How can we use ML to estimate these conditional expectations since when using ML some of the variables originally part of our adjustment set might get coefficient = 0 if we use lasso for example?

Question

How can we use ML to estimate these conditional expectations since when using ML (whose objective is prediction and not causal estimation) some of the variables originally part of our adjustment set might get coefficient = 0 if we use lasso for example. What is the rational to use ML here?