I am struggling currently with constructing a binary glm predictive classifier, due to an issue with dimensionality.

I have a dataset of N samples where each sample has values for M entries (genes) and is either Case or Control. However, each gene consists of 1 … k datapoints for each sample, where each datapoint is a numerical value ranging from 0 to 1. And the number of datapoints varies in turn for each gene (thus meaning that each gene is represented by a dataframe of varying size). The classifier is supposed to be binary, assigning samples as Case/Control based on the values for a set of predictors (genes) which are chosen through cross-validation.

I have previously constructed binary predictive classifiers where one value for each sample was representative of a gene (i.e. from a NxM dataframe), but i am stuck on how to handle the current scenario where each 1…M column in my NxM dataframe could be said to actually be a NxK dataframe rather then a Nx1 column.

I have thought to represent the NxK dataframe as a Nx1 column through taking the median value out of the 1…k datapoints for each of the N samples, or by choosing one column in the NxK dataframe to represent the dataframe (based on selection critera that is fitting for my research question), but ideally I would like to point towards the NxK dataframe representing the gene, when constructing the classifier.

Does anyone have experience with this type of problem, and if so is it even possible to construct a predictive classifier from a dataframe of predictors, where the predictors in turn point towards separate dataframes?
If not, is my approach of taking the median value out of the 1…k datapoints for each sample a valid approach?

I have previously constructed classifiers where each gene was represented by one value, however i have no experience with this type of problem.