With recent advances in computer vision, many applications based on artificial intelligence have been developed to facilitate the diagnosis of rare genetic disorders through the analysis of patients’ two-dimensional frontal images. Some of these have been implemented on online platforms with user-friendly interfaces and provide facial analysis services, such as Face2Gene. However, users cannot run the facial analysis processes in house because the training data and the trained models are unavailable. This article therefore provides an introduction, designed for users with programming backgrounds, to the use of the open-source GestaltMatcher approach to run facial analysis in their local environment. The Basic Protocol provides detailed instructions for applying for access to the trained models and then performing facial analysis to obtain a prediction score for each of the 595 genes in the GestaltMatcher Database. The prediction results can then be used to narrow down the search space of disease-causing mutations or further connect with a variant-prioritization pipeline. © 2023 The Authors. Current Protocols published by Wiley Periodicals LLC.

Basic Protocol : Using the open-source GestaltMatcher approach to perform facial analysis

Recently, many next-generation phenotyping (NGP) approaches based on artificial intelligence (AI) have been developed to analyze patients’ two-dimensional portraits to further facilitate the diagnosis of rare genetic disorders with facial dysmorphism (Dingemans et al., 2022; Dudding-Byth et al., 2017; Gurovich et al., 2019; Hsieh et al., 2022; Hustinx et al., 2023; Porras et al., 2021; Sümer et al., 2023). Among them, DeepGestalt (Gurovich et al., 2019) and GestaltMatcher (Hsieh et al., 2022) have demonstrated not only high accuracy in diagnosing patients but also applicability to exome variant prioritization (Hsieh et al., 2019).

Some of these approaches have been deployed online to provide such services, including FaceMatch (Dudding-Byth et al., 2017) and Face2Gene (Gurovich et al., 2019; Hsieh et al., 2022). Over the past few years, Face2Gene has become well known as a platform providing a user-friendly interface for performing DeepGestalt and GestaltMatcher. However, because its training data and trained models are unavailable, it would be difficult for users with a programming background who might want to run the analysis in-house because of consent restrictions to do this, and also difficult to connect it with an in-house variant-prioritization pipeline. Therefore, this article presents a protocol utilizing the open-source GestaltMatcher approach and the patient dataset from the GestaltMatcher Database (GMDB; https://db.gestaltmatcher.org; Lesmann et al., 2023), a Findable, Accessible, Interoperable, and Reusable (FAIR) database, to run facial analysis in house to obtain prediction results.

NOTE : Appropriate informed consent is necessary for obtaining and use of human study material.

NOTE : All protocols involving animals must be reviewed and approved by the appropriate Animal Care and Use Committee and must follow regulations for the care and use of laboratory animals. Appropriate informed consent is necessary for obtaining and use of human study material.

This protocol introduces the use of the open-source GestaltMatcher approach to perform facial analysis for the diagnosis of rare genetic disorders in house—specifically, with GestaltMatcher-Arc (Hustinx et al., 2023), a newer version of GestaltMatcher. GestaltMatcher-Arc is available in the GitHub repository at https://github.com/igsb/GestaltMatcher-Arc/tree/protocol. This procedure, which is suitable for users familiar with programming in Python, makes it possible to run a facial analysis on a local machine using the command-line steps listed below, and can therefore be applied to data from patients who do not consent to data upload to other websites. The process has two main parts: setting up hosting of the GestaltMatcher service using a REST (representational state transfer) API, and then sending the image and receiving results through the REST API. Ultimately, the gestalt score for each gene can be used for variant prioritization (Fig. 1).

