Optimal Twenty Questions with Biased Priors

You are playing 20 questions to identify a number from $\{1, \ldots, 100\}$. If the prior is uniform, binary search is optimal ($\lceil \log_2 100 \rceil = 7$ questions). But if you know the number is $\leq 10$ with probability 0.9 and uniformly distributed on $\{1, \ldots, 100\}$ otherwise, what strategy minimizes expected questions?