> The team fine-tuned the network by training it on thousands of experimentally determined structures of antibodies attached to their targets

So that should explain the relatively low yield.

To avoid the word "AI", the researchers took a thought-to-be universal model and ran a parameter estimation for a very complex 3D (bio-)chemical interaction model.

Of course this approach will work with enough samples and a model that is universal enough to express the problem solution. But they simply had way fewer samples than, say, a large language model.

I think this approach won't carry that far because of this lack of training data. It might give us some exciting results, but as long as manufacturing of antibodies and testing them takes a (comparatively) long time and huge effort, the model will be much less efficient than a classical algorithmic simulation. But of course, there's a chance that it provides insights into improving such "boring" models.