Dysmorphology in a Genomic Era

Clinical dysmorphology evolved out of the need to standardize descriptive terminology used to define human variation, primarily in the context of malformations and syndromic disorders. The classic dysmorphology approach of establishing a clear history and, through detailed examination, listing abnormal findings ordered in a sequence ranked by perceived importance and then using these data to aid in establishing a working clinical diagnosis, is tried and tested but effective in only a small subset of recognized syndromes. This science and art of syndrome recognition is gradually being lost as newer technologies both blur the margins between various syndromes and enable syndrome recognition through the application of various computational tools. In this article, we review the utility of deep phenotyping and the application of dysmorphology in a genomic era.

DNA analysis may provide the genotype and confirm a diagnosis, but it cannot define the phenotype the advances made in molecular technology have led to the identification and ongoing discovery of the underlying genetic pathoetiology of many syndromes. The field of molecular dysmorphology grew out of the fusion of knowledge of normal human embryologic development and aberrations in gene signaling pathways giving rise to syndromic disorders. Many of these pathways interact, either directly or indirectly, and thus syndromes previously considered distinct entities now share common or overlapping genetic etiologies, which has further compounded the challenge of clearly defining syndromes. As a result, the era of the gene-x-opathy descriptor is our current reality. A commonly used example of this involves the RAS pathway. RASopathies represent disorders impacting the RAS-MEK-ERK pathway, which includes a diverse group of disorders, including neurofibromatosis type I, Legius, Noonan with and without multiple lentigines, Costello, and cardiofaciocutaneous syndromes, as represented in Fig. 1. Noonan syndrome has the greatest overlap, and certain genes within the pathway may present as either Noonan syndrome or cardiofaciocutaneous syndrome. Neurofibromatosis is associated only with neurofibromin, and Costello syndrome is similarly seen only with pathogenic variants in HRAS. In addition to the genetic heterogeneity of common pathways, disorders that use metabolites in signal transduction can share common pathoetiology through these common factors. Disorders of cholesterol metabolism are a good representation of such syndromic overlap. Aside from cholesterol being a key precursor in the general steroid biosynthesis pathway, it is additionally a key signaling molecule in the Sonic Hedgehog pathway, and as such is implicated in related syndromic disorders that include Smith Lemli Opitz, Gorlin syndrome, Rubinstein Taybi, holoprosencephaly, and Pallister Hall and Greig syndromes.

Human variation is estimated at approximately 0.1%, which equates to roughly 6 coding variants per gene. Sequencing of the human genome clarified that humans have close to 23,000 genes, of which only 5500 have corresponding phenotype data. Given that approximately two-thirds of our genes do not have a clear disease association, it is apparent that we need tools to facilitate and assist in the identification of disorders in undiagnosed and rare diseases. There is still much to learn about our genetic code, let alone how tertiary chromatin structure impacts disease and gene regulation. Even if we could perform genomic sequencing on every patient, in most patients the molecular diagnosis would remain elusive. The current diagnostic rate for exome sequencing is approximately 25%, this increases to approximately 30% with exome trio analysis in which selected relatives are sequenced and the additional data are used to assess allele segregation with the phenotype. It should not be inferred from this that genomic sequencing has limited clinical utility, as clinically actionable decisions are made in most cases undergoing sequencing, either through direct pathophysiological mechanisms in a positive result or through exclusion in negative test outcomes.

Sequencing in isolation of supportive phenotypic data significantly limits the analyst’s ability to undertake a comprehensive evaluation of variants identified through sequencing. Several publications have highlighted the value of supportive phenotype data in focusing the analysis to identify causal variants in genomic sequencing.^{,  ,}  Gripp and colleagues published a unique twist on this concept by using artificial intelligence (AI) to aid variant identification. This automated phenotyping through facial recognition software will be discussed in greater detail.

To standardize phenotype terminology and create a structured hierarchical basis for computational analysis, the Human Phenotype Ontology (HPO) was developed in 2008 and through ongoing international collaboration with the Monarch Initiative and other organizations, several tools have been developed to aid diagnostics in undiagnosed diseases. Using HPO terminology, developmental malformations can be cross referenced across a number of different species, which can in turn help define function in human genes. Conversely, genes identified in undiagnosed human malformations can be evaluated for plausible disease association by comparing phenotypic presentations across various species (Fig. 2).

