There is another trap that it is easy to fall into when deciding to use computer models. An obvious way to use modeling is to construct a model that incorporates some particular hypothesis and do "computer experiments" on the model in order to see if it behaves like experiments on the biological system. If it does, you might claim that this is evidence in favor of your hypothesis. The trouble with this is that it suffers from some of the same problems as abstract models: the simulation may be just giving the results that it was designed to give, and you don't know if a different model might have done just as well. With enough assumptions and tweaking of parameters, there are lots of models that could generate reasonable agreement with experiment.
A better approach is to try to ignore your preconceived ideas about the cause of a particular behavior and try to build the best model of the system you are studying that you can. This means incorporating the best physiological data that you can get to model the system in detail. This often means modeling neurons down to fine details of dendritic structure, and modeling each kind of ion channel that is known to exist in the cell. This fosters a close relationship with experiment, because you will soon discover experiments that you need to do in order to get data to use to characterize some channel or membrane parameter.
Once you have done this, you often find that you have to fill in the gaps in your knowledge with some hypotheses. You might postulate some connections between neurons which you feel might be necessary; or the existence of some interneuron in order to provide a needed inhibitory input. Or, you might assume the existence of some type of ionic channel which has been observed in other types of neurons and which seem necessary to explain the behavior of the one you want to model. Then you use your simulation as a sort of breadboard to test out your ideas. If the new feature gives better agreement with experiment, then you have the motivation to perform experiments to see if it really exists.
Of course, there will always be some parameters that you have to estimate. Then you compare the behavior of the model with some "function-neutral" experiments like voltage or current clamp on a cell, or perhaps an electric shock to a nerve in a system. You can then "tune" the model by fitting any unknown parameters with data that isn't directly related to the behavior of interest. If the model passes these tests, then you can have more confidence in it. At this point you can start exploring the more interesting behavior of the model and understanding how the model produces this behavior. You can perform measurements and experiments on your model which might be impossible in the real system. Everything is accessible in a simulation. You can try to simplify your model in order to find out what features are important for the interesting behavior of the system - and which features are merely "icing on the cake".
You may discover that not only does the model behave like the biological system, but that the mechanisms that you postulated are causing that behavior. But it is just as interesting to find that there is another cause, or even to see behavior in the model that you never thought to look for. This can be another useful way to guide the direction of experiments. This kind of realistic modeling opens up a new way to discover, through "computer experiments" the way that neurons process information.