I'm thinking that you should train a new model. Human and yeast genomes are incredibly, incredibly different. Did you take into account things like genome size, gene density, codon usage, and repetitive sequences? So weights trained on human data are likely biased and could hurt performance on yeast.

Using the old weights for transfer learning is maybe possible if your dataset is tiny and the model learns very generic sequence patterns, but it often has negative transfer. Since yeast datasets are smaller and simpler, training from scratch is probably faster and better.