FitSearch: a robust way to interpret a yeast fitness profile in terms of drug's mode-of-action

Abstract

Background: Yeast deletion-mutant collections have been successfully used to infer the mode-of-action of drugs especially by profiling chemical-genetic and genetic-genetic interactions on a genome-wide scale. Although tens of thousands of those profiles are publicly available, a lack of an accurate method for mining such data has been a major bottleneck for more widespread use of these useful resources.

Results: For general usage of those public resources, we designed FitRankDB as a general repository of fitness profiles, and developed a new search algorithm, FitSearch, for identifying the profiles that have a high similarity score with statistical significance for a given fitness profile. We demonstrated that our new repository and algorithm are highly beneficial to researchers who attempting to make hypotheses based on unknown modes-of-action of bioactive compounds, regardless of the types of experiments that have been performed using yeast deletion-mutant collection in various types of different measurement platforms, especially non-chip-based platforms.

Conclusions: We showed that our new database and algorithm are useful when attempting to construct a hypothesis regarding the unknown function of a bioactive compound through small-scale experiments with a yeast deletion collection in a platform independent manner. The FitRankDB and FitSearch enhance the ease of searching public yeast fitness profiles and obtaining insights into unknown mechanisms of action of drugs. FitSearch is freely available at http://fitsearch.kaist.ac.kr.

Figure 1

Overall scheme of FitSearch. Although researchers have only one or two yeast fitness profiles to their drugs of interest that have unknown toxicity mechanisms, they can easily perform data-mining against tens of thousands of public fitness profiles in order to obtain insight into the mechanism through the FitSearch website (http://fitsearch.kaist.ac.kr). When any type of yeast fitness profile is submitted as a query in the website, a similarity search to other public resources is performed by rank-cutoff optimizer through the FitSearch engine, which is a newly developed method using rank-based overlapping statistics (see the details in the Methods). Since available public resources are deposited in FitRankDB as a general repository for the FitSearch engine (see the details in the Methods), the similarity search can be performed more efficiently, thoroughly, and rapidly in the FitSearch website. Finally, users scrutinize characteristics of a list of drugs similar to their drug of interest and obtain clues or plausible hypotheses, which could also help them to design further bioassays.

Figure 2

Toy example showing how the rank-cutoff optimizer works. (A) Ranks of each strain in virtual two query and target yeast fitness profiles to be compared are supposed to be deposited in Fit-RankDB. These profiles are also supposed to be generated using a virtual yeast deletion library comprising strain a to j. (B) Efficient calculation of a match number (or an overlapped strain number) accumulated under all possible rank-cutoffs of the query and the target by Dynamic programming (see the details in the Methods). For this calculation, first, rank matches of each strain should be expressed as the match matrix (M). In the M matrix, its row represents 'ranks in the query', its column 'ranks in the target', and its value 'the strain number with same rank in the query and the target'. Then, the current accumulated match number (in red-colored cell in the A matrix) is calculated by adding the current match number (in the orange-colored cell in the M matrix) to the previous accumulated match number (sky-colored cell plus purple-colored cell minus gray-colored cell in the A matrix). In this way, the accumulated match numbers regarding to all possible rank-cutoffs are efficiently calculated and stored in the A matrix. (C) The matrix of cumulative hyper-geometric p-values (P) is filled by calculating the equation (2) as the objective function (Hp) regarding to all possible rank-cutoffs, and used to find the rank-cutoffs with the minimized p-value as described in the equation (3), called optimal rank-cutoffs. The A matrix provides all of the parameters needed for equations (2) and (3) as follows: Its values represent the overlapped strain number in the equation (2); its row-names, the query strain number; its column-names, the target strain number in their respective rank-cutoffs; and its column or row length, the size of population. When the maximal rank-cutoff is set to 10 in the toy example, the query rank-cutoff 5 and the target rank-cutoff 5 shows the minimal p-value, 0.004. At those optimal rank-cutoffs, overlapping significance (hyper-geometric p-value) and overlapping score (Tanimoto coefficients) can be expressed as the similarity between the query and the target.

Figure 3

Plot of an overlapping score and an overlapping significance as two-way cutoffs to show the most similar chemical or genetic effects to a query's effect. (A) Two-way cutoff plot of the most similar chemical effects to the 5-Fluorouracil's effect. (B) Two-way cutoff plot of similar chemical effects to clotrimazole's effect. (C) Two-way cutoff plot of the most similar genetic effects to clotrimazole's effect. (D) Two-way cutoff plot of the most similar chemical effects to DAPG's effect. Target sources mean public chemical-genetic or genetic-genetic yeast profiles.