EP182: KBG syndrome: Prospective videoconferencing and use of AI-driven facial phenotyping

Abstract

In an ideal world, a facial photograph could be combined with medical records and variant prioritization efforts, after exome or genome sequencing, to more accurately classify new missense and other variants in various genes as likely pathogenic. As a case study in progressing toward this goal, we have studied genetic variants in the gene Ankyrin Repeat Domain 11 (ANKRD11) and deletions in 16q24.3, known to cause KBG syndrome, a rare syndrome associated with craniofacial, intellectual, and neurobehavioral anomalies. It has been diagnosed in at least 200 individuals to date, with potentially many others remaining undiagnosed due to its nonspecific and rare nature. Patients present with skeletal abnormalities including short stature, intellectual disability, and various cardiac, neurological, and endocrine abnormalities. While those with KBG syndrome share similar facial phenotypic traits, the overall features can be non-specific to clinicians and others who are not deeply trained in facial dysmorphology, and the recent development of artificial intelligence facial recognition software can be considered as an adjunct to diagnosis. We created a novel protocol for collection and reporting of data, including prospectively interviewing individuals and their families throughout eight countries via videoconferencing by a single physician, board certified in child and adolescent psychiatry. Interviews took place over the course of eight months and all 33 participants gave IRB approved consent. Interview videos and clinical charts provided by the families were analyzed for human phenotype ontology terms that were input into an open-source database, Human Disease Gene website. Summaries of phenotype information were also used to extract phenotypic features in a beta version of new software in Face2Gene (F2G). Face2Gene (FDNA Inc, USA) and GestaltMatcher, two leading genetic facial recognition algorithms, were used to analyze the facial features of all 33 participants for evidence of KBG syndrome. F2G is designed to compare one image to a constellation of images from many genetic syndromes, whereas GestaltMatcher is able to perform pairwise comparisons of two images. In the latter case, we specifically looked for similarities in facial features within our own cohort. In the end, after we received the blood samples for exome sequencing, we applied the Prioritization of Exome Data by Image Analysis (PEDIA) to integrate phenotypic features, facial images, and exome data for variant prioritization. We have collected a cohort of 33 individuals from 30 families with molecular diagnoses of KBG syndrome. Of these individuals, 28 have truncating (frameshift or nonsense) variants, and five have missense variants. Twenty-seven are de novo variants, three are inherited variants, and one is inherited from a mother exhibiting low-level mosaicism. A vast majority presented with skeletal (94%), neurologic (94%), and gastrointestinal (85%) abnormalities, typical of KBG syndrome. In our cohort, 20% had diagnoses of autism spectrum disorder (ASD) as well; however, the presentation appears to be mild or atypical since several were interactive, social and maintained good eye contact during video conference. Other notable findings included three of seven with a cardiac murmur or arrhythmia and an additional individual reporting pain insensitivity. Overall, six of 33 (18%) report impaired pain and tactile sensation. Neurologic abnormalities including seizures and/or EEG abnormalities were also very common (44%), suggesting that early detection and seizure prophylaxis could be an important point of intervention. Three participants were started on growth hormone with positive results. In terms of facial dysmorphology, a subset of 25 individuals was analyzed thus far, with half of the patients having a triangular face with a broad forehead and pointed chin, common to the disorder, and two additional individuals with microcephaly – an uncommon finding, increasing the total in our cohort to 9%. For Face2Gene, KBG syndrome was ranked as the first and most likely diagnosis for 28% of individuals (n= 7). It was ranked second for 40% (n=10) and third or fourth for 12% (n=3). Overall, 80% (n=20) of patient’s photos analyzed had KBG syndrome ranked in their top five potential diagnoses. For GestaltMatcher, in the gallery of 3,533 images with 816 different disorders and 25 KBG patients, fifteen out of 25 KBG patients had at least one other KBG patient in their top-10 rank, and 21 out of 25 patients had at least one KBG patient in their top-30 rank. For Face2Gene, KBG syndrome was ranked as the first and most likely diagnosis for 28% of individuals (n= 7). Overall, 80% (n=20) of patient’s photos analyzed had KBG syndrome ranked in their top five potential diagnoses. These results suggested that most of the patients described in this analysis share a similar facial phenotype. We demonstrated how to use PEDIA approach as a standard diagnostic workflow to perform variant prioritization that integrates phenotypic features, facial images, and exome data. Facial recognition software can be a good adjunct for diagnosis and allow for more focused variant analysis. Both DeepGestalt and GestaltMatcher have the potential to narrow down a vast range of possible diagnoses from lists of variants that might be generated from exome or genome sequencing. Characterizing and quantifying clinical and facial phenotypes is imperative to the process of ranking possible syndromes during the diagnostic process. There is certainly room to grow when it comes to compiling sufficient quality data to train facial recognition algorithms. This is best done through collaborative data sharing between clinicians, and increased transparency on how facial algorithms work to analyze photos. Therefore, we propose this protocol that enables variant prioritization after sequencing and extensive phenotype data collection, thus empowering the diagnostic process.