Global Optimization of the Sequence-to-Function Network

The complexity of the genome gives rise to global optimization problems that test the limits of our computational power.  One such problem is the inference of isoform function.

Isoform function prediction is the problem of deciding which biological functions are performed by specific gene products.  The lack of labeled training data makes an unsupervised approach that shares information between the gene products highly desirable.  We developed a parallelized Expectation-Maximization algorithm that makes use of mini-batches as an effective strategy for tackling the computational complexity of the problem and the very large input size.  The algorithm optimizes a model that jointly predicts the sequence similarity of all pairs of isoforms given the number of shared functions.