It seems really useful from a biologist standpoint. I don't know anything about protein folding myself though.

From a modeling perspective, I would ask some of the questions below. The article simply leaves me wondering more about the AI modeling.

1. What is the scope of the training data and the model's predictive capability? Meaning, can it predict all sorts of diseases shapes or is it focused on a specific disease?

2. What further data collection is assumed to be needed to predict the behavior of the "all" the diseases of the world?

3. What type of learning model is used?

4. How long did training take? How long is it expected to take to be updated with additional information? When providing more training data, would the learning model start anew or would it tweak a few hyperparameters?

5. How much manual massaging do the hyperparameters need? Being AI, I'm assuming its fully automated, but would just want to ask the question to make sure.

6. What is the false positive rate and the false negative rate? Was this expected? What are the risks associated with being a false positive or negative?

7. How applicable is the model to other fields such as weather, computational fluid dynamics, etc.?

It sounds like a very complex problem and really cool progress for the field.