Necessary Resources

• Computer running operating system such as Windows, Mac OS, or Linux
• Python version 3.7+
• Docker
• Patient photos with only one face in the photo (both eyes, the nose, and the mouth should be visible)
• Model weights for the facial cropper
• GestaltMatcher models (access control)
• Metadata and encodings of GMDB (access control)

1.Apply for access to the model weights, metadata, images, and encodings to the Association for Gene Diagnostics (AGD) by sending a research proposal and General Data Protection Regulation (GDPR) consent form to Prof. Peter Krawitz (pkrawitz@uni-bonn.de). The proposal should describe how you will use the data in your research. The data access committee of the GMDB (https://db.gestaltmatcher.org/about) will review the application and make a decision within 2 to 3 weeks. Details can be found on the GMDB download page (https://db.gestaltmatcher.org/documents).

2.Install the environment by following the instructions listed in the README of the repository (https://github.com/igsb/GestaltMatcher-Arc/tree/protocol). For example, you can run the following commands to create a virtual environment.

• python3 -m venv env_gm
• source env_gm/Scripts/activate
• pip install -r requirements.txt

3.Prepare the gallery encodings and metadata. Be sure to download the following files and place them in the appropriate directories.

• ● Store GMDB_gallery_encodings_v1.0.3.pkl in a folder named./data/gallery_encodings
• ● Store image_gene_and_syndrome_metadata_v1.0.3.p in a folder named ./data/ folder

4.Prepare the Face cropper model. Be sure to download the model weights (Resnet50_Final.pth) in the following path.

• saved_models\ Resnet50_Final.pth

5.Prepare the three GestaltMatcher models. Please ensure that you have accessed the three pre-trained models provided by GMDB and save them in the following three paths.

• saved_models\s1_glint360k_r50_512d_gmdb__v1.0.3_bs64_size112_channels3_last_model.pth
• saved_models\s2_glint360k_r100_512d_gmdb__v1.0.3_bs128_size112_channels3_last_model.pth
• saved_models\glint360k_r100.onnx

6.Built a docker image for the GestaltMatcher REST API.

• docker build -t gm-api

7.Start GestaltMatcher REST API and listen on port 5000.

• docker run -p 5000:5000 gm-api

8.Analyze the image by sending it via REST API and receive the results. You will obtain the results in the output folder shown in Figure 2.If you run the docker image in another server, please use --url and --port to assign to that server and the port. The default URL and port is localhost:5000.

• python send_image_api.py --case_input demo_images/cdls_demo.png --otuput_dir output

Background Information

Rare genetic disorders affect over 6% of the global population (Taruscio et al., 2018). The challenging path toward achieving a diagnosis for such disorders is often described as a “diagnostic odyssey” (Zurynski et al., 2017). Recently, diagnostic yields have improved significantly with the help of whole-exome sequencing analysis. With this type of analysis, however, how to analyze the massive amounts of variants from the whole-exome sequencing data becomes an important concern. Therefore, it is crucial to utilize all available data to facilitate the prioritization and interpretation of variants. Because ∼30% of genetic disorders have a recognizable facial dysmorphic pattern, called a “facial gestalt,” clinicians can narrow the search space by examining the patient's face. However, this approach is highly reliant on clinicians’ experience. If a clinician has never seen a particular rare disorder, it will be difficult for them to recognize it.

With the recent advances in computer vision, many AI-based applications have begun to be used to facilitate diagnosis through facial image analysis (Dingemans et al., 2022; Dudding-Byth et al., 2017; Gurovich et al., 2019; Hsieh et al., 2022; Hustinx et al., 2023; Porras et al., 2021; Sümer et al., 2023) and to further delineate the phenotypes of novel diseases (Aerden et al., 2023; Asif et al., 2022; Averdunk et al., 2023; Ebstein et al., 2023; Guo et al., 2022; Kampmeier et al., 2022; Knaus et al., 2018; Lyon et al., 2023; Marbach et al., 2019; Oppermann et al., 2023; Pantel et al., 2018).

