The advent of genome-wide RNAi-based screens puts us in the position to identify genes for all functions human cells carry out. However, for many functions, assay complexity and cost make genome-scale knockdown experiments impossible. Methods to predict genes required for cell functions are therefore needed to focus RNAi screens from the whole genome to the most likely candidates. While different bioinformatics tools for gene function prediction exist, they lack experimental validation and are therefore rarely used by experimentalists. To address this, we developed an effective computational gene selection strategy that represents public data about genes as graphs and then analyses these graphs using kernels on graph nodes to predict functional relationships. To demonstrate its performance, we predicted human genes required for a poorly understood cellular function, mitotic chromosome condensation, and experimentally validated the top 100 candidates with a focused RNAi screen by automated microscopy. Quantitative analysis of the images demonstrated that the candidates were indeed strongly enriched in condensation genes including the discovery of several new factors. By combining bioinformatics prediction with experimental validation, our study shows that kernels on graph nodes are a powerful tool to integrate public biological data and predict genes involved in cellular functions of interest.