Clojure arithmetic is fast enough if you add optional type annotations or otherwise take advantage of unboxed values. Ordinary calculation-intense code will be hundreds of times faster than in Python which might be enough in most cases.
If you need deeper support for vector math, scientific calculations, and the like then Incanter is probably what you want. It does all the complicated fast math stuff that always benefits from carefully planned and hand-tuned libraries.
To be clear... the benefit of Scipy / Numpy / Matplotlib that I'm alluding to is not (necessarily) runtime efficiency -- though, given that many of the linked libraries are standards (eg. LAPACK) and compiled C / Fortran, it is efficient for many operations.
The benefits to scientists are the developer efficiency (rapid prototyping), comprehensive libraries (scientific functions and plotting), interfaces similar to MatLab (eg. slicing), and its standing in the scientific community. It would sort of defeat the purpose to spend valuable research time to reinvent the wheel in another language -- though I would switch to a LISP variant in a heartbeat if there was comparably functionality.
If you need deeper support for vector math, scientific calculations, and the like then Incanter is probably what you want. It does all the complicated fast math stuff that always benefits from carefully planned and hand-tuned libraries.