Crucially, it depends on jointly modifying the experimental protocol and the training algorithm: on their own, neither modification helps.

This approach lets you focus limited measurements on the most informative datapoints, maximizing information gain without compromising reliability.

Second, modify the training algorithm: compensate for the missing negatives by incorporating the generative variational synthesis model into the objective.

To test, we can deliver billions of designs to different cells. But there is a cost to recovering those designs' function, to obtain (x,y) data.

With variational synthesis, we can now build quadrillions of generative model-designed sequences. The bottleneck is now testing, not synthesis.

We're excited to present LeaVS, a method to scale up learning for protein function models. It is based on the co-design of wet lab experiments and in silico training.