Because GO is a hierarchy, a significant specific term automatically makes its parent terms appear significant too. HSP and SCDP both correct for this redundancy, but in opposite directions. HSP strips the driving genes out of a significant term’s ancestors before testing them — only the most specific term in each branch survives, unless a parent has independent evidence of its own. SCDP does the reverse: it collapses a cluster of significant sibling terms up into the least general ancestor that covers them all. This is why the same term can rank near the top in one method and disappear entirely in the other — the underlying biology hasn’t changed, only how credit for the DE genes is distributed between general and specific terms. A term’s p-value is computed from a different gene set under each scheme, which reorders the results.
Recommendation: For most analyses, HSP is recommended as it surfaces the most specific biological enrichments.