A model learns patterns from the examples it was trained on, and it has no way to know when a new input falls outside that experience. This is the single most important practical limit on everything described so far.
Consider activity prediction. Suppose a model is trained on measured interactions between compounds and a well-studied family of proteins. It will perform well on new compounds against those proteins. Now give it a protein from a family it has never seen. The model still returns a number — it always returns a number — but that number is an extrapolation from unrelated chemistry, and it can be confidently wrong.
The same failure appears with bias rather than sparsity. If the training set contains mostly compounds from one chemical series, the model learns that series' regularities and treats them as general rules. Novel scaffolds that fall outside the training distribution get scored as if they were familiar, and the error is not flagged.
There is a second, subtler problem: the data records what was measured, not what is true. Negative results are under-reported, so a model trained on published data sees a world where most tested compounds worked. That skews its sense of what is normal.
The practical takeaway is not that these models are useless. It is that a prediction is only as trustworthy as the overlap between the new case and the training data, and that overlap is usually invisible from the output alone.