Our analyses provided an accurate quantitative comparison of facial dysmorphologies in Down, Morquio and Noonan syndromes, as well as in Neurofibromatosis type 1, in a Latin-American population from Colombia. An objective and highly detailed description of the facial phenotype is a major improvement over qualitative descriptions of the complex facial dysmorphologies associated with these genetic disorders. We quantified local facial trait differences presented in people diagnosed with these disorders as compared to age matched controls of the same population, localizing the largest statistically significant facial dysmorphologies.
Our results indicated differential facial patterns associated with each disorder, with major significant dysmorphologies in Down, Morquio and Noonan syndromes, and minor facial dysmorphologies associated with NF1. Different types of genetic alterations, which ranged from aneuploidy and overall genetic imbalance as in DS; to point genetic mutations affecting different processes, such as the metabolism of mucopolysaccharides in MS; or signaling pathways, such as the RAS/MAPK pathway in NS and NF1, significantly affected the facial phenotypes. In genetic and rare diseases, genetic alterations deviate the signaling pathways regulating normal facial development [16, 71], altering normal morphogenesis and growth during pre- and postnatal development [15].
Population-specific Facial Traits In Colombian Individuals With Genetic And Rare Disorders
Overall, the facial patterns associated to each syndrome in the Colombian Latin-American population coincide with the descriptions reported in the literature [52, 61, 63, 64]. However, there are specific local traits that differ, suggesting that facial traits associated to disease might be modulated by population ancestry and influenced by different evolutionary and adaptive histories of human populations [33–35].
In Down syndrome, we found a facial dysmorphologies that are consistent with the patterns reported in the literature for populations of European descent. Our analyses detected differences in linear facial measurements that correspond to typical DS traits such as hypertelorism, maxillary hypoplasia and shorter and wider faces associated to a brachycephalic head [16, 72]. Results also suggested other characteristic traits of DS, such as midfacial retrusion and depressed nasal bridge [52]. Open mouth and macroglossia [54, 55] were also observed during the photographic sessions in the participants of our study. However, in contrast to European and North American populations [70], in the Colombian population we detected that the mouth was wider in individuals diagnosed with DS as compared to euploid controls. This difference may be caused by unnatural facial gestures of the participants when asked to close the mouth during the photo shoot, or by facial differences associated to ancestry. Indeed, Kruszka et al. [33, 34, 35] analyzed individuals diagnosed with DS in diverse populations and showed craniofacial differences between individuals from different ethnicities (Africans, Asians, and Latin Americans), demonstrating that ancestry is a relevant factor when assessing craniofacial variation associated to rare diseases.
In Morquio syndrome, the facial dysmorphologies observed in Colombian individuals were consistent with traits reported in the literature, which included hypertelorism, prognathism, wide nose and wide mouth [57, 61]. In our pediatric cohort from Colombia, Morquio syndrome was associated with the most severe facial dysmorphologies. Considering that keratan and chondroitin sulfate alterations associated with MS accumulate over life and cause irreparable damage to the osteoarticular system, facial dysmorphologies associated with MS are expected to increase with age due to an excessively rapid growth of the head, becoming even more severe in adult individuals [57]. Further research is required to assess this hypothesis and to test whether pharmacological treatments can slow down the progression of the disease and reduce the facial dysmorphologies associated with MS. This is especially relevant in Colombia, which is a country with one of the highest prevalence of MS in the world [56].
In Noonan syndrome, we detected hypertelorism, downward slanting palpebral fissures, and midfacial hypoplasia in the Colombian population, as reported in populations of European descent [63]. In addition, our results quantified relative changes in the position of the mouth in Colombian individuals diagnosed with NS not reported before [73].
The facial pattern associated with NF1 in individuals from Colombia was also compatible with some typical traits of NF1, such as midface hypoplasia [62]. However, our results did not underscore facial asymmetry or hypertelorism as prominent facial differences between diagnosed individuals and controls in the Colombian population [62].
Comparative quantitative studies from different world regions are not usually available for most genetic and rare disorders, and reference data for diagnosis is mainly based on phenotypes defined on populations of European descent. In fact, almost no images of individuals of Latin American origin are included in reference medical texts such as Smith's Recognizable Patterns of Human Malformation [16]. However, previous studies have reported that in Noonan and other genetic syndromes, such as Turner, 22q11.2 and Cornelia de Lange syndromes presented distinctive facial traits that were population specific, with clinical features that were significantly different in Africans, Asians, and Latin Americans [34, 35, 36]. Our results in a Colombian population further support this evidence, highlighting the need to extend the analyses to populations from all over the world to achieve a complete and more accurate phenotypic representation of genetic and RD to optimize the diagnostic potential of facial biomarkers in the clinical practice.
