There's more than one way, in fact, and the one you want depends on your purpose and theory. Do you want to treat these combinations as if they're entirely separate treatments all their own, as additive combinations of the separate effects of the treatments that make them up, or, thirdly, as both: as additive combinations but with additional effects that are more than the sum of their parts? As Stephan points out, this is a question of interaction effects - those effects, specifically, would be the "more than the sum of their parts" part of their effects, which a model could separate out for you, and test for the presence of.
The most flexible and comprehensive functional form would be to define separate binary variables marking the presence or absence of each treatment letter, and then, additional and separate variables that are the products of those variables, according to the observed combinations. So you'd have one that's A*B
and another that's A*D
and another that's Z*A*B
. You'd treat them all as separate predictors, and the results would give you a comprehensive picture of all these different effects. But if you're comfortable assuming that no such interactions exist, and treating combinations of treatments as if they're just a different treatment of their own, then you can just treat some or all of the combinations as different levels of the single factor variable.
Note though that you have an unordered categorical variable as your predictor, and so it needs to get broken up into binary dummy variables no matter what you do. And effects will be relative to some base category that you should pick for ease of interpretation.