Methods: We composed a survey of 53 pairs of possibly matching facial photos of diagnosed, undiagnosed, and control patients, asked four geneticists to rate the similarity between each pair, and compared their results to the Patient Matching technology of the Facial Dysmorphology Novel Analysis (FDNA) algorithm. We used the Pearson Correlation coefficient and the area under the curve (AUC) of the ROC curve to assess correlations.

Results: Of the diagnosed pairs, the phenotype comparison performance of the FDNA technology (3.6 on similar pairs, 3.25 on different pairs) and the geneticists (2.5 on similar pairs, 2.25 on different pairs) was comparable, with a slight advantage to the technology. This trend is also seen by AUC: 0.65 for the technology vs. AUC 0.51 for the mean of the clinician’s answers. The overall agreement level between the four geneticists is relatively low (mean correlation 0.52).

Conclusions: The importance of the task and the comparable results is a strong motivator to continue with a second stage of a feasibility study, with more photos, more clinicians, and without the intrinsic biases – the familiarity of geneticists with the patients. Future applications of this technology could complement next-generation sequencing in undiagnosed patients and lead to the discovery of rare novel syndromes

Results: Mean AUC results yielded high values with relatively low STD, showing the best result for RIT1 and the lowest for SOS1. The multiclass challenge is able to assign the case to the correct gene out of five with a mean accuracy result of 61% with low STD (5.9), which is roughly three times bigger than the random chance accuracy of 20%.

Discussion: Facial recognition technology that detects dysmorphic features from 2D photographs holds the promise to assist in deep phenotyping of syndromic patients and automatically associating these clinical findings with disease-causing genes. However, perfect recall is naturally limited by the significant overlap within a clinically and pathogenetically related group of genetic entities.

Co-Authors: Z. Yüksel, Centogene AG, Rostock, Germany; N. Fleischer, FDNA Inc., Boston, MA, USA

Session Type: Poster Presentation only

Description:
Introduction: The use of objective facial analysis using automated facial recognition technology could be considered a form of next-generation phenotyping that complements next-generation sequencing. Initial indications of this have been described during ACMG 2016 by Gripp et al.  In this case study, this technology is implemented through Face2Gene, a novel phenotyping tool that offers gratis to clinicians and combines facial recognition algorithms with clinical feature annotation and anthropometric measurements, enabling the detection of syndrome features from 2D facial photographs. We report here a first-of-its-kind collaboration between the WES sequencing and analysis provided by Centogene AG (Rostock, Germany) and the phenotypic data and analysis provided by Face2Gene (FDNA Inc., Boston, MA), with the aim of finding the diagnosis for a challenging and undiagnosed patient from Australia, selected by a group of expert dysmorphologists from cases posted by clinicians worldwide in the online Unknown Forum.

Case description: The patient is a male child of non-consanguineous parents. He was born via an LSCS at 34 weeks of gestation with birth weight of 1420gm (<10^thpercentile). His head circumference was 27cm (<10^th percentile). He had brachycephaly, low set ears, prominent eyes, depressed nasal bridge, tapering fingers, bilateral clinodactyly, single palmar creases and asymmetric facial movements with decreased movements on left side of his face. He could not pass his initial hearing screening. He had normal genitalia and normal cardiac, chest and neurological examination. His MRI brain and array CGH were normal.

At last examination at 2 years, his weight was 8.5kg cm (<-3.5SD), length was 74.5cm (<-4.2SD) and his head circumference was 46cm (-2SD) for his corrected age. He had his first tooth erupted after 12 months. He has global delay but he has not lost any skills. His recent hearing assessment showed satisfactory hearing for normal development of speech.

Methods: WES trio analysis (CentoXome® Platinum with 100-130x average coverage, > 95% target bp covered >20X and turnaround time of fewer than 15 days) revealed 112.774 variants performed on an Illumina Sequencer (library type). Variant filtration was conducted following different filtering steps and correlated with detected features (in HPO terms) and syndrome matches offered by Face2Gene LABS API as well as case description offered by the clinician. In a second analysis, all available sequences (including the captured intronic regions) for the suggested syndrome matches were analyzed for possible disease-causing variants.

Results: No gene was found with variants of uncertain significance, likely pathogenic or pathogenic significance, and fulfilling the required zygosity criteria for the disease-associated inheritance pattern. At the time of sending this abstract, the deletion/duplication analysis for two candidate genes is being performed. If negative, the follow-up is planned in 1-2 years, to assess if increased knowledge on pathogenic genes, variant annotation, as well as more pronounced clinical and facial features in the patient, will allow a diagnosis. Furthermore, this follow-up is critical considering the ever-improving sequencing technologies as well as Face2Gene’s learning algorithm and increased ability to automatically recognize more syndromes over time.

Conclusions With further training in facial analysis software, the potential to increase diagnostic yield in syndromic cases is clear, especially when clinical examination and variant selection are performed by various physicians/medical geneticists. The inclusion of a phenotypic analysis that is independent of the clinicians’ knowledge and experience in the description of syndrome-specific features is important in the process of variant filtration. Although no definite diagnosis was reached in this case, it is the authors’ belief that this form of next-generation phenotyping is an integral part of the diagnostic quest we embark on with each challenging case.