I understand linkage disequilibrium (LD) versus panmixia, which the technical term for Mendelian randomisation and for humans PLINK is a population package. The best outcome is a correlation between LD and cardiovascular disease. It is however an unlikely outcome.
This really is a machine learning question and you would need to understand "feature selection" and will require very high accuracy scores >0.95 - however you could triple the size of our control group quite easily (recommended). This is a separate question and I'm happy to answer this there.
Just to address the comments. Inheritance probability, if that is the intended statistics, I personally don't think it would work because a 125 sample is low and the patients would need to be in cohorts of familial generations. LD is unlikely the explanation with humans but other non-humans species that might work (selection advantage).
My personal view is this is a straight machine learning calculation, but it's a separate question. If your sample is heavily stratified across each familial generation then Hardy-Weinberg might work, but my experience of patient cohorts is that is really unusual and there could still be confounders. Supervised learning (subset of ML) is 100% the way forward.
.... possibly use GWAS, but you've multiple features/training targets. GWAS is usually when there's one phenotype, ML would account for interaction in a way GWAS could not (it couldn't account for non-sequence features, i.e. mixed data).