The point of this work (well, at least part) is that it doesn't need to build a sequence model (specifically, assemble a sequence model from a multiple sequence

The point of this work (well, at least part) is that it doesn't need to build a sequence model (specifically, assemble a sequence model from a multiple sequence alignment); that model is trained directly into the prediction model itself and so you can skip the MSA step at inference time. It's an improvement over AlphaFold for that reason, although the work has numerous other problems and it doesn't really address any of the challenging bottlenecks in human health.