Skip to main content
Premium Trial:

Request an Annual Quote

FDNA to Release New Version of Facial Analysis Software for Commercial Integration in Q1


NEW YORK (GenomeWeb) – FDNA is planning a soft launch for a new iteration of its Face2Gene software, which integrates information from facial anomalies with phenotypic traits and variant information to diagnose genetic disorders, that is intended for integration with commercial bioinformatics software packages and products.

The planned update is set to launch at the end of the first quarter. FDNA CEO Dekel Gelbman said the company is currently working with early adopters to pilot the yet-to-be named product. He declined to name the early adopters but said that the company's target market for the product includes genetic testing providers and software companies. Integrating the software with existing software products at clinical laboratories, for example, will help them include the results of facial analysis in the genetic test results that they offer to clinicians, Gelbman explained.

FDNA is also finalizing the price for licenses to access the new product. Meanwhile, it is updating the standalone Face2Gene platform to include additional terms for describing human phenotypes. In addition, FDNA is prepping functionality for the solution that will make it possible to link data from other types of images used in healthcare, such as magnetic resonance imaging scans and X-rays, with phenotype and variant information.

Earlier this week, the company published a new study in Nature Medicine that describes DeepGestalt, the deep learning-based algorithm that underlies its Facial Dysmorphology Novel Analysis, or FDNA, technology. According to the paper, FDNA researchers and their collaborators trained the algorithm on data gleaned from over 150,000 patients. This includes 17,000 patient images that represent roughly 200 known genetic syndromes, including Down syndrome, fetal alcohol syndrome, and Rett syndrome.

The company claims that its findings indicate its offering can provide significant value to patients. "With this study, we've shown that adding an automated facial analysis framework, such as DeepGestalt, to the clinical workflow can help achieve earlier diagnosis and treatment, and promises an improved quality of life," Karen Gripp, FDNA's chief medical officer and a co-author on the paper, said in a statement.

FDNA Chief Technology Officer Yaron Gurovich added that the study "demonstrates how one can successfully apply state-of-the-art algorithms, such as deep learning, to a challenging field where the available data is small, unbalanced in terms of available patients per condition, and where the need to support a large amount of conditions is great."

The facial analysis technology is strongest when it is combined with whole-genome or whole-exome sequencing results, according to Gelbman. Currently, the Face2Gene platform uses a version of DeepGestalt that can identify more than 300 genetic disorders from facial anomalies. Face2Gene can also analyze phenotypic attributes associated with thousands of disorders, and the company is piloting a computational model for identifying phenotype information in clinical notes.

Currently, the company is pushing its platform as a complement to exome sequencing but anticipates that the technology could eventually be paired with whole-genome sequencing, as costs associated with the technology are dropping. It is partnering with unnamed third-party laboratories and bioinformatics companies to explore how the Face2Gene technology could help focus the search for variants involved in genetic disorders on specific regions or genes that are most likely to have a role in the disease.

Moreover, Face2Gene fills a gap in parts of the world with limited access to genetic testing and sequencing technologies, Gelbman said. At present, the company's software is used across 2,000 clinical sites in around 130 countries, and the firm is actively working in tackling challenges posed by limited access to representative images from patients of different ethnicities that can be used to train algorithms like DeepGestalt. To that end, FDNA has projects in Africa, Eastern Asia, and Latin America, among other geographic areas, that are focused on collecting more images from the different ethnicities to correct imbalances in the current knowledgebase.

One of these projects focused on analyzing data from children in Thailand. In a paper published last September in the American Journal of Medical Genetics that describes the analysis, the company tested its technology on frontal photographs from children diagnosed with Down syndrome, a control group of children without the condition, and children diagnosed with other syndromes. According to the results, the software recognized the 30 Down syndrome cases, ranking the condition in its top 10 list of suggested disorders. It did classify some non-Down syndrome cases as instances of the disorder, however, leading the researchers to note that although Face2Gene is useful in clinical settings, it is not "a replacement for clinicians' knowledge of phenotypes."

Gelbman attributes the success of his company's software to the way that its underlying deep learning algorithms are trained. According to the Nature Medicine paper, DeepGestalt outperformed clinicians in experiments where it was asked to distinguish patients with a target syndrome from other syndromes, and to distinguish between subtypes of Noonan syndrome. In a separate experiment using 502 images, DeepGestalt listed the correct syndrome in its list of the top 10 likely conditions in 90 percent of cases.

As a first step, the algorithms learn to recognize and characterize normal facial morphology as a baseline before moving on to learning to identify the distinct features associated with genetic disorders, Gurovich, the first author on the paper, explained. This training is sufficient for DeepGestalt to identify disorders even if it receives images from only a small number of patients, as may be the case with rare genetic disorders. According to the paper, the smallest sample size for a disorder was around 10 individuals. "[It] is called transferred learning," he said. "We teach on one task and then you transfer this knowledge into the other domain [or] the other task that you need to work on. [It] is an essential tool that we used in order to achieve … this ability to identify a lot of diseases."

In addition to broadening the diversity of its image database, FDNA is continuing to expand the breadth of genetic disorders that its software can identify. It currently has a backlog of over 2,000 syndromes slated for inclusion in Face2Gene. The company gets suggestions for syndromes to include through third-party collaborations with pharmaceutical companies or with patient advocacy groups that are trying to raise the profile of a specific disease. It then works proactively with these groups to gather the necessary samples to train DeepGestalt.

The primary value for pharmaceutical companies comes from potential applications of Face2Gene in the context of patient discovery and clinical trials recruitment. One of the challenges that pharmaceutical companies with products on the market face is finding patients that have the diseases they are interested in and identifying those patients as early as possible, Gelbman said. Along with several unnamed pharma partners, "we are evaluating together … how we can utilize this technology to identify at-risk populations and then combine that with genome sequencing for diagnostic purposes."

In addition, the company envisions its technology being used for cohort analysis as part of its engagement with pharma. As part of the second phase of its Prioritization of Exome Data by Image Analysis (PEDIA) study, a collaboration between FDNA, the Charité hospital in Berlin, Germany, and others, the partners are working backwards from known phenotypes of a defined set of disorders to identify genes and variants associated with these conditions.