There are too many analytical tools available to allow for a detailed discussion of them all. It would, however, be useful to outline a few of the more common tools used in clinical practice. In general, these tools help define a differential diagnosis based on HPO terminology derived from a thorough dysmorphic evaluation. The details of how to approach a dysmorphic examination have been previously covered^, and are not detailed here, but it is helpful to conceptualize the etiologic basis of the deformity, as outlined in Fig. 3.

To adequately phenotype a patient, it is expected that a detailed head-to-toe evaluation is undertaken, capturing all major (Table 1) and minor anomalies (Table 2). In addition, a complete review of the medical record is required to understand the historical progression and extract the phenotypic elements provided through medical investigation.

An international initiative to standardize the nosology used in clinical dysmorphology has been adapted to the Internet as an online resource supported by the National Human Genome Research Institute: https://elementsofmorphology.nih.gov/. These terms provide the basis for the HPO definitions, which can then be applied to computational tools such as Phenomizer (http://compbio.charite.de/phenomizer/), a tool that creates ranked lists of diseases from OMIM (Online Mendelian Inheritance in Man; www.omim.org) based on the probability of the HPO terms listed presenting concurrently (Fig. 4). It is important to realize that none of the computational tools provide diagnoses, rather they present ranked lists of possible disorders to consider in the differential diagnosis. The onus for making a diagnosis relies on the critical thinking of the physician.

OMIM can be searched directly using any phenotypic descriptor. This character string is processed through a Boolean search engine of its internal database to define the rank list of probable disorders based on the characters used in the search. The lists generated are highly dependent on the search terms used and are, in general, less accurate than implementing specific HPO terms.

Evolving beyond the reliance of the evaluating physician to accurately define dysmorphic ontological terminology, several groups are working on facial recognition software for syndromic identification. These tools too are not diagnostic, but offer rank-listed possibilities for diagnostic consideration. There are various machine learning tools used (eg, Support Vector Machines, KNearest Neighbors, Deformable Models, Hidden Markov Models) that yield variable results. The class of machine learning architectures called Deep Learning, such as convolutional neural networks, have seen substantial gains in recent years. These methods extract features from training images and perform classification of images, often identifying features, such as image texture, that would not have been chosen by diagnosticians. A classic critique of machine learning methods is that they are black-box; meaning that it can be difficult to understand why it arrived at a particular outcome once it has been trained. New algorithms, such as attention-based networks, allow the possibility of opening the box to understand why the classifier made its decision. These new features provide a robust data set for comparative analysis. One of the more widely implemented tools is the free resource: Face2Gene, which in addition to the facial computational analysis, enables adjunct HPO terminology to be entered to further refine the phenotype. The depiction in Fig. 5 outlines the process implemented by FDNA Inc. to establish a bioinformatic reference for comparative image analysis. The machine learning aspect of these tools ensures that, over time, their accuracy will improve. The knowledge base is continually expanded through “crowdsourcing” the expertise of recognized experts who continue to train the system by entering molecularly confirmed diagnoses. The AI utility of Face2Gene has been termed DeepGestalt, and in a recent publication Gurovich and colleagues have shown that this tool is able to list the correct diagnosis within the top 10 differential 91% of the time. The model also has been shown to be able to differentiate the numerous genotypes causing Noonan syndrome with high accuracy in addition to being able to differentiate disorders within a common pathway (eg, RASopathies). Pascolini and colleagues describe similarities in phenotype in various disorders associated with chromatin remodeling that were identified through DeepGestalt. There is some overlap in the context of the epigenetic modification that occurs at a molecular level, but phenotypic overlap had not been previously appreciated. The next evolution of this technology is to incorporate the data derived from phenotyping and apply that directly to machine learning tools used for variant identification and classification. In a recent publication, Hsieh and colleagues described such an approach to variant prioritization termed PEDIA (prioritization of exome data by image analysis) in which data from DeepGestalt is used within the variant analysis pipeline to improve the likelihood of identifying causal variants in genomic analysis.

The development of these various tools and their application to molecular dysmorphology will change the way that we approach phenotyping as a whole. Computer-assisted image processing can be applied to histology, radiology, and any other data sets that allow for a digital reference to be established. This “next-generation phenotyping” will become the basis for comparative analysis to aid further understanding of genomic data. Where is the limit? Vocal analysis, gait analysis, handwriting, visual tracking; there are many avenues for phenotype expansion and with the power of cloud computing and integration of machine learning, we are limited only by our imagination.

D. Basel is an unpaid member of the FDNA Scientific Advisory Board.