Experimental noise cutoff boosts inferability of transcriptional networks in large-scale gene-deletion studies

Generating a comprehensive map of molecular interactions in living cells is difficult and great efforts are undertaken to infer molecular interactions from large-scale perturbation experiments. Here, we develop the analytical and numerical tools to quantify the fundamental limits for inferring transcriptional networks from gene knockout screens and introduce a network inference method that is unbiased with respect to measurement noise and scalable to large network sizes. We show that network asymmetry, knockout coverage and measurement noise are central determinants that limit prediction accuracy, whereas the knowledge about gene-specific variability among biological replicates can be used to eliminate noise-sensitive nodes and thereby boost the performance of network inference algorithms. Reliable inference of gene interactions from perturbation experiments remains a challenge. Here, the authors quantify the upper limits of transcriptional network inference from knockout screens, identify the key determinants of accuracy, and introduce an unbiased and scalable inference algorithm.