A model's accuracy is bounded by its training data in three separate ways, and it is worth keeping them apart.
Quantity: patterns that appear in a handful of examples may be coincidence. With enough examples, the model can distinguish a real regularity from noise. Cancer datasets are often small in exactly the categories that matter most — rare tumor types, unusual patient populations — so the model is weakest where the need is greatest.
Quality: labels are only as trustworthy as the measurement behind them. If activity was assayed under inconsistent lab conditions, or if patient records code the same outcome in different ways, the model learns those inconsistencies as if they were biology.
Representativeness: a model trained mostly on one population or one cancer subtype will perform poorly outside it. This is the bias problem in concrete form — not a moral abstraction, but a measurable drop in accuracy when the model meets patients who were never in its training set.
A useful way to hold this together: the data defines the ceiling, and no amount of clever modeling raises it.