Many aspects of cognition lack neurobiological evidence - consciousness for instance. Now if Cartesian duality is correct (and given its formidable presence in the history of philosophy, it's not a totally insane possibility), it follows that human cognition will never be completely understood within any biological framework. The language faculty can then be one of the 'magical' aspects of cognition, in which case, biological inquiry will be useless. However, this does not preclude the notion that the language faculty employs computational mechanisms like Merge. In light of such possibilities, the computational approach to the language faculty, namely that sentences are formed via Merge, is correct to the extent that we can successfully account for natural language syntax with Merge along with the other assumptions of the minimalist program (e.g. The Strong Minimalist Thesis). I would say that Merge has been very successful in this regard.
On the other hand, it's possible that neurolinguistic research will one day provide a more illuminating account of the language faculty. For instance, perhaps Merge can be localized to Broca's or Wernicke's area. Or as an extreme case, perhaps the language faculty doesn't exist at all, and by extension neither does Merge. That language is nothing more than the result of domain general cognitive processes is a view held by many (e.g. Lakoff and most functionalists). My personal view is that the current generative approach will be vindicated when generative ideas are used to solve the central problem of AI. Once we create robots that speak indistinguishably from humans, using such hated principles of Universal Grammar, generativism's detractors will be at a loss of words.