We can define a roadmap and what we want to achieve.
I suggest we start from VariPred and increase the modality and also tasks for it. After achieving good results there we move to learn joint embedding for protein's different modalities as in ImageBind which will be a very high impact.