I mean it kind of makes sense if viewed like deep learning. If we train a system on one set of data, the virus itself, it will probably learn to recognize similar data. If we then highlight one specific part that's very special in this data, the spike protein, then the system will recognize the pattern much more easily. Dataset I or II alone would both let the system recognize the pattern, but not be as effective, since it's harder to tell what's important and what's not.