You don't.
The native Java driver (https://github.com/datastax/java-driver) addresses this by making everything asynchronous, so you can easily request appropriate columns from a bunch of rows, then wait for all the results. (This has no more overhead than a multiget.)
That said, if you can denormalize into a separate table instead of doing a multiget, that's probably best of all.