Beyond these methodological questions, I think that the Mendelian Randomisation approach works best for situations where the effect of some exposure is chronic

Beyond these methodological questions, I think that the Mendelian Randomisation approach works best for situations where the effect of some exposure is chronic eg LDL and atherosclerosis, particularly where the disease process may begin in childhood or adolescence. For an acute infection I think it is a blunt instrument at best, as there is going to be a lot more noise in the inferred target variable.