Beyond using classification approaches to suggest disorders matched to the facial phenotype, NGP can also speed up the diagnostic process by prioritizing exome and genome data (Pengelly et al., 2017). In the past, such analyses have usually been done by integrating Human Phenotype Ontology (HPO) terminology into the prioritization (Köhler et al., 2021; Smedley et al., 2015). The facial phenotypic scores quantified by DeepGestalt (Gurovich et al., 2019) and GestaltMatcher (Hsieh et al., 2022) can further assist the prioritization of variants with facial image analysis (Brand et al., 2022; Forwood et al., 2023; Hsieh et al., 2019).

Although NGP approaches have proven their utility in facilitating the diagnostic process, researchers often question their explainability and transparency. For example, it is impossible to know why a model predicts that a patient has a specific disorder. Moreover, it is not possible to visualize the patient photos used in the training set, which is crucial to determining whether the prediction is meaningful. Therefore, we introduce here open-source resource GestaltMatcher-Arc, the latest version of GestaltMatcher, and describes its use to perform facial analysis. GestaltMatcher-Arc consists of three steps: face cropping, face encoding, and evaluation. Face cropping is a preprocessing step, in which the face is cropped from the original image into 112 × 112 pixels, and which keeps only the face region and rotates it to the correct angle (step 4). The face cropper requires the pretrained model weights “Resnet50_Final.pth” to perform Retina-Face (Deng et al., 2020). The cropped input image is then encoded using three GestaltMatcher models into 12 encodings (step 5). Because the test-time augmentation is performed with a flag for horizontal flipping, a flag for gray scale or color, and encoding with three different models, this process yields (2 × 2 × 3 = 12) 12 encodings per image. For more detail, please see Figure 3, which is reproduced from the original publication introducing GestaltMatcher-Arc (Hustinx et al., 2023). For each encoding, the distance between the test image and 7459 images related to 449 disorders associated with 595 genes in GMDB is encoded, and the 12 distances are averaged in each case. Ultimately, the averaged distance is used to sort the 7459 images in ascending order. To simplify the process, we have pre-encoded the images into an encoding file (GMDB_gallery_encodings_v1.0.3.pkl), and image_gene_and_syndrome_metadata_v1.0.3.p stores the image_id, subject_id, gene_id, and omim_id information for the mapping.

The distance to the gene or disorder can be considered as the gestalt score. Hence, you can obtain a gestalt score by running this analysis in house, and further integrate it with your variants prioritization pipeline.

Critical Parameters

The quality of the image will influence the results. The quality includes the resolution, angle, lighting, and number of faces in the image.

Resolution

The input image will be cropped and aligned. Only the face will be preserved and rescaled to 112 × 112 pixels. Therefore, the analysis does not require a high-resolution image, but please be careful when you use an image in which the face is too small.

Angle

The angle of the face is very critical. Both eyes, nose, and mouth should be visible. Although having a perfect frontal image is unnecessary, we recommend having the face at as small an angle as possible.

Lightening

Please try to avoid having varying lighting on the face, such as the left part being bright while the right part is dark because the light comes from the left-hand side.

Number of faces in the image

The analysis expects only one face to be shown in the image. Therefore, please be sure to crop images to show only the patient's face when using a photograph containing multiple faces, such as a family photograph.

Troubleshooting

Common problems and their solutions are listed in Table 1.

Understanding Results

To integrate GestaltMatcher into your analysis pipeline, you can use the Basic Protocol to run GestaltMatcher on your local machine. The strategy of GestaltMatcher is to encode each image into a high-dimensional encoding, and the similarity between two images can be quantified by the cosine distance of two encodings. The GestaltMatcher-Arc (Hustinx et al., 2023) resource presented in this protocol is a new version using model ensemble and test-time augmentation, which provides higher prediction accuracy. To make a prediction, we look for similar images in an image collection (gallery) by sorting the distance in ascending order because two images are more similar when the distance is closer.

