High-throughput protein interactome data: minable or not

There is an emerging trend in post-genome biology to study the collection of thousands of protein interaction pairs (protein interactome) derived from high-throughput experiments. However, high-throughput protein interactome data, especially when derived from the Yeast 2-Hybrid (Y2H) method, have been generally believed to be irreproducible and unreliable, with an estimated high "noise ratio" of more than 50%. In this work, we performed a comprehensive study on approximately 70,000 protein interactions derived from a systematic yeast 2-hybrid (SY2H) method. We performed a comprehensive analysis of biases, reproducibility, statistical significance, and biologically significant patterns in this data set. Surprisingly, we found these protein interactions have a much higher quality. The data represented a comprehensive survey of the entire human proteome with no chromosomal location bias. The reproducibility rate of interactions among replicated searches was quite good, i.e., at 78.5%. The false positive rate, 5.5e-5, was two orders of magnitude better than that reported elsewhere. We further developed several statistical measures and concluded that a protein interaction only needs to appear in two different SY2H searches to become significant. We also developed techniques to show supporting evidence that "promiscuous" protein interactions were not random noises; instead, they could be "network hubs" of the cell signaling network. We also attributed the low noise in our data to the adoption of standard control in the experimental data generation process.