What would be enough? 400 patients, 4000, 40k, 400k, 4M?

Well, reading the study, I'm not sure more patients could rescue it from methodological bias. They assumed the premise basically -- we should find a biomarker, which is kind of what this thread is discussing. Then they went trawling for biomarker in a sea of millions of biomarkers. They did this by training an model that produced the desired result, using a grid search for hyper parameters that even further expanded the available degrees of freedom here beyond what they had from the biology. No pre-registration; There are millions of places where the researchers could have made a different decision -- would they still have gotten a publishable result? Oh plus the authors mostly work for the company whose data they use, which is hoping to sell a diagnostic test.

I'm giving you a thorough response because I'm detecting a cavalier anti scientism which I think is sadly becoming more common. This stuff is hard; are you sure you understand it enough to have an informed opinion?