Therefore, you can first use the GestaltMatcher repository to run preprocessing and further calculate the similarities between your patient and the 7459 images related to 449 different disorders associated with 595 genes in the GMDB. Ultimately, you will get the predicted results in aggregated gene level and aggregated disorder level form because sometimes one disorder or patient is represented by multiple images. For the aggregation, we take the image with the closest distance of the disorder or the gene. An example of the result is shown in Figure 2. The predicted gene list can be found in the “suggested_genes_list”, and the predicted syndrome list can be found in the “suggested_syndromes_list”. We focus only on the gene list because we want to prioritize the gene variant, so we only look at the suggested_genes_list here. Each entry in the list is a gene with the gestalt_score, and the list is sorted by the gestalt_score in ascending order. The gestalt_score is the same as the distance. Therefore, a smaller distance (gestalt_score) indicates a higher similarity because it means that the testing image is close to the patient of this gene in the phenotype space. You can then include the gestalt score in your variant prioritization pipeline, as in the example in Figure 1, or integrate the gestalt score with other scores using machine learning methods such as support vector machine (Hsieh et al., 2019).

Moreover, the image_id and subject_id tell us which is the closest image and patient associated with NIPBL in the phenotype space. For example, in the example illustrated in Figures 1 and 2, the first entry is NIPBL , with the gestalt_score 0.44. The closest image is image_id 4883, and the patient is subject_id 3546. Therefore, you can further check the information on the patient to see more details by going to the patient page in GMDB. For example, starting with the subject_id 3546, you can visit the patient page, https://db.gestaltmatcher.org/patients/3546, to visualize the patient images that may aid the interpretation of the results.

Time Considerations

The running time of Basic Protocol can be divided into multiple tasks. When starting the REST API, loading the models and gallery takes ∼1 min. In preprocessing, cropping one image takes 1 s and encoding one image takes ∼3 s. During the evaluation, predicting the results takes ∼3 s. Therefore, testing one patient will take ∼7 s of running time.

Acknowledgments

We would like to thank the University Hospital Bonn, and the patients and their families, for their support for this study.

Open access funding enabled and organized by Projekt DEAL.

Author Contributions

Tzung-Chien Hsieh : Methodology; project administration; software; writing—original draft. Hellen Lesmann : Writing—original draft; writing—review and editing. Peter M. Krawitz : Project administration; writing—review and editing.

Conflict of Interest

The authors have no conflict of interest.

Data Availability Statement

The data supporting the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions. To access the model weights, metadata, images, and encodings, the user should send a research proposal and GDPR consent to the Association for Gene Diagnostics (AGD) by emailing Prof. Peter Krawitz (pkrawitz@uni-bonn.de). The data access committee of the GMDB (https://db.gestaltmatcher.org/about) will review the application and make a decision within 2 to 3 weeks. Details can be found on the GMDB download page (https://db.gestaltmatcher.org/documents).

NOTE : The data in GMDB are for non-commercial usage only. For possible commercial usage, please contact Prof. Peter Krawitz directly for more details.

