We propose improvements to algorithms for gene regulation network inference using gene knockout data. In particular, we attempt to reduce the errors caused due to transitive (indirect) connections on the regulation strength. To address this problem, the matrix of connection strengths (between gene pairs) is first modified by thresholding and normalization with respect to other connection strengths, reducing the impact of indirect connections. This is followed by a cluster normalization step that distinguishes low magnitude direct connections from indirect connections. Our approach combines ranking selection, Gaussian process modeling, and thresholding, thereby enhancing the quality of results obtained with gene regulation network inference. When compared with the top-performing Median Corrected Z scores method, we obtained consistent improvements in the well-known Area Under the Precision Recall Curve (AUPRC) metric, across all ten 100-gene networks from the DREAM3 and DREAM4 contest.