You'd have to figure out how to encode the fitness function into the DNA/environment. We're only scratching the surface of how DNA works in its entirety and are likely decades away from being able to "compile to DNA."
The genetic variability available to the organism can be treated as normalization / quantization of abstract encoding. Map your desired results to this abstract "latent space" and start breeding yeasts. Build environment with various factors' parameters controlled by actuators attached to computer doing ML, maybe even GAs. The environment is the function!
The signal:noise:cost function is prohibitively high.