Have a look at the "TIMi Audit report" (the result of the TIMi univariate analysis). The last section of this report contains the "univariate importance" of each variable. For binary targets, the "univariate importance" of a variable X is just the AUC of the lift of the univariate (non-parametric & nonlinear) model built using only the variable X. The best predictive variables have the best models and thus the highest "univariate importance". This "univariate importance" works for both binary targets and continuous targets (but it's really great for binary targets!). For continuous target, the "univariate importance" of a variable X is just the MAPE (mean absolute prediction error) of the univariate (non-parametric & nonlinear) model built using only the variable X. The measure of "quality" of a variable is a lot more "compute intensive" compared to all the others measures included inside Python but it's also a lot more precise/better.
TIMi can compute the audit report on very large datasets. So, your large matrix will NOT be a problem (although it might take some time to compute). To make the computation faster, use a .gel_anatella file to store your dataset.