• Aerden, M., Denommé-Pichon, A.-S., Bonneau, D., Bruel, A.-L., Delanne, J., Gérard, B., Mazel, B., Philippe, C., Pinson, L., Prouteau, C., Putoux, A., Tran Mau-Them, F., Viora-Dupont, É., Vitobello, A., Ziegler, A., Piton, A., Isidor, B., Francannet, C., Maillard, P.-Y., … van Esch, H. (2023). The neurodevelopmental and facial phenotype in individuals with a TRIP12 variant. European Journal of Human Genetics: EJHG , 31, 461–468. https://doi.org/10.1038/s41431-023-01307-x
• Asif, M., Kaygusuz, E., Shinawi, M., Nickelsen, A., Hsieh, T.-C., Wagle, P., Budde, B., Hochscherf, J., Abdullah, U., Höning, S., Nienberg, C., Lindenblatt, D., Noegel, A. A., Altmüller, J., Thiele, H., Motameny, S., Fleischer, N., Segal, I., Pais, L., … Hussain, M. S. (2022). De novo variants of CSNK2B cause a new intellectual disability-craniodigital syndrome by disrupting the canonical Wnt signaling pathway. Human Genetics and Genomics Advances , 3, 100111. https://doi.org/10.1016/j.xhgg.2022.100111
• Averdunk, L., Huetzen, M. A., Moreno-Andrés, D., Kalb, R., McKee, S., Hsieh, T.-C., Seibt, A., Schouwink, M., Lalani, S., Faqeih, E. A., Brunet, T., Boor, P., Neveling, K., Hoischen, A., Hildebrandt, B., Graf, E., Lu, L., Jin, W., Schaper, J., … Distelmaier, F. (2023). Biallelic variants in CRIPT cause a Rothmund-Thomson-like syndrome with increased cellular senescence. Genetics in Medicine: Official Journal of the American College of Medical Genetics , 25, 100836. https://doi.org/10.1016/j.gim.2023.100836
• Brand, F., Vijayananth, A., Hsieh, T.-C., Schmidt, A., Peters, S., Mangold, E., Cremer, K., Bender, T., Sivalingam, S., Hundertmark, H., Knaus, A., Engels, H., Krawitz, P. M., & Perne, C. (2022). Next-generation phenotyping contributing to the identification of a 4.7 kb deletion in KANSL1 causing Koolen-de Vries syndrome. Human Mutation , 43(11), 1659–1665. https://doi.org/10.1002/humu.24467
• Deng, J., Guo, J., Ververas, E., Kotsia, I., & Zafeiriou, S. (2020). RetinaFace: Single-Shot Multi-Level Face Localisation in the Wild. In IEEE (Ed.), 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA (pp. 5202–5211). Institute of Electrical and Electronics Engineers. doi: 10.1109/CVPR42600.2020.00525
• Dingemans, A. J. M., Hinne, M., Truijen, K. M. G., Goltstein, L., van Reeuwijk, J., de Leeuw, N., Schuurs-Hoeijmakers, J., Pfundt, R., Diets, I. J., den Hoed, J., de Boer, E., Coenen-van der Spek, J., Jansen, S., van Bon, B. W., Jonis, N., Ockeloen, C., Vulto-van Silfhout, A. T., Kleefstra, T., Koolen, D. A., … de Vries, B. B. A. (2022). PhenoScore: AI-based phenomics to quantify rare disease and genetic variation [Preprint]. Retrieved from medRxiv , 22281480. https://doi.org/10.1101/2022.10.24.22281480
• Dudding-Byth, T., Baxter, A., Holliday, E. G., Hackett, A., O'Donnell, S., White, S. M., Attia, J., Brunner, H., de Vries, B., Koolen, D., Kleefstra, T., Ratwatte, S., Riveros, C., Brain, S., & Lovell, B. C. (2017). Computer face-matching technology using two-dimensional photographs accurately matches the facial gestalt of unrelated individuals with the same syndromic form of intellectual disability. BMC Biotechnology , 17(1), 1–9. https://doi.org/10.1186/s12896-017-0410-1
