We study the evoked responses of recurrent neural networks composed of multiple sub-populations of neurons ("cell-types") and a cell-type-specific connectivity rules. Recently we characterized the autonomous dynamics of such networks (Aljadeff, Stern, Sharpee, 2015). Significantly, the block structure of the connectivity matrix diverts the point of bifurcation from the quiescent state to a chaotic state, compared with networks with the same average connectivity strength but without structure (i.e. networks with a single cell-type). Furthermore, the autocorrelation structure of the global dynamics was more rich in the case of multiple sub-populations, compared with networks with a single population. These findings suggest that evoked responses would also depend on the cell-type-specific structure of the network, and not only on the average connectivity as in networks with a single cell-type. Previously it was shown for a network with a single population (and hence a single connectivity rule) that an the irregular spontaneous activity can be suppressed by a small amplitude input as long as its frequency is tuned to a band that depends on the average network gain (Rajan, Sompolinsky, Abbott, 2010). When the network is entrained by the stimulus, the intrinsic spontaneous dynamics are replaced with predictable correlated activity. In the network with multiple sub-populations, suppression of the spontaneous ongoing activity by an input is also observed. However the frequency to which the network is sensitive is now tuned by the parameters of the network: the size of each sub-population and the connectivity gain between every pair of sub-populations. If the connectivity gain in isolated regions of the network changes on a timescale slower than that of the spontaneous activity, for example through the effects of neuromodulation or neuro-glia interactions, this will change the properties of evoked responses in the entire network. The model studied by Rajan et al. is thought to be a step towards understanding the neural basis of stimulus and state-dependent attention. A limitation of this interpretation is that the stimulus that reduces the variability in the network (the "attended stimulus" that lies in a specific frequency band) is modified only if there are global changes to the gain in the network. By extending these results to networks with multiple sub-populations and characterizing their response properties we can develop an understanding of how local changes in a network can modulate the stimulus-dependent suppression of variability, and therefore serve as a more realistic model of the neural basis of attention.