The technological advances underlying the screening of large combinatorial libraries in high-throughput mutational scans deepen our understanding of adaptive protein evolution and boost its applications in protein design. However, many selection signals can be present in one experiment, entangled with the phenotype of interest. I will show how Restricted Boltzmann Machines (RBM), a powerful yet simple generative model from machine learning, can be trained on all the rounds of a sequence screening experiment, capturing an accurate map of the genotype-to-fitness relationship exercised in the experiment. We discuss how to integrate selections with more than one phenotype, and demonstrate that the model can disentangle multiple signals and predict multi-specific variants. The method fits remarkably well several datasets, with robustness to experimental and sampling noise. I will conclude with other applications of RBMs to screening data.