• Ebstein, F., Küry, S., Most, V., Rosenfelt, C., Scott-Boyer, M.-P., van Woerden, G. M., Besnard, T., Papendorf, J. J., Studencka-Turski, M., Wang, T., Hsieh, T.-C., Golnik, R., Baldridge, D., Forster, C., de Konink, C., Teurlings, S. M. W., Vignard, V., van Jaarsveld, R. H., Ades, L., … Bézieau, S. (2023). PSMC3 proteasome subunit variants are associated with neurodevelopmental delay and type I interferon production. Science Translational Medicine , 15(698), eabo3189. https://doi.org/10.1126/scitranslmed.abo3189
• Forwood, C., Ashton, K., Zhu, Y., Zhang, F., Dias, K.-R., Standen, K., Evans, C.-A., Carey, L., Cardamone, M., Shalhoub, C., Katf, H., Riveros, C., Hsieh, T.-C., Krawitz, P., Robinson, P. N., Dudding-Byth, T., Sadikovic, B., Pinner, J., Buckley, M. F., & Roscioli, T. (2023). Integration of EpiSign, facial phenotyping, and likelihood ratio interpretation of clinical abnormalities in the re-classification of an ARID1B missense variant. American Journal of Medical Genetics. Part C, Seminars in Medical Genetics , https://doi.org/10.1002/ajmg.c.32056
• Guo, L., Park, J., Yi, E., Marchi, E., Hsieh, T.-C., Kibalnyk, Y., Moreno-Sáez, Y., Biskup, S., Puk, O., Beger, C., Li, Q., Wang, K., Voronova, A., Krawitz, P. M., & Lyon, G. J. (2022). KBG syndrome: Videoconferencing and use of artificial intelligence driven facial phenotyping in 25 new patients. European Journal of Human Genetics: EJHG , 30(11), 1244–1254. https://doi.org/10.1038/s41431-022-01171-1
• Gurovich, Y., Hanani, Y., Bar, O., Nadav, G., Fleischer, N., Gelbman, D., Basel-Salmon, L., Krawitz, P. M., Kamphausen, S. B., Zenker, M., Bird, L. M., & Gripp, K. W. (2019). Identifying facial phenotypes of genetic disorders using deep learning. Nature Medicine , 25(1), 60–64. https://doi.org/10.1038/s41591-018-0279-0
• Hsieh, T. C., Mensah, M. A., Pantel, J. T., Aguilar, D., Bar, O., Bayat, A., Becerra-Solano, L., Bentzen, H. B., Biskup, S., Borisov, O., Braaten, O., Ciaccio, C., Coutelier, M., Cremer, K., Danyel, M., Daschkey, S., Eden, H. D., Devriendt, K., Wilson, S., … Krawitz, P. M. (2019). PEDIA: Prioritization of exome data by image analysis. Genetics in Medicine: Official Journal of the American College of Medical Genetics , 21(12), 2807–2814. https://doi.org/10.1038/s41436-019-0566-2
• Hsieh, T.-C., Bar-Haim, A., Moosa, S., Ehmke, N., Gripp, K. W., Pantel, J. T., Danyel, M., Mensah, M. A., Horn, D., Rosnev, S., Fleischer, N., Bonini, G., Hustinx, A., Schmid, A., Knaus, A., Javanmardi, B., Klinkhammer, H., Lesmann, H., Sivalingam, S., … Krawitz, P. M. (2022). GestaltMatcher facilitates rare disease matching using facial phenotype descriptors. Nature Genetics , 54(3), 349–357. https://doi.org/10.1038/s41588-021-01010-x
• Hustinx, A., Hellmann, F., Sumer, O., Javanmardi, B., Andre, E., Krawitz, P., & Hsieh, T.-C. (2023). In IEEE (Ed.), Improving deep facial phenotyping for ultra-rare disorder verification using model ensembles. 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), Waikoloa, HI, USA (pp. 5007–5017). Institute of Electrical and Electronics Engineers. https://doi.org/10.1109/wacv56688.2023.00499
