MANUFACTURE

Biological sequences (DNA, RNA, protein) are strings of amino acids or nucleotides. A central paradigm of biology is that the sequence determines the function of the molecule in the organism. However, this mapping is complex and context dependent. Understanding how widely diverse functions are encoded in these sequences is a fundamental open question in biology. Sequence design can be viewed as the inverse map, from function to sequence: Given a desired phenotypic task, which sequences can perform it? Due to the complexity of the relationship between sequence and function, the design of biological sequences to achieve desired functions is a challenging problem, with numerous applications to drug development, industry, basic experimental research, synthetic biology, and others. Throughout evolution, sequence variation is constrained by the conservation of function. Large sequence datasets can now be collected in high-throughput experiments, rich in evolutionary signals that can shed light on the sequence-to-function map. Machine learning models trained on such datasets can be used to make functional predictions and for generation of novel sequences. In this project, we will develop generative models that encode sequences into functionally interpretable representation spaces. Using appropriate constraints during training, we enforce that interesting functional properties are mapped to disentangled coordinates of representation space. These coordinates can then be manipulated without interfering with other attributes, to design sequences with desired properties. Towards this goal, we will develop new machine learning methods, study fundamental limits of representation learning by statistical physics approaches to neural networks, and work with experimental collaborators in the validation of designed sequences.