Summary
Sequences designed by protein language models have the same distribution of pLDDT values, sequence lengths, and in vitro activity as naturally occurring sequences (1). This was measured in amylases and can be overcome with fine-tuning using direct preference optimization.
Figures
Ref (1)
1.
Stocco F, Artigues-Lleixa M, Hunklinger A, Widatalla T, Guell M, Ferruz N. Guiding Generative Protein Language Models with Reinforcement Learning. 2024; Available from: https://arxiv.org/abs/2412.12979