Existence of Multiple Sequence Alignments (MSA) in Training Process of RFdiffusion Model

[Molaison]

I am currently attempting to fine-tune the RFdiffusion model and working on constructing a dataset. However, I noticed that the model does not require Multiple Sequence Alignment (MSA) at inference time. Do I still need to construct MSA during the training process, or should I solely rely on structural information?