This assumption is justified elegantly in Slater and Koster (1954). Essentially, they argue that new orbitals, which are orthogonal and share the symmetry properties of the original orbitals, can be made by linear combinations of non-orthogonal atomic orbitals.