• Kampmeier, A., Leitão, E., Parenti, I., Beygo, J., Depienne, C., Bramswig, N. C., Hsieh, T.-C., Afenjar, A., Beck-Wödl, S., Grasshoff, U., Haack, T. B., Bijlsma, E. K., Ruivenkamp, C., Lausberg, E., Elbracht, M., Haanpää, M. K., Koillinen, H., Heinrich, U., Rost, I., … Kuechler, A. (2022). PHIP-associated Chung-Jansen syndrome: Report of 23 new individuals. Frontiers in Cell and Developmental Biology , 10, 1020609. https://doi.org/10.3389/fcell.2022.1020609
• Knaus, A., Pantel, J. T., Pendziwiat, M., Hajjir, N., Zhao, M., Hsieh, T.-C., Schubach, M., Gurovich, Y., Fleischer, N., Jäger, M., Köhler, S., Muhle, H., Korff, C., Møller, R. S., Bayat, A., Calvas, P., Chassaing, N., Warren, H., Skinner, S., … Krawitz, P. M. (2018). Characterization of glycosylphosphatidylinositol biosynthesis defects by clinical features, flow cytometry, and automated image analysis. Genome Medicine , 10(1), 3. https://doi.org/10.1186/s13073-017-0510-5
• Köhler, S., Gargano, M., Matentzoglu, N., Carmody, L. C., Lewis-Smith, D., Vasilevsky, N. A., Danis, D., Balagura, G., Baynam, G., Brower, A. M., Callahan, T. J., Chute, C. G., Est, J. L., Galer, P. D., Ganesan, S., Griese, M., Haimel, M., Pazmandi, J., Hanauer, M., … Robinson, P. N. (2021). The human phenotype ontology in 2021. Nucleic Acids Research , 49(D1), D1207–D1217. https://doi.org/10.1093/nar/gkaa1043
• Lesmann, H., Lyon, G. J., Caro, P., Abdelrazek, I. M., Moosa, S., Pantel, J. T., ten Hagen, M., Rosnev, S., Kamphans, T., Meiswinkel, W., Li, J.-M., Klinkhammer, H., Hustinx, A., Javanmardi, B., Knaus, A., Uwineza, A., Knopp, C., Marchi, E., Elbracht, M., … Hsieh, T.-C. (2023). GestaltMatcher Database—a FAIR database for medical imaging data of rare disorders [Preptint]. Retrieved from medRxiv , 23290887. https://doi.org/10.1101/2023.06.06.23290887
• Lyon, G. J., Vedaie, M., Besheim, T., Park, A., Marchi, E., Gottlieb, L., Hsieh, T.-C., Klinkhammer, H., Sandomirsky, K., Cheng, H., Starr, L. J., Preddy, I., Tseng, M., Li, Q., Hu, Y., Wang, K., Carvalho, A., Martinez, F., Caro-Llopis, A., … Herr-Israel, E. (2023). Expanding the phenotypic spectrum of NAA10-related neurodevelopmental syndrome and NAA15-related neurodevelopmental syndrome [Preprint]. Retrieved from medRxiv , 22.22279061. https://doi.org/10.1101/2022.08.22.22279061
• Marbach, F., Rustad, C. F., Riess, A., Đukić, D., Hsieh, T. C., Jobani, I., Prescott, T., Bevot, A., Erger, F., Houge, G., Redfors, M., Altmueller, J., Stokowy, T., Gilissen, C., Kubisch, C., Scarano, E., Mazzanti, L., Fiskerstrand, T., Krawitz, P. M., … Netzer, C. (2019). The discovery of a LEMD2-associated nuclear envelopathy with early progeroid appearance suggests advanced applications for AI-driven facial phenotyping. American Journal of Human Genetics , 104(4), 749–757. https://doi.org/10.1016/j.ajhg.2019.02.021
• Oppermann, H., Marcos-Grañeda, E., Weiss, L. A., Gurnett, C. A., Jelsig, A. M., Vineke, S. H., Isidor, B., Mercier, S., Magnussen, K., Zacher, P., Hashim, M., Pagnamenta, A. T., Race, S., Srivastava, S., Frazier, Z., Maiwald, R., Pergande, M., Milani, D., Rinelli, M., … Platzer, K. (2023). CUX1-related neurodevelopmental disorder: Deep insights into phenotype-genotype spectrum and underlying pathology. European Journal of Human Genetics: EJHG , https://doi.org/10.1038/s41431-023-01445-2
