The cool paper is:
Systematic Genetic Analysis with Ordered Arrays of Yeast Deletion Mutants
Amy Hin Yan Tong, Marie Evangelista, Ainslie B. Parsons, Hong Xu, Gary D. Bader, Nicholas Page, Mark Robinson, Sasan Raghibizade, Christopher W. V. Hogue, Howard Bussey, Brenda Andrews, Mike Tyers, Charles Boone
Science 294: 2364-2368 (2001)

It's really the startling statistic that opens the article: "In Saccharomyces cerevisiae, more than 80% of the ~6200 predicted genes are nonessential".

Non-essential, in this context, means that you can grow the yeast in the lab, when that particular gene is deleted; the yeast may not be happy, and may require special care and tenderness, but it'll survive.

It's not surprising that a lot of genes are non-essential, because the more important a gene function is, the more likely it is to be redundant -- or, as the article says, "the genome is buffered from the phenotypic consequences of genetic perturbation"; but (to me, anyway) the scale of the redundancy is surprising.

So does this mean you could knock out 5000 genes from S. cervisiae and it= would keep going? No, because you start cutting out the redundancy; if you've knocked out one gene, and the yeast is flying on one engine, then cutting off the remaining engine will kill the yeast (even though cutting out engine two would not, in itself, be lethal).

But that observation is useful in itself. Even though the genome has been sequenced, only a small fraction of gene functions are known. It's a lot easier to run a DNA sequence through a computer and say, 'There's a gene there,' than it is to say, 'That gene does this'. On the other hand, proteins (the products of genes, of course) don't operate in a vacuum, they interact with each other and work in cascades and pathways.

So let's say you have a gene whose function is known, but which is non-essential. Take the yeast with the gene knocked out. Breed it with an array of yeast with single, non-essential, genes knocked out. And ask which of the double mutants survive. If the genes are networked, or redundant, then the double mutants may be non-viable; if the genes are unrelated, then the double mutants should be no worse off than either single mutant.

So that's what these guys did, on a grand scale. They took a single-gene deletion and crossed it with each of 4700 other single-gene deletions (almost, but not quite, every non-essential gene), and used this to identify "a network of 291 interactions among 204 genes"; some of which were known, some were not known but not surprising, and some were completely unexpected.

It's actually a fairly simple approach--they used pretty standard techniques, the single-gene mutants are (I think) generally available, and the only difficulty is scaling up the system to deal with the thousands of individual mutants and double mutants; which can be automated without too much problem. (These guys don't go into detail on their system, but they thank some people for 'advice on the robotic manipulation of yeast arrays', and one of the authors is from an engineering firm.)

They predict, somehow, that you can use this approach to scan the genome quite efficiently: "... we estimate that on the order of 300 SGA screens covering judiciously selected query genes will provide an effective working genetic scaffold, which should reveal many of the molecular mechanisms behind genetic robustness and buffering". And they note a fairly obvious point, that "because gene function is often highly conserved, a comprehensive functional genetic map of S. cerevisiae will provide a template to understand the relationships among analogous pathways in metazoans": That is, the information derived from the yeast genome will probably be applicable to humans as well. And finally, they note that it's possible to use a variant of this to address even the remaining 20% of genes that are essential.