My guess is that this area is much harder to break into–enzymes facilitate challenging chemical transformations by stabilizing high-energy transition states in chemical reactions. These states are usually highly transient and therefore much harder to capture using the structural biology techniques that generate the structural data that AlphaFold and similar methods are trained on. Even though there are many structures of enzymes in the absence of their substrate, I would imagine that the small number of structures for states that represent actual catalytic intermediates would make it difficult for a model to internalize the features that distinguish a good enzyme/catalyst from a bad one.

Another consideration is that most protein structure prediction methods only generate the backbone, and the sidechains are modeled in afterwards. Enzyme efficiency requires sub-A level structural precision in the sidechains that are actually doing the chemistry involved in catalysis, so it could also be the case that the current backbone-centric methods aren't good enough to predict these fine-tuned interactions.