• Pantel, J. T., Zhao, M., Mensah, M. A., Hajjir, N., Hsieh, T.-C., Hanani, Y., Fleischer, N., Kamphans, T., Mundlos, S., Gurovich, Y., & Krawitz, P. M. (2018). Advances in computer-assisted syndrome recognition by the example of inborn errors of metabolism. Journal of Inherited Metabolic Disease , 41, 533–539. https://doi.org/10.1007/s10545-018-0174-3
• Pengelly, R. J., Alom, T., Zhang, Z., Hunt, D., Ennis, S., & Collins, A. (2017). Evaluating phenotype-driven approaches for genetic diagnoses from exomes in a clinical setting. Scientific Reports , 7(1), 13509. https://doi.org/10.1038/s41598-017-13841-y
• Porras, A. R., Rosenbaum, K., Tor-Diez, C., Summar, M., & Linguraru, M. G. (2021). Development and evaluation of a machine learning-based point-of-care screening tool for genetic syndromes in children: A multinational retrospective study. The Lancet Digital Health , 3, e635–e643. https://doi.org/10.1016/S2589-7500(21)00137-0
• Smedley, D., Jacobsen, J. O. B., Jäger, M., Köhler, S., Holtgrewe, M., Schubach, M., Siragusa, E., Zemojtel, T., Buske, O. J., Washington, N. L., Bone, W. P., Haendel, M. A., & Robinson, P. N. (2015). Next-generation diagnostics and disease-gene discovery with the Exomiser. Nature Protocols , 10(12), 2004–2015. https://doi.org/10.1038/nprot.2015.124
• Sümer, Ö., Hellmann, F., Hustinx, A., Hsieh, T.-C., & Krawitz, P. (2023). Few-shot meta-learning for recognizing facial phenotypes of genetic disorders. In Caring is sharing—exploiting the value in data for health and innovation (pp. 932–936). IOS Press.
• Taruscio, D., Vittozzi, L., Rocchetti, A., Torreri, P., & Ferrari, L. (2018). The occurrence of 275 rare diseases and 47 rare disease groups in Italy. Results from the national registry of rare diseases. International Journal of Environmental Research and Public Health , 15(7), 1470. https://doi.org/10.3390/ijerph15071470
• Zurynski, Y., Deverell, M., Dalkeith, T., Johnson, S., Christodoulou, J., Leonard, H., Elliott, E. J., & APSU Rare diseases impacts on families study group. (2017). Australian children living with rare diseases: Experiences of diagnosis and perceived consequences of diagnostic delays. Orphanet Journal of Rare Diseases , 12(1), 68. https://doi.org/10.1186/s13023-017-0622-4

• https://facematch.org.au/

FaceMatch.

• https://www.face2gene.com/

Face2Gene.

• https://db.gestaltmatcher.org/

GestaltMatcher Database (GMDB).

Number of times cited according to CrossRef: 3

• Meghna Ahuja Bhasin, Alexej Knaus, Pietro Incardona, Alexander Schmid, Manuel Holtgrewe, Miriam Elbracht, Peter M. Krawitz, Tzung-Chien Hsieh, Enhancing Variant Prioritization in VarFish through On-Premise Computational Facial Analysis, Genes, 10.3390/genes15030370, 15 , 3, (370), (2024).
• Mina Grippa, Claudio Graziano, Landscape of Constitutional SOX4 Variation in Human Disorders, Genes, 10.3390/genes15020158, 15 , 2, (158), (2024).
• Annabelle Arlt, Alexej Knaus, Tzung‐Chien Hsieh, Hannah Klinkhammer, Meghna Ahuja Bhasin, Alexander Hustinx, Shahida Moosa, Peter Krawitz, Ekanem Ekure, Next‐generation phenotyping in Nigerian children with Cornelia de Lange syndrome, American Journal of Medical Genetics Part A, 10.1002/ajmg.a.63641, 194 , 9, (2024).