Algorithm 1: 1D-SRA

input:   Feature data X (N x P)      Target class Y (N x 1) constants: P = 11,915,233 // total number of features (SNPs)      S = 250 // reduced number of features      K = 47,660 // number of reduced models step A1: generate reduced data sets and assess performance of reduced models shuffle the vector of all P features for each reduced model k in K // over reduced models   randomly sample S features without replacement from the   shuffled vector   fit a multinomial logistic regression model to N individuals   described by S features (dependent variable–target class   independent variable–features)   collect model_performance_k = cross-entropy loss   for each s in S // over features     feature_performance_k_s = max(SNP effect estimate)     feature_index_k_s = feature Id from feature vector   end for end for step A2: generate model performance matrix C generate a model performance matrix of K rows and P + 1 columns filled with zeros for each k in K // over reduced models   C[k,P + 1] = model_performance_k   for each s in S // over features     C[k,feature_index_k_s] = feature_performance_k_s   end for end for step A3: supervised rank aggregation based on matrix C run LMM (dependent variable–model performance C[:,P + 1], independent variable–feature performance C[:,1:P] collect 1:P vector of LMM parameter estimates step A4: feature selection based on LMM parameter estimates define 2 clusters of LMM parameter estimates based on 1D-K-means clustering