Unfreezing PLM during structure training improves prediction quality

Summary

For structure prediction methods that using language models such as OmegaFold and ESMFold, unfreezing the language model during fine-tuning improves structure prediction TM-score. This was without an Evoformer module which may skew results.

Figures

Method	Embedder Status	Trunk	TMscore
variant1	Freeze	0 Evoformer	0.53
variant2	Trainable	0 Evoformer	0.60
variant3	Freeze	1 Evoformer	0.63
variant4	Trainable	1 Evoformer	0.61

Ref (1)

1.

Lee J, Han K, Kim J, Yu H, Lee Y. Solvent: A Framework for Protein Folding. 2023; Available from: https://arxiv.org/abs/2307.04603