丁香园AV

Synopsis

DNA inside the small volume of the nucleus of a cell is meticulously packaged into a structure known as chromatin. This packaging is highly regulated and depends on the binding of many proteins that interact to compact the DNA and help it form looped structures.  Predicting how DNA folds inside a cell is currently an unsolved problem. Dense neural networks can be trained to solve for the local folding of chromatin, connecting structure, represented as a measured contact map, to the sequence of measured bound proteins.  We can also train a network to solve the inverse problem: given structural information in the form of a contact map, predict the likely distribution of proteins that generated it. Amazingly, these networks are able to learn physical insights that are informative and intuitive about this complex polymer folding problem.