Deep Phenotyping for Deep Learning DPDL Progress Report

Deep Phenotyping for Deep Learning (DPDL): Progress Report Tzung-Chien Hsieh Institut für Genomische Statistik und Bioinformatik Universität Bonn July 2018 1

Deep Gestalt • Facial analysis framework proposed by FDNA which utilizes computer vision and deep learning to quantifies similarities to genetic syndromes by training with over 26, 000 patient photos. Schilbach-Rott Syndrome Rank 1

PEDIA approach • Prioritization of Exome Data with Image Analysis Phenomizer P-Score Symptoms Symptom Feature Match Analysis F-Score Support vector machine Photo Pattern Recognition G-Score Gestalt Match PEDIA Score Exome Variant Filter Scoring CADD Variant Diagnosis 3

DPDL a framework enabling PEDIA and much more • case-based • search capabilities for • • disorders genes features mutations • phenotype space exploration 4

Database structure Features HPO term Phenotypic Scores Disorder Score Cases Case ID Name Disorder to Features: Phenotypic Mutation Annotations: Scores: Gene: Human The We annotated store connect similarity Phenotype thethe scores the Ontology of CADD external disorders. score annotation to (HPO) gene in VCF For example: file, database relationship and store such bythe as Intellectual Gestalt CADD Clinvar importing score for scores, the variant of each Disability, from Phenomizer mutation. classification mapping scores and Seizures db. SNP. OMIM. Disorder_to_Gene Disorder Gene Mutation Scores Mutation Gene Score Annotations Mutation Clinvar db. SNP 5

Application (1) – www. dpdl. org • Web-based exome prioritization platform Case submitter Report HPO terms F 2 G LABS Report VCF Sample Phenotype information NGS Facility Life&Brain VCF DPDL PEDIA score Analysis Team 6

Application – www. dpdl. org 7