Learning faces from DNA

The problem of recovering the geometric structure of the human face from DNA is an important instance of genotype-to-phenotype mapping, one of the Holy Grails of modern genetics. Successfully solving it leads to a breakthrough in many applications, ranging from personalized medicine to biometrics and forensics. However, the problem is complex involving the joint modelling of known and unknown associated genetic interactions with related facial features. To alleviate many of the difficulties involved, we invert the problem, by learning a mapping from the facial geometry to the genetic features. As the facial structure is the overt, integrated, and measureable endpoint of all relevant genetic features and their interactions, the task is no longer to extrapolate partially given genetic features and their interactions, but to learn relevant facial features to classify onto any set of associated genetic features. With advanced deep learning that intrinsically deals with non-Euclidean geometric data, we can learn face-to-DNA mappings. The space of learning parameters can be restricted by integrating biological knowledge into the network’s architectural design. This forward face-to-DNA mapping is of value on its own leading to face-to-DNA biometrics, e.g. confirm whether the face of a suspect matches with DNA found on a crime scene. Also, an actual face will be reconstructed from DNA by “reversing” the neural network in order to synthesize faces with prescribed genetic features.