Continuous Variational Synthesis

Biological machine learning is bottlenecked by our ability to synthesize designed DNA sequences. Variational synthesis creates libraries of quadrillions of designed sequences by steering stochastic synthesis using generative models. However, training these variational synthesis models is challenging: constraints on chemical synthesis can force many parameters into a discrete space. In this article we train ``free'' variational synthesis models using stochastic gradient descent in continuous space, and then discretize with a post-training quantization procedure. We show the method enables variational synthesis models to achieve stringent design objectives, while still producing diverse libraries, achieving a strictly dominating quality-diversity Pareto frontier. We demonstrate by designing libraries of peptides, antibody CDRH3s, regulatory DNA elements, and full length enzymes.