Two-Stage Fine-Tuning Generates Proteins with Targeted Amino-Acid Composition.

Generating protein sequences with precise amino-acid compositions is a significant challenge in biological design, particularly for applications like synthetic feed protein where nutritional value is directly linked to composition. Existing protein language models are powerful priors but lack mechanisms to steer generation towards explicit compositional targets. Researchers have developed a novel two-stage fine-tuning pipeline to address this. The first stage involves domain-adaptive fine-tuning (FT) of a protein language model on a relevant in-domain protein dataset. This initial step brings the average amino-acid composition of generated sequences closer to the desired target. The second stage then employs iterative reward-weighted fine-tuning, utilizing reinforcement learning (RL) and anchoring against the previously fine-tuned model as a stable reference. Evaluations on two distinct amino-acid compositions demonstrate that while the initial FT stage provides a good approximation, the subsequent RL stage is crucial for enforcing specific sequence constraints that FT alone cannot satisfy. The study also explores various design choices for the composition reward term, confirming that this pipeline effectively aligns amino-acid composition without compromising the overall quality or diversity of the generated protein sequences.