquina que podem prever o impacto funcional de qualquer variante de DNA Houve progresso sobretudo na previsão de impacto molecular de qualquer aminoácido que fizesse substituição na função do gene 6 Importante Recentemente as abordagens computacionais abriram as portas para a interpretação dos efeitos moleculares e fenotípicos da regulação de mutações incluindo aquelas nunca antes observadas em populações 7 Explicação resumida da figura 1 que é basicamente o mais importante para o entendimento de todo o artigo A maioria das variantes associadas à doença está fora éxons na porção não codificante do genoma humano que compõe 99 de DNA e não há uma maneira uniforme de decodificar seus impactos Recentes desenvolvimentos em modelos de sequência baseados em aprendizado profundo abordam esse desafio modelando a relação entre sequências não codificantes e propriedades que influenciam a regulação de genes como cromatina modificações acessibilidade do DNA e ligação do fator de transcrição TF 8 Abordagens computacionais fundamentalmente diferentes são necessários para elucidar o impacto de variantes não codificantes As estruturas computacionais para avaliar o impacto de alterações genéticas na porção codificadora do genoma apesar de ser uma área de rápido desenvolvimento são comparativamente maduras em relação às abordagens para interpretar as variantes não codificantes 12156973 9 Como seria o código regulatório para prever a variante e efeitos Os modelos são treinados em perfis de todo o genoma de atributos como modificações de cromatina transcrição sítios de ligação de fator e sítios de ligação de RNA proteína Elas podem então prever os perfis das novas sequências de entrada identificar o efeito de variantes de sequência podem prever o efeito de mutações que raramente ou nunca são vistos em bancos de dados de sequência O modelo prevê o impacto específico de uma determinada variante em atributos como a ligação de ligação de RNA específica proteínas RBPs modificações de histonas e transcrição ligação do fator 10 Exemplo Para avaliar como efeitos bioquímicos relacionados ao impacto da doença um modelo de patogenicidade foi treinado em efeitos bioquímicos previstos de mutações regulatórias que contribuem para a doença humana permitindo a previsão de patogenicidade para qualquer variante Essa estrutura foi aplicada para prever o impacto de mutações de novo em probandos com espectro autista transtorno TEA e irmãos não afetados Mutações em probandos tiveram impacto da doença previsto significativamente maior do que mutações em irmãos não afetados 11 Embora os modelos de sequência possam prever os efeitos moleculares de mutações inclusive em tecidos específicos expressão gênica explicando como essas alterações causam surgimento de fenótipos de doença essa requer uma compreensão das vias e processos desregulados Recente trabalho tentando explicar como mesmo variantes comuns com efeitos fracos ainda pode aumentar a suscetibilidade à doença teoriza um modelo onigênico em que qualquer gene expressa em um tecido relevante para a doença pode afetar o núcleo genes de doenças por meio de interações entre genes que funcionam juntos em processos caminhos e redes maiores 12 O que foi concluído Há uma imensa dificuldade em ligar o fenótipo ao genótipo Todavia os métodos de aprendizado profundo baseados em sequência podem prever o impacto de variantes raras ou nunca vistas que podem ser descobertos com mais sequenciamento fornecendo informações que não podem ser prontamente aprendidas a partir de estudos de associação fenotípica de grandes bancos de dados genômicos A disponibilidade de modelos precisos e específicos do contexto da arquitetura molecular da doença será por sua vez a chave para interpretar um grande número de variantes que afetam diferentes genes e processos 13 Ponto chave da conclusão Avanços na compreensão do impacto da variante levados juntamente com multimodais moleculares e fenotípicos aproximará o campo de uma ampla implementação da medicina de precisão na qual abordagens terapêuticas direcionadas geneticamente ou molecularmente podem ser projetadas em uma maneira que seja responsiva aos processos de doença e drogas sensibilidades em pacientes específicos