Training with linkers

[Amelie-Schreiber]

I was just wondering if perhaps training with linker tuning or by inpainting intrinsically disordered regions (IDRs) using EvoDiff might help with data quality. Using linker tuning as was done in Linker-Tuning: Optimizing Continuous Prompts for Heterodimeric Protein Prediction and showed improved performance as linkers got progressively longer (up to a length of 25 residues). Alternatively, if you inpaint IDR linkers using EvoDiff this might provide the model with some additional context that could improve performance and may be better than simple G-linkers. Thoughts?