Reduced Accuracy Diagnosis In A Colombian Population With Diverse Ancestry
Deep learning algorithms such as Face2gene have shown potential as a reliable and precise tool for genetic diagnosis by image recognition [9, 26, 74, 75]. In the Colombian cohort analyzed here, Face2gene diagnosed with 100% accuracy Down syndrome, which is one of the most common and more accurately represented genetic disorder. In Noonan syndrome, the percentage of top1-accuracy was lower (66.7%), but still correctly identified the disorder in the majority of individuals when considering the top5-accuracy within Noonan syndrome-like disorders (88.9%). However, in Morquio syndrome, despite being associated with the most severe facial dysmorphologies, the top1-accuracy for exact diagnosis of Mucopolysaccharidosis type IVA was 0% and a low percentage of cases (36.4%) were diagnosed with a Mucopolysaccharidosis-like syndrome as the first prediction.
Interestingly, there was a wide range of variation in gestalt similarity for most disorders, even for Down syndrome. Visual inspection of cases suggested that most individuals associated with the lowest gestalt similarity scores were highly admixed, with diverse contributions of Amerindian, African and European ancestry components. Future analyses need to investigate the role of population ancestry and further assess the reliability and validity of automatic diagnostic tools in admixed populations from non-European descent. As evidenced in this study, where syndromes such as MS, NS, NF1 were associated with low accuracy scores versus more frequent syndromes such as DS, this is critical in rare syndromes with heterogenous clinical presentation and phenotype, where clinical diagnosis is a challenging process [5, 6] that may take several years, leading to the so-called diagnostic odyssey [7].
Precise and early diagnosis of genetic and rare disorders is crucial for adequate health care and clinical management. Without a diagnosis, individuals and their families must proceed without basic information regarding their health and future developmental outcomes [6]. Even though gene-based technologies have greatly improved diagnostic procedures [25], the mutations causing many rare diseases are still not known and access to genetic testing is limited [3]. Genetic consultations may become a long process, and broad molecular testing such as exome and genome sequencing represent a high expense that is not affordable for all families and health care systems, especially in low-medium income countries [7].
In this context, faster, non-invasive and low-cost diagnostic methods based on facial phenotypes emerge as complementary tools for providing earlier first reliable diagnoses [9, 10, 25, 26], especially in low-middle income countries with low budget and resources for molecular testing. Qualitative visual assessment of craniofacial dysmorphology is frequently employed for diagnosis, clinical management and treatment monitoring [16]. Experts in dysmorphologies can identify the facial “gestalt” distinctive of many dysmorphic syndromes [16]. However, this facial assessment relies on the expertise of the clinician and is very challenging, as there is no clear one-to-one correspondence between disorders and facial dysmorphologies. Different genetic mutations can cause the same syndrome or similar phenotypes, whereas the same mutation can induce different phenotypes [12, 76]. In addition, within the same rare disease there may be several subtypes, and symptoms may vary even within individuals of the same genetic disorder and the same family [3]. This complex biology generates confusion at the time of diagnosis and warrants the development of efficient, objective and reliable diagnostic methods.
Computer-assisted phenotyping can overcome these pitfalls and provide widely accessible technologies for quick syndrome screening [6]. In this automated approach, methods can be based on 2D or 3D images [9, 10, 26]. The advantage of 2D methods is that data collection is easy and can be readily translated into the clinical practice, as physicians can take facial images even with simple digital cameras or smartphones. The collection of 3D models is more sophisticated and requires specialized equipment but provides more accurate phenotype descriptions by incorporating the depth dimension.
To further improve the methods of craniofacial assessment to diagnose individuals with genetic syndromes and RD that exhibit facial dysmorphologies, it is crucial to assess the large morphological variation displayed by human populations in facial phenotypes. Factors such as age, sex and ancestry should be accounted for in diagnostic methods. Clinical manifestations in some genetic disorders usually begin at an early age, with two thirds of patients expressing symptoms before the second year of birth [3], but in other disorders facial dysmorphologies develop later, during postnatal development. Male and female faces present sexual dimorphism at adulthood [77] and diseases can differently affect the facial phenotype depending on sex differences [78]. As suggested here, differences in population ancestry can also significantly influence the facial phenotype and the potential for clinical diagnosis.
Therefore, recruitment of participants must be expanded to include as many individuals with RD as possible, together with large comparative samples of age-matched controls, from both sexes and from all the populations in the world representing varied ancestries. For instance, the population in Southwestern Colombia is characterized by high levels of admixture from people with Native American, African, and European ancestry [44, 79]. Including the morphological variation of faces from such different ancestry backgrounds is key to pinpoint the facial dysmorphologies associated with diseases in worldwide diverse populations [80]. Our simulation analyses further highlighted the importance of maximizing the recruitment of diagnosed and control individuals, as results considerably changed depending on the cohort and sample sizes.
Our results underscore that facial phenotypes associated with genetic and rare disorders are influenced by population ancestry [34, 35, 36] and that diverse genetic background variation can modulate the phenotypic response to disease, affecting the accuracy of current tools of clinical diagnosis. In the future, deep learning algorithms including a high variety of populations with different ancestry backgrounds will optimize the precision and accuracy of diagnosis in an unbiased approach. Such predictive models will support clinicians in decision-making across the world.