Learning faces from DNA
People
(Responsible)
Frasca F.
(Collaborator)
External participants
Claes Peter
(Third-party beneficiary)
Vandermeulen Dirk
(Third-party beneficiary)
Abstract
The geometric structure of the human face is a complex biological trait strongly influenced by the subject’s genetics that has been recognized as heritable since ancient times. However, the exact understanding of how genetic variations are expressed in phenotypic variations of the facial geometry is still largely lacking: roughly speaking, we know that a child’s face should resemble those of his or her parents, but we have only a rough idea as to which variations in the parents’ DNA code are responsible for a certain shape of the nose, eyebrows, and zygomatic bones of the child. The problem of recovering the facial structure from DNA is a particularly important instance of genotype to phenotype mapping, one of the Holy Grails of modern genetics. Successfully solving it could lead to a breakthrough in numerous applications, ranging from early diagnosis of aging-associated diseases and personalized medicine to forensics.
Our main goal is to apply a novel class of machine learning methods called geometric deep learning in order to associate geometric facial features to genetic features. More specifically, we will use the recently introduced intrinsic convolutional neural networks on 3D scans of human faces to automatically learn the mapping from facial geometry to genetics features, which will then be reversed to reconstruct the face from DNA. The face-to-DNA mapping will be treated by an encoder-decoder architecture, in which the encoder tries to map the geometric facial features (automatically extracted by the intrinsic convolutional neural network) to a genetic feature vector. The decoder performs the inverse mapping, synthesizing 3D patterns associated with prescribed generic features.
Since the face is the “billboard” of our identity, we believe the project will attract a broad range of interest from the society, media, and industry. Better understanding facial morphology would provide important insights to face modelling applications, as well as biometrics such as face recognition. A better understanding of facial genetics will also help in the delineation and diagnosis of genetic disorders and syndromes. In archaeology, ancient remains containing genetic material would allow us to reconstruct the faces of people who lived ages ago. Last but not least, the reconstruction of the face from DNA found at a crime scene could become a ground-breaking instrument that would allow to solve thousands of criminal investigations and cold cases.